Review of multiscale methods for process monitoring, with an emphasis on applications in chemical process systems

Process monitoring has played an increasingly significant role in ensuring safe and efficient manufacturing operations in process industries over the past several years. Chemical process data is highly correlated and has multiscale characteristics in general. To overcome this concern, extensive work has been made for multiscale process monitoring for process plants during the past two decades. The recent success of multiscale methods in monitoring and controlling manufacturing processes has sparked interest in investigating these methods for process monitoring. This article aims to present a concise and critical overview of the applications of multiscale process monitoring methods in chemical processes. First objective is to identify the importance of multiscale methods for process monitoring. The second and main objective is the statistical and critical analysis for methods implementation, application area, types of data used, and various issues mentioned by previous researchers. In addition, the most important critical issues have been identified, and the capabilities and limitations of each method are discussed and highlighted. The reported literature focused mainly on fault detection and did not investigate the root-cause diagnosis of the detected faults. Further, the challenges and prospects in multiscale process monitoring in the chemical process industry have been featured for advancement.

. Features of the process monitoring system [8].
Multivariate statistical process monitoring (MSPM) techniques are capable and are increasingly implemented in monitoring chemical processes [17][18][19]. The key idea behind the MSPM techniques is to extract process features through a specific multivariate analysis process. Highly dimensioned information is then projected into less dimensional space, and the statistics are evaluated. Leading MSPM techniques are the principal component analysis (PCA) [14,20,21] and partial least squares (PLS) [22,23], commonly used for the monitoring of the chemical processes. Although, these techniques for monitoring chemical processes have been very effective. These techniques have certain limitations, such as the presumption of linear relationships among variables, as essential details can be overlooked when nonlinear systems are considered. However, most of these assumptions can easily be infringed in reality. Therefore, several improvements of MSPM techniques for process monitoring have been made in recent years. Although conventional MSPM techniques and extensions have been successful in many practical situations, they are generally limited to the single-scale analysis of events corresponding to the sampled frequency. Most existing methods are based on fixed-scale data, while the multiscale scheme uses decomposition techniques to depict data on several scales. However, a systematic review of these recently developed MSPM methods has not been reported yet.

C. PREVIOUS REVIEWS AND THE AVAILABLE GAP
Many excellent review articles in process monitoring have been published in the past. Fault diagnostic approaches based on quantitative models [6], qualitative models [7], and historical process knowledge [8] have been thoroughly analyzed in a series of papers. Qin [24] has reviewed databased process monitoring methods for fault detection, identification, reconstruction, and diagnosis. Ge et al. [13] reviewed data-based process monitoring methods for nonlinear, non-Gaussian, multimode, and dynamic processes. Gao et al. [25,26] have systematically investigated the use of fault diagnostic methods. A comprehensive bibliometric analysis of data-based fault detection and diagnostic methods for process systems has recently been presented [27]. Nor et al. [28] have reviewed data-based FDD methods for chemical processes. Other significant data-based fault detections and diagnostic studies have also been reported [29][30][31][32][33][34]. The above reviews cover different aspects of data-based fault detection and diagnosis methods by various aspects, as shown in Table 1.
The purpose of this review paper is to guide the selection of multiscale fault detection and diagnosis procedures. As mentioned earlier, none of the available reviews covered the multiscale process monitoring methods. Therefore, this review paper aims to provide a comprehensive insight into multiscale fault detection and diagnosis methods for chemical processes. Thus, this work offers excellent knowledge for those interested in developing a multiscale fault detection and diagnostic framework for the chemical process. It would serve as inspiration for the future valuable addition in the state of knowledge relevant to recent developments in fault detection and diagnosis in chemical processes.
The rest of the paper is structured as follows. The motivation for multiscale process monitoring is provided in Section II. Multiresolution techniques used in multiscale process monitoring methods are discussed in Section III, followed by a statistical analysis of multiscale fault detection and diagnosis methods in Section IV. A detailed review of multiscale process monitoring methods based on the promising issues has been discussed in Section V. Finally, some challenges and future recommendations are discussed in Section VI, followed by the findings of this review article.  Das et al. [30] Review and categorize various process monitoring and fault detection techniques and into data-based, model-based, and hybrid approaches. 2013 Zheng et al. [36] Proton Exchange Membrane Fuel Cell system Review and comparison of non-model-based methods, including artificial intelligence, statistical, and signal processing. 2013 Ge et al. [13] Industrial processes Review of recent developments in data-driven process monitoring methods for industrial processes focused on considering different data characteristics, including non-Gaussian, nonlinear, time-varying, and multimode, dynamic, and batch processes. 2014 Yin et al. [31] Industrial Process Reviews data-driven process monitoring and fault diagnostic methodologies from an application point of view in many industrial processes and provides a basic framework for monitoring under different industrial operating conditions. 2015 Agrawal et al. [37] Coal mills Comparative study of various control and fault diagnostic techniques, including quantitative, signal, qualitative, and process-historical approaches to coal mills and possible directions for future research. 2016 Severson et al. [4] Industrial systems Review of process monitoring methods and discuss current issues and promising future directions. 2016 Tidriri et al. [33] Process industries Comparative analysis of model-based and data-based process monitoring methods highlights their features and points out the benefits and limits of each approach. 2017 Ge [34] Industrial processes Review of data-driven process monitoring methods focused on addressing plant-wide process issues. 2018 Alauddin et al. [27] Process Systems Bibliometric analysis of data-driven FDD approaches addresses key areas, contributing authors, key sources, and actively involving countries in this research area. 2019 Nor et al. [28] Chemical process systems Review data-driven FDD frameworks and their challenges and guide applying such methods in chemical processes. 2019 Jiang et al. [38] Industrial Processes Review of data-driven multivariate process monitoring techniques for industrial plant-wide processes, emphasizing large-scale and multi-unit operations. 2020 Nor et al. [39] Chemical process systems Review of data-driven fault detection and diagnosis methods. 2020 Park et al. [40] Industrial processes Review of current research and developments of FDD approaches for process monitoring. 2021 Taqvi et al. [41] Process industries Overview and categorize data-driven approaches in fault detection and diagnosis in process industries. 2021 This Review Chemical process systems Review of multiscale process monitoring methods and their applications in chemical process systems. Some challenges and future recommendations were also discussed.

II. MOTIVATION FOR MULTISCALE PROCESS MONITORING METHODS
Multiscale process monitoring is an important extension of the statistical process monitoring methods used for highly correlated, noisy data. These methods have been widely used to monitor chemical processes in recent years. The motivation for the multiscale process monitoring is presented in the following sub-sections.

A. UNIVARIATE STATISTICAL PROCESS MONITORING
Univariate statistical process monitoring methods, often known as statistical process control (SPC) methods, evaluate each variable individually [42]. Walter Shewhart invented the first SPC chart, which is used for process monitoring without the usage of a filter. This chart identifies typically significant faults. Other SPC charts, such as the cumulative sum (CUSUM) and exponentially weighted moving average (EWMA), detected minor faults using linear filters. While these techniques continue to be prevalent in the process industry, their efficacy degrades when highly correlated variables are used [43]. Multivariate extensions of Shewhart control charts are used when process parameters of the underlying process are known or unknown. Multivariate CUSUM (MCUSUM) and multivariate EWMA (MEWMA) have been developed for the detection of small changes [44] and have provided unsatisfactory results for highly correlated process variables [45].

B. MULTIVARIATE STATISTICAL PROCESS MONITORING
Multivariate process monitoring (MSPM) methods can be used to evaluate highly correlated and high-dimensional process data. The main concept of MSPM methods is the characteristics of the process that can be achieved through a particular analysis process. Thus, higher dimensional information is projected into a less dimensional space, followed by the evaluation of statistics [46]. Figure 3 shows the classification of well-known MSPM methods. PCA and PLS are the most widely used MSPM methods. PCA-based monitoring methods consider all process faults, while PLS-based monitoring methods emphasis on qualityrelated faults. As the complexity of industrial processes have increased due to the recent technological advances in modern industry, it is necessary to ensure process safety, product quality and production efficiency [2]. Therefore, quality related process monitoring is much more significant than simply monitoring the fluctuations and anomalies of process variables [47,48]. Both methods presume that the Gaussian distribution obeys the data. Independent component analysis (ICA) requires high-level statistics to solve non-Gaussian problems. More important information can be revealed in non-Gaussian data [49]. Gaussian Mixture Model (GMM) is another way of handling non-Gaussian data by treating a complex process as a linear combination of several Gaussian models [50].

C. RECENT DEVELOPMENTS IN MSPM METHODS
Generally, there are two types of processes are used in industry including batch processes and continuous processes. The process data obtained from process industry exhibit multimodal distribution, dynamics, nonlinear relationships between variables, non-Gaussian, and time-varying and multiscale [13]. Several enhancements to conventional MSPM approaches have been made in recent years, and many other data-based methods have also been introduced for process monitoring [51][52][53].
In practical application, the process data are always contaminated by random noises. Therefore, it is required that process monitoring should also be carried out through a statistical manner, and the monitoring decisions are made through a probabilistic way. To address this problem, PCA based monitoring method formulated into a probabilistic framework known as probabilistic PCA [54]. In the probabilistic models a unified likelihood based monitoring statistic is used instead of T 2 and SPE control charts [55]. Furthermore, PPCA framework has also been extended to handle non-Gaussian data to improve the fault detections [56].
The process data collected from chemical processes usually involve high noise levels and autocorrelation and may also vary from normality and impact MSPM process monitoring methods [57]. Such techniques are also based on a single-scale representation of measurements and cannot capture the information from multiscale representations of measurements [58]. Wavelet-based multiscale process monitoring methods have been developed to address these problems. Process monitoring models have been developed in these techniques by using wavelet coefficients at each scale [58][59][60]. Instead of wavelet transforms (WT), some researchers used empirical mode decomposition (EMD) and singular spectrum analysis (SSA) to decompose the process variables before MSPM methods [61,62]. EMD and SSA both are merely relying upon data-adaptive basis functions. Thus, these techniques are more helpful in analyzing nonstationary signals emanating from nonlinear systems [63]. Multiscale process monitoring techniques have effectively been used to analyze chemical processes over the last two decades. Various multiscale process monitoring techniques have been applied based on the process data obtained from different chemical processes.

III. MULTIRESOLUTION TECHNIQUES IN MULTISCALE PROCESS MONITORING METHODS
The demand for operational safety and product quality are critical issues in modern industrial processes. Although, the widespread use of sensor networks, advanced data acquisition technology, and the extensive use of distributed control systems (DCS) have added significant benefits to all process industries [45,64]. They are becoming increasingly integrated, automatic, more complex, and intelligent operations. These developments in modern industrial processes increase process monitoring systems [2]. Conventional MSPM techniques and their extension focus on analyzing single-scale phenomena, typically the sampling frequency. Therefore, the applications of these techniques are restricted to only a single scale and cannot derive the amount of information from process data showing multiscale phenomena [65]. The multiscale approach can obtain information through different decomposition techniques in different scales.
WT is the most effective multiresolution analysis (MRA) tool and helps decompose the original process measurements into their multiscale components according to time and frequency characteristics [14]. The process signals, which have distinct physical patterns or disruptions, decompose, and are viewed as several signals at different resolution scales.
The scaled version of the original signal is achieved by projecting it on an orthogonal signal to obtain coarse approximate coefficient scale and is given as [66]: where, k and j are discretized translation and dilation parameters, respectively. The discrete wavelet function for detail scale is given as [66]: The coarse approximate and detail signal coefficients are computed using the low pass filter (H) and high pass filter (G) given as [67]: where, as and ds are the approximate and detail scale coefficients, respectively. The original signal can be obtained by computing the sum of the last scaled signal and all detail signals: where j and n are the level of decomposition and original signal length, respectively.
In WT, determining the optimal decomposition level is important. At the highest decomposition level, the approximation function adequately reflects the actual deterministic signal with the least amount of noise. Each variable in a multivariate scenario may have a distinct optimum decomposition level. For computational simplicity, only a single decomposition level will be applied to all variables in most practical applications. As a result, the decomposition level chosen must be appropriate or optimal in order to ensure that the underlying characteristics of each variable are appropriately retained in the approximation function with the least amount of noise [68].
Multiscale representation of signal up to level 3 is illustrated in Figure 4. First, the original signal is decomposed into approximation and detail coefficients. The approximation function low-frequency signal, which contains the essential underlying deterministic features. The detail function includes the high-frequency component, which is mainly noises. The approximation function is further decomposed into even coarser approximation until the average signal has been approximated. This reconstruction perfectly composes the original signal if all wavelet coefficients are used. Recently EMD has attracted much attention when decomposing the time-series signal into different time scales. Unlike wavelet-based algorithms where the signal is decomposed in the transform domain, EMD adaptively sets the decomposition functions directly from data instead of using a fixed wavelet function across the entire analysis; therefore, this algorithm is a better choice for handling data collected from non-stationary processes [63]. The following are the two conditions that need to be met for a component to be considered an intrinsic mode function (IMF) [69]: 1. Total zero crossings and the total extrema in the whole data set should be equal or vary by at most one. 2. The mean value of the envelope from maxima and minima should be equal to zero at any component interval.
Among these decomposition frameworks, the WT has dominated the publication landscape over the years and will be referred to more extensively.

IV. STATISTICAL ANALYSIS OF MULTISCALE PROCESS MONITORING METHODS
The available literature based on multiscale process monitoring has been statistically reviewed and summarized in Figure 5. Various fault detection and diagnosis techniques have been used for multiscale process monitoring. These include conventional process monitoring methods such as (CUSUM and EWMA), multivariate (PCA and PLS), and their various extensions. Figure 5(a) shows the most widely used methods involved in multiscale process monitoring. PCA is the most widely used method in multiscale process monitoring, followed by KPCA, PLS, KPLS, NLPCA and KFDA.
Validation of multiscale process monitoring approaches has been done using various applications. Figure 5(b) illustrates the most often used applications in this field of study. The TE process is widely used by researchers, with a share of about 22.22%. The CSTR system, Industrial processes, and simulated numerical data are the next most used application areas, with 15.15%, 16.16%, and 14.14%, respectively.
The performance of the multiscale process monitoring methods was evaluated using process data from various diverse application areas. Figure 5(c) shows the distribution of data types used in multiscale process monitoring. Two types of datasets have been used for multiscale process monitoring, including real-time and simulated data. Figure  5(c) shows that the portion of the real-time dataset used is only 23.23%, acquired from either industrial processes or pilot plants. On the other hand, the rest of the portion includes simulated datasets. The characteristics of simulated datasets usually are known, which can help highlight the effectiveness of a specific method.
Various issues arise in the application of multiscale process monitoring. The major issues identified based on careful study of the research articles related to multiscale process monitoring are presented in Figure 5(d). The figure shows the percentage of papers that dealt with each of them. Although some of these issues are not unique to multiscale process monitoring methods alone, we are reviewing them within the context of multiscale process monitoring. Research articles based on multiscale process monitoring are devoted to discussing these issues. A list of all the research articles reviewed is then provided in Table 2. The table also shows the decomposition technique used, the method used, the case studies used, and, more importantly, the issues addressed. The purpose of this table is to help the reader choose a specific issue of interest and to browse the column for papers that deal with it.

V. REVIEW OF MULTISCALE PROCESS MONITORING METHODS
As identified and presented in Table 2, significant issues related to multiscale process monitoring are discussed one by one in this section. We first motivate why they are important and then give examples of how many researchers have addressed them over the years.

A. MULTISCALE METHODS FOR QUALITY-RELATED PROCESS MONITORING
MSPM methods are more beneficial for extracting meaningful information from the highly correlated process and quality variables because quality variables are measured at lower frequencies and typically have significant time delay [70]. Monitoring quality variables is essential for preventing system breakdowns and substantial financial losses. A few researchers have also developed a quality-related multiscale process monitoring technique. Partial least squares (PLS) is the MSPM method associated with quality-relevant monitoring, and it finds a relationship between the process and quality variables [71]. Teppola and Minkkinen [72] proposed a quality-related multiscale process monitoring scheme combining wavelets with PLS. The PLS model is constructed based on filtered measurements obtained by removing low-frequency scales in this approach. Lee et al. [73] proposed a multiscale technique combining PLS and WT for sensor fault detection. The feasibility of the proposed technique was confirmed by using the real industrial dataset from the biological wastewater treatment process. The monitoring results were also compared to those of the standard PLS method.
Madakyaru et al. [74] proposed a MSPLS model based on generalized likelihood ratio (GLR) tests. In this approach, a modeling framework is created by integrating WT with PLS, and then GLR testing is used to improve the fault detection. The proposed methodology proved immensely influential in the early detection of minor faults with incipient behavior in distillation columns. Similar work is proposed by Botre et al. [75], where efficiency and robustness are demonstrated through simulated continuous stirred tank reactor (CSTR) and Tennessee Eastman process (TEP) data.
Zhang and Hu [76] proposed a multiscale KPLS (MSKPLS) method combining kernel PLS (KPLS) and wavelet analysis for investigating the multiscale nature of the nonlinear process. The feasibility of the proposed method was tested for a real industrial data set, and the process monitoring abilities were compared with the standard KPLS method. VOLUME XX, 2017

B. MULTISCALE METHODS FOR NONLINEAR PROCESS MONITORING
Multiscale process monitoring frameworks using conventional MSPM methods have been used effectively in the process industry. Conventional MSPM methods underperform in complex industrial processes with nonlinear features due to their assumption of linear correlations in the process data. In recent years, nonlinear process monitoring has become a hot area of research in this field, and some nonlinear multiscale approaches have been developed. Detailed reviews of multiscale nonlinear methods are.
Shao et al. [77] proposed a multiscale NLPCA process monitoring approach based on input-training neural network (IT-NN) where non-parametric control limits were employed instead of linear control limits to improve online performance monitoring. This technique was modified using multi-level wavelet decomposition to enhance the process monitoring [78]. Geng and Zhu [79] proposed an adaptive multiscale NLPCA approach to monitor slow and weal changes in process variables. Maulud et al. [68,80] have developed a new multiscale approach using optimal wavelet decomposition and the orthogonal NLPCA. They only use approximation and highest detail functions, simplifying the overall model structure and improving interpretation at each scale. In this work, optimal decomposition level was determined by a PCA based graphical method.
The kernel learning methods recently received significant attention in the chemical industry and have been coupled with conventional MSPM process monitoring methods [81][82][83][84]. Several researchers have proposed KPCA and KPLS based multiscale nonlinear process monitoring methods [85,87]. Choi et al. [85] proposed a new multiscale nonlinear process monitoring technique using KPCA to detect and identify faults. This approach has been extended by Deng and Tian [86] to nonlinear dynamic processes that can effectively extract autocorrelation, cross-correlation, and nonlinearity from the process data. Zhang and Ma [87] further developed this approach to improve diagnostic capabilities. Further study proposed a nonlinear system monitoring approach based on KPLS at different levels. Zhang and Hu [76] have extended this approach to monitoring online processes in nonlinear processes.
The FDA does better than the PCA approach to classification problems in many cases. Although it shows limited performance in nonlinear systems due to its linearity, it is better suited to classification problems [88]. Liu et al. [89] proposed a multiscale classification method to obtain the most discriminatory characteristics of the scale. The effects of feature extraction investigated the classifier performance, and a multiscale classifier was developed to classify the faults better. This method can be applied to relatively large multi-class issues. Nor et al. [90] proposed a novel multiscale approach by combining KFDA with wavelets for nonlinear process monitoring. In this approach, XmR and T 2 statistics used fault detection. This approach was further extended to enhance the performance of fault classification and developed a robust multiscale feature extraction and fault classification method [91].

C. MULTISCALE METHODS FOR DYNAMICS PROCESS MONITORING
Due to random noise and process disturbances, a dynamic relationship exists among process variables in modern chemical processes. Information on this dynamic behaviour is not included in conventional process monitoring methods, leading to misleading results. Changes in dynamic relationships among process variables can not be investigated efficiently, resulting in significant process failure due to dynamic relationships, intermittent noises, and other disturbances. Substantial research has improved monitoring performance in dynamic industrial processes in recent years.
Haitoa et al. [92] proposed a multiscale framework for monitoring dynamic multivariate processes at different scales by combining wavelets and PCA. This framework enhances the suitability of PCA for monitoring processes based on auto-correlated data. Yoo et al. [93] have developed a multiscale approach to dynamic processes using dynamic PCA for WWTP. Similar faults have been detected and isolated by incorporating D statistics into the algorithm. Alabi et al. [94] have been developed a multiscale dynamic process monitoring approach by integrating WT with generic dissimilarity measure (GDM) to improve performance monitoring. Kini and Madakyaru [95] developed a multiscale DPCA framework where T 2 and SPE statistics were used for fault detection. The effectiveness of this framework is demonstrated by using dynamic multivariate data acquired from the TEP.

D. MULTISCALE METHODS FOR INCIPIENT FAULT DETECTION
Early detection of incipient faults in modern chemical process systems is increasingly becoming important, as these faults can slowly develop into severe abnormal events, which leads to failure of critical equipment. It is critical to detect even the most minor irregularities to ensure the safety of the process and the highest level of product quality. Detecting minor or incipient anomalies in modern chemical process systems is essential for process safety and maintaining product quality. Because they are camouflaged by noise and process control, these faults are difficult to detect early. They are common in complex processes and may quickly increase if no action is taken. Multiscale methods for detecting minor faults are reviewed as follows.
Kano et al. [18] proposed a multiscale method for incipient fault detection using dissimilarity analysis (DISSIM). Although DISSIM is mathematically comparable to PCA, its statistical index differs from T 2 . A new multiscale fault detection method based on Ensemble Empirical Mode Decomposition (EEMD) is proposed, effectively detecting three specific faults in the TEP that were previously undetectable using previously reported methods [96]. In this method, faults signatures are extracted using EEMD based PCA, and then half-normal probability and Cumulative Sum (CUSUM) are used for fault detection. The proposed method is further extended where CUSUM based on T 2 and SPE statistics improves fault detection [61]. Recently, a new multiscale framework has been proposed to detect incipient faults. In this framework, wavelet-based PCA is used to extract the fault signatures, and then CUSUM and EWMA based on T2 and SPE statistics are developed to improve the fault detection. The results show that EWMA based SPE statistics successfully detect the incipient faults present in the simulated data obtained from the CSTR system [97]. Žvokelj, et al. proposed a multivariate and multiscale fault detection methods to detect incipient failure of large slewing bearings based on Acoustic Emission (AE) signals by integrating EEMD with PCA [98], KPCA [99] and ICA [100].

E. MULTISCALE METHODS FOR FAULT DIAGNOSIS
Multiscale methods for fault detection have been thoroughly reviewed in previous sections. Although fault diagnosis is essential in process monitoring, it is relatively limited while employing multiscale methods. It is challenging to analyze the simultaneous impact of multiscale variables on monitoring statistics. Generally, fault diagnosis is accomplished via fault identification and classification. In fault identification, the faulty variables are identified based on their influence on the value of the statistical index. Identifying faulty variables is beneficial for highly integrated, large-scale, and complex plants [10]. There is no need for fault information for diagnosis through fault identification. If prior knowledge about the fault is available, the learning problem would be finding the boundary between normal and faulty samples. This learning problem is related to fault classification, and the three common approaches are similarity factors, discriminant analysis, and support vector machines (SVM).
Contribution plots are the most popular tool for identifying which variables push the statistics beyond control limits. Shao et al. [77] proposed a wavelet-based nonlinear PCA algorithm for process monitoring and applied differential contribution plots to find faulty variables of an industrial drying process. Many researchers have also used contribution plots with MSPCA process monitoring approaches to determine the faulty variables [101][102][103]. Zhiqiang and Quanxiong [104] were used contribution plots for fault identification in the wavelet-based adaptive MSPCA method. Many researchers were applied contribution plots to identify the faulty variables using kernel-based nonlinear multiscale techniques [76,85,87,105]. Similarity factor was integrated with MSPCA to identify the fault type and reveal the fault source [86,106].
Lau et al. [107] have implemented Adaptive Neuro-Fuzzy Inference System (ANFIS) fault classification with MSPCA to diagnosis selected fault cases in the TEP. Nor et al. [108] proposed a new multiscale fault diagnosis method by combining the multiscale KFDA and the ANFIS classification model. The fault classification performance was evaluated using the TEP, and the results indicated that the proposed multiscale KFDA-ANFIS framework improved over the multiscale PCA-ANFIS and FDA-ANFIS.
SVM is a well-known classification tool, proposed initially by Cortes and Vapnik [109]. Liu et al. [89] proposed a multiscale fault diagnosis method and applied the SVM classifier based on classification distance, using 4-fold to obtain the optimal parameters. Nor et al. [91] incorporated the SVM classifier with multiscale KFD, and the performance accuracy was compared to the multiscale KFD-GMM of the faults in the TEP.

F. MULTISCALE METHODS FOR BATCH PROCESS MONITORING
Batch processes often operate in different phases of operation. The batch operations are becoming increasingly complicated due to frequent start-ups and shutdowns. As a result, monitoring tasks in batch processes are more challenging to perform. Multiway PCA [110] and multiway PLS [111] are still used to monitor batch processes.
Lee et al. [112] proposed a multiway MSPCA approach for batch processes that combines WT and multiway PCA and has been effectively used in the sequencing batch reactor process for biological wastewater treatment. The proposed approach aids in detecting early faults and detecting less apparent faults. Alawi and Morris [113] proposed a multiscale multi-block modeling approach for batch process monitoring and compared it with the conventional MPCA approach using simulated data obtained from the penicillin fermentation simulation benchmark.

G. MULTISCALE METHODS FOR NON-GAUSSIAN DATA
Contrary to the eminent advances in MSPCA and MSPLS fault detection methods, ICA has received significantly less attention in the field of wavelet-based process monitoring despite ICA being a better choice for monitoring non gaussian data. A few researchers have also developed a multiscale process monitoring methods to handle non gaussian data.
Salahshoor and Kiasi [114] proposed a multiscale-ICA technique by integrating with wavelet analysis and ICA for non-gaussian data. They used Daubechies 3 up to level 3 and found that the proposed technique was effective for TE process data. Zvokelj et al. [100] proposed a new multiscale process monitoring technique by combining EEMD with ICA. They found that this technique is also suitable for detecting incipient faults in large slewing bearing systems. Recently, Kini and Madakyaru [115] proposed a wavelet based multiscale fault detection technique by combing wavelets with ICA. The effectiveness of the proposed technique was illustrated by using three different case studies and found that this technique can enhance the detection rate in noisy process environments. VOLUME XX, 2017 Name of Issue: (A*)-fault detection, (A)-multiscale methods for quality relevant process monitoring, (B)-multiscale methods for nonlinear process monitoring, (C)multiscale methods for dynamic process monitoring, (D)-multiscale methods for incipient fault detection, (E)-multiscale methods for fault diagnosis, (F)-multiscale methods for batch process monitoring, (G)-multiscale methods for non-Gaussian data.

VI. CHALLENGES AND OPPORTUNITIES
The increasing complexity of industrial systems and their related performance requirements have induced the need to develop new approaches for their supervision. This review unravels how multiscale approaches have been applied for process monitoring within various industrial applications. Despite the many advances in multiscale process monitoring research, more challenges are still emerging. Multiscale will likely have a role in addressing these challenges towards safer operations in the industry. A few of these challenges are discussed as follows.

A. ONLINE PROCESS MONITORING
Plant safety and product quality are two essential elements of today's process industry. Implementing a distributed control system and modern mearing techniques adds to the complexity of modern chemical plants. Therefore, it is important to identify and correct anomalies immediately during the process. This issue can be solved by employing online process monitoring, which will be helpful for efficient quality control of final products and process optimization. However, not enough attention is given to the issue of online process monitoring in multiscale methods. Therefore, developing a methodology for online process monitoring is of great interest that needs further research in the future.

B. FAULT IDENTIFICATION AND SMEARING EFFECT
The increasing complexity of chemical process systems makes it much more difficult to diagnose faults. A diagnostic tool is needed for fault identification after the fault has been detected in a process. Identifying a faulty variable is critical in analyzing the causes of abnormalities present in the process. In real systems, there is a possibility that avoiding a specific fault may result in the occurrence of another subsequent fault. Contribution plots are commonly used to determine fault variables, but this technique suffers from the smearing effect, which can mislead the faulty variables of the detected faults. However, insufficient attention is paid to fault identification in multiscale process monitoring. Identifying a faulty variable correctly in multiscale process monitoring is an open question that needs further research in the future.

C. ADAPTIVE FAULT DETECTION AND DIAGNOSIS
One of the most challenging monitoring processes is the detection of minor or incipient irregularities in highly correlated multivariate process data. Indeed, early detection of these incipient irregularities can help prevent significant damages and financial losses. Unfortunately, it is challenging to detect incipient abnormalities as they are too weak to detect conventional MSPM methods. As mentioned in the above discussion, a few techniques attempt to handle such irregularities. The key limitation of these studies is the use of conventional PCA methodology, a linear technique. However, most chemical processes are nonlinear and may have specific dynamic characteristics. Therefore, developing a nonlinear multiscale method for detecting incipient faults is of great interest in the future.

D. MULTIMODE PROCESS MONITORING
The conventional MSPM methods and their extensions assume that the process is operated under single steady state conditions. Since the modern industrial processes are linked with different operations, where operating conditions are change frequently. In this situation, the currently used monitoring technique may not perform well and may cause false alarm. Therefore, to keep the industrial process under control, monitoring process should be updated according to the change of operating conditions.

VII. CONCLUSIONS
This study aims to provide an overview of multiscale process monitoring and its use in chemical process systems. This article firstly discussed the statistical process monitoring and recent developments in MSPM methods. The need for multiscale process monitoring methods are performing feature extraction from industrial plant data within this context. A statistical analysis of the existing literature on multiscale process monitoring methods is also presented, based on the methods used, application area, types of data, and the more important issue addressed in these methods by various researchers. These issues include monitoring qualityrelated processes, monitoring nonlinear processes, handling batch process data, accounting for process dynamics, and performing fault diagnosis. Multiscale process monitoring research has significantly progressed by addressing these issues in the last two decades. Finally, future research prospects for multiscale process monitoring research have been discussed. This article can contribute to a better understanding of the role of multiscale process monitoring and provide new insights for researchers in the field.

ACKNOWLEDGMENT
The authors would like to offer their utmost appreciation to Universiti Teknologi PETRONAS (UTP) for providing technical and administrative support. The Yayasan UTP grant (cost centre: 015LC0-132) for an ongoing research project is also decidedly acknowledged preferred.

APPENDIX
The list of abbreviations is listed in Table 3