Pipeline leakage noise positioning analysis method based on deep learning

Abstract

The application relates to the technical field of pipeline leakage. A pipeline leakage noise positioning analysis method based on deep learning comprises the following steps: collecting data of healthy pipelines and leakage pipelines respectively to obtain a healthy pipeline noise data set and a leakage pipeline noise data set; using a convolutional neural network to extract features of sound samples in the healthy pipeline noise data set and the leakage pipeline noise data set respectively, and classifying the sound samples to obtain healthy pipeline feature data, leakage detection feature data and positioning feature data; using the healthy pipeline feature data, the leakage detection feature data and the positioning feature data to train a deep learning model to obtain a target deep learning model; and using the target deep learning model to monitor a current pipeline to obtain a monitoring result of the current pipeline, wherein the monitoring result at least comprises a judgment of whether leakage exists and positioning information of the leakage. Accurate leakage positioning information is directly outputted, and false positives, false negatives and positioning errors are effectively reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of pipeline leakage, in particular to a pipeline leakage noise positioning analysis method based on deep learning. BACKGROUND

[0002] Real-time detection and accurate positioning of pipeline leakage are the key to ensuring transportation safety. At present, passive detection method based on acoustic or vibration signals is one of the mainstream technologies in this field. The principle is that when a pipeline leaks, the broadband noise and vibration generated by the fluid spouting will propagate along the pipe wall to both ends. Through the sensor array arranged along the pipeline, these signals are captured, and then the leakage position is calculated. However, the actual application effect of this method is seriously restricted by a series of inherent bottlenecks. First of all, the core of traditional positioning algorithm relies on the time difference of signal arrival at different sensors, and geometric calculation is carried out through a simplified sound wave propagation speed model. In the actual complex pipe network environment, the sound wave propagation speed is significantly affected by pipe material, pressure, medium, and even external coating, which is difficult to accurately model, resulting in positioning error often reaching tens or even hundreds of meters, which cannot meet the demand of accurate excavation and repair.

[0003] Secondly, this method is vulnerable when facing complex environmental noise. The background noise (such as traffic vibration, pump operation, construction interference) of urban underground pipe network or industrial site has complex spectrum and variable intensity, which is easily confused with real leakage signal. Traditional methods usually rely on manually set fixed threshold or simple frequency band based filtering to distinguish noise and leakage signal, which has high false positive rate and false negative rate in non-stationary noise environment, seriously affecting the reliability and credibility of the monitoring system. SUMMARY

[0004] The main purpose of the present application is to provide a pipeline leakage noise positioning analysis method based on deep learning to solve the problem of poor existing pipe network monitoring effect.

[0005] In order to achieve the above purpose, the first aspect of the present application provides a pipeline leakage noise positioning analysis method based on deep learning, comprising:

[0006] Data acquisition is performed on healthy pipelines and leaking pipelines respectively to obtain healthy pipeline noise data set and leaking pipeline noise data set, wherein the samples obtained by data acquisition on the healthy pipelines and the leaking pipelines include sound samples of pipe diameter, pressure, material and environmental conditions;

[0007] Feature extraction is performed on the sound samples in the healthy pipeline noise data set and the leaking pipeline noise data set respectively by using convolutional neural network, and classification is performed to obtain healthy pipeline feature data, leakage detection feature data and positioning feature data;

[0008] training a deep learning model by using the health pipeline feature data, the leakage detection feature data and the positioning feature data to obtain a target deep learning model;

[0009] monitoring the current pipeline by using the target deep learning model to obtain a monitoring result of the current pipeline, wherein the monitoring result at least includes a judgment of whether there is leakage and positioning information of leakage occurrence.

[0010] In some implementable manners, the step of respectively collecting data of the health pipeline and the leakage pipeline to obtain a health pipeline noise data set and a leakage pipeline noise data set includes:

[0011] collecting data during operation of the health pipeline to obtain the health pipeline noise data set, wherein the collecting data during operation of the health pipeline includes passive collection of the health pipeline and active collection of the health pipeline, corresponding to passive health pipeline noise data and active health pipeline noise data, the passive health pipeline noise data representing data collected from the health pipeline in continuous time, and the active health pipeline noise data representing data collected from the health pipeline at a preset time interval;

[0012] collecting data of the leakage pipeline to obtain the leakage pipeline noise data set, wherein the leakage pipeline noise data set includes a set of leakage pipeline noise data collected by passive collection of the leakage pipeline, and the passive collection of the leakage pipeline represents data collected from the leakage pipeline in continuous time.

[0013] In some implementable manners, the step of training a deep learning model by using the health pipeline feature data, the leakage detection feature data and the positioning feature data to obtain a target deep learning model includes:

[0014] constructing and training a health pipeline learning model to obtain a target health pipeline learning model, wherein the health pipeline learning model is an unsupervised model based on an autoencoder, representing training by using passive health pipeline noise data, for calculating reconstruction error of input noise data, and capable of outputting an abnormal attention map of noise data, the abnormal attention map being a multi-dimensional matrix representing abnormal degree of time, frequency and space dimensions;

[0015] constructing and training a leakage pipeline learning model to obtain a target leakage pipeline learning model, wherein the leakage pipeline learning model is a supervised model based on a deep neural network, representing training by using leakage detection feature data and positioning feature data, for judging whether there is leakage and outputting leakage position information;

[0016] combining the health pipeline learning model and the leakage pipeline learning model in sequence to obtain the deep learning model;

[0017] combining the target health pipeline learning model and the target leakage pipeline learning model in sequence to obtain a target deep learning model, wherein when the target health pipeline learning model determines that the state of the current pipeline is abnormal, it outputs a trigger to start the target leakage pipeline learning model.

[0018] In some implementable manners, the step of monitoring the current pipeline using the target deep learning model to obtain a monitoring result of the current pipeline comprises:

[0019] For the current pipeline, a sensor is used to implement passive collection of original pipeline passive noise data, and an active acoustic emission instrument is used to implement active collection of original pipeline active noise data, and data synchronization, denoising and normalization processing are performed to form preprocessed pipeline passive noise data and preprocessed pipeline active noise data;

[0020] The preprocessed pipeline passive noise data in a continuous time period is input into the target health pipeline learning model to obtain a reconstruction error of the current pipeline, and a spatial feature distance calculation is performed between the reconstruction error and a reference value of the health pipeline to obtain a pattern novelty index sequence;

[0021] The preprocessed pipeline active noise data in a continuous time period is compared with an active detection reference value of the health pipeline to obtain a dynamic restoring force index sequence;

[0022] According to the pattern novelty index sequence and the dynamic restoring force index sequence, risk scores are respectively calculated to obtain a comprehensive risk score.

[0023] In some implementable manners, the step of monitoring the current pipeline using the target deep learning model to obtain a monitoring result of the current pipeline further comprises:

[0024] When the comprehensive risk score exceeds a first threshold value, a risk situation deep analysis step is performed:

[0025] According to the pattern novelty index sequence and the dynamic restoring force index sequence, sliding window time series analysis is performed to obtain a change trend and a curvature of the pattern novelty index sequence and a change trend and a curvature of the dynamic restoring force index sequence;

[0026] According to the pattern novelty index sequence and the dynamic restoring force index sequence, a correlation coefficient is calculated to obtain a trend of cooperative rise or fall of the pattern novelty index sequence and the dynamic restoring force index sequence;

[0027] According to the trend of the index of adjacent sensors in the current pipeline, the spatial consistency of the mode novelty index sequence and the dynamic resilience index sequence is obtained;

[0028] According to the change trend and curvature of the mode novelty index sequence, the change trend and curvature of the dynamic resilience index sequence, the trend of the mode novelty index sequence and the dynamic resilience index sequence rising or falling together, and the spatial consistency of the mode novelty index sequence and the dynamic resilience index sequence, a risk situation description is obtained.

[0029] In some implementable manners, the step of monitoring the current pipeline by using the target deep learning model to obtain the monitoring result of the current pipeline further includes:

[0030] According to the risk situation description, a leakage prediction level assessment of the current pipeline is performed to obtain a leakage risk level; when the leakage risk level is greater than or equal to a preset leakage risk level, the target leakage pipeline learning model is started;

[0031] The passive noise data of the current pipeline in the latest continuous time window when the risk situation description is formed is obtained as to-be-analyzed data;

[0032] From the target healthy pipeline learning model, the abnormal attention map corresponding to the to-be-analyzed data is obtained;

[0033] According to the abnormal attention map, one or more key time segments with an abnormal attention value greater than or equal to a preset threshold are selected from the to-be-analyzed data;

[0034] The noise data corresponding to one or more key time segments is input into the target leakage pipeline learning model for leakage judgment to obtain the monitoring result of the current pipeline, wherein the monitoring result at least includes a judgment of whether there is leakage and positioning information of leakage occurrence.

[0035] In some implementable manners, the step of selecting one or more key time segments with an abnormal attention value greater than or equal to a preset threshold from the to-be-analyzed data according to the abnormal attention map includes:

[0036] The abnormal attention values of the abnormal attention map in the time dimension are integrated or clustered for analysis to identify an abnormal attention continuous period;

[0037] The abnormal attention continuous period greater than or equal to the preset threshold is taken as the key time segment.

[0038] In a second aspect, the application provides a pipeline leakage noise positioning analysis system based on deep learning, which is applied to the pipeline leakage noise positioning analysis method based on deep learning described above, and comprises:

[0039] A collection unit is configured to collect data of healthy pipelines and leakage pipelines respectively to obtain a healthy pipeline noise data set and a leakage pipeline noise data set, wherein the samples obtained by collecting data of the healthy pipelines and the leakage pipelines include sound samples of pipe diameters, pressures, materials and environmental conditions;

[0040] A feature extraction unit is configured to extract features of the sound samples in the healthy pipeline noise data set and the leakage pipeline noise data set respectively by using a convolutional neural network, and classify the sound samples to obtain healthy pipeline feature data, leakage detection feature data and positioning feature data;

[0041] A model training unit is configured to train a deep learning model by using the healthy pipeline feature data, the leakage detection feature data and the positioning feature data to obtain a target deep learning model;

[0042] A result unit is configured to monitor a current pipeline by using the target deep learning model to obtain a monitoring result of the current pipeline, wherein the monitoring result at least includes a judgment of whether there is leakage and positioning information of the leakage.

[0043] In a third aspect, the application provides a computer storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps of the pipeline leakage noise positioning analysis method based on deep learning described above.

[0044] In a fourth aspect, the application provides a computer program, which is executed by a processor to implement the steps of the pipeline leakage noise positioning analysis method based on deep learning described above.

[0045] The technical scheme provided by the embodiments of the application can have the following beneficial effects:

[0046] In the pipeline leakage noise positioning analysis method based on deep learning, a data-driven deep learning framework is adopted, a convolutional neural network is used to automatically extract and classify acoustic samples covering multiple working conditions to obtain distinctive feature data, and a special deep learning model is trained. This method replaces the steps of signal discrimination by manually setting thresholds and position calculation by simplified physical models in the traditional method, so that the leakage event can be more reliably distinguished from complex background noise, and more accurate leakage positioning information can be directly output, thereby effectively reducing false positives, false negatives and positioning errors. BRIEF DESCRIPTION OF DRAWINGS

[0047] The accompanying drawings, which form a part of this application, are included to provide a further understanding of the application and are incorporated in and constitute a part of this application. The illustrations, together with their description, serve to explain the application. In the drawings:

[0048] Figure 1 A flow chart of a pipeline leakage noise positioning analysis method based on deep learning provided in the present application. DETAILED DESCRIPTION

[0049] In order to enable persons skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by persons skilled in the art without creative labor should fall within the scope of protection of the present application.

[0050] It should be noted that the terms “first”, “second”, and the like in the specification of the present application, the claims, and the above-described drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or a chronological sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented. In addition, the terms “include” and “have” and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but can include other steps or units that are not clearly listed or inherent to these processes, methods, products, or devices.

[0051] In the present application, the terms “up”, “down”, “left”, “right”, “front”, “back”, “top”, “bottom”, “inner”, “outer”, “middle”, “vertical”, “horizontal”, “lateral”, “longitudinal”, and the like indicate the orientation or positional relationship shown in the drawings. These terms are mainly used to better describe the present application and its embodiments, and are not intended to limit the indicated devices, elements or components to have a specific orientation, or to be constructed and operated in a specific orientation.

[0052] In addition, in addition to indicating the orientation or positional relationship, some of the above-mentioned terms can also be used to indicate other meanings, for example, the term “up” may also be used to indicate a certain attachment relationship or connection relationship in some cases. For persons skilled in the art, the specific meaning of these terms in the present application can be understood according to the specific circumstances.

[0053] In addition, the terms "mounting", "arrangement", "provided with", "connected", "linked", "sleeved" should be broadly interpreted. For example, "connected" can be fixed connection, detachable connection, or integral structure; can be mechanical connection, or electrical connection; can be direct connection, or indirect connection through intermediate medium, or internal communication between two devices, elements or components. The specific meaning of the above terms in the present application can be understood by the person skilled in the art according to the specific circumstances.

[0054] The terms related to the present application are explained below to understand the technical solutions of the present application:

[0055] The novelty index of mode is a quantitative index for evaluating the abnormal mode caused by the current data, and the strangeness or novelty degree compared with the known healthy baseline mode of the system. It does not focus on the overall strength of the anomaly, but on whether the internal distribution form in the feature space is unprecedented.

[0056] The dynamic resilience index is an index for quantitatively evaluating the deviation degree of the dynamic mechanical characteristics of the pipeline (or general structure) from the baseline state after being subjected to standard excitation. It directly reflects the changes of the internal properties such as integrity, stiffness or damping of the pipeline structure itself by comparing the current active detection response with the healthy baseline response.

[0057] As shown in Figure 1 The first aspect, the present application provides a pipeline leakage noise positioning analysis method based on deep learning, comprising:

[0058] S100, respectively, data acquisition is carried out on the healthy pipeline and the leakage pipeline, and the healthy pipeline noise data set and the leakage pipeline noise data set are obtained.

[0059] Among them, the samples obtained by data acquisition on the healthy pipeline and the leakage pipeline include sound samples of pipe diameter, pressure, material and environmental conditions.

[0060] It should be noted that the purpose of S100 step is to collect sound samples of pipeline under different internal states (healthy / leakage) and different external operating conditions, so as to ensure the diversity and representativeness of the data set, so that the model trained subsequently has good generalization ability. The sound samples need to be associated with the data of pipe diameter, internal pressure, pipeline material and external environmental conditions corresponding to the generation of the sound samples, which will be used as key features to describe the samples.

[0061] Specifically, obtaining the healthy pipeline noise data set and the leakage pipeline noise data set can include the following steps:

[0062] S101, collecting data during the health pipeline operation to obtain the health pipeline noise data set.

[0063] Wherein, the collecting data during the health pipeline operation includes health pipeline passive collection and health pipeline active collection, corresponding to form health pipeline passive noise data and health pipeline active noise data, the health pipeline passive noise data represents the data collected from the health pipeline in continuous time, and the health pipeline active noise data represents the data collected from the health pipeline at a preset time interval.

[0064] Specifically, the step S101 aims to systematically establish the acoustic reference of the pipeline in the normal state without leakage. To ensure that the subsequent deep learning model can not only sensitively identify leakage anomalies, but also effectively distinguish the interference caused by normal working condition fluctuations, the collection work adopts a cooperative collection mode combining passive and active collection.

[0065] Health pipeline passive collection: this mode aims to record the sound generated by the pipeline in the natural running state. Specifically, during the normal transportation of the pipeline (such as water flow), acoustic or vibration sensors are used to continuously record at a pre-set fixed high sampling rate (for example, 44.1 kHz or higher) at the sensor installation point. The health pipeline passive noise data formed by the collection is essentially a time series of wideband continuous sound waves and vibration signals generated by the fluid flow in the pipeline, the operation of equipment (such as pumps, valves), and the coupling effect with the environmental background (such as soil friction, distant traffic). The health pipeline passive noise data constitutes a reference library representing the normal acoustic background of the pipeline under various flow rates, pressure and other operating conditions.

[0066] Healthy pipeline active acquisition: This mode aims to provide a physical probe independent of the fluid working condition to verify whether the pipeline structure state has changed. Specifically, the healthy pipeline is intermittently or synchronously passively acquired at preset regular time intervals, a standard, known characteristic short excitation signal (such as a pulse) is applied to the pipeline by the active acoustic emission instrument, and the transient response of the pipeline is synchronously acquired. The healthy pipeline active noise data formed thereby is essentially the impulse response function of the pipeline structure in the current state. The core purpose is that when the healthy pipeline passive noise data appears abnormal in subsequent monitoring, the difference is compared by comparing the reference value (the reference value is a reference standard previously established by statistically analyzing (such as calculating the mean value, standard deviation, or extracting the typical range of the feature vector) the data obtained by multiple active acquisitions of the pipeline in the healthy state) of the healthy pipeline active noise data at this time, to determine whether the abnormality is caused by physical changes of the pipeline structure (such as the appearance of a leakage crack, whose response will change) or only by non-structural operating parameter fluctuations (such as valve adjustment, whose response is basically unchanged), thereby assisting in verifying or correcting the suspected leakage judgment based on passive noise alone to reduce false positives.

[0067] Need to be explained is that for the acquisition of healthy pipeline active noise data, it can be based on the aforementioned preset regular time interval, and on the basis of the preset regular time interval, when the healthy pipeline passive noise data appears abnormal, a signal is sent to the active acoustic emission instrument, so that the active acoustic emission instrument responds to the acquisition of the healthy pipeline at the time to form the healthy pipeline active noise data, the healthy pipeline active noise data, and the current healthy pipeline active noise data is compared with the reference value to determine the cause of the abnormality of the healthy pipeline passive noise data.

[0068] S102, data acquisition is performed on the leakage pipeline to obtain a leakage pipeline noise data set.

[0069] The leakage pipeline noise data set includes a collection of leakage pipeline noise data passively acquired by the leakage pipeline. The leakage pipeline passive acquisition means that data is acquired from the leakage pipeline over a continuous period of time.

[0070] Specifically, the purpose of step S102 is to construct a sample library for training a model to identify leaks by simulating or recording leakage states at different positions and different leakage diameters on a controllable experimental pipeline or an in-service pipeline that has been confirmed to have a leak.

[0071] The passive collection of the leaking pipeline is to perform the same operation as the passive collection mode in S101 under the set or confirmed leakage condition, that is, to continuously collect the acoustic or vibration signals during the operation of the leaking pipeline. The passive noise data of the leaking pipeline thus formed has the core feature that the noise generated by the leakage is superimposed on the background noise. When constructing the data set, each noise data sample of the leaking pipeline must be associated with its corresponding true leakage working condition label. For example, "when collecting the noise data of the leaking pipeline, record and associate the corresponding true leakage position, leakage aperture (or leakage rate), and pipeline pressure at the time of collection, and other label information. These labels will serve as the true values of the supervised training model.

[0072] S200, using a convolutional neural network, respectively extracts features from the sound samples in the healthy pipeline noise data set and the leaking pipeline noise data set, and classifies to obtain healthy pipeline feature data, leakage detection feature data, and positioning feature data.

[0073] Specifically, using an existing convolutional neural network, the sound samples in the healthy pipeline noise data set and the leaking pipeline noise data set are respectively extracted and classified, that is, the noise data formed is not noise formed by a single signal, but a mixture of acoustic signals generated by multiple physical processes. For example, the noise data can include fluid dynamics noise: turbulent flow, cavitation sound generated by the flow of the medium (water, oil, gas) in the pipeline; equipment operation noise: mechanical vibration and electromagnetic noise inherent to pumps, compressors, valves, etc.; structural vibration noise: structural sound generated by the pipeline under external excitation (such as soil pressure, traffic) or internal fluid impact; environmental background noise: interference from the environment outside the pipeline (such as wind, rain, distant construction); abnormal event noise: injection, turbulent flow, cavitation mixed sound generated by leakage, as well as friction, impact and other abnormal sound. Therefore, it is necessary to classify these non-single signals in the noise data to obtain classified acoustic signals, and further obtain healthy pipeline feature data, leakage detection feature data, and positioning feature data from the classified acoustic signals, wherein the healthy pipeline feature data represents the features extracted from the healthy pipeline noise data set using the convolutional neural network and classified according to the extracted features. In addition, the leakage detection feature data and the positioning feature data represent the features extracted from the leaking pipeline noise data set using the convolutional neural network and classified according to the extracted features.

[0074] It should be noted that from the healthy pipeline noise data set, the convolutional neural network learns and separates the features of all noise components that may occur under normal operating conditions, such as smooth flow sound, pump fundamental frequency and harmonics, regular valve action sound, etc. The total representation of these features is summarized as healthy pipeline feature data. This constitutes a normal noise component dictionary.

[0075] From the leaky pipeline noise data set, the convolutional neural network learns and separates the characteristics of the abnormal noise components generated by the specific physical process of "leakage" in the leaky pipeline noise data set. According to different uses of these characteristics, further organization into two categories:

[0076] Leak detection feature data: This is the most representative of the essence of the leak attribute, and can distinguish it from healthy noise components and other abnormalities (such as non-leakage impact). For example, features sensitive to the combination of persistent broadband random characteristics and specific resonance modes.

[0077] Positioning feature data: This is a set of features that can reflect the spatial propagation properties of leak noise components. For example, features sensitive to the correlation between different sensor signal pairs, time delay differences, which directly encode the location information of the sound source.

[0078] The healthy pipeline feature data, leak detection feature data, and positioning feature data are all high-dimensional feature vectors extracted by the convolutional neural network. They will be used as inputs for training the healthy pipeline learning model and the leaky pipeline learning model in subsequent steps.

[0079] S300, training a deep learning model using the healthy pipeline feature data, the leak detection feature data, and the positioning feature data to obtain a target deep learning model.

[0080] In the S300 step, the previously extracted healthy pipeline feature data, leak detection feature data, and positioning feature data are used to train the deep learning model through a "module training + sequential combination" approach, ultimately obtaining a target deep learning model that can achieve "first determine pipeline abnormalities, then identify leaks and locate them."

[0081] Specifically, obtaining the target deep learning model can include the following steps:

[0082] S301, build a healthy pipeline learning model and train it to obtain a target healthy pipeline learning model.

[0083] The healthy pipeline learning model is an unsupervised model based on an autoencoder, which is trained using passive noise data from a healthy pipeline. It is used to calculate the reconstruction error of the input noise data and can output an anomaly attention map for the noise data. The anomaly attention map is a multi-dimensional matrix representing the degree of anomaly in time, frequency, and space dimensions.

[0084] Specifically, the health pipeline learning model is an unsupervised learning model based on an autoencoder, wherein the autoencoder refers to encoding compression first and decoding reconstruction, the model automatically learns the inherent law of the health pipeline feature data, compresses the input health feature data into low-dimensional feature codes, and restores the reconstruction data from the feature codes as close as possible to the input data. The purpose of the unsupervised learning model is to complete the training without manual labeling of normal / abnormal labels, and only rely on the law of the health pipeline passive noise data itself.

[0085] The health pipeline feature data obtained by feature extraction on the health pipeline passive noise data in the S200 step covers health features under different pipe diameters, pressures, materials and environmental conditions, ensuring that the model learns the health data law in all scenarios. The health pipeline learning model is trained to enable the health pipeline learning model to accurately reconstruct the health pipeline feature data, so that the reconstruction error (difference between the input data and the reconstructed data) after the health pipeline feature data is input is in a low and stable range, and the health pipeline learning model has the ability to output an "abnormal attention map".

[0086] It should be noted that the reconstruction error represents the deviation of the input data from the health baseline. If the input is health data, the reconstruction error is small, and if the input contains abnormal features such as leakage, the reconstruction error will significantly exceed the health baseline range. The abnormal attention map is a three-dimensional matrix (time dimension × frequency dimension × space dimension), and each value in the matrix represents the abnormality of the corresponding dimension (the higher the value, the more the signal at the time point, the frequency band, and the sensor collection position deviates from the health state).

[0087] Exemplarily, the training process is as follows:

[0088] The health pipeline feature data is input into the autoencoder model (health pipeline learning model) in batches;

[0089] The parameters of the autoencoder model are iteratively adjusted until the reconstruction error of the health pipeline feature data converges to a preset low threshold (such as the error mean being stable within 0.01);

[0090] Based on the trained target health pipeline learning model, the "baseline value" (i.e., the average reconstruction error of health data) is solidified, and the generation logic of the abnormal attention map is configured (the target health pipeline learning model automatically identifies the dimensions with large reconstruction errors and labels the abnormality). The target health pipeline learning model is used to determine whether the input data deviates from the health state, and does not directly determine whether there is leakage.

[0091] S302, construct a leakage pipeline learning model and train to obtain a target leakage pipeline learning model.

[0092] The leakage pipeline learning model is a supervised model based on a deep neural network, indicating that the leakage detection feature data and the positioning feature data are trained to determine whether there is leakage and output leakage position information.

[0093] Specifically, the leakage pipeline learning model is a supervised learning model based on a deep neural network. The deep neural network is a high-order correlation rule for extracting features through multiple network levels. The core of the supervised learning is to label the training data with clear labels. The leakage pipeline learning model learns the judgment rule through the corresponding relationship between the input features and the labels.

[0094] For example, the leakage detection feature data (used to output the type or shape of the leakage, such as a hole, a crack, etc.) and the positioning feature data (used to output the leakage position) extracted in S200 need to cover the spatial propagation characteristics of different leakage positions, the exclusive acoustic characteristics of different leakage shapes, and the labels. The labeled labels are divided into two categories, both of which are supervised training classification labels: a leakage type label (such as “hole leakage / crack leakage / other preset leakage types”, which needs to correspond to the leakage detection feature data), and a regression label: a leakage point spatial coordinate label (such as an absolute coordinate based on the pipeline mileage or a relative position coordinate based on the sensor array, which needs to correspond to the positioning feature data).

[0095] Through training of the leakage pipeline learning model, the target leakage pipeline learning model obtained can identify the specific type of the input leakage signal based on the leakage detection feature data, and accurately output the spatial position information of the leakage point based on the positioning feature data, ensuring that the type identification accuracy and the positioning accuracy meet the preset practical requirements.

[0096] The training process is to divide the leakage detection feature data and the positioning feature data with the above-mentioned double labels into a training set and a validation set according to a proportion, input the deep neural network model (leakage pipeline learning model) in batches, and iteratively adjust the deep neural network model parameters, taking the type identification accuracy and the positioning error as the core evaluation indexes, until the indexes converge to the preset target (such as type identification accuracy ≥ 90%, positioning error ≤ preset threshold). After the training is completed, the target leakage pipeline learning model is obtained. The target leakage pipeline learning model is started only when the “abnormal trigger signal” of the target health pipeline learning model is received, and performs the “type identification + position positioning” dual tasks on the pipeline noise data corresponding to the abnormal attention map.

[0097] S303, sequentially combining the health pipeline learning model and the leakage pipeline learning model to obtain the deep learning model.

[0098] Specifically, the sequence combination architecture of "pre-determination-post-special processing" is adopted to clarify the function division and trigger logic of the two sub-models, ensuring that the process is free of redundancy and fault.

[0099] S304, a target deep learning model is obtained by sequentially combining a target health pipeline learning model and a target leakage pipeline learning model.

[0100] When the target health pipeline learning model determines that the state of the current pipeline is abnormal, it outputs a trigger to start the target leakage pipeline learning model.

[0101] Specifically, the target health pipeline learning model trained in S301 and the target leakage pipeline learning model trained in S302 are integrated according to the "sequential trigger combination" architecture defined in S303 to form a target deep learning model that can be directly used for monitoring.

[0102] For example, the trigger and workflow are as follows:

[0103] First-level continuous monitoring: The target health pipeline learning model continuously receives feature data extracted from real-time noise of the current pipeline, calculates its reconstruction error and abnormal attention map.

[0104] Abnormality determination and triggering: The system presets a threshold based on the reconstruction error (or a health index derived therefrom). When the reconstruction error output by the target health pipeline learning model first exceeds the preset preliminary warning threshold, the target leakage pipeline learning model is not triggered immediately, but the current state is marked as "preliminary abnormality" and the time is recorded.

[0105] After determining "preliminary abnormality", the active acquisition verification process is started, the active acoustic emission instrument sends a standard detection signal to the pipeline and synchronously acquires the response to obtain the health pipeline active noise data at the current time (i.e. the dynamic response of the structure). The features extracted from the current health pipeline active noise data are compared with the baseline values of the health pipeline active noise data established under the health state, and the difference between them (such as the change of dynamic response decay rate, the shift of characteristic frequency, etc.) is calculated.

[0106] Verification logic, if the difference of the active features also exceeds the verification threshold, it indicates that the state of the pipeline structure has indeed changed significantly, which is mutually confirmed with the passive abnormality, and is confirmed as "high-confidence abnormality". In addition, for high-confidence abnormality, the active acoustic emission instrument can perform multiple verifications at a preset time, and if multiple verifications all exceed the verification threshold, it is confirmed as "high-confidence abnormality", and if the measurement result of less than or equal to the set threshold number exceeds the verification threshold, it is confirmed as "low-confidence abnormality".

[0107] If the difference of the health pipeline active noise data features does not exceed the verification threshold, it indicates that the current passive noise anomaly is more likely to be caused by non-structural working condition fluctuations (such as flow changes), and is marked as "low credibility anomaly" or "suspected working condition interference".

[0108] Further, the target leakage pipeline learning model is triggered (started) only in the case of high credibility anomaly.

[0109] S400, using the target deep learning model, monitoring the current pipeline to obtain a monitoring result of the current pipeline.

[0110] The monitoring result at least includes a determination of whether there is a leakage and positioning information of the leakage occurrence.

[0111] Specifically, obtaining the monitoring result of the current pipeline can include the following steps:

[0112] S401, for the current pipeline, using a sensor to realize passive collection of original pipeline passive noise data of a health pipeline, and using an active acoustic emission instrument to realize active collection of original pipeline active noise data of the pipeline, and performing data synchronization, denoising and normalization processing to form preprocessed pipeline passive noise data and preprocessed pipeline active noise data.

[0113] Specifically, a sound or vibration sensor deployed on the pipeline continuously records the sound wave or vibration signal when the pipeline is running at a high sampling rate to form the original pipeline passive noise data. At the same time, a standard mechanical or acoustic pulse signal with a known characteristic is applied to the pipeline by an active acoustic emission instrument (such as an acoustic emission instrument) at a predetermined time interval or under a specific instruction, and its response is recorded synchronously to form the original pipeline active noise data.

[0114] It should be noted that during the collection process, it is necessary to ensure that the continuous data stream collected passively is aligned with the pulse response data triggered actively on the same time reference, so that the two types of data have comparability and can accurately locate the time of anomaly occurrence.

[0115] It should also be noted that the original pipeline active noise data and the original pipeline passive noise data need to be denoised and normalized. Specifically, denoising is to suppress background environmental noise interference in a specific frequency band that is irrelevant to the pipeline state by using filtering and other methods. Normalization is to adjust the amplitude of the original pipeline active noise data and the original pipeline passive noise data to a unified numerical range (for example, between -1 and 1) to eliminate signal intensity fluctuations caused by differences in sensor sensitivity, different transmission distances or working condition changes, and ensure the stability of the target deep learning model analysis.

[0116] S402, input the preprocessed passive noise data of the pipeline within a continuous time period into the target healthy pipeline learning model to obtain the reconstruction error of the current pipeline, and calculate the spatial feature distance between the reconstruction error and the benchmark value of the healthy pipeline to obtain the pattern novelty index sequence.

[0117] Specifically, a segment of passive noise data from the pipeline (e.g., data from the past 10 minutes), collected and preprocessed continuously, is input into the target healthy pipeline learning model. This model attempts to reconstruct the signal; the reconstruction error is the overall difference between the original input data and the reconstructed output data from the model. This reconstruction error directly measures how much the current signal pattern deviates from its learned healthy pipeline baseline.

[0118] It should be noted that, to further refine the evaluation of the reconstruction error generated by the input data to the target healthy pipeline learning model, the distribution pattern in the feature space (multi-dimensional, including time, frequency, and sensor location) is examined to determine whether it deviates from the typical error distribution pattern under healthy conditions. Therefore, it is necessary to analyze not only the overall error but also the distribution characteristics of the pipeline's passive noise data. By calculating the distribution of the reconstruction error in the signal feature space (correlated with dimensions such as time, frequency, and sensor location) and comparing it with a baseline distribution (benchmark value) obtained statistically from a large amount of healthy data, a pattern novelty index sequence is obtained. This pattern novelty index sequence changes over time; an increase in the value indicates an anomaly in the current signal, rather than just random fluctuations.

[0119] It should also be noted that spatial feature distance calculation involves mapping the reconstruction error generated by the current data into a feature vector in a multi-dimensional feature space (such as time, frequency, sensor). Then, the quantized distance between this vector and the health baseline vector (or distribution) is calculated (e.g., using Mahalanobis distance to account for correlations between features, or calculating KL divergence to measure differences in probability distributions). This calculated distance value, after standardization, is output as the pattern novelty index. Continuous calculations in this way will form a sequence of pattern novelty indices.

[0120] S403, compare the pre-processed active noise data of the pipeline with the active detection benchmark value of the healthy pipeline within a continuous time period to obtain the dynamic resilience index sequence.

[0121] Specifically, in step S403, the pipeline active noise data is periodically actively collected, for example, with a period of 1 hour, to form noise data, so as to establish a baseline value, and at the same time, in response to the occurrence of multiple isolated peaks in the pattern novelty index sequence, the active noise data is actively collected in time by using the active acoustic emission instrument. The preprocessed pipeline active noise data collected at intervals of a preset time period is compared with the baseline value of the active detection established under the condition that the pipeline is completely healthy. The baseline value represents the standard dynamic fingerprint of the pipeline structure when it is healthy, such as the characteristic decay time and the main frequency of the impulse response. Next, by quantifying the deviation of the current pipeline active noise data from the baseline value (such as calculating the correlation coefficient or similarity of the current pipeline active noise data and the baseline value in some characteristics), a dynamic resilience index sequence is generated. The decrease of the dynamic resilience index indicates that the dynamic characteristics (such as the stiffness) of the pipeline structure have changed, which is usually an indication of structural damage (such as cracks or corrosion leading to local softening) of the pipeline structure, rather than temporary changes in fluid working conditions.

[0122] S404, according to the pattern novelty index sequence and the dynamic resilience index sequence, respectively, risk scoring is performed to obtain a comprehensive risk score.

[0123] Specifically, the pattern novelty index sequence and the dynamic resilience index sequence obtained in steps S402 and S403 are respectively analyzed and scored. The scoring rules can be based on the numerical level, the change trend (such as whether it is continuously rising) and the synergistic relationship (such as whether they deteriorate at the same time) of the two indexes.

[0124] Next, the two independent scores are integrated according to a preset fusion rule (such as weighted average, maximum value or combination based on rule logic) to obtain a single comprehensive risk score. The score is a quantitative output of the system on whether the current pipeline state is in a high leakage risk, which provides a decision basis for whether to trigger subsequent diagnosis (start the target leakage pipeline learning model). The higher the score, the greater the possibility of leakage.

[0125] The following is described with an example:

[0126] First, the system calculates independent scores for the two pattern novelty indexes and the dynamic resilience index according to preset rules (assuming that the score range is 0-100 points).

[0127] Calculate PNI (Pattern Novelty Index) risk score: set the early warning threshold of PNI as 60. Take the current latest value 90, calculate its "level score" as (90-60) / (100-60)*100 = 75 points. At the same time, analyze the pattern novelty index sequence with strong upward trend, calculate its "trend score" as 30 points. Weight the level score (weight 0.7) and trend score (weight 0.3) to get PNI risk score = 75 * 0.7 + 30 * 0.3 = 61.5 points.

[0128] Calculate DRI (Dynamic Resilience Index) risk score: set the early warning threshold of DRI as 85. Take the current latest value 72, calculate its "level score" as (85-72) / (85-60)*100 = 52 points. Its downward trend is obvious, calculate "trend score" as 25 points. After weighting, DRI risk score = 52 * 0.7 + 25 * 0.3 = 43.9 points.

[0129] Next, the sensor detects that PNI has been rising continuously in the past 3 days (trend deterioration), while DRI has been falling continuously (trend also deteriorated), and determines that there is a "coordinated deterioration" relationship between the two. Therefore, activate the coordinated deterioration coefficient K = 1.5. Then, the system calculates the comprehensive risk score by the fusion formula:

[0130] ;

[0131] Among them, represents the comprehensive risk score, represents the pattern novelty risk score, represents the dynamic resilience risk score, represents the coordinated deterioration coefficient, preferably 1.5, represents the upper limit of the score. In this way, after substituting the specific numerical value into the formula, the obtained , if the 70 points is set as the high risk threshold, the score exceeds the preset 70 points high risk threshold. According to this, it is determined that the current is "high confidence structural anomaly", and the next stage of leakage accurate identification and positioning process is automatically triggered. This process converts multiple index information into reliable action instructions through quantitative evaluation and logical fusion.

[0132] Need to explain, for the start target leakage pipeline learning model, can be divided into two steps, the first step is according to the mode novelty index sequence appears multiple isolated peaks (multiple isolated peaks can be in 30 minutes, mode novelty index sequence appears ≥3 peak more than PNI warning threshold (60) signal peak, and the time interval of adjacent peak ≥5 minutes (avoid the same abnormal continuous trigger), that is, judged as abnormal, start DRI auxiliary verification, in addition, also can be set to appear a isolated peak is judged as abnormal, start DRI auxiliary verification), that is, after the abnormal, the second step is to use the pipeline active noise data for auxiliary verification, comprehensive judgment, form comprehensive risk score, if the comprehensive risk score is lower than the threshold, for example, 70 minutes, then do not start the target leakage pipeline learning model, on the contrary, if according to the mode novelty index sequence appears multiple isolated peaks, that is, after the abnormal, the pipeline active noise data is used for auxiliary verification, comprehensive risk score, comprehensive risk score is higher than the threshold, then start the target leakage pipeline learning model.

[0133] In one embodiment, the steps of S401 to S404 get the comprehensive risk score, but according to the comprehensive risk score, only the basic monitoring and warning are realized, and the deep diagnosis and evaluation are not realized, therefore, S405 to S408 still need to be executed in order to understand the nature, range and evolution trend of the comprehensive risk score. That is to say, although the steps of S401 to S404 give the comprehensive risk score, the speed and severity of development still need further analysis, thus:

[0134] S400, the step of monitoring the current pipeline by using the target deep learning model to get the monitoring result of the current pipeline, further comprises:

[0135] When the comprehensive risk score exceeds the first threshold (which can be set according to the need), the risk situation deep analysis step is executed:

[0136] S405, according to the mode novelty index sequence and the dynamic resilience index sequence, respectively, sliding window time series analysis is carried out to get the change trend and curvature of the mode novelty index sequence and the change trend and curvature of the dynamic resilience index sequence.

[0137] Specifically, when the comprehensive risk score exceeds the first threshold value, it indicates that the pipeline state has deviated from the healthy benchmark value preliminarily, further triggering the execution of the risk posture deep analysis step. In step S405, a sliding window of a length of time is set and continuously slides on the time axis. For the data segment within the sliding window, mathematical fitting analysis is respectively performed on the pattern novelty index sequence and the dynamic resilience index sequence. Considering that the pipeline state change is generally gradual, the length of the sliding window is usually set to several hours to several days (for example, 24 hours, 48 hours) to effectively capture long-term trends reflecting structural evolution while smoothing short-term random fluctuations. Exemplarily, a one-dimensional quadratic polynomial based on the least squares method can be used to fit the index sequence within the sliding window;

[0138] ;

[0139] wherein, represents the fitted value of the index (PNI or DRI) at time , is the relative time within the window, represents the quadratic coefficient, quantifying the curvature, represents the linear coefficient, quantifying the trend, represents the constant term representing the fitted value of the index at the start of the window .

[0140] Trend: quantified by the linear coefficient , for example, indicates that the pattern novelty index is linearly rising at a rate of 0.01 units per hour over the past day. A positive trend indicates that the abnormality is intensifying, and a negative trend indicates that it is mitigating.

[0141] Curvature: quantified by the quadratic coefficient . For example, indicates that the rising speed of the pattern novelty index is accelerating (positive curvature). The curvature reveals whether the abnormality is accelerating, uniform, or decelerating, which is crucial for judging the urgency of the situation. For example, “There are signs of slight acceleration in the rising rate of the pattern novelty index over the past three days (positive curvature)” or “The decreasing rate of the dynamic resilience index tends to be stable (curvature close to zero)”.

[0142] S406, according to the pattern novelty index sequence and the dynamic resilience index sequence, calculating a correlation coefficient to obtain the trend of the pattern novelty index sequence and the dynamic resilience index sequence rising or falling together.

[0143] Specifically, the linear correlation coefficient between the pattern novelty index sequence and the dynamic resilience index sequence is calculated within the same sliding time window. The coefficient is a value between -1 and 1.

[0144] Exemplarily, the specific calculation method of the correlation coefficient can be:

[0145] The Pearson product-moment correlation coefficient is used for calculation. For a sliding time window with a length of , let the pattern novelty index sequence be , and the dynamic resilience index sequence be , and the correlation coefficient between the two sequences be given by the following formula:

[0146] ;

[0147] wherein and are the arithmetic mean values of the sequences and in the window, and the correlation coefficient has a value range of .

[0148] Next, the synergistic trend is determined:

[0149] If the correlation coefficient is a significant negative value (for example, less than -0.7), it indicates that the pattern novelty index rises (acoustic anomaly intensifies) while the dynamic resilience index decreases (structural performance deteriorates), and the two show a negative synergy. This is a typical sign of structural damage (such as leakage development and crack propagation), because physical damage usually causes both acoustic anomalies and deterioration of structural dynamic characteristics.

[0150] If the correlation coefficient is a significant positive value, it indicates that both rise and fall, which may be directed to a common, non-structural external interference factor (such as overall fluctuations in system pressure).

[0151] If the correlation coefficient is close to zero, it indicates that there is no clear statistical correlation between the changes, which may mean independent disturbance sources or local sensor problems.

[0152] S407, according to the index trends of adjacent sensors in the current pipeline, the spatial consistency of the pattern novelty index sequence and the dynamic resilience index sequence is obtained.

[0153] Specifically, in step S407, considering multiple monitoring points (sensors) arranged along the pipeline, each point independently generates a respective pattern novelty index sequence and a dynamic resilience index sequence. The trend of the pattern novelty index calculated by adjacent or similar sensors (the change trend obtained in step S405) is compared. If multiple adjacent sensors detect similar change direction and amplitude in the same time period (for example, three consecutive sensors upstream, midstream, and downstream all show that the pattern novelty index significantly rises), it is determined to have high spatial consistency. This indicates that the abnormal event has a wide range of influence, which may involve a continuous section of the pipeline. Conversely, if the abnormal change trend only appears on an isolated sensor, the spatial consistency is low, which may indicate that the sensor is malfunctioning or a very local disturbance.

[0154] In step S408, the risk situation description is obtained according to the change trend and curvature of the pattern novelty index sequence, the change trend and curvature of the dynamic resilience index sequence, the trend of the coordinated rise or fall of the pattern novelty index sequence and the dynamic resilience index sequence, and the spatial consistency of the pattern novelty index sequence and the dynamic resilience index sequence.

[0155] Specifically, step S408 is a multi-dimensional information fusion that takes all the quantitative indicators (the trend and curvature of the two indexes, their coordination, and spatial consistency) output by steps S405 to S407 as input. Based on a pre-set rule knowledge base or a lightweight classification model, the above fusion information is interpreted to generate a qualitative risk situation description text. This is specifically implemented by a rule engine: the engine pre-defines a series of rules, each rule is composed of a combination of threshold conditions of multiple indicators, and corresponds to a description template. If the current indicator values are substituted into the rule engine, they are matched in priority, the first rule that meets all the conditions is triggered, and the corresponding risk description text is generated. For example:

[0156] Rule and output example:

[0157] When the pattern novelty index rapidly rises and accelerates, the dynamic resilience index rapidly falls and accelerates, there is strong negative correlation between the two, and the spatial consistency is high, the "high risk-accelerating deterioration" description is triggered, indicating that the structural damage is expanding.

[0158] When the pattern novelty index moderately rises, the dynamic resilience index does not change significantly, the correlation between the two is weak, and the spatial consistency is low, the "medium risk-local disturbance" description is triggered, indicating that it may be a local non-structural disturbance.

[0159] When the two indexes change in the same direction and have positive correlation, the spatial consistency is high, and the change is smooth, the "low risk-system fluctuation" description is triggered, indicating that it is a system overall working condition fluctuation.

[0160] This description goes beyond a single composite risk score, providing a comprehensive judgment of the nature, development speed, and impact range of the anomaly.

[0161] In one embodiment, the step of monitoring the current pipeline using the target deep learning model to obtain a monitoring result of the current pipeline further comprises:

[0162] S409, according to the risk situation description, performing leakage prediction level assessment on the current pipeline to obtain a leakage risk level; when the leakage risk level is greater than or equal to a preset leakage risk level, starting the target leakage pipeline learning model.

[0163] Specifically, the risk situation description is a comprehensive output containing a qualitative conclusion (such as "high risk-accelerating deterioration") and multi-dimensional quantitative indicators (such as trend, synergy, and spatial consistency). In the S409 step, the risk situation description is analyzed and quantitatively rated by a preset rule logic.

[0164] Leakage risk level determination: according to the risk situation description, it is mapped to a preset leakage risk level. For example, if the risk situation description indicates that the anomaly has structural damage characteristics, and shows an accelerating or continuous deterioration trend, and has high spatial consistency, it will be evaluated as a high leakage risk level; if the description indicates that the anomaly may be caused by local, non-structural transient interference, it will be evaluated as a medium or low leakage risk level.

[0165] Triggering start decision: according to the evaluated leakage risk level, it is determined whether the current situation is a real leakage.

[0166] It should be noted that the leakage risk level can be divided into three levels: high risk, medium risk, and low risk; the level mapping rule is: high risk: structural damage characteristics + accelerating / continuous deterioration + high spatial consistency; medium risk: non-structural interference + moderate change + low spatial consistency; low risk: system fluctuation + stable change + high spatial consistency; only when the leakage risk level is "high risk", the target leakage pipeline learning model is started.

[0167] S410, obtaining the current pipeline passive noise data in the latest continuous time window when forming the risk situation description as the data to be analyzed.

[0168] Specifically, after determining the leakage risk level in step S409, further analysis of the data related to the judgment of the leakage risk level is needed, that is, the risk situation description is obtained based on the analysis of the pattern novelty index sequence and the dynamic resilience index sequence of a specific length of time in the past. Step S410 performs data backtracking to extract the original data on which the description is based from the real-time data cache or database. That is, the most relevant original evidence that leads to the final risk level assessment is found.

[0169] For example, the extracted data to be analyzed has a time range (start and end time) corresponding to the deep analysis window of the risk situation. For example, if the risk situation is obtained based on the analysis of the index sequence of the past 5 days, the original pipeline passive noise data (or its high-fidelity preprocessed version) in the same 5 days is extracted here. It should be noted that the time range of the data to be analyzed is equal to the length of the sliding window in step S405, ensuring that the data to be analyzed is completely consistent with the analysis basis of the risk situation description, avoiding data misplacement.

[0170] S411, from the target health pipeline learning model, obtaining the abnormal attention map corresponding to the data to be analyzed.

[0171] Specifically, when the data to be analyzed is processed in step S411 to calculate the reconstruction error and the pattern novelty index from the target health pipeline learning model, the abnormal attention map describing the distribution of abnormal areas in the data has been generated internally. Step S411 directly calls or outputs the abnormal attention map generated by the target health pipeline learning model in the previous calculation, which strictly corresponds to the data to be analyzed.

[0172] Information inheritance: This operation avoids repeated running of the health model for the same data, directly inherits the intelligent analysis results completed in the first level monitoring, and realizes efficient transmission of information. That is, the data to be analyzed is used as a condition for generating an abnormal attention map, and the abnormal attention map is generated according to the data to be analyzed. The abnormal attention map previously calculated and cached corresponding to the data to be analyzed is directly called.

[0173] S412, according to the abnormal attention map, selecting one or more key time segments with abnormal attention values greater than or equal to a preset threshold from the data to be analyzed.

[0174] Specifically, step S412 uses the guidance of the abnormal attention map to intelligently focus on the most suspicious segments from long-time data, realizes data dimension reduction and feature enhancement, and thus greatly reduces the amount of data needed to input the subsequent leakage positioning model, improves processing efficiency and signal-to-noise ratio.

[0175] It should be noted that step S412 can include the following steps:

[0176] S4121, integrating or clustering the abnormal attention values of the abnormal attention graph in the time dimension, to identify time periods of continuous abnormal attention.

[0177] Specifically, the purpose of integration or clustering analysis is:

[0178] For the integration analysis process, the abnormal attention graph in the time dimension is regarded as a continuous curve. The sliding window (sliding window length such as 1 hour, 5 hours) is integrated to calculate the sum of abnormal attention values in each sliding window. The purpose is to smooth the transient fluctuations and highlight those time periods where the abnormal attention values are continuously maintained at a high level. A short strong pulse may not have a high integral value, while a long duration of moderate abnormal area will obtain a high integral value and be identified.

[0179] For the clustering analysis process, each high value point of the abnormal attention graph on the time axis is regarded as an "event". Through clustering algorithm, these adjacent discrete high value points in time are merged into the same cluster (for example, the specific time length adjacent in time can be 1 hour). Each cluster represents a continuous or nearly continuous abnormal event active period in time. This method can directly output several independent and clear boundary abnormal time periods.

[0180] S4122, the time period of continuous abnormal attention greater than or equal to the preset threshold is taken as the key time segment.

[0181] Specifically, the preset threshold (such as 0.7, 0.8) is the key criterion to distinguish "significant abnormal period" from "general background fluctuations or noise". The preset threshold can be based on the benchmark value determined by historical health data statistics.

[0182] For the window identified by the integration analysis of S4121 step, or the abnormal event cluster identified by the clustering analysis, it is judged whether the integral energy or the average attention value (or peak value) in the cluster is greater than or equal to the preset threshold. All continuous time periods that meet the threshold condition are finally determined as key time segments. These key time segments are the subset of the original long time data that is considered most likely to contain effective leakage features and most worthy of in-depth analysis after evaluation. They will be extracted as the input of the subsequent leakage positioning model.

[0183] Exemplarily, the abnormal attention graph marks the abnormal degree of each time (or short period) in the time dimension. Next, the abnormal attention graph is traversed to identify continuous time periods where the attention value is continuously higher than the preset threshold. The original noise data signal corresponding to these time periods is most likely to contain abnormal events (such as leakage sound emission) that lead to high risk determination.

[0184] All identified continuous time segments that meet the threshold condition are determined as key time segments. The set of these key time segments represents the "suspected moments" in the entire analysis window that are most worthy of in-depth examination, and the total length of which is less than the length of the original data to be analyzed.

[0185] S413, input the noise data corresponding to one or more of the key time segments into the target leakage pipeline learning model for leakage judgment to obtain the monitoring result of the current pipeline.

[0186] The monitoring result at least includes the judgment of whether there is leakage and the positioning information of the leakage.

[0187] Specifically, the passive noise data of the pipeline corresponding to the key time segments selected in S412 (usually the same preprocessed segments) is taken as input and provided to the target leakage pipeline learning model that has been started. If the length of the selected key time segments is inconsistent with the standard input length of the target leakage pipeline learning model, the system will perform standardization processing such as truncation or splicing on it. The target leakage pipeline learning model is trained to directly process such noise segments.

[0188] The target leakage pipeline learning model independently or jointly analyzes one or more key time segments, and outputs two core judgments: 1) whether there is leakage (binary classification or probability of yes / no); 2) the positioning information of the leakage (such as the distance from the reference sensor, the mileage coordinate of the pipeline). This output is the final executable monitoring result.

[0189] Finally, the monitoring result and the previously generated risk situation description and other information are packaged to form a complete diagnosis report, triggering corresponding alarms, log records or maintenance work orders.

[0190] It should be noted that in the target leakage pipeline learning model, the target leakage pipeline learning model can calculate the spatial position information of the leakage point according to the formation time of the key time segment or the multi-sensor signal features (such as signal arrival time difference) contained in the key time segment data, map it to the specific position of the pipeline, and thus complete the positioning of the leakage position of the pipeline. In this way, the target leakage pipeline learning model only needs to monitor the leakage positioning position corresponding to the key time segment in subsequent work, without the need to monitor the entire pipeline, saving computing power.

[0191] Finally, it should be noted that the threshold values mentioned in the present application can be adjusted according to actual conditions, and the specific numerical values of the threshold values are not limited in the present application.

[0192] In a second aspect, the application provides a pipeline leakage noise positioning analysis system based on deep learning, applied to the pipeline leakage noise positioning analysis method based on deep learning, comprising:

[0193] A collection unit is configured to collect data of healthy pipelines and leakage pipelines respectively to obtain a healthy pipeline noise data set and a leakage pipeline noise data set, wherein the samples obtained by collecting data of the healthy pipelines and the leakage pipelines include sound samples of pipe diameter, pressure, material and environmental conditions;

[0194] A feature extraction unit is configured to extract features of the sound samples in the healthy pipeline noise data set and the leakage pipeline noise data set respectively by using a convolutional neural network, and classify to obtain healthy pipeline feature data, leakage detection feature data and positioning feature data;

[0195] A model training unit is configured to train a deep learning model by using the healthy pipeline feature data, the leakage detection feature data and the positioning feature data to obtain a target deep learning model;

[0196] A result unit is configured to monitor a current pipeline by using the target deep learning model to obtain a monitoring result of the current pipeline, wherein the monitoring result at least includes a judgment of whether there is leakage and positioning information of leakage occurrence.

[0197] In a third aspect, the application provides a computer storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps of the pipeline leakage noise positioning analysis method based on deep learning.

[0198] In a fourth aspect, the application provides a computer program, wherein the computer program is executed by a processor to implement the steps of the pipeline leakage noise positioning analysis method based on deep learning.

[0199] It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0200] Obviously, those skilled in the art should understand that the units or steps of the present application described above can be realized by universal computing devices, and they can be centralized on a single computing device or distributed on a network composed of multiple computing devices. Alternatively, they can be realized by program codes executable by the computing devices, so that they can be stored in storage devices and executed by the computing devices, or they can be respectively manufactured into individual integrated circuit modules, or multiple modules or steps among them can be manufactured into a single integrated circuit module to realize. Thus, the present application is not limited to any particular combination of hardware and software.

[0201] The preferred embodiments of the present application are described above, but the present application is not limited to the above. Those skilled in the art can make various modifications and changes to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for deep learning-based pipeline leak noise localization analysis, characterized in that, The application comprises the following steps: Collecting data of healthy pipelines and leaking pipelines respectively to obtain healthy pipeline noise data set and leaking pipeline noise data set, wherein the samples obtained by collecting data of the healthy pipelines and the leaking pipelines include sound samples of pipe diameter, pressure, material and environmental conditions; Extracting features of the sound samples in the healthy pipeline noise data set and the leaking pipeline noise data set respectively by using a convolutional neural network, and classifying to obtain healthy pipeline feature data, leakage detection feature data and positioning feature data; Training a deep learning model by using the healthy pipeline feature data, the leakage detection feature data and the positioning feature data to obtain a target deep learning model; Monitoring the current pipeline by using the target deep learning model to obtain a monitoring result of the current pipeline, wherein the monitoring result at least includes a judgment of whether there is leakage and positioning information of leakage occurrence; wherein The step of training a deep learning model by using the healthy pipeline feature data, the leakage detection feature data and the positioning feature data to obtain a target deep learning model comprises the following steps: Constructing a healthy pipeline learning model and training to obtain a target healthy pipeline learning model, wherein the healthy pipeline learning model is an unsupervised model based on an autoencoder, which is trained by using healthy pipeline passive noise data, is used to calculate reconstruction error of input noise data, and can output an abnormal attention map of noise data, the abnormal attention map is a multi-dimensional matrix representing the abnormal degree of time, frequency and space dimensions; Constructing a leakage pipeline learning model and training to obtain a target leakage pipeline learning model, wherein the leakage pipeline learning model is a supervised model based on a deep neural network, which is trained by using leakage detection feature data and positioning feature data, is used to judge whether there is leakage and output leakage position information; Sequentially combining the healthy pipeline learning model and the leakage pipeline learning model to obtain the deep learning model; Sequentially combining the target healthy pipeline learning model and the target leakage pipeline learning model to obtain the target deep learning model, wherein when the target healthy pipeline learning model judges that the state of the current pipeline is abnormal, it outputs to trigger the target leakage pipeline learning model to start.

2. The deep learning-based pipeline leak noise localization analysis method of claim 1, wherein, The step of collecting data of healthy pipelines and leaking pipelines respectively to obtain healthy pipeline noise data set and leaking pipeline noise data set comprises the following steps: Collecting data during operation of the healthy pipelines to obtain the healthy pipeline noise data set, wherein collecting data during operation of the healthy pipelines includes healthy pipeline passive collection and healthy pipeline active collection, corresponding to form healthy pipeline passive noise data and healthy pipeline active noise data, the healthy pipeline passive noise data represents data collected from the healthy pipeline in continuous time, and the healthy pipeline active noise data represents data collected from the healthy pipeline at a preset time interval; The data of the leakage pipeline is collected to obtain a leakage pipeline noise data set, wherein the leakage pipeline noise data set includes a set of leakage pipeline noise data collected passively from the leakage pipeline, and the passive collection of the leakage pipeline means collecting data from the leakage pipeline in a continuous time period. 3.The deep learning-based pipeline leak noise localization analysis method of claim 2, wherein, The step of monitoring the current pipeline by using the target deep learning model to obtain a monitoring result of the current pipeline comprises: For the current pipeline, passive collection of original pipeline passive noise data is realized by using a sensor, and active collection of original pipeline active noise data is realized by using an active acoustic emission instrument, and data synchronization, denoising and normalization processing are performed to form preprocessed pipeline passive noise data and preprocessed pipeline active noise data; The preprocessed pipeline passive noise data in a continuous time period is input into the target health pipeline learning model to obtain a reconstruction error of the current pipeline, and spatial feature distance calculation is performed between the reconstruction error and a reference value of the health pipeline to obtain a pattern novelty index sequence; The preprocessed pipeline active noise data in a continuous time period is compared with the active detection reference value of the health pipeline to obtain a dynamic restoring force index sequence; According to the pattern novelty index sequence and the dynamic restoring force index sequence, risk scores are respectively obtained to obtain a comprehensive risk score. 4.The deep learning-based pipeline leak noise localization analysis method of claim 3, wherein, The step of monitoring the current pipeline by using the target deep learning model to obtain a monitoring result of the current pipeline further comprises: When the comprehensive risk score exceeds a first threshold, a risk situation deep analysis step is performed: According to the pattern novelty index sequence and the dynamic restoring force index sequence, sliding window time series analysis is respectively performed to obtain a change trend and a curvature of the pattern novelty index sequence and a change trend and a curvature of the dynamic restoring force index sequence; According to the pattern novelty index sequence and the dynamic restoring force index sequence, a correlation coefficient is calculated to obtain a trend of cooperative rise or fall of the pattern novelty index sequence and the dynamic restoring force index sequence; According to the index trend of adjacent sensors in the current pipeline, spatial consistency of the pattern novelty index sequence and the dynamic restoring force index sequence is obtained; According to the change trend and the curvature of the pattern novelty index sequence, the change trend and the curvature of the dynamic restoring force index sequence, the trend of cooperative rise or fall of the pattern novelty index sequence and the dynamic restoring force index sequence, and the spatial consistency of the pattern novelty index sequence and the dynamic restoring force index sequence, a risk situation description is obtained.

5. The deep learning-based pipeline leak noise localization analysis method of claim 4, wherein, The step of monitoring the current pipeline by using the target deep learning model to obtain a monitoring result of the current pipeline further comprises: According to the risk situation description, a leakage prediction level of the current pipeline is evaluated to obtain a leakage risk level; when the leakage risk level is greater than or equal to a preset leakage risk level, the target leakage pipeline learning model is started. acquire current pipeline passive noise data in a latest continuous time window when the risk situation description is formed as to-be-analyzed data; acquire the abnormal attention map corresponding to the to-be-analyzed data from the target health pipeline learning model; select one or more key time segments with abnormal attention values greater than or equal to a preset threshold from the to-be-analyzed data according to the abnormal attention map; input noise data corresponding to the one or more key time segments into the target leakage pipeline learning model for leakage judgment to obtain a monitoring result of the current pipeline, wherein the monitoring result at least includes a judgment of whether there is leakage and positioning information of leakage occurrence.

6. The deep learning-based pipeline leak noise localization analysis method of claim 5, wherein, The step of selecting one or more key time segments with abnormal attention values greater than or equal to a preset threshold from the to-be-analyzed data according to the abnormal attention map comprises: integrating or clustering abnormal attention values of the abnormal attention map in the time dimension to identify abnormal attention continuous time periods; abnormal attention continuous time periods greater than or equal to a preset threshold are taken as the key time segments.

7. A deep learning-based pipeline leak noise localization analysis system, characterized by, The deep learning-based pipeline leakage noise positioning analysis method of any one of claims 1-6 comprises: a collection unit configured to collect data of a health pipeline and a leakage pipeline respectively to obtain a health pipeline noise data set and a leakage pipeline noise data set, wherein samples obtained by collecting data of the health pipeline and the leakage pipeline include sound samples of pipe diameter, pressure, material, and environmental conditions; a feature extraction unit configured to extract features of the sound samples in the health pipeline noise data set and the leakage pipeline noise data set respectively by using a convolutional neural network and classify the features to obtain health pipeline feature data, leakage detection feature data, and positioning feature data; a model training unit configured to train a deep learning model by using the health pipeline feature data, the leakage detection feature data, and the positioning feature data to obtain a target deep learning model; a result unit configured to monitor a current pipeline by using the target deep learning model to obtain a monitoring result of the current pipeline, wherein the monitoring result at least includes a judgment of whether there is leakage and positioning information of leakage occurrence.

8. A computer storage medium having stored thereon a computer program, characterized in that The computer program is executed by a processor to implement the steps of the deep learning-based pipeline leakage noise positioning analysis method of any one of claims 1-6.

9. A computer program product, characterised in that, The computer program is executed by a processor to implement the steps of the deep learning-based pipeline leakage noise positioning analysis method of any one of claims 1-6.

Images