Detection of structural anomalies in a pipeline network
Methods of processing a data signal obtained from a sensor sensing a dynamic signal to detect a structural anomaly event are disclosed. In one embodiment, a method includes obtaining signal components attributable to fluid flow at a location within an operational pipeline network; processing the data signal to extract one or more features; characterising the one or more extracted features; and detecting an indication of a structural anomaly event proximal the location depending on the characterisation; wherein the structural anomaly event includes an occurrence and/or further development of a structural anomaly.
1 . A method of detecting an indication of a structural anomaly event in an operational mains water distribution pipeline network, the method including:
providing, by a sensor, a data signal including signal components caused by vibro-acoustic energy generated by fluid flow at a location within the operational pipeline network;
periodically sampling the data signal at a predetermined time interval to obtain a set of sampled data segments(S) for a time period;
processing the data segments to determine power spectral density values for each data segment;
processing at least the determined power spectral density values
using a machine learning model trained to
detect an indication of the structural anomaly event proximal the location and classify the structural anomaly event as either an occurrence and/or further development of a circumferential crack or a longitudinal crack, the machine learning model trained using temporal changes in one or more frequency domain features derived from power spectrum density values for previous time periods to distinguish between circumferential and longitudinal cracks; and
generating an alert in response to the machine learning model detecting an indication of a structural anomaly event, the alert identifying the location and classification of the structural anomaly event.
2 . A method according to claim 1 wherein the operational mains water distribution pipeline network includes a network of cast iron pipes, and wherein the alert indicates a required urgency of an operational response.
3 . A method according to claim 1 wherein the one or more frequency domain features include a statistical value.
4 . A method according to claim 1 wherein each data segment has an associated duration, and wherein the duration of each data segment is less than about 60 seconds.
5 . A method according to claim 4 wherein the duration of each data segment is less than about 30 seconds.
6 . A method according to claim 5 wherein the duration of each data segment is less than about 20 seconds.
7 . A method according to claim 6 wherein the duration of each data segment is less than about 15 seconds.
8 . A method according to claim 1 wherein the data signal is an acoustic signal and wherein each data segment is an uncompressed wave file.
9 . A method according to claim 1 wherein the one or more frequency domain features includes at least one of:
a value of a median frequency;
a value of a mean frequency;
a value of a peak frequency;
a standard deviation of a power spectral density;
a kurtosis of a power spectral density;
a skewness of a power spectral density;
a spectral flatness;
a spectral centroid;
a spectral spread; and
a distribution of the power spectral density values.
10 . A method according to claim 1 further comprising processing the data signal to extract one or more time domain features of the data signal and performing an analysis to characterise changes in at least one of the one or more time domain features.
11 . A method according to claim 1 wherein the processing of at least the determined power spectral density values by the machine learning model comprises one or more of:
a. generating a prediction for at least one of the one or more frequency domain features for comparison with an instance of the at least one frequency domain feature;
b. recognising a trend or pattern associated with at least one of the one or more frequency domain features over the time period as indicative of a structural anomaly event; and
c. identifying a predetermined behaviour and/or attribute of at least one of the at least one frequency domain features as indicative of a structural anomaly event.
12 . A method according to claim 1 , wherein processing the data segments to determine power spectral density values for each data segment includes:
framing each data segment as a sequence of data frames; and
processing a selection of the data frames for each data segment to determine separate respective sets of power spectral density values for each data segment.
13 . The method according to claim 12 wherein deriving one or more frequency-domain features from the power spectrum density values for previous time periods includes:
processing each separate set of power spectral density values to determine a respective first signal parameter value; and
processing a set of the determined first signal parameter values to determine the one or more frequency domain features.
14 . The method according to claim 13 wherein the first signal parameter includes a median frequency of one or more the power spectral densities derived from or for the data signal.
15 . The method of claim 14 , wherein the median frequency is a normalised median frequency.
16 . The method of claim 13 wherein the first signal parameter includes a root-mean-squared noise derived for or from the data signal.
17 . The method of claim 13 wherein processing a set of determined first signal parameter values to identify an indicator value includes identifying the indicator value as a first signal parameter value which n % of an ordered set of the first signal parameter values exceed.
18 . The method of claim 17 wherein n % is between about 50% and about 75%.
19 . The method of claim 17 wherein n % is between about 75% and about 90%.
20 . The method of claim 17 wherein n % is about 90%.
21 . The method according to claim 1 further including detecting a change in a profile of the distribution of the power spectral density values for a current data segment of the set as an indicator of a structural anomaly event proximal the location.
22 . A method according to claim 1 wherein the temporal changes include one or more new peaks in a power spectral density associated with the determined power spectral density values, each one or more new peak having a power density value above a threshold established using the power spectral density values from the previous time periods known to not be indicative of a circumferential or longitudinal crack.
23 . A method according to claim 1 further comprising validating the detection by inspection of the mains water distribution pipeline network at the location and retraining the machine learning model depending on the presence of a circumferential or longitudinal crack at the location to improve the accuracy of the detection.
24 . A system for processing a data signal obtained from a sensor sensing a dynamic signal including signal components caused by vibro-acoustic energy generated by fluid flow at a location within an operational mains water distribution pipeline network to detect an indication of a structural anomaly event, the system including:
a memory;
a set of computer readable instructions stored in the memory;
a processor coupled to the memory, the processor configured to execute the set of computer readable instructions and use a machine learning model trained to detect an indication of the structural anomaly event proximal the location and classify the structural anomaly event as either an occurrence and/or further development of a circumferential crack or a longitudinal crack, to:
periodically sample the data signal at a predetermined time interval to obtain a set of sampled data segments(S) for a time period;
process the data segments to determine power spectral density values for each data-segment;
process at least the determined power spectral density values using the neural network; and
generate an alert in response to the machine learning model detecting an indication of a structural anomaly event, the alert identifying the location and classification of the structural anomaly event,
wherein the machine learning model is trained using temporal changes in one or more frequency domain features derived from power spectrum density values for previous time periods to distinguish between circumferential and longitudinal cracks.
25 . An apparatus for processing a data signal obtained from a sensor sensing a dynamic signal including signal components caused by vibro-acoustic energy generated by fluid flow at a location within an operational mains water distribution pipeline network to detect an indication of a structural anomaly event, the apparatus including:
means for periodically sampling the data signal at a predetermined time interval to obtain a set of sampled data segments (S) for a time period;
means for processing the data segments to determine power spectral density values for each data segment;
means for processing at least the determined power spectral density values using a machine learning model trained to detect an indication of the structural anomaly event proximal the location and classify the structural anomaly event as either an occurrence and/or further development of a circumferential crack or a longitudinal crack, the machine learning model trained using temporal changes in one or more frequency domain features derived from power spectrum density values for previous time periods to distinguish between circumferential and longitudinal cracks; and
means for generating an alert in response to the machine learning model detecting the indication of a structural anomaly event, the alert identifying the location and classification of the structural anomaly event.
26 . A non-transitory computer-readable storage device storing instructions that, when executed by a processor, cause the processor to perform operations comprising:
obtaining a data signal from a sensor sensing a dynamic signal including signal components caused by vibro-acoustic energy generated by to fluid flow at a location within an operational pipeline network;
processing the data signal to determine power spectral density values for one or more frequency bands of the data signal;
processing at least the determined power spectral density values using a machine learning model trained to detect an indication of the structural anomaly event proximal the location and classify the structural anomaly event as either an occurrence and/or further development of a circumferential crack or a longitudinal crack, the machine learning model trained using temporal changes in one or more frequency domain features derived from power spectrum density values for previous time periods to distinguish between circumferential and longitudinal cracks; and
generating an alert in response to the machine learning model detecting an indication of a structural anomaly event, the alert identifying the location and classification of the structural anomaly event.