Civil air defense construction concealed conduit leakage positioning method based on acoustic characteristics

Abstract

The invention provides a civil air defense engineering concealed conduit leakage positioning method based on acoustic characteristics, and belongs to the technical field of civil air defense engineering. An acoustic signal acquisition network is constructed by deploying a high-sensitivity hydrophone array, and signals are preprocessed by using a time division multiple access technology and a pulse compression technology; and carrying out time-frequency analysis by applying wavelet transform to extract acoustic features. A time reversal mirror technology is introduced to identify direct propagation and multipath reflection signals, a fluid acoustic coupling propagation equation is constructed to analyze a leakage sound source mechanism, and pipe network topological information and a sound wave speed correction function are combined to compensate a measurement error. Wherein the deep learning pipe network acoustic propagation multi-modal model is fused with a pipeline structure encoder, an acoustic feature extractor and a position prediction decoder, and high-precision leakage positioning and degree evaluation in a complex environment are realized through a three-stage pre-training strategy; the technical problem that it is difficult to accurately locate the position of a leakage point in the complex pipe network environment of civil air defense engineering is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of civil air defense engineering. Specifically, it relates to a method for locating hidden pipe leakage in civil air defense engineering based on acoustic characteristics. Background Technique

[0002] The pipeline system of civil air defense engineering is an important part of the urban underground protection project, and its airtightness and integrity are directly related to the overall protection function. Traditional leakage detection mainly relies on means such as pressure testing, tracer injection, and visual inspection, and usually can only determine the existence of pipeline leakage but it is difficult to accurately locate the leakage point. The application of these methods is limited in the special environment of civil air defense engineering and it is difficult to adapt to its characteristics of complex pipe network structure, narrow space, and coexistence of multiple media.

[0003] With the development of acoustic detection technology, the characteristics of acoustic wave propagation are used for leakage point detection, and the leakage location is inferred by analyzing the acoustic signals generated by leakage. However, the existing acoustic detection methods are insufficient in accuracy when dealing with problems such as the high-density pipe network layout, multiple acoustic wave reflections, and environmental noise interference unique to civil air defense engineering. Especially in a complex system with multi-branches and interlaced pipes of different materials, the multi-path propagation effect of acoustic waves leads to signal overlap and interference, resulting in a significant decrease in the accuracy of leakage point location.

[0004] Therefore, how to overcome the influence of acoustic wave multi-path propagation, environmental noise interference, and complex pipe network structure on acoustic signals in the special environment of civil air defense engineering and achieve high-precision leakage point location has become a technical problem to be solved urgently. The existing single signal processing method is difficult to comprehensively consider the physical characteristics of acoustic wave propagation and the characteristics of the pipe network structure and cannot meet the requirements of accurate location in complex environments. That is to say, there is a technical problem in the prior art that it is difficult to accurately locate the leakage point in the complex pipe network environment of civil air defense engineering. Summary of the Invention

[0005] In view of this, the present invention provides a method for locating hidden pipe leakage in civil air defense engineering based on acoustic characteristics, which can solve the technical problem in the prior art that it is difficult to accurately locate the leakage point in the complex pipe network environment of civil air defense engineering.

[0006] The present invention is implemented as follows: The present invention provides a method for locating hidden pipe leakage in civil air defense projects based on acoustic features, including: deploying a hydrophone array at key nodes of the pipeline system to construct an acoustic signal acquisition network; exciting internal acoustic waves in the pipeline and preprocessing the acquired signals using time division multiple access technology; applying pulse compression technology to improve the time-domain resolution of the signals; performing time-frequency analysis on the acoustic signals based on wavelet transform to extract acoustic features; introducing the time reversal mirror technology to reconstruct the acoustic wave propagation path; constructing a fluid-acoustic coupling propagation equation to analyze the generation mechanism of the leakage sound source; combining the pipeline network topology structure information and using an acoustic wave velocity correction function to compensate for the influence of the medium; and using a pre-trained multi-modal model for acoustic propagation in the pipeline network to locate the sound source.

[0007] Among them, the hydrophone array is a high-sensitivity hydrophone array deployed at key nodes of the pipeline system in civil air defense projects.

[0008] Among them, the time division multiple access technology refers to a multiple access technology that enables multiple signals to share the same frequency band without interfering with each other by allocating channels at different times, and is applicable to the situation where multiple sensors in a complex pipeline network collect signals simultaneously.

[0009] Among them, the pulse compression technology refers to a method of converting a broadband signal into a narrow pulse signal through processing by a matched filter, which can improve the time-domain resolution while maintaining the signal energy and enhance the distinguishability between weak leakage sound signals and background noise.

[0010] Among them, the time reversal mirror technology refers to a method of reversing the received acoustic signal in the time domain and retransmitting it, and using the reversibility of acoustic wave propagation to refocus the acoustic energy on the original sound source position, effectively solving the multi-path propagation problem in a complex pipeline network.

[0011] Among them, the inputs of the fluid-acoustic coupling propagation equation include: the fluid density measured by the hydrophone array, the acoustic wave propagation distance identified by the time reversal mirror technology, the sound source flow function extracted by wavelet transform, the acoustic wave propagation speed corrected by the acoustic wave velocity correction function, and the attenuation coefficient measured by the pulse compression technology, and the output is the sound pressure level at the distance and time point.

[0012] Among them, the acoustic wave velocity correction function refers to a mathematical expression that dynamically adjusts the acoustic wave propagation speed calculation model according to parameters such as the temperature, pressure, density, and flow velocity of the medium in the pipeline, improving the positioning accuracy.

[0013] Among them, the specific structure of the multi-modal model for acoustic propagation in the pipeline network is an acoustic wave propagation feature extraction network based on the fusion of a deep convolutional neural network and a self-attention mechanism, including three main modules: a pipeline structure encoder, an acoustic feature extractor, and a position prediction decoder.

[0014] Among them, the number of self-attention heads of the self-attention mechanism is determined according to the number of main branches of the pipeline system, and the number of main branches of the pipeline system is derived from the constructed acoustic signal acquisition network.

[0015] Among them, the pipeline structure encoder uses a graph convolutional network structure to convert the pipe network topology information into a high-dimensional feature vector; the acoustic feature extractor uses a multi-layer one-dimensional convolutional network to extract time-domain and frequency-domain features; the position prediction decoder fuses the pipeline structure and acoustic feature information through a multi-head cross-attention mechanism to output the estimated result of the leakage point position.

[0016] The present invention constructs a complete technical route from signal acquisition, preprocessing to feature extraction and position calculation. By deploying a high-sensitivity hydrophone array and introducing a time division multiple access acquisition strategy, the system can obtain high-quality leakage acoustic signals in a complex noise environment; combining pulse compression and wavelet transform technologies can improve the time-frequency resolution and enhance the recognition ability of weak leakage signals.

[0017] The present invention uniquely introduces the time reversal mirror technology and combines it with the fluid acoustic coupling propagation equation to effectively solve the problem of multi-path propagation of sound waves in the complex pipe network of civil air defense projects, and accurately separates the direct propagation signal and the reflected signal. By constructing a sound wave velocity correction function to compensate for the influence of medium parameter changes on the measurement accuracy, the defect of insufficient accuracy of traditional acoustic positioning methods in complex pipe network structures is overcome.

[0018] In particular, the multi-modal model of pipe network acoustic propagation developed by the present invention integrates the pipe network topology structure information and acoustic features, can achieve high-precision leakage positioning in a complex civil air defense project environment, improves the three-dimensional coordinate positioning accuracy of the leakage point to the centimeter level, and at the same time provides a leakage degree grading assessment, solving the technical problem of accurate positioning of leakage points in the complex pipe network environment of civil air defense projects. Brief Description of the Drawings

[0019] Figure 1 It is a flowchart of the method of the present invention. Detailed Embodiments

[0020] To make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0021] As Figure 1 shown, it is a flowchart of a method for locating hidden pipe leakage in a civil air defense project based on acoustic features provided by the present invention. The method includes the following steps:

[0022] S01. Deploy a high-sensitivity hydrophone array at key nodes of the pipeline system in the civil air defense project to construct an acoustic signal acquisition network;

[0023] S02. Excite the internal acoustic waves of the pipeline, and use time-division multiple access technology to preprocess the collected acoustic signals to eliminate environmental noise interference;

[0024] S03. Apply pulse compression technology to improve the time-domain resolution of the signal and enhance the recognition of weak acoustic signals generated by leaks;

[0025] S04. Perform time-frequency analysis on the acoustic signals based on wavelet transform, and extract acoustic features including frequency center, energy distribution, and propagation characteristics;

[0026] S05. Introduce time reversal mirror technology to reconstruct the acoustic wave propagation path and identify direct propagation signals and multipath reflection signals;

[0027] S06. Construct a fluid-acoustic coupling propagation equation to analyze the generation mechanism of leak sound sources, and determine the relationship between leak fluid flow state parameters and acoustic features by solving the fluid-acoustic coupling propagation equation;

[0028] S07. Combine the pipeline network topology structure information, and use the acoustic wave velocity correction function to compensate for the influence of medium density and flow velocity on the measurement accuracy;

[0029] S08. Use a pre-trained multi-modal model of pipeline network acoustic propagation to optimize the sound source localization process, and output the three-dimensional coordinates and confidence level of the leak point;

[0030] S09. Optionally, it further includes establishing a leak acoustic feature database, and grading and evaluating the leak degree through machine learning algorithms to form maintenance decision-making suggestions.

[0031] Among them, time-division multiple access technology refers to a multiple access technology that enables multiple signals to share the same frequency band without mutual interference by allocating channels at different times, and is applicable to the situation where multiple sensors simultaneously collect signals in a complex pipeline network.

[0032] Among them, pulse compression technology refers to a method of converting a broadband signal into a narrow pulse signal through matched filter processing, which can improve the time-domain resolution while maintaining the signal energy, and enhance the distinguishability between weak leak sound signals and background noise.

[0033] Among them, time reversal mirror technology refers to a method of reversing the received acoustic signal in the time domain and retransmitting it, and using the reversibility of acoustic wave propagation to refocus the acoustic energy on the original sound source position, effectively solving the multipath propagation problem in a complex pipeline network.

[0034] Among them, the fluid-acoustic coupling propagation equation refers to a set of equations that describe the physical mechanism of the generation and propagation of sound sources at leakage points in a pipeline system. The fluid-acoustic coupling propagation equation is used to calculate the variation law of the sound pressure level at the leakage point with distance and time. The inputs include the fluid density measured by the hydrophone array in step S01, the sound wave propagation distance identified in step S05, the sound source flow function extracted in step S04, the corrected sound wave propagation speed in step S07, and the attenuation coefficient measured in step S03. The output is the sound pressure level at the distance and time points for source localization in step S08.

[0035] Among them, the sound wave velocity correction function refers to a mathematical expression that dynamically adjusts the sound wave propagation speed calculation model according to parameters such as the medium temperature, pressure, density, and flow velocity in the pipeline, so as to improve the positioning accuracy.

[0036] The specific structure of the multi-modal model for acoustic propagation in the pipe network is a sound wave propagation feature extraction network based on the fusion of a deep convolutional neural network and a self-attention mechanism, which includes three main modules: a pipeline structure encoder, an acoustic feature extractor, and a position prediction decoder. Among them, the number of self-attention heads is determined according to the number of main branches of the pipeline system, and the number of main branches of the pipeline system comes from the acoustic signal acquisition network constructed in step S01; the pipeline structure encoder uses a graph convolutional network structure to convert the pipe network topology information into a high-dimensional feature vector; the acoustic feature extractor uses a multi-layer one-dimensional convolutional network to extract time-domain and frequency-domain features; the position prediction decoder fuses the pipeline structure and acoustic feature information through a multi-head cross-attention mechanism and outputs the estimated result of the leakage point position.

[0037] The steps for establishing the training data set of the multi-modal model for acoustic propagation in the pipe network specifically include first constructing various pipeline structure models in a laboratory environment, simulating different types of leakage situations and recording complete acoustic signal data; then using computational fluid dynamics software to simulate the fluid-acoustic coupling characteristics under various leakage scenarios, generating virtual acoustic signal data, and annotating accurate leakage source position information; finally, mixing the laboratory data and the simulation data to construct a pipe network leakage acoustic feature data set under various working conditions, ensuring that the multi-modal model for acoustic propagation in the pipe network has the generalization ability to cope with complex situations in the actual engineering environment.

[0038] The steps for pre-training the multi-modal model for acoustic propagation in the pipeline network to be monitored specifically include adopting a three-stage training strategy. In the first stage, unsupervised training is carried out using a large amount of pipeline network topology data, enabling the multi-modal model for acoustic propagation in the pipeline network to master the pipeline network structure encoding ability. In the second stage, supervised fine-tuning is performed using acoustic feature data to optimize the parameters of the multi-modal model for acoustic propagation in the pipeline network for the leakage point feature recognition task. In the third stage, the leakage detection and localization performance are simultaneously optimized through multi-task learning, and a weighted loss function is used to balance the gradient contributions between different tasks, where the weight coefficient is dynamically adjusted according to the importance of leakage detection and precise localization in actual engineering applications. After training is completed, the multi-modal model for acoustic propagation in the pipeline network adapts to pipeline network systems of different scales and complexities, realizing the function of high-precision leakage localization.

[0039] The following describes in detail the specific implementation manners of the above steps.

[0040] The specific implementation manner of step S01 is to deploy a high-sensitivity hydrophone array inside the pipeline system of the civil air defense project. First, determine the positions of key nodes according to the pipeline network topology, including pipeline intersections, turning points, and pipe diameter change points. Then, install hydrophones with a sensitivity of not less than -180 dB (referenced to 1 V / μPa) at each key node, ensuring that its frequency response range covers 20 Hz to 20 kHz. Fix the hydrophones on the inner side of the pipe wall to avoid direct contact with the fluid and reduce flow noise interference. Adopt synchronous sampling technology to ensure that the time synchronization error of all hydrophones is controlled within 10 μs. Finally, transmit the signals of each acquisition point to the central processing system through a high-speed data bus to form an acoustic signal acquisition network covering the entire pipeline network. This step aims to establish a complete acoustic data acquisition infrastructure to provide raw data support for subsequent leakage analysis.

[0041] The specific implementation manner of step S02 is to preprocess the acquired acoustic signals using time division multiple access technology. First, install an acoustic wave excitation device at a suitable position in the pipeline system to generate a broadband acoustic pulse signal with a bandwidth of 1 kHz to 5 kHz. Then, assign a unique time code to the signals acquired by each hydrophone node, and use an orthogonal time code sequence to ensure that the cross-correlation between each signal channel is less than 0.2. Next, limit the signals within the effective frequency band through a digital band-pass filter. The lower limit frequency of the filter is set to 500 Hz to remove low-frequency ambient noise, and the upper limit frequency is set to 10 kHz to retain the acoustic features of leakage. Finally, apply an adaptive noise cancellation algorithm to filter out the background noise, and the noise cancellation threshold is set to a signal-to-noise ratio of 5 dB. The purpose of this step is to separate the effective acoustic signals from the ambient noise and improve the signal quality for subsequent analysis.

[0042] The specific implementation of step S03 is to apply pulse compression technology to improve the time-domain resolution of the signal. First, a linear frequency modulation signal is designed as the reference signal, with a frequency range of 1 kHz to 8 kHz and a duration of 50 ms. Then, the collected acoustic signal is subjected to matched filtering with the reference signal, and the matched filtering process is implemented using the fast convolution algorithm. Next, the signal pulse compression ratio is calculated to ensure that the main lobe width after compression is less than 1 ms and the sidelobe suppression ratio is greater than 20 dB. Finally, the compressed signal is smoothed using the Hanning window function to reduce the influence of Gibbs phenomenon. This step aims to compress the broadband signal into a narrow pulse signal, enhance the distinguishability between weak leakage sound signals and background noise, and improve the sensitivity of subsequent leakage detection.

[0043] The specific implementation of step S04 is to perform time-frequency analysis on the acoustic signal based on wavelet transform. First, a wavelet basis function with good time-frequency localization performance, such as the Debyeche 5th-order wavelet, is selected to perform continuous wavelet transform on the preprocessed acoustic signal. Then, the energy features in the wavelet coefficient matrix are extracted, and the wavelet energy spectrum at different scales is calculated. Next, the energy aggregation interval is identified to determine the frequency center of the leakage signal, and the typical frequency center range of the leakage signal is between 2 kHz and 6 kHz. Subsequently, the time-varying characteristics of the energy are analyzed, and the change rate of the wavelet energy with time is calculated. The leakage signal usually exhibits a continuous and stable energy output, and the change rate is lower than that of the normal flow signal. Finally, the acoustic propagation characteristic parameters, including group velocity and phase velocity, are extracted for subsequent propagation path analysis. The purpose of this step is to obtain the complete time-frequency characteristics of the acoustic signal and provide multi-dimensional feature support for leakage sound source identification.

[0044] The specific implementation of step S05 is to introduce the time reversal mirror technology to reconstruct the acoustic wave propagation path. First, the collected acoustic signal is reversed in the time domain to generate a time reversal sequence. Then, a virtual sound source array model is established, and multiple possible sound source position points are set in the pipe network system. Next, the time reversal signal is propagated in the virtual sound source array, and the energy focusing degree is calculated. The energy focusing degree is defined as the ratio of the peak value to the average value of the reversed signal. Generally, the energy focusing degree at the leakage point position is higher than 10 dB. Subsequently, the potential leakage point position is identified according to the maximum energy focusing criterion. Finally, the direct propagation signal and the multipath reflection signal are separated. The direct propagation signal usually appears as the first high-energy signal to arrive, and the subsequent lower-energy signal is the multipath reflection signal. This step aims to utilize the reversibility of acoustic wave propagation to solve the multipath propagation problem in complex pipe networks and improve the positioning accuracy.

[0045] The specific implementation of step S06 is to construct a fluid-acoustic coupling propagation equation to analyze the generation mechanism of the leakage sound source. First, establish a set of equations that describe the physical mechanism of the generation and propagation of the sound source at the leakage point in the pipeline system, including the continuity equation, momentum equation, and energy equation. Then, input the fluid density measured by the hydrophone array, the identified acoustic wave propagation distance, the extracted sound source flow function, the corrected acoustic wave propagation speed, and the measured attenuation coefficient into the set of equations. Next, use the finite difference method to numerically solve the fluid-acoustic coupling propagation equation, with the time step set to 10 μs and the spatial step set to 5 mm. Finally, calculate the sound pressure level at different distances and time points, and establish a quantitative relationship between the leakage fluid flow state parameters and the acoustic characteristics. The purpose of this step is to understand the characteristics of the leakage sound source from the physical mechanism level and provide a theoretical basis for sound source localization.

[0046] The specific implementation of step S07 is to combine the pipeline network topology structure information and use the acoustic wave velocity correction function to compensate for the influence of medium density and flow velocity on the measurement accuracy. First, establish the acoustic wave velocity correction function, considering parameters such as the temperature, pressure, density, and flow velocity of the medium in the pipeline, with the temperature measurement accuracy of ±0.5 °C and the pressure measurement accuracy of ±0.5 kPa. Then, use the parameters collected by the hydrophone array to update the acoustic wave velocity calculation model in real time. The form of the acoustic wave velocity correction function is a non-linear mathematical expression considering the physical parameters of the medium and the flow velocity. Next, calculate the acoustic wave propagation speed in different pipe segments. The sound speed in typical water media is about 1450 - 1550 m / s. Finally, apply the corrected acoustic wave velocity to the leakage localization calculation to reduce the measurement error. The purpose of this step is to improve the calculation accuracy of the acoustic wave propagation speed and eliminate the influence of medium characteristic changes on the localization accuracy.

[0047] The specific implementation of step S08 is to use a pre-trained multi-modal model for pipeline network acoustic propagation to optimize the sound source localization process. First, input the acoustic characteristics and pipeline network topology information extracted in the previous steps into the multi-modal model for pipeline network acoustic propagation, which is designed based on a deep convolutional neural network and a self-attention mechanism. Then, convert the pipeline network topology information into a high-dimensional feature vector through the pipeline structure encoder of the model. Next, use the acoustic feature extractor to process the time-frequency features. Subsequently, the position prediction decoder fuses the pipeline structure and acoustic feature information through a multi-head cross-attention mechanism. The number of attention heads is set according to the number of main branches in the pipeline system, usually 4 - 8. Finally, the model outputs the three-dimensional coordinates of the leakage point and the corresponding confidence level, with the confidence level threshold set to 0.85. Prediction results below this threshold need to be further verified. The purpose of this step is to use deep learning technology to comprehensively analyze various acoustic characteristics and pipeline network information to achieve high-precision leakage localization.

[0048] The specific implementation of step S09 is to establish a leakage acoustic feature database and grade and evaluate the leakage degree through machine learning algorithms. First, collect acoustic feature data of different types and degrees of leakage, including spectral features, energy features, and time-domain features. Then, use the hierarchical clustering algorithm to group the leakage features according to similarity to form a leakage feature spectrum. Next, train a support vector machine classifier to divide the leakage degree into five levels, from slight leakage to severe leakage, with a classification accuracy of not less than 90%. Subsequently, generate maintenance priority suggestions based on the leakage degree and location. The maintenance priority considers the severity of leakage, the importance of location, and the potential impact range. Finally, feed the newly detected leakage data back to the database to continuously expand and optimize the leakage acoustic feature library. This step aims to evaluate and classify the detected leakage and provide a scientific basis for maintenance decisions.

[0049] The detailed structure of the multi-modal model for acoustic propagation in pipe networks is a feature extraction network for acoustic wave propagation based on the fusion of a deep convolutional neural network and a self-attention mechanism, including three main modules: The pipeline structure encoder adopts a graph convolutional network structure, including 3 layers of graph convolutional layers, with 64 feature channels in each layer. The activation function is LeakyReLU, and the pooling adopts the edge aggregation method to convert the pipe network topology information into a 512-dimensional feature vector; The acoustic feature extractor uses a 5-layer one-dimensional convolutional network, with convolutional kernel sizes of 3, 5, 7, 9, and 11, and the number of channels are 32, 64, 128, 256, and 512 respectively. After each layer, batch normalization and max pooling are connected to extract time-domain and frequency-domain features; The location prediction decoder uses a multi-head cross-attention mechanism, and the number of attention heads is equal to the number of main branches of the pipeline system. After the attention layer, 3 layers of fully connected layers are connected, with the number of neurons being 512, 256, and 128. Finally, the output layer has 3 neurons representing the three-dimensional coordinates of the leakage point and 1 neuron representing the confidence level. The model input includes the information of the pipe network structure diagram and the acoustic feature vector, and the output is the estimated result of the leakage point location. The total number of model parameters is about 2.5×10 6 pieces, and the inference time is controlled within 100 ms to meet the real-time positioning requirements.

[0050] The steps for establishing the training dataset of the multi-modal model for acoustic propagation in pipe networks are detailed as follows: First, construct various pipeline structure models in the laboratory environment, including straight pipe sections, elbows, T-joints, and cross pipe sections, with a pipe diameter range of 50 - 300 mm, and the pipe materials cover steel pipes, cast iron pipes, and concrete pipes. Simulate different types of leakage situations, including crack-type, hole-type, and interface loosening-type leakage, with a leakage diameter of 0.5 - 5 mm and a pressure gradient of 10 - 100 kPa. Use a 24-bit high-precision data acquisition device to record complete acoustic signal data, and set the sampling rate to 96 kHz; Then, use computational fluid dynamics software to simulate various leakage scenarios, establish a simulation environment including pipe network geometric models, fluid parameter settings, boundary condition definitions, and mesh divisions, and the number of mesh cells is not less than 106 pieces, with a simulation time step of 10 -5 seconds, generate virtual acoustic signal data, and mark the accurate leakage source location information with a location marking accuracy of ±1 cm; finally, mix the laboratory data and the simulation data in a ratio of 7:3 to construct a dataset containing various working conditions, and the total sample size is not less than 10 4 pieces, and each sample contains acoustic signal time series data, spectrum data, pipe network topology information, and leakage location labels. The 5-fold cross-validation method is used to ensure the quality of the dataset and ensure that the multi-modal model of acoustic propagation in the pipe network has the generalization ability to cope with complex situations in the actual engineering environment.

[0051] The following is a detailed description of the mathematical models or calculation processes involved in the present invention.

[0052] In step S02, the time division multiple access technology preprocesses the collected acoustic signals, and the specific expression of the adaptive noise cancellation algorithm is as follows:

[0053] y(n) = s(n) + v(n);

[0054]

[0055]

[0056] w i (n + 1) = w i (n) + μ·e(n)·r(n - i);

[0057] In the formula, y(n) is the original signal collected by the hydrophone; s(n) is the effective acoustic signal; v(n) is the noise signal; e(n) is the estimated effective signal after noise cancellation; is the estimated noise signal; r(n) is the reference noise signal; w i (n) is the value of the i-th coefficient of the adaptive filter at time n; L is the filter order, usually taking values from 32 to 128; μ is the learning rate, and the value range is 0.01 to 0.1, which is adjusted according to the convergence speed and stability requirements.

[0058] The parameter acquisition method is: y(n) is directly collected by the hydrophone; r(n) is collected by an additional reference hydrophone set in the non-leakage area of the pipeline system; w i (0) The initial values are all set to 0 and the optimal values are obtained through iterative updates. This equation is based on the principle of minimum mean square error. By continuously adjusting the filter coefficients, the estimated noise signal is made similar to the actual noise, so as to suppress the noise while retaining the effective signal and improve the signal quality of subsequent analysis.

[0059] In step S03, the specific expression of the pulse compression technology to improve the signal time-domain resolution is as follows:

[0060]

[0061] where x r (t) is the reference signal, i.e., the chirp signal; A is the signal amplitude, usually normalized to 1; f0 is the starting frequency, with a value of 1 kHz; k is the chirp rate, B is the bandwidth, with a value of 7 kHz; t is the signal duration, with a value of 50 ms; x(t) is the received acoustic signal; is the complex conjugate of the reference signal; y(t) is the output signal after matched filtering; R(τ) is the autocorrelation function; τ0 is the signal delay; B is the signal bandwidth; j is the imaginary unit.

[0062] The method for obtaining the parameters is as follows: x(t) is obtained by collecting through a hydrophone and preprocessing through step S02; x r (t) generates a chirp signal through a signal generator; τ0 is estimated by the propagation time of sound waves in the pipeline, and the calculation formula is where d is the propagation distance and c is the sound wave velocity. This equation is based on the matched filtering theory. By calculating the correlation between the received signal and the reference signal, the broadband signal is compressed into a narrow pulse signal, enhancing the distinguishability between the weak leakage sound signal and the background noise and improving the leakage detection sensitivity.

[0063] In step S04, the specific representation of the time-frequency analysis of the acoustic signal by wavelet transform is as follows:

[0064]

[0065] where W f (a, b) is the continuous wavelet transform coefficient; f(t) is the preprocessed acoustic signal; ψ(t) is the wavelet basis function, the Debyeche 5th-order wavelet; a is the scale parameter, which determines the frequency resolution; b is the translation parameter, which determines the time position; E(a, b) is the time-frequency energy distribution; E s (a) is the energy spectrum at scale a; b1 and b2 are the time window boundaries; is the energy change rate; Δt is the time step, with a value of 10 ms.

[0066] The method for obtaining the parameters is as follows: f(t) is obtained by processing through step S03; the value range of a is determined according to the frequency range of interest, usually 1 - 64; the value range of b covers the entire signal duration; ψ(t) selects the Debyeche 5th-order wavelet, and its mathematical expression is obtained by solving the scale equation. This equation is based on the wavelet analysis theory. By analyzing the time-frequency characteristics of the signal at multiple scales, acoustic features including the frequency center, energy distribution, and propagation characteristics are extracted, providing multi-dimensional feature support for leakage sound source identification.

[0067] In step S05, the specific representation of the time-reversal mirror technology for reconstructing the acoustic wave propagation path is as follows:

[0068] f TR (t) = f(T - t);

[0069]

[0070] In the formula, f TR (t) is the time-reversal signal; f(t) is the original received signal; T is the total signal duration; p(r, t) is the acoustic pressure at spatial position r at time t; G(r, r i , t) is the acoustic Green's function from position r i to position r; N is the number of receivers; FC(r) is the energy focusing degree; t max is the time corresponding to the maximum acoustic pressure.

[0071] The method for obtaining the parameters is as follows: f(t) is obtained after being processed by step S04; G(r, r i , t) is calculated through the pipeline acoustic wave propagation model, considering the pipeline geometric structure, wall reflection, and medium characteristics; r i is the hydrophone position coordinate; r is the possible leakage point position coordinate in the pipe network. This equation is based on the time-reversal acoustics theory, utilizes the reversibility of acoustic wave propagation, and identifies the sound source position by calculating the energy focusing degree, effectively solving the multipath propagation problem in complex pipe networks and improving the positioning accuracy.

[0072] In step S06, the specific representation of the fluid-acoustic coupling propagation equation is as follows:

[0073]

[0074] In the formula, ρ is the fluid density; is the fluid velocity vector; p is the fluid pressure; μ is the fluid dynamic viscosity; c is the acoustic wave propagation velocity; p' is the acoustic pressure perturbation; ρ0 is the fluid static density; Q is the volume flow function representing the leakage source strength; e is the distance from the leakage point; is the gradient operator; is the divergence operator; is the Laplace operator.

[0075] The method for obtaining the parameters is as follows: ρ is measured by the density sensor attached to the hydrophone array; is measured by the flowmeter inside the pipeline; p is measured by the pressure sensor; μ is obtained by looking up the table according to the medium temperature; c is calculated through the acoustic wave velocity correction function in step S07; Q is estimated by analyzing the fluid outflow characteristics at the leakage opening, and its expression is where Cd is the flow coefficient, with a value range of 0.6 to 0.9, A is the leakage orifice area, and Δp is the pressure difference inside and outside the leakage location. This set of equations is based on the basic principles of fluid mechanics and acoustics, describes the physical mechanism of the generation and propagation of the sound source at the leakage point in the pipeline system, and obtains the sound pressure level at the distance and time point through solution, providing a theoretical basis for sound source localization.

[0076] In step S07, the specific representation of the acoustic wave velocity correction function is as follows:

[0077] c eff = c0·[1 + α T (T - T0) + α p (p - p0) + α ρ (ρ - ρ0)]·(1 ± M·cosθ);

[0078] In the formula, c eff is the corrected effective acoustic wave velocity; c0 is the acoustic wave velocity under reference conditions, with a value of 1500 m / s; T is the current temperature; T0 is the reference temperature, with a value of 20 °C; p is the current pressure; p0 is the reference pressure, with a value of 101.3 kPa; ρ is the current density; ρ0 is the reference density, with a value of 1000 kg / m 3 ; α T is the temperature correction coefficient, with a value of 2.4×10 -3 / °C; α p is the pressure correction coefficient, with a value of 1.8×10 -5 / kPa; α ρ is the density correction coefficient, with a value of -3.0×10 -4 / (kg / m 3 ); M is the Mach number, θ is the angle between the acoustic wave propagation direction and the fluid flow direction.

[0079] The parameter acquisition method is as follows: T is measured by a temperature sensor with an accuracy of ±0.5 °C; p is measured by a pressure sensor with an accuracy of ±0.5 kPa; ρ is measured by a densitometer; is measured by a flow meter; θ is calculated from the pipeline geometric structure and the acoustic wave propagation path. This equation takes into account the effects of temperature, pressure, density, and flow velocity on the acoustic wave propagation speed, improves the positioning accuracy by correcting the acoustic wave velocity, and eliminates the influence of the change in medium characteristics on the measurement results. The linear correction of the temperature, pressure, and density terms is based on the approximate linear relationship of the influence of these parameters within the engineering application range; the Mach number correction of the flow velocity term takes into account the Doppler effect of the acoustic wave in the flowing medium, and the cos term reflects the relationship between the acoustic wave propagation direction and the flow direction.

[0080] In step S08, the confidence calculation formula used by the pipeline network acoustic propagation multimodal model is specifically expressed as follows:

[0081]

[0082] In the formula, C is the confidence of the positioning result; S is the similarity score output by the model; S th is the similarity threshold, with a value of 0.75; β is the sensitivity coefficient, with a value of 10.

[0083] The parameter acquisition method is as follows: S is calculated by the pipeline network acoustic propagation multimodal model, representing the similarity between the predicted location and the actual leakage location; S th and β are optimized through the model verification stage. This equation uses the Sigmoid function to convert the similarity score into a confidence value between 0 and 1. When the similarity score is much higher than the threshold, the confidence is close to 1; when it is much lower than the threshold, the confidence is close to 0, and there is a smooth transition near the threshold, facilitating the setting of the confidence threshold of 0.85 to screen the positioning results.

[0084] In step S09, the decision function of the support vector machine classifier used for leakage degree classification and evaluation is specifically expressed as follows:

[0085]

[0086] K(x i , x) = exp(-γ||x i - x|| 2 );

[0087]

[0088] In the formula, f(x) is the support vector machine decision function; x is the input acoustic feature vector; x i is the support vector; y i is the class label corresponding to the support vector; α i is the Lagrange multiplier; b is the bias term; K(x i , x) is the kernel function, using the radial basis function; γ is the kernel function parameter, determining the influence radius; L is the comprehensive leakage level; P j is the probability of the j-th level of leakage; w j is the weight coefficient, reflecting the importance of different levels, generally increasing with the level. w1 to w5 take values of 0.1, 0.2, 0.3, 0.5, and 0.9 respectively.

[0089] The parameter acquisition method is as follows: x is composed of the acoustic features extracted in the previous steps, including the center frequency, energy distribution, duration, etc.; x i , α iα and b are obtained through the support vector machine training process, using the leakage acoustic feature database established in step S09 as the training set; γ is optimized through cross-validation, and the typical value is 0.1 - 1.0. This equation is based on the support vector machine theory. By calculating the similarity of the kernel function between the input features and the support vectors, the classification of the leakage degree is realized. The radial basis function is adopted as the kernel function considering the non-linear characteristics of the class boundaries in the acoustic feature space. The comprehensive level calculation formula takes into account the probability distributions of multiple levels, improving the robustness of the evaluation.

[0090] Optionally, the graph convolution calculation formula of the pipeline structure encoder is specifically expressed as follows:

[0091]

[0092] In the formula, H (l) is the node feature matrix of the l-th layer; A is the adjacency matrix of the pipe network topology; D is the degree matrix, and the diagonal element D ii is equal to the degree of node i; W (l) is the trainable weight matrix of the l-th layer; σ is the activation function, and LeakyReLU is adopted.

[0093] The parameter acquisition method is: A is constructed through the pipe network topology information, representing the connection relationship between pipeline nodes; H (0) is the input node feature, including information such as node type, pipe diameter, and material; W (l) is obtained through the model training process. This equation is based on the graph convolution network theory. By aggregating and transmitting information on the pipe network topology, the pipe network topology information is converted into a high-dimensional feature vector, effectively capturing the structural characteristics of the pipeline system.

[0094] Optionally, the one-dimensional convolution calculation formula of the acoustic feature extractor is specifically expressed as follows:

[0095] F (l+1) = Pool(BN(σ(Conv1D(F (l) , K (l) )))));

[0096]

[0097] In the formula, F (l) is the acoustic feature map of the l-th layer; K (l) is the convolution kernel parameter of the l-th layer; Conv1D is the one-dimensional convolution operation; BN is the batch normalization operation; Pool is the max pooling operation; σ is the activation function; k is the convolution kernel size, which is 3, 5, 7, 9, 11 respectively for each layer.

[0098] The parameter acquisition method is: F (0)is the input acoustic signal and the time-frequency features obtained by wavelet transform in step S04; K (l) It is obtained by optimizing through the model training process. This equation is based on the convolutional neural network theory, and extracts the time-domain and frequency-domain features of the acoustic signal through multi-layer one-dimensional convolution. Convolution kernels of different sizes can capture acoustic feature patterns of different scales.

[0099] Optionally, the calculation formula of the multi-head cross-attention mechanism of the position prediction decoder is specifically expressed as follows:

[0100]

[0101] MultiHead(Q, K, V) = Concat(head1, head2,..., head h )W O ;

[0102] head i = Attention(QW i Q , KW i K , VW i V );

[0103] Z = MultiHead(H structure , F acoustic , F acoustic );

[0104] In the formula, Q, K, and V are the query, key, and value matrices respectively; d k is the dimension of the key vector; W i Q , W i K , W i V are the linear mapping matrices of the i-th attention head; W O is the output mapping matrix; h is the number of attention heads, which is determined according to the number of main branches of the pipeline system; H structure is the feature output by the pipeline structure encoder; F acoustic is the feature output by the acoustic feature extractor; Z is the fused feature output by the multi-head attention.

[0105] The parameter acquisition method is: H structure is calculated through the pipeline structure encoder; F acoustic is calculated through the acoustic feature extractor; W i Q , W i K , W i V and WO It is obtained through the optimization of the model training process. This equation is based on the multi-head attention mechanism in Transformer. By calculating the correlation between the pipeline structure features and the acoustic features, it effectively fuses the information of two different modalities and improves the leakage location accuracy.

[0106] Specifically, the principle of the present invention is as follows: The core principle of the present invention lies in constructing a multi-dimensional fusion analysis framework for acoustic signals and pipeline networks, and solving the leakage location problem in complex environments by comprehensively utilizing acoustic propagation characteristics, fluid mechanics principles, and deep learning techniques. First, based on the original data obtained by the acoustic signal acquisition network, the time division multiple access technology is applied to achieve interference-free acquisition of multi-sensor signals, avoiding mutual interference between signals in the same frequency band and improving the signal quality; while the pulse compression technology converts the broadband signal into a narrow pulse form, improving the time domain resolution while maintaining the signal energy, enabling weak leakage signals to be effectively separated from background noise.

[0107] The time reversal mirror technology is an important innovation of the present invention. It utilizes the time reversibility of acoustic wave propagation, reverses the received acoustic signal in the time domain and re-transmits it, so that the acoustic energy refocuses at the original sound source position. This technology, combined with the fluid-acoustic coupling propagation equation, can accurately describe the physical mechanism of the generation and propagation of the sound source at the leakage point, and solves the positioning error problem caused by multi-path propagation of acoustic waves in complex pipe networks. The fluid-acoustic coupling propagation equation combines the hydrodynamic characteristics and the acoustic propagation characteristics, and provides a theoretical basis for sound source positioning by calculating the sound pressure level distribution under different conditions.

[0108] Another key innovation of the present invention is the multi-modal model of pipeline acoustic propagation. This model integrates three core modules: a pipeline structure encoder, an acoustic feature extractor, and a position prediction decoder, and processes heterogeneous information through a deep convolutional neural network and a self-attention mechanism. The model adopts a three-stage training strategy to gradually master the pipeline structure encoding ability, leakage feature recognition ability, and precise positioning ability. The acoustic wave velocity correction function dynamically adjusts the acoustic wave propagation velocity calculation model according to the medium parameters in the pipeline, further improving the positioning accuracy. This method of multi-technology fusion can effectively overcome various interference factors in complex pipeline network environments and achieve high-precision leakage location.

[0109] A specific embodiment 1 of the present invention is provided below. The specific implementation of each step in this embodiment 1 is described in detail as follows.

[0110] In this embodiment, the specific implementation of step S01 is the same as that described above, and will not be elaborated here.

[0111] In this embodiment, the specific implementation of step S02 is to preprocess the collected acoustic signals using time-division multiple access technology. First, a suitable position is selected in the pipeline system to install an acoustic wave excitation device to generate a broadband acoustic pulse signal with a bandwidth of 1 kHz to 5 kHz. Then, a unique time code is assigned to the signals collected by each hydrophone node, and an orthogonal time code sequence is used to ensure that the cross-correlation between signal channels is less than 0.2. Next, the signals are restricted within the effective frequency band through a digital band-pass filter. The lower limit frequency of the filter is set to 500 Hz to remove low-frequency environmental noise, and the upper limit frequency is set to 10 kHz to retain the acoustic characteristics of leakage. Finally, an adaptive noise cancellation algorithm is applied to filter out background noise. The algorithm is based on the principle of minimum mean square error and recursively adjusts the filter coefficients to make the estimated noise similar to the actual noise, thereby improving the signal quality. The specific implementation is as follows:

[0112] y(n) = s(n) + v(n);

[0113]

[0114] w i (n + 1) = w i (n) + μ·e(n)·r(n - i);

[0115] In the formula, y(n) is the original signal collected by the hydrophone; s(n) is the effective acoustic signal; v(n) is the noise signal; e(n) is the estimated effective signal after noise cancellation; is the estimated noise signal; r(n) is the reference noise signal, which is collected by an additional reference hydrophone in the non-leakage area; w i (n) is the value of the i-th coefficient of the adaptive filter at time n; L is the filter order, usually taking values from 32 to 128; μ is the learning rate, with a value range of 0.01 to 0.1, adjusted according to the convergence speed and stability requirements; the noise cancellation threshold is set to a signal-to-noise ratio of 5 dB. The purpose of this step is to separate the effective acoustic signal from the environmental noise and improve the signal quality for subsequent analysis.

[0116] The specific implementation of step S03 is to apply pulse compression technology to improve the time-domain resolution of the signal. First, a chirp signal is designed as the reference signal, with a frequency range of 1 kHz to 8 kHz and a duration of 50 ms. The signal expression is:

[0117]

[0118] In the formula, A is the signal amplitude, usually normalized to 1; f0 is the starting frequency, taking a value of 1 kHz; k is the frequency modulation rate, B is the bandwidth, taking a value of 7 kHz; T is the signal duration, taking a value of 50 ms. Then, the collected acoustic signal is subjected to matched filtering with the reference signal. The expression of the matched filtering output is:

[0119]

[0120] Wherein, x(t) is the received acoustic signal; is the complex conjugate of the reference signal; y(t) is the output signal after matched filtering. The matched filtering process is implemented by a fast convolution algorithm, and then the signal pulse compression ratio is calculated to ensure that the main lobe width after compression is less than 1 ms and the sidelobe suppression ratio is greater than 20 dB. The expression of the autocorrelation function after pulse compression is:

[0121]

[0122] Wherein, R(τ) is the autocorrelation function; τ0 is the signal delay, which is calculated from the propagation time of the sound wave in the pipeline, where d is the propagation distance, c is the sound wave velocity; B is the signal bandwidth; j is the imaginary unit. Finally, the compressed signal is smoothed by a Hanning window function to reduce the influence of Gibbs phenomenon. This step aims to compress the broadband signal into a narrow pulse signal, enhance the distinguishability between the weak leakage sound signal and the background noise, and improve the sensitivity of subsequent leakage detection.

[0123] The specific implementation of step S04 is to perform time-frequency analysis on the acoustic signal based on wavelet transform. First, a wavelet basis function with good time-frequency localization performance, such as the Debyeche 5th-order wavelet, is selected to perform continuous wavelet transform on the preprocessed acoustic signal. The transform expression is:

[0124]

[0125] Wherein, W f (a, b) is the continuous wavelet transform coefficient; f(t) is the preprocessed acoustic signal; ψ(t) is the wavelet basis function, the Debyeche 5th-order wavelet; a is the scale parameter, which determines the frequency resolution, and its value range is 1 to 64; b is the translation parameter, which determines the time position, and its value range covers the entire signal duration. Then, the energy characteristics in the wavelet coefficient matrix are extracted to calculate the time-frequency energy distribution:

[0126] E(a, b) = |W f (a, b)| 2 ;

[0127] Wherein, E(a, b) is the time-frequency energy distribution. Then, the wavelet energy spectrum at different scales is calculated:

[0128]

[0129] Wherein, E s(a) is the energy spectrum at scale a; b1 and b2 are the time window boundaries. Then, identify the energy aggregation interval to determine the frequency center of the leakage signal. The typical frequency center range of the leakage signal is between 2 kHz and 6 kHz. Subsequently, analyze the time-varying characteristics of the energy and calculate the rate of change of wavelet energy with time:

[0130]

[0131] In the formula, is the rate of change of energy; Δt is the time step, with a value of 10 ms. The leakage signal usually shows a continuous and stable energy output, and the rate of change is lower than that of the normal flow signal. Finally, extract the acoustic propagation characteristic parameters, including group velocity and phase velocity, for subsequent propagation path analysis. The purpose of this step is to obtain the complete time-frequency characteristics of the acoustic signal and provide multi-dimensional feature support for leakage sound source identification.

[0132] The specific implementation of step S05 is to introduce the time reversal mirror technology to reconstruct the acoustic wave propagation path. First, perform a time reversal operation on the collected acoustic signal in the time domain to generate a time reversal sequence:

[0133] f TR (t) = f(T - t);

[0134] In the formula, f TR (t) is the time reversal signal; f(t) is the original received signal; T is the total duration of the signal. Then, establish a virtual sound source array model, set multiple possible sound source position points in the pipe network system, and then propagate the time reversal signal in the virtual sound source array. The acoustic field calculation expression is:

[0135]

[0136] In the formula, p(r, t) is the sound pressure at spatial position r at time t; G(r, r i , t) is the acoustic Green's function from position r i to position r, which is calculated through the pipe acoustic wave propagation model; N is the number of receivers; r i is the hydrophone position coordinate; r is the possible leakage point position coordinate in the pipe network. Subsequently, calculate the energy focusing degree:

[0137]

[0138] In the formula, FC(r) is the energy focusing degree; t maxis the time corresponding to the maximum sound pressure. The potential leakage point location is identified by the maximum energy focusing criterion. Generally, the energy focusing degree at the leakage point location is higher than 10 dB. Finally, the direct propagation signal and the multipath reflection signal are separated. The direct propagation signal usually appears as the first high-energy signal arriving, and the subsequent lower-energy signals are multipath reflection signals. This step aims to utilize the reversibility of sound wave propagation to solve the multipath propagation problem in complex pipe networks and improve the positioning accuracy.

[0139] The specific implementation of step S06 is to construct a fluid-acoustic coupling propagation equation to analyze the generation mechanism of the leakage sound source. First, a set of equations describing the physical mechanism of the generation and propagation of the leakage sound source in the pipeline system is established, including the continuity equation, the momentum equation, and the energy equation:

[0140]

[0141] In the formula, ρ is the fluid density, which is measured by the density sensor attached to the hydrophone array; is the fluid velocity vector, which is measured by the flow meter in the pipeline; p is the fluid pressure, which is measured by the pressure sensor; μ is the dynamic viscosity of the fluid, which is obtained by looking up the table according to the medium temperature; c is the sound wave propagation velocity, which is calculated by the sound wave velocity correction function in step S07; is the gradient operator; is the divergence operator; is the Laplace operator. Then, the sound wave propagation equation is established:

[0142]

[0143] In the formula, p′ is the sound pressure perturbation; ρ0 is the static density of the fluid; Q is the volume flow function, representing the leakage source strength, which is estimated by analyzing the fluid outflow characteristics at the leakage opening, where C d is the flow coefficient, with a value range of 0.6 to 0.9, A is the leakage opening area, and Δp is the pressure difference inside and outside the leakage. The sound pressure expression is:

[0144]

[0145] In the formula, r is the distance from the leakage point. The finite difference method is used to numerically solve the fluid-acoustic coupling propagation equation. The time step is set to 10 μs, and the space step is set to 5 mm. Finally, the sound pressure level at the distance and time point is calculated, and a quantitative relationship between the leakage fluid flow state parameters and the acoustic characteristics is established. The purpose of this step is to understand the characteristics of the leakage sound source from the physical mechanism level and provide a theoretical basis for sound source positioning.

[0146] The specific implementation of step S07 is to combine the pipeline network topology structure information and use the acoustic wave velocity correction function to compensate for the influence of medium density and flow velocity on the measurement accuracy. First, an acoustic wave velocity correction function is established, considering parameters such as the temperature, pressure, density, and flow velocity of the medium in the pipeline:

[0147] c eff =c0·[1+α T (T-T0)+α p (p-p0)+α ρ (ρ-ρ0)]·(1±M·cosθ);

[0148] In the formula, c eff is the corrected effective acoustic wave velocity; c0 is the acoustic wave velocity under reference conditions, with a value of 1500 m / s; T is the current temperature, obtained by measuring with a temperature sensor, with an accuracy of ±0.5 °C; T0 is the reference temperature, with a value of 20 °C; p is the current pressure, obtained by measuring with a pressure sensor, with an accuracy of ±0.5 kPa; p0 is the reference pressure, with a value of 101.3 kPa; ρ is the current density, obtained by measuring with a densitometer; ρ0 is the reference density, with a value of 1000 kg / m 3 ; α T is the temperature correction coefficient, with a value of 2.4×10 -3 / °C; α p is the pressure correction coefficient, with a value of 1.8×10 -5 / kPa; α ρ is the density correction coefficient, with a value of -3.0×10 -4 / (kg / m 3 ); M is the Mach number, θ is the angle between the acoustic wave propagation direction and the fluid flow direction, calculated from the pipeline geometric structure and the acoustic wave propagation path. Then, the parameters collected by the hydrophone array are used to update the acoustic wave velocity calculation model in real time. Next, the acoustic wave propagation velocities of different pipe segments are calculated. The sound speed in typical water media is about 1450 - 1550 m / s. Finally, the corrected acoustic wave velocity is applied to the leakage location calculation to reduce the measurement error. This step aims to improve the calculation accuracy of the acoustic wave propagation velocity and eliminate the influence of changes in medium characteristics on the location accuracy.

[0149] The specific implementation of step S08 is to use a pre-trained multi-modal model for acoustic propagation in the pipe network to optimize the sound source location process. First, the acoustic features and pipe network topology information extracted in the previous steps are input into the multi-modal model for acoustic propagation in the pipe network, which is designed based on a deep convolutional neural network and a self-attention mechanism. The pipeline structure encoder uses a graph convolutional network structure to encode the pipe network topology information, and the calculation formula is:

[0150]

[0151] where H (l) is the node feature matrix of the l-th layer; A is the adjacency matrix of the pipe network topology, constructed from the pipe network topology information; D is the degree matrix, and the diagonal element D ii is equal to the degree of node i; W (l) is the trainable weight matrix of the l-th layer, obtained by optimizing the model training process; σ is the activation function, using LeakyReLU. When the acoustic feature extractor uses a multi-layer one-dimensional convolutional network to extract time-frequency features, the calculation formula is:

[0152] F (l+1) = Pool(BN(σ(Conv1D(F (l) , K (l) )))));

[0153]

[0154] where F (l) is the acoustic feature map of the l-th layer, F (0) is the input acoustic signal, the time-frequency features obtained by the wavelet transform in step S04; K (l) is the convolutional kernel parameter of the l-th layer, obtained by optimizing the model training process; Conv1D is the one-dimensional convolutional operation; BN is the batch normalization operation; Pool is the max pooling operation; k is the convolutional kernel size, which is 3, 5, 7, 9, 11 for each layer respectively. The position prediction decoder fuses the pipeline structure and acoustic feature information through the multi-head cross-attention mechanism, and the calculation formula is:

[0155]

[0156] MultiHead(Q, K, V) = Concat(head1, head2,..., head h )W O ;

[0157] head i = Attention(QW i Q , KW i K , VW i V );

[0158] Z = MultiHead(H structure , F acoustic , F acoustic );

[0159] where Q, K, and V are the query, key, and value matrices respectively; d k is the dimension of the key vector; W iQ , W i K , W i V is the linear mapping matrix of the i-th attention head, which is optimized and obtained through the model training process; W O is the output mapping matrix, which is optimized and obtained through the model training process; h is the number of attention heads, which is determined according to the number of main branches of the pipeline system and is usually 4 to 8; H structure is the feature output by the pipeline structure encoder; F acoustic is the feature output by the acoustic feature extractor; Z is the fused feature output by the multi-head attention. Finally, the three-dimensional coordinates of the leakage point and the corresponding confidence are output by the model. The confidence calculation expression is:

[0160]

[0161] In the formula, C is the confidence of the positioning result; S is the similarity score output by the model, which is calculated by the multi-modal model of pipeline acoustic propagation; S th is the similarity threshold, with a value of 0.75; β is the sensitivity coefficient, with a value of 10. The confidence threshold is set to 0.85, and the prediction results below this threshold need to be further verified. The purpose of this step is to comprehensively analyze various acoustic features and pipeline network information using deep learning technology to achieve high-precision leakage positioning.

[0162] The specific implementation of step S09 is to establish a leakage acoustic feature database and classify and evaluate the leakage degree through machine learning algorithms. First, collect the acoustic feature data of different types and degrees of leakage, including spectral features, energy features, and time-domain features. Then, use the hierarchical clustering algorithm to group the leakage features according to similarity to form a leakage feature spectrum. Next, train a support vector machine classifier to classify the leakage degree into five levels, from slight leakage to severe leakage. The classification decision function is:

[0163]

[0164] K(x i , x) = exp(-γ||x i - x|| 2 );

[0165] In the formula, g(x) is the support vector machine decision function; x is the input acoustic feature vector, which is composed of the acoustic features extracted in the previous steps, including the center frequency, energy distribution, duration, etc.; x i is the support vector, which is obtained through the support vector machine training process; y i is the class label corresponding to the support vector; α iis the Lagrange multiplier obtained through the support vector machine training process; b is the bias term obtained through the support vector machine training process; K(x i , x) is the kernel function, and the radial basis function is adopted; γ is the kernel function parameter that determines the influence radius and is optimized through cross-validation. The typical value is 0.1 - 1.0. The classification accuracy is not less than 90%. Subsequently, the comprehensive leakage level is calculated:

[0166]

[0167] In the formula, L is the comprehensive leakage level; P j is the probability of leakage at the j-th level; w j is the weight coefficient that reflects the importance of different levels and generally increases with the increase of the level. The values of w1 to w5 are 0.1, 0.2, 0.3, 0.5, and 0.9 respectively. Suggestions for the maintenance priority are generated according to the leakage degree and location. The maintenance priority considers the severity of leakage, the importance of location, and the potential impact range. Finally, the newly detected leakage data is fed back to the database to continuously expand and optimize the leakage acoustic feature library. This step aims to evaluate and classify the detected leakage and provide a scientific basis for maintenance decision-making.

[0168] In this embodiment, the detailed structure of the multi-modal model for acoustic propagation in pipe networks is a feature extraction network for acoustic wave propagation based on the fusion of a deep convolutional neural network and a self-attention mechanism, which includes three main modules: The pipeline structure encoder adopts a graph convolutional network structure, including 3 layers of graph convolutional layers, with 64 feature channels in each layer. The activation function is LeakyReLU, and the pooling adopts the edge aggregation method to convert the pipe network topology information into a 512-dimensional feature vector; The acoustic feature extractor uses a 5-layer one-dimensional convolutional network, with convolutional kernel sizes of 3, 5, 7, 9, and 11, and the number of channels are 32, 64, 128, 256, and 512 respectively. After each layer, batch normalization and max pooling are connected to extract time-domain and frequency-domain features; The position prediction decoder uses a multi-head cross-attention mechanism, and the number of attention heads is equal to the number of main branches of the pipeline system. After the attention layer, 3 layers of fully connected layers are connected, with the number of neurons being 512, 256, and 128. Finally, the output layer has 3 neurons representing the three-dimensional coordinates of the leakage point and 1 neuron representing the confidence level. The model input includes the pipe network structure diagram information and the acoustic feature vector, and the output is the estimated result of the leakage point position. The total number of model parameters is about 2.5×10 6 pieces, and the inference time is controlled within 100 ms, meeting the real-time positioning requirements.

[0169] In this embodiment, the steps for establishing the training dataset of the multi-modal model for acoustic propagation in pipe networks are specifically as follows: First, various pipe structure models are constructed in a laboratory environment, including straight pipe segments, elbows, T-joints, and cross-shaped pipe segments. The pipe diameter ranges from 50 to 300 mm, and the pipe materials cover steel pipes, cast iron pipes, and concrete pipes. Different types of leakage situations are simulated, including crack-type, hole-type, and interface-looseness-type leakage. The leakage aperture is 0.5 to 5 mm, and the pressure gradient is 10 to 100 kPa. A 24-bit high-precision data acquisition device is used to record the complete acoustic signal data, and the sampling rate is set to 96 kHz. Then, computational fluid dynamics software is used to simulate various leakage scenarios, and a simulation environment including the pipe network geometric model, fluid parameter settings, boundary condition definitions, and mesh generation is established. The number of mesh cells is not less than 10 6 pieces, and the simulation time step is 10 -5 seconds to generate virtual acoustic signal data, and the accurate leakage source location information is marked. The location marking accuracy is ±1 cm. Finally, the laboratory data and the simulation data are mixed in a ratio of 7:3 to construct a dataset containing various working conditions. The total sample size is not less than 10 4 pieces. Each sample contains acoustic signal time series data, spectrum data, pipe network topology information, and leakage location labels. The 5-fold cross-validation method is used to ensure the quality of the dataset and ensure that the model has the generalization ability to handle complex situations in the actual engineering environment.

[0170] In this embodiment, the steps for pre-training the multi-modal model for acoustic propagation in pipe networks specifically include adopting a three-stage training strategy. In the first stage, unsupervised training is performed using large-scale pipe network topology data to enable the model to master the pipe network structure encoding ability. In the second stage, supervised fine-tuning is performed using acoustic feature data to optimize the model parameters for the leakage point feature recognition task. In the third stage, the leakage detection and localization performance are simultaneously optimized through multi-task learning, and a weighted loss function is used to balance the gradient contributions between different tasks, where the weight coefficient is dynamically adjusted according to the importance of leakage detection and precise localization in actual engineering applications. After training, the model adapts to pipe network systems of different scales and complexities in civil air defense projects and realizes the function of high-precision leakage localization.

[0171] To better understand and implement the present invention, the following provides an embodiment of a specific application scenario of the present invention: Researchers conducted a leakage localization experiment in a pipe network system of a large underground civil air defense project. The total length of the pipes in this system is about 5.2 km, including various materials such as steel pipes, cast iron pipes, and concrete pipes with different pipe diameters (100 mm to 400 mm). The service life of the system has exceeded 25 years, and leakage problems have occurred continuously in recent years. It is difficult to detect leakage points in hidden locations using traditional manual inspection methods. Therefore, the researchers used the method for leak location of hidden pipes in civil air defense projects based on acoustic features of the present invention for verification.

[0172] First, 48 high-sensitivity hydrophones were deployed inside the system. The sensitivity is -178 dB (referenced to 1 V / μPa), and the frequency response range is 50 Hz to 15 kHz. The hydrophones are located at key nodes of the pipeline system, including 28 pipeline intersections, 12 bends, and 8 pipe diameter changes, forming a complete acoustic signal acquisition network. The sampling frequency is set to 48 kHz, and the synchronization accuracy is 5 μs. The data is transmitted to the central processing system through an optical fiber bus.

[0173] The system was tested by acoustic wave excitation. A broadband acoustic pulse signal with a center frequency of 3 kHz and a bandwidth of 4 kHz was used. The signals collected by the hydrophones were separated by channels through time division multiple access technology, and the channel cross-correlation was less than 0.15. The adaptive noise cancellation algorithm was applied to process the original signals. A 96-order filter was used with a learning rate of 0.05. The noise cancellation effect is shown in Table 1:

[0174] Table 1 Comparison table of noise cancellation effects

[0175] Frequency Band (Hz) Original Signal-to-Noise Ratio (dB) Signal-to-Noise Ratio after Processing (dB) Improvement Amount (dB) 500-1000 3.2 9.8 6.6 1000-2000 4.5 12.7 8.2 2000-5000 5.8 15.2 9.4 5000-10000 2.9 8.4 5.5

[0176] The pulse compression technology was applied to the preprocessed signals. A chirp signal with a frequency of 1.5 kHz to 7.5 kHz and a duration of 40 ms was designed as the reference signal. Pulse compression was achieved through matched filtering. The main lobe width after compression is 0.83 ms, and the sidelobe suppression ratio is 23.6 dB, improving the recognition ability of weak leakage acoustic signals.

[0177] The Debye - 5th order wavelet was selected to perform continuous wavelet transform on the acoustic signals to analyze the time - frequency characteristics. The wavelet scale parameter a ranges from 1 to 48, with a total of 24 scales. The energy distribution characteristics of the wavelet coefficients were extracted, and the energy change rates of different pipe segments were compared and analyzed. As a result, 3 suspicious leakage points were found, and their energy change rates were 0.036 / s, 0.029 / s, and 0.044 / s respectively, which were significantly lower than 0.185 / s - 0.326 / s of the normal flow pipe segments.

[0178] The time - reversal mirror technology was introduced to reconstruct the acoustic wave propagation path to accurately locate the 3 suspicious leakage points. First, a virtual sound source array containing 825 potential leakage location points was constructed to cover the main area of the pipe network system. The energy focusing degree of each point was calculated, and the results are shown in Table 2:

[0179] Table 2 Energy focusing degree table of suspicious leakage points

[0180] Suspicious Location Number Maximum Energy Focus Degree (dB) Position Coordinate (m) Confidence Level L1 14.3 (127.4,85.2,-8.6) 0.92 L2 8.7 (358.9,204.5,-6.2) 0.74 L3 12.5 (462.1,138.7,-12.3) 0.88

[0181] According to the fluid - acoustic coupling propagation equation, the leakage mechanism was analyzed and the sound pressure distribution was calculated. The fluid parameters inside the pipeline were measured: density 998.2 kg / m 3, the pressure is 375.8 kPa, the temperature is 17.3 °C, and the flow velocity is 0.68 m / s. According to the acoustic wave velocity correction function, the corrected acoustic velocity is 1478.5 m / s. The finite difference method is used to solve the equation, with a time step of 8 μs and a spatial step of 4.5 mm. The relationship between the sound pressure at the leakage point and the distance is calculated as shown in Table 3:

[0182] Table 3 Sound Pressure Attenuation Characteristics Table at the Leakage Point

[0183] Distance (m) L1 Sound Pressure Level (dB) L2 Sound Pressure Level (dB) L3 Sound Pressure Level (dB) 0.5 82.6 75.3 80.1 1.0 76.8 69.5 74.2 2.0 70.9 63.7 68.4 5.0 62.4 55.2 59.8 10.0 56.1 49.0 53.5

[0184] The above data is input into a pre-trained multi-modal model for acoustic propagation in pipe networks for optimized positioning. The model includes a pipe structure encoder with 3 layers of graph convolutional networks and an acoustic feature extractor with 5 layers of one-dimensional convolutional networks, and uses a cross-attention mechanism with 6 attention heads to fuse features. The model positioning results are shown in Table 4:

[0185] Table 4 Optimized Leakage Location Results Table

[0186]

[0187]

[0188] The leakage degree of the located leakage point is evaluated. Ten feature dimensions including the center frequency, energy distribution, and duration are extracted from the acoustic features, and a support vector machine classifier is used for five-level classification, with the kernel function parameter γ being 0.35. The leakage degree evaluation results are shown in Table 5:

[0189] Table 5 Leakage Degree Evaluation Results Table

[0190]

[0191] The engineering verification results show that this method accurately locates L1 and L3. For L2, due to its location near the elbow of the pipeline, there is an error of about 1.2 m between the actual leakage point and the located point. Through on-site excavation, it is found that L1 is a leakage caused by a crack in the steel pipe weld, about 31 mm long; L3 is a leakage caused by corrosion perforation on the pipeline wall, with a diameter of about 2.6 mm; L2 is a leakage caused by aging of the seal at the pipeline joint. Compared with traditional detection methods, this method significantly improves the positioning efficiency and accuracy of hidden pipe leakage points in civil air defense projects.

[0192] The traditional method for locating hidden pipe leaks in civil air defense projects mainly relies on manual inspections, pressure testing methods, and tracer methods. Manual inspections are time-consuming and laborious and cannot detect leaks in hidden pipes; the pressure testing method can only determine the general area with low positioning accuracy; the tracer method poses environmental pollution risks and is complex to operate. The present invention uses an acoustic feature analysis method, combined with time division multiple access technology, wavelet transform, time reversal mirror technology, and fluid-acoustic coupling propagation analysis, to achieve high-precision positioning through a multi-modal model of acoustic propagation in pipe networks. Compared with traditional methods, the present invention has the following advantages: First, the positioning accuracy is high, with an average error of less than 60 cm; second, detection can be achieved without excavation, significantly reducing maintenance costs; third, it can evaluate the degree of leakage and provide maintenance suggestions, scientifically guiding maintenance decisions; fourth, it constructs an acoustic feature database for leaks, and as the scope of use expands, the system performance will continue to improve; fifth, it is applicable to complex pipe network environments, solving the problem of difficult positioning in traditional methods at elbows, joints, etc.

[0193] It should be noted that the detailed explanations of the variables involved in the present invention are shown in Tables 6, 7, and 8 below.

[0194] Table 6 Variable Explanation Table (Part 1)

[0195]

[0196]

[0197] Table 7 Variable Explanation Table (Part 2)

[0198]

[0199] Table 8 Variable Explanation Table (Part 3)

[0200]

[0201]

[0202] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered within the protection scope of the present invention.

Claims

1. A method for locating leakage of concealed pipes in civil air defense projects based on acoustic characteristics, characterized in that: include: Deploy hydrophone arrays at key nodes of the pipeline system to build an acoustic signal collection network; stimulate sound waves inside the pipeline and use time division multiple access technology to pre-process the collected signals; Apply pulse compression technology to improve the time domain resolution of signals; perform time-frequency analysis on acoustic signals based on wavelet transform to extract acoustic features; introduce Time reversal mirror technology reconstructs the path of sound wave propagation; The fluid-acoustic coupled propagation equation is constructed to analyze the generation mechanism of leakage sound sources. The sound wave velocity correction function is used to compensate for the influence of the medium in combination with the pipeline network topology information. The pre-trained pipeline network acoustic propagation multimodal model is used to locate the sound source.

2. The method according to claim 1, characterized in that The hydrophone array is a high-sensitivity hydrophone array deployed at key nodes of the civil air defense project pipeline system.

3. The method according to claim 2, characterized in that Time division multiple access technology refers to a multiple access technology that enables multiple signals to share the same frequency band without interfering with each other by allocating channels at different times. It is suitable for situations where multiple sensors in complex pipe networks collect signals at the same time.

4. The method according to claim 3, characterized in that Pulse compression technology refers to a method of converting broadband signals into narrow pulse signals through matched filter processing. It can improve the time domain resolution while maintaining signal energy and enhance the distinction between weak leakage sound signals and background noise.

5. The method according to claim 4, characterized in that Time reversal mirror technology refers to a method of reversing and re-transmitting the received acoustic signal in the time domain, using the reversibility of sound wave propagation to refocus the sound energy on the original sound source location, effectively solving the multipath propagation problem in complex pipe networks.

6. The method according to claim 5, characterized in that The inputs of the fluid acoustic coupling propagation equation include: fluid density measured by the hydrophone array, sound wave propagation distance identified by time reversal mirror technology, sound source flow function extracted by wavelet transform, sound wave propagation velocity corrected by the sound wave velocity correction function, and attenuation coefficient measured by pulse compression technology. The output is the sound pressure level at the distance and time point.

7. The method according to claim 6, characterized in that The sound wave velocity correction function refers to the mathematical expression of the sound wave propagation velocity calculation model that dynamically adjusts the parameters such as the medium temperature, pressure, density and flow rate in the pipeline to improve the positioning accuracy.

8. The method according to claim 7, characterized in that The specific structure of the multimodal model of pipe network acoustic propagation is a sound wave propagation feature extraction network based on the fusion of deep convolutional neural network and self-attention mechanism, which includes three main modules: pipe structure encoder, acoustic feature extractor and position prediction decoder.

9. The method according to claim 8, characterized in that The number of self-attention heads of the self-attention mechanism is determined according to the number of main branches of the pipeline system, and the number of main branches of the pipeline system comes from the constructed acoustic signal acquisition network.

10. The method according to claim 9, characterized in that The pipeline structure encoder uses a graph convolutional network structure to convert the pipeline network topology information into a high-dimensional feature vector; the acoustic feature extractor uses a multi-layer one-dimensional convolutional network to extract time domain and frequency domain features; The position prediction decoder fuses the pipeline structure and acoustic feature information through a multi-head cross-attention mechanism and outputs the leakage point location estimation result.

Images