[go: up one dir, main page]

WO2022187898A1 - Procédé et système de détection d'anomalie de pipeline - Google Patents

Procédé et système de détection d'anomalie de pipeline Download PDF

Info

Publication number
WO2022187898A1
WO2022187898A1 PCT/AU2022/050194 AU2022050194W WO2022187898A1 WO 2022187898 A1 WO2022187898 A1 WO 2022187898A1 AU 2022050194 W AU2022050194 W AU 2022050194W WO 2022187898 A1 WO2022187898 A1 WO 2022187898A1
Authority
WO
WIPO (PCT)
Prior art keywords
pressure wave
anomaly
noise
pipeline
transient
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/AU2022/050194
Other languages
English (en)
Inventor
Jessica Maria BOHORQUEZ AREVALO
Martin Francis LAMBERT
Bradley James ALEXANDER
Angus Ross SIMPSON
Derek Abbott
Sylvan ELHAY
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Adelaide University
Original Assignee
University of Adelaide
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Priority claimed from AU2021900659A external-priority patent/AU2021900659A0/en
Application filed by University of Adelaide filed Critical University of Adelaide
Priority to AU2022234947A priority Critical patent/AU2022234947A1/en
Publication of WO2022187898A1 publication Critical patent/WO2022187898A1/fr
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N29/00—Investigating or analysing materials by the use of ultrasonic, sonic or infrasonic waves; Visualisation of the interior of objects by transmitting ultrasonic or sonic waves through the object
    • G01N29/02—Analysing fluids
    • G01N29/032—Analysing fluids by measuring attenuation of acoustic waves
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
    • G01M3/00—Investigating fluid-tightness of structures
    • G01M3/02—Investigating fluid-tightness of structures by using fluid or vacuum
    • G01M3/26—Investigating fluid-tightness of structures by using fluid or vacuum by measuring rate of loss or gain of fluid, e.g. by pressure-responsive devices, by flow detectors
    • G01M3/28—Investigating fluid-tightness of structures by using fluid or vacuum by measuring rate of loss or gain of fluid, e.g. by pressure-responsive devices, by flow detectors for pipes, cables or tubes; for pipe joints or seals; for valves ; for welds
    • G01M3/2807—Investigating fluid-tightness of structures by using fluid or vacuum by measuring rate of loss or gain of fluid, e.g. by pressure-responsive devices, by flow detectors for pipes, cables or tubes; for pipe joints or seals; for valves ; for welds for pipes
    • G01M3/2815—Investigating fluid-tightness of structures by using fluid or vacuum by measuring rate of loss or gain of fluid, e.g. by pressure-responsive devices, by flow detectors for pipes, cables or tubes; for pipe joints or seals; for valves ; for welds for pipes using pressure measurements
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
    • G01M3/00—Investigating fluid-tightness of structures
    • G01M3/02—Investigating fluid-tightness of structures by using fluid or vacuum
    • G01M3/04—Investigating fluid-tightness of structures by using fluid or vacuum by detecting the presence of fluid at the leakage point
    • G01M3/24—Investigating fluid-tightness of structures by using fluid or vacuum by detecting the presence of fluid at the leakage point using infrasonic, sonic, or ultrasonic vibrations
    • G01M3/243—Investigating fluid-tightness of structures by using fluid or vacuum by detecting the presence of fluid at the leakage point using infrasonic, sonic, or ultrasonic vibrations for pipes
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N29/00—Investigating or analysing materials by the use of ultrasonic, sonic or infrasonic waves; Visualisation of the interior of objects by transmitting ultrasonic or sonic waves through the object
    • G01N29/44—Processing the detected response signal, e.g. electronic circuits specially adapted therefor
    • G01N29/4454—Signal recognition, e.g. specific values or portions, signal events, signatures
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N29/00—Investigating or analysing materials by the use of ultrasonic, sonic or infrasonic waves; Visualisation of the interior of objects by transmitting ultrasonic or sonic waves through the object
    • G01N29/44—Processing the detected response signal, e.g. electronic circuits specially adapted therefor
    • G01N29/4463—Signal correction, e.g. distance amplitude correction [DAC], distance gain size [DGS], noise filtering
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N29/00—Investigating or analysing materials by the use of ultrasonic, sonic or infrasonic waves; Visualisation of the interior of objects by transmitting ultrasonic or sonic waves through the object
    • G01N29/44—Processing the detected response signal, e.g. electronic circuits specially adapted therefor
    • G01N29/4481—Neural networks
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01V—GEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
    • G01V1/00—Seismology; Seismic or acoustic prospecting or detecting
    • G01V1/003—Seismic data acquisition in general, e.g. survey design
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/09—Supervised learning
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N2291/00—Indexing codes associated with group G01N29/00
    • G01N2291/01—Indexing codes associated with the measuring variable
    • G01N2291/015—Attenuation, scattering
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N2291/00—Indexing codes associated with group G01N29/00
    • G01N2291/26—Scanned objects
    • G01N2291/263—Surfaces
    • G01N2291/2636—Surfaces cylindrical from inside
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
    • G01R29/00—Arrangements for measuring or indicating electric quantities not covered by groups G01R19/00 - G01R27/00

Definitions

  • the present disclosure relates to the detection of anomalies in a pipeline carrying a fluid.
  • the present disclosure relates to the use of artificial neural networks to analyse transient pressure information to detect anomalies in a pipeline.
  • the Applicant has previously developed methods and systems for determining the condition of a pipeline based on measuring transient pressure information from the pipeline and then processing this information to determine the presence of an anomaly or fault based on specifically configured artificial neural networks (ANN).
  • ANN artificial neural networks
  • the detection system involves the generation of a transient pressure wave in the pipeline and then the real time processing of the resulting transient pressure wave interaction signal measured in the pipeline by an ANN trained to detect the presence of the anomaly.
  • the anomaly detector ANN is both trained and operates on downsampled pressure information which allows the detection process to occur in real time.
  • the detection system functions to monitor continuously pipeline pressure information and applies a first classifier ANN to this information to determine the overall condition of the pipeline and if an anomaly is detected, a second anomaly detector ANN functions to determine the type of anomaly and its associated anomaly characteristics.
  • the analysed system might be exposed to background noise, as a result preventing existing techniques from performing as well as required.
  • the focus may be further improving the performance of an ANN for a specific anomaly detection task.
  • the present disclosure provides a method for detecting an anomaly in a pipeline, comprising: generating a transient pressure wave in a fluid carried along the pipeline; measuring a transient pressure wave interaction signal corresponding to a response from the anomaly to the generated transient pressure wave; processing the transient pressure wave interaction signal to detect the anomaly by a series of artificial neural networks (ANNs) each trained to detect the anomaly on respective datasets of noise modulated training data; and determining a performance measure characterising the detection of the anomaly based on anomaly detection results from each of the ANNs.
  • ANNs artificial neural networks
  • a dataset of noise modulated training data is generated in accordance with a characteristic pressure noise signal that is based on the pipeline and an anomaly type of the anomaly that the ANN is trained to detect.
  • generating a dataset of noise modulated training data in accordance with the characteristic pressure noise signal comprises: generating for the anomaly type a plurality of pressure wave interaction signals covering a range of anomaly characteristics associated with the anomaly type; determining a representative pressure fluctuation value based on the pipeline and the anomaly type; determining a characteristic pressure noise signal based on the representative pressure fluctuation value; and modulating the plurality of pressure wave interaction signals by the characteristic pressure noise signal to generate the dataset of noise modulated training data.
  • determining a representative pressure fluctuation value comprises comparing a first transient pressure wave interaction signal where there is no anomaly in the pipeline with a second transient pressure wave interaction signal where the pipeline has the anomaly.
  • determining the characteristic pressure noise signal based on the representative pressure fluctuation value comprises forming a Gaussian distribution function having a zero mean and a standard deviation defined as a multiple of the representative pressure fluctuation value.
  • additional datasets of noise modulated training data that comprise the respective datasets of noise modulated training data are generated by additionally modulating the plurality of pressure wave interaction signals by the characteristic pressure noise signal to generate each additional dataset of noise modulated training data.
  • each of the respective datasets of noise modulated training data is divided into a training dataset for training a respective ANN of the series of ANNs and a testing dataset for testing and/or validating the respective ANN.
  • the method further comprises pre-processing the transient pressure wave interaction signal following measurement of the pressure wave interaction signal to remove signal artefacts arising from generating the transient pressure wave.
  • the signal artefact arising from generating of the transient pressure wave is an increased pressure variability in a portion of the pressure wave interaction signal corresponding to prior to generation of the transient pressure wave and pre-processing the transient pressure wave interaction signal comprises reducing the pressure variability in that portion of the pressure wave interaction signal.
  • the method further comprises pre-processing the transient pressure wave interaction signal following measurement by offsetting the pressure wave interaction signal to more closely conform with the pressure wave interaction signals used to train the ANNs.
  • the performance measure assesses a variability of the anomaly detection results.
  • the method further comprises verifying the anomaly detection results.
  • the present disclosure provides a system for detecting an anomaly in a pipeline, the system including: a transient pressure wave generator for generating a transient pressure wave in fluid carried along the pipeline; a pressure detector for measuring a transient pressure wave interaction signal corresponding to a response from the anomaly to the generated transient pressure wave; an analysis module comprising one or more data processors configured for: processing the transient pressure wave interaction signal to detect the anomaly by a series of artificial neural networks (ANNs) each trained to detect the anomaly on respective sets of noise modulated training data; and determining a performance measure characterising the detection of the anomaly based on anomaly detection results from each of the ANNs.
  • ANNs artificial neural networks
  • a dataset of noise modulated training data is generated in accordance with a characteristic pressure noise signal that is based on the pipeline and an anomaly type of the anomaly that the ANN is trained to detect.
  • generating a dataset of noise modulated training data in accordance with the characteristic pressure noise signal comprises: generating for the anomaly type a plurality of pressure wave interaction signals covering a range of anomaly characteristics associated with the anomaly type; determining a representative pressure fluctuation value based on the pipeline and the anomaly type; determining a characteristic pressure noise signal based on the representative pressure fluctuation value; modulating the plurality of pressure wave interaction signals by the characteristic pressure noise signal to generate the dataset of noise modulated training data.
  • determining a representative pressure fluctuation value comprises comparing a first transient pressure wave interaction signal where there is no anomaly in the pipeline with a second transient pressure wave interaction signal where the pipeline has the anomaly.
  • determining the characteristic pressure noise signal based on the representative pressure fluctuation value comprises forming a Gaussian distribution function having a zero mean and a standard deviation defined as a multiple of the representative pressure fluctuation value.
  • additional datasets of noise modulated training data that comprise the respective datasets of noise modulated training data are generated by additionally modulating the plurality of pressure wave interaction signals by the characteristic pressure noise signal to generate each additional dataset of noise modulated training data.
  • each of the respective datasets of noise modulated training data is divided into a training dataset for training a respective ANN of the series of ANNs and a testing dataset for testing and/or validating the respective ANN.
  • system further comprises pre-processing the transient pressure wave interaction signal following measurement of the pressure wave interaction signal to remove signal artefacts arising from generating the transient pressure wave.
  • the signal artefact arising from generating of the transient pressure wave is an increased pressure variability in a portion of the pressure wave interaction signal corresponding to prior to generation of the transient pressure wave and pre-processing the transient pressure wave interaction signal comprises reducing the pressure variability in that portion of the pressure wave interaction signal.
  • system further comprises pre-processing the transient pressure wave interaction signal following measurement by offsetting the pressure wave interaction signal to more closely conform with the pressure wave interaction signals used to train the ANNs.
  • the performance measure assesses a variability of anomaly detection results.
  • the analysis module is further configured for verifying the anomaly detection results.
  • the present disclosure provides a system for detecting an anomaly in a pipeline comprising means configured to carry out a method in accordance with the first aspect.
  • Figure 1 is a flowchart of a method for detecting an anomaly in a pipeline in accordance with an illustrative embodiment
  • Figure 2 is a system overview diagram of an anomaly detection system for detecting an anomaly in a pipeline in accordance with an illustrative embodiment
  • Figure 3 is a pipeline model for training the ANN to determine an anomaly in a pipeline in accordance with an illustrative embodiment
  • Figure 4 is a flowchart of a method for generating noise modulated training data in accordance with an illustrative embodiment
  • Figure 5 is a plot of two transient pressure wave interaction signals comparing the pipeline response for an intact pipeline with a pipeline having a small leak to determine a representative noise in accordance with an illustrative embodiment
  • Figures 6a and 6b are plots of the average leak location organised by leak size for an ANN trained with original training data (ie, Figure 6a) and an ANN trained with noise modulated data (ie, Figure 6b) in accordance with an illustrative embodiment;
  • Figures 7a-d are a series of plots depicting the original transient pressure wave interaction signal overlayed with a corresponding noise modulated transient pressure wave interaction signal in accordance with an illustrative embodiment
  • Figure 8 is a data flow diagram showing figuratively the training of a series of ANNs on respective sets of noise modulated training data in accordance with an illustrative embodiment
  • Figure 9 is a flowchart of a method for detecting an anomaly in a pipeline in accordance with another illustrative embodiment
  • Figure 10 is a flowchart showing example pre-processing steps for the transient pressure wave interaction signal in accordance with an illustrative embodiment
  • Figure 11 is a plot showing the results of a sensitivity analysis conducted on an ANN trained to detect an anomaly in accordance with an illustrative embodiment
  • Figure 12 is a plot of a measured transient pressure wave interaction signal following the generation of a transient pressure wave showing in detail the pressure fluctuations before and after generation of the transient pressure wave in accordance with an illustrative embodiment
  • Figure 13 is a plot of the distributions of pressure head for the (a) and (b) regions shown in Figure
  • Figure 14 is a plot of the cumulative distribution function (CDF) for the (a) region shown in Figure 12 and a modified CDF having the same mean but with the standard deviation of the (b) region;
  • CDF cumulative distribution function
  • Figure 15 is a figurative view of a pipeline arrangement for experimental confirmation in accordance with an illustrative embodiment
  • Figure 16 is a plot of the measured transient pressure wave interaction signals obtained from the pipeline arrangement illustrated in Figure 15 in accordance with an illustrative embodiment
  • Figure 17 is a figurative view of the different series of ANNs trained on respective sets of noise modulated training data in accordance with an illustrative embodiment
  • Figures 18a-g is a series of plots illustrating the percentage of exceedance associated with the absolute average error for the location of a leak following training and testing for each noise intensity according to an illustrative embodiment
  • Figure 19 is a plot showing the measured transient pressure wave interaction signals in Figure 16 following pre-processing in accordance with an illustrative embodiment
  • Figures 20 to 26 are plots of the distribution of predictions for determining the location of a leak for each of the transient pressure wave signals illustrated in Figure 19 following the processing by a series of ANNs trained on noise modulated training data having different noise intensities in accordance with an illustrative embodiment
  • Figure 27 is a plot of the performance of the trained ANNs on the measured and processed transient pressure wave signals illustrated in Figure 19 for each noise intensity in accordance with an illustrative embodiment
  • Figure 28 is a plot of the average values (in circles) and distributions of the Root Mean Square Error (RMSE) for the training (in lighter shade) and the testing (in black) of the trained anomaly detector ANNs for each noise intensity in accordance with an illustrative embodiment
  • RMSE Root Mean Square Error
  • Figure 29 is a plot of the error in RMSE (in circles) and the range for the prediction of a leak location (whiskers) for the training (in lighter shade) and the testing (in black) of the trained anomaly detector ANNs for each noise intensity in accordance with an illustrative embodiment
  • Figure 30 is a plot comparing the predicted transient pressure wave interaction signal obtained by numerically generating the expected trace based on the determined anomaly characteristics with the measured transient pressure wave interaction signal used to determine the anomaly characteristics in accordance with an illustrative embodiment.
  • FIG. 1 there is shown a flowchart of a method 100 for detecting an anomaly in a pipeline.
  • the present method and system is applicable to any fluid carrying pipeline system or network including, but not limited to, water transmission pipes, pipelines in chemical plants, wastewater pumping pipelines or oil and gas pipelines.
  • anomaly is taken to mean any feature or component of the pipeline that affects the hydraulic performance of the pipeline.
  • An anomaly may be classified under different anomaly types, including, but not limited to, the following anomaly types:
  • An anomaly may also involve associated characteristics that characterise the anomaly.
  • the anomaly may be of the type “leak” and the associate anomaly characteristics may include the location of the “leak”, the size of the “leak” and the flow rate of fluid exiting the pipeline as a result of the “leak”.
  • a transient pressure wave is generated in the fluid carried along the pipeline by in this example a pressure wave generator.
  • an anomaly detection system 200 for detecting an anomaly in a pipeline according to an illustrative embodiment operable to implement method 100.
  • anomaly detection system 200 includes a transient pressure wave generator 205 for generating a transient pressure wave in the fluid of the pipeline and a pressure sensor or detector 210 in the form of a pressure or acoustic transducer for detecting the transient pressure wave interaction signal.
  • System 200 further includes an analysis and control module 220 for processing of the transient pressure wave interaction signal to analyse a region of interest of the pipeline.
  • anomaly detection system 200 may also control the operation of pressure wave generator 205.
  • This transient pressure wave may be generated in the fluid by any one of a number of techniques.
  • a transient pressure wave may be generated at a device attached to, for example, an existing scour or fire plug air valve or offtake valve and then abruptly stopping the flow of water. This has the effect of progressively stopping or altering the flow of water along the pipe that had been previously established. This progressive stopping or alteration of the flow of water along the pipeline is equivalent to the generation of a transient pressure wave resulting in the propagation of a transient wavefront along the pipeline.
  • transient pressure wave examples include, but are not limited to, inline valve closure devices, side discharge valves and piston chambers where an amount of fluid is drawn into a chamber containing a piston which is then operated.
  • inline valve closure devices side discharge valves
  • piston chambers where an amount of fluid is drawn into a chamber containing a piston which is then operated.
  • One example system for generating a transient pressure wave in fluid carried by the pipeline is described in PCT Application No PCT/AU2016/000246 (W02017008100) titled “SYSTEM AND METHOD FOR GENERATION OF A PRESSURE SIGNAL”, filed by the Applicant here, and whose entire contents are incorporated by reference in their entirety in the present disclosure.
  • a popular method for generating the transient pressure wave consists of generating a single step pulse created by the fast closure of a valve within the pipeline system or attached to the system.
  • the typical useful bandwidth of this method may be less than 100 Hz, which means that, for some applications, a single pulse may not allow the extraction of enough information from the transient pressure wave interaction signal recorded for the pipeline system.
  • Another transient pressure wave generation method consists of a pulse generation or sine wave stepping technique.
  • the sine wave stepping technique uses a single frequency sinusoidal oscillatory signal as the input, and this frequency is adjusted to cover the range of frequencies required.
  • generating a transient pressure wave may include the generation of persistent signals known as pseudo-random binary sequences (PRBS).
  • PRBS pseudo-random binary sequences
  • These signals consist of randomly spaced and equal magnitude pulses that are set to repeat periodically, and have a spectrum similar to that of a single input pulse.
  • This generation method can use Maximum-Length Binary Sequences (MLBS) or Inverse Repeated Sequences (IRS).
  • the hydraulic noise of the system may be used to generate the transient pressure waves in the pipeline for analysis of the pipeline in accordance with the present disclosure.
  • customized and small amplitude pressure signals may be obtained from a piezoelectric actuator driven by a linear power amplifier to generate the transient pressure wave.
  • controlled electrical sparks are employed to generate a vapour cavity that then collapses. An electrical spark surrounded by water causes the development of a localized vapour cavity, the collapse of which induces a transient pressure wave into the surrounding body of fluid having the characteristics of an extremely sharp pressure pulse. This typically results in high frequency pressure waves that can improve the incident signal bandwidth.
  • a transient pressure wave interaction signal that corresponds to a response from the anomaly to the generated transient pressure wave is measured by in one example pressure sensor or detector 210.
  • the time duration of this transient pressure wave interaction signal that is detected may be selected to cover the first complete cycle of reflections of the transient pressure wave (4 L/a) seconds where L is the length of region of interest of the pipeline and a is the transient wave speed in the fluid.
  • the time duration is selected to be between 2 L/a and 4L/a seconds, where again L is the length of region of interest of the pipeline and a is the transient wave speed in the fluid.
  • the time window covers at least 2.5 L/a.
  • the transient pressure wave interaction signal covers at least 3 L/a.
  • the signal covers at least 3.5 L/a.
  • L may be selected to be the entire length of the pipeline and in one example, the boundaries of the analysed transient pressure wave interaction signal would include at least L/a seconds before the generation of the transient pressure wave and at least 2.0 L/a seconds after the transient pressure wave to cover the complete length of the pipeline.
  • a pressure detector 210 is employed in the form of a pressure or acoustic transducer in combination with a data acquisition capability.
  • any type of high frequency response pressure detector, optical fibre sensor or transducer configured to record the transient pressure wave interaction signal of the pipeline following initiation of a transient pressure wave for a time duration as described above at a selected detection sampling rate or frequency typically between 2,000 Hz and 10,000 Hz may be used.
  • the selection of a detection sampling frequency depends on the pipe wall properties of the pipeline, the wave speed of the fluid and the expected speed of occurrence of the anomaly.
  • the detection sampling frequency for detecting the transient pressure wave interaction signal may be selected from the following frequency ranges, including, but not limited to, greater than 2 kHz, 2 kHz - 5 kHz, 5 kHz - 10 kHz, 2 kHz - 3 kHz, 3 kHz - 4 kHz, 4 kHz - 5 kHz, 5 kHz - 6 kHz, 6 kHz - 7 kHz, 7 kHz - 8 kHz, 8 kHz - 9 kHz, 9 kHz - 10 kHz, or greater than 10 kHz.
  • analysis module 220 includes a customised data logging and analysis arrangement comprising a timing module 222 or other clock arrangement which may be GPS based, a data acquisition module 224, data processing module 226 and a remote communications module 228 to convey analysis results to a central location as required.
  • a timing module 222 or other clock arrangement which may be GPS based
  • data acquisition module 224 data acquisition module
  • data processing module 226 data processing module 226
  • remote communications module 228 to convey analysis results to a central location as required.
  • the functionality of the various modules may be implemented primarily in hardware or in a combination of hardware and software or primarily in software.
  • the pressure wave generator 205 is deployed remotely from the pressure detector 210 and analysis module 220.
  • the pressure detector 210 and analysis module 220 may together form a “measurement station”.
  • the pressure wave generator 205 may be co-located together with the pressure detector 210 and/or analysis module 220.
  • other implementations may include multiple measurement stations which will detect the transient pressure wave interaction signal at different locations along the pipeline.
  • the transient pressure wave interaction signal is processed by a series of artificial neural networks (ANNs) that have each been trained to detect the anomaly of interest on respective sets of noise modulated training data.
  • ANNs artificial neural networks
  • ID Convolution Network and Dense Network architectures were described.
  • a ID Convolutional Network architecture with three layers for the ANN was adopted as appropriate for characterising the condition of a pipeline.
  • a ID Convolutional ANN architecture is also adopted but modified to include: a) four convolutional layers, b) use of Leak Rectified Linear Unit as activation function, c) 20 filters that increase to 30 filters in the last convolutional layer; and d) three dense layers of size 14, 6 and 2.
  • the four convolutional layers contain weights connecting each neuron in a layer to neurons in the corresponding neighbourhood in the subsequent layer while in the dense layers each neuron in a layer is connected to every neuron in the subsequent layer.
  • the final shape of the initial convolutional layer depends on the downsampling frequency selected for the generation of the transient wave pressure signal for the ANN training.
  • the rest of the architecture of the ID Convolutional ANN depends on this initial shape following this structure: a) convolutional layer #1 with L neurons corresponding to L values of the downsampled transient pressure wave signals and 20 filters, b) convolutional layer #2 with the same shape size of L neurons corresponding to L values of the downsampled transient pressure wave signals and 20 filters, c) max pooling layer where the resulting shape is U2 neurons corresponding to L values of the downsampled transient pressure wave signals and 20 filters, d) convolutional layer #3 with a shape size of U2 neurons corresponding to L values of the downsampled transient pressure wave signals and 20 filters, e) max pooling layer where the resulting shape is U4 neurons corresponding to L values of the downsampled transient pressure wave signals and 20 filters, f) convolutional layer #4 with a shape size of U4 neurons corresponding to L values of the downsampled transient pressure wave signals and 30 filters, g) max pooling layer where the resulting shape is L/8 neurons corresponding to
  • a new ANN architecture was designed with the analysis of five different alternatives of ID-convolutional neural networks using different number of layers and filters in each layer. These five ANN architectures were trained using two input datasets: one using numerically generated transient pressure wave signals and a second one using these numerical transient pressure wave signals including the addition of a Gaussian distribution of noise. The analysis of the results of these ANN training procedures demonstrated that out of these five ANN architectures, a ID-convolutional ANN with the architecture described above showed more potential of achieving consistent and accurate results for a leak location in a real pipeline and was thus selected as the final ANN architecture.
  • this ANN architecture evaluation exercise may be carried out for other anomaly types to determine the ANN architecture suitable for the particular anomaly detection requirement. Accordingly, the resulting ANN architecture might have a different number of convolutional, max pooling and/or dense layers, a different distribution of filters across the convolutional layers or a different length of filters. Variations in these parameters will affect the total number of weights as described in Equation 1 below and will modify the predictions of the anomaly location and size.
  • Equation 1 the resulting total number of weights, W, for the ID Convolutional ANN is given by: Equation 1
  • the first term represents the weights in the convolutional layers (n) and the second term the weights in the dense layers (/).
  • w and h are the width and height of the filters
  • f n is the number of filters in the convolutional layer n
  • C j are the number of neurons in the dense layer j.
  • c ⁇ _ 1 will depend on the dimensions of the ANN input set which will correspond to the number of individual data points in the transient pressure wave interaction signal after appropriate sampling.
  • a set of training data corresponding to the anomaly being detected is generated.
  • the anomaly is a leak in the pipeline and the associated anomaly characteristics include the location of the leak as defined with respect to some reference point and the size of leak which is given by an equivalent circular diameter, D L , of the orifice functioning as the leak source.
  • pipeline 310 for generating numerical training data for training the ANN to detect an anomaly 320 in the pipeline 310 according to an illustrative embodiment.
  • pipeline 310 is supplied by a reservoir 330 with an initial head pressure H 0 (in metres), where the pipeline 310 has an internal diameter D and a total length L T .
  • H 0 head pressure
  • L T total length
  • a side discharge valve G that is initially open with a flow Q 0 and which functions to generate a transient pressure wave in the fluid in the pipeline on closing.
  • the transient pressure wave response information or interaction signal is obtained from a measurement point (M), which in this example is positioned at the same location as the side discharge valve G.
  • M measurement point
  • the numerical model used for the generation of the training data is created to replicate the specific characteristics of the analysed pipeline if these are known.
  • a particular numerical model can undergo a further non-dimensional transformation which allows the ANN to determine results for any pipeline configuration regardless of its dimensions.
  • One example system for transforming a transient pressure wave interaction signal into a non-dimensional form is described in PCT Application No PCT/AU2019/000148 (W02020102846) titled “METHOD AND SYSTEM TO ANALYSE PIPELINE CONDITION”, filed by the Applicant here, and whose entire contents are incorporated by reference in their entirety in the present disclosure.
  • an anomaly in the form of a leak could be present at any location along pipeline 310 and additionally the leak could have a range of leak sizes which are modelled as circular orifices having a diameter D L so the training data is generated to covering the potential range of locations of the leak and sizes of the leak.
  • anomaly training data would be generated covering a range of associated anomaly characteristics such as the location of the anomaly and any other relevant associated anomaly characteristics that quantify the anomaly and its effect on the pipeline. These characteristics are set out below in a non-exhaustive list of some anomaly characteristics.
  • the training data is a set of numerically generated transient pressure wave interaction signals based on the modelled pipeline response to the generated transient pressure wave from the anomaly by a computational hydrodynamic model.
  • the computational hydrodynamic model employs the Method of Characteristics (MOC).
  • MOC Method of Characteristics
  • input transient pressure wave is modelled as the instantaneous closure of side discharge valve G.
  • the transient characteristics of the generated transient pressure wave may be more closely modelled.
  • the closure characteristics of the side discharge valve G may be experimentally determined and incorporated into the computational hydrodynamic model.
  • the experimentally determined input transient characteristics of the generated transient pressure wave may be incorporated into the computational hydrodynamic model.
  • 50,000 different transient pressure wave interaction signals are generated by the computational hydrodynamic model corresponding to 5,000 leak locations randomly selected from 5,000 segments into which pipeline 310 has been divided and leak sizes randomly selected from a range of 10 predetermined leak sizes. This range would generally be chosen to cover the expected range of leak sizes given the physical configuration of the pipeline, including its carrying capacity. While in this example, the leak has been parameterised as a circular orifice in other examples, the leak could be parameterised in terms of a flow rate or other suitable physical characteristic of the leak.
  • a dataset of noise modulated training data is generated in accordance with a characteristic pressure noise signal that is based on the pipeline that is being analysed and the type of anomaly of the anomaly that the ANN will be trained to detect.
  • training data for the anomaly detection ANN is generated as has been described above.
  • the training data may be determined empirically by measurements of pipeline systems exhibiting the relevant anomaly which in another example may be combined with numerically generated training data.
  • a plurality of pressure wave interaction signals are generated that cover the expected range of the anomaly characteristics associated with the anomaly type.
  • a representative pressure variation is determined based on the pipeline and the type of anomaly that the ANN is being trained to detect.
  • the representative pressure variation is a measure of the variation in pressure that would be expected as a result of the anomaly and provides a basis for determining a characteristic pressure noise signal that may be introduced into the training data.
  • a representative pressure fluctuation may be determined by modelling the change in pressure resulting from the anomaly and/or by empirical measurement and/or by theoretical calculation.
  • FIG. 5 there is shown a plot 500 of two example transient pressure wave interaction signals overlayed with respect to each other where the unit is head pressure in metres (see left hand axis).
  • the first transient pressure wave interaction signal 510 (shown in continuous line) corresponds to the response from a generated transient pressure wave in the form of the closure of the side discharge valve G where there is no anomaly in pipeline 310, ie the pipeline is intact with no anomaly.
  • the second overlayed transient pressure wave interaction signal 520 (shown in dash-dotted line) corresponds to the transient pressure wave response from a generated transient pressure wave in the form of the closure of the side discharge valve G, except now there is a leak 320 in pipeline.
  • the leak is chosen to be the smallest leak that is of interest to be detected by the anomaly detection method.
  • typical historical anomalies detected in a particular pipeline or anomalies with the severity of interest to be detected can be used to determine the pressure variation used for the noise generation.
  • the pressure difference 530 Ah between the first and second transient pressure wave interaction signals shown in dashed line using right hand axis.
  • first and second transient pressure wave interaction signals 510, 520 have been numerically generated based on pipeline model 300 shown in Figure 3, in other embodiments, these transient pressure wave interaction signals and their difference may be determined empirically by measurement of a pipeline either as deployed or in a laboratory setting or obtained from historical data available from the analysed pipeline to determine a representative pressure fluctuation.
  • Plot 610 as shown in Figure 6a is the average leak location error (as a percentage) when an ANN is trained with the original training data without the addition of any noise modulated data. This shows that the ANN performs satisfactorily in determining the location of leaks in numerical pipelines for all leak sizes.
  • Plot 620 in Figure 6b shows the average leak location error (as a percentage) for an ANN trained with the original training data combined with a pressure noise signal with a standard deviation equal to Ah 0 .
  • the ability of the ANN to detect leaks with the smaller sizes from the numerical pipeline decreases when noise is added to the signal.
  • the relative error in the leak location is less than 5% and for the rest of the leak sizes it is around 0.5%. This shows that even though the addition of noise to the training data can affect the accuracy of the detection of small leaks, this effect is not significant and the ANN is able to predict the location of anomalies in pipelines experiencing background noise.
  • the representative pressure fluctuation is determined by comparing a first transient pressure wave interaction signal where there is no anomaly in the pipeline with a second transient pressure wave interaction signal where the pipeline has the anomaly.
  • the next step is to determine a characteristic pressure noise signal that will be used to modulate the original training data to provide the noise modulated training data referred to above.
  • the noise modulated training data comprises the original training data combined with a characteristic pressure noise signal having a variability corresponding to the representative pressure variation of the pipeline as determined above.
  • the form of the characteristic pressure noise signal is a Gaussian distribution with a zero mean and having a standard deviation s defined by the representative pressure variation referred to above, ie Ah 0 .
  • the standard deviation s is equivalent to a noise intensity as it quantifies the variation of pressure about the mean value.
  • the noise modulated training data is generated by taking one pressure data point from one transient pressure wave interaction signal and adding to that data point a random noise value sampled from the Gaussian distribution forming the characteristic pressure noise signal. This process is then repeated for every pressure data point in each transient pressure wave interaction signal of the original training data to form multiple sets of noise modulated training data.
  • the form of the characteristic pressure noise signal can be a random, pseudo-random or a chaotic time series. While a random signal with a Gaussian distribution is a good working choice, other signal types and distributions can be used such as Gamma, Beta, Poisson or negative binomial. Essentially those signal distributions that most efficiently cover the parameter phase space are the ones that provide the best performance.
  • FIG. 7 a to 7d there are shown a series of plots depicting the original transient pressure wave interaction signal comprising the original training data overlayed with the noise modulated transient pressure wave interaction signal which comprises the noise modulated training data that will be used to train the ANNs according to an illustrative embodiment.
  • the value for s, or equivalently the noise intensity is increasing.
  • FIG. 8 there is shown a data flow diagram 700 showing figuratively the training of a series of ANNs 711, 712, 713, 714, 715 on respective sets of noise modulated training data 720 that in this example has been generated from original training data 710 in accordance with an illustrative embodiment.
  • the original training data 710 will comprise a first set of training data consisting of multiple transient pressure wave interaction signals (eg, 50,000) that characterise the anomaly which has either been generated numerically as described above or obtained empirically or a combination of these two sources.
  • transient pressure wave interaction signals eg, 50,000
  • Multiple sets of noise modulated training data are then generated each based on the original set of training data.
  • Multiple ANNs are then trained on respective sets of noise modulated training data.
  • five separate ANNs are trained on five respective sets of noise modulated training data 720 each, in this example, resulting from the originally generated training data 710.
  • additional datasets of noise modulated training data are generated by additionally modulating the plurality of pressure wave interaction signals comprising the original set of training data by the characteristic pressure noise signal.
  • the original noise modulated training data is randomly divided into groups of equal size comprising a reduced set of noise modulated training data and an equal sized set of noise modulated testing data for testing the ANNs once they have been trained with reduced set of noise modulated training data.
  • a loss function metric is computed using different metrics by comparing the predicted results obtained from the ANN and the real value of the location and characteristics of the anomaly.
  • this metric can be the Mean Absolute Error (MAE).
  • MSE Mean Square Error
  • this metric can be the logarithm of the MSE (Log MSE).
  • An iterative process is included to repeat this batch training until a threshold in the loss function is met or the maximum time is reached.
  • the complete noise modulated training data is used to calculate the ANN weights and then validated with the testing data.
  • different metrics are used and the performance of the ANNs are inspected with noise modulated validation data.
  • the learning rate specified during the training process is increased to provide for a quicker training process in each iteration and to increase the variation in the network weights between iterations in order to improve the global search of weight configurations produced by the iterative process.
  • the testing dataset is used to obtain predictions of leak location and size in the dataset that the trained ANNs have not been exposed to. These predictions are then compared to the input anomaly locations and sizes for the testing data.
  • An ANN that has been successfully trained should present with a similar distribution of errors in the training and the testing datasets.
  • an ANN test measure is determined based on the performance of the trained ANN for the training and the testing dataset. In one example, the Root Mean Square Error (RMSE) from the results of the trained ANN for the training and the testing dataset are computed.
  • RMSE Root Mean Square Error
  • the leak location and size error are both individually computed. Once all the individual errors for leak location and size have been calculated, the RMSE is computed as a unified metric of the ANN performance on the training and the testing dataset.
  • This process assists, in combination with the results from the leak location and size predictions obtained from a measured transient pressure wave interaction signal, in choosing the correct noise intensity for noise modulation of the original training data.
  • the RMSE obtained for the training dataset and the testing dataset are similar as this demonstrates that the ANN is able to predict the characteristics of the leak similarly for a set of transient pressure wave interaction signals that have not been analysed by the ANN before.
  • a performance measure is determined that characterises the detection of the anomaly based on the results from each of the ANNs.
  • the performance measure assesses the variability of the anomaly detection results from the series of ANNs by determining the variability in the associated anomaly characteristics for the anomaly type being detected.
  • the performance measure comprises determining the variability between the different leak locations that are determined by each of the ANNs that are trained on respective sets of noise modulated data.
  • the performance measure comprises determining the variability between the different leak apertures that are determined by each of the ANNs that are trained on respective sets of noise modulated data.
  • the performance measure may be a combined measure over a number of associated characteristics.
  • a box whisker plot may be generated summarising the distribution of the predicted leak locations determined by each of the ANNs that are trained on respective sets of noise modulated data.
  • a box whisker plot is a standardized way of displaying the variability in a dataset by showing the minimum and maximum values of the dataset as well as the median value and first and third quartiles of the dataset.
  • the leak location predictions can be adjusted to a probability distribution to characterise this distribution.
  • a histogram can be built based on the ANN predictions organised by each noise intensity.
  • violin plots a plot combining a box plot and a kernel density plot, can be used to assess the performance of the ANNs trained on respective sets of noise modulated data considering that this type of plot allows for the analysis of the density curve of the ANNs training on different noise intensities.
  • Method 800 is essentially equivalent to method 100 illustrated in Figure 1 except that it includes a pre-processing step of pre processing the measured transient pressure wave interaction signal at step 825 prior to the processing of the transient pressure wave signal by the trained ANNs at step 830 (equivalent to step 130 of method 100) as well as an optional verification step at step 850 where the results from the anomaly detection are verified.
  • FIG. 10 there is shown a flowchart 900 depicting example pre processing methods or approaches that may be adopted either solely or in combination in accordance with pre-processing step 825.
  • pre-processing the transient pressure wave interaction signal at step 825 comprises at step 910 the downsampling of the transient pressure wave interaction signal to generate a downsampled transient pressure wave interaction signal.
  • a is again the wave speed in the pipeline. This implies that for a desired spatial resolution, ie, Dc, a specific time resolution needs to be selected for a pipeline for a given wave speed a.
  • the transient pressure wave interaction signal may be downsampled in the time domain as referred to above.
  • the measured transient pressure wave interaction signal may be downsampled to an equivalent downsampled sampling frequency using a uniform selection of the n-th sample of the transient pressure wave interaction signal.
  • the size of the resulting downsampled transient pressure wave interaction signal in this example depends on the size of the original pressure trace and the selected n.
  • the measured transient pressure wave interaction signal may be downsampled to an equivalent downsampled sampling frequency by averaging the values of an n-th block of sampled pressure values into one value of pressure.
  • the sampling frequency and the frequency used for the training of the ANNs need to be related by an integer n.
  • the measured transient pressure wave interaction signal may be downsampled to an equivalent downsampled sampling frequency by defining a new sample grid that matches the one used for the training of the ANN.
  • the pressure value in the new grid is calculated by interpolation (eg, linear, quadratic, cubic, Gaussian, nearest neighbour, etc).
  • the downsampling frequency eg, selecting every n-th sample or averaging over every n-th sample block or grouping
  • the downsampling frequency does not need to be explicitly related to the frequency used for training the ANN by an integer factor.
  • the final size of the downsampled transient pressure wave interaction signal and, therefore, the size of the input for the ANN may be selected depending on the desired resolution for the identification of the features. As would be appreciated, there is a trade-off between the equivalent downsampled sampling frequency of the downsampled transient pressure wave signal and the computational time required to develop the training and testing of the ANN. A larger input dataset for the ANN will require in general more time to train, however, the testing time is not affected to the same extent.
  • the downsampled sampling frequency is selected from the following ranges, including, but not limited to: greater than 200 Hz, 200 Hz - 250 Hz, 250 Hz - 300 Hz, 300 Hz - 350 Hz, 350 Hz - 400 Hz, 400 Hz - 450 Hz, 450 Hz - 500 Hz, greater than 500 Hz, 500 Hz - 550 Hz, 550 Hz - 600 Hz, 600 Hz - 650 Hz, 650 Hz - 700 Hz, 700 Hz - 750 Hz, 750 Hz - 800 Hz, 800 Hz - 850 Hz,
  • the ratio of the downsampled sampling frequency to the detection sampling frequency is selected from the following ranges, including, but not limited to: 0.01 - 0.025, 0.025 - 0.05, 0.05 - 0.075, 0.075 - 0.1, 0.1 - 0.15, 0.15 - 0.2, 0.2 - 0.25, less than 0.25, 0.25 - 0.3, 0.3 - 0.35, 0.35 - 0.4, 0.4 - 0.45, 0.45 - 0.50, less than 0.5, 0.5 - 0.55, 0.55 - 0.6, 0.6 - 0.65, 0.65 - 0.7, 0.7 - 0.75, less than 0.75, 0.75 - 0.8, 0.8 - 0.85, 0.85 - 0.9 or 0.9 - 0.95.
  • the 825 comprises at step 920 pre-processing the transient pressure wave interaction signal to remove signal artefacts introduced or arising from the original generation of the transient pressure wave in the pipeline.
  • the signal artefacts removed are based on those identified by an ANN sensitivity analysis.
  • FIG. 11 there is shown a plot 1000 of the results of a sensitivity analysis conducted on an ANN trained to detect an anomaly to in turn identify a region of interest in the transient pressure wave interaction signal.
  • the anomaly is a leak in a pipeline and the associated characteristic being tested for sensitivity is the leak location.
  • Plot 1000 shows the results of successive testing of the ANNs where the transient pressure wave interaction signal that is being processed by the ANN has been perturbed at each point along the transient pressure trace 1050.
  • a 0.1 m perturbation has been applied to the measured pressure head at each pressure perturbation location and five ANNs each trained on the original training dataset (with no noise modulated data) are applied to these perturbed input pressure traces to provide a distribution of errors in leak location 1010.
  • the plotted errors in leak location 1010 show that perturbations in a region 1020 defined by the first 60 points of the transient pressure head trace, which corresponds to the steady state pressure head before the generation of the transient pressure wave which in this case results from the closure of valve G (see Figure 3), induce a considerably larger error in the distribution of the leak location predictions.
  • the ANN predictions of leak location are more consistent, although errors are also present due to the perturbation.
  • This sensitivity analysis demonstrates that the pressure variation in a region of interest defined by the initial steady state portion of the measured transient pressure wave interaction signal from prior to generation of the transient pressure wave will potentially induce the largest errors in the overall ANN performance.
  • FIG. 12 there is shown a plot 1100 of a measured transient pressure wave interaction signal 1110 following the generation of a transient pressure wave 1111 showing in detail the pr5essure fluctuations before and after the transient event according to an illustrative embodiment.
  • a) shows the pressure fluctuations before the generation of the transient pressure wave, corresponding to region of interest 1020 illustrated in Figure 11, by in this case valve closure
  • subplot b) shows the pressure fluctuations after the dissipation of the transient event created by the generated transient pressure wave.
  • the increased pressure fluctuations prior to the generated pressure wave relative to the pressure fluctuations after the pressure event are due to the interaction that the transient or pressure wave generator, ie, in this example an open valve initially allowing fluid to exit the pipeline, has with the pipeline.
  • the operating characteristics of the pressure wave generator being an open valve that is suddenly shut, contributes an initial pressure fluctuation to the transient pressure wave interaction signal prior to the generation of the transient pressure wave which is independent of the leak.
  • pressure fluctuations or artefacts in this region of interest prior to the generation of the transient pressure wave additionally have an enhanced capability to cause errors in the ANNs determination of the anomaly.
  • the artefacts introduced or arising from the original generation of the transient pressure wave in the pipeline correspond to an additional pressure fluctuation or “noise” in the transient pressure wave interaction signal that are assumed to have a constant standard deviation.
  • pre-processing at step 920 to remove the artefacts comprises reducing the pressure fluctuation or variation in a region of interest or portion of the transient pressure wave interaction signal so that fluctuations or variations in pressure are those associated with the presence of the anomaly, ie in this case a leak in the pipeline.
  • FIG. 13 there is shown a plot 1200 of the distributions of pressure head 1210, 1220 for the (a) and (b) regions shown in Figure 12 as fitted with corresponding normal distributions having mean and standard deviations of ( i a , s a ) and (ji b , a b ) respectively.
  • the (b) region following stabilisation after the transient pressure wave has a larger mean fi b given that the pipeline pressure will stabilise at a higher pressure head value due to the reduction in total flow in the pipeline following the closure of the discharge valve, but still overall has a smaller standard deviation a b when compared to the pressure fluctuation or s a before generation of the transient pressure wave, which as discussed above results from the flow through the discharge valve prior to the closure.
  • the pressure fluctuation in the region of interest having characteristics ( i a , s a ) is processed to have a similar pressure fluctuation characteristics to region (b) corresponding to when the transient pressure wave interaction signal goes back to the steady state following generation of the transient pressure wave.
  • FIG 14 there is shown a plot 1300 of the cumulative distribution function (CDF) 1310 for the (a) region shown in Figure 12 having characteristics (m a , s a ) and a modified or transformed CDF 1320 having the same mean but with the standard deviation of the (b) region, ie having characteristics (c3 ⁇ 4).
  • CDF 1320 preserves the mean value of the pressure fluctuations before the generation of the transient pressure wave but its distribution is modified to match the pressure fluctuations after the transient test.
  • the region of interest is then processed by selecting a pressure value from the region of interest prior to the generation of the transient pressure wave and then computing the cumulative probability based on original CDF 1310. With this value of probability, a new value for the pressure is found using the modified CDF 1320. These two steps are then repeated for each value of pressure in the region of interest.
  • pre-processing the transient pressure wave interaction signal at step 825 comprises at step 930 pre-processing the transient pressure wave interaction signal by shifting or offsetting the initial pressure of the measured transient pressure wave interaction signals in line with the initial pressure of the ANNs training samples.
  • This shifting includes determining the difference between the mean pressure in the measured transient pressure wave interaction signal before the generation of the transient pressure wave and the steady state pressure used in the training of the anomaly detection ANNs and then transformation of the transient pressure head traces by adding or subtracting this difference.
  • method 800 also comprises an optional verification of the results of the anomaly detection at step 850.
  • verification of the results of the anomaly detection involves determining whether the determined associated characteristics are physically consistent with the pipeline. In one example, this verification could include determining whether a location of the anomaly is consistent with a length of pipeline.
  • verification of whether the anomaly has been detected in the transient pressure wave interaction signal comprises numerically generating a theoretical transient pressure signal based on the detected anomaly and the associated anomaly characteristics and then comparing the measured transient pressure wave interaction signal with the numerically generated theoretical transient pressure signal to determine a comparison measure. A comparison threshold may then be applied to the comparison measure to verify that the anomaly has been detected in the transient pressure signal.
  • FIG. 15 there is shown a pipeline arrangement 1400 for experimental confirmation of the methods and systems disclosed the present disclosure.
  • a pipeline 1410 having a length of 37.24 metres and an internal diameter of 22.14 mm with a wall thickness of 1.63 mm is originally connected at both ends to two pressurized tanks 1460, 1465. This corresponds to a wave speed a for the pipe of 1305 m/s and a corresponding L T /a time of 0.029 s.
  • An inline valve 1420 has been closed on the downstream of the pipeline to allow flow only through a solenoid valve 1440 installed right before the end of the pipeline at location Gl.
  • a circular orifice of size 2.2 mm has been installed at 28.52 m downstream of source tank 1465 to simulate a leak.
  • the transient pressure wave in the fluid is generated by the fast closure of the solenoid valve Gl with a closure time of 5 ms and the pressure has been measured with a PDCR 810 pressure sensor or detector at point 1440 with a 10 kHz sampling rate.
  • Each of the transient pressure waves was generated and measured using the same pipeline configuration and under the similar initial conditions.
  • the initial pressure at the end of pipeline 1410 was set between 20.0 and 23.9 m and the transient pressure wave interaction signal has been measured from almost 7 L T /a (0.2 seconds) before the generation of the transient pressure wave (ie, valve closure) for a duration of 103L r /a (3 seconds) following generation.
  • the time window has been limited to a duration of 1 second.
  • next step involves the processing the transient pressure wave interaction signals by a series of ANNs each trained on respective sets of noise modulated training data the question arises as to what is the appropriate level of noise modulation that should be adopted.
  • a 1-D convolutional ANNs has been adopted in accordance with the architecture previously described having four convolutional layers, 20 to 30 filters and three dense layers.
  • a total of 50,000 numerically generated transient pressure wave interaction signals were generated based on a MOC based computation hydrodynamic model by modelling 10 leaks at random locations each 7.45 mm along pipeline 1410 with different diameters ranging between 0.4 and 3.5 mm.
  • the total simulation time was set to 0.09 s which corresponds to 3.15L/a seconds, ie, L/a seconds before the generation of the transient pressure wave and 2.15 L/a seconds after transient pressure wave have been generated to account for the effects of in this case the valve closure curve in the computed pressure.
  • the time resolution of the MOC hydrodynamic model is required to be at least 0.006 ms; therefore, the total size of the input dataset before any downsampling process would be 788 million transient pressure head values where each time window has almost 16,000 pressure head values.
  • the transient pressure wave signals are downsampled to generate a downsampled transient pressure wave interaction signal.
  • a downsampling frequency of 5 kHz is selected. This downsampling frequency is selected based on a consideration of the dimensions of pipeline 1210 and the potential number of weights to train in the resulting ANNs.
  • an overly small downsampling frequency would create a very small ANN that would not be able to learn enough information from the transient pressure head traces forming the training dataset. Smaller downsampling frequencies may be selected for larger pipelines with larger L/a characteristics.
  • the resulting number of weights for the anomaly detection ANNs is 13,868 and the input dataset contains 8.55 million individual transient pressure head values for the 50,000 transient pressure wave signals.
  • the next step is the generation of noise modulated training data for the ANNs.
  • a representative pressure fluctuation is determined based on the pipeline configuration and the type of anomaly.
  • the smallest leak drop ( h 0 as shown in Figure 5) corresponding to the smallest leak considered was determined to be 0.1238 m.
  • six different noise intensities were considered in accordance with Equation 2 and the selected values of k n and resulting standard deviation s h are presented in Table 1.
  • ANNs with the ability to find leaks across the complete defined leak size range without significantly decreasing performance with the addition of noise.
  • the number of different noise intensities that may be analysed will be determined depending on the pipeline system and potentially the available computing resources. [00162] The selection of the number of noise intensities in this example and their magnitude is defined based on observations of the pressure noise in the system, the normal operational changes in the pipeline that might create pressure fluctuations and the expected background noise.
  • ANNs trained on respective datasets of noise modulated training dataset of noise modulated training data according to an illustrative embodiment.
  • Table 1 This results in the training and testing of a total of 35 ANNs where 5 ANNs are trained on the original training data and the remaining 30 ANNs are trained in series of 5 ANNs on respective sets of noise modulated training data with six levels of increasing noise intensity.
  • each dataset of 250,000 transient pressure wave signals was divided into a training dataset of 125,000 transient pressure wave signals for training and a second testing dataset for testing of the respective trained anomaly detector ANN. This division process is repeated for each of the 5 ANNs trained with the noise modulated training data with six level of noise intensity.
  • each ANN is trained and tested with a different dataset.
  • FIG. 18a-g there is shown a series of plots illustrating the percentage of exceedance associated with the absolute average error for the location of a leak following training and testing for each noise intensity according to an illustrative embodiment.
  • a single ANN for each noise intensity is shown from the five trained and tested ANNs as the distribution of errors was found to be consistent across the five ANNs for each noise intensity.
  • Each plot shows two lines.
  • the solid line corresponds to the distribution of the absolute average leak location error on the dataset used for the ANN training while the dotted line shows the leak location error for the dataset used during the ANN testing.
  • the percentage of exceedance can be interpreted as the proportion of the total trained or tested samples that the average leak location exceeded a certain error size.
  • An average error in the predictions is presented because in some cases two or more traces with the same leak location and size have been used either for the training or the testing.
  • Results from Figure 18 demonstrate that the ANN training methodology using noise modulated data was successful when tested with numerical data. However, the selection of a noise modulated ANN to predict the location and characteristics of an anomaly depends on the performance of the ANN when tested with transient pressure wave interaction signal obtained from the analysed pipeline, not from numerical data. These results demonstrate that the range of noise intensities was selected appropriately and ANNs can still learn about the presence of anomalies despite the background noise.
  • the transient pressure wave interaction signals are processed to reduce any pressure fluctuations associated with the generation of the transient pressure wave (see step 920 of Figure 10).
  • pressure fluctuations arising from the flow through the solenoid valve installed at the end of the pipeline prior before it is closed to generate the transient pressure wave are reduced.
  • the resulting transient pressure wave interaction signals have been further processed by being shifted to be aligned to one initial average steady state pressure (see step 930 of Figure 10).
  • the initial pressure value of each time window was slightly different in a 3.9 m range; therefore, all transient pressure wave signals have been aligned to an average steady state pressure of 21.16 m, which corresponds to the initial pressure considered for the numerically generated transient pressure wave interaction signals.
  • the resulting shifted transient pressure signals are also trimmed to select only the region of interest corresponding to L/ a seconds before the generation of the transient pressure wave and 2.15 L/a seconds after the generation of the pressure wave.
  • FIG 19 there is shown a plot 1800 of transient pressure wave interaction signals illustrated in Figure 16 following pre-processing according to an illustrative embodiment.
  • the initial steady state pressure 1810 for the transient pressure wave interaction signals has been shifted to the same value.
  • the increase in pressure 1820 after the generation is different. This is expected due to the small differences in the resulting flow in the pipeline given different initial steady state pressures, however, in this example there was no further pre-processing except for downsampling of the original transient pressure signal to a downsampled frequency of 5 kHz to match the input size of the trained anomaly detection ANNs.
  • FIG. 20 to 26 there are shown plots of the distribution of predictions for determining the location of a leak for each of the transient pressure wave interaction signals illustrated in Figure 19 following processing by a series of ANNs trained on noise modulated training each having different noise intensities according to an illustrative embodiment.
  • Each of the plots indicates the end of the pipeline 1910 as well as the location 1920 of the leak in the pipeline at 28.05 m.
  • the ANNs have been trained with smooth numerical samples, the predictions when the analysed transient pressure head traces have some pressure fluctuations result in illogical predictions for the leak location.
  • the ANNs trained with the original training data might not be able to locate the anomaly without the addition of noise in the training samples.
  • the ANNs trained with the original data will be able to accurately predict the location of the anomaly.
  • Figures 21 to 26 demonstrate that as the noise intensity s h increases the distribution of the leak locations are more compact and are, in general, closer to the actual leak location.
  • Predictions from ANNs trained with noise intensities s 2 and s 3 are within the length of the pipeline but vary considerably between the different measured transient pressure signals.
  • leak location predictions obtained from the last three noise intensities (s 4-6 ) oscillate between 26.05 and 31.85 m with a couple of predictions outside the physical length of the pipelines for s 4 .
  • transient pressure wave interaction signals #1 and #12 generally resulted in more scattered leak location predictions.
  • Leak location predictions for time window #1 are less satisfactory as there is a more significant difference between the initial steady state pressure and the resulting pressure increase after the generation of the transient wave as it can be seen in Figure 19 (see pressure trace 1840).
  • transient pressure wave interaction signal #12 does not present with any particular differences in comparison with the other transient tests, it has produced less consistent results for all the noise intensities indicating that there might have been additional noise during this test. Considering these results, conducting multiple tests are likely to provide additional information about the performance of the performance of the trained ANNs for a selected noise intensity for the noise modulated training data and may be recommended.
  • the effectiveness or performance of a group of ANNs may be assessed by determining a performance measure that characterises the detection of the anomaly that is based on the results from each of the ANNs where there is a higher performance when the results from each of the ANNs are consistent with each other.
  • FIG. 27 there is a plot 2600 of the performance of the trained ANNs on the measured and processed transient pressure wave signals #1 to #14 illustrated in Figure 19 for each noise intensity according to an illustrative embodiment.
  • the average of the absolute median predicted leak location error computed for transient pressure wave signals #1 to #14 is shown for each corresponding noise intensity (as indicated by the respective associated standard deviation on the x-axis) (eg, 2610) as well as the distribution of the absolute median predicted leak location error (eg, 2620).
  • the median predicted leak location for a given noise intensity and a given measured transient pressure wave signal will be determined from the five ANNs trained on the corresponding noise modulated training set having that noise intensity.
  • FIG. 28 there is shown a plot 2700 of the median values (in circles) and distributions of the Root Mean Square Error (RMSE) for the training (in lighter shade) and the testing (in black) of the trained anomaly detector ANNs for each noise intensity as shown in Table 1.
  • RMSE Root Mean Square Error
  • Table 1 the distribution of the RMSE has been computed using the leak location error from each of the 125,000 samples used for the training of the ANNs (or 25,000 for the case of the ANNs trained without any noise). It is expected that good ANN performance would correspond to obtaining low values of RMSE and consistent magnitudes between the training and the testing RMSE.
  • FIG. 29 there is shown a plot 2800 of the average error in RMSE (in circles) and the range for the prediction of a leak location (whiskers) for the training (in lighter shade) and the testing (in black) of the trained anomaly detector ANNs for each noise intensity as shown in Table 1.
  • the optimum noise intensity for the noise modulated training data for training of the leak detection ANN is obtained when the noise has a Gaussian noise distribution having a standard deviation of 0.0619 m (ie, s 4 ) as the median leak location prediction for this set of ANNs was 28.74 m compared to an actual leak location of 28.52 m and the median predicted leak size was 2.32 mm compared to an actual leak size of 2.2 mm.
  • the relevant trained ANNs that would be adopted for anomaly detection would be those that have been trained on a respective sets of noise modulated training date where the noise intensity would correspond to a standard deviation of 0.0619 m.
  • the selection of the relevant trained group of ANNs that is adopted for anomaly detection depends on the distribution of the anomaly locations and the performance of the ANNs during training and testing. In one example, all ANNs trained have a successful performance during training and testing (with no evidence of overfitting). In this case, the group of ANNs adopted for anomaly detection is the one with the smallest distribution of anomaly locations. In another example, some ANNs trained do not present with a successful training because the testing stage demonstrates overfitting. In this case, these groups of ANNs are discarded and the group of ANNs adopted for anomaly detection is the one with the smallest distribution of anomaly locations from the remaining groups.
  • the determined anomaly is then subject to a further verification stage where the determined characteristics of the anomaly form the inputs for numerically generating a transient pressure wave interaction signal responsive to a generated transient wave which can then be compared to the measured transient pressure wave signal.
  • FIG. 30 there is shown a plot 2900 comparing the predicted transient pressure wave signal 2910 obtained by numerically generating the expected trace based on the determined anomaly characteristics with the measured transient pressure wave signal 2920 used to determine the anomaly characteristics according to an illustrative embodiment.
  • measured transient pressure wave signal 2920 corresponds to transient pressure wave signal #8. It can be observed in this figure that there is reasonable concordance between the pressure values for the measured and predicted transient pressure wave signals pointing to a successful prediction of the location and size of the leak using the series of ANNs trained on noise modulated data with a standard deviation of 0.0619 m (ie, s 4 ) as described above.
  • the Normalised Root Mean Square Error (NRMSE) has been computed between these two transient pressure head traces obtaining a value of 2.06% demonstrating again the accuracy of the methodology proposed.
  • NPMSE Normalised Root Mean Square Error
  • method and systems in accordance with the present disclosure for detecting anomalies in a pipeline where noise modulated training data are employed to train a series of ANNs not only improve the results but also provide a performance measure that allows the performance of the detection task to be characterised more fully.
  • the results from the series of ANNs may concentrate around two different values and then the verification step may then be used to determine which set of results are correct.
  • methods and systems in accordance with the present disclosure have the advantage that a generated characteristic pressure noise signal may be added to the training data for the ANNs without the requirement to specifically replicate the background noise of the pipeline system.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Theoretical Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Analytical Chemistry (AREA)
  • Biochemistry (AREA)
  • Chemical & Material Sciences (AREA)
  • Immunology (AREA)
  • Pathology (AREA)
  • Molecular Biology (AREA)
  • Mathematical Physics (AREA)
  • Signal Processing (AREA)
  • Software Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Remote Sensing (AREA)
  • Acoustics & Sound (AREA)
  • Environmental & Geological Engineering (AREA)
  • Geology (AREA)
  • Geophysics (AREA)
  • General Life Sciences & Earth Sciences (AREA)
  • Measuring Fluid Pressure (AREA)
  • Pipeline Systems (AREA)

Abstract

Un procédé et un système de détection d'une anomalie dans un pipeline sont divulgués. Le procédé et le système consistent à générer une onde de pression transitoire dans un fluide transporté le long du pipeline et à mesurer un signal d'interaction d'onde de pression transitoire correspondant à une réponse de l'anomalie à l'onde de pression transitoire générée. Le procédé et le système consistent ensuite à traiter le signal d'interaction d'onde de pression transitoire pour détecter l'anomalie au moyen d'une série de réseaux de neurones artificiels (ANN), chacun d'eux étant entraîné pour détecter l'anomalie sur des jeux de données respectifs de données d'entraînement modulées par bruit, et à déterminer une mesure de performance caractérisant la détection de l'anomalie sur la base de résultats de détection d'anomalie de chacun des ANN.
PCT/AU2022/050194 2021-03-09 2022-03-09 Procédé et système de détection d'anomalie de pipeline Ceased WO2022187898A1 (fr)

Priority Applications (1)

Application Number Priority Date Filing Date Title
AU2022234947A AU2022234947A1 (en) 2021-03-09 2022-03-09 Pipeline anomaly detection method and system

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
AU2021900659 2021-03-09
AU2021900659A AU2021900659A0 (en) 2021-03-09 Pipeline anomaly detection method and system

Publications (1)

Publication Number Publication Date
WO2022187898A1 true WO2022187898A1 (fr) 2022-09-15

Family

ID=83226080

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/AU2022/050194 Ceased WO2022187898A1 (fr) 2021-03-09 2022-03-09 Procédé et système de détection d'anomalie de pipeline

Country Status (2)

Country Link
AU (1) AU2022234947A1 (fr)
WO (1) WO2022187898A1 (fr)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117520996A (zh) * 2024-01-04 2024-02-06 长沙小枰建材有限公司 一种用于涵管的实时压力测试方法
US11953161B1 (en) 2023-04-18 2024-04-09 Intelcon System C.A. Monitoring and detecting pipeline leaks and spills
CN118393311A (zh) * 2024-06-28 2024-07-26 南通华隆微电子股份有限公司 用于半导体封装的高温反偏检测装置
CN120721322A (zh) * 2025-09-01 2025-09-30 重庆文灿压铸有限公司 一种新能源汽车压铸件气密性自动检测装置

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2007056803A1 (fr) * 2005-11-15 2007-05-24 Bernadette Garner Methode pour entrainer des reseaux neuraux
CN101008992A (zh) * 2006-12-30 2007-08-01 北京市劳动保护科学研究所 基于人工神经网络的管道泄漏的检测方法
WO2017008100A1 (fr) * 2015-07-10 2017-01-19 The University Of Adelaide Système et procédé de génération d'un signal de pression
US20190324432A1 (en) * 2017-08-02 2019-10-24 Strong Force Iot Portfolio 2016, Llc Data monitoring systems and methods to update input channel routing in response to an alarm state
CN111783945A (zh) * 2020-06-22 2020-10-16 东北石油大学 基于迁移学习lstm的长输油管道泄漏实时检测系统
WO2021022315A1 (fr) * 2019-08-02 2021-02-11 The University Of Adelaide Procédé et système permettant de surveiller l'état d'un pipeline

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2007056803A1 (fr) * 2005-11-15 2007-05-24 Bernadette Garner Methode pour entrainer des reseaux neuraux
CN101008992A (zh) * 2006-12-30 2007-08-01 北京市劳动保护科学研究所 基于人工神经网络的管道泄漏的检测方法
WO2017008100A1 (fr) * 2015-07-10 2017-01-19 The University Of Adelaide Système et procédé de génération d'un signal de pression
US20190324432A1 (en) * 2017-08-02 2019-10-24 Strong Force Iot Portfolio 2016, Llc Data monitoring systems and methods to update input channel routing in response to an alarm state
WO2021022315A1 (fr) * 2019-08-02 2021-02-11 The University Of Adelaide Procédé et système permettant de surveiller l'état d'un pipeline
CN111783945A (zh) * 2020-06-22 2020-10-16 东北石油大学 基于迁移学习lstm的长输油管道泄漏实时检测系统

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11953161B1 (en) 2023-04-18 2024-04-09 Intelcon System C.A. Monitoring and detecting pipeline leaks and spills
CN117520996A (zh) * 2024-01-04 2024-02-06 长沙小枰建材有限公司 一种用于涵管的实时压力测试方法
CN117520996B (zh) * 2024-01-04 2024-04-02 长沙小枰建材有限公司 一种用于涵管的实时压力测试方法
CN118393311A (zh) * 2024-06-28 2024-07-26 南通华隆微电子股份有限公司 用于半导体封装的高温反偏检测装置
CN120721322A (zh) * 2025-09-01 2025-09-30 重庆文灿压铸有限公司 一种新能源汽车压铸件气密性自动检测装置

Also Published As

Publication number Publication date
AU2022234947A1 (en) 2023-10-19

Similar Documents

Publication Publication Date Title
WO2022187898A1 (fr) Procédé et système de détection d'anomalie de pipeline
Wang et al. Pipeline leak detection method based on acoustic-pressure information fusion
Wang et al. Pipeline leak detection by using time-domain statistical features
US12091842B2 (en) Method and system to monitor pipeline condition
Oliver et al. Calibration of imperfect models to biased observations
US12140570B2 (en) Method and system to analyse pipeline condition
CA3110977C (fr) Detection et quantification de flaques de liquide dans des pipelines de d'hydrocarbure fluide
ZIO Model output uncertainty in risk assessment
CN107917341B (zh) 一种输油管道泄漏检测方法及装置
Gong et al. Experimental validation of gas leak detection in screw thread connections of galvanized pipe based on acoustic emission and neural network
CN112539887A (zh) 一种基于wt-lcd-wd的管道泄漏信号去噪方法
Smith et al. Pipeline rupture detection using real-time transient modelling and convolutional neural networks
Ye et al. Multi-variable classification model for valve internal leakage based on acoustic emission time–frequency domain characteristics and random forest
Wang et al. Natural gas transmission pipeline leak detection model based on acoustic emission and machine learning
Lanjile et al. Exploiting multi-modal sensing for increased detection fidelity of pipeline leakage
Romano et al. Burst detection and location in water distribution systems
Oladele et al. Detection of leakages in a pipeline network based on hydraulic laboratory modelling with artificial intelligence
Kampelopoulos et al. Applying one class classification for leak detection in noisy industrial pipelines
Kampelopoulos Development of Alternative Methods for Leak Detection in Industrial Pipelines
Gong et al. Advances in Hydraulic Transient-Based Pipeline Condition Assessment and Feature Diagnosis
Martinez-Ríos et al. Parameter Selection of Generalized Morse Wavelets for Water Leakage Classification
Jara‐Arriagada et al. Spatial interpolation of pressure transient metrics for improved water distribution network asset management
Zheng et al. Machine Learning-Enhanced Leakage Detection in District Heating Networks: Integrating Improved Hydraulic Modeling and Signal Denoising for High-Accuracy Localization
Wang et al. Two-Stage Transient-Based Leak Detection in Water Supply Pipe Networks
Zhao et al. Research on Leak Localization Method for Water Supply Pipelines Based on Optimized Successive Variational Mode Decomposition and Generalized Cross-Correlation

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22765990

Country of ref document: EP

Kind code of ref document: A1

WWE Wipo information: entry into national phase

Ref document number: AU2022234947

Country of ref document: AU

Ref document number: 2022234947

Country of ref document: AU

NENP Non-entry into the national phase

Ref country code: DE

ENP Entry into the national phase

Ref document number: 2022234947

Country of ref document: AU

Date of ref document: 20220309

Kind code of ref document: A

122 Ep: pct application non-entry in european phase

Ref document number: 22765990

Country of ref document: EP

Kind code of ref document: A1