Optical fiber vibration auscultation method and system for water supply network leakage

Abstract

The invention provides an optical fiber vibration auscultation method and system for water supply network leakage, and relates to the technical field of water supply network monitoring, and the method comprises the steps: carrying out the synchronous sampling of a dual-channel optical fiber vibration signal through an FPGA main clock, so as to obtain a pipeline vibration signal and an environment noise signal which are aligned in time; performing multi-dimensional feature fusion on the pipeline vibration signal and the environmental noise signal to obtain a pipeline leakage feature vector; and inputting the pipeline leakage feature vector into the CNN-LSTM hybrid model to obtain pipeline leakage information output after the CNN-LSTM hybrid model performs feature analysis and prediction processing on the pipeline leakage feature vector. The technical problem that the recognition accuracy of water supply pipeline leakage in the complex environment is low in the prior art is solved, and the technical effect of improving the recognition accuracy of water supply pipeline leakage in the complex environment is achieved.

Description

Technical Field

[0001] The present application relates to the technical field of water supply network monitoring, and in particular to a fiber optic vibration auscultation method and system for detecting water supply network leakage. Background Art

[0002] As urbanization accelerates, the scale of water supply networks continues to expand. As the core artery of urban infrastructure, water supply networks bear the heavy responsibility of allocating water resources for residential use, industrial production, and public services. However, water supply networks face numerous severe challenges in their long-term operation. Many aging pipes, which have been in service for over 20 years, suffer from material degradation, such as rusted cast iron pipes and brittle PE pipes. This leads to leakage rates as high as 10%-30%, resulting in over 20 billion tons of water waste annually.

[0003] Among related technologies, leak detection in water supply networks mainly relies on three methods. The first is the pressure wave method, which determines leaks by sudden changes in pipeline pressure. However, it is easily disturbed by water turbulence, pump and valve start-up and shutdown, and cannot achieve continuous monitoring in large-diameter trunk pipes above DN800 due to rapid pressure decay, resulting in a missed detection rate of over 40%. The second is the sound listening method, which relies on manual nighttime inspections with a leak detector. However, it is affected by traffic noise and complex vibrations of underground pipelines, resulting in a leak point identification accuracy rate of less than 30%, and a positioning error of up to 5-8 meters. The third is the fiber optic sensing method. Although it can monitor pipeline vibration, the signal collected by a single sensing fiber seriously overlaps with the environmental noise spectrum (10-500Hz) and the leakage characteristic frequency band (0.1-200Hz), making it difficult to distinguish between effective leakage signals and traffic / construction noise. This is especially true in areas of surface subsidence (annual subsidence >3cm). Pipe deformation causes the sound wave propagation delay error to exceed ±15ms, further amplifying the positioning error.

[0004] However, the above-mentioned water supply network leakage detection method lacks a targeted environmental noise elimination mechanism and the existing signal processing method cannot effectively separate overlapping spectra, resulting in the effective leakage signal extraction being seriously interfered with by environmental noise, which in turn leads to the low recognition accuracy of water supply pipeline leakage in complex environments in related technologies. Summary of the Invention

[0005] The present application provides a fiber optic vibration auscultation method and system for detecting water supply network leaks, which are used to improve the accuracy of identifying water supply pipeline leaks in complex environments.

[0006] In the first aspect, the present application provides a fiber optic vibration auscultation method for water supply network leakage, which is applied to the above-mentioned fiber optic vibration auscultation system. The method includes: using the FPGA master clock to synchronously sample the dual-channel fiber optic vibration signal to obtain time-aligned pipeline vibration signals and environmental noise signals; performing multi-dimensional feature fusion on the pipeline vibration signal and the environmental noise signal to obtain a pipeline leakage feature vector; inputting the pipeline leakage feature vector into the CNN-LSTM hybrid model to obtain the pipeline leakage information output after the CNN-LSTM hybrid model performs feature analysis and prediction processing on the pipeline leakage feature vector.

[0007] By adopting this technical solution, the dual-channel fiber optic vibration signals are synchronously sampled using the FPGA master clock, directly acquiring time-aligned pipeline vibration and ambient noise signals. This avoids the signal misalignment problem caused by clock asynchrony in traditional sampling methods, laying the foundation for subsequent accurate analysis of pipeline vibration characteristics and distinguishing valid signals from ambient noise, thereby ensuring improved validity and reliability of leak detection data from the sampling source. This solves the technical problem of low accuracy in identifying water supply pipeline leaks in complex environments, achieving the technical effect of improving the accuracy of identifying water supply pipeline leaks in complex environments.

[0008] In a second aspect, an embodiment of the present application provides a fiber optic vibration stethoscope system, which includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code includes computer instructions, and the one or more processors call the computer instructions to enable the fiber optic vibration stethoscope system to execute the method described in the first aspect and any possible implementation method of the first aspect.

[0009] In a third aspect, an embodiment of the present application provides a computer program product comprising instructions. When the computer program product is run on a fiber optic vibration stethoscope system, the fiber optic vibration stethoscope system executes the method described in the first aspect and any possible implementation of the first aspect.

[0010] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium comprising instructions. When the instructions are executed on a fiber optic vibration stethoscope system, the fiber optic vibration stethoscope system executes the method described in the first aspect and any possible implementation of the first aspect. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 This is a flow chart of a fiber optic vibration auscultation method for detecting water supply network leakage in an embodiment of the present application; Figure 2 This is a schematic diagram of a hierarchical architecture of the optical fiber vibration stethoscope system in an embodiment of the present application; Figure 3 This is a schematic diagram of the hardware composition of the optical fiber vibration stethoscope system in an embodiment of the present application; Figure 4 This is a schematic diagram of a functional module of the optical fiber vibration stethoscope system in an embodiment of the present application; Figure 5 This is a functional diagram of the alarm access module in an embodiment of the present application; Figure 6 This is a functional diagram of the diagnostic analysis module in an embodiment of the present application; Figure 7 1 is a schematic diagram of the principle of optical fiber Rayleigh scattering measurement in an embodiment of the present application; Figure 8 This is a structural diagram of a CNN-LSTM hybrid model in an embodiment of the present application; Figure 9 Schematic diagram of a convolution kernel operation in an embodiment of the present application; Figure 10 This is a structural diagram of LSTM time series modeling in an embodiment of the present application; Figure 11 This is a schematic diagram of the architecture of a multi-task learning neural network in an embodiment of the present application; Figure 12 This is a schematic diagram of the physical device structure of the optical fiber vibration stethoscope system in the embodiment of the present application. DETAILED DESCRIPTION

[0012] The terms used in the following examples of the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application. As used in the specification and appended claims of this application, the singular expressions "a," "an," "said," "above," "the," and "this" are intended to include plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used in this application refers to any or all possible combinations comprising one or more of the listed items.

[0013] In the following, the terms "first" and "second" are used for descriptive purposes only and should not be understood to imply or suggest relative importance or implicitly indicate the number of the technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the embodiments of this application, unless otherwise specified, "plurality" means two or more.

[0014] This application provides a fiber optic vibration auscultation method for detecting water supply network leakage. Figure 1 , Figure 1 This is a flow chart of a fiber optic vibration auscultation method for detecting water supply network leakage in an embodiment of the present application, comprising the following steps: Step S101, synchronously sampling the dual-channel optical fiber vibration signal using the FPGA master clock to obtain a time-aligned pipeline vibration signal and an ambient noise signal; Step S102: performing multi-dimensional feature fusion on the pipeline vibration signal and the ambient noise signal to obtain a pipeline leakage feature vector; Step S103: Input the pipeline leakage feature vector into the CNN-LSTM hybrid model to obtain pipeline leakage information output by the CNN-LSTM hybrid model after performing feature analysis and prediction processing on the pipeline leakage feature vector.

[0015] In the above embodiment, the FPGA master clock refers to the master clock circuit in the field programmable gate array, which is used to provide a unified clock reference signal for the fiber optic vibration stethoscope system. A dual-channel fiber optic vibration signal represents a vibration signal collected simultaneously by two independent fiber optic channels, including a pipeline vibration signal and an ambient noise signal. Synchronous sampling refers to sampling the signals of the two channels at the same clock rate. Time alignment means that the signals collected by the two channels have a strict correspondence in the time dimension. Multidimensional feature fusion refers to the combined processing of feature information from multiple dimensions. The CNN-LSTM hybrid model represents a composite deep learning model that combines a convolutional neural network and a long short-term memory network.

[0016] In the above embodiment, in a large urban water supply system, in order to promptly detect leaks in the water supply network, a fiber optic vibration auscultation method is used for monitoring. Key nodes are selected in the water supply area, and dual-channel fiber optic sensors are installed on the surface of the water supply pipe and in the environment away from the pipe. The FPGA master clock synchronously samples the pipeline vibration signal collected by the fiber optic sensor from the pipe surface and the environmental noise signal collected by the environmental fiber optic sensor at a sampling frequency of 100kHz. After A / D conversion, a time-aligned digital signal sequence is obtained. For example, on a 5-kilometer-long DN800 water supply pipe, a monitoring point is set every 500 meters, and each monitoring point performs dual-channel signal acquisition to ensure that the pipeline vibration information can be fully captured. The sampled pipeline vibration signal and environmental noise signal are input into the feature extraction module. The signal is analyzed in the time domain to extract features such as peak value, mean, and variance. In the frequency domain, a fast Fourier transform (FFT) is used to obtain the signal's spectral characteristics, such as energy distribution and dominant frequency. Simultaneously, a time-frequency domain analysis is performed using a short-time Fourier transform (STFT) or wavelet transform to obtain signal characteristics at different times and frequencies. These time, frequency, and time-frequency domain features are fused to construct a pipeline leakage feature vector containing 20 dimensions. For example, when analyzing a particular acquired signal, it was found that the pipeline vibration signal exhibited an abnormal energy concentration in the low-frequency band, while the ambient noise signal had lower energy in this frequency band. Feature fusion clearly reflected this difference in the feature vector. The constructed pipeline leakage feature vector was then input into a pre-trained CNN-LSTM hybrid model. The CNN first performs a convolution operation on the feature vector, extracting local features using convolution kernels of different sizes. The pooling layer reduces the data dimension while retaining key features. The features extracted by the CNN are then passed to the LSTM layer. The LSTM layer uses its memory cells to analyze and learn the time series information of the features, predicting whether a leak has occurred in the pipeline and the location and extent of the leak.

[0017] In the above embodiment, an adaptive filtering algorithm, such as the least mean square (LMS) algorithm, is introduced during the signal acquisition phase. Based on the real-time collected ambient noise signal, the filter parameters are dynamically adjusted to reduce noise on the pipeline vibration signal. This allows for better adaptation to complex and variable ambient noise, further improving the quality of the pipeline vibration signal and providing more accurate data for subsequent feature extraction and analysis. For example, when strong noise interference is generated by construction near the monitoring area, the adaptive filtering algorithm can quickly adjust to effectively suppress the noise's impact on the pipeline vibration signal. During the multidimensional feature fusion process, a multi-scale analysis approach is employed. In addition to conventional time, frequency, and time-frequency domain features, the signal is analyzed at different scales, such as using multi-scale decomposition techniques to obtain signal features at different resolutions. These multi-scale features are fused with the original features to construct a richer and more representative pipeline leakage feature vector. This helps capture subtle differences in the characteristics of leaks of varying severity and types, improving the model's ability to identify various leak scenarios. For example, multi-scale feature fusion can more accurately distinguish between slow leaks in small-diameter pipelines and sudden leaks in large-diameter pipelines, improving detection accuracy. To address the ever-changing pipeline conditions and environmental conditions during water supply network operation, a CNN-LSTM hybrid model uses a dynamic update mechanism. New detection data is regularly collected, added to the training set, and the model is retrained. Furthermore, an online learning algorithm is employed to enable the CNN-LSTM hybrid model to adjust parameters in real time based on the new data, adapting to changes in the network's operating conditions. This ensures that the model maintains high detection accuracy and reliability over the long term, preventing performance degradation due to network changes. For example, if the vibration characteristics of a particular pipeline section change due to aging, the model's dynamic update mechanism can promptly learn these changes and adjust the detection strategy to ensure leak detection accuracy. The fiber optic vibration auscultation method is integrated with a geographic information system (GIS). Upon detecting a pipeline leak, the CNN-LSTM hybrid model outputs leak location information that is directly marked on a GIS map, along with relevant pipeline attributes such as diameter, material, and burial depth. Workers can visually view the leak location and surrounding environment through the GIS system, quickly developing repair plans and improving maintenance efficiency. In addition, the GIS system can also analyze historical leakage data, predict areas where leakage may occur in the future, carry out prevention and maintenance in advance, and further ensure the safe operation of the water supply network.

[0018] Through the above steps, the dual-channel fiber optic vibration signals are synchronously sampled using the FPGA master clock, directly acquiring time-aligned pipeline vibration and ambient noise signals. This avoids the signal misalignment caused by clock asynchrony in traditional sampling methods, laying the foundation for subsequent accurate analysis of pipeline vibration characteristics and distinguishing valid signals from ambient noise, thereby ensuring improved validity and reliability of leak detection data from the sampling source. This solves the technical problem of low accuracy in identifying water supply pipeline leaks in complex environments, achieving the technical effect of improving the accuracy of identifying water supply pipeline leaks in complex environments.

[0019] Among them, the executor of the above steps can be a system with the ability to detect water supply network leakage, such as a fiber optic vibration stethoscope system, or a device with the ability to detect water supply network leakage, or a controller or processor in the device or system, or a separate controller or processor, or other processing equipment or processing units with similar processing functions, but not limited to these.

[0020] In an optional embodiment, multi-dimensional feature fusion is performed on the pipeline vibration signal and the ambient noise signal to obtain a pipeline leakage feature vector, specifically including: performing optical time domain reflection processing on the pipeline vibration signal to obtain a phase delay variation of the back Rayleigh scattered light; using an ultra-narrow linewidth pulse laser to generate local oscillator light, and superimposing the back Rayleigh scattered light and the local oscillator light on a photodetector to obtain an interference light intensity signal; using an improved LMS adaptive filtering algorithm to filter the interference light intensity signal and the ambient noise signal to obtain a leakage vibration signal; determining the optical fiber strain characteristics of the first optical fiber channel in the dual-channel optical fiber vibration signal based on the phase delay variation, wherein the first optical fiber channel is tightly coupled with the water supply pipeline and is used to collect the pipeline vibration signal, and the second optical fiber channel in the dual-channel optical fiber vibration signal maintains a preset distance from the water supply pipeline and is used to collect the ambient noise signal; performing time domain feature extraction on the leakage vibration signal to obtain a leakage impact feature, and performing frequency domain feature extraction on the leakage vibration signal to obtain a leakage soundprint feature; and performing multi-dimensional feature fusion on the optical fiber strain feature, the leakage impact feature, and the leakage soundprint feature to obtain a pipeline leakage feature vector.

[0021] In the above embodiment, optical time domain reflectometry refers to the measurement and analysis of optical signals transmitted in an optical fiber channel using an optical time domain reflectometer. Backward Rayleigh scattered light refers to scattered light caused by fluctuations in material density when light is transmitted in an optical fiber channel. The phase delay variation is used to represent the phase variation during the transmission of the optical signal. An ultra-narrow linewidth pulse laser refers to a device that can generate pulsed lasers with an extremely narrow linewidth. Local oscillator light refers to a stable optical signal used as a reference. A photodetector refers to a device that converts an optical signal into an electrical signal. An interference light intensity signal refers to a light intensity signal generated by the interference generated by the superposition of two beams of light. The improved LMS adaptive filtering algorithm represents an optimized minimum mean square adaptive filtering method.

[0022] In the above embodiment, in a coastal city's water supply network, due to long-term erosion from sea breezes and seawater backflow, some water supply pipes have aged severely, leading to frequent leaks. To achieve efficient and accurate leak detection, the above-mentioned fiber-optic vibration-based detection solution was adopted. Monitoring points were set up every 200 meters along key sections of the city's water supply network, each equipped with a dual-channel fiber-optic sensor. The first fiber-optic channel was tightly coupled to the surface of the water supply pipe, enabling it to accurately capture pipe vibration signals. The second fiber-optic channel, maintained at a preset distance of 10 cm from the pipe, was used to collect ambient noise signals. At one monitoring point, an ultra-narrow linewidth pulsed laser (wavelength 1550 nm, linewidth 5 kHz) generated light pulses, which were injected into the first fiber-optic channel for optical time-domain reflectometry processing. As the light pulses propagate through the optical fiber, they generate Rayleigh backscattered light. Simultaneously, the local oscillator light generated by the laser and the backscattered light are superimposed in a photodetector, generating an interference light intensity signal. By analyzing the phase delay change resulting from the interference between the backscattered light and the local oscillator light, a preliminary determination can be made as to whether the pipeline is experiencing fiber strain caused by abnormal vibration. For example, when a small leak occurs in a section of the pipeline, the impact of the water flow will cause the pipeline to vibrate, which in turn causes a small deformation of the optical fiber coupled to it, resulting in a change in the phase of the backscattered Rayleigh light.

[0023] In the above embodiment, an improved LMS adaptive filtering algorithm is used to process the interference light intensity signal and the ambient noise signal collected from the second channel. Traditional LMS algorithms are prone to slow convergence and poor filtering performance in complex and variable ambient noise environments. This improved algorithm, however, introduces a variable step size mechanism that dynamically adjusts the step size based on real-time signal changes. When noise intensity suddenly increases, the algorithm rapidly increases the step size to accelerate noise suppression. When the signal stabilizes, the step size is reduced to improve filtering accuracy, ultimately obtaining a pure leakage vibration signal. Time and frequency domain features are extracted from the resulting leakage vibration signal. In the time domain, leakage impact characteristics such as peak value, rise time, and pulse width are extracted. These characteristics can intuitively reflect the intensity and duration of the leak at the moment of occurrence. In the frequency domain, leakage soundprint features are derived using a fast Fourier transform (FFT). Different types and degrees of leakage produce unique frequency domain distributions. For example, small crack leaks may have energy concentrated in the high-frequency band, while large hole leaks have distinct energy signatures in the low-frequency band. Simultaneously, the optical fiber strain signature of the first optical fiber channel is determined based on the previously acquired phase delay variation and the optical fiber's physical parameters. For example, a specific calculation formula is used to convert the phase variation into an optical fiber strain value, reflecting the degree of impact of pipeline vibration on the optical fiber. Multidimensional feature fusion is then performed on the optical fiber strain signature, leakage impact signature, and leakage acoustic signature. Principal component analysis (PCA) combined with an artificial neural network can be employed. PCA first reduces the dimensionality of the multidimensional features, removing redundant information and retaining key features. The artificial neural network then further learns the complex relationships between the features, ultimately generating a pipeline leakage feature vector containing key information about the pipeline leak.

[0024] In the above embodiment, in addition to monitoring fiber optic vibration signals, fiber optic temperature sensors and pressure sensors can also be deployed at monitoring points. When a pipeline leak occurs, the water temperature and pressure near the leak point will change. These changes are detected by the fiber optic temperature and pressure sensors and then fused with the features extracted from the fiber optic vibration signals. For example, when the vibration signal indicates a possible leak, the temperature sensor simultaneously detects a local drop in water temperature, and the pressure sensor detects an abnormal pressure fluctuation. These three signals corroborate each other, enabling a more accurate leak determination and reducing false alarm rates. Because actual pipeline leakage data is difficult to obtain and the number of samples is limited, a generative adversarial network (GAN) is used to generate virtual pipeline leakage data. By training the GAN, it learns the characteristic distribution of real leakage data, generating a large amount of simulated leakage data. This simulated data is combined with the actual collected data to expand the training dataset, which is used to optimize the subsequent leak detection model, improving its generalization and adaptability to different leakage scenarios. Considering that fiber optic sensors may experience performance drift over long-term use due to environmental factors (such as temperature and humidity changes), a self-calibration dynamic compensation mechanism is designed. The sensor performs regular self-calibration. The built-in calibration module compares the difference between the standard signal and the sensor's collected signal and automatically adjusts the sensor's parameters. Simultaneously, the collected signal is dynamically compensated based on environmental monitoring data (such as real-time temperature and humidity) to ensure the sensor's long-term stable and accurate operation.

[0025] In an optional embodiment, the pipeline leakage feature vector is input into a CNN-LSTM hybrid model to obtain pipeline leakage information output by the CNN-LSTM hybrid model after the CNN-LSTM hybrid model performs feature analysis and prediction processing on the pipeline leakage feature vector. Specifically, the pipeline leakage feature vector is input into the CNN-LSTM hybrid model so that the CNN-LSTM hybrid model performs the following operations: when the CNN-LSTM hybrid model determines that the pipeline leakage feature vector has been received, the pipeline leakage feature vector is subjected to multi-scale time series modeling processing to obtain an optimized feature vector; and the CNN-LSTM hybrid model performs multi-task joint prediction processing on the optimized feature vector to obtain pipeline leakage information.

[0026] In the above embodiment, the CNN-LSTM hybrid model refers to a deep learning model that combines a convolutional neural network (CNN) with a long short-term memory network (LSTM). CNN is used to extract spatial features, and LSTM is used to process time series data. Multi-scale time series modeling processing refers to feature extraction and time series analysis of feature vectors at different time scales. Optimizing feature vectors refers to the feature representation that integrates spatiotemporal information after multi-scale processing. Multi-task joint prediction processing refers to the model simultaneously completing the prediction of multiple related tasks, such as leakage probability and leakage level prediction. Pipeline leakage information is used to represent comprehensive detection results including leakage probability, level, etc.

[0027] In the above embodiment, the fused pipeline leakage feature vector is input into a CNN-LSTM hybrid model. This model performs multi-scale time series modeling on the feature vector: three convolution kernels of different sizes are used in parallel to process the feature vector. A small convolution kernel (3×1) captures short-term burst features, such as high-frequency vibration at the moment of leakage; a medium convolution kernel (11×1) extracts medium-term pattern features, such as periodic vibration caused by water flow fluctuations; and a large convolution kernel (19×1) captures long-term trend features, such as chronic leakage caused by pipeline aging. A two-layer bidirectional LSTM network is used to perform time series modeling on the multi-scale features. The forward LSTM learns temporal dependencies from the past to the present, while the backward LSTM captures temporal information from the present to the future. An attention mechanism automatically assigns weights to different time steps. For example, when processing the feature vector of a monitoring point, the model, through multi-scale analysis, discovers high-frequency vibration features in the short term, periodic fluctuations in the medium term, and a gradually increasing trend in the long term, comprehensively judging the presence of a developing leakage fault.

[0028] In the above embodiment, the CNN-LSTM hybrid model decomposes the optimized feature vector into three tasks: leak detection, determining the presence of a leak (binary classification); leak location, determining the leak location (regression prediction); and leak severity assessment, assessing the severity of the leak (multi-classification). The optimized feature vector is input into a shared fully connected layer, where a Swish activation function is used to enhance nonlinear representation and a Dropout layer is added to reduce overfitting. The leak detection layer uses a Sigmoid activation function to output the leak probability, the leak location layer uses a linear activation function to predict the leak location coordinates, and the leak severity assessment layer uses a Softmax activation function to output the severity probability distribution. The weights of each task are dynamically adjusted based on the network's operating status, increasing the weight of the leak detection task during peak water usage periods and the weights of the leak location and severity assessment tasks during nighttime off-peak periods. For example, in one prediction, the model output: leak probability 0.92 (high risk), leak location 235 meters from the monitoring point, severity "medium," and immediate investigation recommended.

[0029] In the above embodiment, the topology of the water supply network is constructed using a graph neural network, using the feature vectors of each monitoring point as node features and pipeline connections as edges. A graph convolutional network (GCN) is used to learn the spatial dependencies between monitoring points, and combined with an LSTM to process time series information, a spatiotemporal graph neural network is formed. This structure can capture the mutual influence between distant monitoring points in the network, improving detection capabilities for complex leak scenarios. The same CNN-LSTM model is deployed in water supply networks across multiple cities, and collaborative training is performed using a federated learning framework. Each city trains the model using local data, only uploading model parameter updates rather than raw data, protecting privacy while sharing knowledge. Through federated learning, the model learns common features across different city networks while retaining local adaptability, significantly improving cross-regional leak detection performance. Monte Carlo dropout technology is introduced to quantify the uncertainty of model predictions and calculate the confidence level of each prediction result. When the prediction uncertainty exceeds a threshold, an active learning mechanism is triggered, automatically labeling these samples and requesting manual verification. The verified data is then added to the training set. This approach efficiently utilizes limited annotation resources and gradually improves model performance on difficult samples. Based on multi-task prediction, a causal reasoning model is introduced to analyze leak causes. A causal graph of the water supply network is constructed, including nodes such as pipe material, age, pressure changes, and environmental factors. When a leak is detected, causal reasoning is used to trace the most likely cause, such as "pipeline aging" or "abnormal pressure," providing a more comprehensive basis for repair decisions.

[0030] In an optional embodiment, when the CNN-LSTM hybrid model determines that a pipeline leakage feature vector has been received, the CNN-LSTM hybrid model performs multi-scale time series modeling on the pipeline leakage feature vector to obtain an optimized feature vector, specifically including: the CNN-LSTM hybrid model uses three groups of parallel convolution kernels of different scales to perform multi-scale feature extraction on the pipeline leakage feature vector to obtain local time-frequency features; the CNN-LSTM hybrid model performs a maximum pooling operation on the local time-frequency features to obtain dimensionality reduction features; the CNN-LSTM hybrid model uses a two-layer bidirectional LSTM network to perform time series modeling on the dimensionality reduction features to obtain time series correlation features, wherein the two-layer bidirectional LSTM network includes a first LSTM layer for forward propagation and a second LSTM layer for back propagation; the CNN-LSTM hybrid model uses a self-attention layer to perform weighted processing on the time series correlation features to obtain weight features; the CNN-LSTM hybrid model uses a fully connected layer to perform nonlinear mapping on the weight features to obtain an optimized feature vector.

[0031] In the above embodiment, parallel convolution kernels represent multiple groups of convolution kernels of different scales (such as 1×3, 1×11, 1×19, etc.) working simultaneously in CNN, which are used to extract multi-scale features; local time-frequency features refer to the frequency distribution characteristics of the signal in the local time window; the maximum pooling operation refers to taking the maximum value of the feature map to reduce the dimension and retain the key features; the dimensionality reduction feature is used to represent the low-dimensional feature representation after pooling; the bidirectional LSTM network refers to a network composed of forward and backward LSTM layers, which can simultaneously capture past and future time series information; the time series correlation feature represents the dependency relationship feature of the signal in the time series; the self-attention layer refers to a mechanism for weighting features by calculating the correlation between features; the weight feature refers to the feature after attention weighting; the fully connected layer refers to a network layer in which all neurons are connected, which is used for nonlinear mapping.

[0032] In the aforementioned embodiment, during water supply network monitoring in the coastal city, the pipeline leakage feature vector obtained by fusion of multidimensional features was input into a CNN-LSTM hybrid model. The CNN-LSTM hybrid model activates three sets of parallel convolution kernels of different sizes. Small convolution kernels sensitively capture short-term local features, such as high-frequency vibrations generated by sudden pipeline leaks; medium convolution kernels focus on analyzing medium-term pattern features, such as periodic vibrations caused by unstable water flow; and large convolution kernels focus on exploring long-term trend features of chronic leaks caused by pipeline aging. This generates rich local time-frequency features. A maximum pooling operation processes the local time-frequency features, reducing the data dimension while retaining key information, generating reduced-dimensionality features. The reduced-dimensionality features are then fed into a two-layer bidirectional LSTM network. The forward LSTM layer learns the dependencies between features from the past to the present, while the backward LSTM layer captures time series information from the present to the future. Together, the two layers perform deep time series modeling on the features, generating features rich in temporal correlations. A self-attention layer weights the temporal correlation features, assigning weights based on the importance of features at different time steps to leak detection, highlighting key information and forming weighted features. The fully connected layer implements nonlinear mapping on the weight features and converts them into the final optimized feature vector through complex calculations of multiple layers of neurons, providing an accurate basis for subsequent judgment of pipeline leakage.

[0033] In the above-described embodiment, an adaptive convolution adjustment strategy is implemented during the operation of the CNN-LSTM hybrid model. The model analyzes the fluctuations, frequency distribution, and complexity of the input feature vector in real time. If an increase in high-frequency components in the feature vector is detected, indicating a suspected sudden leak, the weight of the small convolution kernel in multi-scale feature extraction is automatically increased to enhance the capture of high-frequency details. If the feature changes are relatively gradual, indicating the possibility of chronic leaks, the weight of the large convolution kernel is increased. A hierarchical attention enhancement mechanism is also introduced, establishing a special connection between the two layers of bidirectional LSTM networks. The feature information output by the first-layer LSTM is computationally filtered to identify key components, forming a context vector. This vector, along with the output of the first layer, serves as the input to the second-layer LSTM, allowing the second-layer network to focus more on features that are important for leak detection, thereby improving the model's sensitivity to critical leak information. A dynamic timing adaptation scheme is designed. The CNN-LSTM hybrid model automatically identifies periodic patterns in the signal based on the autocorrelation characteristics of the input features and adjusts the time step of the LSTM network processing data. When a leakage vibration signal with significant periodicity is detected, the time step is adjusted to a value that matches the signal period, allowing the CNN-LSTM hybrid model to more efficiently capture the changing patterns of the leakage signal. An adversarial perturbation training phase has also been added. During the CNN-LSTM hybrid model training process, varying degrees of simulated noise and interference are added to the input data. This allows the CNN-LSTM hybrid model to learn to accurately extract leakage features in complex interference environments, enhancing the model's anti-interference capabilities and detection stability under actual complex working conditions. This allows the model to accurately determine pipeline leakage even in the presence of strong environmental noise.

[0034] In an optional embodiment, the CNN-LSTM hybrid model performs multi-task joint prediction processing on the optimized feature vector to obtain pipeline leakage information, specifically including: the CNN-LSTM hybrid model flattens the optimized feature vector and inputs it into the shared fully connected layer for a first nonlinear mapping to obtain shared features; the CNN-LSTM hybrid model inputs the shared features into the first task-specific layer for a second nonlinear transformation to obtain first task features; the CNN-LSTM hybrid model uses a Sigmoid activation function to perform binary classification processing on the first task features to output the pipeline leakage probability, wherein the pipeline leakage information includes the pipeline leakage probability; the CNN-LSTM hybrid model inputs the shared features into the second task-specific layer for a third nonlinear transformation to obtain the second task features; the CNN-LSTM hybrid model uses a Softmax activation function to perform multi-classification processing on the second task features to output the pipeline leakage level, wherein the pipeline leakage information includes the pipeline leakage level.

[0035] In the above embodiment, the flattening operation refers to converting a multi-dimensional feature vector into a one-dimensional vector for processing by the fully connected layer; the shared fully connected layer refers to a fully connected layer used by multiple tasks in multi-task learning to extract common features; shared features refer to common feature representations; the task-specific layer refers to a dedicated network layer designed for different tasks (such as binary classification and multi-classification); nonlinear transformation refers to the introduction of nonlinear relationships through activation functions (such as ReLU functions, etc.); the Sigmoid activation function refers to a function that maps the output to the interval [0,1] and is suitable for binary classification; the Softmax activation function refers to a function that maps the output to a probability distribution and is suitable for multi-classification; the pipeline leakage probability is used to represent the numerical probability of pipeline leakage; the pipeline leakage level refers to the classification of the leakage severity into multiple levels (such as mild, moderate, and severe).

[0036] In the aforementioned embodiment, in the coastal city's water supply network monitoring system, after the CNN-LSTM hybrid model completes multi-scale time series modeling of the pipeline leakage feature vector and generates the optimized feature vector, the prediction and decision-making phase begins. The optimized feature vector is flattened and then fed into the shared fully connected layer. This layer acts as an "information filter." Through a first nonlinear mapping, it extracts common key information related to pipeline leakage from complex features, forming shared features. These shared features contain essential clues for determining whether a pipeline is leaking, locating the leak, and assessing the leak severity. The shared features are divided into two paths. One path enters the first task-specific layer, where it undergoes a second nonlinear transformation to further mine and enhance relevant features for the task of determining whether a pipeline is leaking, resulting in the first task feature. The first task feature is processed using a sigmoid activation function, which maps the feature value to a range between 0 and 1 and outputs a specific value as the pipeline leakage probability. For example, a value close to 1 indicates a high probability of pipeline leakage; a value close to 0 indicates minimal leakage risk. The other shared feature enters the second task-specific layer and undergoes a third nonlinear transformation to produce a second task feature specifically designed for assessing pipeline leak levels. A softmax activation function then transforms the second task feature into a probability distribution for each leak level, outputting the pipeline leak level. For example, the output might show a probability of 0.2 for a minor leak, 0.6 for a moderate leak, and 0.2 for a severe leak, providing a clear assessment of the severity of the pipeline leak.

[0037] In the above embodiment, a dynamic threshold adjustment mechanism is introduced in the shared fully connected layer. The threshold output by the shared fully connected layer is dynamically adjusted based on the real-time operating status of the water supply network, such as peak and low water usage periods, as well as the distribution patterns of historical leakage data. During peak water usage periods, due to complex pipeline pressure fluctuations, the threshold for determining leak probability may be raised to avoid misjudgments caused by pressure fluctuations. During low water usage periods, the threshold is appropriately lowered to capture even the slightest leak sign. A feature interaction module is implemented between the task-specific layers, enabling the features generated by the first and second task-specific layers to communicate with each other. For example, some feature information obtained when determining leak status is passed to the second task-specific layer assessing leak severity, assisting it in more accurately determining the leak severity, and vice versa. This feature interaction breaks down information barriers between tasks, allowing the model to share key clues when making predictions for different tasks, improving prediction accuracy and relevance. Furthermore, the parameters of the activation function are adaptively adjusted. The Sigmoid and Softmax functions are not static; instead, they automatically adjust their parameters based on the distribution of input features. When the input features fluctuate greatly, the Sigmoid function adjusts the slope of the curve to make it more sensitive to small changes, thereby outputting the leakage probability more accurately; the Softmax function optimizes the calculation method of the probability distribution based on the degree of difference in the features of different leakage levels, ensuring that the judgment of the leakage level is more reasonable and reliable.

[0038] In an optional embodiment, after the CNN-LSTM hybrid model uses the Softmax activation function to perform multi-classification processing on the second task feature to output the pipeline leakage level, the method also includes: extracting a standard leakage voiceprint feature corresponding to the pipeline leakage level from a preset leakage voiceprint feature library; matching the leakage voiceprint feature in the pipeline leakage feature vector with the standard leakage voiceprint feature to obtain the pipeline leakage type; when it is determined that the pipeline leakage type is a micro leakage type, determining the low-frequency energy proportion in the leakage voiceprint feature to obtain a low-frequency energy proportion result; when it is determined that the low-frequency energy proportion is greater than a first preset threshold according to the low-frequency energy proportion result, determining that the cause of the pipeline leakage is aging of the interface rubber seal; when it is determined that the low-frequency energy proportion is less than or equal to the first preset threshold according to the low-frequency energy proportion result. When the pipeline leakage type is determined to be a medium leakage type, the mid-frequency energy distribution in the leakage soundprint feature is determined to obtain the mid-frequency energy distribution result; when the standard deviation of the mid-frequency energy distribution is determined to be greater than the second preset threshold value according to the mid-frequency energy distribution result, the leakage cause is determined to be pipeline corrosion pits; when the pipeline leakage type is determined to be a severe leakage type, whether there is a high-frequency burst peak in the leakage soundprint feature is detected; when a high-frequency burst peak is detected in the leakage soundprint feature and the amplitude of the high-frequency burst peak is greater than the third preset threshold value, the leakage cause is determined to be pipe body rupture; fault propagation prediction is performed according to the pipeline leakage cause to obtain the leakage expansion risk level; a priority-ordered repair strategy set is generated according to the pipeline leakage level and the leakage expansion risk level.

[0039] In the above embodiment, the preset leakage voiceprint feature library refers to a pre-established voiceprint feature database containing different leakage types and levels; the standard leakage voiceprint feature represents the typical voiceprint feature corresponding to the leakage level; the pipeline leakage type refers to the category divided according to the leakage characteristics (such as micro leakage, medium leakage, severe leakage); the low-frequency band energy proportion is used to indicate the proportion of the energy of the low-frequency part (such as 0-50Hz, etc.) in the leakage voiceprint to the total energy; the medium-frequency band energy distribution refers to the energy distribution characteristics of the medium-frequency part (such as 50-500Hz, etc.); the high-frequency burst peak represents the instantaneous energy peak in the high-frequency band (such as above 500Hz, etc.); the fault propagation prediction refers to the prediction of the leakage development trend based on the cause of the leakage; the leakage expansion risk level refers to the assessment of the degree of harm that the leakage may cause; the repair strategy set refers to the priority list of maintenance plans generated according to the leakage level and risk.

[0040] In the above embodiment, during routine monitoring of a coastal city's water supply network, pipeline leak feature vectors obtained through fiber optic vibration acquisition and multidimensional feature fusion are input into the subsequent analysis process. A standard leak soundprint signature corresponding to the currently detected pipeline leak level is quickly extracted from a preset leak soundprint signature library. The standard leak soundprint signature is carefully matched with the leak soundprint signature in the pipeline leak feature vector. When the pipeline leak is determined to be a microleak, the low-frequency band of the leak soundprint signature is focused on and its energy contribution is accurately calculated. When the calculated low-frequency band energy contribution exceeds a first preset threshold, the cause of the pipeline leak is determined to be aging of the rubber seal at the interface. Long-term wind erosion and seawater backflow accelerate the aging of rubber seals, resulting in tiny cracks at the interface. The resulting vibration signal exhibits a significant energy concentration in the low-frequency band. If the low-frequency band energy contribution is less than or equal to the first preset threshold, the cause of the leak is determined to be micropores in the pipeline weld. This is because the vibration signal generated by the water pressure from the micropores in the weld has relatively weak energy in the low-frequency band.

[0041] In the above embodiment, assuming that a certain detection determines that the pipeline leak type is a medium leak, the mid-frequency energy distribution of the leakage soundprint feature is analyzed. When it is found that the standard deviation of the mid-frequency energy distribution is greater than the second preset threshold, it means that the energy distribution fluctuates greatly. Combined with the actual working conditions, it is determined that the leak is caused by pipeline corrosion pits. Over time, the pits formed by pipeline corrosion continue to expand, and the vibrations generated by the impact of water flow on the inner wall of the pit present an unstable energy distribution in the mid-frequency band. If the pipeline leak type is detected as a severe leak, the leakage soundprint feature will be carefully checked for high-frequency burst peaks. Once a high-frequency burst peak is captured and its amplitude exceeds the third preset threshold, the cause of the leak is immediately determined to be a pipe rupture. Under the impact of strong water flow, strong vibrations will be generated at the moment of pipe rupture, and this vibration will appear in the soundprint feature in the form of a high-frequency burst peak.

[0042] In the above embodiment, after determining the cause of the pipeline leakage, the fault propagation is predicted based on the actual conditions such as the pipeline material, service life, and surrounding water pressure. Through comprehensive evaluation, the leakage expansion risk level is obtained. For example, for serious leakage caused by pipe rupture, considering that the surrounding water pressure is large and the pipeline material is aged, the leakage expansion risk level is assessed to be high. Finally, a priority-ordered repair strategy set is generated based on the pipeline leakage level and the leakage expansion risk level. For high-risk serious leaks, professional repair teams are given priority to carry emergency pipes to the site for emergency plugging and replacement; for minor leaks with lower risks, a detailed maintenance plan is formulated to arrange for seal replacement or weld repair work during non-water usage peak periods.

[0043] In the above embodiment, a dynamic weight adjustment strategy is introduced into the matching process. The weights of various features in each dimension are automatically adjusted when matching the pipeline leakage feature vector with the standard leak voiceprint signature based on different time periods (e.g., peak daytime water usage and low nighttime water usage) and different environmental conditions (e.g., heavy rain and high temperatures). For example, during heavy rain, ambient noise may interfere with the voiceprint signature. In this case, the weights of features that are most affected by noise are reduced, while the weights of relatively stable features are increased, thereby more accurately determining the leak type. During the fault propagation prediction phase, geographic information surrounding the pipeline is incorporated. If there are engineering activities such as subway construction or road excavation near the pipeline, these external factors are taken into account and the leak propagation risk level is reassessed. For example, if subway construction is ongoing near a leaking pipeline with a rupture, the vibrations generated by the construction may accelerate pipeline damage. The system will accordingly increase the leak propagation risk level and prioritize protective measures in the repair strategy set to prevent the interaction between construction and the leak from causing a more serious accident. When generating the repair strategy set, the real-time scheduling of maintenance resources is taken into account. When the number of repair teams is limited, the priority of repair strategies is optimized based on the distance between each leak point and the repair base and the inventory of required repair materials. Priority will be given to repairing leaks that are close and have sufficient repair materials. At the same time, repair resources from other areas will be coordinated to support urgent and resource-scarce leaks, thereby improving overall repair efficiency.

[0044] In an optional embodiment, the dual-channel optical fiber vibration signal is synchronously sampled using the FPGA master clock to obtain time-aligned pipeline vibration signals and environmental noise signals, specifically including: controlling the FPGA master clock to send a Sync message to the vibration acquisition terminal, and recording the first sending timestamp of the Sync message sent by the FPGA master clock; when it is determined that the vibration acquisition terminal receives the Sync message, recording the first receiving timestamp of the Sync message received by the vibration acquisition terminal; controlling the vibration acquisition terminal to return a Delay_Req message to the FPGA master clock, and recording the second sending timestamp of the Delay_Req message returned by the vibration acquisition terminal; when it is determined that the vibration acquisition terminal receives the Sync message, recording the first receiving timestamp of the Sync message received by the vibration acquisition terminal; controlling the vibration acquisition terminal to return a Delay_Req message to the FPGA master clock, and recording the second sending timestamp of the Delay_Req message returned by the vibration acquisition terminal; When the FPGA master clock receives the Delay_Req message, the second reception timestamp of the FPGA master clock receiving the Delay_Req message is recorded; the link delay time is calculated based on the first transmission timestamp, the first reception timestamp, the second transmission timestamp, and the second reception timestamp, where the link delay time is used to represent the signal transmission delay between the FPGA master clock and the vibration acquisition terminal; the initial pipeline vibration signal and the initial ambient noise signal are time-compensated based on the link delay time, and the compensated initial pipeline vibration signal and the initial ambient noise signal are interpolated and aligned to obtain time-aligned pipeline vibration signal and ambient noise signal.

[0045] In the above embodiment, the Sync message refers to the synchronization signal message sent by the FPGA master clock, which is used to trigger sampling and record the timestamp; the Delay_Req message refers to the delay request message returned by the vibration acquisition terminal, which is used to measure the round-trip time of the signal; the first transmission timestamp indicates the moment when the FPGA master clock sends the Sync message; the first reception timestamp indicates the moment when the terminal receives the Sync message; the second transmission timestamp indicates the moment when the terminal returns the Delay_Req message; the second reception timestamp indicates the moment when the FPGA master clock receives the Delay_Req message; the link delay time refers to the transmission delay of the signal between the FPGA master clock and the vibration acquisition terminal; time compensation processing refers to the time offset correction of the sampled data based on the delay time; interpolation alignment processing refers to the time axis alignment of the compensated signal using an interpolation algorithm to ensure strict correspondence between the sampling points. In the above coastal city water supply network monitoring system, precise time synchronization between the vibration acquisition terminals distributed at each monitoring point and the FPGA master clock is key to ensuring effective data acquisition. The FPGA master clock first sends the Sync message to the vibration acquisition terminal, and the FPGA master clock's high-precision timer immediately records this transmission time, generating the first transmission timestamp. Once the vibration collection terminal receives the Sync message, it responds quickly. Its internal time recording module immediately records the moment of receipt, forming the first receive timestamp. The vibration collection terminal then transmits the Delay_Req message back to the FPGA master clock and records the second transmit timestamp. After receiving the Delay_Req message, the FPGA master clock again uses a timer to record the second receive timestamp. This series of operations fully records the key timing points of the signal's round-trip transmission between the two terminals. Subsequently, based on these four timestamps, a specific calculation logic is used to accurately calculate the link delay. This time accurately reflects the signal transmission delay between the FPGA master clock and the vibration collection terminal. With this link delay, time compensation can be performed on the initial pipeline vibration signal and the initial ambient noise signal. By correcting the time offset caused by transmission delay, interpolation and alignment are then performed to obtain perfectly time-aligned pipeline vibration and ambient noise signals, laying a solid foundation for subsequent signal analysis.

[0046] In the above embodiment, a dynamic link monitoring mechanism is introduced. The changing trend of the link delay time is continuously monitored. Once an abnormal fluctuation in the delay time is found, such as a substantial increase in a short period of time, more frequent Sync message and Delay_Req message interactions are automatically triggered to obtain a more accurate real-time link delay time. At the same time, combined with the environmental factors around the water supply network, such as whether there is large-scale construction nearby that causes electromagnetic interference, the parameters of time compensation and interpolation alignment are dynamically adjusted to ensure that accurate time alignment of signals can be achieved even in complex environments. A multi-link redundant backup strategy can also be added. Multiple communication links are set up at each monitoring point to connect the vibration collection terminal and the FPGA main clock. When the link delay time of the main link exceeds a certain threshold, it automatically switches to a backup link with lower delay and greater stability to ensure the timeliness and stability of signal transmission. In addition, after switching the link, the delay time of the new link is quickly calculated and the signal time compensation is adjusted to ensure that data collection is not affected by link switching.

[0047] It should be noted that the above-described embodiments are only part of the embodiments of the present application, rather than all the embodiments. The present application will be described in detail below with reference to specific embodiments.

[0048] The embodiment of the present application provides a system architecture. Figure 2 This is a schematic diagram of a hierarchical architecture of the optical fiber vibration auscultation system in the embodiment of the present application, see Figure 2 , the system consists of three parts: The vibration sensing layer (using IoT optical fibers to obtain electrical signals converted from Rayleigh scattered light, with fixed laser wavelength and fiber refractive index) is installed along the water supply network. One adhesive vibration optical fiber, encapsulated with high-pressure silicone (compression resistance ≥ 2 MPa), adheres closely to the pipe surface and senses low-frequency vibrations (0.1-200 Hz) induced by leaking water pressure. The other optical fiber, installed overhead at a distance of 10 cm (9 cm, 10.5 cm, 11 cm, etc., are also possible; this is not a limitation), collects ambient noise and uses an improved LMS adaptive filtering algorithm to offset common-mode interference. FPGA-based time synchronization is implemented at the μs level to ensure time-domain alignment of vibration and noise signals.

[0049] FPGA acts as the master clock, sending Sync messages to the vibration / noise acquisition terminal through the optical fiber network and recording the sending timestamp. , record the reception timestamp from the device (such as a vibration sensor) , and reply with a Delay_Req message, recording the reply timestamp , FPGA records the receiving timestamp , calculate the link delay using the following formula :

[0050] To compensate for time deviation, the vibration and noise signals are interpolated and aligned using the following formula:

[0051] in, is the aligned signal sequence, which represents the signal value after time compensation and interpolation processing; is the original sampling signal sequence, which represents the unprocessed vibration or noise signal value; is the index of the sampling point, used to traverse the original signal sequence; is the new sampling point index, indicating the time point of the aligned signal; is the time deviation, which indicates the amount of time delay that needs to be compensated; is the sampling interval, which represents the time interval between adjacent sampling points. sinc is the Sinker function, which is used for signal interpolation and reconstruction.

[0052] Data extraction layer (data cleaning of collected signals, identification and removal of erroneous data; preprocessing to provide standardized data for subsequent analysis; storage of data considering flexibility and analytical capabilities; data management to ensure data security, accessibility and efficient use): Extract the multi-dimensional features of the vibration signal, analyze the leakage impact intensity and frequency through time domain features, analyze the leakage soundprint spectrum characteristics through frequency domain features, and analyze the phase change caused by optical fiber strain through optical features.

[0053] Analysis and decision layer (feature recognition through algorithms, stacking bidirectional LSTM to learn the before and after correlation features of leakage events, nonlinear mapping through fully connected layers, using sigmoid activation function, and outputting binary classification probabilities): AI feature fusion is used to construct a CNN-LSTM hybrid neural network. The input layer is a multi-source feature vector, and the output layer is the leakage probability and leakage level. The training data set contains typical working condition samples.

[0054] Figure 3 This is a hardware diagram of the optical fiber vibration auscultation system in the embodiment of the present application, see Figure 3The hardware components include: Laser transmitter: used to generate high-precision laser pulse signals and control the wavelength, power and pulse characteristics of the laser. Distributed optical fiber: collects vibration signals and monitors the status of the water supply pipeline by sensing changes in the vibration, sound and other aspects of the water supply pipeline. Water supply pipeline: monitoring object, used to transport and distribute water supply, in which possible leakage, damage and other problems will generate characteristic vibrations and acoustic signals. The material, pressure, flow and other characteristics of the pipeline will affect the characteristics of the vibration signal. Spectrum analyzer: used to receive the scattered light signal returned from the optical fiber, analyze the spectral characteristics of the optical signal, process the Rayleigh scattering signal, and extract vibration information. Photoelectric converter: used to convert the optical signal into an electrical signal, perform signal conditioning and preprocessing, and provide standard electrical signal input for subsequent analysis. The stethoscope system (the terminal analysis platform in the optical fiber vibration stethoscope system) is used to receive the signal after photoelectric conversion, perform signal processing and analysis, realize leak detection and positioning, and provide data display and early warning functions. Signal flow: Distributed optical fiber is laid along the water supply network. A pulsed laser transmitter is used to emit periodic pulses into the optical fiber. The scattered light (Rayleigh) information is obtained through a light wave analyzer. The optical signal is converted into an electrical signal through a photoelectric converter and provided to the stethoscope system for leak diagnosis.

[0055] Figure 4 This is a functional module diagram of the optical fiber vibration auscultation system in the embodiment of the present application, see Figure 4 , the functional module includes: Data acquisition module: collects on-site signal data and transmits it to the stethoscope system in the form of electrical signals.

[0056] Data extraction module: Analyze and screen signal features to identify key features.

[0057] Analysis and identification module: identifies normal and different leakage states.

[0058] Alarm contact module: monitors and analyzes identification results. If there are any anomalies, an alarm message is generated and pushed to the contact platform for reminders; Figure 5 This is a functional diagram of the alarm access module in the embodiment of the present application, see Figure 5This module includes three sub-functions: Alarm monitoring and triggering function: real-time monitoring of auscultation system data, setting alarm thresholds and triggering conditions, and automatically triggering the alarm mechanism when abnormal data is detected to ensure that abnormal situations can be discovered in time. Alarm classification analysis function: classify the severity of alarm events (such as critical, important, and ordinary levels), analyze the scale and potential impact of leaks, evaluate the processing priority, and provide processing suggestions at different levels. Alarm contact notification function: select different notification methods (such as SMS, email, APP push, etc.) according to the alarm level, and send the alarm information to the relevant responsible persons. Ensure that the information is delivered to the corresponding processing personnel in a timely manner and track the alarm processing status.

[0059] Diagnosis and analysis module: Analyze the cause of leakage according to the leakage situation; Figure 6 This is a functional diagram of the diagnostic analysis module in the embodiment of the present application, see Figure 6 This module includes three sub-functions: Operation abnormality monitoring function: monitors the operating status of the water supply pipeline, identifies abnormal conditions that deviate from normal operating parameters, tracks changes in vibration, acoustics and other signals of the water supply pipeline in real time, and determines the time and location of the abnormality. Diagnostic result analysis function: conducts in-depth analysis of detected water supply pipeline abnormalities, determines the type and severity of the leak, evaluates the impact of the leak on the water supply, and generates a diagnostic report. Event root cause analysis function: traces the cause of the water supply pipeline leak, analyzes the influence of factors such as pipeline material, water pressure, and flow, determines the specific cause of the leak, and provides a basis for subsequent maintenance and prevention.

[0060] Recommendation decision module: recommends corresponding corrective measures based on the cause of the leak.

[0061] Data visualization module: Visual display of abnormal operation data and system diagnosis results.

[0062] The present application also provides a fiber optic vibration auscultation process for water supply network leakage, which includes the following steps: Step 1: Water supply network leak detection based on phase-sensitive optical time domain reflectometry (Φ-OTDR). Figure 7 This is a schematic diagram of the principle of optical fiber Rayleigh scattering measurement in the embodiment of the present application, see Figure 7, demonstrating the core principle of phase-sensitive optical time-domain reflectometry (Φ-OTDR) in water supply pipeline leak detection. That is, a pulsed laser emits a narrow-linewidth pulsed laser into the sensing optical fiber. When the light propagates inside the optical fiber, it elastically collides with the molecular structure and produces Rayleigh scattering. When a leak occurs in the pipeline, the water pressure vibration at the leak point is transmitted through the pipe wall to the optical fiber close to the pipe, causing local strain in the optical fiber. According to the elasto-optic effect, the strain causes the refractive index of the optical fiber to change, thereby modulating the phase of the backscattered Rayleigh scattered light. The backscattered light carrying the phase change information returns along the original path and undergoes differential interference with the local oscillator light through the photodetector, converting the phase difference Δφ into an interference light intensity fluctuation signal. The analyzer accurately calculates the leak location based on the time difference between the light pulse emission and the scattered light return, and simultaneously analyzes the vibration characteristics in the interference light intensity signal to achieve highly sensitive identification and positioning of the leak event: When a water supply pipe leaks, the vibrations generated by the water pressure at the leaking point are transmitted through the pipe wall to the sensing fiber, causing local strain in the fiber. According to the elasto-optical effect of the fiber, the strain changes the refractive index distribution of the fiber, thereby causing a phase delay change in the backscattered Rayleigh light. The relationship between the strain and the

[0063] in, is the laser wavelength, is the refractive index of the optical fiber, is the initial length of the optical fiber, is the change in fiber length, is the refractive index change.

[0064] The phase-intensity conversion is achieved by using differential delayed interferometry technology. Ultra-narrow linewidth pulsed lasers are used, and their high coherence enables the scattered light waves in the pulse region to form stable interference. The backscattered Rayleigh light and the local oscillator light (delay time matching) are superimposed on the photodetector, and the interference light intensity is:

[0065] in, Interference light intensity, which represents the final measured light intensity signal, is the scattered light field intensity, is the local oscillator light field intensity, is the light field amplitude, is the cosine function of the phase difference.

[0066] According to the light pulse propagation time ( ) determines the vibration location and constructs a time-domain vibration signal from a continuous pulse sequence. Noise suppression uses an improved LMS adaptive filtering algorithm, utilizing suspended noisy optical fibers to collect ambient noise and dynamically cancel common-mode interference. A pre-trained pipeline vibration signature library supports adaptive updates for cast iron and PE pipeline noise.

[0067] The output time series data of the Φ-OTDR is generated by measuring the phase and intensity changes of the Rayleigh backscattered light, and includes the following: Backscattered Rayleigh light intensity sequence: When the probe light pulse propagates in the optical fiber, the backscattered light intensity caused by Rayleigh scattering changes with time, reflecting the scattering characteristics of each position in the optical fiber.

[0068] Phase change sequence: External vibration or strain causes the refractive index or length of the optical fiber to change, causing phase fluctuations in the backscattered Rayleigh light. , and the phase time series is obtained through coherent detection demodulation.

[0069] Time domain positioning information: Each data point in the time series corresponds to a spatial position on the optical fiber, and the distance is calculated by the speed of light and the time delay. ), the spatial features of the target, such as contour features and amplitude features, can be extracted using the CNN convolutional neural network. However, due to the fixed length of the window function of the traditional CNN, there will be a problem that the frequency resolution and time resolution cannot be well balanced. If the scale and span of the convolution kernel are too small, the time resolution of the signal is better and it is more sensitive to changes in high-frequency features, but it cannot learn the low-frequency features in the signal well. On the contrary, a larger-scale convolution kernel corresponds to a larger span, which can learn information over a longer time range, that is, the low-frequency features in the signal, but cannot reflect the high-frequency characteristics therein well. Therefore, the present invention adopts multiple one-dimensional convolution kernels of different scales, and performs parallel convolution with the input information to achieve the extraction of signal features at different time scales.

[0070] In addition to spatial features, the Φ-OTDR's output signal also exhibits temporal characteristics. The characteristic pulses on the curve are closely related to their time and preceding and following states. Therefore, based on the CNN model, the LSTM model is further employed to effectively extract the signal's temporal characteristics. The LSTM uses a gating mechanism to control the flow of information, allowing the model to decide what information to retain, update, and forget. It maintains long-term memory through cellular states, using sampling points within a time window as network input. The network then obtains predicted values based on the sampling points within the window. By moving the window with a set step size, the network continuously predicts the sequence, thereby extracting the signal's temporal characteristics.

[0071] Step 2, Figure 8 This is a structural diagram of the CNN-LSTM hybrid model in the embodiment of the present application, see Figure 8 The CNN-LSTM hybrid model combines the strengths of CNN and LSTM, and can simultaneously obtain the signal's contour features, amplitude features, and time series features. Through feature fusion, it effectively improves the model's sensitivity and accuracy, and achieves leakage / non-leakage event classification and leakage level prediction through multi-task learning: 1) Input layer processing: The top Input layer ( ) is the input of time series data (i.e., the time series of the Φ-OTDR output signal). This time series data undergoes preliminary feature extraction through two layers of CNN and Pooling. CNN Layer 1 (the first convolutional layer) uses convolution kernels (filters) to extract local spatial features of the input data, enabling the identification of low-level features. Pooling Layer 1 downsamples the features output by the convolutional layer, reducing the spatial dimension and thus reducing computational effort. CNN Layer 2 (the second convolutional layer) further extracts higher-order spatial features. By stacking multiple layers of convolution, it combines low-level features (such as edges) to form more complex patterns. Pooling Layer 2 compresses features, further reducing parameters and improving the model's robustness to small changes in the input. Deep pooling captures more global spatiotemporal features. The fully-connected layer maps the features of the pooling output to the target space, integrating all features through a weight matrix to achieve nonlinear transformation.

[0072] 2) LSTM Time Series Modeling: Two LSTM layers (LSTM Layer 1 and LSTM Layer 2); each LSTM layer processes sequence data to achieve forward and backward propagation; Represents the hidden state of LSTM and records timing information; the stacking of two layers of LSTM enhances the model's ability to capture long-term dependencies.

[0073] 3) Self-Attention Layer: The circle in the self-attention layer Represents the characteristics of each time step; Represents the attention weight of each time step; the "+" sign indicates that the weighted features are merged.

[0074] 4) Multi-task output processing: The output nodes of the bottom Output layer are actually divided into two groups: A set of binary classifications for leakage / non-leakage (Sigmoid function activation); Another set is used for leaky class multi-classification (Softmax function activation).

[0075] The following is a detailed explanation of CNN feature extraction (layer-by-layer extraction of spatial features (CNN→Pooling→CNN→Pooling)): Three parallel convolution kernels (kernel 1, size 1×3, stride 1; kernel 2, size 1×11, stride 4; kernel 3, size 1×19, stride 7) are used to extract local time-frequency features, capturing the local correlation between high-frequency shocks and low-frequency pressure fluctuations in the leakage signal. Large convolution kernels cover a wide time window of low-frequency pressure fluctuations, extracting long-term pressure fluctuation features. Medium convolution kernels focus on medium-frequency transient events, analyzing localized signal mutations at a medium scale to enhance time-frequency resolution. Small convolution kernels capture high-frequency shock components, using a narrow window to extract microsecond-level transient features.

[0076] Cross-scale feature concatenation: concatenate the three sets of convolution outputs along the channel dimension to form a multi-scale feature tensor , which is fed into the LSTM layer to model temporal dependencies. The SE (Squeeze-Excitation) module dynamically assigns weights to each frequency band to enhance key features.

[0077] High-frequency shock detection: The output of the small core is Hilbert transformed to obtain the envelope. If the envelope peak exceeds three standard deviations of the baseline, a high-frequency event flag is triggered. The envelope mean (baseline_mean) is calculated as the baseline, and the envelope standard deviation (baseline_std) is calculated. The trigger threshold is set to baseline_mean + 3baseline_std. The three standard deviation threshold is based on the statistical principle of "small probability events" and is considered an anomaly when the signal exceeds this threshold.

[0078] Low-frequency correlation analysis: Calculate the cross-correlation function between the large core output and the signal. If the correlation coefficient is greater than 0.7 (of course, it can also be 0.8, 0.85, 0.9, etc., which are not limited here), it is determined to be a leakage correlation event.

[0079] Max pooling (pooling window 2×1) is used to reduce dimensionality while retaining key features (such as the peak of high-frequency impulses, which highlight the amplitude of transient events; the apex of leakage pulses; and the extreme points of low-frequency waveforms, which mark the critical phases of pressure fluctuations, such as peaks and troughs. These key features together form the basis for leakage signal identification). This reduces computational effort.

[0080] Figure 9 This is a schematic diagram of a convolution kernel operation in the embodiment of the present application, see Figure 9 , showing the process of one-dimensional convolution operation: 1) Input part: Input signal: The input signal is a 2×3 matrix with the upper row of data being [abc] and the lower row of data being [xy z]; Convolution kernel (filter) parameters: kernel_size=1×3 (one-dimensional convolution kernel, size is 3), stride=1 (step size is 1, indicating that the convolution kernel moves 1 unit each time).

[0081] 2) Convolution process: The convolution kernel slides from left to right, and each time it calculates the sum of the products of the value in the current window and the convolution kernel. The dotted arrows indicate how the calculation results at each position are obtained.

[0082] 3) Calculation formula: ; ; ; .

[0083] 4) Output: Result: Generates a one-dimensional array containing 4 elements [r1 r2 r3 r4]. It should also be noted that the convolution kernel operation process is existing technology and will not be repeated here.

[0084] The following is a detailed description of LSTM time series modeling (used to receive the flattened feature sequence output by CNN and model dependencies): Figure 10 This is a structural diagram of LSTM time series modeling in the embodiment of the present application, see Figure 10 , stacking two layers of bidirectional LSTM (128 units) captures long-term dependencies, addresses the limitations of traditional LSTM unidirectional propagation, and simultaneously learns the contextual features of leakage events. A self-attention layer is introduced to adaptively assign feature weights, enhance the salience of key time steps, and improve the ability to identify weak leakage signals. The bottom layer input is the input layer: Represents time series data (i.e., the output signal of Φ-OTDR - time series); the middle L1 layer is the first LSTM layer; the upper L2 layer is the second LSTM layer; the top output layer is the output layer: ; Represents the weight connections between layers.

[0085] The following is a detailed description of multi-task output processing: Figure 11 This is a schematic diagram of the architecture of a multi-task learning neural network in the embodiment of the present application, see Figure 11 , the architecture is divided into three main parts: Input layer: Input layer, used to receive raw data.

[0086] Common hidden layers: Shared hidden layers (such as CNN, LSTM, or fully connected layers) extract common features of the input data for use by all tasks.

[0087] Task-specific hidden layers: Task-specific layers (consisting of two fully connected layers, performing task-specific nonlinear transformations on shared features and learning the classification decision boundary) perform specific processing for different tasks. Task 1 is a multi-label classification task, and Task 2 is a classification task. Each task has its own unique hidden layer structure. Arrows indicate the forward propagation direction of feature extraction. Dark gray dots indicate activated output nodes.

[0088] After flattening the LSTM output, the fully connected layer (64 neurons, dropout = 0.5) performs nonlinear mapping. Two output heads are then designed: one for binary classification (the goal of a binary classification task is to classify the input data into two mutually exclusive categories, "positive" and "negative"). The output layer uses a sigmoid activation function and outputs binary classification probabilities (leakage / non-leakage). The other for multi-classification, using a softmax activation function, outputs multi-class probabilities (leakage level). By sharing the model parameters of the hidden layer, these two output heads learn common features, reduce overfitting, improve model generalization and training efficiency, and reduce computational complexity.

[0089] Step three: Develop an engineering early warning system for leaks and provide real-time warnings. A built-in voiceprint feature library containing a variety of typical leak voiceprints and standardized repair plans enables closed-loop management of the entire process from hidden danger identification to intelligent disposal. Micro-leakage (Level I): Fault characteristics: weak leakage vibration disturbance, sound waves manifest as low-frequency energy; Cause analysis: aging of interface rubber seals (Shore hardness change > 20%), micro-pores in pipeline welds (<0.5mm²), etc.; Recommended treatment plan: Use non-curing rubber asphalt grouting to seal (permeability <1×10⁻ 6 cm / s).

[0090] Moderate leakage (Level II): Fault characteristics: Moderate leakage vibration disturbance, sound waves manifest as medium-frequency energy; Cause analysis: Pipeline corrosion pit depth >1mm, etc.; Recommended treatment solution: Injection polyurethane elastomer sealing (curing time 4-6h, elastic modulus ≥10MPa).

[0091] Severe leakage (Level III): Fault characteristics: Severe leakage vibration disturbance, sound waves manifested as high-frequency energy and sudden peaks; Cause analysis: Pipe rupture (crack width > 2mm), etc.; Recommended treatment plan: Trigger the emergency valve closing protocol, close the upstream and downstream butterfly valves, and quickly repair using prefabricated pipe sections.

[0092] The fiber optic vibration auscultation system in the embodiment of the present invention is described below from the perspective of hardware processing. Figure 12 , Figure 12 This is a schematic diagram of the physical device structure of the optical fiber vibration stethoscope system in the embodiment of the present application.

[0093] It should be noted that Figure 12 The structure of the fiber optic vibration stethoscope system shown is only an example and should not limit the functions and scope of use of the embodiments of the present invention.

[0094] like Figure 12 As shown, the fiber optic vibration auscultation system includes a central processing unit (CPU) 1201, which can perform various appropriate actions and processes, such as the methods described in the above embodiments, based on programs stored in a read-only memory (ROM) 1202 or programs loaded from a storage unit 1208 into a random access memory (RAM) 1203. RAM 1203 also stores various programs and data required for system operation. CPU 1201, ROM 1202, and RAM 1203 are interconnected via a bus 1204. An input / output (I / O) interface 1205 is also connected to bus 1204.

[0095] The following components are connected to the I / O interface 1205: an input section 1206 including an audio input device, push button switches, and the like; an output section 1207 including a liquid crystal display (LCD), an audio output device, indicator lights, and the like; a storage section 1208 including a hard disk and the like; and a communication section 1209 including a network interface card such as a LAN (Local Area Network) card or a modem. The communication section 1209 performs communication processing via a network such as the Internet. A drive 1210 is also connected to the I / O interface 1205 as needed. Removable media 1211, such as a magnetic disk, an optical disk, a magneto-optical disk, or a semiconductor memory, is installed in the drive 1210 as needed, so that computer programs read from the removable media can be installed in the storage section 1208 as needed.

[0096] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program including a computer program for executing the methods illustrated in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via the communication section 1209 and / or installed from removable media 1211. When executed by the central processing unit (CPU) 1201, the computer program performs the various functions defined in the present invention.

[0097] It should be noted that specific examples of computer-readable storage media may include, but are not limited to, an electrical connection having one or more conductors, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0098] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present invention. Each box in the flowchart or block diagram can represent a module, program segment, or part of the code, and the above-mentioned module, program segment, or part of the code contains one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in an order different from that marked in the accompanying drawings.

[0099] Specifically, the fiber optic vibration auscultation system of this embodiment includes a processor and a memory. The memory stores a computer program. When the computer program is executed by the processor, the fiber optic vibration auscultation method for water supply network leakage provided in the above embodiment is implemented.

[0100] As another aspect, the present invention further provides a computer-readable storage medium, which may be included in the fiber optic vibration auscultation system described in the above embodiments, or may exist independently and not be incorporated into the fiber optic vibration auscultation system. The storage medium carries one or more computer programs, and when the one or more computer programs are executed by a processor of the fiber optic vibration auscultation system, the fiber optic vibration auscultation system implements the fiber optic vibration auscultation method for detecting water supply network leaks provided in the above embodiments.

[0101] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.

[0102] Those skilled in the art will appreciate that all or part of the process steps in the above-described method embodiments can be implemented by a computer program instructing the relevant hardware. The program can be stored in a computer-readable storage medium, and when executed, the program can include the process steps in the above-described method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM or random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A fiber optic vibration auscultation method for detecting water supply network leakage, characterized in that: include: The dual-channel fiber optic vibration signal is synchronously sampled using the FPGA master clock to obtain time-aligned pipeline vibration signals and ambient noise signals. Performing multi-dimensional feature fusion on the pipeline vibration signal and the ambient noise signal to obtain a pipeline leakage feature vector; The pipeline leakage feature vector is input into a CNN-LSTM hybrid model to obtain pipeline leakage information output after the CNN-LSTM hybrid model performs feature analysis and prediction processing on the pipeline leakage feature vector.

2. The method according to claim 1, characterized in that The performing multi-dimensional feature fusion on the pipeline vibration signal and the ambient noise signal to obtain a pipeline leakage feature vector specifically includes: performing optical time domain reflectometry processing on the pipeline vibration signal to obtain a phase delay variation of the backscattered Rayleigh light; Using an ultra-narrow linewidth pulse laser to generate local oscillation light, and superimposing the back-Rayleigh scattered light with the local oscillation light on a photodetector to obtain an interference light intensity signal; Using an improved LMS adaptive filtering algorithm to filter the interference light intensity signal and the environmental noise signal to obtain a leakage vibration signal; determining, based on the phase delay variation, a fiber strain characteristic of a first fiber channel in the dual-channel fiber optic vibration signal, wherein the first fiber channel is tightly coupled to a water supply pipe and is used to collect the pipe vibration signal, and a second fiber channel in the dual-channel fiber optic vibration signal maintains a preset distance from the water supply pipe and is used to collect the ambient noise signal; Performing time domain feature extraction on the leakage vibration signal to obtain leakage impact features, and performing frequency domain feature extraction on the leakage vibration signal to obtain leakage voiceprint features; The optical fiber strain feature, the leakage impact feature and the leakage soundprint feature are multi-dimensionally fused to obtain the pipeline leakage feature vector.

3. The method according to claim 1, characterized in that Inputting the pipeline leakage feature vector into the CNN-LSTM hybrid model to obtain pipeline leakage information output after the CNN-LSTM hybrid model performs feature analysis and prediction processing on the pipeline leakage feature vector specifically includes: The pipeline leakage feature vector is input into the CNN-LSTM hybrid model, so that the CNN-LSTM hybrid model performs the following operations: When the CNN-LSTM hybrid model determines that the pipeline leakage feature vector is received, the CNN-LSTM hybrid model performs multi-scale time series modeling processing on the pipeline leakage feature vector to obtain an optimized feature vector; The CNN-LSTM hybrid model performs multi-task joint prediction processing on the optimized feature vector to obtain the pipeline leakage information.

4. The method according to claim 3, characterized in that When the CNN-LSTM hybrid model determines that the pipeline leakage feature vector is received, the CNN-LSTM hybrid model performs multi-scale time series modeling processing on the pipeline leakage feature vector to obtain an optimized feature vector, specifically including: The CNN-LSTM hybrid model uses three sets of parallel convolution kernels of different scales to perform multi-scale feature extraction on the pipeline leakage feature vector to obtain local time-frequency features; The CNN-LSTM hybrid model performs a maximum pooling operation on the local time-frequency features to obtain dimensionality reduction features; The CNN-LSTM hybrid model uses a two-layer bidirectional LSTM network to perform time series modeling on the dimensionality reduction features to obtain time series correlation features, wherein the two-layer bidirectional LSTM network includes a first LSTM layer for forward propagation and a second LSTM layer for backward propagation; The CNN-LSTM hybrid model uses a self-attention layer to perform weighted processing on the temporal correlation features to obtain weighted features; The CNN-LSTM hybrid model uses a fully connected layer to perform nonlinear mapping on the weight features to obtain the optimized feature vector.

5. The method according to claim 3, characterized in that The CNN-LSTM hybrid model performs multi-task joint prediction processing on the optimized feature vector to obtain the pipeline leakage information, specifically including: The CNN-LSTM hybrid model flattens the optimized feature vector and inputs it into a shared fully connected layer for a first nonlinear mapping to obtain shared features; The CNN-LSTM hybrid model inputs the shared features into the first task-specific layer for a second nonlinear transformation to obtain the first task features; The CNN-LSTM hybrid model performs binary classification processing on the first task feature using a Sigmoid activation function to output a pipeline leakage probability, wherein the pipeline leakage information includes the pipeline leakage probability; The CNN-LSTM hybrid model inputs the shared features into the second task-specific layer for a third nonlinear transformation to obtain the second task features; The CNN-LSTM hybrid model uses a Softmax activation function to perform multi-classification processing on the second task feature to output a pipeline leakage level, wherein the pipeline leakage information includes the pipeline leakage level.

6. The method according to any one of claims 1 to 5, characterized in that After the CNN-LSTM hybrid model performs multi-classification processing on the second task features using the Softmax activation function to output the pipeline leakage level, the method further includes: Extracting standard leakage soundprint features corresponding to the pipeline leakage level from a preset leakage soundprint feature library; Match the leakage soundprint features in the pipeline leakage feature vector with the standard leakage soundprint features to obtain the pipeline leakage type; When it is determined that the pipeline leakage type is a micro-leakage type, determining the low-frequency energy ratio in the leakage soundprint feature to obtain a low-frequency energy ratio result; If it is determined according to the low-frequency energy proportion result that the low-frequency energy proportion is greater than a first preset threshold, determining that the cause of the pipeline leakage is aging of the interface rubber seal; If it is determined according to the low-frequency energy proportion result that the low-frequency energy proportion is less than or equal to the first preset threshold, determining that the cause of the pipeline leakage is micro-pores in the pipeline weld; When it is determined that the pipeline leakage type is a medium leakage type, determining the mid-frequency energy distribution in the leakage soundprint feature to obtain a mid-frequency energy distribution result; If it is determined according to the mid-frequency energy distribution result that the standard deviation of the mid-frequency energy distribution is greater than a second preset threshold, determining that the cause of the leakage is a pipeline corrosion pit; If it is determined that the pipeline leakage type is a serious leakage type, detecting whether there is a high-frequency burst peak in the leakage soundprint feature; When the high-frequency burst peak is detected in the leakage voiceprint feature and the amplitude of the high-frequency burst peak is greater than a third preset threshold, determining that the cause of the leakage is a tube rupture; Perform fault propagation prediction based on the pipeline leakage cause to obtain the leakage expansion risk level; A prioritized repair strategy set is generated according to the pipeline leakage level and the leakage extension risk level.

7. The method according to claim 1, characterized in that The method of synchronously sampling the dual-channel optical fiber vibration signal using the FPGA master clock to obtain the time-aligned pipeline vibration signal and environmental noise signal specifically includes: Control the FPGA master clock to send a Sync message to the vibration collection terminal, and record a first sending timestamp of the FPGA master clock sending the Sync message; When it is determined that the vibration collection terminal receives the Sync message, recording a first receiving timestamp of the vibration collection terminal receiving the Sync message; Controlling the vibration acquisition terminal to return a Delay_Req message to the FPGA master clock, and recording a second sending timestamp of the Delay_Req message returned by the vibration acquisition terminal; When it is determined that the FPGA master clock receives the Delay_Req message, recording a second reception timestamp of the FPGA master clock receiving the Delay_Req message; Calculating a link delay time according to the first sending timestamp, the first receiving timestamp, the second sending timestamp, and the second receiving timestamp, wherein the link delay time is used to characterize a signal transmission delay between the FPGA master clock and the vibration acquisition terminal; Time compensation processing is performed on the initial pipeline vibration signal and the initial ambient noise signal according to the link delay time, and interpolation alignment processing is performed on the compensated initial pipeline vibration signal and the initial ambient noise signal to obtain the time-aligned pipeline vibration signal and the ambient noise signal.

8. An optical fiber vibration stethoscope system, characterized in that: The fiber optic vibration stethoscope system includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code includes computer instructions, and the one or more processors call the computer instructions to enable the fiber optic vibration stethoscope system to execute the method described in any one of claims 1-7.

9. A computer-readable storage medium comprising instructions, characterized in that: When the instructions are executed on the fiber optic vibration stethoscope system, the fiber optic vibration stethoscope system is caused to execute the method according to any one of claims 1 to 7.

10. A computer program product, characterized in that When the computer program product is run on a fiber optic vibration auscultation system, the fiber optic vibration auscultation system is caused to execute the method according to any one of claims 1 to 7.

Images