Pipeline Leak Detection Systems and Data Fusion: A Survey

The pipeline leakage problem is a very challenging and critical issue. Solving this problem will save the nation a lot of money, resources and more importantly, it will save the environment. This paper discusses the state of the art of leak detection systems (LDSs) and data fusion approaches that are applicable to pipeline monitoring. A comparison of LDSs is performed based on well-defined criteria. We have classified and critically reviewed these techniques. A thorough analysis and comparison of all the recent works have been provided.


Introduction
Pipeline links are vital for a nation's infrastructure and social and economic well-being.Damaged water pipes deteriorate the quality of the transported commodity, resulting in taste, odor, and aesthetic problems in the water supply as well as public health problems.Oil spills are known to destroy ecosystems and kill scores of aquatic organisms.Pipe damage results in other losses as well, such as increased operational and maintenance costs, loss of transported commodities (including oil, water, and gas), damage to property, disruption of service, disruption of industrial processes, increased environmental hazards, and imbalances in ecosystems.There is no doubt that efficient leak detection in pipelines can conserve a large amount of resources, save money, reduce carbon footprints, and achieve high levels of operational efficiency [28].
Pipeline deterioration is caused by static factors, such as soil type, pipe material, size, etc., and dynamic factors, such as changes in pressure zones and climate.Little is known about the breaking modes of buried pipes, and the physical mechanism is not completely understood.The broad aspects of pipeline leak detection encompass physical modeling of the pipe in the soil, understanding the nature of pipe failure, empirical and/or statistical modeling of historical failures, inspecting pipes to identify stress factors, rating the pipe conditions, and modeling the deterioration to forecast future failures and residual life.
The length and size of the pipeline, type of product carried, proximity of the pipeline to a high consequence area, swiftness of leak detection, location of nearest response personnel, leak history, and risk assessment results, etc., determine the efficiency of a leak detection system.The parameters for the evaluation of a leak detection system (LDS) are derived based on API1995b and the Alaska Department of Environmental Conservation 1999.Generally, for any good LDS, the most important four criteria are [24], [38]: Reliability, Sensitivity, Accuracy, and Robustness, and these criteria are what we use in this work.
The rest of the paper is organized as follows.In the next section, we discuss the leak characterization under different environments; Section 3 presents the leak detection techniques as mentioned in recent works.Section 4 discusses the WSN-based techniques that are used in monitoring pipelines.Finally, we discuss in detail the data fusion in pipeline monitoring.

Characteristics of Leak Detection Systems
Type of fluids: Pipelines transport a variety of fluids, such as gases, crude oil, petroleum products, steam, carbon dioxide, water, wastewater, etc.
Type of operation: Pipelines may operate in single-batch or multi-batch mode.In the single-batch mode of operation, pipelines operate continuously around the clock.I n multi-batch mode, the pipelines function is based on a time schedule.

Characteristics of leaks:
Leaks can occur suddenly or gradually depending on the causes and circumstances.
Sudden leaks occur due to external damage, resulting in a significant change in the temperature, flow, pressure, etc. Gradual leaks may occur due to corrosion.Sudden leaks may be successfully detected using an internally based LDS.In contrast, gradual leaks have very low magnitudes, and dedicated equipment, such as externally based LDSs, may be required to identify such leaks.
Operational phase: Pipeline conditions vary.The pumping condition involves the transport of fluid, whereas in the paused flow condition, the fluid flow is zero.Sometimes, valves will be used to block the fluid flow in a given segment.This special flow phase is known as the shut-in or blocked-line condition.

State-of-the-Art o f Leak Detection Technologies
The related work on leak detection systems as in [53] and [20] classified leak detection systems into visual, internal, and external based on the physical principles involved in t h e leak detection process.
Monitoring can be continuous or non-continuous.In the classification by [34], non-continuous inspection includes acoustic and non-acoustic methods, whereas continuous monitoring includes measurement and model-based methods.R e f e r e n c e [49] classified technologies based on the area of inspection, such as internal pipe surface, pipe wall integrity, and pipe bedding/void conditions.R e f e r e n c e s [35], [44], and [51] classified leak detection systems into non-technical and hardware-and software-based methods.Nontechnical methods do not involve any devices and use only natural senses, such as hearing and smelling, to identify a leak, whereas the technical methods use special devices to identify leaks; in the hardware methods, these devices include liquid sensing cables, vapor sampling, etc., and in the software methods, these devices include negative pressure waves, pressure point analysis, etc. Re f er ence [3] divided the leak detection systems into visual, physical, acoustic, ultra-spectrum, and electromagnetic.A similar classification by [22] divides LDS into visual, acoustic, and Electromagnetic-Radio Frequency (EM-RF) techniques.Figure 1 depicts the LDS classification.
LDS can be broadly classified into continuous or non-continuous monitoring systems.In non-continuous monitoring systems, the inspection is performed at regular intervals.Depending on the mode of inspection, pipeline operations can either continue or need to stop.For example, visual inspection or a helicopter survey does not require pipeline operations to be stopped, whereas an intelligent pigging system may require the operations to be stopped.The remote sensing of liquid hydrocarbons u s i n g aircraft mounted gas remote sensing is given in [48].This system detects evaporative plumes from pools of oil, gasoline, condensate, or pentane.Continuous monitoring systems monitor pipelines around the clock and are based on a physical principle.This approach can further be classified into external and internal systems.

Visual Technologies
Visual Manual Inspection: Visual inspection requires the manual patrolling of the pipeline for leaks.
Patrolling can be performed by any means (e.g., walking, in a vehicle, or from a helicopter).The operator examines the area for stains or other evidence of leaks.The leak detection capability depends on the ability of the inspection team, frequency of inspection, and the size of the leak.Smoke/gas testing: A smoke bomb is placed inside a water pipe with a blower to push the smoke.The smoke filters out through any cracks, thus exposing them.Water utilities used Formier10 gas (10% hydrogen and 90% nitrogen) for approximately 20 years [18].Hydrogen is a very lightweight gas and easily escapes through small cracks.The time taken for the gas to reach the surface depends on the depth of the pipe, soil conditions, and size of the leak.The gas detector is sensitive to small leaks.This method is not usable in large pipe mains due to the larger volume of gas required.The operator later looks for defects in the pipe from the recorded images.

Internal Systems
Internal systems use field sensors to monitor the operational and hydraulic conditions of the pipeline, e.g., measurements of the flow, pressure and temperature.The normal working parameters of the pipeline are determined either manually by pipeline controllers or based on sophisticated algorithms and hydraulic models.
A difference between the measured and predicted operational parameters indicates a leak.Typically, the remote field sensors provide data to a centralized monitoring station, where the data undergo filtering, signal processing and modules with leak detection algorithms to identify a leak.Internal systems generally do not require the installation of extensive hardware throughout the pipeline.Volume balance: Volume balance identifies the imbalance between incoming and outgoing volumes.Volume balance can detect catastrophic failures; however, its usage is rare due to its limited performance.

Rate of pressure/flow change:
The rate of pressure or flow change is based on the principle that a leak causes a rapid change in pressure.First, a sudden pressure drop can also be due to transient conditions.Filtering techniques need to be used to differentiate operational conditions from leak conditions.Second, pressure waves damp out as they traverse a longer length and thus additional pressure sensors need to be installed along the pipelines.This method is o n l y effective for large leaks, and transient conditions may trigger false alarms.
Negative Pressure Wave (NPW): Sudden leaks create a negative pressure wave or rarefaction wave, which propagates in both directions from the leak.NPW is easy to install and maintain and capable of continuously monitoring pipelines.However, the system cannot distinguish b e t w e e n leak scenarios and normal operations, thereby giving raise to false alarms.ATMOS, a novel technique that was developed recently and is based on the rarefaction method, shows tolerance to transient, shut-in, and slack flow conditions, thus triggering few false alarms [13].

Computational Pipeline Monitoring (CPM)
CPM detects hydraulic anomalies in pipeline operating parameters [5].t h e pipeline controller, etc.The method is retrofit table but less adaptable to complex pipeline configurations.

Real Time Transient Modeling (RTTM):
T h e parameters derived from a simulation model are compared with actual field data to look for discrepancies.Leaks occurring under all flow conditions can be modeled u s i n g this software, and small leaks can be detected in seconds.However, RTTM needs extensive training and skilled workers to operate and maintain.

External Systems
External systems use local sensors to detect fluids escaping from pipes.Impedance methods use cables with fiber optic or electro-chemical detection to sense liquids.Sniffing methods depend on vapor sensing through tubes.Acoustic methods depend on sensing noises induced by leaks.These systems are highly sensitive to leaks and c a n accurately locate them [24], [21].However, due to th e high costs, these methods are employed only in sensitive locations or for short pipeline segments [41] [38].
Liquid sensing cables: Similar to optical fiber methods, liquid sensing probes or cables are laid throughout the pipeline.Leaking fluids come into contact with cables and change their electrical properties, such as impedance, electrical resistance, dielectric constant, etc.A dedicated evaluation unit connected to the cable identifies the changes to the cable and detects a leak.Liquid cables can continuously monitor and accurately locate leaks.As with optical fibers, cable replacement may be required after a leak occurs.
Vapor sensing cables [21]: In the vapor sensing method, a highly permeable, pressure-tight air tube is fitted along the entire length of the pipe.When a leak occurs, the leaked material diffuses into the tube due to the concentration gradient.After a certain time, an accurate image of the substance surrounding the tube is obtained.A column of air that is pumped at constant speed passes through a gas sensor, and the substance produces a peak, indicating a leak.The increase in the gas concentration produces a leak peak, and the height of the peak is proportional to the concentration (Figure 3).Initially, a test gas is injected from one end to pass through the entire length of the pipe.The detector unit marks the start peak and the end peak to calculate the length of the pipe.When a leak occurs, the ratio of leaked distance to the overall distance is calculated to identify the location of the leak.Fiber optic sensing cables: In this method, a fiber op-tic cable is installed along the entire length of the pipeline.When a liquid comes into contact with the cable, the transmission characteristics of the fiber change.
While a pulsed laser propagates through the fiber, any changes to the density or composition of the fiber cause the light to scatter backwards.Spectral analysis reveals the temperature profile, leading to leak detection and localization [21], [33].The process is depicted in Figure 4. Cable replacement may be required after a leak.Recent advances in fiber optic sensors include quasi-distributed sensing, e.g., integrated Bragg gratings, and distributed sensing, e.g., interferometry and fiber sheath sensors ( [52], [25]).
When a pipeline leak occurs, the liquid moves from a high-pressure area to a low-pressure area and a turbulent flow is generated.This flow generates a characteristic sound that can be picked up by a specially designed hydrophone.Using sophisticated software, Leak ACO detects this signal, analyzes it, and evaluates the measurement results, thus identifying and providing the location of the leak.

Acoustic Emission (AE) testing:
When a pipeline leak occurs, turbulent liquid flow occurs in a high-pressure to a low-pressure area, creating a low frequency sound signal.Acoustic sensors affixed to the outside of the pipe, e.g., accelerometers, hydrophones, piezoelectric transducers, etc., pick up these signals.The deviation of the sound signals from a baseline fingerprint triggers a leak alarm.The received signal is stronger near the leak site, enabling localization of t h e leak.For pipes, such as PCCP, wire breaks release energy and cause a series of discrete events.AE monitoring of wire breaks is limited only to on-going wire events and cannot be used to detect already broken wires.

Acoustic Technologies
Listening rods: Acoustic signals from leaks propagate through the soil to the surface.Listening rods placed on the ground in the proximity of the leak pick up these noises and determine the location of the leak.
Inline acoustic leak detection: Inline acoustic leak detection sensors pass through pipes while in service and detect sounds due to leaks.The equipment can be tethered or free swimming.This method detects leaking joints and welds very well.
Leak noise correlator: Acoustic sensors placed on either side of a suspected leak transmit leak signals to a noise correlator.The correlator is typically a computer that analyzes the input sound spectrum and pinpoints leaks based on the time lag and sensor-to-sensor spacing.
Sonar or ultrasound: A major drawback of ultrasonic devices is that they cannot be operated above and below the water line simultaneously.To overcome this limitation, CCTV and sonar can be mounted on the same carrier vehicle so that CCTV can capture the information from above the waterline, while sonar captures the information below the waterline to account for the shortcomings of both systems [50].

Electromagnetic Systems
Magnetic Flux Leakage (MFL): MFL analyzes the flux leakage in a magnetic field when magnetized by strong, powerful magnets.A flawless pipe exhibits a homogeneous magnetic flux distribution, while a damaged pipe causes a flux leakage, as shown in Figure 5. DC inspection of pipes can be performed using Hall Effect devices and magneto resistive materials, while AC inspection can be performed using pick up coils.This testing mode is non-invasive and accurately detects cracks, corrosion, and the thinning of pipe walls.However, MFL is usable only on ferrous pipes and requires access to the surface of the pipe.T h e analysis of test results requires experienced personnel.Traditional MFL only d e t e c t s defects perpendicular to the magnetic field and cannot identify defects parallel to it.To overcome this, a new inspection method called Traverse Field Inspection (TFI) is employed in t h e Spiral MFL tool [27].

Remote Field Eddy Current (RFEC):
In BEM, a solenoid exciter probe generates pulsed eddy currents and magnetic flux lines within the pipe.Anomalies such as cracks or defects disrupt the current flow, which is captured by a receiving probe placed at a distance of 2.5 pipe diameters.The contour maps obtained after intensive post processing reveal the corrosion and thickness of the pipe wall [50].
Remote Field Transformer Coupling (RFTC): RFTC detects any broken wires in pre-stressed concrete cylinder pipes (PCCP) and holes or perforations in the steel used in PCCP [50].

Broadband Electromagnetic (BEM):
In BEM, a primary winding or exciter coil generates a short burst of pulsed waves in t he broadband frequency range.Eddy currents are induced in the adjacent ferrous conductive material shortly after the excitation pulses have been turned off; these eddy currents create a time varying magnetic field.The varying magnetic field induces a time varying voltage on the secondary winding or the receiver coil, which is correlated to the thickness of the pipe.BEM is similar to RFEC, but the signal transmitted covers a broad frequency spectrum [50].BEM is immune to electromagnetic interference and differs from other electromagnetic inspection methods because of its frequency independence.
Ground Penetrating Radar (GPR): GPR transducers radiate a short burst of varying radio frequencies into the ground and identify buried objects based on the s c a t t e r i n g o f t h e EM waves.The propagation of EM waves in soils is governed by parameters such as permittivity, magnetic permeability and conductivity.
The occurrences of leaks increase the moisture content of the soil nearby and cause dielectric variation.
Reflections occur at the interfaces between media with different electrical properties.The time lag between the transmitted and reflected waves determines the depth of the objects.An array of antennae attached to a survey vehicle driven along the transmission main detects the pipe anomalies.A three-dimensional (3D) GPR image is obtained using the raw field data after software processing.Example GPR data before and after interpretation are shown in Figure 6.Highly skilled expertise is needed to interpret the data.From the perspective of system design, GPR falls into three main categories ( [32]):

Pipeline Monitoring using WSN
A sensor node in Wireless Sensor Network (WSN) typically consists of transducers (to determine variations in temperature, pressure, strain, etc.), analog-to-digital converters, signal processing, power sources, memory, etc.Typical sensors used in pipeline monitoring are discussed in this section.
Acoustic sensors: Acoustic sensors are based on the principle that a liquid escaping through a hole in a pipeline produces a detectable sound.Acoustic sensors are easy to install and maintain and can continuously monitor a long pipeline.An important drawback is their high susceptibility to noise sources, such as system noises, environment noises, radio chatter, wind, Doppler effects, etc.To eliminate system noises, various techniques, such as band pass filtering [17], Fast Fourier Transform (FFT) and time-averaging Wigner-Ville distribution [54], can be used.Acoustic sensors can be used along with other sensors to overcome these limitations.I n [46] t h e y used piezoelectric sensors along with acoustic sensors to identify leaks and other pipe defects.
Piezoelectric sensors: Piezoelectric sensors (PZT) or lead zirconate titanate sensors can monitor t h e physical properties of pipelines, such as pressure, acceleration, vibration, acoustic waves, etc., and convert them into electric signals.The strength of the signal is determined by factors, such as the amount of energy released, distance from the monitored event, orientation of the sensor, transmission media, etc. [40], [39], [29].Piezoelectric sensors are a suitable candidate for pipeline monitoring because they exhibit high tolerance to harsh conditions and are not sensitive to electromagnetic radiation.However, piezoelectric sensors are not free from generating false alarms because a sensor deployed to detect one physical quantity may be affected by another, e.g., a pressure sensor may be affected by vibrations in the pipeline.To compensate for this limitation, secondary sensors can be used where the pressure sensors are used along with accelerometer piezoelectric sensors to detect pressure transients.
Chemical sensors: Chemical sensors determine a defect based on a change in chemical composition.Oxygen, carbon monoxide, and mercury vapor sensors are some examples of chemical sensors.The parameter to be detected and the resultant effect vary between different types of chemical sensors.For example, mercury sensors cause a change in resistance in the case of a gas leak [10].In another approach, the weight of the material changes considerably.Chemical sensors are very handy device in hazardous environments.
WSNs provide effective solutions for pipeline monitoring, due to its low cost, flexibility and ease of deployment in inaccessible terrain.However, some design issue need to be addressed before selecting wireless deployment.The major design issues that should be taken in consideration when using WSN for monitoring pipelines are: power source, communication standard, node antenna, communication protocol, localizations, network reliability, density of sensor nodes, packet loss control and network congestion control.

Data Fusion in Pipeline Monitoring
This section classify and describe the data fusion approaches in pipeline monitoring, and provide the relevant architecture models.

Classification and Architecture Models
The fusion of data from multiple sensors, called multi-sensor data fusion, provides more information than a single sensor.Multi-sensor data fusion can also include fusing overlapping measurements from the same sensor obtained at different times.Data fusion improves performance i n at least four ways: representation, accuracy, certainty, and completeness [1].Durrant-Whyte classified data fusion based on the relationship among the sources, such as complementary, competitive, and co-operative [15].
Complementary: Non-redundant data from different sensors can be fused to provide a complete view.

Redundant(competitive):
The same pieces of data from a single sensor or multiple sensors can be fused to increase the associated confidence.
Co-operative: Different data can be fused to provide a realistic view.
The abstraction levels of the input and output in the fusion process, including the measurement, signal, feature, and decision, can also form a basis for classification.Re f e r e nc e [31] applied these levels to classify fusion into signal fusion, pixel fusion, feature-level fusion, and symbol fusion.Boudjemaa and Forbes classified data fusion based on time, domains, attributes, and sensors [7].DaSarathy classified data fusion according to its input and output characteristics [12]: DaI-DaO (Data Input/Data Output), DaI-FeO (Data Input/Feature Output), FeI-FeO (Feature Input/Feature Output), FeI-DeO (Feature Input/Decision Output), DeI-DeO (Decision Input/Decision Output).
Data fusion architecture models can be data based, such as JDL [36] and DaSarathy [11], activity based, such as Boyd control loop, intelligence cycle, and t h e omnibus model [6], or role based, such as object-oriented and Frankel-Bedworth [19].Reference [26] provided a data centric taxonomy of data fusion methodologies and discussed the fusion of imperfect data, the fusion of correlated data, and the fusion of inconsistent data.Imperfect data can be fused using probabilistic, evidential, fuzzy reasoning, possibility theory, rough set theory, random set, and hybridization approaches.The fusion of correlated data can be achieved using correlation elimination and correlation presence.The fusion of inconsistent data focuses on removing the outliers, disorders, and conflicts.Reference [4] evaluated the performance of data fusion algorithms based on the Kalman filter, i.e., state vector fusion, measurement fusion, and gain fusion.

Existing Data Fusion (DF) Approaches
Some of the data fusion schemes applied in pipeline monitoring are discussed as follows: Homogeneous DF of seismic pulses: Reference [47] used Dempster-Shafer (D-S) method to fuse data from multiple seismic sensors in a proactive pipeline monitoring system.The data fusion increased the accuracy of the decisions by 8-25%.To detect the seismic pulses, geophones were deployed along the length of the pipeline at a depth of approximately half a meter, with varied sensor spacing.Different sources, such as people walking, driving a car, manual digging, etc., generated seismic signals with different frequencies, as shown in Figure 7a.The signals were then amplified, filtered, and A/D converted, followed by the extraction of the features.Figure 7b shows the decomposition of the original time series data into intrinsic oscillation mode functions (IMF) using empirical mode decomposition (EMD).Each IMF component was subjected to a Hilbert-Huang Transform (HHT) to obtain the amplitude and frequency.Normalized Kurtosis gives the feature vectors of different signals, as shown in Figure 7c, and extracts the features of each target.Data fusion was applied to the inputs from multiple seismic sensors with the D-S method using the rule given in Eq. 1.
Data driven framework using DF: Reference [55] proposed a data driven framework that used piezoelectric wafers to generate and sense ultrasonic waves.Multiple signal processing techniques were applied to extract as many as 365 features.The wave patterns were then checked u s i n g an adaptive boosting algorithm and five machine learning classifiers for damage detection.A voting strategy-based data fusion in a spatially distributed sensor network is given in [37].Certain features, such as the normalized mean squared error (MSE), correlation coefficient, curve length, loss of local coherence, etc., were used to detect the damage.For data fusion, the independent decisions were fused to arrive at the outcome, i.e., decision-level data fusion was employed.The features for all monitored signals were compared against the threshold, as shown in Figure 9.A lower false alarm rate means t h a t the result is highly accurate.For each transducer pair, a voting strategy was used to increase the credibility.
The system seemed to increase the detection probability to more than 90% and reduced the false alarms to under 5%.
Figure 9: Sensor and Feature fusion at the decision level [37].
Fusion of GPR and EMI for buried pipes: A multisensory system was used to fuse data from seismic, GPR, and EMI sensors to reduce false alarms in landmine detection [45].The responses of the three types of sensors to the soil condition differed from each other.EMI sensors were sensitive to soil conductivity, while the seismic sensor was sensitive to the difference between the mechanical properties of the soil and t h e landmine; GPR was sensitive to dielectric properties.A good use of the complementary features from sensors reduced the false alarm rate significantly.Reference [2] proposed a multisensory data fusion architecture to assess the locations and structural conditions of the buried pipes.
Using a combination of ground penetrating radar and electromagnetic waves, inspection can be performed without draining the pipelines.GPR can detect the presence and depth information of buried pipes but cannot distinguish metallic and plastic pipes.Electromagnetic sensors can detect the condition of metallic pipes but cannot measure the depth.A data fusion algorithm that is based on artificial neural networks and uses a combination of inputs from GPR and EMI can detect and classify various defects, such as major cracks and leaks in pipelines.However, this architecture is only conceptual, and the implementation results are n o t yet known.
Heterogeneous DF of NDE methods using geometric transformation: I n [47] t h e y employed a neural networkb a s e d g eometric transformation algorithm to fuse data from images obtained from three NDE methods: IR thermal imaging, magnetic flux leakage (MFL), and ultrasonic testing (UT).Given a training data set, the radial basis function identified redundant and complementary features using artificial neural networks (ANN).
Redundancy increased the reliability of defect characterization by identifying the common information in different NDE methods.Complementarity improved the accuracy of defect characterization by identifying the defect characteristics unique to each inspection method.
Let x 1 (r, c 1 ) and x 2 (r, c 2 ) denote the two different NDE images, where r represents the redundancy feature and c1 and c2 represent the complementary features.Then, the redundancy (hr) and complementary information (gr) are defined in Eq 2 and 3: The redundant relation between the data are given by Eq 4 In Eq 4, ⋄ represents a homomorphic operator and g 1 (x 1 ) is a radial basis function that takes the training data set as an input and outputs the best function approximation for x 1 .g 2 (x 2 ) is a conditioning function and application dependent.For example, if the data x 2 is spread over a wide range, a logarithmic function can be used for g 2 ().From Eq 4, if the homomorphic operator is chosen as an addition operator + and g 2 is assumed as the identity function, then h 1 (r) is given by Eq 5.
Similarly, the complementary relationship can also be defined as in Eq 6 and h 2 (c 1 , c 2 ) is obtained, but the neural network is trained with the complementary data.
The training of the artificial neural network with diverse and sufficient NDE signatures is essential for improved fusion.The data fusion for different NDE combinations is shown in Figure 10.

Conclusion:
Each leak detection system is a unique system and designed based on the pipeline f o r w h i c h it is developed for.The choice of LDS should be based on a fit-for-purpose approach.T h e operating parameters, such as the pipeline size, length, instrumentation design, etc., dictate the applicability of an approach.The time taken to detect a leak, number of false alarms, accuracy of the installed instrumentation, and many other factors influence the performance of a leak detection system.The capabilities of each LDS and the degree to which they mitigate the risks discussed in this paper can be used as a guideline when choosing a leak detection approach.The rule of thumb is that field test results from similar applications always provide the best recommendations.T h e use of information from multiple LDSs increases the detection accuracy.
There are multiple uncertainties in data sources, including hydraulic noise, errors in analog-to-digital conversion, the non-repeatability of field sensors, data communication errors, timing, drift, transient conditions, etc.The uncertainty in the data is a crucial issue because, without proper inputs, a correct output cannot be achieved no matter how efficient the filtering, signal processing or data fusion algorithms may be.The quality of a data fusion framework depends on good input data and the performance of the fusion system.A thorough review of the available historical data regarding pipe performance and failure can give greater insight into applying data fusion and accurately predicting pipe deterioration.
Future research directions for developing and improving leak detection systems should focus on the following critical issues: enhancing the accuracy and fidelity of collected information by developing more accurate mechanisms such as optical means, for example.Second, developing advanced data fusion techniques for minimizing false alarms and precisely determining the leak position; deep learning is a very potential approach in this direction.Finally, the development of robust and secure wireless sensor networks that are able to continuously collect, process and make a decision is very critical; and it is important exploit the advancement in Internet of Things, cheap drones for providing a comprehensive and integrated solution.
Figure 2 illustrates different types of internal LDS techniques.

Figure 2 :
Figure 2: Internal LDSs line pack correction: The changes to a line pack are observed by various sensors, e.g., pressure, temperature and densitometers, at multiple locations between the inlet and outlet flow meters.The pipeline is divided into multiple segments based on certain factors, such as elevation profile, location of instruments, desired level of accuracy, etc.The changes measured by various sensors are adjusted in the mass balance to account for transient flows, anticipated fluid changes, and other flow conditions.The capability depends on the selection of t h e alarm set points, repeatability of t h e instrumentation, skill of
a) Seismic signals from three sources (b) EMD for a signal (c) Feature vector for three signals
The system was shown to exhibit an average ac c ura c y of 84.2-89%.As shown in Figure8, the received signals were preprocessed to remove low/high frequency vibrations via a band pass filter of t h e range f r o m 190 to 450 kHz.By normalizing the signals, the ambient effects were contained to a certain degree.Various signal processing methods, such as the wavelet transform, Hilbert transform, Mellin transform, etc., were then applied to extract features.
Balance with Line Pack compensationExisting leaks and leaks for shut in and transient conditions can be detected Able to detect 1% leak in minutes Method adaptable to any pipeline configuration