Contents Introduction
While wind power production keeps rising worldwide, wind turbines are playing an increasingly major role in the present and future of renewable power generation. Yet, in the current wind production landscape, two trends seem to jeopardize the fulfillment of this global role. On the one hand, a significant share of the existing wind turbines has already reached its 20-year estimated lifetime, which requires additional maintenance services; on the other hand, new wind turbine technology is evolving toward larger wind turbines in remote offshore locations, which poses new accessibility issues for inspection and maintenance. Consequently, greater attention has recently been paid to the soaring operation and maintenance (O&M) costs of wind turbines (both onshore and offshore), which must be addressed for wind energy to remain financially viable [1]–[3]. From a technical point of view, wind turbines are complex aeroelectromechanical systems that are becoming larger and more flexible in design and increasingly digitalized and automated in operation. Unlike conventional power plants, they are often installed in remote locations and, thus, constantly undergo harsh weather conditions and largely variable aerodynamic, gravitational, centrifugal, and gyroscopic loads. Taken together, these elements induce a high frequency of faults and failures in wind turbines. In addition, access for maintenance is often troublesome and costly due not only to the large dimension of wind turbines but also their safety regulations, which restricts maintenance service to daytime hours, a relatively low air humidity, temperatures above 10 °C, and wind speeds below 8–12 m/s [4]. Such issues are exacerbated in offshore wind farms due to the harsh marine environment, which causes higher rates of failures in wind turbine components and additional maintenance complexities, whether in terms of difficult accessibility, higher logistic costs, or lower skilled manpower [5], [6]. As a result of poor equipment reliability and impaired wind turbine availability for power generation (unscheduled downtime/stoppages), O&M costs are quickly rising, now representing an increasingly significant part of the total wind energy generation cost [3], [7]–[9]. Currently, the O&M costs account for approximately 10%–30% of the total energy generation cost of an onshore wind farm after it has become operational [10]. In comparison, although offshore wind farms can generate wind energy more efficiently, their O&M costs can surge up to 25%–50% of the total energy generation cost, which is a considerable increase [5], [11]. Lowering the ever-soaring O&M costs in wind energy production requires improving the critical aspects of wind turbine reliability and availability (uptime) using appropriate condition monitoring solutions.
A. Wind Turbine Condition Monitoring
Wind turbine condition monitoring involves the process of monitoring (and analyzing) the operating parameter(s) of condition in a wind turbine system or its components, aiming at identifying any abnormal change(s) in the condition or specific events that can indicate developing fault(s) in the system. Using the monitored change(s) in the parameter(s), the fault diagnosis (detection, isolation, and identification) and lifetime prognosis (LTP) (remaining useful life (RUL) estimation) can be accomplished before a failure or a critical malfunction occurs in the wind turbine. This allows a very cost-effective and optimal type of preventive maintenance (before a failure), which is often referred to as “condition-based maintenance” instead of resorting to costly time-based maintenance (fixed service intervals).
Currently, wind turbine condition monitoring can be performed either offline or online (in real time). Off-line condition monitoring involves periodic inspections in which a wind turbine needs to be shut down and/or requires the attention of an operator. This condition monitoring method suffers from important shortcomings. It is, indeed, an expensive monitoring method since it requires the wind turbine to stop working, which results in energy production loss, not to mention the additional costs incurred during off-line inspections. Moreover, off-line condition monitoring obviously falls short of detecting and reporting the faults or damages that happen between the inspection intervals. The mentioned shortcomings become increasingly serious, especially in offshore wind farms for which the inspection intervals are longer than those of onshore wind farms [12], [13]. Therefore, real-time (online) condition monitoring techniques are rapidly evolving to address the issues that come with off-line techniques. Indeed, real-time condition monitoring continuously provides a deeper insight into the real-time health status of a wind turbine or its main components, while it is in service. It relies on appropriate sensors or data acquisition systems to obtain continuous raw measurements (e.g., vibration, strain, temperature, voltage, and current) from wind turbine components and may even incorporate onboard models and/or signal processing functionalities for enhanced data reduction and analysis. Thanks to the wide variety of sensors and data acquisition systems available for wind turbine components, real-time condition monitoring is highly customizable. Factors such as the wind farm’s environment (onshore/offshore, cold/hot, humid/dry, and so on), the size and design specifications of its wind turbines, the types, and characteristics of target components to be monitored, together with the capabilities, limitations, and costs of available real-time monitoring technologies, should all be considered when choosing a real-time condition monitoring system (CMS).
Given the significance, plurality, and evolution of real-time condition monitoring for wind turbines, this article introduces and discusses the details of well-known technologies and solutions in this area and provides an up-to-date comprehensive review of the available literature, covering both technical aspects of fault diagnosis and LTP in wind turbines.
B. Bibliographical Status of Literature and Existing Reviews/Surveys
Wind turbine condition monitoring, including both aspects of fault diagnosis and prognosis, has become an active research area over the last two decades. Fig. 1 shows the number and trend of published articles in this area from 2000 to 2020. It should be noted that this information is extracted from the Web of Science (WOS), which only includes a part of the available literature in this area. Therefore, the number of the entire publications on wind turbine condition monitoring, including articles, conference proceedings, reports, press releases, and so on, is, indeed, much higher than shown in the graphic in Fig. 1. Having said that, the reliable information extracted from the WOS and reflected in this graphic clearly indicates an overall rising trend for research articles on wind turbine condition monitoring by the end of 2020.
Number and trend of published articles on wind turbine condition monitoring indexed in the WOS since 2000 (as of the beginning of 2021). Source: WOS.
Given the rapid growth of research activities in this area, several literature reviews or survey papers have already been published on the different aspects of condition monitoring in wind turbines. For instance, Table 1 provides a chronological list of the existing review/survey journal publications in this area. Many of the reviews/surveys listed in the table are excellent in some facets and include very useful details. However, according to the indicated focus category and the brief description of each publication listed in Table 1, the following conclusions can be drawn regarding the existing reviews/surveys.
Some of them are relatively old, going back to 2006, which obviously obliterates the great advancement of wind turbine condition monitoring technology ever since (see Fig. 1).
Some are mainly focused on the monitoring of a particular wind turbine component, such as blade or gearbox, while a more comprehensive review/survey would provide broader knowledge of the appropriateness of technology, including its capabilities, limitations, and cost for monitoring different components at the same time.
Most of these reviews/surveys focus either only on the diagnosis or on the prognosis of wind turbines, not both together, which would enable a broader, more comprehensive perspective in the literature.
Those that discuss both aspects of diagnosis and prognosis in wind turbines are still limited in their attempts to offer a meaningful and comprehensive review of every available technique or solution, especially when it comes to prognosis.
The condition monitoring techniques reviewed are often restricted to those using either the signal-based (data-driven) approaches or mathematical model-based approaches but not both approaches in a broader perspective, nor using the hybrid methodologies that emerged out of the combination of the mentioned approaches.
The wind turbine nondestructive condition monitoring technologies (such as machine vision, ultrasound, and thermography) are often ignored, or, if reviewed, there is a tendency to only focus on a specific wind turbine component, such as blades, while providing little context about which component a technology is suited for, and with scarce details on capabilities, limitations, and costs.
In addition, as it came to the attention of the authors while reviewing the literature, the cited references in some of the existing reviews/surveys in this area are not always necessarily related to the problem of condition monitoring in wind turbines. For instance, some of the techniques or results reported for the condition monitoring of bearings or generators in engineering applications other than wind turbines were mistakenly cited and classified among the techniques specifically developed for wind turbines. This is not necessarily correct or reasonable since the problem of condition monitoring in wind turbines involves its own particular aspects and challenges that differ from those found in other engineering applications since they originate from the specific operating conditions and load variations of a wind turbine. Having said that, another problem with the references of some of the existing reviews/surveys is that, despite the new results and publications constantly emerging in this active research area, some recent reviews/surveys keep citing the same old or even obsolete references invocated in former reviews or surveys for years. This shortcoming has been carefully avoided in this review paper by citing the most important and relevant research papers with particular attention given to the recently published results, that is, in the last ten years.
The shortcomings and loopholes found in the existing literature reviews, along with the constant advancement of wind turbine condition monitoring technology that has inspired a considerable amount of new publications in this area, all together motivate the authors to present an updated and more comprehensive literature review in this article.
C. Overview of the Present Comprehensive Review
This article aims at providing an up-to-date comprehensive literature review on real-time condition monitoring of wind turbines, spanning both fault diagnosis and prognosis aspects while exploring signal- and model-based approaches as well. Furthermore, the common fault and failure modes along with their root causes, traced to different wind turbine components, are discussed and categorized based on their severity; likewise, the updated results of recent reliability studies on both onshore and offshore wind turbines are carefully organized and presented. Compared with the existing literature reviews/surveys, the main features and contributions of this review paper specifically include the following.
A simple, straightforward language along with many informative and easy-to-understand figures, schematics, and tables is used to provide a systematic, comprehensive literature review in a way that is both appropriate and accessible to students, researchers, and any other practitioners, whether from academia or industry.
Both aspects of fault detection and diagnosis (FDD) and LTP are considered and reviewed for a deeper insight into the conditions of wind turbine components and subsystems.
All critical components of wind turbines are considered, and the advantages, disadvantages, limitations, challenges, costs, and trends of each condition monitoring technique or technology are carefully reviewed and explicitly discussed.
Given the in-depth literature in this field, and since a considerable share of studies has been published in the past ten years (see Fig. 1), the main focus is placed on the most recent journal publications, namely, those produced approximately between 2010 and 2021, unless the citation of some other references or recourses is utterly relevant to the area reviewed.
The most important, relevant, and up-to-date research results are carefully selected and cited to avoid the invocation of the same old references that used to be mentioned repeatedly in some of the existing reviews or surveys, disregarding the constant emergence of new results and publications in this active research area. That being said, this review paper explores various well-known and impactful papers that have significantly contributed to the development of this area since the 1990s.
A larger number of references are analyzed in order to achieve a more comprehensive literature review spanning an extended number of available techniques and technologies that include the SCADA-based techniques, 12 different varieties of dedicated condition-specific-based techniques (e.g., oil quality, vibration, and acoustic emission), mathematical model-based techniques, and the hybrid techniques as the combination of other techniques.
Throughout the review, the cited references are carefully selected and categorized with a special attention given to their true contents rather than the simple reformulation of their abstracts.
The cited references categorized in each table are sorted according to their publication dates to better represent the evolutive trend of a technology or approach over time and with respect to the monitored components.
The remainder of this article is organized into six different sections. The cutting edge of wind energy and wind turbines, with an emphasis on monitoring and control subsystems in wind turbines, is briefly discussed in Section II. The faults and failure modes, along with their root causes in different wind turbine components, are detailed and categorized in Section III. Wind turbine condition monitoring and condition-based maintenance techniques are described in Section IV. Hardware signal-based condition monitoring, including SCADA-based techniques and dedicated condition-specific-based techniques, is reviewed in Section V. Mathematical model-based condition monitoring is reviewed in Section VI. Finally, the summary, concluding remarks, and future trends are outlined in Section VII.
State of the Art of Wind Turbines
A. Wind Energy and Wind Turbines
In recent years, the demand for renewable/green energy has significantly increased. Among the different types of renewable energies, wind energy is the cleanest, a factor that contributes to making it the world’s fastest-growing renewable energy source. Presenting the global cumulative installed wind power capacity from 2000 to 2020, Fig. 2 shows that, globally, the cumulative installed wind power capacity reached nearly 744 GW in 2020, which amounts to about 7% of the global electricity demand [50]. Following that trend, the mentioned wind power contribution is expected to increase up to 25%–30% by 2050 according to the Global Wind Energy Council [51].
Trend of the global cumulative installed wind power capacity 2000–2020. Source: The World Wind Energy Association (WWEA), March 2021.
As their names indicate, wind turbines harness the power of the wind. They do so by converting the wind’s kinetic energy into mechanical energy, which is, in turn, converted into electricity using a generator. From a design viewpoint, wind turbines are classified into two different types: the vertical axis and the horizontal axis (see Fig. 3). Vertical axis wind turbines [see Fig. 3(a)] have a set of rotor blades that spin around a vertical axis, while other major components are located at the base of the wind turbine, which facilitates maintenance services. However, compared with horizontal axis wind turbines, they are known to be less efficient and are associated with higher O&M costs. This is partly due to the low rotor height that cannot harness greater wind speeds often found at higher levels, the less efficient rotor design that prevents the blades on the vertical axis rotor from receiving incoming wind at the same time, and more wind turbulence and structure vibrations that altogether cause higher component wear-down. On the contrary, with three rotor blades and presenting numerous advantages, such as access to the stronger wind (thanks to their tall towers), higher efficiency (since the blades always move perpendicularly to the wind), and receiving power through the whole rotation, the horizontal axis wind turbine [see Fig. 3(b)] has gradually dominated the commercial market of wind energy [52]. Given that, in Section II-B provides more details about the components and subsystems of this common type of wind turbine.
B. Wind Turbine Components and Subsystems
A cross-sectional view of a typical three-bladed horizontal-axis wind turbine is shown in Fig. 4. The figure depicts the most important wind turbine components and briefly describes their functionalities. Although absent from this figure, a large and strong foundation always finds itself under the wind turbine tower to withstand all the forces from the wind and hold the turbine upright. Section III provides detailed information about the significance, failure rates, and general reliability aspects of these wind turbine components. To further reinforce background knowledge of modern wind turbine systems, Fig. 5 shows the basic configuration of a wind turbine with its mechanical, electrical, and monitoring and control subsystems that are appropriately integrated for electricity generation in a controlled and reliable manner. Table 2 presents a more detailed list of major components under each category of mechanical, electrical, and monitoring and control subsystems. Note that some of the components in this list can be optional depending on the wind turbine’s drivetrain design. As shown in Fig. 6, different alternative designs and arrangements for the drivetrains of wind turbines exist. As seen in this figure, depending on the type of the generator used and the design specifications of the turbine, the drivetrain can either be direct-drive (without gearbox) or geared-drive (with gearbox). The common alternative current (ac) generators are also categorized under two types of synchronous generators (which can be conventional or multipolar) and asynchronous (induction) generators with different rotor types, as also shown in Fig. 6. One should also mention that new designs can be expected as wind turbine technologies keep developing. Readers seeking further details on generators and power electronics for wind turbines are referred to [53] and references therein.
C. Monitoring and Control: Critical Subsystems With Intertwined Functions
As they continuously ensure the safe and efficient operation of a wind turbine as a whole system, the monitoring and control equipment, among other subsystems, are of critical importance. They can be considered under each individual wind turbine and an entire wind farm (group of wind turbines) level. Fig. 7 illustrates this important classification of monitoring and control functionalities in wind turbines and wind farms. From a larger perspective, wind turbine condition monitoring, when conducted in real time, not only can enable condition-based maintenance (discussed in Section IV) but also the emerging concepts of “condition-based control” and “fault-tolerant control” to safely increase wind farm availability. This ultimately results in a health management scheme, as outlined in Fig. 8. One can see that the information obtained from real-time condition monitoring is used in a high-level supervisory control system, which accordingly determines if it is safe and reasonable for wind turbines to continue producing power in the event of faults and before the essential condition-based maintenance. In addition, wind turbines’ operational uncertainties can be alleviated using condition-based and fault-tolerant control algorithms at the levels of any individual wind turbine or the entire wind farm. More precisely, condition-based control keeps track of the real-time condition of wind turbine components and adapts the control actions to modify the loading of components depending on their health condition and, thus, to delay/avoid failure. Fault-tolerant control, on the other hand, aims at accommodating the effects of noncritical faults in wind turbines by reconfiguring the control algorithms to avoid unnecessary shutdowns and missed production. When supported by a reliable and effective CMS, condition-based and fault-tolerant control mechanisms lower the probability of wind turbine unexpected maintenance and help enjoy improved supply certainty (availability). Let us note that such benefits come on top of the benefits from condition-based maintenance itself [54], [55]. Accounting for these benefits altogether leads to significant revenue improvements over the lifetime of a wind farm, especially if located in less accessible offshore regions.
Schematic representation of an integrated health management scheme for wind turbines.
Fault and Failure Modes in Wind Turbines
Wind turbines constantly undergo a wide range of changing loads and operating conditions, which results in their components experiencing considerably high failure rates. The most common root causes of the fault and failure modes in wind turbines are shown in Fig. 9.
Unexpected faults may occur in any wind turbine components, such as sensors, actuators, rotor blades, generators, gearboxes, electric systems, and electronic control units, to name but a few. Fig. 10 shows the most important wind turbine faults and component failures according to their severity. The more damaging the ramification, the higher the level of severity assigned to the fault/failure effects. The most severe faults/failures can lead to the complete shutdown of the turbine. As for the severe and less severe faults, they usually partially affect the ability of a wind turbine to produce its nominal power, but urgent repairs may still be required. Notice that the low level of severity considered for sensor faults relates to the physical redundancy applied when installing sensors, which aims at facilitating the detection and accommodation of sensor fault effects. However, these sensor faults may become critical if their effects are not handled in a timely manner, which is before going through feedback loops into the wind turbine’s control system. Indeed, a fault might not be severe in the first place, but its effects might quickly propagate in the wind turbine and lead to the catastrophic failure of its rotor, drivetrain, or power generation systems. Fig. 11 illustrates this process by showing how various components’ fault effects can propagate through the wind turbine subsystems.
Several independent studies on the reliability of wind turbines (mostly onshore) are already available in the literature (for instance, see [3], [9], and [56]). Also, relatively solid reviews, such as those in [7], [8], and [57], bring together and compare data from a selection of major studies in the literature. Table 3 presents the main results and conclusions extracted and compiled from these studies for both onshore and offshore wind turbines. According to reliability studies, as wind turbines are becoming larger, more flexible, and located further from shore, their O&M costs rapidly rise unless their reliability is improved through effective condition monitoring solutions.
Wind Turbine Condition Monitoring And Condition-Based Maintenance
It is quite important to detect, diagnose, and prognose any types of abnormalities and faults as early as possible before they propagate to major damage or severe failure. As shown in Fig. 12, the advanced techniques for FDD and LTP, included in a CMS, enable a very cost-effective type of preventive maintenance (before a failure), which is often referred to as “condition-based maintenance.” When a fault occurs, depending on its type and location, it takes a certain time to develop before it can interrupt or stop the operation of the wind turbine. To be effective, CMS must take into account that time span. For instance, some faults occur within a very short timeframe, in the order of seconds, to grow from inception to failure (e.g., generator earth fault), whereas others may take up to months before causing a failure (e.g., fatigue and fracture). Thus, the former may provide sufficient time for detection but probably not for complete diagnosis, prognosis, and maintenance action, while the latter provides enough time not only for detection but also for effective diagnosis, prognosis, and successful maintenance action. The time period from fault detection to maintenance action is usually referred to as the “prognostic horizon.” It is important to accomplish the FDD and LTP as effectively and automatically as possible to alleviate manpower and enable efficient condition-based maintenance. Fig. 13 exhibits the general maintenance objectives and a refined classification of the existing maintenance strategies, including condition-based maintenance. Compared with other maintenance strategies, condition-based maintenance reduces the number of maintenance visits and ensures that the overall maintenance is necessary and truly worthwhile. As a result, O&M costs decline in terms of labor, materials, and machine downtimes.
As aforementioned and shown in Fig. 13, CMS plays a key role in implementing condition-based maintenance. In recent years, efforts to develop efficient and cost-effective CMS for wind turbines have increased significantly. Several commercial systems for wind turbine monitoring, most of them developed based on existing techniques from other rotating machine industries, are already available on the market (for instance, see [23] and [58]). Also, many innovative projects and techniques have been introduced and are being explored in both academia (for instance, see [59] and [60]) and industry (for instance, see [61] and [62]). In general, the monitoring techniques for wind turbines can be divided into two categories: 1) offline and 2) online (in real time). On the one hand, off-line techniques involve periodic inspections during which the machine needs to be shut down, and/or the attention of an operator is necessary. Although these techniques are suitable for the design and certification process of new wind turbines, they often require the attention of an operator and are unable to determine the real-time condition of a working wind turbine [12]. On the other hand, real-time condition monitoring based on online techniques is becoming increasingly important, especially for offshore wind farms where the inspection intervals are longer than those in onshore wind farms [13]. These online techniques continuously provide the real-time monitoring of a machine during its operation. They can automatically report continuous raw measurements and may even incorporate onboard models and/or signal processing functionalities for enhanced data reduction and analysis. Although there is no unified categorization of online techniques in the literature, they can generally be further categorized as: 1) hardware signal-based techniques; 2) mathematical model-based techniques; and 3) hybrid techniques that refer to any combinations or integrations of both hardware signal- and mathematical model-based techniques together.
Hardware signal-based techniques refer to any techniques utilizing the output signals from hardware sensors in a wind turbine. These signals can come from either a “standard” supervisory control and data acquisition (SCADA) system or any “standalone condition-specific” sensing and data acquisition systems adopted to monitor the status of a specific condition variable (e.g., oil quality, vibration, and acoustic emission) using oil quality sensors, strain sensors, thermal sensors, infrared sensing devices, acoustic emission sensors, and so on. Therefore, signal-based techniques can be further categorized as: 1) SCADA- and 2) condition-specific-based techniques.
Mathematical model-based techniques employ the “mathematical models” of a wind turbine or its major components without requiring the high-resolution condition-specific signals used in signal-based techniques. Indeed, model-based techniques mainly require the mathematical models of the process and the input–output (I-O) information commonly available from a wind turbine and mainly related to the wind turbine’s control system.
Conceived from the reviewed literature, Fig. 14 shows the important milestones in the evolution of the mentioned wind turbine condition monitoring techniques since 1990. In addition, Table 4 classifies a detailed list of fault and failure modes associated with wind turbine components and their relevant condition monitoring techniques. As observed in the table, although SCADA- and model-based techniques are potentially (or rather ideally) applicable to most wind turbine components, some other techniques, especially the condition-specific-based ones, can only be used for certain wind turbine components. Sections V and VI are devoted to the comprehensive review of each technique, individually.
Hardware Signal-Based Condition Monitoring
As its name suggests, signal-based condition monitoring usually involves measurement signals and signal processing methods under a data-driven approach designed to obtain useful FDD and/or LTP information from a large amount of observed data. Data-driven approaches typically rely on the time domain, the frequency domain, or the time–frequency domain data analyses to extract the fault-related features in the signals (observed data) and enable FDD or LTP without using a physics-based (or an explicit I-O) model. It is worth noting that, although a physics-based model is not used here, data-driven approaches may still use some sort of data-based models (in terms of mathematical functions/relations), which work only for a particular system under monitoring.
The major steps typically involved in the signal-based condition monitoring process and data-driven approaches are shown in Fig. 15. As seen in this figure, a set of (raw) signals is first obtained from either a standard SCADA system (i.e., the SCADA signals in Fig. 15) or other standalone condition-specific sensing and data acquisition systems (i.e., the condition-specific signals in Fig. 15). Typically, for data acquisition, the signals need to be simultaneously collected from several sensors installed at different locations and various wind turbine components. The collected signals are then preprocessed using appropriate methods for signal conditioning. Signals from multiple sensors and different sources are integrated (sensor fusion), and their features necessary to relate the process signals to the system conditions are generated. In most cases, data analysis using signal feature extraction/selection is needed to extract more useful (informative) features and reduce the dimensionality of data that are highly sensitive to the system’s conditions (see Fig. 15). Often, choosing an optimal method for “feature extraction” is a challenging task that depends on the problem and classifier design. The main objective is to achieve independent and discriminative features. In general, the methods can be categorized into time, frequency, and time–frequency domains.
In time domain (time-domain analysis), the methods deal with extracting signal features directly from the time-domain representation of signals. These time-domain methods usually involve time-domain indices, such as peak level, variance, root mean squared (rms) value, shock pulse counting, kurtosis, crest factor, envelope analysis, time series averaging, and many more.
In the frequency domain (frequency-domain analysis), the methods focus on extracting signal features from the frequency-domain representation of signals, which is obtained from the Fourier analysis or fast Fourier transforms (FFTs). These frequency-domain methods usually involve frequency-domain indices, such as mean frequency, rms frequency, root variance frequency, spectral skewness, spectral kurtosis, spectral entropy, the Shannon entropy, and many more.
In the time–frequency domain (and the analysis of the same name), the methods deal with extracting signal features from the time–frequency domain representation of signals obtained from the mapping of 1-D time-domain signals to a 2-D function of time and frequency. The said mapping can be achieved using some common techniques, such as the short-time Fourier transform (STFT), the wavelet transform, and the Wigner–Ville distribution.
It should be noted that, due to their inherent dynamic nature, and particularly under the time-varying operating conditions, most of the signals encountered in engineering applications are nonstationary and vary over time. The time- or frequency-domain analysis alone is not ideal to deal with those nonstationary signals because they cannot reveal the important features in both time and frequency domains simultaneously. To overcome this problem, the joint time–frequency analysis is an effective solution, which has been used as the main feature extraction method in wind turbine CMS [44]. A detailed review of time–frequency analysis methods for signal feature extraction is presented in [90].
Beside feature extraction, another useful approach for feature dimensionality reduction (reduction of the number of features) is referred to as “feature selection.” Unlike feature extraction that involves the transformation of features into a higher dimensional space, feature selection works by sorting out a subset of the existing features to remove those that are not relevant or that are redundant. Feature selection techniques are divided into three categories.
Filters that work by ranking the features according to some characteristics of the data, such as distance, correlation, information gain, and fisher score. The filter techniques are computationally efficient since they are independent of the classifier performance [91]. However, this results in a feature subset, which is not tuned to a specific type of classifier and usually gives lower prediction performance.
Wrappers that work by searching for a useful subset of the features according to a learning algorithm, which involves the classifier itself as part of the evaluation function. Compared to the filters, the wrapper techniques tend to perform better in selecting features. However, these techniques are computationally intensive, especially as the feature space grows [92].
Embedded techniques that are introduced to bridge the gap between the filter and wrapper techniques. They work by embedding feature selection with classifier construction to attain the benefits of wrappers (as they involve the interaction with the classification model) and filters (as they are computationally efficient) [93].
A comprehensive review of feature selection techniques in each of the abovementioned categories is presented in [94].
In continuation of the steps shown in Fig. 15, the signal features obtained from the process of data analysis are accordingly considered by the concluding process of condition classification/regression. In this process, the classification-based and/or regression-based data-driven approaches are used to predict a categorical variable (classification) or a numeric variable (regression), respectively. More precisely, a classification-based data-driven algorithm leads to a mapping (called classifier) from some input space (the space of feature values) to a discrete output in the form of class labels, while a regression-based data-driven algorithm results in a mapping (called regressor) from some input space to a continuous output in the form of real numbers. Indeed, the data-driven algorithm utilizes an adequate set of data to understand (learn) and evaluate the mapping relations between the input and output spaces. The set of data is representative of various operating conditions and is typically split into training (usually, about 65% of the data) and testing (usually, about 35% of the data) sets [44]. After parameter adjustment and learning from the training set, the performance of the resulting mapping relations should be evaluated using the testing set. In the case of classification, the performance is typically evaluated according to several measures, such as accuracy, sensitivity, specificity, and F1-measure. However, the evaluation measures for regression are different and may include measures, such as the mean absolute error, rms error, and r-squared. When a classifier or regressor is accurately trained and evaluated, it can be eventually used in the FDD or LTP process.
As seen in Fig. 15, for classification or regression purposes, there is a whole variety of algorithms in data-driven approaches that are generally included under two main categories: 1) artificial intelligence (AI) algorithms (e.g., artificial neural networks (ANNs), fuzzy logic, and expert systems) and 2) statistical algorithms (e.g., least-squares regression, relevance/support vector machine, and the stochastic processes, such as the gamma process, the Wiener process, and the hidden Markov model). Among these algorithms, the AI ones based on different types of ANNs have been widely used for both classification and regression in different fields and applications. Comprehensive reviews on various data-driven algorithms with general application to FDD and LTP can be found in [95] and [96], respectively.
Fig. 16 shows a typical schematic of a signal-based condition monitoring scheme but from a machine learning perspective. Indeed, a signal-based condition monitoring scheme can be designed using either a machine-learning-based approach (also known as a knowledge-based approach) or a simple nonmachine-learning-based approach. The nonmachine-learning-based approach only relies on a plant’s measured outputs whose signal patterns under healthy conditions are
It is worth emphasizing that the schematic shown in Fig. 16 is for illustration purposes only. Depending on the specific requirements and circumstances of the problem addressed, the shown schematic may vary in some parts or steps (e.g., pattern recognition methods may act directly on the raw data without signal preprocessing and/or feature extraction).
With respect to the type of signals used in the signal-based condition monitoring of wind turbines, Sections V-A and V-B review the SCADA-based techniques and condition-specific-based techniques, respectively.
A. SCADA-Based Techniques
All large utility-scale wind turbines use a SCADA system, mainly for monitoring the overall performance of wind turbines and their major components. A standard SCADA system provides an important source of information at both wind turbine and wind farm levels. This information includes operational and availability data (typically recorded at 10-min intervals), as well as instantaneous alarms data (whenever alarms or warning messages are generated). Although the exact information (and data configuration) obtainable from the SCADA system depends on the supplier, a list of signals typically recorded by such a system is given in Table 5 [23], [33], [79]. Note that several additional parameters, such as vibrations, oil pressure levels, and filter status, can also be recorded by some up-to-date SCADA systems. However, in order to reduce the transmitted data bandwidth from a wind farm, the recorded data are generally 10-min averages of 1-Hz sampled values, while other forms, such as minimum, maximum, or standard deviation of live values recorded in a 10-min period, are also possible [97].
Despite the basic idea of using the SCADA system, rigorous analysis of SCADA data can potentially result in an efficient understanding of wind turbine’s health condition. Indeed, using appropriate algorithms, the basic idea of “overall performance monitoring” by the SCADA system can be matured into efficient condition monitoring schemes with FDD and LTP capabilities for both rotating and nonrotating subassemblies of wind turbines. More precisely, “performance” is described in connection with the underlying process physics of a system—in this case, a wind turbine and its components. As wind turbine components deteriorate over time, the efficiency with which wind kinetic energy is converted to useful electrical energy decreases, and the performance of the wind turbine degrades. Therefore, the performance degradation can be a sign or symptom of many diverse problems in the wind turbine and its components, ranging from aerodynamic degradation of rotor blades (due to erosion, dirt, or ice buildup on blades) to any other component malfunctions due to faults, damages, or wear and tear.
Since condition monitoring of wind turbines using SCADA data is a potentially low-cost solution, requiring no additional sensors, several SCADA-based performance monitoring systems are already researched and developed in both academia and industry (for instance, see [19], [21], [33], and references therein). From a design viewpoint, a suitable performance parameter is first computed according to the measured SCADA signals [84]. This could be any performance-representative parameter or fault indicator chosen from raw sensor signals (or their features), sensor signals corrected for environmental conditions, component efficiencies or aerodynamic parameters, and so on. Then, using different methods, one or more such parameters are carefully considered to assess a wind turbine’s condition and determine whether the wind turbine is behaving within its normal bounds. Several methods can be used for this purpose. Indeed, one of the first steps in the evolution of condition monitoring with SCADA data was based on simple trending methods, for instance, using regression lines in scatter diagrams of temperature versus power or 3-D visualizations, including the ambient temperature. Past studies involved such trending methods trying to detect faults or early signs of degradation (i.e., the evolution of damage) (e.g., see [98]–[100]). However, the main challenge lies in how to accurately interpret “trends” of SCADA parameters given the variability of the operating conditions of modern wind turbines, as a change in the value of a SCADA parameter is not necessarily evidence of a fault. Most studies have shown that automated online monitoring based on trending methods will most likely fall short of required accuracy since the problem is highly case-specific and usually requires an off-line visual interpretation of the trends. Such a problem can be very difficult to tackle, especially when dealing with monitoring a large farm of wind turbines that operate under very different conditions. To address this shortcoming of trending methods and as another step in the evolution of condition monitoring with SCADA data, clustering algorithms are introduced to automate the classification of “normal/healthy” and “faulty” observations (for instance, see [101]–[103]). Yet, according to the reviewed literature and compared with trending methods, the overall idea of clustering the observations does not exhibit a clear or noticeable advantage since the interpretation of results for condition monitoring is again difficult [33]. In addition, extensive historical fault data are required for training purposes—the data of faulty operation, which ideally include the full range of fault dynamics. However, it is not necessarily feasible to access such fault data in practice. To avoid such problems, more advanced methods, such as normal behavior modeling or damage modeling, are proposed in the literature. In normal behavior modeling methods, models of normal behavior, empirically “trained” on historical data (under normal conditions), are used to generate an error or residual signals with respect to the measured performance parameters. Accordingly, any serious deviations of the residuals (from the vicinity of zero) indicate the occurrence of faults or failures in a wind turbine component. For instance, SCADA data can be used to model and monitor the wind turbine “power curve” that relates the power output to the wind speed, thus giving an important measure of power generation performance by the wind turbine (e.g., see [104]–[107]). In addition, there are several other examples based on modeling approaches but for monitoring other wind turbine operational parameters (e.g., see [108]–[111]). From a design viewpoint, the normal model can be created based on either the Full Signal ReConstruction (FSRC) concept, for which only the signals other than those of a target parameter are used to predict the target, or AutoRegressive with eXogenous (ARX) input modeling, for which past values of the target are used as well. According to the reviewed literature, models created using polynomial equations, ANNs, adaptive neurofuzzy inference systems (ANFISs), or nonlinear state estimation techniques (NSETs) have demonstrated more satisfactory condition monitoring results [33]. However, there is no solid study comparing these modeling techniques in terms of their true value for the provided modeling accuracy in return for the imposed computational burden. This could help conclude which technique has superior performance over others for normal behavior modeling purposes. In addition, the majority of studies in this area are focused on modeling the normal behavior of “single-target” variables of interest. As a result, a growing number of models are developed to describe the normal behaviors of wind turbines’ specific components (e.g., a normal model for the temperature behavior of the main bearing, a second model for the generated active power, and so on). In practice, one needs to both keep track of and maintain all of these models and to update them and/or their related threshold values when required, as the normal behavior may change over time due to sensor recalibrations, part replacements, software updates, and so on. Therefore, each additional normal behavior model further increases the burden for wind farm operators. To address this issue, multitarget normal behavior modeling approaches (such as the one recently proposed in [112]) can be considered. Such approaches enable, at once, simultaneous SCADA-based monitoring of multiple state variables using a single (multitarget) normal model. In contrast to the normal behavior modeling approach that presents a “black-box” solution with little or no insight into the physical processes that drive faults and failures, the damage modeling approach focuses on developing a damage model based on a physical understanding of one particular failure mode of interest. Thus, the latter approach can potentially better represent damage development and provide more accurate results. However, the development of reliable and sufficiently accurate damage models for all failure modes of a wind turbine will be very demanding. This is especially due to the lack of knowledge needed on the sufficiently large numbers of failures, which themselves can vary depending on the type, manufacturer, or location of wind turbines. Table 6 presents an updated summary of the literature produced on the so-called SCADA-based condition monitoring techniques.
In summary, motivated by the lower costs offered by the system not requiring additional hardware investment, a large number of studies in the literature have been devoted to SCADA-based condition monitoring in wind turbines. However, since not initially designed for condition monitoring purposes, the provided monitoring performance by SCADA-based CMS is limited, mainly due to the following concerns and shortcomings.
Although SCADA data provide relatively extensive information that can be useful to identify abnormal wind turbines in a wind farm, the data usually do not include all the necessary information for a full (detailed) condition monitoring of wind turbine subsystems and components.
The distribution of SCADA data is generally imbalanced, and the anomalous data mining is usually insufficient. This means that the amount of normal data is typically much higher than that of abnormal data, which might result in poor condition monitoring performance, since the data-driven models tend to be biased toward the majority class (i.e., normal data).
The sampling frequency of SCADA data is much slower than those required for some condition monitoring techniques.
Data quality is usually a concern. For instance, 10-min logs of SCADA data are commonly affected by problems such as “missing values,” “NULL entries,” or “zeros,” “plausibility limits exceeded,” “statistical outliers,” “large blocks of identical values consecutively,” “incorrect data format,” and so on. Accordingly, any necessary corrections made to the data, such as those using linear or exponential interpolation, extreme values to limit value, or even removing the problematic channels of data, can decrease the required level of accuracy.
Given the variability in the operating conditions of modern wind turbines, rapid fluctuations of environmental conditions (wind speed/direction, air density, turbulence, and so on), and the low sampling frequency of typical SCADA systems, it is difficult to detect, diagnose and prognose incipient faults in a timely manner.
Because of the abovementioned limitations, SCADA-based condition monitoring cannot currently replace a “professional” CMS, which is especially designed for condition monitoring purposes in wind turbines. As discussed in Section V-B, such a system employs additional hardware in the form of different standalone condition-specific sensing devices that are purposely dedicated to precisely monitor-specific condition variables, such as oil quality, vibration, and acoustic emission. Compared with SCADA systems, the measured signals are more precise and sampled at higher frequencies, which provide richer information, enabling greater insights into the wind turbine health condition. However, this comes with a higher cost, which depends on measurement accuracy, sampling frequency, system functionality, and application environment.
B. Condition-Specific-Based Techniques
According to the additional hardware used in a professional CMS, a wide range of condition-specific signals is available for analysis. Table 7 provides a list of those signals and their related information, including the features of monitoring hardware, possible condition monitoring capabilities, and relevant components monitored in a wind turbine.
With respect to the condition-specific signals listed in Table 7, the following paragraphs briefly review the available literature for detection principles, development methods, pros and cons, and challenges and limitations of each technique.
1) Vibration:
Many wind turbine components after being affected by faults or defects produce a new vibration behavior, which can be monitored using the signals obtained from vibration sensors. Thus, any abnormal variations in the measured vibration signals can indicate faults in a monitored component. For instance, in the case of moderate to high-speed bearings, each and every time a roller element passes over a defect, an impulse of vibration is generated, which can act as a fault indicator [29]. Also, the vibration signal’s amplitude can indicate the fault’s severity [28]. Indeed, appropriate signal processing methods can identify the fault frequencies and, accordingly, the location and type of fault [137].
Given the significance of vibration monitoring, most of the commercially available CMSs mainly rely on this type of monitoring using vibration sensors, often installed on the surfaces of rotor blades, and the casings of internal components, such as generator, gearbox, main shaft, and bearings. Depending on the frequency range and operating conditions of the monitored components, various types of vibration sensors are available, including: 1) displacement sensors in the low-frequency range; 2) velocity sensors in the middle-frequency range; 3) accelerometers in the high-frequency range; and 4) spectral emitted energy (SEE) sensors in very high frequencies (acoustic vibrations). Among them, accelerometers cover the widest working frequency range from 1 to 30 kHz, which makes them the most popular vibration sensors in the condition monitoring of wind turbines. In addition to the mentioned sensors, ground-based radar (GBR) is also used as a vibration-based noncontact remote sensor for structural condition monitoring of in-field wind turbine blades [138].
As listed in Table 7, the vibration signal-based technique is appropriate for monitoring the health condition of rotor blades, generator, gearbox, bearings, and other selected wind turbine components, such as a tower. Table 8 summarizes the advantages and disadvantages of this technology and provides a list of selected references categorized based upon the name of components being monitored by vibration in wind turbines. The main challenges of this technology relate to the complications that come with the components’ different frequency ranges and the distinction of the acquired vibration signals generated by the faults from those caused by the environmental and operating conditions.
2) Strain:
A wind turbine structure is made of materials that deform under applied loads. These structural deformations can be characterized by a dimensionless quantity known as strain. The measured strain signals can be effectively used to monitor structural health conditions against faults in the form of structural defects (e.g., blade icing and mass unbalance) or damages. Strain sensors are often mounted on the surface or embedded in the layers of a structure. Generally, there are two popular types of sensors for strain measurement: 1) traditional electrical sensors and 2) relatively modern optical fiber sensors. The electrical sensors include several types involving capacitance, inductance, semiconductor, or resistance. Among them, the resistance strain gauge is the most popular with a well-established mature technology. However, strain gauges can suffer from several issues, such as easy degradation and failure over long-term operations, and vulnerability to lightning strikes, electromagnetic interference, and variations in temperature, which necessitates careful compensations in the results [17], [46]. As modern alternatives based on fiber optics technology, optical fiber sensors are being developed [164]. Specifically, the so-called fiber Bragg gratings (FBGs) have become popular for offering high sensitivity through a direct physical correlation between wavelength and strain. They have long-term durability under hostile operating conditions and use nonelectrically conducting transmission lines, which ensures lightning safety and neutrality against electromagnetic interference [17], [165]. Having said that, FBG sensors are still expensive although efforts are made to make their application more cost-effective [15].
As listed in Table 7, the strain signal-based technique is mainly appropriate for monitoring the health condition of wind turbine structures, including rotor blades, towers, and foundations. Table 9 summarizes the advantages and disadvantages of this technology and provides a list of selected references, which individually address the monitoring of a specific component by strain in wind turbines. The main challenges of this technology relate to the need for a large number of sensors (as each sensor can measure at one local point) and enough initial knowledge of critical points and high-strain areas for effective sensor placement.
3) Torque:
Torque (also known as moment or moment of force) is the tendency of a force to cause an object to rotate around an axis or other point. It is a vector quantity with both a direction and a magnitude. The faults and defects in mechanical components usually leave signatures in measured torque signals. Thus, it is possible to detect and diagnose these faults and defects by monitoring the torque signals obtained from wind turbine components. For instance, malfunctions such as rotor imbalance and aerodynamic asymmetries can be diagnosed by analyzing the torque experienced by a wind turbine’s tower [70]. Likewise, gear defects, especially for the low-speed gear in a gearbox, can be diagnosed using the envelope spectrum of the torque measurements in the gearbox [83].
A torque sensor is indeed a transducer that converts a torsional mechanical force into an electrical signal. Generally, there are two major types of torque sensors that can be installed on the components being monitored, such as rotor blades, gearbox, generator, and tower. The mentioned sensors include rotary torque sensors for measurement of rotational torque (when there is an axle or pivot to be turned around) and reaction torque sensors for measurement of bending moment (when there is an element to be bent). In the case of a generator, the torque can be also estimated based on the generator speed and electrical outputs without using torque transducers.
As listed in Table 7, the torque signal-based technique is available for monitoring the health condition of wind turbine components, including rotor, generator, drivetrain (gearbox, shaft, and so on), and tower. Table 10 summarizes the advantages and disadvantages of this technology and provides a list of selected references categorized based on the name of components being torque-monitored in wind turbines. This technology usually requires more complicated signal processing algorithms because of torque signal modulation problems with dominant components related to the load. Having said that, the main challenges still relate to the high cost and installation complexities of torque transducers, especially for new wind turbines with more compact designs.
4) Shock Pulse:
Shock pulse monitoring, referred to as shock pulse method (SPM) technology, was first introduced in 1969 to determine the condition of rolling element bearings or any piece of machinery with continuous metal-to-metal contact, which gives off shock pulse signals [192]. In simple terms, at the instantaneous moment of mechanical impact between two masses, the molecular contact happens, and a compression (shock) wave/pulse develops in each mass. In a bearing (whether new or old), the mechanical impacts happen during the rotation of the bearing and due to its natural surface roughness or surface defects/damages. These mechanical impacts generate shock pulses in the interface between the loaded roller element and the race way, which, in turn, results in the bearing acting as a “shock pulse generator.” The magnitude of shock pulses depends on the bearing’s surface condition (i.e., roughness, stress, damages, and oil film thickness) and its peripheral velocity (i.e., rotational speed, size, and so on). These shock pulses have an ultrasonic frequency band and typically occur around a center frequency of 32 kHz [29].
In SPM, shock pulses are measured using specially designed piezoelectric accelerometers equipped with filters for reducing the influence of environmental factors, such as background vibration and noise. Such accelerometers are tuned mechanically and electrically to a resonant frequency of 32 kHz [192]. The measured shock pulses are typically recorded per second, and two amplitude levels are extracted: 1) the decibel carpet value of 200 shock pulses per second and 2) the peak value of incoming shock pulse under 2 s. The decibel carpet value provides an indication of the lubrication condition, and the peak value provides the extent of bearing damage. Indeed, the measured shock pulse amplitudes are subtracted from the expected shock values in a healthy bearing at a similar speed. Accordingly, an indication of the bearing health condition is obtained. Typically, there are three condition regions/scales, namely, “Green” for good condition, “Yellow” for warning, and “Red” for damaged condition. This provides the operator with the status information of the machine and the zone it belongs to.
As listed in Table 7, the shock pulse signal-based technique is mainly applied for the condition monitoring of wind turbine bearings used in the gearbox, yaw system, or blades. Unlike other alternative techniques, such as vibration analysis, SPM can uniquely analyze and display the state of bearing lubrication and bearing mechanical condition without needing baseline data development for trending [192], [193]. Also, this technology can isolate (locate) the faults/damages since the damaged bearings generate shock pulses with a pattern that corresponds to the frequency of the balls passing over the damaged part. Table 11 summarizes the advantages and disadvantages of SPM as referenced in a selected list of studies focusing on its implementation in wind turbines. Although the commercial SPM instrument is typically hand-held and simple to use in real time, the presence of semiskilled personnel for holding the device and pressing the sensor to the bearing cover with the probe is necessary. Therefore, the fleet-level implementation of this technology might become complicated and expensive compared with vibration monitoring. Finally, as a result of the technology’s proprietary nature, the published literature on this subject, especially for wind turbines, is scarce, and there is almost no evidence of prognosis capabilities for SPM [29].
5) Acoustic Emission:
Acoustic emission is a transient impulse caused by a rapid release of strain energy in the form of transient elastic waves within a solid material when it undergoes stress/strain conditions through mechanical or thermal loadings. According to this phenomenon, any alterations in a structure excite acoustic emission signals, which may be monitored by appropriate sensors. Indeed, the information about the occurrence and propagation of surface or subsurface structure damages can be extracted using the signals’ waveform characteristics, such as acoustic energy, rise time, duration, amplitude, kurtosis, and rms values [17], [46], [198], [199].
Basically, acoustic emission monitoring is very similar to vibration monitoring in terms of the nature of the collected data, which, for both, originates from the alterations in a material structure. Indeed, the former employs acoustic sensors to “listens” to the material alterations using sound level meters, while the latter associates such alterations to the material vibrations measured using the vibration sensors [1], [30]. In contrast to the vibration sensors that are rigidly mounted to the monitored components, the acoustic sensors are flexibly attached to the components using a viscous couplant (i.e., a coupling medium, such as silicone grease or adhesive) and a mechanical clamp offering a constant coupling pressure [30], [200]. To widely monitor the whole structure and identify the location of a damage or fault, multiple acoustic emission sensors (i.e., sensor arrays) are required. The frequency of acoustic emission signals typically lies between 100 kHz and 1 MHz recorded by a wide range of sensor types, such as piezoelectric, resonant, or wideband transducers [200], [201].
As listed in Table 7, the acoustic emission signal-based technique is mainly applied for the condition monitoring of wind turbine components, including rotor blades, generator, drivetrain (gearbox, bearings, and so on), and tower. Table 12 summarizes the advantages and disadvantages of acoustic emission monitoring and provides a list of selected references categorized according to the names of the components being monitored in wind turbines. Compared with vibration monitoring, acoustic emission technology works with signals of much higher frequencies that can provide more effective monitoring performance, especially for incipient structure defects, damages, or faults at an early stage. Yet, such high sampling frequencies can complicate the required signal processing and increase its computational cost. In addition, this technology usually relies on the installation of a large number of sensors, including dedicated data acquisition devices (for signal sensing, processing, and transfer), which increases condition monitoring costs and causes additional reliability issues due to the use of supplementary sensors. Finally, it is also challenging to correctly distinguish signals from acoustic emissions and noisy background environments, not to mention the other factors, such as temperature, lubrication, and loading, which are found to have a significant influence on the acoustic emissions [28], [29], [198].
6) Temperature:
By definition, the temperature is the degree or intensity of heat present in a material or object. It is especially expressed using a comparative scale and indicated by a sensor. A variety of different sensors, such as optical pyrometers, resistant thermometers, and thermocouples, are available for temperature measurement [15], [30]. Among them, thermocouples are widely used in wind turbines given their very low cost and good reliability [21].
During the normal operation of a wind turbine, each component’s temperature must remain within its allowable range. An abnormal temperature can possibly indicate faults due to problems such as equipment degradation, low lubrication oil or inefficient lubricant properties (excessive frictions), generator winding short circuits, and shafts over speed. Therefore, measurement and monitoring of temperatures in a wind turbine are among the most common techniques, which can provide useful information on the machine’s health condition. For this technique, measurement of temperatures in each individual component and subcomponent (e.g., bearings and shafts) is essential since it enables better diagnosis performance based on larger information available. Also, it should be noticed that the changes in component temperatures can be correlated with variations in surroundings temperature (e.g., nacelle temperature) or the wind turbine’s normal operating conditions, including rotational speed and loads [21], [30].
As listed in Table 7, the temperature signal-based technique is mainly used in the condition monitoring of wind turbine components, including generator, converter, drivetrain (gearbox, bearings, and so on), nacelle, and transformer. Table 13 summarizes the advantages and disadvantages of this technology and provides a list of selected references categorized according to the name of the components being monitored by temperature in wind turbines. In addition to concerns related to the measurement accuracy of a component’s temperature (due to environmental and operational effects), the main challenge of this technology lies in the fact that “temperature” (containing fault-related information) develops slowly and sometimes too late compared with other condition-specific signals. As a result, to enable early and precise detection and diagnosis of faults, the temperature signal is rarely used alone but rather often with another source of information such as vibration signals [15].
7) Oil Debris/Quality Parameters:
The lubrication oil plays an important role in reducing the friction, heat, and wear between mechanical components that are in contact, especially those inside the wind turbine’s rotating subsystems and components, such as gearboxes, generators, and bearings. The aims of the oil monitoring and analysis are twofold: 1) to ensure an appropriate oil quality to optimize oil changing schedule and prevent equipment damage caused by poor oil quality and 2) to monitor and estimate the health condition and wear of the oil-lubricated equipment. Besides monitoring the oil pressure and temperature, its samples can be analyzed to assess the oil debris/quality parameters, such as viscosity, levels of contaminants (e.g., water content, coolant, and fuel), and the size, shape, composition, and count of solid particles (debris). Indeed, an excessive number of particles, the presence of large particles, or those of a particular shape can indicate abnormal wear conditions, faults, or an impending failure [26], [227]–[229]. Therefore, the precise monitoring and analysis of oil debris/quality parameters can provide very useful information about the health condition of the wind turbine’s oil-lubricated components and their potential faults at an early stage [227]–[230]. Although such an analysis is typically performed offline by taking periodic samples, recent advances in online oil sensing technology have enabled continuous online oil analysis tools [231], [232]. For online oil condition monitoring, there are several sensors available, such as the oil humidity (water) sensor, sensors for viscosity, particle concentration, quality and properties, conductivity, thermometer, and level sensor [35], [227], [233]. Indeed, the sensors’ applicability depends on the oil type and the sensor measurement range and accuracy.
As listed in Table 7, the oil-debris/quality signal-based technique is mainly used for the condition monitoring of oil-lubricated wind turbine components, including generator and drivetrain (gearbox, bearings, and so on). Table 14 summarizes the advantages and disadvantages of oil debris/quality monitoring and provides a list of selected references categorized according to the names of the components being monitored in wind turbines. For this technology, the well-chosen monitored oil parameters, the relevant set of sensors, and their accuracy all together play a vital role in the reliability of detection and diagnosis results. However, the use of additional oil sensors not only is intrusive in terms of installations but also increases condition monitoring costs and leads to additional reliability issues caused by supplementary sensors. In addition, since the operation of a wind turbine has various impacts on the oil condition, it is usually challenging to correctly interpret the real-time measurements or determine the root causes of faults or abnormalities.
8) Electrical Effects:
Electric current and voltage signals constitute the electrical effects (signatures) of the wind turbine generator. On the basis of electromechanical coupling between the generator and other wind turbine components/structures, several studies (e.g., Schoen et al. [238], Marzebali et al. [239], and Douglas et al. [240]) demonstrated that vibrations inside the mechanical components (shaft, gearbox, bearings, and so on) appear in the electrical effects of the wind turbine generator as well. More precisely, these electrical effects can include stator current, stator voltage, rotor current, and so on. Having said that, as listed in Table 7, the electrical effects’ analysis provides a unique and nonintrusive signal-based technique to monitor wind turbine components and structure since their faults usually induce vibrations; such fault-induced vibrations appear in the electrical effects accordingly [241], [242]. For instance, Kia et al. [243] analytically derived the relationship between the characteristic frequencies of gearbox faults in vibration and electric current signals. With respect to the bearing faults in the wind turbine’s gearbox or generator, Gong and Qiao [244] analyze the amplitude and phase spectra of electric current signals to diagnose bearing faults from an early stage. Another work in [245] studies the effectiveness of this technique in the detection of wind turbine rotor blade imbalances.
Table 15 summarizes the advantages and disadvantages of this technology and provides a list of selected references categorized based on the name of components being monitored by electrical effects analysis in wind turbines. The key advantage of this technology is that it does not need any additional sensors or costly data acquisition systems since the current and voltage signals used for signal analysis can be the same as those used in the wind turbine’s existing control/protection schemes. This provides significant benefits in terms of system costs, hardware complexities, implementations, and overall reliability. Having said that, the real-life application of this technology in wind turbines is still challenging mainly due to the time-varying (nonstationary) nature and, thus, a low signal-to-noise ratio of the electric signals. Therefore, the fault information (i.e., fault signatures or features) hidden in a nonstationary signal obtained from the wind turbine cannot be directly extracted by the classical signal frequency analysis [28], [35].
9) Machine Vision:
Machine vision inspection, also known as the remote machine vision-based monitoring approach, detects structure damages and defects externally visible: surface cracks, scratches, displacements, deformations, deflection, and so on. Obtaining the targeted object’s information through sequences of 2-D or 3-D digital images from different locations and perspectives, this technology relies on principles similar to those of the stereoscopic view of human biological vision [46], [264]. For instance, two cameras can be located at known distances to take simultaneous images of the measured object from different positions. Based on the parallax principle and the geometrical relationship between the two cameras with respect to the object, the images can be combined and processed to obtain a clear sense of depth and surface geometry information [46], [265].
The machine vision-based technology typically consists of a sensing (measurement) system, including image acquisition devices (high-resolution digital cameras, lens, and so on) together with appropriate image processing and damage identification software/hardware platforms [266]. Online inspection and monitoring can be accomplished using remotely installed ground-based or airborne high-resolution image acquisition devices having sampling frequencies over 125 frames per second (i.e., >125 Hz), which would be much higher than those, for instance, typically adopted for strain measurement (about 20 Hz) in the commercial monitoring of wind turbine blades [31], [267]. Yet, the imaging results can be affected by weather conditions [31]. In this technology, especially for online applications, the digital image processing integrated with the damage identification algorithms plays an important role in the effective monitoring and detection of structure faults, defects, or damages [266], [268], [269]. The overall steps typically include: 1) capturing object images (either in 2-D or 3-D); 2) obtaining binary (gray scale) images from the original ones; 3) using edge segmentation techniques (e.g., threshold and edge detectors) and binary morphology to distinguish the defect/damage from the background; 4) extracting important fault indicator features, such as structure deformation, deflection, displacement, distributed strain, and modal parameters; and 5) analyzing the extracted features to assess the structural health [46], [266], [269]. Furthermore, the accuracy of feature extraction highly depends on the quality of image processing, which may involve a wide variety of techniques such as image restoration, reconstruction, segmentation, and recognition. Having said that, it would be yet necessary to comprehensively assess the knowledge behind the relationship between the extracted features and related damages [46].
As listed in Table 7, the machine vision signal-based technique is mainly applied for the condition monitoring of wind turbine structural components, such as rotor blades, nacelle, tower, and foundation. Table 16 summarizes the advantages and disadvantages of machine vision-based monitoring and provides a list of selected references categorized according to the names of the components being monitored in wind turbines. With the fast development of computer science and optics devices in recent years, machine vision technology shows a growing potential for structural health monitoring in the coming years [46], [266], [270], [271]. Also, through the power of AI and the latest autonomous system technologies, such as those using unmanned aerial vehicles (UAVs), new horizons for the autonomous machine vision-based monitoring of wind turbines are opening [4], [268], [271]–[273]. However, to improve detection accuracy, speed, and online computational efficiency, further studies need to focus on image processing algorithms, simultaneous localization and mapping (SLAM), and machine learning (pattern recognition) for correct damage recognition.
10) Ultrasound:
Ultrasound, also known as ultrasonic scanning/testing, is a nondestructive inspection technique that relies on the propagation and reflection of ultrasonic waves within a material. Indeed, the amplitude attenuation and phase shift of the ultrasonic waves are affected differently depending on the differences of the material or any inner material discontinuities [43], [46], [283]. Therefore, structure faults cause different reflection, attenuation, resonance, and transmission patterns, whose analysis through signal processing algorithms enables effective detection and diagnosis of surface or subsurface material defects and damages.
Owing to its efficiency and reliability, ultrasonic scanning is one of the most common nondestructive inspection techniques used in the wind energy industry for the structural monitoring of wind turbine components [15], [46]. In general, the scanning mechanism can be based on the capture and quantification of either the reflected waves (known as the pulse-echo mechanism) or the transmitted waves (known as the through-transmission mechanism). Although each of these mechanisms is used in certain applications, the pulse-echo mechanism is usually more useful as it only requires one-sided access to the component being monitored.
As for the monitoring process, an ultrasound transducer (or probe connected to a diagnostic machine) is passed over the target component being monitored. A thin film of coupling materials (usually a liquid such as oil and grease) is typically used to remove the air gap between the ultrasound transducer and the component monitored. This facilitates the efficient transmission of ultrasonic energy from the transducer into the component. Ultrasound transducers are available in a variety of configurations depending on the application. To optimize the monitoring capability, it is important to select a proper type of transducer that suits the application requirements and has the required frequency, bandwidth, and focusing properties. In contrast to contact transducers, noncontact transducers (such as air-coupled transducers (ACTs), electromagnetic acoustic transducers (EMATs), and lasers) are especially suited to implement continuous remote monitoring of structures that are inaccessible or located in hostile environments [284]. This enables the noncontact automated ultrasonic scanning during which ambient air is the only acoustic coupling medium and the wind turbine is in operation. For instance, Park et al. [285], [286] employ the noncontact laser ultrasonic scanning in the structure monitoring of wind turbine blades.
As listed in Table 7, the ultrasonic signal-based technique is mainly applied for the condition monitoring of wind turbine structures: rotor blades, nacelle, drivetrain, and tower. In addition to the detection and diagnosis of material defects and damages, ultrasonic scanning is shown to be an effective tool to detect, locate, and characterize the icing on the rotor blades as well [43]. Table 17 summarizes the advantages and disadvantages of this technology as referenced in a selected list of studies focusing on its implementation in wind turbines. With respect to its main challenges and technological limitations, the ultrasonic scanning of some complex geometries may be challenging, and initial preparation and equipment calibrations are required especially for contact ultrasonic scanning. Also, more sophisticated signal processing techniques are usually needed to isolate signals from noise particularly in noncontact ultrasonic scanning using EMATs.
11) Thermography:
Thermography, also known as thermal imaging, is a nondestructive inspection technique that uses a special camera (infrared-based camera) to produce thermal images, known as thermograms, showing patterns of heat (temperature) on the surface of objects [300]. It can be effectively used to detect (near-surface) “thermal transients” at a target material. These thermal transients may indicate faults due to different defects or problems, such as structural defects or damages, poor wiring or electrical connections, unbalanced loads, deteriorated insulation, or other potential problems in energized electrical components. More precisely, the faults can be detected, located, and characterized by the analysis of disturbances generated in the local thermal properties (e.g., temperature, thermal capacity, conduction, diffusion, and interface thermal resistance) of a structure or an object [300]. Another monitoring approach relies on the fact that any mechanical phenomenon is accompanied by correlated thermal effects (i.e., thermomechanical coupling), which can be reversible (linked to the strain (and stress) states of structure) or irreversible (linked to the occurrence of damages). Therefore, mapping the thermal state of a structure can also be a tool to detect abnormalities in the mechanical behavior of that structure [300].
As listed in Table 7, the thermographic signal-based technique is mainly applied for the condition monitoring of wind turbine components: generator, drivetrain (gearbox, bearings, and so on), tower, and especially rotor blades. In addition, thermography is shown to be an effective technique to detect and locate the accumulation of ice on a blade surface while distinguishing the types and conditions of icing on the blades [43], [301]. Table 18 summarizes the advantages and disadvantages of thermographic monitoring and provides a list of selected references classified according to the names of components being monitored by thermography in wind turbines. For this technology, “image processing” plays a vital role in the accuracy of detection and diagnosis results. However, it often encounters issues such as motion blurs of moving target objects (e.g., rotor blades when in service) and environmental conditions (e.g., wind speed, ambient temperature, air humidity, reflections, dirt, and prolusion) [302]. Other challenges relate to the nature of this technology, essentially sensitive to near-surface thermal transients (unable to detect or identify deep defects within a material). Also, it is difficult to detect thermal signals generated by incipient defects; the inspection speed is slow and still needs to be improved [303].
12) Radiography:
Radiography, also known as X-ray imaging, is a nondestructive inspection technique that uses an X-ray scanner system to produce radiographic images (X-rays) of the interior structure of objects. This reveals structural variations of the materials, which are caused by changes in material properties, internal delamination, or cracks [15], [312], [313]. Indeed, X-ray transmission data provide quantitative information about those structural variations and enable effective detection and diagnosis of structure faults (i.e., surface or subsurface material defects and damages) in wind turbine components [69], [312]. Although there are some overlaps, the combined application of X-ray and ultrasound is proven to provide complementary capabilities, allowing the detection and diagnosis of a wider range of damages in wind turbine blades (e.g., cracks in adhesive joints, delamination, and laminate failure) [69].
Real-time radiographic inspection is a practical, accurate, and effective technique not only to detect and locate damage (especially laminate damage) but also to determine the size of the damage [69]. In contrast to film radiography, real-time radiography (also called digital radiography) employs a planer array of digital radiation-sensitive sensors (instead of a traditional radiation-sensitive film) and allows immediate image preview and availability [314]. Therefore, the radiographic images (X-ray data) can be quickly displayed, processed, and stored in digital format on a computer. Such a digital technology enables more advanced and flexible algorithms for image processing, real-time analysis, and storage. Recently, portable radiography equipment has been studied for wind turbines [315], [316]. When placed on a robot platform, equipped with robotic arms, the X-ray system can climb wind turbine towers to deploy the X-ray inspection and scan a large area of structure automatedly [61], [317].
As listed in Table 7, the radiographic signal-based technique is mainly applied for the condition monitoring of wind turbine structures: rotor blades, nacelle, and tower. It is worth mentioning that, in addition to X-ray, Gamma-ray, as another radioactive source, can also be applied in the radiography of a wind turbine’s structure [318]. Table 19 summarizes the advantages and disadvantages of this technology as referenced in a selected list of studies, focusing on its implementation in wind turbines. In addition to health safety concerns, this technology presents certain challenges and limitations, mainly regarding possible technical complications and equipment-related costs. These shortcomings make difficult the development of a complete, commercially viable system composed of an automated platform able to scan a large-scale wind turbine structure.
In summary, the signal-based condition monitoring when applied in wind turbines, using the reviewed standalone condition-specific sensing and data acquisition technologies, can provide highly precise and specialized condition monitoring capabilities that are usually much beyond those of the SCADA-based techniques. However, this requires additional hardware investment, which increases condition monitoring costs, not to mention the measurement reliability concerns and common complications of equipment calibration, installation, or implementation. In addition, it is worth noting that the full signal-based condition monitoring of a wind turbine, including all its components, would require a well-selected combination of several condition-specific monitoring equipment and technologies rather than only one. From a cost-benefit viewpoint, using standalone condition-specific monitoring equipment in onshore wind turbines can be justified if the costs associated with the equipment replacement, labor, and lost production are fully considered, especially if generator and gearbox failures are prevented. In the case of offshore wind turbines, using such condition monitoring equipment can be justified if the abovementioned costs plus the site access/logistic costs are considered, especially if large subassembly failures (e.g., rotor, foundation, gearbox, and generator failures) are prevented. Having said that, the ultimate investment justification in professional condition-specific monitoring systems highly depends on their practical capability in detecting incipient defects, damages, or faults at an early stage to avoid full subassembly replacements, which is the costliest aspect of failures.
Mathematical Model-Based Condition Monitoring
Compared with signal-based techniques, model-based techniques require the “mathematical models” of a wind turbine or its major components. For this, one does not need to use high-resolution condition-specific signals, such as those used in signal-based techniques. Indeed, model-based techniques mainly require mathematical models of the process and the I-O information, commonly available from a wind turbine and mostly associated with the wind turbine’s control system. The required mathematical models are typically developed either as the so-called “nominal models” to describe the fault-free process for FDD purposes or as “degradation models” to describe the degraded process for LTP purposes. As summarized in Fig. 17, both nominal and degradation models can be mathematically obtained using either the so-called “theoretical/physical modeling” or “experimental modeling” approaches. In the theoretical modeling approach, the process model is derived on the basis of mathematically formulated physical principles or laws of nature. However, in the experimental modeling, the model is identified from the process measurements using system identification techniques in a way that the process I-O relationship is expressed in a mathematical model [322]. Depending on the problem to address, one can choose between either of these modeling approaches.
Once the best-suited modeling approach is selected, different design approaches, as listed in Fig. 17, can be used to develop a model-based condition monitoring framework. The most important model-based design approaches for FDD and LTP are briefly reviewed in the following.
A. Model-Based FDD Design
Model-based FDD schemes are commonly designed using the so-called residual-based, fault estimation, or set-membership approaches.
1) Residual-Based Approach:
Residual (or symptom signal) is a fault indicator computed in real time as a deviation between measurements and outputs of the mathematical model. Most often, a model-based FDD scheme is developed based on residual generation to simply monitor the level (or trend) of the residual signal through the so-called residual evaluation and to act when the signal reaches a prescribed threshold value [322]. The residual evaluation is carried out by different methods, such as fixed or adaptive threshold testing on instantaneous or moving average values of the residuals, statistical methods (e.g., generalized likelihood ratio tests), and fuzzy logic approaches. Indeed, the residual is designed to be small (“ideally” zero if no noise and model uncertainty are involved) under fault-free (normal) conditions and to deviate significantly from zero when a fault occurs. The residual-based approach is relatively simple to implement although, when designing the residual generation, the main challenge is to address the process noise and modeling uncertainty, not to mention achieving the necessary disturbance decoupling (i.e., to ensure that the residual is not affected by unknown inputs other than faults) [322]–[324]. The most frequently used residual-based methods include parity equations, state estimation methods (observers in deterministic problem formulation framework and Kalman filters in stochastic/random process problem formulation framework), parameter estimation methods (e.g., least-squares or recursive least-squares and regression analysis), joint state-parameter estimation methods (e.g., a two-stage Kalman filter and an extended Kalman filter), and I-O representations (also known as “data-driven model-based”) using neural network models, fuzzy models, neurofuzzy models, and so on. Most of the abovementioned methods are well established and extensively cited and explained in several reference books (e.g., see [322], [325], and [326]). In wind turbines, the residual-based FDD approach using the already mentioned methods has been largely studied. Table 20 presents an updated summary of the literature on the model-based FDD techniques, including the residual-based approach in wind turbines.
2) Fault Estimation Approach:
In most cases, it is either difficult or inadequate to use residuals alone to determine the magnitude of faults. The fault estimation approach involves the online estimation of faults and is usually more challenging than the residual generation approach. Such an estimation approach not only can detect fault signals but can also effectively estimate their magnitudes to provide accurate fault information to active fault-tolerant control systems [327]. Fault estimation can be achieved using a variety of observers (e.g., an adaptive observer [328], a sliding mode observer [329], an unknown input observer [330], and an extended state observer [331]) or Kalman filters (e.g., a zonotopic Kalman filter [332], a two-stage exogenous Kalman filter [333], and a three-stage Kalman filter [334]). When appropriately designed for a particular fault scenario, a bank of observers or Kalman filters can be used to isolate the faults from each other in different components. For instance, the application of fault estimation in wind turbine diagnosis and fault-tolerant control has been recently studied using an adaptive sliding mode observer [335], the Takagi–Sugeno sliding mode observer [336], and an extended state observer [337]. More references are also provided in Table 20. For further details on fault estimation methods, interested readers are referred to [327] and [338].
3) Set-Membership Approach:
This approach, also known as the “error-bounded approach,” relies on the assumption under which the noise, disturbance, and uncertainty in the model’s parameters are unknown but bounded with
B. Model-Based LTP Design
If the degradation model originates from the black-box modeling methods, such as neural networks, the obtained LTP is commonly classified under the data-driven LTP, as explained in Section V. However, when a degradation model other than black-box models becomes available, the measured monitoring data can be used to identify (or calibrate) model parameters. Once these model parameters are identified, it is possible to predict how faults or damages will grow in the future and thus obtain the RUL. In reality, however, the degradation model is not perfect, as the data used for model identification are always corrupted by measurement errors, noises, and variabilities; future loading or operating conditions are uncertain. These significant sources of uncertainty contribute to the ultimate uncertainty in estimated model parameters and, thus, the RUL prediction. Therefore, the key issue in model-based LTP is to find ways to improve the accuracy of the degradation model while incorporating uncertainty in the future [366], [367]. To address this issue, parameter estimation algorithms based on the so-called Bayesian approach are often used for the real-time estimation of the degradation model parameters. A major advantage of the Bayesian approach over other parameter estimation methods, such as the least-squares method and the maximum likelihood estimation method, is its capability to estimate the uncertainty structures of the identified model parameters [366]. Such uncertainty structures depend on those of the prior information and likelihood functions. Indeed, the Bayesian approach employs the Bayesian statistics and measurement data to probabilistically identify unknown parameters of the degradation model and reduce their uncertainty in real time. Therefore, most model-based LTP schemes have their foundation in Bayesian statistics. Among Bayesian-based algorithms, the overall Bayesian method [368] and filtering-based techniques, such as a Kalman filter [369], an extended/unscented Kalman filter [370], [371], and especially the particle filter [372], are more commonly known. In the overall Bayesian method, the model’s unknown parameters are estimated in the form of a posterior probability distribution, which is proportional to the likelihood of observed data multiplied by the prior probability distribution [368]. When applied, this method’s main challenge is to appropriately choose the right options for the sampling process (e.g., the initial values of unknown parameters and the width of proposal distribution in the Markov chain Monte Carlo method) [366].
As for filtering-based techniques, they rely on a recursive Bayesian update process, under which model parameters are updated recursively by accepting one new measurement data at a time. The performance of the Kalman filter and its modern extensions highly depends on the initial condition and variance of model parameters, as well as the type of nonlinearity and errors in linearization. However, the particle filter is much more flexible and easier to design with no restrictions on the type of system (nonlinearity) or noise. Indeed, it employs particles (samples) to represent the prior and posterior probability density function (pdf) of model parameters. More precisely, when a new measurement becomes available, the posterior pdf at the previous time step is used as the prior pdf at the current time step. Accordingly, the parameters are updated by multiplying the prior pdf with the likelihood from the new measurement. Thus, particle filtering is also referred to as the “sequential Monte Carlo method,” known to be the most popular method in model-based prognosis. Having said that, the computational cost and especially the so-called “particle depletion” phenomenon (i.e., accumulated sampling error during the updating process) have been major problems when using particle filtering [366], [372].
Compared with data-driven prognosis methods, model-based LTP has several advantages. First, model-based methods enable long-term prediction. Indeed, when the parameters of the degradation model are properly identified, the RUL can be well-predicted by propagating the model until degradation reaches a prescribed threshold. Second, the model-based methods mainly rely on the physics-based model and require a relatively small amount of data for the parameter estimation. However, the challenging issues to address include the adequacy of the degradation model, the accuracy and efficiency of parameter estimation, and the quality of required degradation data [366], [367].
In wind turbines, the application of model-based LTP is relatively recent and limited. This is mainly due to the particularly challenging aspects of prognosis in wind turbines. To name a few, the high nonlinearity and strong coupling of wind turbine components and their operation under the wide range of changing loads and operating conditions, along with the highly uncertain and variable onshore or especially offshore environmental conditions, are the most challenging problems to tackle. Table 21 presents an updated summary of the literature on model-based LTP techniques in wind turbines.
C. Model-Based Condition Monitoring Framework
A complete condition monitoring framework involves all aspects of data acquisition, FDD, and LTP to provide the overall health assessment of a plant such as a wind turbine. For instance, Fig. 18 shows a typical schematic of a complete model-based condition monitoring framework, including the data acquisition and both model-based FDD and LTP schemes via the commonly used residual-based and Bayesian approaches, respectively. Under a mathematical model-based framework, data acquisition is achieved by collecting measurement data usually from the sensors of a wind turbine SCADA system (especially the sensors associated with the wind turbine’s control system) and by processing them to obtain useful features for FDD. The model-based FDD is developed using an appropriate design approach (e.g., residual-based and fault estimation-based) to detect, isolate, and identify early symptoms of any fault or anomaly in the wind turbine components. According to the FDD information, the severity or size of a detected fault is usually quantified in the form of a normalized degrading health index (or health measure) for LTP purposes. Using such a health index, the model-based LTP is achieved based on the modeling of the degradation process in any wind turbine components while considering an appropriate Bayesian method to address the inherent large degree of uncertainty associated with the long-term predictions of a component RUL. As shown in Fig. 18, in addition to the FDD information, the LTP may also use some upcoming information expected about changes in the system’s operating conditions in terms of variations in the environmental conditions, usage patterns, loadings, and so on.
In summary, the most important advantage of mathematical model-based condition monitoring is its low development and running costs since there is no need for any additional hardware components. In many cases, the measurement data used to control the process also suffice for the model-based FDD and/or LTP algorithms, meaning that no additional sensors must be installed. Also, the model-based algorithms can be finally implemented via software on a process control microcomputer [322], [325], [366]. Another advantage of the mathematical model-based approach compared with its signal-based counterpart is that the FDD/LTP information under a mathematical model-based scheme can be more useful for real-time control reconfiguration and fault-tolerant control purposes [322], [384]. This enables the timely accommodation/compensation of faults in such a way that they do not lead to wind turbine failures. Having said that, it is worth noting the particular challenges and limitations of the design and widespread adoption of mathematical model-based condition monitoring schemes. The high nonlinearity of wind turbine components—especially the aerodynamic subsystems—together with the stochastic wind fluctuations and turbulences, measurement noises, and external disturbances under harsh environmental conditions all make the model-based condition monitoring of wind turbines very difficult in practice. It is, indeed, challenging to derive robust and accurate models that can be used for FDD or LTP purposes. Given this main limitation of model-based approaches, various combinations of hardware signal-based and mathematical model-based techniques can also be considered. Such combinations are sometimes referred to as “hybrid” techniques, crossing the boundary between model- and signal-based techniques, in the hope of integrating both techniques’ advantages to overcome the challenges and limitations of each technique alone. For instance, from a design architecture viewpoint, the knowledge of a system’s physical behavior can be utilized to determine a mathematical model (e.g., determining the order of exponential or polynomial functions) in a data-driven framework. Also, it is possible to use a system’s data-driven model along with a physics-based fault model or vice versa. It is worth mentioning that, when possible, the information used in hardware signal-based techniques can be added to the information about control signals and the wind turbine model to enable more optimized and comprehensive monitoring capabilities. This helps improve not only the reliability and performance of the CMS but also the feasibility of fault-tolerant control designs in wind turbines. Finally, a number of examples of hybrid techniques for wind turbine FDD and LTP are reported in [75], [385]–[388] and [76], [77], [389], respectively.
Summary, Conclusion, and Future Trends
Due to their complex integrated nature and wide range of changing loads and operating conditions, onshore and especially offshore wind turbines are prone to component faults and premature failures that jeopardize their reliability and availability (uptime) for efficient energy production [3], [9], [57]. In order to detect, diagnose, and prognose any type of abnormalities or faults in wind turbines’ components before they can propagate and cause major damage or severe failure, it is crucial to effectively monitor wind turbines’ conditions in real time. Motivated by the significance of this issue and the ever-increasing role of wind turbines in the modern world’s power grid, wind turbine condition monitoring has been at the forefront of both academia and industry over the past two decades. Considering the great diversity of approaches and techniques developed in wind turbine condition monitoring and the large number of results disseminated in this active research area as of 2020, this article aims at providing an up-to-date, comprehensive review of the available literature with particular attention paid to the results reported in the last decade.
This article summarized the major fault and failure modes observed in wind turbines and comprehensively reviewed the available techniques and strategies for wind turbine condition monitoring from signal- to model-based perspectives. Both aspects of condition monitoring, that is, fault diagnosis and prognosis, were considered, and the respective functionalities, capabilities, advantages, flaws, and challenges/limitations of each available technique were carefully outlined and explored, especially in view of wind turbine condition monitoring issue. According to the reviewed literature, the following holds.
Signal-based condition monitoring: This mainly involves measurement signals and signal processing methods under a data-driven approach designed to obtain useful FDD and/or LTP information from a large amount of observed data. Depending upon the measurement signals available from a wind turbine, this article categorized and reviewed the techniques for signal-based condition monitoring under the terms of “SCADA-based” and “condition-specific-based” techniques that rely on the signals coming from the SCADA and standalone condition-specific data acquisition systems, respectively. This article reviewed several SCADA-based techniques, such as trending, clustering, and data-driven normal/damage behavior modeling. In addition, this article outlined and reviewed a wide range of condition-specific techniques using vibration, strain, torque, shock pulse, acoustic emission, temperature, oil debris/quality parameters, electrical effects, machine vision, ultrasound, thermography, and radiography according to the additional hardware used in a professional CMS.
Model-based condition monitoring: This employs the mathematical models of a wind turbine or its major components without requiring high-resolution, condition-specific signals, such as those used in signal-based techniques. Indeed, the model-based techniques mainly rely on the processing of mathematical models and the I-O information that can be easily retrieved from a wind turbine as they are mostly associated with its control system. This article outlined the commonly used design approaches for both model-based FDD and LTP in wind turbines. Regarding the model-based FDD design, the so-called residual-based, fault estimation, and set-membership approaches were described and reviewed. Furthermore, this article examined the specific application of the Bayesian approach to the real-time estimation of the degradation model parameters in the model-based LTP design for wind turbines.
In addition, any combination of both signal- and model-based techniques was categorized under hybrid techniques, leveraging the monitoring performance by integrating together the advantages of both signal- and model-based techniques to overcome the challenges and limitations of each technique alone.
Drawing on the reviewed literature, the following challenges and shortcomings call for additional research and development in the future:
Data acquisition: This is the first, thus essential, the step of any condition monitoring scheme. In wind turbines, a wide range of sensors and devices are used to measure the wind turbine’s environmental, operational, and performance parameters. Optimal selection, placement, and implementation of these sensors and devices play an important role in the overall monitoring performance, and in both capital and O&M costs. Improper sensor selection or placement can easily degrade the monitoring performance. Likewise, inaccurate calibration or implementation/installation of sensors may cause measurement errors, serious reliability concerns, or even equipment failures. When designing any CMS, the sensor reliability issue and the possibility of sensor failures must be considered. When possible and appropriate, depending on the type of measurement, efficient utilization and management of redundancy (in hardware, software, or even communication networks) can be contemplated. In addition, the application of smart sensor technology along with the Internet of Things (IoT) can offer significant advantages for remote real-time data acquisition and transmission throughout large wind farms.
Data analysis: To enable meaningful and effective condition monitoring, especially when based on data-driven approaches, it is important to collect and analyze a sufficiently large and representative (complete) amount of observed data. With an average lifespan of twenty years, systems such as wind turbines presumably offer huge amounts of easily collectable data, which makes the adoption of data-driven monitoring approaches all the more appealing. Having said that, beyond the challenges commonly posed by big datasets and the extraction of useful features, the quality and completeness of such data usually represent the major points of concern, particularly when originating from SCADA systems. These concerns become even more significant in new wind turbines, operating most often normally at the beginning of their operational life. In that case, the available historical databases are usually limited (incomplete) and do not cover the entire range of fault-related features useful for designing FDD and especially LTP solutions since the recorded data only characterize the wind turbine’s normal operation without any information about other operating modes (i.e., faults and failures). When possible, using appropriate physical models, referred to as “digital twins,” to generate the database that covers the system’s useful features in normal and faulty operations can be considered a potential solution to tackle this issue [390]. In addition, the emerging paradigm of data-centric AI, where giant databases simply do not exist, could be part of the solution [391].
Signal-based condition monitoring: SCADA-based techniques are useful to identify abnormal turbines within a wind farm through the overall monitoring and tracking of key environmental, operational, and performance parameters but are usually limited when performing a full detailed) condition monitoring of a single wind turbine’s subsystems and components. Given this shortcoming and the diversity of failure modes in wind turbine components, an integration approach starting with a digest of SCADA data and accordingly fusing several dedicated condition-specific-based techniques considering their capabilities, advantages, and disadvantages is recommended. With an ever-increasing number of offshore wind turbines and their inevitable accessibility limitations in harsh offshore conditions, it is crucial to embrace the power of AI and advanced machine learning capabilities (e.g., deep learning). Along with the latest advancements in robotics, UAVs or drones, and other autonomous system technologies, they remotely enable autonomous data collection, and fleet-wide condition monitoring and asset management using effective and fully autonomous condition-specific-based techniques, such as machine vision, thermography, and radiology, to name only a few.
Model-based condition monitoring: Model-based techniques do not require any additional hardware components and can provide very cost-effective solutions for wind turbine condition monitoring. However, the diagnosis and prognosis performance obtained by model-based techniques is strongly tied to the accuracy, thoroughness, and robustness of their mathematical models, which are used to describe the nominal or degradation behaviors of a wind turbine or its major components. Since a wind turbine is an integrated complex system built with highly nonlinear components working under stochastic wind fluctuations and turbulences, external disturbances, measurement noises, and harsh environmental conditions, it is often challenging or even impossible to identify system dynamics across the entire wind turbine’s operation regime. This makes the wind turbine modeling problem a serious challenge. Depending on the problem tackled, advanced modeling approaches, such as those using integrated multiple linear models or those based on hybrid modeling approaches (merging white and black-box modeling), can be considered as potential solutions to cover all possible system operating ranges. This does not mean that new modeling methods are no longer needed. In fact, recent studies have explored the design of model-based FDD with regard to nonlinear dynamic systems. However, it should be noted that most nonlinear designs are quite complicated and can only be applied to a very limited class of nonlinearities while relying on highly strict assumptions about the system nonlinearity to be implemented.
Hybrid approaches: Given the limitations of signal- and model-based approaches, their various “hybrid” combinations can also be considered, crossing the boundary between the signal- and model-based approaches in the hope of integrating the advantages of both approaches together and overcoming the challenges posed by each approach alone. By establishing a promising framework to leverage the merits of different condition monitoring techniques, hybrid approaches can be more appealing to the industry in the short run while providing both companies and researchers with valuable practical experience in the long run when it comes to the real-life evolution of fully model-based solutions for wind turbine condition monitoring.
Monitoring system architecture: A barrier to a hybrid application of different condition monitoring approaches under a truly integrated system is the common architecture, where available CMSs are segregated from each other. This phenomenon is mainly due to a dissociation between the original manufacturers of wind turbine components and those of their monitoring equipment. For instance, SCADA-based signals and alarms are both generated from within the industrial control system network of wind turbines, supplied by turbine manufacturers, whereas professional CMSs are purchased separately and installed on wind turbines independently of their control system. Therefore, it is physically difficult to integrate the SCADA-based and professional CMS signals, in spite of their different bandwidths. Having said that, some wind turbine control system manufacturers have been expanding SCADA and professional CMS signal facilities within their products, where fault monitoring algorithms and alarm handlers can operate based upon both SCADA-based and professional CMS signals data. As such, more flexible system architecture and integrated condition monitoring opportunities can be achieved.
Multiparameter monitoring: An important aspect of condition monitoring that concerns wind farm operators is the reliability of the monitoring equipment/technique and the quality of its generated FDD/LTP information. The former can be addressed by experience and the appropriate selection of sensing and data acquisition systems, whereas the latter depends on the accuracy and the way the condition data (information) are presented to the outside world. It is obvious that, when a number of monitoring signals from different sources (e.g., both vibration and oil debris/quality for gearbox bearings) present confirmatory fault information, this is useful and builds confidence among operators and O&M crew. Any type of condition monitoring sensor signal (e.g., vibration, strain, and temperature) has a probability of detecting and identifying faults in a wind turbine component. Indeed, the probability of accurate FDD/LTP depends in part on the sensor location and in part on the reliability and accuracy of the sensor. According to the reviewed literature, relying on more than one monitoring sensor (e.g., multiple sensors in different locations and of different types) or multiparameter monitoring often improves the chances of successful detection and diagnosis of incipient damages or faults at an early stage. However, this can result in a data overload (as commonly seen in the wind industry), not to mention that its benefits may diminish if more than a sufficient (optimum) number of sensors are used. For instance, although two sensors may improve the probability of accurate FDD/LTP results, increasing, say, from six to seven sensors provides a much smaller improvement. Therefore, wind farm operators are advised to reduce the number of sensors (when possible) but to increase their quality and reliability. Overall, it can be reasonable to explore higher integration of the interpretation of monitoring signals among different systems (e.g., between SCADA-based and professional CMSs, such as in [392]) with the objective of enabling accurate and early fault detections and enhancing the prognostic horizon.
Prognosis: This will be an important function of future CMSs. Compared to wind turbine diagnosis, the studies on prognosis are still in their infancy. This is mainly due to the challenging nature of prognosis problems in the view of the high complexity, nonlinearity, and uncertainty of wind turbine systems, not to mention the technical challenges to obtain the useful historical data (especially up-to-failure data) needed to identify the degradation progress. This research area certainly deserves further attention and requires more research efforts since prognosis presents a huge potential for enabling more effective condition-based maintenance and the reduction of O&M costs in both onshore and offshore wind farms.
Evolving new technologies and offshore wind: Wind turbine technology is rapidly evolving to reduce weight, control loads, and improve energy production. With the increasing tendency toward larger and more flexible wind turbines in offshore installations, the O&M costs will quickly increase unless reliability and availability are improved through the remote real-time condition monitoring and health management of wind turbines. Indeed, the harsh offshore conditions may impose unknown complexities and new challenges for the condition monitoring of wind turbines, especially in terms of accessibility and logistics. Therefore, adapting and tailoring wind turbines’ condition monitoring solutions to specific offshore conditions need to be researched in an exclusive and extensive manner. Having said that, recent advances in wind farm digitalization, wireless communications, supercomputing technology, and sophisticated atmospheric measurement capabilities and large real-time streams of data (big data) collected from turbine-based systems (both the turbines and meteorological measurement equipment) provide excellent opportunities to significantly improve wind turbine condition monitoring and control as never before.
Farm-level condition monitoring: Depending on the type of fault and failure modes, condition monitoring strategies can be carried out either at the levels of an individual wind turbine or an entire wind farm. However, thanks to the multiple measurements offered by the different wind turbines found in a wind farm and the simultaneous and collective consideration of those measurements in the condition monitoring process, some faults may be diagnosed more easily at the wind farm level (e.g., the rotor aerodynamic malfunctions or faults related to icing or debris built up on the rotor blades). An example of this approach is presented in [360] although there is much more room for future research in this area.
Cybersecurity: Wind farms’ cybersecurity is becoming increasingly important. While wind farm digitalization can offer various benefits, it can also make wind farms more vulnerable to “cyberthreats,” which constitutes a new dimension of health risks associated with wind farms operation, the others being brought by physical faults or damages. As a matter of fact, cyberthreats in the form of cyberintrusions or cyberattacks on wind energy systems have been already reported in recent years. For instance, a cyberattack in March 2019 exploited a vulnerability within a firewall and resulted in a denial-of-service (DoS) condition and, consequently, the disruption of communications between a control center and wind and solar generation sites for a large wind owner/operator in Utah, USA [393]. Given the expansional role of digitalization of wind farms empowered by wireless communications that come with inherent cybervulnerabilities, such examples of cyberattacks will probably continue to increase in sophistication and number, resulting in severe cascading failures impacting not only the cyber and physical components and operations of a wind farm, but also the reliability of the entire power grid. It is worth noting that, although faults and cyberattacks originate from different sources, they both might have similar signatures that ultimately lead to the system’s failure. Consequently, it is essential to consider appropriate technologies exclusively developed for the detection, identification, and mitigation of cyberattacks, especially in order to differentiate cyberattack- and noncyber-related/operational incidents in wind energy systems. This helps to correctly identify the type of incident and its root cause(s), which is necessary to achieve effective and reliable condition monitoring, control, and health management of wind turbines. Additional considerations, requirements, and guidelines regarding the cybersecurity aspects of wind turbines and wind farms can be found in a recently published roadmap for wind cybersecurity in [394].
Condition monitoring and system protections: Wind turbine system’s protections are necessary to ensure safety and system integrity in the event of serious uncontrollable fault effects or emergencies. However, a very conservative application of system protections may shut down a wind turbine too early, even before it is necessary and justified by the condition-based maintenance enabled by the condition monitoring process. Such early shutdowns of wind turbines by system protections may potentially restrict the benefits obtained from condition monitoring and condition-based maintenance. To further explain, an effective and reliable condition monitoring scheme can detect, diagnose, and prognose incipient signs of degradation or faults before they propagate to major damage or severe condition where system protections are inevitable. In other words, wind turbine condition monitoring information provides wind farm operators with enough time span to plan condition-based maintenance actions (see Fig. 12) and even enables control adaptation and reconfiguration mechanisms (see Fig. 8) to safely increase wind farm availability. As a result, wind turbine condition monitoring can enhance confidence among operators to possibly further fine-tune the tolerance for activation of some protections and, thus, lower the probability of unnecessary wind turbine shutdowns to help enjoy improved production (availability).
Health management: Finally, it is worth mentioning that the complete and autonomous health management of wind turbines can be fully realized through a smart and integrated design of condition monitoring, control, and intelligent decision-making strategies under a unified framework (see Fig. 8). Currently, the FDD and/or LTP information obtained from CMSs is mainly used for wind turbine general health assessment and possible maintenance recommendations after being analyzed and interpreted by expert assistance. Yet, the exploitation of real-time condition monitoring information when a fault happens in a wind turbine component still shows great potential to enable appropriate control reconfiguration under active fault-tolerant (or self-healing) control strategies [42], [323], [356], [357], [360], [361], [395]–[397]. This offers an inexpensive technology that enables large wind farm operators to monitor and extend their wind turbines’ operations (availability) by accommodating fault effects before they propagate to failures, which improves reliability, productivity, and planning for condition-based maintenance. Although some studies have already initiated the exploration of the abovementioned potential to some extent, the research on wind turbine fault-tolerant control, particularly the robust integrated design of model-based FDD/LTP and control reconfiguration, is yet scarce and open. In addition, the research areas dealing with intelligent strategies for data interpretation, cautioning, and automated decision-making (e.g., using expert systems), which aim at delivering meaningful condition-based maintenance, are yet open to further explorations.
ACKNOWLEDGMENT
As is the case in many collaborative efforts, this work has also beneficiated from formal and informal discussions with several colleagues throughout industry and academia. The authors wish to express their appreciation to the cited and noncited researchers around the world, who all contributed to the thriving field of wind turbine condition monitoring. They extend their sincere thanks and appreciation to the Editor-in-Chief, the Associate Editor, and the five anonymous reviewers whose constructive comments and suggestions during the review process helped to improve the quality of this article considerably.