Multisensor method and system for characterizing a pipeline

Abstract

A method and system for determining a condition of a pipeline is disclosed, comprising receiving pressure flow information corresponding to one or more pipeline locations on the pipeline and segmenting the pressure flow information to generate a selection of pressure flow information data segments corresponding to at least one or more pipeline locations determined at two or more time periods, each of the pressure flow information data segments corresponding to pressure flow information received for a unique time interval and pipeline location combination. A respective segment condition measure is then determined for each of the pressure flow information data segments in the selection of pressure flow information data segments and the pipeline condition measure is determined based on the respective segment condition measures.

Description

MULTISENSOR METHOD AND SYSTEM FOR CHARACTERIZING A PIPELINEPRIORITY DOCUMENTS

[0001] The present application claims priority from Australian Provisional Patent Application No. 2023900262 titled “MULTISENSOR METHOD AND SYSTEM FOR CHARACTERIZING A PIPELINE” and filed on 3 February 2023, the content of which is incorporated by reference in its entirety.INCORPORATION BY REFERENCE

[0002] The following publications are referred to in the present application and their contents are incorporated by reference in their entirety:International Patent Application No. PCT / AU2020 / 000036 (W02020215117) titled “METHOD AND SYSTEM FOR DETECTING A STRUCTURAL ANOMALY IN A PIPELINE NETWORK” in the name of The University of Adelaide and filed on 24 April 2020.International Patent Application No. PCT / AU2020 / 000035 (W02020215116) titled “DETECTION OF STRUCTURAL ANOMALIES IN A PIPELINE NETWORK” in the name of The University of Adelaide and filed on 24 April 2020.

[0003] The contents of the above publications are incorporated by reference in their entirety.TECHNICAL FIELD

[0004] The present disclosure relates to determining the condition of a pipeline carrying a fluid. In a particular form, the present disclosure relates to determining the condition of pipeline based on using multiple pressure flow sensors.BACKGROUND

[0005] The sensing of pressure flow information from a pipeline is an important tool for determining the condition of a pipeline as any deterioration in the condition, eg, the forming of leaks, can have significant impacts on the users or customers of the pipeline. An important non-limiting example of the importance of this activity is the detection and characterisation of leaks or breakages in a utility scale pipeline. These pipelines can be used to convey water over long distances or in an urban setting to multiple delivery points by a pipeline network where a leak can also cause significant traffic interruptions, general property damage as well as potentially damage to third party buried telecommunication and power infrastructure. Further, in currently less common instances, a leak orbreakage can result in conditions which risk the safety of the public. Of course, characterisation of pipeline condition can be an important requirement irrespective of the media being conveyed by the pipeline, eg, in the case of wastewater, oil or gas to name a few examples.

[0006] Pipeline leak detection systems have been devised that involve a network of individual pressure flow sensors or sensing region distributed at different locations along a pipeline to measure how the pressure flow information changes as a function of time at or within the detection range of a given sensor. This information is analysed for each sensor to determine whether a leak has occurred for a given location. While these pipeline detection systems have represented some progress in monitoring for leaks they are still focused on individual pipeline locations and how a leak may develop at a given location without having regard to the surrounding pipeline.

[0007] In other developments, the use of a spare installed optical telecommunications fibre (ie, “dark” fibre) that is located proximal to a pipeline has been proposed as a means of determining the condition of a pipeline. In these systems, an optical signal is sent down the dark fibre and reflections are measured from the fibre. This reflection signal may be processed to identify a location along the fibre where the fibre has been subject to the presence of vibroacoustic energy such as may arise from a nearby leak in the pipeline. While there is a beneficial ability to employ already installed telecommunications infrastructure to obtain pressure flow information at multiple locations along a pipeline, there is still the issue that current analysis techniques concentrate on the development of a leak (or other pipeline anomaly) using the data at a particular location.

[0008] Against this background, it would be desirable to provide a method and system for determining the condition of a pipeline that recognises that the pipeline is an extended entity and to assess not only the temporal behaviour but also the behaviour of the pipeline as it varies along the pipeline.SUMMARY

[0009] In one aspect, the present disclosure provides a method for determining a condition of a pipeline, comprising: receiving pressure flow information corresponding to one or more pipeline locations on the pipeline; segmenting the pressure flow information to generate a selection of pressure flow information data segments corresponding to at least one or more pipeline locations determined at two or more time periods, each of the pressure flow information data segments corresponding to pressure flow information received for a unique time interval and pipeline location combination; determine a respective segment condition measure for each of the pressure flow information data segments in the selection of pressure flow information data segments; anddetermining a pipeline condition measure based on the respective segment condition measures.

[0010] In another form, the pressure flow information corresponds to three or more pipeline locations on the pipeline, and wherein segmenting the pressure flow information comprises generating the selection of pressure flow information data segments to correspond to at least three or more pipeline locations determined at two or more time periods.

[0011] In another form, determining a segment condition measure for the pressure flow information data segment comprises determining a signal intensity measure of the pressure flow information data segment.

[0012] In another form, determining a signal intensity measure comprises determining a root-meansquare (RMS) value for the pressure flow information data segment.

[0013] In another form, determining a signal intensity measure comprises: dividing the pressure flow information data segment into a number of frames; determining a respective RMS value for each frame of the number of frames; and determining a lower bound RMS value for which a selected percentage of the respective RMS values exceed.

[0014] In another form, determining the segment condition measure for the pressure flow information data segment comprises determining a spectral / frequency content measure of the pressure flow information data segment.

[0015] In another form, determining the spectral / frequency content measure comprises determining a power spectrum density (PSD) of the pressure flow information data segment.

[0016] In another form, determining the spectral / frequency content measure comprises determining a spectral heat map of the pressure flow information data segment.

[0017] In another form, determining the spectral content measure comprises determining a median frequency (MF) value for the pressure flow information data segment.

[0018] In another form, determining the spectral content measure comprises: dividing the pressure flow information data segment into a number of frames; determining a respective MF value for each frame of the number of frames; anddetermining a lower bound MF value for which a selected percentage of the respective MF values exceed.

[0019] In another form, determining the segment condition measure for the pressure flow information data segment comprises processing the pressure flow information data segment by a machine learning model trained to identify a pipeline anomaly condition.

[0020] In another form, the machine learning model is an artificial neural network (ANN) or a recurrent neural network (RNN).

[0021] In another form, processing the pressure flow information data segment by a machine learning model trained to identify a pipeline anomaly condition comprises: extracting one or more time domain and / or frequency domain features from the pressure flow information data segment; and processing the one or more extracted time domain and / or frequency domain features by a feature based machine learning classifier trained to identify the pipeline anomaly condition based on the one or more extracted time domain and / or frequency domain features.

[0022] In another form, the feature based machine learning classifier is a decision tree model or a support vector machine.

[0023] In another form, processing the pressure flow information data segment by a machine learning model trained to identify a pipeline anomaly condition comprises: forming a 2D spectrogram from the pressure flow information data segment; and processing the 2D spectrogram by a machine learning classifier trained to identify the pipeline anomaly condition based on the 2D spectrogram.

[0024] In another form, the machine learning classifier is a convolutional neural network.

[0025] In another form, determining the segment condition measure for the pressure flow information data segment comprises processing the pressure flow information data segment by a statistical model configured to identify a pipeline anomaly condition.

[0026] In another form, processing the pressure flow information data segment by a statistical model configured to identify a pipeline anomaly condition comprises: extracting one or more time domain and / or frequency domain features from the pressure flow information data segment; andprocessing the one or more extracted time domain and / or frequency domain features by the statistical model configured to receive the one or more extracted time domain and / or frequency domain features as inputs.

[0027] In another form, determining the segment condition measure for the pressure flow information data segment comprises: identifying the associated pipeline location and time period for the pressure flow information data segment; and determining respective coherence measures between the pressure flow information data segment and further pressure flow information data segments from the same or similar time period and corresponding to two or more of the remaining pipeline locations in the selection of pressure flow information data segments.

[0028] In another form, one or more of the remaining pipeline locations correspond to adjacent pipeline locations to the associated pipeline location of the pressure flow information data segment.

[0029] In another form, one or more of the remaining pipeline locations correspond to non-adjacent pipeline locations to the associated pipeline location of the pressure flow information data segment.

[0030] In another form, determining the pipeline condition measure based on the respective segment condition measures comprises: forming an assembly of segment condition measures ordered with respect to time and pipeline location.

[0031] In another form, determining the pipeline condition measure based on the respective segment condition measures comprises: determining whether one or more of the respective segment condition measures satisfy a threshold condition.

[0032] In another form, determining the pipeline condition measure based on the respective segment condition measures comprises: determining whether two or more of the respective segment condition measures corresponding to the same pipeline location satisfy a temporal rate of change condition.

[0033] In another form, determining whether two or more of the respective segment condition measures corresponding to the same pipeline location satisfy a temporal rate of change condition comprises determining whether a temporal benchmark is exceeded.

[0034] In another form, determining whether two or more of the respective segment condition measures corresponding to the same pipeline location satisfy a temporal rate of change condition comprises determining whether a temporal rate of change of power density in one or more selected frequency bands is exceeded.

[0035] In another form, determining whether two or more of the respective segment condition measures corresponding to the same pipeline location satisfy a temporal rate of change condition comprises determining whether a temporal rate of change of coherence level is exceeded.

[0036] In another form, determining the pipeline condition measure based on the respective segment condition measures comprises: determining whether two or more of the respective segment condition measures corresponding to the same or similar time period satisfy a spatial rate of change condition.

[0037] In another form, determining whether two or more of the respective segment condition measures corresponding to the same or similar time period satisfy a spatial rate of change condition comprises determining whether a spatial benchmark is exceeded.

[0038] In another form, determining whether two or more of the respective segment condition measures corresponding to the same or similar time period satisfy a spatial rate of change condition comprises determining whether a spatial rate of change of power density in one or more selected frequency bands is exceeded.

[0039] In another form, determining whether two or more of the respective segment condition measures corresponding to the same or similar time period satisfy a spatial rate of change condition comprises determining whether a spatial rate of change of coherence level is exceeded.

[0040] In another form, determining the pipeline condition measure based on the respective segment condition measures comprises: determining whether two or more of the respective segment condition measures corresponding to the same or similar time period and two or more of the respective segment condition measures corresponding to the same pipeline location satisfy a combined temporal / spatial rate of change condition.

[0041] In another form, determining the pipeline condition measure based on the respective segment condition measures comprises processing the combined respective segment condition measures by a machine learning model trained to identify a pipeline anomaly condition.

[0042] In another form, processing the combined respective segment condition measures by a machine learning model trained to identify a pipeline anomaly condition comprises the machine learning model operating on an assembly of segment condition measures ordered with respect to time and pipeline location.

[0043] In another form, the method further comprises pre-processing a pressure flow information data segment for a selected pipeline location to remove historical background noise associated with the selected pipeline location.

[0044] In another form, the method further comprises pre-processing a pressure flow information data segment for a selected time period to remove historical background noise associated with the selected time period.

[0045] In another form, the method further comprises benchmarking a pressure flow information data segment for a selected pipeline location against an earlier pressure flow information data segment associated with an earlier time period for the selected pipeline location.

[0046] In another form, the method further comprises pre-processing a pressure flow information data segment for a selected pipeline location in accordance with a coherence analysis to compensate for differences in sensor gain and / or environmental noise.

[0047] In another form, the coherence analysis comprises: processing the pressure flow information data segment to form a PSD; determining the maximum coherence level between the pressure flow information data segment and further pressure flow information data segments from the same or similar time period and corresponding to two or more of the remaining pipeline locations in the selection of pressure flow information data segments; and weighting the PSD with the maximum coherence level.

[0048] In another form, the coherence analysis further comprises: normalising the PSD to form a normalised PSD; forming a filtered PSD by filtering the normalised PSD; and weighting the normalised PSD with the maximum coherence level.

[0049] In another form, the method further comprises pre-processing a pressure flow information data segment for a selected pipeline location in accordance with a beamforming analysis to improve the vibroacoustic signal to noise.

[0050] In another form, the beamforming analysis comprises: determining the phase differences between the pressure flow information data segment and further pressure flow information data segments from the same or similar time period and corresponding to two or more of the remaining pipeline locations in the selection of pressure flow information data segments; phase shifting the further pressure flow information data segments to have the same phase as the pressure flow information data segment to form phase shifted further pressure flow information data segments; weighting the phase shifted further pressure flow information data segments in accordance with their level of coherence with the pressure flow information data segment to form weighted phase shifted further pressure flow information data segments; and forming a beamformed pressure flow information data segment by summing in frequency space the weighted phase shifted further pressure flow information data segments to the pressure flow information data segment.

[0051] In another form, phase shifting the further pressure flow information data segments comprises only phase shifting those pressure flow information data segments having a coherence greater than a predetermined coherence threshold.

[0052] In another form, the pressure flow sensing arrangement comprises one or more individual pressure flow sensors each associated with a corresponding location on the pipeline.

[0053] In another form, the pressure flow sensing arrangement comprises a distributed pressure flow sensing apparatus having one or more sensing regions each associated with a corresponding location on the pipeline.

[0054] In another form, the distributed pressure flow sensing apparatus is an optical fibre and the one or more sensing regions are fibre Bragg gratings located along the optical fibre.

[0055] In another form, the distributed flow sensing apparatus is an optical fibre and the one or more sensing regions correspond to regions located along the fibre identified by optical time domain reflectometry.

[0056] In another form, the fibre is separated from the pipeline and wherein a vibroacoustic coupling member is introduced between the pipeline and the fibre to transfer vibroacoustic energy from the pipeline to the fibre.

[0057] In another form, the vibroacoustic coupling member is a metal rod or bar.

[0058] In another form, the vibroacoustic coupling member is a fibre segment spliced into the fibre.

[0059] In a second aspect, the present disclosure provides a system for determining a condition of a pipeline, comprising: a pressure flow sensing arrangement configmed to measure pressure flow information from the pipeline; and one or more data processors configured to carry out the method of any one of the forms of the first aspect.

[0060] In a third aspect, the present disclosure provides a pipeline condition determining system comprising: one or more processors; and a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, wherein the one or more processors are enabled to implement the method according any one of the forms of the first aspect.

[0061] In a fourth aspect, the present disclosure provides a non-transitory computer-readable medium comprising instructions stored thereon, that when executed on a processor, perform the steps of the method according to any one of the forms of the first aspect.

[0062] In a fifth aspect, the present disclosure provides a system for determining a condition of a pipeline, comprising: a pressure flow sensing arrangement configmed to measure pressure flow information from the pipeline; and one or more data processors configured for: segmenting the pressure flow information to generate a selection of pressure flow information data segments corresponding to at least one or more pipeline locations determined at two or more time periods, each of the pressure flow information data segments corresponding to pressure flow information received for a unique time interval and pipeline location combination; and determine a respective segment condition measure for each of the pressure flow information data segments in the selection of pressure flow information data segments; and determining a pipeline condition measure based on the respective segment condition measures.

[0063] In another form, the pressure flow information corresponds to three or more pipeline locations on the pipeline, and wherein segmenting the pressure flow information comprises generating the selection of pressure flow information data segments to correspond to at least three or more pipeline locations determined at two or more time periods.

[0064] In another form, determining a segment condition measure for the pressure flow information data segment comprises determining a signal intensity measure of the pressure flow information data segment.

[0065] In another form, determining a signal intensity measure comprises determining a root-meansquare (RMS) value for the pressure flow information data segment.

[0066] In another form, determining a signal intensity measure comprises: dividing the pressure flow information data segment into a number of frames; determining a respective RMS value for each frame of the number of frames; and determining a lower bound RMS value for which a selected percentage of the respective RMS values exceed.

[0067] In another form, determining the segment condition measure for the pressure flow information data segment comprises determining a spectral / frequency content measure of the pressure flow information data segment.

[0068] In another form, determining the spectral / frequency content measure comprises determining a power spectrum density (PSD) of the pressure flow information data segment.

[0069] In another form, determining the spectral / frequency content measure comprises determining a spectral heat map of the pressure flow information data segment.

[0070] In another form, determining the spectral content measure comprises determining a median frequency (MF) value for the pressure flow information data segment.

[0071] In another form, determining the spectral content measure comprises: dividing the pressure flow information data segment into a number of frames; determining a respective MF value for each frame of the number of frames; and determining a lower bound MF value for which a selected percentage of the respective MF values exceed.

[0072] In another form, determining the segment condition measure for the pressure flow information data segment comprises processing the pressure flow information data segment by a machine learning model trained to identify a pipeline anomaly condition.

[0073] In another form, the machine learning model is an artificial neural network (ANN) or a recurrent neural network (RNN).

[0074] In another form, processing the pressure flow information data segment by a machine learning model trained to identify a pipeline anomaly condition comprises: extracting one or more time domain and / or frequency domain features from the pressure flow information data segment; and processing the one or more extracted time domain and / or frequency domain features by a feature based machine learning classifier trained to identify the pipeline anomaly condition based on the one or more extracted time domain and / or frequency domain features.

[0075] In another form, the feature based machine learning classifier is a decision tree model or a support vector machine.

[0076] In another form, processing the pressure flow information data segment by a machine learning model trained to identify a pipeline anomaly condition comprises: forming a 2D spectrogram from the pressure flow information data segment; and processing the 2D spectrogram by a machine learning classifier trained to identify the pipeline anomaly condition based on the 2D spectrogram.

[0077] In another form, the machine learning classifier is a convolutional neural network.

[0078] In another form, determining the segment condition measure for the pressure flow information data segment comprises processing the pressure flow information data segment by a statistical model configured to identify a pipeline anomaly condition.

[0079] In another form, processing the pressure flow information data segment by a statistical model configured to identify a pipeline anomaly condition comprises: extracting one or more time domain and / or frequency domain features from the pressure flow information data segment; and processing the one or more extracted time domain and / or frequency domain features by the statistical model configured to receive the one or more extracted time domain and / or frequency domain features as inputs.

[0080] In another form, determining the segment condition measure for the pressure flow information data segment comprises: identifying the associated pipeline location and time period for the pressure flow information data segment; and determining respective coherence measures between the pressure flow information data segment and further pressure flow information data segments from the same or similar time period andcorresponding to two or more of the remaining pipeline locations in the selection of pressure flow information data segments.

[0081] In another form, one or more of the remaining pipeline locations correspond to adjacent pipeline locations to the associated pipeline location of the pressure flow information data segment.

[0082] In another form, one or more of the remaining pipeline locations correspond to non-adjacent pipeline locations to the associated pipeline location of the pressure flow information data segment.

[0083] In another form, determining the pipeline condition measure based on the respective segment condition measures comprises: forming an assembly of segment condition measures ordered with respect to time and pipeline location.

[0084] In another form, determining the pipeline condition measure based on the respective segment condition measures comprises: determining whether one or more of the respective segment condition measures satisfy a threshold condition.

[0085] In another form, determining the pipeline condition measure based on the respective segment condition measures comprises: determining whether two or more of the respective segment condition measures corresponding to the same pipeline location satisfy a temporal rate of change condition.

[0086] In another form, determining whether two or more of the respective segment condition measures corresponding to the same pipeline location satisfy a temporal rate of change condition comprises determining whether a temporal benchmark is exceeded.

[0087] In another form, determining whether two or more of the respective segment condition measures corresponding to the same pipeline location satisfy a temporal rate of change condition comprises determining whether a temporal rate of change of power density in one or more selected frequency bands is exceeded.

[0088] In another form, determining whether two or more of the respective segment condition measures corresponding to the same pipeline location satisfy a temporal rate of change condition comprises determining whether a temporal rate of change of coherence level is exceeded.

[0089] In another form, determining the pipeline condition measure based on the respective segment condition measures comprises: determining whether two or more of the respective segment condition measures corresponding to the same or similar time period satisfy a spatial rate of change condition.

[0090] In another form, determining whether two or more of the respective segment condition measures corresponding to the same or similar time period satisfy a spatial rate of change condition comprises determining whether a spatial benchmark is exceeded.

[0091] In another form, determining whether two or more of the respective segment condition measures corresponding to the same or similar time period satisfy a spatial rate of change condition comprises determining whether a spatial rate of change of power density in one or more selected frequency bands is exceeded.

[0092] In another form, determining whether two or more of the respective segment condition measures corresponding to the same or similar time period satisfy a spatial rate of change condition comprises determining whether a spatial rate of change of coherence level is exceeded.

[0093] In another form, determining the pipeline condition measure based on the respective segment condition measures comprises: determining whether two or more of the respective segment condition measures corresponding to the same or similar time period and two or more of the respective segment condition measures corresponding to the same pipeline location satisfy a combined temporal / spatial rate of change condition.

[0094] In another form, determining the pipeline condition measure based on the respective segment condition measures comprises processing the combined respective segment condition measures by a machine learning model trained to identify a pipeline anomaly condition.

[0095] In another form, processing the combined respective segment condition measures by a machine learning model trained to identify a pipeline anomaly condition comprises the machine learning model operating on an assembly of segment condition measures ordered with respect to time and pipeline location.

[0096] In another form, the method further comprises pre-processing a pressure flow information data segment for a selected pipeline location to remove historical background noise associated with the selected pipeline location.

[0097] In another form, the method further comprises pre-processing a pressure flow information data segment for a selected time period to remove historical background noise associated with the selected time period.

[0098] In another form, the method further comprises benchmarking a pressure flow information data segment for a selected pipeline location against an earlier pressure flow information data segment associated with an earlier time period for the selected pipeline location.

[0099] In another form, the method further comprises pre-processing a pressure flow information data segment for a selected pipeline location in accordance with a coherence analysis to compensate for differences in sensor gain and / or environmental noise.

[0100] In another form, the coherence analysis comprises: processing the pressure flow information data segment to form a PSD; determining the maximum coherence level between the pressure flow information data segment and further pressure flow information data segments from the same or similar time period and corresponding to two or more of the remaining pipeline locations in the selection of pressure flow information data segments; and weighting the PSD with the maximum coherence level.

[0101] In another form, the coherence analysis further comprises: normalising the PSD to form a normalised PSD; forming a filtered PSD by filtering the normalised PSD; and weighting the normalised PSD with the maximum coherence level.

[0102] In another form, the method further comprises pre-processing a pressure flow information data segment for a selected pipeline location in accordance with a beamforming analysis to improve the vibroacoustic signal to noise.

[0103] In another form, the beamforming analysis comprises: determining the phase differences between the pressure flow information data segment and further pressure flow information data segments from the same or similar time period and corresponding to two or more of the remaining pipeline locations in the selection of pressure flow information data segments; phase shifting the further pressure flow information data segments to have the same phase as the pressure flow information data segment to form phase shifted further pressure flow information data segments;weighting the phase shifted further pressure flow information data segments in accordance with their level of coherence with the pressure flow information data segment to form weighted phase shifted further pressure flow information data segments; and forming a beamformed pressure flow information data segment by summing in frequency space the weighted phase shifted further pressure flow information data segments to the pressure flow information data segment.

[0104] In another form, phase shifting the further pressure flow information data segments comprises only phase shifting those pressure flow information data segments having a coherence greater than a predetermined coherence threshold.

[0105] In another form, the pressure flow sensing arrangement comprises one or more individual pressure flow sensors each associated with a corresponding location on the pipeline.

[0106] In another form, the pressure flow sensing arrangement comprises a distributed pressure flow sensing apparatus having one or more sensing regions each associated with a corresponding location on the pipeline.

[0107] In another form, the distributed pressure flow sensing apparatus is an optical fibre and the one or more sensing regions are fibre Bragg gratings located along the optical fibre.

[0108] In another form, the distributed flow sensing apparatus is an optical fibre and the one or more sensing regions correspond to regions located along the fibre identified by optical time domain reflectometry.

[0109] In another form, the fibre is separated from the pipeline and wherein a vibroacoustic coupling member is introduced between the pipeline and the fibre to transfer vibroacoustic energy from the pipeline to the fibre.

[0110] In another form, the vibroacoustic coupling member is a metal rod or bar.

[0111] In another form, the vibroacoustic coupling member is a fibre segment spliced into the fibre.

[0112] In yet another aspect, the present disclosure provides a method, system and non-transitory computer-readable medium for determining a condition of a pipeline, comprising: receiving pressure flow information corresponding to three or more pipeline locations on the pipeline;segmenting the pressure flow information to generate a selection of pressure flow information data segments corresponding to at least three or more pipeline locations determined at two or more time periods, each of the pressure flow information data segments corresponding to pressure flow information received for a unique time interval and pipeline location combination; and determining a pipeline condition measure by processing the selection of pressure flow information data segments.BRIEF DESCRIPTION OF DRAWINGS

[0113] Embodiments of the present disclosure will be discussed with reference to the accompanying drawings wherein:

[0114] FIG. 1A is a figurative view of an example pipeline whose condition may be determined in accordance with the present disclosure;

[0115] FIG. IB is a figurative view of another example pipeline whose condition may be determined in accordance with the present disclosure;

[0116] FIG. 1C is a diagram of an actual water pipeline indicating 10 pipeline locations of interest as indicated by the arrowed indications on “Greenhill Road”;

[0117] FIG. 2 is a flowchart of a method for determining a condition of a pipeline in accordance with an illustrative embodiment;

[0118] FIG. 3 is a system overview diagram of a pipeline condition determining system in accordance with an illustrative embodiment;

[0119] FIG. 4 is a figurative view showing the operating principles of a Fibre Bragg Grating (FBG) based pressure flow sensor;

[0120] FIG. 5 is a figurative view showing where a distributed pressure flow sensing apparatus in the form of an optical fibre may be positioned with respect to a pipeline such as that illustrated in FIG. 1 A;

[0121] FIG. 6 is a figurative view of optical time domain reflectometry (OTDR) based distributed flow pressure sensing apparatus in accordance with an illustrative embodiment;

[0122] FIG.7 is a figurative view of an OTDR based distributed flow sensing apparatus employing a pre-existing optical telecommunications fibre in accordance with an illustrative embodiment;

[0123] FIG. 8 is a figurative diagram illustrating the generation of pressure flow data segments from pressure flow information received from three or more locations according to an illustrative embodiment;

[0124] FIG. 9 is a data structure diagram of an example data structure for a pressure flow data segment information in accordance with an illustrative embodiment;

[0125] FIG. 10 is a flowchart of a method for determining a condition of a pipeline based on a selection of pressure flow information data segments (PFIDS) in accordance with an illustrative embodiment;

[0126] FIG. 11 is a data structure diagram of a pressure flow information data segment following division into a number of frames in accordance with an illustrative embodiment;

[0127] FIG. 12A is a flowchart of a method for processing a pressure flow information data segment by a machine learning model or statistical model to determine a segment condition measure in accordance with an illustrative embodiment;

[0128] FIG 12B is a series of plots corresponding to the median PSD power levels for a number of frequency bands for pressure flow information data segments corresponding to a range of different pipeline locations;

[0129] FIG. 13A is an example stmcture of a Convolutional Neural Network (CNN) model for processing a pressure flow information data segment (or segments) to determine a segment condition measure in accordance with an illustrative embodiment;

[0130] FIG. 13B is a plot of the relationship between the mel frequency and standard frequency;

[0131] FIG. 13C is an example of the convolution process forming part of the CNN model illustrated in FIG. 13A;

[0132] FIG. 13D shows an example of the pooling process forming part of the CNN model illustrated in FIG. 13A;

[0133] FIG. 13E shows example confusion matrices following assessment of the output from a CNN model of the type shown in FIG. 13 A in accordance with an illustrative embodiment;

[0134] FIG. 13F is an example structure of a CNN Siamese Twin architecture 1390 according to an illustrative embodiment;

[0135] FIG. 13G is a series of spectrograms and PSDs (top row), along with time series data and histogram representations (bottom row), corresponding to a pipeline location but for different time periods showing a noise signal that periodically reoccurs;

[0136] FIG. 14A is a PSD heat map following coherence processing to reinforce the detected signal power from a pipeline anomaly in the form of a leak in accordance with an illustrative embodiment;

[0137] FIG. 14B is a 2D representation of the heat map information forming the basis of the PSD heat map shown in FIG. 14A prior to coherence processing;

[0138] FIG. 14C is a 2D representation of the heat map information shown in FIG. 14B following coherence processing in accordance with an illustrative embodiment;

[0139] FIG. 14D is a flowchart of the steps for coherence processing of pressure flow information data segment(s) in accordance with an illustrative embodiment;

[0140] FIG. 15A is a plot of the summed coherent power levels within the coherence processed PSD derived from the pressure flow information data segment pairs illustrated in FIG. 14A;

[0141] FIG. 15B is a table showing the determination of the Spearman rank for the sensor pairs and their summed coherence values for the coherence values shown in FIG. 15 A;

[0142] FIG. 16 is a coherence heat map representation of the coherence level between pressure flow information data segment pairs corresponding to both adjacent pipeline locations and non-adjacent pipeline locations in accordance with an illustrative embodiment;

[0143] FIG. 17 is a coherence heat map representation similar to that illustrated in FIG. 16 showing the effect of incidental or environmental noise in the pressure flow information data segment corresponding to pipeline location 4;

[0144] FIG. 18A is a spectral heat map comprising an assembly of individual spectrograms for a selection of individual pressure flow information data segments covering a range of pipeline locations and time periods in accordance with an illustrative embodiment;

[0145] FIG. 18B is a series of diagrams showing segment condition measures in the form of PSD heat maps covering multiple days, a difference PSD heat map showing the difference between a PSD heat map for a given day and a benchmark PSD heat map and associated pipeline condition measure in accordance with an illustrative embodiment;

[0146] FIG. 18C a series of heat maps 1850 of the Spearman rank corresponding to a vibroacoustic event involving a pipe leak evolving over a period of 60-70 days for window lengths of 20 (plot(a)), 30 (plot(b)), 40 (plot(c)) and 50 (plot(d)) days in accordance with an illustrative embodiment;

[0147] FIG. 18D shows plots of the maximum r value (1860a) and the corresponding WL (1860b) for an example vibroacoustic event involving a pipe leak evolving over a period of 60-70 days in accordance with an illustrative embodiment;

[0148] FIG. 18E shows two example clusterings of frequency bins having increased PSD power levels in accordance with an illustrative embodiment;

[0149] FIG. 18F shows a table of some example parameter values for use in determining an anomaly condition measure based on increasing PSD over time in accordance with an illustrative embodiment;

[0150] FIG. 19A is an example stmcture of a CNN model for determining whether two or more respective segment condition measures corresponding to the same pipeline location satisfy a temporal rate of change condition in accordance with an illustrative embodiment;

[0151] FIG. 19B is a table showing the various convolution, pooling and activation stages of the CNN model illustrated in FIG. 19A;

[0152] FIG. 20 is a plot of the pressure flow information signal as a function of distance from a source of vibroacoustic energy in the form of a leak;

[0153] FIG. 21 is a plot of the results of a time based “Rate of Change” (ROC) pipeline condition measure applied to pressure flow information data segments from the 10 sensor configurations illustrated in FIG. 1C provided daily for an extended period in accordance with an illustrative embodiment;

[0154] FIG. 22 is a heatmap representation of coherence levels where the individual segment condition measures are the coherence levels for pressure flow data segments corresponding to adjacent pipeline locations as determined daily in accordance with an illustrative embodiment;

[0155] FIG. 23 is an example structure of a CNN model showing the processing of the selection of segment condition measures (eg, input spectrograms) in combination by a machine learning model trained to identify a pipeline anomaly condition according to an illustrative embodiment;

[0156] FIG. 24 A is an example stmcture of a CNN model similar to FIG. 23 but in this case the selection of segment condition measures in combination are in the form of the 2D representation shown in FIG.22;

[0157] FIG. 24B is a flowchart of a method for determining a condition of a pipeline in accordance with another illustrative embodiment;

[0158] FIG. 25 is a flowchart of a method for pre-processing a pressure flow information data segment to remove historical background noise data in accordance with an illustrative embodiment;

[0159] FIG. 26 is a flowchart of a method for determining a spectral content difference in accordance with an illustrative embodiment;

[0160] FIG. 27 is a flowchart of a method for removing the spectral content difference from the pressure flow information data segment in accordance with an illustrative embodiment;

[0161] FIG. 28 is a flowchart of a method for benchmarking a pressure flow information data segment against an earlier pressure flow information data segment corresponding to an earlier period in accordance with an illustrative embodiment;

[0162] FIG. 29 is a diagram showing a series of pressure flow sensors that are spaced apart from one another and offset from a pipeline in accordance with an illustrative embodiment;

[0163] FIG. 30A shows a streetscape 3000 showing the configuration of a pipeline and a distributed pressure flow sensing apparatus in the form of a fibre and also individual pressure flow sensors in the form of accelerometers in accordance with an illustrative embodiment;

[0164] FIG. 30B shows an experimental pipeline condition determining system comprising a distributed flow sensing apparatus in accordance with an illustrative embodiment;

[0165] FIG. 31A is a plot of the PSD following the beamforming process (“Beamformed”) applied with respect to the pipeline condition determining system shown in FIG. 30B and having a leak of 0.07 L / s as indicated and adopting pressure flow sensor #1 as the reference;

[0166] FIG. 3 IB is a plot of the PSD following the beamforming process (“Beamformed”) applied with respect to the pipeline condition determining system shown in FIG. 30B and having a leak of 0.07 L / s as indicated and adopting pressure flow sensor #6 as the reference;

[0167] FIG. 32 is a data processing diagram of a method for isolating the coherent energy from a source and reject noise that is not coherent for a three sensor arrangement;

[0168] FIG. 33 is a plot of the determined correlated energy for a selected pressure flow sensor based on the requirement that it be also correlated to adjacent pressure flow sensors for the pipeline condition determining system illustrated in FIG. 3 OB in accordance with an illustrative embodiment;

[0169] FIG. 34 is a plan view of a spatial arrangement of vibroacoustic coupling members located between a pipeline and an adjacent non-parallel fibre based distributed flow sensing apparatus in accordance with an illustrative embodiment;

[0170] FIG. 35 is a series of plots showing the relative frequency dependent attenuation (acoustic power loss) along a pipeline, a vibroacoustic coupling member and through soil in accordance with an illustrative embodiment;

[0171] FIG. 36 is a flowchart of a method for determining the occurrence and location of a leak along the pipeline adopting one or more vibroacoustic coupling members according to an illustrative embodiment;

[0172] FIG. 37 comprises plots showing the original PSDs for first and second pipeline coupling locations and the respective PSDs of the transferred vibroacoustic signals as well as a plot of the coherence level between the transferred vibroacoustic signals and a plot of their phase relationship as determined in accordance with the method illustrated in FIG. 36; and

[0173] FIG. 38 is a plot 3800 of the correlation level between the transferred vibroacoustic signals as measured by the fibre based distributed pressure flow information sensing apparatus following energy transfer by respective acoustic coupling members in accordance with an illustrative embodiment.

[0174] In the following description, like reference characters designate like or corresponding parts throughout the figures.DESCRIPTION OF EMBODIMENTS

[0175] Referring now to FIG. 1A, there is shown a figurative view of a pipeline 100 configured to supply fluid in the direction indicated by the arrow whose condition may be assessed in accordance with methods and systems described in the present disclosure. In one example, pipeline 100 may be a water transmission pipeline operable to convey water over large distances such as between a water source (eg, a reservoir or river) to an urban centre. Indicated on pipeline 100 are three indicative pipeline locations 111, 112 and 113.

[0176] Referring now to FIG. IB, there is shown a figurative view of a pipeline 150 which may also be assessed with the methods and systems described in the present disclosure. As can be seen, and differing to pipeline 100, pipeline 150 in this example is configured to supply fluid to multiple receiving locations. In this example, pipeline 150 consists of a pipe network 160 designed to deliver water to receiving locations such as buildings and homes and the like and where the pipe network 160 is located below ground and generally aligns with the road and street system for traffic. As shown, there may be many different pipeline locations of interest in pipeline 150 and three are indicated, these being pipeline locations 161, 162 and 163.

[0177] Referring now to FIG. 1C, there is shown a diagram of an actual water pipeline 180 indicating 10 pipeline locations of interest 183 A-183 J as indicated by the arrowed indications on “Greenhill Road” 187 which is a main road in Adelaide, South Australia.

[0178] As would be appreciated, while the present disclosure is described in the context of utility scale pipelines for the supply of water, the methods and systems disclosed may also be applied to any pipeline which is designed to supply fluid to a single or multiple receiving locations. Examples include, but are not limited to, petroleum refinery, storage and supply, processing factories, wine making and storage, minerals processing, gas supply networks, hot water heating supply networks or airport fuel systems.

[0179] Referring now to FIG. 2, there is shown a flowchart 200 of a method for determining the condition of a pipeline according to an illustrative embodiment. By way of overview, method 200 comprises at step 210 receiving pressure flow information corresponding to three or more locations (eg, 111, 112 and 113 illustrated in FIG 1A; 161, 162 and 163 illustrated in FIG. IB; or 183A-183J illustrated in FIG. 1C respectively) on a pipeline where the pressure information is related to the fluid flow characteristics of a fluid being conveyed by the pipeline at the given location.

[0180] At step 220, the pressure flow information is segmented to generate a selection of pressure flow information data segments corresponding to at least three or more pipeline locations determined at two or more time periods where each of these segments corresponds to a unique time interval or period (eg, between a respective initial time t0and t0+ At for each segment) and pipeline location combination. At step 230, the pipeline condition measure is determined (eg, such as for the presence of a structural anomaly in the pipeline) by processing the selection of pressure flow information data segments corresponding to at least three or more pipeline locations determined at two or more time periods.

[0181] Throughout the specification the term “pressure flow information” is defined to mean any measured dynamic or time varying signal data related to the flow of fluid in the pipeline measured by a pressure flow sensor. As would be appreciated, pressure flow information may result, in some examples,from direct measurement of fluid flow and / or alternatively, in other examples, from the measurement of vibroacoustic energy generated by the fluid flow in the pipeline.

[0182] In various examples, vibroacoustic energy may arise from a number of sources including, but not limited to:• the flow of the fluid through the pipeline or pipeline network including the interaction of the liquid with various components of the pipeline, eg, pipe wall, valves, distribution points, water meters etc; or• the interaction of fluid flowing through the pipeline network with any structural anomaly such as a crack / leak event resulting in additional flows of fluid departing the network; or• other sources of acoustic energy eg, general environmental noise such as traffic, plant equipment, night clubs, tram operation, bus operation, train operation, functions, excavation work, rainfall (or runoff from it), construction site activities, concerts, car races and industrial water use; or• the response of the pipeline or pipeline network to an transient pressure wave introduced into the pipeline to probe pipeline condition.

[0183] As would be appreciated, the use of terms such as fluid-borne, ground-borne or air-borne as used throughout the specification indicate that the respective pressure flow information or pressure flow sensor is configured to measure predominantly fluid-borne, ground-borne or air-borne vibroacoustic energy respectively. It is the nature of these vibroacoustic measuring systems that the components of the system will be coupled physically to measure vibroacoustic energy and in that sense a pressure flow sensor configured to measure primarily fluid-borne vibroacoustic energy may also sense other forms of vibroacoustic energy transmitted through the ground, air, pipe wall and the like.

[0184] Referring also to FIG. 3, there is shown a system view of a pipeline condition determining system 300 operable in one example to implement a pipeline condition determining method 200. Again by way of overview, pipeline condition determining system 300 includes a data processor 320 configured to receive pressure flow information 310 corresponding to three or more locations on the pipeline indicated as Li, L2, L3, •, LN in FIG. 3 (eg, corresponding to pipeline locations 111, 112 and 113 or 161, 162 and 163 illustrated inFIGs 1A and IB respectively) and operable to segment the pressure flow information as sensed in this example by a pressure flow sensing arrangement 350. This generates a selection of pressure flow information data segments corresponding to at least three or more pipeline locations determined at two or more time periods where each of the pressure flow information data segments corresponds to pressure flow information received for a unique time interval and pipeline location combination.

[0185] As an example, a particular pressure flow information data segment may correspond to pressure flow information corresponding to pipeline location L2and for a time interval defined between t0and t0+ At where another pressure flow information data segment may correspond to pressure flow information corresponding to pipeline location L3and for time interval defined between t0+ At and t0+ 2At.

[0186] The selection of these pressure flow information data segments corresponding to at least three or more pipeline locations determined at two or more time periods is then processed to determine the pipeline condition measure. In one example, the results may be displayed on optional display 340 and input may be provided by optional user input 330 as required.

[0187] As will be become apparent in the discussion below, the pressure flow information that corresponds to a given location need not be generated by a pressure flow sensing arrangement involving sensors or sensing regions located at the respective location on the pipeline but could be generated by a pressure flow sensing arrangement where three or more of the sensors or sensing regions are located at a distance from the pipeline but which otherwise will be still associated with a given pipeline location.

[0188] Referring again to FIG. 2, at step 210 in one example the pressure flow information may be received from a pressure flow sensing arrangement 350 comprising three or more individual standalone pressure flow sensors each associated with a corresponding location on the pipeline. As would be appreciated, the individual pressure flow sensors may be located within the pipeline, connected to the external wall of the pipeline, or even located at a distance from the pipeline but still are able to sense pressure flow information through any intervening media between the sensor and the pipeline.

[0189] In various examples, the standalone pressure flow sensor is an acoustic vibrometer or accelerometer. In another example, the pressure flow sensor is an acoustic hydrophone. In yet another example, the standalone pressure flow sensor is a pressure transducer. In a further example, the standalone pressure sensor may be an optical sensor and in one example the optical sensor is a Fibre Bragg Grating (FBG) based pressure flow sensor comprising a light source and an optical fibre having a FBG reflection region.

[0190] Referring now to FIG. 4, there is shown a figurative view showing the operating principles of a FBG based pressure flow sensor 400. In this example, sensor 400 is formed from a fibre 410 comprising a FBG reflection region 420 in the fibre core 430 formed by a periodic variation in refractive index 470 where sensor 400 is configured so that an external vibration functions to induce a variable strain in the FBG reflection region 420. This results in modifying or shifting the Bragg wavelength 480 of the grating in accordance with the external vibration. This modulation of the Bragg wavelength 480 may then be detected in either the reflection spectrum 492 or transmission spectrum 491 following input of a broadspectrum input light source 490 to provide pressure flow information originating from the FBG reflection region 420.

[0191] As an example, consider a structural anomaly in the form of a pipeline leak where water is flowing from the pipeline through an aperture in the pipeline. In one example, the vibroacoustic energy from the leak could propagate through the pipeline wall to a directly attached FBG based pressure flow sensor where in this case the FBG based pressure flow sensor is acting as a form of accelerometer at the attachment location on the pipeline. In another example, the FBG based pressure flow sensor may be installed within the fluid being conveyed by the pipeline at a particular installation location and in this case the FBG based pressure sensor is functioning as a form of hydrophone sensor at the location of the FBG sensor. In other examples, the FBG based pressure sensor may be located remote from, but be associated with, a corresponding location on the pipeline and pressure flow information in the form of ground-borne vibroacoustic energy may be sensed through the soil or other intervening media by the fibre based pressure flow sensor.

[0192] In one example, the pressure flow sensor will comprise a physical sensing unit and an electronic data acquisition (DAQ) module operable to convert the output from the physical sensing unit to a data signal which may comprise magnitude and / or frequency data attributable to vibroacoustic energy that constitutes the pressure flow information.

[0193] In one example, the DAQ simultaneously receives and applies an appropriate time stamp to the data signal forming the pressure flow information allowing the pressure flow information from respective pressure flow sensors to be synchronised for later processing. In one example, the pressure flow sensor may incorporate an appropriate timer synchronised to the other pressure flow sensors and the locally synchronised data may be transferred without requiring further timestamping. In one example, one or more of the pressure flow sensors may include a GPS module capable of providing this time synchronisation capability.

[0194] In various examples, the pressure flow information may originate from the pressure flow sensors sampling the environment at different rates. Exemplary sampling rate ranges include, but are not limited to, greater than 0kHz and less than 1kHz, 1kHz - 2kHz, 2kHz - 3kHz, 3kHz - 4kHz, 4kHz - 5kHz, 5kHz - 6kHz, 6kHz - 7kHz, 7kHz - 8kHz, 8kHz - 9kHz, 9kHz - 10kHz, or greater than 10kHz. In another example embodiment, one or more of the pressure flow sensors may be configmed to sample synchronously, ie, at the same frequency.

[0195] In some examples, the pressure flow sensor may be configured to sense or store the magnitude and / or frequency data constituting the pressure flow information commencing at a predetermined time for a predetermined time period and either store the pressure flow information locally at the respective sensoror DAQ module and / or transmit this data to a remote location (eg, server or cloud storage) by wireless or other means as will be discussed below. In another example, one or more of the pressure flow sensors may be configured to buffer data locally and then transmit this data for further storage and processing.

[0196] As would be appreciated, the pressure flow information may be transmitted by the pressure flow sensors using any known data transfer arrangements, including, but not limited to, radio, cell phone (GSM) and / or other physical data communication systems. In various examples, the pressure flow information may be transmitted at varying rates from, for example, once per day to, for example, every five minutes, or continuously, depending on the type of data collected and whether threshold conditions have been met requiring the transmission of data from a pressure flow sensor.

[0197] In another embodiment, pressure flow information is received from a pressure flow sensing arrangement 350 comprising a distributed pressure flow sensing apparatus that covers an extended distance and which comprises three or more pressure flow sensing regions each associated with a corresponding location on the pipeline. In one example, the distributed flow sensing apparatus is an optical fibre incorporating associated sensing regions.

[0198] Referring now to FIG. 5, there is shown a figurative view showing where a distributed pressure flow sensing apparatus 500 in the form of an optical fibre may be positioned with respect to a pipeline such as that illustrated in FIG. 1 A. In scenario A, the optical fibre 500 may be positioned offset and external to the pipeline 100. In scenario B, the optical fibre 500 is attached to an external (or internal) wall of the pipeline 100. In scenario C, the optical fibre 500 resides within the fluid being conveyed by the pipeline 100.

[0199] In one example, an optical fibre may have discrete pressure flow sensing regions corresponding to locations on the pipeline such as a FBG that measures pressure flow information at the respective location by varying a reflection wavelength of FBG as described above to form a FBG based distributed pressure flow sensing apparatus. In this example, each spatially distributed FBG reflection region along the fibre will have its own respective Bragg wavelength that is frequency separated from the Bragg wavelengths of the other reflection regions. Following illumination by a broadband light source directed along the fibre, the variation in the respective Bragg wavelengths, each corresponding to its own FBG reflection region and associated location, may be processed by a FBG customised interrogator to provide pressure flow information corresponding to each FBG reflection region.

[0200] In one example, the Bragg wavelengths of the reflection regions are selected to be within the wavelength range of 1500nm - 1600nm with a wavelength separation between adjacent reflection regions in the order of 2-5nm.

[0201] Referring now to FIG. 6, there is shown an optical time domain reflectometry (OTDR) based distributed pressure flow sensing apparatus 600 according to an illustrative embodiment. In this example, OTDR based apparatus 600 comprises a fibre 610, a signal generator 620 and an optical signal detector 630 where the signal generator 620 generates a series of light pulses 625 along optical fibre 610 that in turn produces an optical reflection signal 635 which is collected by optical signal detector 630 such as a photodetector where the signal is amplified and digitised. The presence of vibroacoustic energy 650 at a particular location will cause a modification of the detected optical reflection signal 635 whose intensity and location may be determined using OTDR.

[0202] The light pulses 625 mentioned above are in one example a continuous series of pulses in the time domain. These pulses 625 are subject to Rayleigh backscattering from inherent manufacturing variations distributed along the length of the optical fibre which form scattering centres along the fibre generating respective optical reflection signals 635. The power of Rayleigh backscattering from these scattering centres along the fibre may be measured as described above. The backscattering remains coherent for relatively short distances for broadband input but for longer distances for narrow band input. The pulse length can be set such that it is longer than the coherence distance and thus allow for faults relating to the fibre itself to be detected (at specific backscattering locations).

[0203] The cell resolution length for the fibre is a function of the pulse rate and time to travel along the length of fibre being sensed at the speed of light - with each pulse needing to travel out and back along the fibre before the next pulse. A typical pulse would be 20-100 ns long with a maximum cell resolution length of approximately 1-2 m (although this depends on the length of fibre being sensed with long sensing lengths requiring longer cells). The backscattered power level from a cell is proportional to the cell size (and hence pulse length) and In turn the number of cells versus length of fibre being sensed.

[0204] Typical sensing rates of approximately 2 kHz may be achieved with attenuation of high frequency energy pressure flow information such as arising from a leak in a pipeline leading to typical observed frequencies in the order of 10-500 Hz (below the Nyquist frequency for the sensing rate).

[0205] Phase-OTDR (or coherent-OTDR (COTDR)) is a form of OTDR that employs a coherence length that is greater than the length of the resolution cell along a fibre. A coherent narrow-band input (single-frequency wave) is injected into the fibre which enables it to act as a form of distributed interferometer. This enables the power from multiple backscattering locations along the fibre to be aggregated (based on same phase) and the detection of external energy sources (eg, the presence of vibroacoustic energy 650) affecting greater lengths of fibre. All the reflections within the resolution cell are coherent, which means the phase differences between the waves do not change in different repeated tests as long as the injected waves are coherent and stable and the properties of the fibre do not change.

[0206] Reflected waves have reversed phases (backscattering from inherent imperfections along the fibre), and if they all have the same phase, the energy from the reflected waves can be amplified. Compared with standard OTDR techniques, the reflection signal will generally be noisier but also more sensitive to any changes that affect the normal backscattering levels along the fibre.

[0207] External sources of vibroacoustic energy at a particular location will change the normal backscattering levels from locations along the fibre at the point of vibration by inducing variable strain in the fibre which was not present without the energy from the vibration (eg, as an example this vibroacoustic energy could be caused by the propagation of leak energy either within a pipe to an internal fibre (ie, scenario “C” illustrated in FIG. 6), through a pipe wall to an attached fibre (ie, scenario “B” illustrated in FIG. 6) and / or through soil or other media to a fibre spatially offset from a leaking pipe (ie, scenario “A” illustrated in FIG. 6).

[0208] In one example, comparison with a reference signal from the fibre (with backscatter for the case without any external source of vibroacoustic energy) enables detected new power within a cell (or equivalently a pressure flow sensing region), corresponding to a pipeline location near a pipeline anomaly, such as pipe leak, to be identified.

[0209] COTDR sensing ranges are typically in the order of 50 kms before the input pulses 625 are attenuated to a level that prevents a clear signal from a local source of vibration noting that additional losses may occur from splicing of the fibre resulting from branches from, or connections to, the main fibre. While the magnitude of the vibration level being sensed is relevant, as well as the input injected wave power, a range of typically 50 kms remains applicable.

[0210] Referring now to FIG. 7, there is shown a figurative view of an OTDR based distributed flow sensing apparatus 700 employing a pre-existing optical telecommunications fibre 730 according to an illustrative embodiment positioned with respect to the pipe network 160 of pipeline 150 illustrated in FIG. IB. As referred to above, the pipe network 160 will generally align with the road and street system of an urban area being serviced by pipe network 160 and in many cases the fibre network will also generally align with the road and street system of an urban area. In the example depicted in FIG. 7, optical telecommunications fibre 700 is located proximal to pipeline locations 161, 162 and 163.

[0211] In many instances, an urban fibre network will also include spare or “dark” fibre components for redundancy purposes that form part of the fibre network but which is otherwise not utilised for telecommunication traffic. In these circumstances, a dark fibre (eg, fibre 730) may be employed as a OTDR based distributed flow sensing apparatus using optical time domain reflectometry (OTDR) techniques as discussed above where a pressure flow signal may be determined corresponding to pipelinelocations 161, 162 and 163 (in this example) by suitable analysis of the optical reflection signal from optical telecommunications fibre 730.

[0212] The use of OTDR fibre based distributed pressure flow sensing apparatus based on preexisting spare (or dark) optical telecommunication fibre networks is particularly advantageous for urban environments where it may be difficult to position standalone or distributed pressure flow sensors with respect to pipeline infrastructure or such sensors may be susceptible to interference from environmental factors (eg, electrical interference). In this case, a pre-existing optical telecommunications fibre may be used to provide pressure flow information in relation to locations on a pipeline that are within the sensing range of the telecommunications fibre.

[0213] As would be appreciated, in one example the distributed pressure flow sensing apparatus may further comprise an electronic data acquisition (DAQ) capability operable to convert the output from the individual sensing regions to a data signal which may comprise magnitude and / or frequency data attributable to vibroacoustic energy that constitutes the pressure flow information where the DAQ capability incorporates one or more of the electronic sampling, storage, communication and / or processing capabilities described above with respect to a standalone pressure flow sensor.

[0214] Longer pulse lengths (lower pulse rates) result in more back scattered energy (and power) returning to the source because the energy under each pulse interacting with the fibre and back scattering is greater. Hence the amount of back scatter is proportional to the total energy in each pulse and not the peak power level of the pulse. The pulsed source laser and the power intensity back scattering from the fibre to the source result in pressure flow information data in the general form of a power versus time data series which is sensed and recorded at the source.

[0215] In one example, the power of the backscattered and / or reflected optical energy recorded at the source can be recorded in 16 -bit or 32-bit integer format as strings of data versus time of collection (which also corresponds to the distance along the fibre from which the returning power / energy is measured at the source). The returning pulses (or waves) are arriving at uniformly spaced time intervals but with varying power levels (due to varying backscattering and / or varying acoustic excitation (vibration) of the fibre at unique distances from the source).

[0216] In one convention, the maximum expected power level is used as a zero for the recording such that all waveforms that backscatter (or also reflect) from imperfections or specific points of vibroacoustic excitation along the fibre, and also reflections from specific physical elements or features along the fibre (eg, splices), are presented as zero (maximum) or otherwise negative dB levels (which is analogous to other signal dB reporting conventions (eg, for GSM communications). The result is an optical powerversus time measurement at the source containing the backscattering information (natural or including specific vibroacoustic sources that vibrate the fibre at specific locations).

[0217] This is a form of vibroacoustic response signal from the fibre and is analogous conceptually to other forms of representation of vibroacoustic response signals over time that provide pressure flow information data. For example, an acoustic accelerometer (or vibrometer) also presents vibroacoustic power levels versus time after recording (in typically 8-bit or 16-bit format). The source of the measured data for an accelerometer, for example, is a uniaxial (single axis) IEPE (Integrated Electronics Piezo- Electric) type device with a typical sensitivity of 1000 mV / g (millivolts per g of acceleration). The piezoelectric resistance of the accelerometer is sensitive to the level of vibroacoustic vibration it is subject to. The optical counterpart is the level of Rayleigh backscattering from a narrow laser pulsed source being sensitive to the level of vibroacoustic vibration that sections of the fibre are subject to.

[0218] Once a measurement of the vibroacoustic intensity experienced by a fibre is obtained in dB (power) versus time convention it is susceptible to the derivation of spectrograms, power density spectrums (PSDs) and coherence analysis using the methods mathematically described throughout the present disclosure including, but not limited to, the use of RMS power level determination, fast Fourier transforms, discrete fast Fourier transforms, coherence determination for two or many more spatially distributed channels, spectral differencing, beamforming (signal reinforcement), coherent signal noise reduction and others). The time domain OTDR data may also be combined over time and space intervals before analysis.

[0219] In all the above pressure flow sensing arrangements, pressure flow information corresponding to three or more locations on the pipeline may be provided. As discussed above, this may involve the use of individual discrete pressure flow sensors or the use of a distributed fibre based sensing apparatus that involves discrete sensing regions (eg, FBG reflection regions) or the use of standard telecommunication fibres and OTDR based signal processing techniques to determine pressure flow information at sensing regions along the fibre.

[0220] Referring to FIG. 2, at step 220, the pressure flow information is segmented to generate a selection of pressure flow information data segments corresponding to at least three or more pipeline locations determined at two or more time periods where each pressure flow information segment that corresponds to pressure flow information is received for a particular pipeline location and for a particular time interval. Referring now to FIG. 8 there is shown a figurative diagram 800 illustrating the generation of the selection of pressure flow information data segments 820 from pressure flow information 810 received from three or more locations LI, L2 and L3 according to an illustrative embodiment.

[0221] In this example, pressure flow information 810 comprises three individual pressure flow information signals 811, 812 and 813 corresponding to three locations on a pipeline designated as Li, L2 and L3. In this example, the segmenting process involves separating each of the pressure flow signals 811, 812 and 813 into two time intervals, ie,and At2, in this manner forming six pressure flow information data segments 821, 822, 823, 824, 825 and 826 which are designated by the unique time interval and pipeline location combinations (At^ L , (At2, Lj). (At L2), (At2, L2), (At L3) and (At2, L3).

[0222] While in this example, the defined time intervals Atxand At2were used for each of the pressure flow signals 811, 812 and 813 it would be understood that different time intervals may be used for each of the pressure flow signals. Additionally, while in this case the pressure flow signal is seen as occupying the entire respective time interval in other examples, the pressure flow signal may only occupy a portion of the time interval, ie, where there are time gaps in the pressure flow signal arising from a corresponding location on the pipeline. Additionally, the time intervals used for a given pipeline location may not necessarily be continuous with each other or indeed continuous with one or more time intervals used when forming pressure flow information data segments relating to other pipeline locations.

[0223] As can be seen, the pressure flow information data segments will each contain pressure flow information corresponding to a unique time interval and pipeline location combination, ie, the combination is unique. Accordingly, for a given time period the pressure flow information data segments relating to that time period, but for different pipeline locations, may be chosen to characterise the pipeline spatially over a distance spanned by the different pipeline locations at the given time period. Similarly, for a given pipeline location the pressure flow information data segments may be chosen and ordered in time to characterise the pipeline over an extended time period for the given pipeline location.

[0224] The pressure flow information and consequently the pressure flow information data segments, S, may have any suitable data format. Referring now to FIG. 9, there is a data structure diagram of an example data structure 900 for pressure flow information data segments S' in accordance with an illustrative embodiment. In this embodiment, segment S is in the form of an acoustic wave file in UINT8 data (d) type (unsigned integer stored with 8-bits, having a minimum possible value of 0 and a maximum possible value of 255). It will be appreciated that other file types and indeed other data formats may be used.

[0225] In one example, the wave file corresponding to the pressure flow information data segments has a sampling rate of 4681Hz and an approximate duration (foe®) of 10 seconds resulting in (A) 46,786 data points (dx).

[0226] In one example, the pressure flow information data segment is normalised. Taking the example of the pressure flow information data segment being the in the form of UINT8 data, this datamay be converted into double precision values and then normalised so as to have a minimum possible value of -1 and a maximum possible valve of +1. Normalisation normalises the sensed pressure flow information relative to a maximum possible dynamic range of the pressure flow sensor. For example, if the output of a pressure flow sensor has a minimum possible value of Xiim minand a maximum possible value of Xiim max, normalisation could be performed using Equation 1 as follows:Equation 1

[0227] where x is the original pressure flow information measurement and x is the normalised data with a minimum possible value of -1 and a maximum possible value of 1.

[0228] In one example, the pressure flow information data segment may be fdtered such as by a high- pass filter having a cut-off frequency and an in-built anti-aliasing filter. In an embodiment, a high-pass filter having a cut-off frequency of 30 Hz may be adopted. However, it will be appreciated that the type and configuration of the filter may vary according to implementation considerations.

[0229] As would be appreciated, any normalisation and filtering process may be applied initially to the original pressure flow information or there may be a combination where some pre-processing steps may occur prior to the segmentation process and others will occur after segmentation.

[0230] Referring back to FIG. 2, at step 230 the selection of pressure flow information data segments corresponding to at least three or more pipeline locations determined at two or more time periods are processed to determine a pipeline condition measure.

[0231] Referring now to FIG. 10, there is shown a flowchart of a method 1000 for determining the pipeline condition measure based on the pressure information data segments according to an illustrative embodiment.

[0232] At step 1010 a respective segment condition measure is determined for each of the pressure flow information data segments that have been selected.

[0233] In one example, determining a segment condition measure for a pressure flow information data segment comprises determining a signal intensity measure of the pressure flow information data segment.

[0234] In one example, determining a signal intensity measure comprises determining a RMS value for the pressure flow information data segment in accordance with Equation 2 as follows:Equation 2

[0235] where N is the total number of data points within a data frame a, and Xi is the value of thedata point of the pressure flow information data segment.

[0236] In other examples, the signal intensity measure may include, but not be limited to, median amplitude values, interquartile ranges or persistence values derived from the pressure flow information data segment.

[0237] In one example, determining a signal intensity measure comprises dividing the pressure flow information data segment into a number of frames, determining a respective RMS value for each frame in the number of frames and then determining a lower bound RMS value for which a selected percentage of the respective RMS values exceed.

[0238] Referring now to FIG. 11, there is shown a data structure diagram of a pressure flow information data segment 1100 following division into a number of frames according to an illustrative embodiment. In this example, the frame division or windowing process involves the pressure flow information data segment S' being divided into a number of time-domain data frames {ai, a2, ... a„}. Data frames may be formed as overlapping (ie, with adjacent frames overlapping by an overlapping ratio) or non-overlapping frames. In one particular example, and as is shown in FIG. 11, the set of data frames A={aj, a2, ... a„} is formed using a sliding rectangular window having a data length of 256 (ie, each window has a duration of approximately 55 ms for a sampling rate of 4681 Hz) with an overlapping ratio of 80%.

[0239] A respective RMS value for each frame is then determined for each of the data frames of the set A={aj, a2, ... a„} to provide a corresponding number of RMS values, ie, {RMSi, RMS2, ... RMSn}. The RMS values may then be ranked from lowest to highest to determine a lower bound RMS value for which a selected percentage of the respective RMS values exceed. As an example, the lower bound RMS value for which 90% of the RMS values exceeds may be referred to as the “L90 RMS value” indicator meaning that 90% of the RMS values are higher than the L90 RMS value. In other words, in some embodiments, other LXX values (where XX is a two-digit number) may be extracted, such as L95, L85, L80 or L75 as further non-limiting examples. Other LXX values may be determined.

[0240] In certain cases, at least some of the data frames of the set A.={ai, a2, ... a„} will include data which has been contaminated by vibroacoustic energy from sources not relating to the pipeline condition (eg, leaks or cracks in the pipeline) such as traffic noise, electrical noise, mechanical device noise, watermeter ticking noise or other sources of incoherent noise. In a particular example, data frames having features which may indicate that the data frame includes data which has been significantly contaminated by other non-pipeline condition related sources may be attenuated or removed prior to further processing by any suitable technique.

[0241] One example of a suitable technique includes setting a signal threshold value such that any data frames having a peak signal value which exceeds the signal threshold value are taken to have been contaminated by incoherent noise and are removed before any following analysis. For example, a signal threshold value may be selected as 0.95 (for normalised data that has a possible maximum of 1). It will of course be appreciated that other signal threshold values may be used. In this respect, lowering the threshold for the peak-based pre-processing may remove more contaminated data frames. However, this may also increase the risk of removing useful data.

[0242] In another example, a suitable technique for attenuating or removing data frames having features which may indicate that the data frame includes data which has been contaminated may be based on the respective RMS value associated with the data frame. In this example, data frames having an RMS value which exceed an RMS threshold value are removed from the set of data frames A = {ai, 02, as, ... , a„} to thereby provide a subset of data frames B. It will be appreciated that this will also affect the LXX RMS value that may be calculated.

[0243] As would be appreciated, the above framing process may be applied to other signal intensity measures as referred to above.

[0244] In one example, determining the segment condition measure for the pressure flow information data segment comprises determining a spectral / frequency content measure of the pressure flow information data segment.

[0245] In one example, determining the spectral / frequency content measure comprises determining the power spectrum density (PSD) of the pressure flow information data segment. As would be appreciated converting a time-domain data signal into a frequency domain signal data representing a PSD involves first taking of a fast Fourier transform of the discrete time based data xn(n = 0, ... , N - 1) where N is the total length (data points) of the data which is defined by Equation 3 : Equation 3

[0246] where Pk(k = 0, ... , N - 1) are the transformed results in the frequency domain and j is the imaginary unit.

[0247] In a particular variation, the discrete fast Fourier transform (DFFT) is determined for the data xn(n = 0, ... , N - 1) which may also be a data series that is partitioned into K segments or batches with M data points in each segment (and optionally S data points for overlapping segments). For each segment k=l, , K the discrete fast Fourier transform (DFFT) is determined (eg, for non-overlapping segments) as: Equation 4

[0248] where / = i / M and a>m= a window function

[0249] The power in each segment is then determined using the obtained DFFT as: as follows: Equation 5

[0250] where W = m=o2(m) for the window function.

[0251] The average of the power level values in each segment is then determined to give the Welch’s method estimate of the PSD as:Equation 6

[0252] The PSD is accordingly determined by squaring each FFT (or DFFT) value (squared value of each frequency data point) to obtain power values for each point for each frequency bin (window) used in the determination of the FFT and an average for any frame size applied that is greater than the frequency bin width.

[0253] The power values (squared FFT frequency data point values) are typically normalised by dividing by the signal recording rate in Hertz (sampling frequency Hz) such that the final representation of and units for the PSD values are in decibels (power) per Hertz (Hz) or dB / Hz and vary with frequency over the sampling frequency range for the data. The above method is modified in cases where there is significant environmental noise such that the process of determining the average of the power level values in each segment is modified such that a median value is determined (rather than average value).

[0254] In one example, the PSD may be separated into distinct frequency bands and the power densities of all the frequency bins that comprise a frequency band may be combined to represent the power density in that frequency band. In this manner, the segment condition measure may comprise a series of values corresponding to the individual power densities of the different frequency bands.

[0255] In one example, determining the spectral content measure comprises determining a median frequency (MF) for the pressure flow information data segment. In one example, determining the MF for the pressure flow information data segment comprises determining the frequency bin M that divides the PSD into halves in accordance with Equation 7 as follows: Equation 7

[0256] where the MF is the frequency corresponding to the MA frequency bin. In one example, the MF may be normalised by dividing by the Nyquist frequency, which is half of the sampling frequency.

[0257] In one example, the pressure flow information data segment may be initially windowed by applying a suitable tapering window function to the pressure flow information segment data to enhance frequency resolution and reduce spectral leakage prior to applying a fast Fourier transform (FFT) in accordance with Equation 2.

[0258] One example of a suitable tapering window function is a Hanning window defined by Equation 8 as follows:2irn w[n] = 0.54 - 0.46cos (— — -) Equation 8

[0259] where n is an integer from 0 to A - 1, IV is the total length (data points) of the window, w[n] is the nth value in the window and where the window is applied by multiplying the window function with the pressure flow information data segment.

[0260] In other examples, the frequency content measure may include, but not be limited to, dominant frequency components, spectral entropy features, mel-frequency cepstral coefficients, persistence spectrum related features, dominant frequency magnitude, centroid, spectral spread or spectral flatness derived from the pressure flow information data segment.

[0261] In one example, determining a spectral content measure comprises dividing the pressure flow information data segment into a number of frames, determining a respective MF value for each frame in the number of frames and then determining a lower bound MF value for which a selected percentage of the respective MF values exceed.

[0262] As discussed previously, and referring again to FIG. 11, the pressure flow information data segment S' may be divided into a number of time-domain data frames {a2, a2, ... a„}. A respective MF value for each frame is then determined for each of the data frames of the set A ={ai, a2, ... a„} to provide a corresponding number of MF values, ie, {MFi, MF2, ... MFn}. The MF values may then be ranked fromlowest to highest to determine a lower bound MF value for which a selected percentage of the respective MF values exceed.

[0263] As an example, the lower bound MF value for which 90% of the MF values exceeds may be referred to as the “L90 MF value” indicator meaning that 90% of the MF values are higher than the L90 MF value. In other words, in some embodiments, other LXX values (where XX is a two-digit number) may be extracted, such as L95, L85, L80 or L75 as further non-limiting examples. Other LXX values may be determined.

[0264] As would be appreciated, the above framing process may be applied to other signal intensity measures as referred to above.

[0265] As will be seen below, once the above form of spectral content measure has been determined for an individual pressure flow information data segment measured at a point in time for a particular pipeline location it can be assembled with other spectral content (in heatmap spectrograph, PSD or other format) from other spatial locations along a pipe or other element being subject to measurement (through spatially discrete (eg, accelerometer) or continuous (eg, FBG or DAS fibre) to form greater assemblages of spatially and temporally varying spectral content from individual or combined data segments.

[0266] Spectral content measures for spatially and temporally varying individual or combined segment data include spectrogram PSDs (power densities) which can be determined at maximum, minimum, average and median levels for each data segment as divided into frames, windows or bins for frequency analysis processing. Tracking median versus average power levels has the advantage of excluding some environmental flow and pressure signals which detract from the quantification of the pipe or other element being subject to sensor measurement data acquisition. In other examples, tracking maximum or minimum powers levels, or the envelope formed between them, the standard deviation and / or the variance of power levels across measured frequency ranges, with different frame, window and / or bin setups, may be insightful.

[0267] In another example, the spectral content measure of a pressure flow information data segment can be directly represented as a spectrogram or PSD (average or median power based) over a spatial range (a composite heatmap formed in space over a range of corresponding pipeline locations), over a series of time intervals (a composite heatmap formed over time) and / or over both a spatial range and time interval simultaneously to form a version of a “waterfall” representation of the pressure flow data information segments as referred to above (and as shown, for example, in FIG.18 A).

[0268] In one example, determining the segment condition measure for the pressure flow information data segment comprises processing the pressure flow information data segment by a machine learning model trained to identify a pipeline anomaly condition.

[0269] Referring now to FIG. 12A, there is a shown a flowchart of a method 1200 for processing a pressure flow information data segment (ie, PFIDS) by either a machine learning 1220 or statistical model 1230 to determine a segment condition measure according to an illustrative embodiment.

[0270] At step 1210, method 1200 comprises extracting one or more time domain and / or frequency domain features from the pressure flow information data segment (eg, corresponding to an acoustic recording by a pressure flow sensor from a leak at a pipe location). In various examples, the time domain features may include, but not be limited to, median amplitude values, interquartile ranges, persistence values, RMS values (such as mean and median RMS values). In various examples, frequency domain features may include, but not be limited to, dominant frequency components, median frequency values, spectral entropy features, mel-frequency cepstral coefficients, skewness, interquartile ranges, persistence spectrum related features, dominant frequency magnitude, spectral features such as spectral centroid, spectral spread and spectral flatness.

[0271] At step 1220, method 1200 comprises processing the one or more extracted time domain and / or frequency domain features by a feature based machine learning classifier trained to identify the pipeline anomaly condition, in space or time, based on the one or more extracted time domain and / or frequency domain features.

[0272] At step 1230, method 1200 comprises processing the one or more extracted time domain and / or frequency domain features by a statistical model comprising statistical analysis and trending to identify the pipeline anomaly condition, in space or time, based on the one or more extracted time domain and / or frequency domain features.

[0273] In one example, the segment condition measure may indicate whether a pipeline anomaly condition has been identified 1240 or not following processing of the pressure flow information data segment. In another example, the segment condition measure may indicate the type of pipeline anomaly, eg, pipeline leak or pipeline crack.

[0274] As would be appreciated, training of feature based machine learning classifier will involve an initial determination of those time domain and / or frequency domain features that contribute to the classification and identification of the pipeline anomaly condition and which will be extracted at step 1210 in analysing a pressure flow information data segment. In one example, features may be selected by a filter method (features are evaluated individually and selected before running a machine learningmodel), wrapped method (feature selection is treated as an optimization problem where a machine learning model is run iteratively to search for the optimal combination of features) or embedded method (the feature selection is based on the feature importance which is a result of machine learning model).

[0275] In one example, the pipeline anomaly may be the presence of a leak or crack. In this case, the feature based machine learning classifier will be trained on labelled power flow information data comprising examples where the pipeline anomaly is present in contrast to examples where there is no pipeline anomaly but other environmental noises sources may or may not be present in the pressure flow information data.

[0276] In one example, the feature based machine learning classifier is a decision tree model. In another example, the feature based machine learning classifier is a Support Vector Machine (SVM) model.

[0277] In another example, the information from a pressure flow information data segment may be processed or framed in the frequency domain (or mixed time and frequency domains) before machine learning and / or statistical training feature extraction. Referring now to FIG 12B, there is shown a series of plots 1280 corresponding to the median PSD power levels for a number of frequencies for pressure flow information data segments corresponding to a range of different pipeline locations. As shown in FIG.12B, the median PSD power levels varying in space (along a set of acoustic accelerometers (in this case 10) corresponding to 10 different pipeline locations all recording the same leak noise in this example) may be segmented to represent power level variations in frequency bands and show variation in these bands in space across sensor pairs (adjacent and non-adjacent) such that mean, median, standard deviation, variance and other indicators can be applied before direct statistical trending and alert setting and / or feature collation and feeding into decision tree or SVM machine learning models.

[0278] In one example, processing the pressure flow information data segment by a machine learning model trained to identify a pipeline anomaly condition comprises forming a 2D spectrogram from the pressure flow information data segment and processing the 2D spectrogram by a machine learning classifier trained to identify the pipeline anomaly condition based on the 2D spectrogram. In one example, the machine learning classifier is a convolutional neural network (CNN) model.

[0279] Referring now to FIG. 13A, there is shown an example structure of a CNN model 1300 for processing a pressure flow information data segment to determine a segment condition measure according to an illustrative embodiment. As would be appreciated, machine learning classifier such as a CNN does not require “hand-designed” low-level features and it is capable of learning mid to high-level information from the input data through the millions of parameters inside the deep neural network.

[0280] In one example, the CNN model processes the 2D spectrogram as the input, so that the loss of information from the input pressure flow information data segment is minimal. By way of overview, in this embodiment CNN model 1300 processes the pressure flow information data segment 1310 to form an equivalent mel spectrogram 1315 which is then processed by a machine learning step 1320 comprising multiple convolution 1321 and pooling 1322 steps followed by a classification step 1330 comprising forming a flattened feature vector 1331 which is then input into a fully connected neural network before connection to assigned outcomes.

[0281] In one example, an input pressure flow information data segment 1310 (eg, acoustic sound file) is broken into frames with minimum time lengths (eg, for a 10 second sound file it may be re-framed into 4 x2.5 second frames). The minimum frame time length is determined by the data capture practicalities and resolution (including sampling frequency and length) of the pressure flow sensing arrangement and by the need for a minimum time period of pressure flow information data and sampling frequency within it to reliably generate a spectrogram (or PSD). A spectrogram 1315 (or mel spectrogram) is then derived for pressure flow information data segment 1310 following the fast Fourier transform (FFT) process with parameters set to output a 96*64 spectrogram or mel spectrogram where 96 is the number of frames and 64 is the number of mel frequency bins in the image. It will be appreciated that other image parameters (eg, including other than a dimension of 96*64) can be applied in processing the pressure flow information data segment before passing into the model.

[0282] In one example, the training of the CNN deep neural network has been accelerated by using transferred learning from the VGGish model which has a neural network structure used for visual recognition training and is pre-trained on public big vision data with labelling (of the images) previously completed by the Visual Geometry Group (VGG). This pre-training embeds diverse feature recognisers for visual images and by transferring into the CNN model provides much improve model training initialisation. The input for the VGGish model used is the 96*64 mel spectrogram.

[0283] Referring now to FIG. 13B, there is shown a plot of the relationship between the mel frequency, m, and standard frequency, f. As can be seen, the mel scale is a non-linear transformation of the normal frequency scale. The mel scale has been derived such that sounds of equal distance from each other on the mel scale also sound to human hearing as if they are equal distances from each other. This can be compared with the standard frequency Hz scale where the difference between 500 Hz and 1000 Hz is obvious to human hearing but the difference between 7500 Hz and 8000 Hz is not. After mel transformation the frequency range 7500 Hz - 8000 Hz will sit in the 2500-3000 Hz mel scale which remains distinguishable to human hearing.

[0284] In relation to the convolution process 1321 within the machine learning step 1320, this process involves taking a kernel or filter matrix and passing it over the input mel or normal spectrogram 1315 totransform the part of the input image passed over based on the kernel or filter values. This process of passing over an input image such as the mel spectrogram 1315 is sometimes referred to as striding with the stride able to be set in relation to the kernel or filter size.

[0285] Referring now to FIG. 13C, there is shown an example of the convolution process 1321 forming part of the CNN model 1300 illustrated in FIG. 13 A. In this example, a 4 x 4 kernel “Kernel” is applied with a stride of one to the “Input” to result in convolved “Output”. The kernels or filters (the actual values) need to be stored as parameters in the overall CNN model. In the case of the application of the VGGish transferred learning, it is the kemels / filters from previous image feature recognition training that are being incorporated and applied to the classification of the input mel or normal spectrogram image 1315.

[0286] Referring now to FIG. 13D, there are shown examples of the pooling process 1322 forming part of the CNN model 1300 illustrated in FIG. 13A. As shown in FIG. 13D, the left hand side is an example of “Max” pooling and the right hand side is an example of “Average” pooling, which for both processes involves taking either a maximum or average value for an area covered by the pooling kemel / filter which in this example has a size of 2 * 2. The pooling process does not add parameters for training and are preset dimensions.

[0287] In addition to the convolution 1321 and pooling 1322 processes referred to above, an activation function may also be adopted to manage or control the dimensionality of the network being trained. In particular, negative input, convolved or pooled values emerging from the training process can be managed using what is termed a Rectified Linear Unit (ReLU) activation function which allows only positive units to remain active.

[0288] The combination of the convolution 1321, pooling 1322 and activation processes occurs in hidden layers within the CNN training as visually depicted in FIG.13 A in machine learning step 1320. Once all hidden layers are established, the input image is reduced as it passes through them to a final flattened vector representation 1331 of the final convolved and pooled units 1329.

[0289] This flattened vector 1331 is then fully connected to a further vector layer(s) 1332 which in turn is connected to a “Softmax” vector layer 1335 containing the actual image classifications for which decimal probabilities are assigned (adding up to 1.0).

[0290] The structuring after the vector flattening within the CNN model 1300 is similar to typical ANN structures with the input layer (flattened vector) connected to one or more hidden and fully connected layers before connection to the output layer (the actual classes used in back propagation training of the CNN model parameters). Each of the input file frames are subject to classification throughthe CNN model 1300 and a voting system is then applied to the results from each separate frame to derive a final classification result (with the highest overall probability of being the identified class).

[0291] The goal is to establish an architecture in the model (combination of hidden layers with convolution and pooling parameters) to achieve the best model image recognition accuracy. The output from the model is typically assessed using a confusion matrix. In one example, the CNN model is trained to identify four different conditions comprising a “pipe crack / leak” class (C) (also an anomaly condition), a persistent environmental noise class” (P), an intermittent environmental noise” class (I), and a “background noise / quiet” class (Q).

[0292] Referring now to FIG. 13E, there are shown some example confusion matrices 1380 following assessment of the output from trained CNN models 1300 according to an illustrative embodiment.

[0293] In this example, the result for the aggregated single pressure flow information data segment frames (ie, in the case of pipe leaks an acoustic sound file) are shown on the left side and on the right the confusion matrix is shown after voting combination (after re-combination of the 4 x frame splitting) for all frames on the right. Specifically, the confusion matrix on the right side of FIG. 13E results from training using 10 second long pressure flow information data segments whereas the confusion matrix on the left side of FIG. 13E results from training using the 10 second long pressure flow information data segments broken into 4 x 2.5s long segments, and the four CNN models being trained and the model accuracies then recombined to form the confusion matrix. The splitting assists with model accuracy because the occurrence of transient environmental noise (typically for l-2s from, for example, vehicles) acts to vary the overall data over 10s whereas over 4 x 2.5s periods the impact of transient environmental noise is contained to a specific pressure flow information data segment and is able to be labelled and classified accordingly.

[0294] Referring now to FIG. 13F, there is shown an example structure of a CNN Siamese Twin architecture 1390 according to an illustrative embodiment. CNN Siamese Twin Architecture is employed to identify when a particular spectrogram image has been seen before. This can occur where pressure flow information data segments (and their derived spectrograms and / or PSDs) are being classified from different time periods and where an acoustic noise source at a particular time occurs and then ceases but then reoccurs at some later time (or times).

[0295] Referring now to FIG. 13G, there is shown a series 1395 of spectrograms and PSDs corresponding to a pipeline location but for different time periods showing a noise signal that periodically reoccurs. In this example, the spectrogram and PSD on the 5thDec ceases on the 6thDec but then reoccurs on 7thDec and ceases again on the 8thDec. As would be appreciated this behaviour is not possible for a pipeline anomaly such as a leak even though when considered individually the pressureflow information data segments (or equivalent spectrograms / PSD) for the time periods on the 5thand 7thDec are similar to a leak and may possibly be classified as a leak.

[0296] The CNN Siamese Twin architecture is formed by developing and training two identically structured CNN models (eg, CNN type model 1300) on pressure flow information data sets and then running both on incoming (new) pressure flow information data segments (eg, acoustic sound files from an acoustic sensor processed into spectrogram (or mel-spectrogram) form to ascertain whether the input pressure flow information data segment has been seen before and has a similarity score above a defined threshold. As shown in FIG. 13F, CNN Siamese Twin architecture not only determines the likely acoustic source class (ie, “Background”, “Intermittent”, “Persistent” or “Pipe Break”) but also whether it has been seen before (and non-continuously) in the case of an anomaly in the form of “Pipe Break” being identified based on the similarity score.

[0297] Following processing by the CNN model 1300 or CNN Siamese Twin architecture 1390, the pressure flow information data segment is classified according to the class having the highest probability, and an anomaly condition alert (see step 1240 of FIG. 12A) is generated if the pressure flow information data segment is classified as, for example, a pipe crack.

[0298] As another example, the spectral content measure could be in spectrogram form and be susceptible to direct machine learning classification processing using spectrogram and / or mel- spectrogram generation from an individual data segment (eg, an acoustic noise file measured at a point in time and space (at a location)). This spectrogram can be directly encoded and input into convolutional neural network (CNN) machine learning models. Alternatively, PSD heatmap representations of spectral content (power levels) can be directly encoded and input into CNN machine learning models.

[0299] Alternatively, the spectrogram and / or PSD processing may be occurring in combination with correlation and / or coherence analysis of spatial pairs (or more) of pressure and flow data segment source sensors with CNN models applied to the resulting framed pressure and flow segment data to train to patterns in the visual representations of the data in space and time. Further, these PSD heatmaps can be generated using the average power levels from the framed (windowed) frequency analysis undertaken and / or generated using a MF (median) approach.

[0300] The construction of sequential coherence plots in heatmap form as well as PSD format is more specific to correlation and / or coherent analysis of three or more spatially related but distinct pressure flow information data segment collection locations. In one example, the segment condition measure for the pressure flow information data segment comprises identifying the associated pipeline location and time period for the pressure flow information data segment and determining respective coherence measures between the pressure flow information data segment and further pressure flow information data segmentsfrom the same or similar time period and corresponding to two or more of the remaining pipeline locations in the selection of pressure flow information data segments.

[0301] The coherence measure between a first pressure flow information data segment (ie, x(t)) and a second pressure flow information data segment (ie, y(t + r)) may be expressed using a form of coherence function or measure (where T is a time shift (to ensure the coherent components in the signals are overlapped when there is a known delay in the arrival of signal y) as follows: Equation 9

[0302] where y^y is the magnitude-squared coherence (or coherence function) between the pressure flow information data segments x(t) and y(t + r), Gxyis the cross-spectral density between the pressure flow information data segments and Gxxand Gyyare the auto spectral densities for the pressure flow information data segments x(t) and y(t + r), respectively.

[0303] Referring now to FIG. 1C, consider in one example pressure flow information that corresponds to fO pipeline locations (eg, as indicated by arrows on “Greenhill Rd”) that is measured over two or more time periods and there being a pipeline leak located in the pipeline 187. In one example, coherence measures are determined between a selected pressure flow information data segment and those pressure flow information data segments that correspond to adjacent pipeline locations to that of the pipeline locations associated with the selected pressure flow data segment.

[0304] In one example, the generation of spectrograms and / or PSDs for individual pressure flow information data segments measured, or processed into, continuous or discrete time intervals for adjacent or non-adjacent spatial locations (at which the individual pressure flow information data segments are measured), are replaced by correlation and / or coherence level determinations between adjacent and / or non-adjacent spatial locations at which individual pressure flow information data segments are measured.

[0305] Referring now to FIG. 14A, there is shown the variation in a coherence heatmap 1400 between adjacent sensor pairs versus frequency for the pressure flow information data segments obtained for the real system shown in FIG. 1C (at a point in time to reinforce the detected signal power from a pipeline anomaly in the form of a leak). Referring now to FIG. 14B, there is shown a 2D representation 1420 of the PSDs forming the basis of the coherence heatmap shown in FIG. 14A prior to coherence processing. FIG. 14B shows the 2D PSDs for the same time for each of the 10 pressure flow information data segment locations used to derive coherence levels in FIG. 14 A.

[0306] The pressure flow information data segments used to form the heatmap in FIG. 14A are affected by variability in gain levels from sensors (particularly for accelerometers, hydrophone and / or pressure transducer type), resonance induced via the method of installation of sensors and general environmental noise (not related to the leak or other condition issue of interest and with the environmental noise itself either coherently received at multiple spatially separated locations or not). This is reflected in the PSDs shown in FIG. 14B which exhibit variable baseline power densities across each of the spatially separate pressure flow information data segment locations.

[0307] In this example, heatmap 1400 uses the coherence level between pressure flow information data segments referred to above, and corresponding to adjacent pipeline locations, for developing a coherence focussed set of 2D series PSD representations as shown in FIG. 14C. In this example, pressure flow information data segment for sensor pairs plotted on the x-axis correspond to adjacent pipeline location pairings 1 and 2 (sensor pair 1), 2 and 3 (sensor pair 2), 3 and 4 (sensor pair 3), 4 and 5 (sensor pair 4), 5 and 6 (sensor pair 5), 6 and 7 (sensor pair 6), 7 and 8 (sensor pair 7), 8 and 9 (sensor pair 8) and 9 and 10 (sensor pair 9).

[0308] Referring now to FIG. 14C, there is shown a 2D representation 1440 of the heatmap information shown in FIG. 14B following coherence processing in accordance with an illustrative embodiment. The process for re-framing the PSD power densities shown in FIG. 14B obtains PSD power densities that are not affected by sensor gain, installation effects and / or environmental noise by refocussing on coherent energy. In this manner, the heatmap information following coherence processing is then available for further processing for machine leaning pattern and / or statistical analysis for event occurrence and localisation analysis.

[0309] Referring now to FIG. 14D, there is shown a flowchart 1450 of the steps for coherence processing of pressure flow information data segment according to an illustrative embodiment. The steps in the process for re-framing the PSD power densities based on the use of coherent signal content as shown in FIG. 14D are described as follows:

[0310] At step 1451, a PSD is formed from the pressure flow information data segment.

[0311] At step 1452, a normalised PSD is formed by normalising the PSD based on the minimumPSD value. This functions to compensate for any variability of gain (eg, in response to incidental noise levels) of the pressure flow sensor that generates the pressure flow information data segment. This corrects for the effect of variable gain which would otherwise differently emphasise bands where energy arising from a leak is present on a sensor by sensor basis (and the intensity of leak or other incidental noises recorded at each sensor).

[0312] At step 1453, a filtered PSD is formed by filtering the normalised PSD by applying a moving window of configurable width (eg, a window length of 55Hz) to remove sharper and shorter spikes in the PSD. This has effect of removing local resonances which present as spikes that may resulting from sensor fitting and connection variability.

[0313] At step 1454, a coherence processed PSD is formed by determining the maximum coherence between sensor i-1 and sensor i and sensor i and sensor i+1 and so on (ie, between the pressure flow information data segment and further pressure flow information data segments from the same or similar time period that correspond to the two or more remaining pipeline locations in the selection of pressure flow information data segments) and applying the coherence as a weighting function to the PSD. This has the effect of reducing the power density not associated with the leak (being sensed across all the sensors). In this manner, the vibroacoustic signal energy corresponding to the leak is enhanced (and non-coherent non-leak energy) due to incidental noise at each sensor being reduced.

[0314] As would be appreciated, the 2D coherence processed PSDs 1440 is clearer than an equivalent unprocessed PSD counterpart 1420 as it is focussed on coherent noise (and excludes other incidental noise sources).

[0315] In another example, the strength and spread of the power density from the leak noise at each sensor location is obtained by summing the coherence values obtained, after the above described processing and illustrated in FIG. 14D, for each sensor and using the trends and variability in this plot to identify the occurrence and location of a leak. As would be appreciated, steps 1452 and 1453 may be optional depending on the configuration of the pressure flow sensing arrangement and local environmental conditions.

[0316] In this example, summed weighted coherence levels, derived via steps in the process summarised in FIG. 14D, are able to be determined for adjacent and non-adjacent pressure flow information data segment locations. An example of spatial framing of coherent energy is shown in the 2D formation for summed coherence levels in adjacent pressure flow data segment pairs is shown in FIG. 15A.

[0317] Referring now to FIG. 15A, there is shown a plot 1500 of the summed coherent power levels within the framed PSDs derived from the pressure flow information data segment pairs illustrated in FIG. 14A. Plot 1500 shows the spatially separated pressure flow information data segments that are processed in adjacent pairs (can also be non-adjacent pairs) such that variations in the summed coherence levels in space can be visualised and statistically quantified. As can be seen by inspection, the summation of the coherence values maximises for adjacent pair 4 (ie, corresponding to pipeline locations 4 and 5) which indicates that these pipeline locations are located either side of the leak and which most closely bracket it.The sum of the coherence values generally falls as the relevant pipeline location pair becomes more distant from the leak.

[0318] In one example, Spearman rank coefficients may be used to quantify increasing / decreasing PSD levels (after processing to account for coherence in the data segments) as shown for the adjacent pipeline locations to the “west” and “east” of a pipeline anomaly such as a leak. The Spearman ranking coefficient is determined for the levels of coherence between pressure flow (eg, acoustic) data segment pairs (adjacent or non-adjacent). The Spearman ranking coefficient is determined as:6 V d r = 1 - Equation 10 n(nz— 1)

[0319] where d, is the difference between the spatially ordered rank of the segment pairs and what their rank would be based on their summed coherence values and n is the number of pairs.

[0320] Referring now to FIG. 15B, there is shown a table 1550 showing the determination of the Spearman rank for the sensor pairs and their summed coherence values for the coherence values shown in FIG. 15A.

[0321] In another example, the 2D heatmap representation of coherence level relationships between adjacent and non-adjacent pressure flow information data sensing locations is formed such that losses in coherence along a spatial section of pipe and associated sensing accelerometers or fibre can be visualised and formulated for analysis while including effects from environmental noise, media signal loss and / or variable separation between pipe and sensing fibre in the framed data representation.

[0322] Referring now to FIG. 16, there is shown a coherence heat map representation 1600 of the coherence measure between pressure flow information data segment pairs corresponding to both adjacent pipeline locations and non-adjacent pipeline locations according to an illustrative embodiment. In this example, the coherence measure has been normalised to have a value between 0-100 with 100 indicating auto-coherence.

[0323] As can be seen, for the case of a pressure flow information data segments corresponding to 10 pipeline locations there will be 100 determined coherence measures for all the adjacent and non-adjacent pipeline location pairings. Once again, in this example the maximum coherence measure corresponds to the pair of pipeline locations closest to and also either side of the leak location (ie, in this case pipeline locations 4 and 5 as shown physically in FIG. 1C).

[0324] As would be appreciated, the advantages of having a spatially distributed sensing arrangement having sensors or sensing regions all able to record a pressure flow information concurrently can be demonstrated through coherence analysis.

[0325] One advantage is the ability to diagnose an anomaly in the pressure flow data segments (eg, a leak) and its location, and any intensification of coherence over time, if one or more of the sensors or sensing regions are either faulty or not communicating. Another advantage is the ability to diagnose a leak and its location, and any intensification of coherence over time, even if one or more of the sensors in the string is significantly affected by either incidental environmental noise and / or gain / resonance issues (as introduced previously).

[0326] Another advantage, in the case of pressure flow anomaly detection using fibre located remotely from the monitored physical system (eg, a pipeline) is the ability for the analysis and representation of the coherence measures along the fibre string to overcome situations where the media between the pipe and fibre absorbs all, or a proportion, of the energy from the pressure flow anomaly and / or the distance between the pipe and fibre varies (from near zero to where the pipe and fibre cross over to tens or hundreds of metres at other locations) such that the level of energy from the pipe pressure flow anomaly reaching the fibre is variably reduced (or can reach zero).

[0327] Referring now to FIG. 17, there is shown a coherence heat map representation 1700 similar to that illustrated in FIG. 16 showing the effect of incidental or environmental noise in the pressure flow information data segment corresponding to pipeline location 4. As can be seen, the pressure flow information data from pipeline location 4 has been contaminated (overlaid) with incidental noise (eg, see loss of coherence in region 1720 due to environmental noise). The coherence relationships between the remaining pipeline locations, however, are sufficient, based on a symmetrical pattern of coherent pipeline location pairs around pairs with weaker coherence, to continue to provide coherence strength trend and leak location information.

[0328] In another example, the appearance of two or more energy (eg, vibroacoustic power) sources can occur and may be diagnosed, without further information and analysis, as coming from two or more independent sources. However, if the coherent analysis described is undertaken then the energy from two or more sources will be able to be analysed for levels of coherence and a single source identified and located even if the sensors or fibre closest to the actual source (eg, a leak) are not able to measure the pressure flow anomaly. The energy sources can be confirmed as independent or related using the analysis and representations illustrated in FIG. 16 and FIG. 17.

[0329] As will be seen, the types of coherence heatmap patterns presented in FIGS. 16 and 17 are susceptible to interpretation and quantification using both statistical (eg, reduction of the informationdistributed in space across any adjacent or non-adjacent press flow segment sensing pair into 2D forms makes it susceptible to the determination of Spearman ranking coefficients and any other statistical methods determining maximum, minimum, average, median, variance or other features) and machine learning methods (eg, through the application of CNN type models to the derived coherence heatmaps for classification training based on different forms of labelled heatmap patterns and their physical meanings).

[0330] Referring back to FIG. 10, at step 1020 the pipeline condition measure is determined based on the respective segment condition measures corresponding to at least three or more pipeline location determined at two or more time periods.

[0331] In one example, a pipeline condition measure in the form of a “heatmap” plot may be formed comprising multiple segment condition measures where an assembly of segment condition measures ordered with respect to time and pipeline location is formed. In one example, the heatmap plot is a spectral heatmap comprising spectrograms or PSDs generated from individual pressure flow information data segments corresponding to different time periods and pipeline locations. In one example, a spectral heatmap may be formed from an assembly of spectrograms determined for individual pressure flow information data segments measured continuously at discrete time intervals (which may or may not form a continuous time sequence). In another example, the spectral heatmap may be formed from an assembly of PSDs determined for individual pressure flow information data segments measured continuously at discrete time intervals (which may or may not form a continuous time sequence).

[0332] In another example, the spectral heatmap may be formed from an assembly of spectrograms determined for individual pressure flow information data segments measured at discrete adjacent or non- adjacent pipeline locations. The spectrograms assembled spatially along the pipe may or may not be derived from individual pressure flow information data segments measured continuously or at discrete tine intervals. Similarly, the spectral heatmap can be formed from an assembly of PSDs determined for individual pressure flow information data segments measured as discrete adjacent or non-adjacent locations along a pipe or other element of interest. Further, these discrete adjacent or non-adjacent locations may have pressure flow information data segments measured at continuous and sequential time intervals (smaller or larger).

[0333] Referring now to FIG. 18A, there is shown a spectral heatmap 1800 comprising an assembly of individual spectrograms for a selection of individual pressure flow information data segments covering a range of pipeline locations and time periods according to an illustrative embodiment. In this example, spectral heatmap 1800 covers 10 pipeline locations shown across the x-axis 1820 and 14 time periods shown across the y-axis 1810. As would be appreciated, spectral heatmap 1800 is in the form of a “waterfall” map that is generated for time varying pressure flow information signals from spatiallydistributed sensors (or a distributed sensing arrangement) which are also measuring either continuously or at discrete time intervals.

[0334] In this example, as shown in spectral heatmap 1800, an anomaly in the form of a leak on a pipe is introducing vibroacoustic energy with a maximum in the spectrogram corresponding to the fifth spatial pipeline location 1822 as determined from the corresponding pressure flow information data segment and the leak then ceasing in the tenth time period 1812 within the overall assemblage.

[0335] Apart from the main leak, other vibroacoustic energy sources can be seen over more limited spatial ranges (sometimes just one pipeline location) and more limited time ranges (sometimes just one time period) within the individual spectrogram representations of each individual pressure flow information data segment. In another example, a time and space representation of individual data segments formed from PSDs, correlation and / or coherence processed segments, or pairs or more of adjacent or non-adjacent spatially distinct data segments can be generated.

[0336] Similarly, segment condition measures in the form of correlation or coherence content measure for pairs or greater numbers of sensors and / or along a continuous sensing system (eg, FBG or DAS fibre) for pressure flow information data segments varying in space and / or time can be represented to form an assembly of segment condition measures ordered with respect to time and pipeline location. It is important to note that while the optimal configuration of a FBG and / or DAS fibre sensing system would provide spatially continuous data in a classic “waterfall” representation (eg, for cases where the sensing fibre is mounted in direct contact continuously along a pipe wall externally and / or internally within the pipe) this does not occur in reality when using in-situ spare fibre (“dark fibre”) which is spatially separated from the pipe or other element being monitored (with the sensing fibre and pipe not on the same alignments with variable distances separating the target pipe and the nearest length of sensing fibre).

[0337] The assembly of spectrogram, PSD, correlation and / or coherence pressure flow information from data segments in “waterfall” format is one way to process data ready for analytic interpretation and acoustic leak energy identification and account for temporal and spatial variability (including damping and dispersion of pressure flow information along pipes or through media between pipes and fibres).

[0338] In one example, determining the pipeline condition measure based on the respective segment condition measures comprises determining whether one or more of the respective segment condition measures satisfy a threshold condition. In various examples, the threshold condition will depend on the respective segment condition measure.

[0339] In one example, where the segment condition measure for a pressure flow information data segment comprises a single value such as the L90 RMS and / or L90 MF value for the respective power flow information data segment, satisfaction of the threshold condition may comprise one or more of the determined L90 RMS or L90 MF values that have been determined exceeding a reference value, or falling within a range that indicates the presence of a pipeline anomaly or fault.

[0340] In other examples, where the pipeline condition measure comprises a set of values such as the calculated power densities for a pressure flow information data segment corresponding to a selected set of frequency bands, then satisfaction of the threshold condition can involve a comparison between whether one or more of the calculated power densities exceed a corresponding reference range, or fall within a range for the particular frequency band that indicates the presence of a pipeline anomaly or fault.

[0341] In another example, where the segment condition measure is the result of the application of a machine learning model, that following processing of a respective pressure flow information data segment indicates that there is a pipeline anomaly condition, then satisfaction of the threshold condition is simply the binary determination of whether the machine learning classification has identified the presence of the pipeline anomaly condition.

[0342] As would be appreciated, threshold conditions may require the experience of relevant skilled persons based on an understanding of the physical anomalies being monitored. This may be done after statistical and / or machine learning processing based on empirical experience with the data and / or partial or fully deterministic methods (based on physical feasibility and / or calibration results).

[0343] Where one or more threshold conditions are satisfied, after statistical and / or machine learning processing, methods of combination and weighting of the combined results (to determine a level of response prioritisation and urgency), are required. These combinations can be achieved based on fixed empirical experience of the relative importance (and accuracy) of each statistical and / or machine learning method applied and the resulting level of threshold exceedance or not.

[0344] In cases where some of the segment condition measures as determined in accordance with the present disclosure are exceeding thresholds, but others are not, the structure and operation of the combination and weighting process may be important in determining whether 1) any alert or other response action will be raised and 2) what the priority and / or urgency is given to the alert or other response action.

[0345] In another example, threshold conditions can be determined by a range of experienced users based on physical understanding and human cognition and the results weighted and averaged to derive a combination function for the various statistical and / or machine learning process outcomes. An additionallayer of machine learning training can be implemented after the first statistical and / or machine learning processing of the pressure flow data segments to generate the segment condition measures (eg, CNN type image classification model as trained at fixed points in time) which incorporates human assessment of the statistical and / or machine learning outcomes (before or after combination and weighting) and adds a further layer of insight or classification adjustment.

[0346] In another example, non-fixed combination processes (versus fixed empirical settings) are implemented, wherein feedback from pressure flow data segment analysis and alerts, through to actual physical anomaly detection and repair outcomes is incorporated into an updatable combination and weighting system such that the accuracy of the pipeline condition determining system is improved.

[0347] In one example, determining the pipeline condition measure based on the respective segment condition measures comprises determining whether two or more of the respective segment condition measures corresponding to the same pipeline location satisfy a temporal rate of change condition.

[0348] In one example, where the segment condition measure is a single value parameter such as the L90 RMS or L90 MF, satisfaction of the temporal rate of change condition may involve determining the differences between these parameters for successive time periods and determining whether the difference exceeds a reference value, or falls within a range, for the difference in time that would indicate a pipeline anomaly associated with the pipeline location.

[0349] For example, in certain embodiments a temporal rate of change (ROC) of a peak and / or average percentage above the upper bound of the occurrences and / or the integrated total RMS / MF value, also above the upper bound, for the new or next “current” RMS / MF values may be determined by establishing differences between each parameter for the new or next “current” (time = t + At), relative to the updated reference RMS / MF value (including the effect of the changes at time t), and quantifying a percentage shift and / or percentage shift in time between segment condition measures.

[0350] The above process may then be repeated for time periods corresponding to t + 2At, t + 3At .... t + nAt. In another example, the difference between the time periods of the respective segment condition measures may be variable.

[0351] Depending on the difference At between the time periods, a selectable number of segment condition measures corresponding to a given pipeline location may be used to assess the ROC of the segment condition measure, such as indicative statistic parameters in order to classify a pressure flow anomaly (eg, a leak / crack event for a pipe). For example, if the ROC:• reduces to 0 (± a % tolerance), after an initial change and rate of change is determined from pressure flow data segments, and after a further selectable number of pressure flow data segmentsand associated segment condition measures corresponding to consecutive time periods (the assessment period), then the fault (in the context of the range of faults that occur for leaks and pipes) may be assigned a diagnosis as either a likely joint leak or circumferential crack and a repair priority (and forecast date) can be established; or• continues to increase (selectable bands can be set - eg, 10% growth rate for pre-determined number of pressure flow data segments and associated segment condition measures corresponding to consecutive time periods), after starting and / or an initial change and rate of change is determined from pressure flow data segments, then the fault may be assigned a diagnosis as a likely longitudinal crack and a repair priority (and forecast date) can be established.

[0352] In various examples, determining whether two or more of the respective segment condition measures corresponding to the same pipeline location satisfy a temporal rate of change condition will comprise initially determining one or more comparison temporal benchmark (BMs) and determining whether a temporal BM has been exceeded. In one example, determining the benchmarks (BM) may involve establishing short period BMs and / or longer period BMs. As an example, short period BMs may be based on, for example, a daily or weekly period, whereas longer period BMs may be based on a monthly period. It will of course be appreciated that other suitable periods may be used to establish the temporal benchmarks. For example, a suitable period may be set, or adjusted, depending on parameters such as the time of year, weather, or water demand / usage patterns.

[0353] In one example where the time periods correspond to successive days, benchmarks are established using the power levels (P) recorded in each frequency bin for each PSD for each day. Equation 11 shows the method of calculation of the benchmark (BM) for current day (n) in a frequency bin (i) as the averaged daily power densities from the previous day (n-1) back to the start of the benchmarking period (n-LBM) where LBM equals the length of the benchmark period: Equation 11

[0354] The benchmarks may then be used to calculate the difference between the segment condition measures in the form of measured PSDs of corresponding pressure flow information data segments and the benchmarks (with the difference also plotted in the form of a heatmap).

[0355] Equation 12 shows the method of calculation of the difference for the current day (APSDn) in a frequency bin (i) as the power density for the current day (Pn) minus the benchmark value (BMn) previously calculated using Equation 11 :Equation 12

[0356] In one example, a sliding window approach is applied where frequency bands (ie, windows) of different selectable widths may be moved across the overall sensor frequency range to identify the bands which, when overlapped with the sliding window, best highlight the change in vibroacoustic energy level within that window width. The change in vibroacoustic energy level is derived using the benchmarks described above. It is important to note that the width of the band with the greatest vibroacoustic energy increase is not typically known initially, and that a range of window sizes need to be moved across the overall sensor range to find the window width and position within the overall frequency range that best focuses on the frequency bands, with the energy increase signifying the anomaly.

[0357] Power spectral density difference (APSD) maps can be visualised as a 2D matrix usingEquation 13 below in which, for example, I = 129 (frequency bins) and K = total number of days (for an example with a daily time interval between pressure flow data segments (eg, acoustic recordings)) in the heatmap and Apsdi,k = the power density change for any particular day (k) or frequency bin (i):Equation 13

[0358] The cumulated PSD change (typically growth) using different sliding window widths (SWWs) from 1 to 129 bins, for example, can be visualised as a 3-D matrix (APSD) with elements Apsdij.k where i = 1, ..., 129, j = the SWW and k = the relevant day within the overall number of days (K) comprising the map. The SWW is split evenly either side of the central frequency bin (i) during the determination of Apsdij,k. The elements in the 3D matrix Apsdij.k can be determined using Equation 14 below for i = 1, ... , 129 and k=l, ..., K: Equation 14

[0359] Only the cumulated PSD changes for each of the sliding windows adopted (say number n) are stored per day (for example), resulting in a reduction from a 3-D to 2D matrix representation of the maximum cumulated (me or MC) PSD changes over the sliding windows as shown in Equation 15 where mcApsdj,k = max (i e I) Apsdi,j,k :Equation 15

[0360] The maximum cumulated acoustic energy changes for each SWW from the 2D matrix represented in Equation 15 are then normalised against the row maximum within the matrix (ie, the maximum of the total number of days K used in the analysis) as shown in Equation 16 where nmcApsdi.k = me Apsdi,k / [max(k e K) mcApsIi,k with i = 1 , ... ,n] :

[0361] Comparison of on-going data with the temporal BMs may involve using measurements over, for example, a previous week, month, or plural weeks or months, depending on the model structure established. In some examples, processing using multiple BMs may be used to provide a continuous ROC indication.

[0362] The ROC of the peak and average percentage above the upper bound of the occurrences and the integrated total RMS / MF value, also above the upper bound, for the new or next “current” RMS / MF values may also be determined by establishing the differences between each parameter for the current (time = f) and new or next “current” (time = t + At) and quantifying a percentage shift and / or percentage shift in time.

[0363] In certain embodiments, the ROC is assessed relative to the established comparison benchmark RMS / MF values within their respective frames (eg, frequency windows) or for benchmarks established using integrated totals across all frames. For example, benchmarks may be established for each pipeline location and can be established over durations of days, weeks or months.

[0364] Benchmarks established over shorter periods may reflect (and account for) dynamic changes in system and environmental noise at specific pipeline locations. Benchmarks established over longer periods are less reactive to dynamic changes in system and environmental noise at specific pipeline locations and are more sensitive to changes in RMS / MF within their respective frames and can be used to detect pipeline anomalies such as early crack development.

[0365] Measurement data for a period of days, weeks or months may be compared with the relevant established benchmarks to determine quantified statistics for the ROC that is occurring.

[0366] In one example, weekly comparisons are used as a detector for more rapidly evolving pipe faults such as circumferential cracks. Equations 17, 18 and 19 are used, with Equation 17 used to establish the level in the current week (days 1-7), Equation 18 used to establish the level in the reference week 1 month earlier (days 21-28) and Equation 19 used to establish the alert ratio between the current and reference weeks with an alert triggered based on exceedance of a set threshold: Equation 17 Equation 18anmchpsdi wk4alert ratio)i week= - - — - - Equation 19 anmc psdi wk4

[0367] where anmcApsd = average normalised maximum cumulated APSD for the relevant SWW (i) and time period in weeks as per the subscripts.

[0368] In another example, monthly comparisons are used as a detector for relatively slower evolving pipe faults such as longitudinal cracks. Equations 20, 21 and 22 are used, with Equation 20 used to establish the level in the current month (days 1-28), Equation 21 used to establish the level in the reference month which is 4 months earlier (days 84-112) and Equation 22 used to establish the alert ratio between the current and reference months with an alert triggered based on exceedance of a set threshold: as follows:Equation 20Equation 21, , , . anmcpsdi mthl(alert ratio)i montfl= - Equation 22 anmcpsdimtk4

[0369] where anmcApsd = average normalised maximum cumulated APSD for the relevant SWW(i) and time period in months as per the subscripts.

[0370] Referring now to FIG. 18B, there is shown a series of diagrams 1830 showing segment condition measures in the form of PSD heatmaps 1835 covering multiple days, a difference PSD heatmap 1838 showing the difference between a PSD heatmap for a given day, and a benchmark PSD heatmap and an associated pipeline condition measure 1840 in accordance with an illustrative embodiment.

[0371] In this example, the pipeline condition measure 1840 comprises a short time based ROC alert 1841 that has been triggered over a 24 day period as indicated, and a long time based ROC alert 1842 that has been triggered over a 44 (on-going) period as indicated. In this example, the anomaly was a circumferential crack in the pipeline which was repaired.

[0372] In another example, the consistency of the rate of change of power density in frequency bands can be determined using the Spearman’s rank coefficient previously introduced. Spearman’s rank coefficient provides a measure of the consistency of the monotonic increase in the power density over a window of time and in a frequency range. The determination of Spearman’s rank coefficient has been previously described and is defined as the Pearson correlation coefficient between the rank variables of two sets of data.

[0373] In this example, for n continuous days, the number of days X and the PSD in a specific frequency band Y are converted to the ranks R(X) and R(Y). The Spearman rank r between them can be determined as follows: cov (P(X),P(T)') r = PR - Equation 23

[0374] where p denotes the usual Pearson correlation coefficient but applied to the rank variables, cov / ?( / )) is the covariance of the rank variables, and aRX) and oR(y are the standard deviations of the rank variables.

[0375] The determination of the Spearman rank for PSD power density in a specific frequency band is determined by applying a fixed temporal window length (WL) and moving the window at a time step (eg, day by day for acoustic data). As the above process is applied to different frequency bands, covering the range of measured frequencies for the pressure flow information data segments, a map of the Spearman ranks for different window lengths (WL) versus frequency can be developed.

[0376] Referring now to FIG. 18C, there is shown a series of heatmaps 1850 of the Spearman rank corresponding to a vibroacoustic event involving a pipe leak evolving over a period of 60-70 days for window lengths of 20 (1850a), 30 (1850b), 40 (1850c) and 50 (1850d) days according to an illustrative embodiment. In this example, there are 129 frequency bins (bands) each 18.3 Hz wide over a range of 2340.5 Hz for the pressure flow information data segments (ie, acoustic sound files measured once per day).

[0377] Plot 1850a of FIG. 18C shows an increase in PSD power density between 1000-1500 Hz at day 35 for the Spearman rank coefficient map with the shortest WL of 20 days. The longer the WL, the longer the delay for the increasing power density to become apparent. For the longest WL, the increase in PSD power density is only apparent after 50 days. However, there is a trade-off between the early detection of a power level increase (and possible leak alert), and the confidence with which any continuity in power increase can be confirmed. FIG. 18C shows that in this example a WL of 50 days provides clear confirmation of the continuous growth in the PSD power density increase.

[0378] In one example, the use of short and long WLs can be combined to achieve the benefit of early detection (short WL configuration) and the certainty of confirmed continuity of PSD power density increase (long WL configmation). In this example, WLs of duration including, but not limited to, less than 10 days, 10, 12, 15, 18, 22, 26, 30, 35, 41, 47, 54, 61, 70, 80, 90, 100 and greater than 100 days, may be flexibly adopted based on empirical experience with developing pipe leaks. Spearman’s rank coefficient r is calculated for each WL and for each pressure flow information data segment recordinginterval or duration (eg, typically a day for acoustic sound files but can be shorter for fibre acquired data). The maximum r value is then retained for each frequency bin range.

[0379] Referring now to FIG. 18D, there are shown plots of the maximum r value (1860a) and the corresponding WL (1860b) for an example vibroacoustic event involving a pipe leak evolving over a period of 60-70 days according to an illustrative embodiment. In this example, a threshold is applied to the maximum r value such that only clear increasing PSD power level trends are displayed. In this example, an r value threshold of 0.5 is applied and must be exceeded before the r value will be shown in 1860a but this can be varied.

[0380] As the number of frequency bins (bands) that have growing PSD power levels increases, so does the confidence that a pipe leak is occurring. Another example statistical indicator is the degree of spreading of the frequency bins containing increasing PSD power levels. By assuming the positions of these effective bins as a vector B e {1: 129} which contains K (< 129) bin positions, the bin spreading factor can be defined as: Equation 24

[0381] where K2is a constant and:G1(l: K — 1) = B(2: K) — B(l: K — 1) Equation 25

[0382] The vector Gi depicts the gaps between two adjacent bins among all the K frequency bins. The bin spreading factor equals the averaged gap before eliminating the largest I<2 gaps.

[0383] Referring now to FIG. 18E, there is shown two example clusterings of frequency bins 1870 having increased PSD power levels according to an illustrative embodiment. In this example, two illustrative cases with K2=1 are shown to illustrate the physical meaning of this factor. In 1870a, the frequency bins are clustered in two regions (1 2 3 4) and (9 10 11), and the gaps are mostly 1 with one exception. By eliminating this exception, the averaged gap for 1870a is then equal to 1. Analogously, the number of exceptions will be K2- 1 if these bins are clustered in K2regions.

[0384] From the PSD data collected from different sensors with a pipe leak involved, it has been found that these frequency bins with an increasing trend caused by a developing pipe leak are mostly located in one or two regions (ie, when more clustered). Thus, the parameter K2is set as 4 to allow for 3 separated regions for these bins to distribute across. For 1870b, when these frequency bins spread across a large range, a large averaged bin gap of 2.2 after eliminating the largest value is obtained.

[0385] In one example, an alert score A considering all the criteria can be formulated as:

[0386] where WFis the weight associated with the frequency bins, WTis the weight associated with the temporal window, K is the number of effective frequency bins with the coefficient r larger than the threshold rs, and fc3is the exponential factor that needs to be selected.

[0387] When this alert score A is larger than a threshold As, then a pipeline anomaly condition will be indicated. The alert score also indicates how strong and consistent the increasing trend is, and thus it is an indicator of the confidence of the detection.

[0388] It can be seen that a temporal window weight WT is introduced in Equation 26. A positive relationship between the temporal window weight WT and the length of the window WL is assigned given a longer window length means higher confidence of the detection. First, a weight of 1 is assigned for the shortest window length (10 days) and a weight of 2 for the longest window length (100 days). This means that an increasing trend in a specific frequency bin in a period of 100 days will be twice as likely to be caused by a pipe leak than that in 10 days in the same frequency bin and with the same r value.

[0389] A non-linear relationship between the window length WL and the temporal window weight WT which satisfies WL = 10 days and WT = 1 and WL = 100 days and WT = 2 is assigned as follows: / (WL- 1O)\0 4c.. „WT= 1 + I — - - - 1 Equation 27T1 100 - 10

[0390] A frequency weight WF is also introduced as: Equation 28

[0391] where K defines the minimum number of frequency bins (bands) required to raise an alert as, in this example, 15 out of 129 bins and G is the previously defined bin spreading factor.

[0392] The frequency weight WF is positively correlated with the number of bins and negatively correlated with the bin spreading factor. Other key parameters to select include the exponential factor k3of the r value (as mentioned above), the threshold of the r value (rs), the threshold to trigger the alarm (As), and the options for the flexible window length.

[0393] Referring now to FIG. 18F, there is shown a table 1880 of some example parameter values for use in determining an anomaly condition measure based on increasing PSD over time according to an illustrative embodiment.

[0394] While the above examples relate to frequency and / or MF processed data, analogous methods are applicable to RMS and other magnitude data.

[0395] In another example, the determined ROC value for a single value parameter such as L90 RMS or L90 MF for a given time period may be compared with a machine learning generated value from a machine learning system (eg, ANN or RNN), trained on historical pressure flow information data corresponding to the pipeline location and time periods, and taking into account variations that occur in the pressure flow information on a monthly, weekly and / or daily basis.

[0396] Referring now to FIG. 19 A, there is shown an example structure of a CNN model 1900 for determining whether two or more respective segment condition measures corresponding to the same pipeline location satisfy a temporal rate of change condition according to an illustrative embodiment. In this example, an assembly of segment condition measures in the form of spectrograms or median derived PSDs corresponding to different time periods are assembled to form a heatmap 1910 which is used for CNN model 1900 training and assessment.

[0397] As would be appreciated, CNN model 1900 is different to CNN model 1300 (see FIG. 13A) where the input is for a single point in time period derived mel-spectrogram.

[0398] Referring now to FIG. 19B, there is shown a table 1950 detailing the various convolution, pooling and activation stages in the hidden layers of the CNN model 1900 for a model with one extra hidden layer (see rectangle 1955), and another with one less hidden layer, to assess the effect of model configuration complexity against accuracy but also overfitting (parameter redundancy) problems. The stages of convolution and pooling, numbers of model parameters and then flattening and connection to a fully connected layer (dense layer) are illustrated.

[0399] In one example, determining the pipeline condition measure based on the respective segment condition measures comprises determining whether two or more of the respective segment condition measures corresponding to the same or similar time period satisfy a spatial rate of change condition.

[0400] By analogy to the disclosed methods and processes to determine whether two or more of the respective segment condition measures corresponding to the same pipeline location satisfy a temporal rate of change condition, these same processes may be adopted to determine whether two or more of the respective segment condition measures corresponding to the same or similar time period satisfy a spatialrate of change condition, noting in this case that instead of analysing for temporal variation the analysis is conducted with respect to spatial variation.

[0401] Accordingly, and by contrast to temporal ROC, instead of establishing a temporal benchmark at a fixed point (ie, pipeline location) for a pressure flow information data segment series of measurements covering different time periods, and then processing to determine the rate and characteristics of change in time of that data, the process is applied spatially to form a spatial benchmark.

[0402] In one example, a spatial benchmark is established for multiple adjacent and non-adjacent pressure flow data segment acquisition locations along a pressure flow information sensing arrangement comprising a string of pressure flow sensors or a fibre comprising sensing regions. The form that the pressure flow information data segment that the spatial benchmark can be based on includes any of the previously disclosed forms including 2D processed (eg, coherent power focussed) PSDs, or summed coherence levels across a spatial section of monitored infrastructure (eg, a pipeline with accelerometers, hydrophones and / or pressure transducer string spatially distributed along it through to a pressure flow information sensing arrangement comprising offset fibre sensing). In another example, the spatially varying pressure flow information data segments, in raw or processed forms, may be processed to form segment condition measures comprising heatmaps showing PSD power level or coherence (or summed coherence) levels.

[0403] A spatial benchmark for a single set of a substantially simultaneously measured and spatially distributed pressure flow information data segments allows diagnosis of single or multiple pressure flow data segment sources of vibroacoustic energy such as a leak or environmental noise sources, the locations of each source of vibroacoustic energy, and the coherence levels between the sources of vibroacoustic energy. This provides an ability to manage the loss of data or low signal of interest to environmental noise ratios along the spatial section of monitored infrastructure as discussed above in relation to determining a segment condition measure in the form of a coherence measure. This spatial benchmark can be compared at a single point in time (with no other temporal context) with expected levels of spatial signal loss either along the infrastructure (eg, along a pipe from a leak to sensing locations) or through media between the infrastructure and sensing locations (eg, transmission from a pipe with a leak to fibre through ground and other media).

[0404] In one example, the expected levels of spatial signal loss are determinable theoretically and / or through databases for different physical configurations. Referring now to FIG. 20, there is shown a plot 2000 of the pressure flow information signal (dB) as a function of distance from a source of vibroacoustic energy in the form of a leak. In one example, the pressure flow information data segment may be benchmarked compared with the expected transmission range based on factors such as pipe diameter, material jointing, embedment conditions age and condition (eg, see FIG. 20).

[0405] In another example, another set of benchmark database criteria would include those specific to the device or system being used to provide the measured pressure flow segment data including:• In one example, the type and sensitivity of the pressure flow sensors (eg, accelerometers) used to acquire acoustic data, their method of connection to a pipe (resonance issues) and gain settings in the signal acquisition.• In another example, the type and sensitivity of fibre used to acquire acoustic data in a distributed pressure flow sensing apparatus, the number of splices along any fibre and its physical installation conditions (including depth and whether in a secondary conduit, buried in soil, or located within below ground chambers or tunnels).

[0406] In another example, determining the pipeline condition measure based on the respective segment condition measures comprises determining whether two or more of the respective segment condition measures corresponding to the same or similar time period and two or more of the respective segment condition measures corresponding to the same pipeline location satisfy a combined temporal / spatial rate of change condition.

[0407] In this manner, in one example both the use of spatial and temporal benchmarking and associated Rate of Change (ROC) analysis is applied to the selection of pressure flow information data segments (and associated segment condition measures) to determine a combined temporal / spatial rate of change condition . The processing of the pressure flow information data segments takes place as per any of the following (non-exhaustive list):• 2D PSDs or spectrograms for pressure flow (including acoustic) data segments.• Statistically reduced trends of maximum, minimum, average, median, variance and / or other (eg, Spearman Rack) indicators.• PSD heatmaps assembled over time.• 2D heatmaps in space for adjacent and non-adjacent sensors (or strings) based on either PSD power or coherence level inputs.• Spatial coherence values filtered and / or summed.• 2D plot and / or the assembly of waterfall plots of varying time or frequency domain processing of pressure and flow segment data.• Alert state mapping (eg, from the application of the CNN classification through to ROC in time methods).

[0408] In one example, the ROC method for determining changes in PSD power levels in frequency bands within pressure flow data segments (ie, acoustic noise files), as they vary over time, as applied to pressure flow data segments acquired at a specific spatial location (eg, a sensor location or fibre channel cross-over with a pipe), can be applied at the same point in time across a string of sensors and / or along afibre to determine simultaneously at that point in time whether adjacent or non-adjacent spatial sensing locations are alerting to a change in power levels at their locations that are above thresholds.

[0409] Referring now to FIG. 21, there is shown a plot 2100 of the results of a time based ROC pipeline condition measure applied to pressure flow information data segments from the 10 sensor configuration illustrated in FIG. 1C provided daily for an extended period according to an illustrative embodiment. In this example, the anomaly is a pipe leak between the 4thand 5thsensors (as per the corresponding columns depicted in FIG. 21). The threshold alert for the ROC method is triggered relatively uniformly for most of the spatially separated sensing locations around the date of 14 / 15thof August 2020. The diagnosis of a leak, its location, its time and duration of occurrence and any issues with specific sensing locations (in terms of sensor state or environmental noise) can be reinforced as illustrated in FIG. 21.

[0410] Referring now to FIG. 22, there is shown a heatmap representation 2200 where the individual segment condition measures are the coherence levels for pressure flow data segments corresponding to adjacent pairs of pipeline locations (ie, the spatial dimension) as determined daily according to an illustrative embodiment. In another example, a heatmap representation may be formed for non-adjacent pairs of pipeline locations.

[0411] For the heatmap 2200 shown in FIG. 22, the time is shown on the horizontal axis and the spatial dimension (sensor adjacent pairs in this example) is shown on the vertical axis. This coherence level in time versus space (sensor pairs) “waterfall” type heatmap framing of the pressure flow segment data is susceptible to statistical analysis techniques for analysis of maximum, minimum, average, median, variance and / or other measures over all or regions of the heatmap and / or within particular temporal and / or spatial bands across the heatmap allowing comparative analysis of coherence levels across such temporal or spatial bands.

[0412] In one example, determining the pipeline condition measure based on the respective segment condition measures comprises processing the selection of respective segment condition measures in combination by a machine learning model trained to identify a pipeline anomaly condition.

[0413] In one example, the machine learning model is based on feature extraction from the combined respective segment condition measures and the machine learning model may be a decision tree model or a SVM model.

[0414] Referring now to FIG. 23, there is shown an example structure of a CNN model 2300 showing the processing of the selection of respective segment condition measures in combination 2310 by amachine learning model 2320 trained to identify a pipeline anomaly condition 2330 according to an illustrative embodiment.

[0415] In this example, the individual segment condition measures correspond to the spectrograms for each of the pressure flow information data segments and the selection of respective segment condition measures in combination has a 2D representation where the segment condition measures are arranged in a 2D array with the x-axis corresponding to pipeline location and the y-axis corresponding to increasing time (eg, see “waterfall” type configuration of heatmap 1800 illustrated in FIG. 18A).

[0416] In one example, an individual pressure flow information data segment may correspond to a 10 second acoustic sound file as determined by a pressure flow sensor and the associated segment condition measure is the spectrogram generated from the sound file and the heatmap is the collection of spectrograms for a range of different sensors and time periods.

[0417] In this example, the machine learning model is a CNN model 2300 that is trained to classify the 2D representation of the spectrograms arranged in time and space. As such, in this example instead of the machine learning model being operative to determine a segment condition measure for an individual pressure flow information data segment, the machine learning model 2300 is operable to process the respective segment condition measures (in the form of spectrograms) of the selection of pressure flow information data segments in combination covering a range of pipeline locations and time periods that are being processed to then determine the pipeline condition measure. In this manner, the machine learning model has regard to both temporal and spatial information in its determination of the pipeline condition measure.

[0418] Referring now to FIG. 24 A, there is shown an example structure of a CNN model 2400 similar to FIG. 23, but in this case the combined respective segment condition measures 2410 are in the form of the 2D representation 2200 shown in FIG. 22 (ie, coherence intensity levels for adjacent pairs of pressure flow data segments are the segment condition measures shown in the 2D heatmap representation rather than spectrograms for the individual pressure flow information data segments).

[0419] Once again, in this example the machine learning model 2420 is a CNN that is trained to classify the 2D representation arranged in time and space but in this case the segment condition measure is the coherence level, as represented in a “waterfall” time and space varying heatmap, for adjacent pipeline locations. As such, in this example instead of the machine learning model being operative to determine a segment condition measure, the machine learning classifier is operable to process the selection of segment condition measures of the selection of pressure flow information data segments covering a range of pipeline locations and time periods in combination that are being processed to thendetermine the pipeline condition measure 2430. In this manner, the machine learning model has regard to both temporal and spatial information in its determination of the pipeline condition measure 2450.

[0420] Accordingly, in various examples segment condition measures in the form of signal intensity or frequency content measures that are susceptible to the disclosed processing include spectral power density (as a derivative of a spectrogram in the form of a power spectrum density) with either average or median values used per divided frame, signal correlation levels (where signal data from three or more sensing locations are available and processed) with either average or median values used per divided frames, signal coherence levels (where signal data from three or more sensing locations are available and processed) with either average or median values used per divided frame and / or other physical or statistical measures determined for the data (or processed data) in each divided frame including maximum, minimum, average, median, standard deviation and variance.

[0421] Referring now to FIG. 24B, there is shown a flowchart of a method 2450 for determining the condition of a pipeline in accordance with another illustrative embodiment. By way of overview, method 2450 comprises at step 2451 receiving pressure flow information corresponding to one or more pipeline locations on the pipeline. At step 2452, method 2450 comprises segmenting the pressure flow information to generate a selection of pressure flow information data segments corresponding to at least one or more pipeline locations determined at two or more time periods where each of the pressure flow information data segments corresponding to pressure flow information received for a unique time interval and pipeline location combination. In this example embodiment, at step 2453 a respective segment condition measure is determined for each of the pressure flow information data segments in the selection of pressure flow information data segments. This determination may be carried out in accordance with the various methods described in the present disclosure. Finally, at step 2454 a pipeline condition is determined measure based on the respective segment condition measures.

[0422] Referring now to FIG. 25, there is shown a flowchart of a method 2500 for pre-processing a pressure flow information data segment for a selected pipeline location to remove historical background noise data associated with the selected pipeline location according to an illustrative embodiment. As will be seen in the discussion below, a noise removal process in accordance with the present disclosure may be applied to any two sets of pressure flow information data.

[0423] In one example, the historical background noise data is based on noise data measured at the selected location in the pipeline supply network 100 corresponding to the location of the sensors. In another example, the historical background noise data may be based on one or more measurements taken from sensors located throughout the pipeline supply network 100. In another example, the historical background noise data is selected from a predetermined time period relative to the pressure flow information data segment.

[0424] Examples of this selection of historical background noise data may include, but not be limited to:• background noise data chosen from a period 30 days prior to the time period of the pressure flow information data segment as a result taking into account potential monthly variations in the background; or• background noise data chosen from a period 7 days prior to the time period of the pressure flow information data segment as a result taking into account potential weekly variations in the background.

[0425] In one example, the historical background noise data is taken from a time period substantially matching the time of day that corresponds to the time period of the pressure flow information data segment as a result taking into account potential daily variations in the background.

[0426] In another example, the historical background noise data is formed from a number of days of historical noise data. Examples include, but are not limited to:• historical noise data formed from the first 7 days or otherwise comprising a weekly characterisation of the pipeline supply network; or• historical noise data formed over the last 30 days or from the first 15 days of the last 30 days as a result providing a longer term characterisation of the pipeline supply network.

[0427] At step 2530, the spectral content difference is determined between the pressure flow information data segment 2520 and the historical background noise data 2510.

[0428] Referring now to FIG. 26, there is shown an example method 2530 for determining the spectral content difference corresponding to step 2530 of FIG. 26 according to an illustrative embodiment.

[0429] At step 2531, the spectral content of the historical background noise data is determined.

[0430] In one embodiment, historical background noise data xb(t) may comprise a vector xb(n) having time index n and containing Nrsamples as would be acquired by the sensors and associated DAQ system corresponding to a time period [0, T and where the spectral content is determined by forming a spectrogram or spectrum. As such, the relevant sampling frequency Fsis then N1 / T1.

[0431] In one example, the vector xb(n) comprises multiple different segments of historical acoustic signal data that characterise the pipeline supply network as discussed above which are then concatenated together to form an overall vector whose spectral content is to be determined. As an example, xb(ri) maybe formed to characterise the pipeline supply network over a particular time of day, a whole day, a week, a month or even a year.

[0432] The spectrogram SXbof xbis formed by first taking the discrete Fourier Transform using: Equation 29

[0433] of the consecutive short windows of the data which might be [% / >(l), xb(2),xb(3), , % / ,(M)] and then the window [% / ,(l + As), xb(2 + s),xb(3 + As), --- , xb(M + As)] and so on, with M being the length of the sliding window, and As being the step between consecutive windows. Then the absolute value of Xb( / <) is squared for all the consecutive short windows to yield the spectrogram SXb. Accordingly, the spectrogram SXbof xbwith window size M is then a matrix whose columns are the DFT of the windows xbwhere the rows of SXbare indexed by frequency and the columns are indexed by time.

[0434] A window size is determined that allows the spectral content to be characterised in time and depends on the sample frequency. In one non-limiting example, a window size of 1024 may be adopted where a sampling frequency of 4681Hz has been originally employed. As this process is not a coherent method, long windows are not necessarily required to capture the impulse response function and as such the window size is only required for resolving broad frequency structures in the spectrum.

[0435] At step 2532, the spectral content of the first time window of acoustic signal data is determined. By analogy, the first time window of acoustic signal data xp(t) may comprise a vector xp(n) having time index n and containing N2samples as would be acquired by the sensors and associated DAQ system corresponding to a time period [0, T2] and the spectral content is determined by forming a spectrogram or spectrum. In this case, the relevant sampling frequency Fsis then N2 / T2. As would be appreciated, interpolation techniques may be used to vary the size of vectors and the associated sampling frequencies as required.

[0436] The spectrogram SXpof xpis then formed by first taking the discrete Fourier Transform using: Equation 30

[0437] of the consecutive short windows of the data which might be [xp(1), xp(2), xp(3), • • • , xp(M)] and then the window [xp(1 + As), xp(2 + As),xp(3 + As), ••• , xpM + As)] and so on, with M being the length of the sliding window, and As being the step between consecutive windows. Then the absolute value of Xpk) is squared for all the consecutive short windows to yield the spectrogram SXp.Accordingly, the spectrogram SXpof xpwith window size M is then a matrix whose columns are the DFT of the windows xpwhere the rows of SXpare indexed by frequency and the columns are indexed by time.

[0438] At step 2533, the difference between the spectral content of the background acoustic signal data 2510 and the spectral content of the pressure flow information data segment 2520 is determined. In one example, the expected background spectrum / iXbis first calculated by normalising SXb. In one embodiment, this involves calculating the complex magnitude of the spectrogram SXbafter which the mean of each column is determined which is then used to scale the values in each of the columns of SXb. In another example, a low pass filter such with an auto-regressive filter may be applied to the values.

[0439] Once / Xbhas been determined, then the spectral content difference D ( / <) may be determined as follows: Equation 31

[0440] Referring back to FIG. 25, once the spectral content difference has been determined at step 2530, at step 2540 of FIG. 25 this determined spectral content difference is then removed from the pressure flow information data segment 2520 to generate the noise reduced pressure flow information data segment 2550.

[0441] Referring now to FIG. 27, there is shown an example method 2540 for removing the spectral content difference from the pressure flow information data segment corresponding to step 2540 of Figure 25 according to an illustrative embodiment.

[0442] At step 2541, the spectral content difference D ( / <) is applied in the frequency space to the spectral content of the pressure flow information data segment to generate a spectrally modified first time window Xs( / c) as follows:Xs(k) = max(Xp(k) - £>(k), 0) Equation 32

[0443] As the removal of spectral energy from Xp( / <) cannot result in energy levels that drop below 0 the “max” function is used to set an effective floor for Xs( / <).

[0444] At step 2542, the spectrally modified pressure flow information data segment is transformed to the time domain to generate the noise reduced pressure flow information data segment, ie xs. In one example, the inverse discrete Fourier Transform is applied to each column in the Xsk) to reconstruct inthe time domain the background benchmarked first time window, xs, as follows by adopting the windowing used previously: Equation 33

[0445] Referring now to FIG. 28, there is shown a flowchart of a method 4500 for benchmarking a pressure flow information data segment against an earlier pressure flow information data segment corresponding to an earlier period in accordance with an illustrative embodiment. In some examples, it may be appropriate to determine a segment condition measure for a given pressure flow information data segment following compensating for the difference between the given pressure flow information data segment and a pressure flow information data segment corresponding to an earlier time period but for the same location.

[0446] At step 4530, the spectral content difference is determined between the pressure flow information data segment for the selected location 4520 and the pressure flow information data segment for the selected location at an earlier period 4510. Once the spectral content difference has been determined at step 4530, at step 4540 of FIG. 28 this determined spectral content difference is then removed from the pressure flow information data segment 2520 to generate the benchmarked flow information data segment 4550.

[0447] As can be seen in FIG. 28, the process is similar to that shown in FIG. 25 except that here instead of one of the inputs being the background noise data for the selected location it is the pressure flow information data segment from the earlier time period.

[0448] Referring now to FIG. 29, there is shown a diagram 2900 depicting a series of pressure flow sensors that are spaced apart or offset from a pipeline 2910 in accordance with an illustrative embodiment. In this case, where the source of vibroacoustic energy is a leak 2915 on pipeline 2910 the transfer mechanism between the signal source 2915 and series of pressure flow sensors 2930 may be complex with direct paths, as well as potentially indirect paths as indicated. This is irrespective of the type of pressure flow sensor which in this example is a fibre based sensor 2940. In a relatively ideal arrangement (other than the most ideal being a continuous string of accelerometers, hydrophones and / or pressure transducers along a pipe or FBG or DAS fibre directly in contact with a pipe wall or inside a pipe), the dominant path for the transfer of vibroacoustic energy is the direct signal path through the intervening material (eg, soil or equivalent) between the pipeline at the location of the signal source (ie, the leak 2915) and the respective sensor 2930, as compared to along the pipeline via pipeline wall coupling and then indirect transmission through the intervening material to the pressure flow sensor 2930.

[0449] However, it is common for a non-ideal state to exist where either there is not a continuous string of accelerometers, hydrophones and / or pressure transducers along a pipe (which represents normal practice with spatially separate (sparse) point sensors) and / or not a fibre in contact with a pipe wall (unless retrospectively fitted), within a pipe (unless retrospectively fitted) and / or at a uniform parallel offset from a pipe and its direction.

[0450] Instead, in the case of accelerometers, hydrophones and / or pressure transducers, the pressure flow information sensing devices may be separated anywhere between 10 - 100 m, 100 - 1000 m or greater than 1000 m.

[0451] In the case of FBG or DAS fibre, which is not specifically fitted either in contact with a pipe wall, inside a pipe or at a uniform parallel offset from the fibre, and which is an in-situ dark or spare fibre that is being interrogated to obtain power level and frequency data as a form of pressure flow sensor, the spatial relationship between the fibre and pipeline 2910 being monitored is generally variable.

[0452] Referring now to FIG. 30A, there is shown a streetscape 3000 showing the configuration of a pipeline 3010 and a distributed pressure flow sensing apparatus 3040 in the form of a fibre and also individual pressure flow sensors in the form of accelerometers according to an illustrative embodiment. As can be seen by inspection, the pipeline 3010 and fibre 3040 can be distant, parallel or non-parallel, cross over or under each other, have both variable vertical and horizontal separation and come close or remain distant. Further, the pipeline 3010 can have connected pipes 3018 which intersect each other and come close to, far from or cross over / under in-situ fibre sensing systems.

[0453] In the case of a non-uniform spatial separation between a spatially distributed sensing system and the pipe or other element being monitored, the energy from a pressure flow information data source (eg, acoustic energy from a leak on a pipe) does not transmit uniformly, or sometimes even at all, between a pipe or other element and sensing fibre system due to damping and dispersion in the physical media between the pipe or other element and the sensing fibre. This is illustrated in FIG. 30A, with the acoustic energy from a pipe leak arriving at the closest fibre with variable signal power across the frequency spectrum due to varying direct distance between the leak on the pipe, varying indirect distance between leak acoustic energy emitting from locations along the pipe not at the leak location and fibre media energy absorption between the fibre and the pipe.

[0454] FIG. 30A shows that the fibre can be parallel to the pipe, non-parallel and / or cross over or under the pipe. Where the fibre crosses over or under the pipe the potential for energy transmission from the pipe to the fibre is increased and the location acts as a “hotpot” and a pseudo “point sensing” location (ie, similar to an installed accelerometer, hydrophone or pressure transducer). Noting also that the energy transmission from the pipe leak directly to the fibre is important where the fibre is close enough, and themedia between the pipe and fibre is transmissive enough. Energy transmission along the pipe is important where the fibre crosses the pipe at some point remote from the leak and the leak energy needs to be relayed along the pipe before it is able to be sensed by the fibre.

[0455] In the case of non-parallel and variable separations between sensed pipe(s) and sensors (including variable sensor placements along pipes) the propagation of acoustic energy along the pipe may be important. In the case of accelerometers, hydrophones and / or pressure transducers, the propagation of the frequency dependent power from a pressure flow data segment source (eg, pipe leak acoustic source) along the pipe and within the contained fluid is directly related to the typical detection and sensor spacing distances referred to above. Forms of symmetric (axial) and asymmetric (flexural) acoustic wave energy transmission occur along the coupled pipe wall-fluid (accelerometer sensitive) and within the pipe fluid (hydrophone sensitive) systems (with a mixture of waves but at low frequencies and high amplitudes for pressure / acoustic wave transmission).

[0456] In the case of FBG or DAS fibre sensing, where the direct propagation path from a pipe leak through media to an offset fibre is too far (and weak or no energy is sensed) the transmission of acoustic energy along the pipe (or pipes in a network) to a location where the fibre is closer and / or crosses over / under the pipe may become important. In this arrangement, the conceptual sensing action of the fibre changes from “parallel to pipe” continuous sensing to picking up pipe transmitted energy at a close approach and / or cross over point such. In this case, the fibre sensing is susceptible to analytic processing as if it were an accelerometer, hydrophone and / or pressure transducer with all methods of framing and analysis of the pressure flow data segments as described applicable interchangeably.

[0457] Beamforming is the process of phase shifting the signals from a series or array of sensors (eg, accelerometers, Fibre Bragg Gratings (FBGs) and / or Distributed Acoustic Sensing (DAS) channels along an optical fibre) until they are coherent (or reach maximum coherence) with one another for a selected location of interest along the series or array of sensors (eg, at or adjacent to the location of a leak along a pipe). In the example where a string of sensors is installed along or parallel to a pipe (at least for a while), and / or where a FBG or DAS fibre string comes close to and / or crosses over / under a pipe(s), beamforming may be applied to enhance the pressure flow information signal strength (using coherent noise processing methods) versus incoherent environment noise (forms of coherent environmental noise can also be separated).

[0458] Beam forming can increase the signal to noise ratio for a series or array of sensors by a maximum of 10 x loglO (N) if all the sensors are able to sense the source noise and the transmission and loss between the source and sensors can be fully accounted for (where N = the number of sensors).

[0459] In one example, referring again to FIG. 29, there is shown a series of pressure flow sensors spaced apart from a pipeline and operable to measure pressure flow information data from the pipeline.

[0460] In the case where there is a direct path through a media from a signal source to the series of pressure flow sensors, the phase shift is e~lkx, where x is the distance from a pressure flow sensor to the signal source and k is the wave number. The estimation of the signal source (eg, a leak) is then determined using:Equation 34

[0461] in which T is a transfer function between the signal source and the series of sensors.

[0462] The following approach is based on the analysis of the arrival of coherent signal energy directly through the intervening material from the leak location (as a point signal source). The coherent power from the leak signal source can be determined using the relationship below with the phase difference and shift incorporated in the determination of Gyy, weighting based on coherence incorporated through the |Gxy|2 / GxxGyy term, and the power then summed via the overall expression (below).

[0463] The location of interest for determining the coherent output power is x (noting that this can be moved along the pipeline). The coherent power from locations y to yNis added to provide the total at location x (the reference location). The approach has similarities to the use of the coherent power spectrum density (CPSD) method where phase shifting is employed to focus on coherent energy.

[0464] The coherence between two signals x(t) and y(t) may be defined as: Equation 35

[0465] which can be represented as shown below where y2yis the magnitude-squared coherence (or coherence function) between the signals as previously defined: Equation 36

[0466] The method can be applied to broader band leak noise with different levels of noise (at different frequencies) reaching different pressure flow sensors or sensing regions due to soil embedment conditions and the different phase shifts occurring across frequency (as also affected by the soil conditions).

[0467] This technique takes advantage of (ie, weights in favour of) correlated acoustic energy, as measured (and varying) along the series of sensors, after frequency dependent phase shifting and involves the following steps:• Determining the phase differences between the pressure flow information data segment (ie, corresponding to the reference sensor) and further pressure flow information data segments from the same or similar time period and corresponding to two or more of the remaining pipeline locations in the selection of pressure flow information data segments.• Phase shifting the further pressure flow information data segments to have the same phase as the pressure flow information data segment to form phase shifted further pressure flow information data segments.• Weighting the phase shifted further pressure flow information data segments in accordance with their level of coherence with the pressure flow information data segment to form weighted phase shifted further pressure flow information data segments.• Forming a beamformed pressure flow information data segment by summing in frequency space the weighted phase shifted further pressure flow information data segments to the pressure flow information data segment.

[0468] A further modification to the method has been developed such that energy is only accumulated once a minimum level of threshold coherence is reached.

[0469] The method involves phase shifting the pressure flow information data signals across the frequency range where the coherence is greater than a threshold (eg, 0.3) (and not phase shifting the nonreference pressure flow information data signals where their coherence with the reference signal is < 0.3). Phase shifted pressure flow information data signals are then again weighted by their level of coherence with the reference signal (> 0.3) and summed to obtain the accumulated correlated energy. This has the effect of focussing the process on frequency ranges with higher coherence and removing the influence of other low coherence signals.

[0470] Referring now to FIG. 30B, there is shown an experimental pipeline condition determining system 3050 comprising a distributed flow sensing apparatus 3060 according to an illustrative embodiment. In this example, distributed flow sensing apparatus 3060 comprises an arrangement of 24 pressure flow sensors (in this case FBGs) as indicated embedded in soil at 5.5 cm intervals between fibre gratings (individual pressure flow sensing locations) and located parallel to a 40 mm high density polyethylene (HDPE) pipe 3070 at an offset of 30 cm. The FBG grating locations are marked #1 through to #24 along with the location of a leak generated through a small ball valve located adjacent to the offset FBG #1.

[0471] Referring now to FIG. 31 A, there is shown a plot 3100 of the P SD following the beamforming process (“Beamformed”) applied with respect to pipeline condition determining system shown in FIG. 30B and having a leak of 0.07 L / s as indicated and adopting pressure flow sensor #1 as the reference.

[0472] Referring now to FIG. 3 IB, there is shown a plot 3150 of the PSD following the beamforming process (“Beamformed”) applied with respect to pipeline condition determining system shown in FIG. 30B and having a leak of 0.07 L / s as indicated and adopting pressure flow sensor #6 as the reference.

[0473] In this example, pressure flow sensor #1 (closest to the leak location) and pressure flow sensor #6 have been shown because, even though pressure flow sensor #1 is closest to the leak, pressure flow sensor #6 has stronger signal energy (clear in the lower frequency range around 100 Hz) due to soil embedment and acoustic noise transmission variabilities both at the individual pressure flow sensors and through the soil pathways between the leak signal source and each pressure flow sensor. This behaviour is characteristic of the type of variable signal transmission expected between a pressure flow sensor and offset pipe with a leak through the separating media between them.

[0474] The results from the application of the original beamforming method (a coherence cut off level of 0.3) are included for comparison (labelled as “Beamformed 2”).

[0475] The process of frequency dependent phase shifting of the signals from the non-reference pressure flow sensors (where the coherence between them and the reference is > 0.3) and weighting the signals by their coherence (when > 0.3) before summing, results in the reinforcement of coherent energy from the qualifying pressure flow sensors along the string (and suppression of non-coherent energy). This can be seen by inspection of FIG. 31 A and FIG. 3 IB for pressure flow sensor #1 and pressure flow sensor #6 references labelled “Beamformed” with greater variance in the coherently reinforced noise power levels across frequency after the non-coherent noise is removed.

[0476] Noise reduction (or signal enhancement) methods based on the use of the coherent components of a noise measured at multiple locations (via multiple devices), as distinct from beamforming reinforcement of coherent signals are known. These methods have been motivated by the need to reduce noise independently arising in measured signals at sensing locations (for either a single source noise or a noise source that is multi-source but coherent).

[0477] Referring now to FIG. 32, there is shown a data processing diagram of a method 3200 for isolating the coherent energy from a source and rejecting noise that is not coherent for a three sensor arrangement. As can be seen from method 3200, the contribution of the input x(t) to each of the three outputs can be calculated using the three output measurements (only). For example, with noise at thethree measurement locations due to x(t), the output for the second location (for example) can be calculated using: follows: Equatio _n 37

[0478] for which the various terms involved have been previously discussed above.

[0479] In one example, this method has been applied to the first three FBGs (#1, #2 and #3) in the 24 FBG string laboratory apparatus illustrated in FIG. 30B for a leak of 0.07L / s (the singular noise source). As previously discussed, different levels of noise (across different frequencies) reach the different FBGs (due to soil (sand) embedment conditions) with different levels of phase shifting and transformation. It is uncertain what proportion of independent noise combines with the weakening noise from the leak as the distance between the location of the leak and the FBGs along the string increases (ie, for the FBGs further from the leak location the signal is so weak it can be difficult to determine whether it is coherent with the locations closer to the leak or not).

[0480] Referring now to FIG. 33, there is shown a plot 3300 of the determined correlated energy 3310 for a selected pressure flow sensor (eg, FBG1) based on the requirement that it be also correlated to adjacent pressure flow sensors (eg, FBG2 an FBG3) for the pipeline condition determining system illustrated in FIG. 30B according to an illustrative embodiment. In this example, FIG. 33 shows reduction in the energy in FBG#1 (ie, 3310) based on the requirement that it also be coherent with energy in FBG#2 and FBG#3. The method can be extended to include more FBGs or pressure flow sensors along the pressure flow information sensing arrangement up to a limit beyond which the pressure flow information signals from the respective pressure flow information signals no longer contain coherent energy (at a measurable level).

[0481] In various examples, the modified beamforming (and signal enhancement) methods described above may be pre-applied to measured pressure flow information data from an arrangement of point pressure flow sensors (eg, accelerometers, hydrophones and / or pressure transducers), a FBG string (deployed inside, on the wall of, or offset parallel to a pipe) and / or a DAS system (spare or dark fibre type) with spatial offset from a pipe to improve the signal to noise ratio. The processed data may then be subject to the various processing methods in accordance with the present disclosure to determine initially the segment condition measures and then a pipeline condition measure based on the segment condition measures.

[0482] In the circumstances where a distributed pressure flow sensing apparatus is employed based on a fibre that is physically separated or offset from the pipeline of interest, then there may be a fundamental limit to vibroacoustic energy that can be transmitted through the intervening material ormedia between the pipeline and a respective sensing location on the fibre. This reduction in vibroacoustic signal will be dependent on the distance between the fibre and the pipeline and the energy absorbing characteristics of the intervening media between the pipeline and the sensing region of the fibre.

[0483] In one example, a vibroacoustic coupling member may be introduced between the pipeline and the fibre, where the vibroacoustic coupling member is configured to transfer vibroacoustic energy between the pipeline and the fibre. In various embodiments, the vibroacoustic coupling member may be configured to extend either wholly or partway through the intervening media between the pipeline and the sensing region or more generally the fibre.

[0484] In one example, the vibroacoustic coupling member is configured to transfer vibroacoustic energy from the pipeline through the intervening media to a location much closer to (say within 1 m of) a distributed pressure flow sensing apparatus in the form of a fibre based sensing FBG string and / or DAS system (dark fibre). As would be appreciated, the acoustic energy arriving at a point location nearer the sensing FBG string and / or DAS system (dark fibre) may be significantly increased relative to the acoustic energy that would otherwise arrive at the in-situ sensing fibres.

[0485] In one example, the vibroacoustic coupling member is configured as a thin metal bar or rod.

[0486] In another example, the vibroacoustic coupling member is a portion of fibre spliced into the original fibre and extending towards the pipeline.

[0487] Regardless of whether one or more physical acoustic energy transfer mechanisms are connected to a leaking pipe (to increase the magnitude of leak signal reaching in-situ fibre), or additional fibre is placed (to get closer to the leaking pipe and increase the magnitude of leak signal reaching the in- situ fibre), the result is that the acoustic energy reaching the fibre at the location of the physical transfer mechanism or closer fibre path (retrofitted) will be significantly stronger than at other locations along the leaking pipe. In some circumstances, no acoustic energy from the leaking pipe will reach the in-situ fibre until the location at which the physical transfer mechanism or closer fibre path (retrofitted) is reached. In this situation, the acoustic information arriving at the in-situ fibre will be more similar in nature to point excitation arrangement rather than continuous distributed sensing (DAS).

[0488] The deployment of point acoustic sensing devices (eg, accelerometers, hydrophones and / or pressure transducers) and measurement of acoustic energy in in-situ fibres, at locations where a physical transfer mechanism has been fitted between the pipe to a location near the in-situ fibre (or additional fibre has been spliced into in-situ fibre to reduce the distance between the leaking pipe and sensing fibre), share significant similarities. The methods described above may be applicable to the processing and interpretation of the physically collected pressure flow information data (using energy transfer fittingsand / or additional closer fibre) at the relevant discrete locations that are realised through the described installations.

[0489] Referring now to FIG. 34, there is a plan view of a spatial arrangement 3400 of vibroacoustic coupling members 3410A-3410D located between a pipeline 3450 and an adjacent non-parallel fibre based distributed flow sensing apparatus 3470 according to an illustrative embodiment. In this example physical pressure flow energy transfer mechanisms or vibroacoustic coupling members 3410 are in the form of a bar or rod) coupled in this example to the pipe wall at a respective pipeline coupling location (eg, “1”) and extending to within 1 m of in-situ fibre 3470 and defining a sensor coupling location (eg, “A”). The end of the bar or rod 3410 can be custom formed to enhance pressure flow energy (eg, leak acoustic energy) from it to the in-situ fibre 3470 through the separating media.

[0490] In one example, the separating media can remain undisturbed or be replaced with controlled energy transfer enhancing media for a 1 m by 1 m by depth to bar or road and fibre deep volume (typically 1000-2000 L or material to be removed and replaced).

[0491] A leak 3460 is shown in FIG. 34 along the pipe between the bar or rod to pipe coupling locations marked as 1 for the longest bar or rod 3410A and 2 for the second longest bar or rod 3410B, with these two bars or rods bracketing the leak location on the pipe (as the nearest coupled energy transfer mechanisms either side of the leak location). These bars or rods then transfer leak energy, with attenuation, from pipe coupling locations 1 and 2 to corresponding sensor coupling locations A and B on the in-situ sensing fibre 3470, after transmission along the bars / rods 3410A, 3410B, and through the media between the ends of the bars / rods and the in-situ fibre 3470.

[0492] Referring now to FIG. 35, there is shown a series of plots 3500 showing the relative frequency dependent attenuation (acoustic power loss) along a pipeline 3510 (Hl), a vibroacoustic coupling member 3520 (HA) and soil 3530 (HS) (eg, between the vibroacoustic coupling member and the fibre) according to an illustrative embodiment.

[0493] Referring now to FIG. 36, there is shown a flowchart of a method 3600 for determining the occurrence and location of a leak along the pipeline adopting one or more vibroacoustic coupling members (eg, 3410A-3410B) according to an illustrative embodiment.

[0494] At step 3610, determine a first transferred vibroacoustic signal at a first sensor coupling location corresponding to a first vibroacoustic coupling member. With reference to FIG. 34, transfer the time domain leak signal (LT) to the frequency domain (LF) via the application of a fast Fourier transform FFT and then apply the equation AF = LF * Hl * HA * HS to obtain the frequency domain leaksignal (AF) at the sensor coupling location corresponding to Point A and then transfer the signal back to the time domain (AT) using an inverse fast Fourier transform (IFFT).

[0495] At step 3620, determine a second transferred vibroacoustic signal at a second sensor coupling location corresponding to a second vibroacoustic coupling member. With reference to FIG. 34, repeat step 3610 to obtain the frequency domain leak signal (BF) at a second sensor coupling location point B and then the time domain leak signal (BT).

[0496] At step 3630, determine the level of correlation and coherence between the first and second transferred vibroacoustic signals. With reference to FIG. 34, determine the level of correlation and coherence between the signals A T and BT, as transmitted to the sensing in-situ fibre after routing from the pipe through the energy transfer mechanisms (eg, bars or rods coupled to the pipe wall) and the media between the end of the mechanisms and the in-situ fibre, and as determined from steps 3610 and 3620 respectively and diagnose the occurrence and location of the leak on the pipe by sensing through the in- situ fibre after signal enhancement via the energy transfer mechanisms illustrated.

[0497] Referring now to FIG. 37, there are shown plots 3700 of the original PSDs 3710, 3720 for first and second pipeline coupling locations and the respective PSDs of the transferred vibroacoustic signals 3715, 3725 (after loss and attenuation) as well as a plot 3740 of the coherence level between the transferred vibroacoustic signals and a plot 3750 of their phase relationship as determined in accordance with the method illustrated in FIG. 36. FIG. 37 shows the PSD as if the leak signal was measured on the pipe at location 1 and then after loss and attenuation as measurable at location A on the in-situ sensing fibre (both in top left plot), as well as the PSD as if the leak signal was measured on the pipe at location 2 and then after loss and attenuation as measurable at location B on the in-situ sensing fibre (both in the top right plot).

[0498] Referring now to FIG. 38, there is shown a plot 3800 of the correlation level between the transferred vibroacoustic signals as measured by the fibre based distributed pressure flow information sensing apparatus following energy transfer by respective acoustic coupling members according to an illustrative embodiment. As can be seen by inspection, the result of correlation of the in-situ sensing fibre measured pressure flow (eg, vibroacoustic) signals, after transmission (and attenuation) along the energy transfer mechanisms from the pipe to near the fibre (eg, bars or rods), is shown confirming that the leak can be diagnosed and located via the correlation result.

[0499] Methods and systems in accordance with the present disclosure operate to determine the condition of a pipeline. In various example, this comprises determining whether a structural anomaly is present in the pipeline. The term “anomaly” is taken to mean a non-standard operation, functioning or characteristic of the pipeline. An anomaly may comprise different types, including, but not limited to, aburst, a recently formed leak, the sudden closure or opening of a valve, an unusual high demand in the pipeline or the failure of a pump. Another example of an anomaly corresponding to a non-standard operation of the pipeline would be a firefighting equipment test procedure where valves and pumps designed to provide water in the event of a fire are tested. Other example structural anomalies include a circumferential crack in a pipe; a longitudinal crack in a pipe; a joint leak in a jointed pipe or a small blow-out in the pipeline.

[0500] An anomaly may also involve characteristics that characterise the anomaly. As a non-limiting example, an anomaly may be of the type “burst’ ’ and the associated anomaly characteristics may include the location of the “burst”, the size of the “burst” and the flow rate of fluid exiting the pipeline as a result of the “burst”.

[0501] As would be appreciated, method and systems for determining the condition of a pipeline in accordance with the present disclosure are directed to not only determining the condition based on temporal variations in measured pressure flow information from a pipeline but with combining this with assessing spatial variations in measure pressure flow information that occur at different pipeline locations. This provides an enhanced understanding of the pipeline behaviour as pipeline anomalies not only have an associated temporal behaviour but also can affect a pipeline along an extended length resulting in an associated spatial signature which can be identified in accordance with the present disclosure.Additionally, by determining the condition based on pressure flow information data corresponding to multiple pipeline locations this allows coherence, correlation and beamforming techniques to be used in accordance with the present disclosure both to enhance vibroacoustic signal corresponding to an anomaly and to indicate that there is a pipeline anomaly.

[0502] Those of skill in the art would appreciate that the various illustrative logical blocks, modules, circuits, and algorithm steps described in connection with the embodiments disclosed may be implemented as electronic hardware, computer software or instructions, or combinations of both. To clearly illustrate this interchangeability of hardware and software, various illustrative components, blocks, modules, circuits, and steps have been described above generally in terms of their functionality. Whether the disclosed functionality is implemented as hardware or software will depend upon the requirements of the application and design constraints imposed on the overall detection system. As an example, data processor 320 may encompass one or more separate applications on a single digital computer or processor or be distributed over a number of digital computers or processors depending on requirements.

[0503] The reference to any prior art in this specification is not, and should not be taken as, an acknowledgement or any form of suggestion that such prior art forms part of the common general knowledge.

[0504] It will be understood that the terms “comprise” and “include” and any of their derivatives (eg comprises, comprising, includes, including) as used in this specification, and the claims that follow, is to be taken to be inclusive of features to which the term refers, and is not meant to exclude the presence of any additional features unless otherwise stated or implied.

[0505] In some cases, a single embodiment may, for succinctness and / or to assist in understanding the scope of the disclosure, combine multiple features. It is to be understood that in such a case, these multiple features may be provided separately (in separate embodiments), or in any other suitable combination. Alternatively, where separate features are described in separate embodiments, these separate features may be combined into a single embodiment unless otherwise stated or implied. This also applies to the claims which can be recombined in any combination. That is a claim may be amended to include a feature defined in any other claim. Further a phrase referring to “at least one of’ a list of items refers to any combination of those items, including single members. As an example, “at least one of: a, b, or c” is intended to cover: a, b, c, a-b, a-c, b-c, and a-b-c.

[0506] It will be appreciated by those skilled in the art that the disclosure is not restricted in its use to the particular application or applications described. Neither is the present disclosure restricted in its preferred embodiment with regard to the particular elements and / or features described or depicted herein. It will be appreciated that the disclosure is not limited to the embodiment or embodiments disclosed, but is capable of numerous rearrangements, modifications and substitutions without departing from the scope as set forth and defined by the following claims.

Claims

CLAIMS1. A method for determining a condition of a pipeline, comprising: receiving pressure flow information corresponding to one or more pipeline locations on the pipeline; segmenting the pressure flow information to generate a selection of pressure flow information data segments corresponding to at least one or more pipeline locations determined at two or more time periods, each of the pressure flow information data segments corresponding to pressure flow information received for a unique time interval and pipeline location combination; determine a respective segment condition measure for each of the pressure flow information data segments in the selection of pressure flow information data segments; and determining a pipeline condition measure based on the respective segment condition measures.

2. The method of claim 1, wherein the pressure flow information corresponds to three or more pipeline locations on the pipeline, and wherein segmenting the pressure flow information comprises generating the selection of pressure flow information data segments to correspond to at least three or more pipeline locations determined at two or more time periods.

3. The method of claim 1 or 2, wherein determining a segment condition measure for the pressure flow information data segment comprises determining a signal intensity measure of the pressure flow information data segment.

4. The method of claim 3, wherein determining a signal intensity measure comprises determining a root-mean-square (RMS) value for the pressure flow information data segment.

5. The method of claim 3, wherein determining a signal intensity measure comprises: dividing the pressure flow information data segment into a number of frames; determining a respective RMS value for each frame of the number of frames; and determining a lower bound RMS value for which a selected percentage of the respective RMS values exceed.

6. The method of claim 1 or 2, wherein determining the segment condition measure for the pressure flow information data segment comprises determining a spectral / frequency content measure of the pressure flow information data segment.

7. The method of claim 6, wherein determining the spectral / frequency content measure comprises determining a power spectrum density (PSD) of the pressure flow information data segment.

8. The method of claim 6, wherein determining the spectral / frequency content measure comprises determining a spectral heat map of the pressure flow information data segment.

9. The method of claim 6, wherein determining the spectral content measure comprises determining a median frequency (MF) value for the pressure flow information data segment.

10. The method of claim 6, wherein determining the spectral content measure comprises: dividing the pressure flow information data segment into a number of frames; determining a respective MF value for each frame of the number of frames; and determining a lower bound MF value for which a selected percentage of the respective MF values exceed.

11. The method of claim 1 or 2, wherein determining the segment condition measure for the pressure flow information data segment comprises processing the pressure flow information data segment by a machine learning model trained to identify a pipeline anomaly condition.

12. The method of claim 11, wherein the machine learning model is an artificial neural network (ANN) or a recurrent neural network (RNN).

13. The method of claim 11, wherein processing the pressure flow information data segment by a machine learning model trained to identify a pipeline anomaly condition comprises: extracting one or more time domain and / or frequency domain features from the pressure flow information data segment; and processing the one or more extracted time domain and / or frequency domain features by a feature based machine learning classifier trained to identify the pipeline anomaly condition based on the one or more extracted time domain and / or frequency domain features.

14. The method of claim 13, wherein the feature based machine learning classifier is a decision tree model or a support vector machine.

15. The method of claim 11, wherein processing the pressure flow information data segment by a machine learning model trained to identify a pipeline anomaly condition comprises: forming a 2D spectrogram from the pressure flow information data segment; and processing the 2D spectrogram by a machine learning classifier trained to identify the pipeline anomaly condition based on the 2D spectrogram.

16. The method of claim 15, wherein the machine learning classifier is a convolutional neural network.

17. The method of claim 1 or 2, wherein determining the segment condition measure for the pressure flow information data segment comprises processing the pressure flow information data segment by a statistical model configured to identify a pipeline anomaly condition.

18. The method of claim 17, wherein processing the pressure flow information data segment by a statistical model configured to identify a pipeline anomaly condition comprises: extracting one or more time domain and / or frequency domain features from the pressure flow information data segment; and processing the one or more extracted time domain and / or frequency domain features by the statistical model configured to receive the one or more extracted time domain and / or frequency domain features as inputs.

19. The method of claim 2, wherein determining the segment condition measure for the pressure flow information data segment comprises: identifying the associated pipeline location and time period for the pressure flow information data segment; and determining respective coherence measures between the pressure flow information data segment and further pressure flow information data segments from the same or similar time period and corresponding to two or more of the remaining pipeline locations in the selection of pressure flow information data segments.

20. The method of claim 19, wherein one or more of the remaining pipeline locations correspond to adjacent pipeline locations to the associated pipeline location of the pressure flow information data segment.

21. The method of claim 19, wherein one or more of the remaining pipeline locations correspond to non-adjacent pipeline locations to the associated pipeline location of the pressure flow information data segment.

22. The method of any one of claims 1 to 21, wherein determining the pipeline condition measure based on the respective segment condition measures comprises: forming an assembly of segment condition measures ordered with respect to time and pipeline location.

23. The method of any one of claims 1 to 21, wherein determining the pipeline condition measure based on the respective segment condition measures comprises: determining whether one or more of the respective segment condition measures satisfy a threshold condition.

24. The method of any one of claims 1 to 21, wherein determining the pipeline condition measure based on the respective segment condition measures comprises: determining whether two or more of the respective segment condition measures corresponding to the same pipeline location satisfy a temporal rate of change condition.

25. The method of claim 24, wherein determining whether two or more of the respective segment condition measures corresponding to the same pipeline location satisfy a temporal rate of change condition comprises determining whether a temporal benchmark is exceeded.

26. The method of claim 24, wherein determining whether two or more of the respective segment condition measures corresponding to the same pipeline location satisfy a temporal rate of change condition comprises determining whether a temporal rate of change of power density in one or more selected frequency bands is exceeded.

27. The method of claim 24, wherein determining whether two or more of the respective segment condition measures corresponding to the same pipeline location satisfy a temporal rate of change condition comprises determining whether a temporal rate of change of coherence level is exceeded.

28. The method of any one of claims 2 to 21, wherein determining the pipeline condition measure based on the respective segment condition measures comprises: determining whether two or more of the respective segment condition measures corresponding to the same or similar time period satisfy a spatial rate of change condition.

29. The method of claim 28, wherein determining whether two or more of the respective segment condition measures corresponding to the same or similar time period satisfy a spatial rate of change condition comprises determining whether a spatial benchmark is exceeded.

30. The method of claim 28, wherein determining whether two or more of the respective segment condition measures corresponding to the same or similar time period satisfy a spatial rate of change condition comprises determining whether a spatial rate of change of power density in one or more selected frequency bands is exceeded.

31. The method of claim 28, wherein determining whether two or more of the respective segment condition measures corresponding to the same or similar time period satisfy a spatial rate of change condition comprises determining whether a spatial rate of change of coherence level is exceeded.

32. The method of any one of claims 2 to 21, wherein determining the pipeline condition measure based on the respective segment condition measures comprises:determining whether two or more of the respective segment condition measures corresponding to the same or similar time period and two or more of the respective segment condition measures corresponding to the same pipeline location satisfy a combined temporal / spatial rate of change condition.

33. The method of any one of claims 1 to 21, wherein determining the pipeline condition measure based on the respective segment condition measures comprises processing the combined respective segment condition measures by a machine learning model trained to identify a pipeline anomaly condition.

34. The method of claim 33, wherein processing the combined respective segment condition measures by a machine learning model trained to identify a pipeline anomaly condition comprises the machine learning model operating on an assembly of segment condition measures ordered with respect to time and pipeline location.

35. The method of any one of claims 1 to 34, further comprising pre-processing a pressure flow information data segment for a selected pipeline location to remove historical background noise associated with the selected pipeline location.

36. The method of any one of claims 1 to 34, further comprising pre-processing a pressure flow information data segment for a selected time period to remove historical background noise associated with the selected time period.

37. The method of any one of claims 1 to 34, further comprising benchmarking a pressure flow information data segment for a selected pipeline location against an earlier pressure flow information data segment associated with an earlier time period for the selected pipeline location.

38. The method of any one of claims 1 to 37, further comprising pre-processing a pressure flow information data segment for a selected pipeline location in accordance with a coherence analysis to compensate for differences in sensor gain and / or environmental noise.

39. The method of claim 38, wherein the coherence analysis comprises: processing the pressure flow information data segment to form a PSD; determining the maximum coherence level between the pressure flow information data segment and further pressure flow information data segments from the same or similar time period and corresponding to two or more of the remaining pipeline locations in the selection of pressure flow information data segments; and weighting the PSD with the maximum coherence level.

40. The method of claim 39, wherein the coherence analysis further comprises: normalising the PSD to form a normalised PSD; forming a filtered PSD by filtering the normalised PSD; and weighting the normalised PSD with the maximum coherence level.

41. A system for determining a condition of a pipeline, comprising: a pressure flow sensing arrangement configmed to measure pressure flow information from the pipeline; and one or more data processors configured to carry out the method of any one of claims 1 to 40.

42. A pipeline condition determining system comprising: one or more processors; and a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, wherein the one or more processors are enabled to implement the method according to any one of claims 1 to 40.

43. A non-transitory computer-readable medium comprising instructions stored thereon, that when executed on a processor, perform the steps of the method according to any one of claims 1 to 40.

Images