A multi-modal pipe network leakage detection and positioning method based on a CNN-SVM hybrid architecture

Abstract

The application relates to the technical field of pipeline safety monitoring, and discloses a multi-modal pipeline network leakage detection and positioning method based on a CNN-SVM hybrid architecture, which synchronously collects and aligns an acoustic emission signal sequence and a pressure transient signal sequence, inputs the two sequences into a double-flow neural network after dimension conversion, extracts acoustic depth features and hydraulic transient features, and carries out weighted splicing to obtain a fusion feature vector; a support vector machine is used to classify the vector in a leakage state, and upstream and downstream node features and a time difference are obtained when leakage is determined; the features are input into a support vector regression model for nonlinear wave velocity dispersion compensation, and high-precision predicted leakage coordinates are output; and finally, a difficult example sample library is constructed through actual residuals to continuously update the model. The application effectively overcomes positioning errors caused by single-mode limitations and wave velocity dispersion under complex working conditions, and realizes self-evolution of model environmental adaptability.

Description

Technical Field

[0001] This invention relates to the field of pipeline safety monitoring technology, and in particular to a multimodal pipeline leak detection and localization method based on a CNN-SVM hybrid architecture. Background Technology

[0002] As a fundamental infrastructure of cities and industries, the safe operation of pipeline networks is crucial. Pipeline leaks not only waste resources but can also trigger serious safety accidents. Therefore, early detection and precise location of pipeline leaks have always been a key research area in the field of pipeline safety monitoring technology.

[0003] Existing pipeline leak detection technologies primarily rely on single-mode data sources, such as low-frequency hydraulic pressure signals or high-frequency acoustic emission signals. However, in complex real-world pipeline conditions (e.g., baseline drift during peak water usage periods, strong traffic noise interference), single-mode methods often struggle to balance detection sensitivity and interference resistance. Furthermore, when the high-frequency elastic shock waves generated by a leak propagate through the pipeline medium, they exhibit significant nonlinear frequency-varying attenuation and group velocity dispersion effects. Traditional time-difference-of-arrival (TDOA) positioning methods typically assume a constant sound wave propagation speed, making it difficult to effectively compensate for this nonlinear wave velocity dispersion, thus leading to substantial errors in absolute spatial positioning.

[0004] In recent years, although deep learning has been introduced into this field, conventional convolutional neural networks (CNNs) mostly use black-box feature extraction and lack deep coupling with the physical attenuation law of pipeline networks. Moreover, statically trained models will rapidly degrade their generalization ability when faced with dynamic data drift such as long-term pipeline aging and soil impedance changes, making it difficult to meet the requirements of practical engineering for high accuracy and continuous robustness. Summary of the Invention

[0005] This invention provides a multimodal pipeline leak detection and localization method based on a CNN-SVM hybrid architecture. By using deep alignment fusion of multimodal features, L2-Margin boundary constraint classification, SVR nonlinear dispersion compensation localization, and a closed-loop self-evolution mechanism for difficult examples, it achieves high-precision and robust detection and localization of pipeline leaks under complex operating conditions.

[0006] This invention provides a multimodal pipeline leak detection and localization method based on a CNN-SVM hybrid architecture, comprising:

[0007] S1. Collect the acoustic emission signal sequence and pressure transient signal sequence of the monitoring nodes in the pipeline network and add timestamps. Based on the timestamps, align the acoustic emission signal sequence and pressure transient signal sequence with time windows and perform data dimension transformation respectively to obtain the acoustic feature matrix and hydraulic time series feature tensor.

[0008] S2. Input the acoustic feature matrix and the hydraulic time series feature tensor into two independent network branches of a pre-constructed two-stream neural network for feature extraction. Then, concatenate the extracted acoustic depth features and hydraulic transient features to obtain a fused feature vector.

[0009] S3. Input the fused feature vector into a pre-constructed support vector machine for state classification. When the state classification result is a leakage state, obtain the fused feature vectors corresponding to the upstream monitoring node and the downstream monitoring node on both sides of the leakage location, and calculate the time difference between the leakage signal propagating to the upstream monitoring node and the downstream monitoring node.

[0010] S4. Input the time difference, the fusion feature vector corresponding to the upstream monitoring node, and the fusion feature vector corresponding to the downstream monitoring node into the pre-constructed support vector regression model to perform regression calculation for nonlinear wave velocity dispersion compensation, and output the predicted leakage coordinates.

[0011] S5. Determine difficult case samples based on the residual between the predicted leakage coordinates and the actual leakage coordinates, and update the support vector machine and support vector regression model based on the difficult case samples; wherein, the difficult case samples include the acoustic emission signal sequence, pressure transient signal sequence and actual leakage coordinates corresponding to the residual.

[0012] Furthermore, S1 specifically includes:

[0013] The acoustic emission signal sequence and the pressure transient signal sequence are acquired synchronously, and an absolute timestamp is injected into the acoustic emission signal sequence and the pressure transient signal sequence using a high-precision time synchronization module;

[0014] Each sampling point of the pressure transient signal sequence is defined as a discrete reference time anchor point; based on the preset observation time window width, with the absolute timestamp of each reference time anchor point as the center, the acoustic emission signal subsequence falling within the corresponding time window is extracted from the acoustic emission signal sequence to complete the time window alignment.

[0015] The acoustic emission signal subsequence is processed by continuous wavelet transform to generate a two-dimensional energy spectrum matrix, which is then mapped into an image tensor and normalized to obtain the acoustic feature matrix.

[0016] The aligned pressure transient signal sequence is normalized and then feature-mapped using a one-dimensional convolution operator to generate a hydraulic time series feature tensor.

[0017] Furthermore, S2 specifically includes:

[0018] The acoustic feature matrix is ​​input into the two-dimensional convolutional branch of the two-stream neural network, and acoustic depth features are extracted through multi-layer two-dimensional convolution and global average pooling layer operations.

[0019] The hydraulic time series feature tensor is input into the parallel one-dimensional convolutional branch of the two-stream neural network. High-frequency local mutation points and low-frequency global reflection periods in the hydraulic time series feature tensor are extracted through multiple parallel one-dimensional convolutional layers. After being concatenated in the channel dimension, the hydraulic transient features are output after dimensionality reduction.

[0020] The acoustic depth features and the hydraulic transient features are mapped to a shared latent space through a linear mapping layer to obtain the query vector and the key vector, respectively.

[0021] Calculate the outer product between the query vector and the key vector, and obtain the cross-covariance weight matrix after scaling and normalization. Use the cross-covariance weight matrix to extract the global channel weights, and obtain the acoustic attention vector and hydraulic attention vector by dimensionality reduction through a linear projection layer.

[0022] The acoustic depth feature is weighted element-wise using the acoustic attention vector and the hydraulic transient feature is weighted using the hydraulic attention vector, and then concatenated along the feature dimension to generate a fused feature vector.

[0023] Furthermore, S3 specifically includes:

[0024] The fused feature vector is input into a pre-constructed support vector machine, spatial mapping is performed through a radial basis kernel function, and the state classification result is output based on the decision hyperplane;

[0025] When the state classification result is a leak state, the topology association matrix of the pipeline geographic information system is retrieved, and bidirectional backtracking is performed along the physical topology of the pipeline with the current monitoring node as the center. The two nodes closest to the leak location are marked as the upstream monitoring node and the downstream monitoring node, respectively.

[0026] Data from the upstream and downstream monitoring nodes within the same absolute time window are extracted respectively, and the time window alignment, data dimension transformation, and feature extraction and splicing processes are reused to generate fused feature vectors for the upstream and downstream monitoring nodes respectively.

[0027] Extract the acoustic emission signal sequences corresponding to the upstream and downstream monitoring nodes, and perform fast Fourier transform on each to obtain the corresponding frequency domain signals;

[0028] The cross-power spectral density of the frequency domain signal is calculated, and a phase transform weighting function is introduced for whitening. The generalized cross-correlation function is obtained by inverse fast Fourier transform.

[0029] The time delay independent variable corresponding to the global maximum peak value on the generalized cross-correlation function curve is extracted and used as the time difference for the leakage signal to propagate to the upstream and downstream monitoring nodes.

[0030] Furthermore, the decision classification flow formula for the decision hyperplane is:

[0031]

[0032] In the formula, This is a symbolic function used to output discrete classification labels; The total number of support vectors retained by the support vector machine during the training phase; For the sequence index of support vectors, ; The i-th support vector is determined in advance during training; Let Lagrange multipliers be the values ​​corresponding to the i-th support vector; The category label corresponding to the i-th support vector; The bandwidth parameter of the radial basis kernel function controls the nonlinear range of the eigenmap. denoted by L2 norm between feature vectors; b is the bias term of the decision classification hyperplane;

[0033] The formula for the generalized cross-correlation function is:

[0034]

[0035] In the formula, for The complex conjugate; Indicates the modulus of a complex number; For a continuously sliding time delay independent variable, search for a generalized cross-correlation function. Extract the global maximum peak value on the curve and the corresponding horizontal coordinate time delay variable. The time difference between the propagation of the acoustic emission shock wave from the leak to the upstream and downstream monitoring nodes .

[0036] Furthermore, S4 specifically includes:

[0037] The fusion feature vectors corresponding to the upstream monitoring nodes, the fusion feature vectors corresponding to the downstream monitoring nodes, and the time difference are flattened and stitched together to generate a joint input vector.

[0038] The joint input vector is input into a pre-constructed support vector regression model. The joint input vector is mapped to a high-dimensional feature space using a regression kernel function. The continuous output value of the support vector regression model is extracted. The continuous output value is used as the relative one-dimensional curve distance of the leakage source from the upstream monitoring node along the physical pipeline of the pipeline network.

[0039] Retrieve the topological correlation matrix of the pipeline network geographic information system to obtain the three-dimensional spatial absolute coordinates of the upstream monitoring node and the three-dimensional spatial trajectory equation of the pipeline network connecting the upstream monitoring node and the downstream monitoring node;

[0040] Starting from the absolute coordinates in the three-dimensional space, a curve segment with a length equal to the relative one-dimensional curve distance is intercepted along the three-dimensional trajectory equation of the pipeline space towards the downstream monitoring node;

[0041] The endpoint of the curve segment is calculated using three-dimensional coordinates to obtain a spatial location vector containing longitude, latitude, and elevation information. This vector is then used as the predicted leakage coordinates and output.

[0042] Furthermore, the support vector regression model employs a nonlinear wave velocity dispersion compensation regression function to output continuous output values. :

[0043]

[0044] In the formula, is the total number of support vectors in the support vector regression model; j is the index of the support vector. This refers to the j-th regression support vector determined during the training phase. and Let Lagrange multipliers be the values ​​corresponding to the j-th support vector. Here is the regression kernel function for SVR; This is the bias term of the regression model.

[0045] Furthermore, S5 specifically includes:

[0046] Obtain the actual leakage coordinates from the field feedback, and calculate the spatial distance residual between the predicted leakage coordinates and the actual leakage coordinates as the positioning residual; if the field feedback result is a false alarm, assign the positioning residual a preset maximum penalty value;

[0047] Determine whether the positioning residual exceeds a preset tolerance threshold. If the positioning residual exceeds the tolerance threshold, extract the corresponding acoustic emission signal sequence and pressure transient signal sequence, and bind and encapsulate the acoustic emission signal sequence and pressure transient signal sequence with the actual leakage coordinates to construct a difficult case sample and store it in the difficult case database.

[0048] The database of difficult examples is monitored. When the total number of accumulated samples reaches the preset model update trigger threshold, the acoustic emission signal sequence and pressure transient signal sequence in the difficult example samples are used to fine-tune the convolution kernel parameters of the two-stream neural network in reverse.

[0049] Based on the updated feature vector after parameter fine-tuning, the Lagrange multipliers of the support vector machine are recalculated to adjust the classification hyperplane;

[0050] Using the actual leakage coordinates in the difficult sample as the regression target, the kernel function weights and Lagrange multipliers of the support vector regression model are re-optimized.

[0051] The present invention also provides a multimodal pipeline leak detection and localization device based on a CNN-SVM hybrid architecture. Based on the multimodal pipeline leak detection and localization method based on a CNN-SVM hybrid architecture described above, the device includes:

[0052] The preprocessing module is used to collect the acoustic emission signal sequence and pressure transient signal sequence of the monitoring nodes in the pipeline network and add timestamps. Based on the timestamps, the acoustic emission signal sequence and pressure transient signal sequence are aligned with time windows and the data dimensions are transformed respectively to obtain the acoustic feature matrix and hydraulic time series feature tensor.

[0053] The feature extraction module is used to input the acoustic feature matrix and the hydraulic time series feature tensor into two independent network branches of a pre-constructed two-stream neural network for feature extraction, and to concatenate the extracted acoustic depth features and hydraulic transient features to obtain a fused feature vector.

[0054] The state classification module is used to input the fused feature vector into a pre-constructed support vector machine for state classification. When the state classification result is a leakage state, the module obtains the fused feature vectors corresponding to the upstream monitoring nodes and downstream monitoring nodes on both sides of the leakage location, and calculates the time difference between the leakage signal propagating to the upstream monitoring nodes and downstream monitoring nodes.

[0055] The coordinate prediction module is used to input the time difference, the fused feature vector corresponding to the upstream monitoring node, and the fused feature vector corresponding to the downstream monitoring node into a pre-constructed support vector regression model to perform regression calculations for nonlinear wave velocity dispersion compensation, and output the predicted leakage coordinates.

[0056] An iterative update module is used to determine difficult case samples based on the residual between the predicted leakage coordinates and the actual leakage coordinates, so as to update the support vector machine and support vector regression model based on the difficult case samples; wherein, the difficult case samples include the acoustic emission signal sequence, pressure transient signal sequence and actual leakage coordinates corresponding to the residual.

[0057] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.

[0058] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method.

[0059] The beneficial effects of this invention are as follows:

[0060] This invention effectively integrates the characteristics of high-frequency acoustic shock waves and low-frequency hydraulic reflections by combining acoustic emission and pressure transient signals and using a reference time anchor for time window alignment and dimensional transformation, overcoming the deficiency of single-mode representation capability under complex working conditions. It utilizes a two-stream neural network combined with a self-attention mechanism to extract and dynamically weight cross-modal features, while introducing an L2-Margin loss function to constrain and maximize the Euclidean distance of outlier features, greatly improving the robustness and accuracy of large-boundary classification by Support Vector Machine (SVM). In the localization stage, it reuses multi-dimensional deep fusion features from upstream and downstream nodes, combined with the initial time difference, and feeds them into the Support Vector Regression (SVR) model for mapping calculation, achieving implicit analysis and deep compensation for nonlinear wave velocity dispersion effects, improving the accuracy of physical coordinate localization of leak points. Furthermore, it introduces a closed-loop feedback and dynamic retraining mechanism based on absolute spatial positioning residuals for difficult samples, enabling the underlying network and regression model to evolve and iterate with changes in the pipeline network's service environment, achieving high-precision localization of pipeline leaks under complex working conditions. Attached Figure Description

[0061] Figure 1 This is a schematic diagram of a method flow according to an embodiment of the present invention.

[0062] Figure 2 This is a schematic diagram of the device structure according to an embodiment of the present invention.

[0063] Figure 3 This is a schematic diagram of the internal structure of a computer device according to an embodiment of the present invention.

[0064] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0065] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0066] like Figure 1 As shown, this invention provides a multimodal pipeline leak detection and localization method based on a CNN-SVM hybrid architecture, comprising:

[0067] S1. Acquire the acoustic emission signal sequence and pressure transient signal sequence of the monitoring nodes in the pipeline network and add timestamps. Based on the timestamps, align the acoustic emission signal sequence and pressure transient signal sequence with time windows and perform data dimension transformation respectively to obtain the acoustic feature matrix and hydraulic time series feature tensor.

[0068] In one specific embodiment, step S1 includes the following sub-steps:

[0069] S101. Acquire acoustic emission signal sequences and pressure transient signal sequences simultaneously through multi-source sensing hardware, and inject absolute timestamps.

[0070] Specifically, acoustic emission sensors (e.g., piezoelectric ceramic hydrophones) and high-frequency pressure transmitters are deployed at various monitoring nodes in the pipeline network. The acoustic emission sensors are used to capture high-frequency elastic shock waves propagating on the surface of the pipeline network or in the fluid, generating a one-dimensional acoustic emission signal sequence with a high sampling rate (e.g., 10kHz~100kHz); the pressure transmitters are used to capture low-frequency hydrodynamic parameter changes in the fluid inside the pipe, generating a one-dimensional pressure transient signal sequence with a low sampling rate (e.g., 100Hz~500Hz).

[0071] During data acquisition, a high-precision time synchronization module (such as a PTP precise time protocol module based on BeiDou or GPS) built into the monitoring node is used to assign a unified absolute timestamp to each data packet of the acquired acoustic emission signal sequence and pressure transient signal sequence. The timestamping action ensures that, even with minor delays and jitters in the heterogeneous sensor hardware acquisition link, the two types of raw signals have an absolute reference on the physical time axis, providing a basis for signal phase alignment.

[0072] S102. Using the sampling points of the pressure transient signal sequence as the reference time anchor, extract the corresponding acoustic emission signal sub-sequence based on the timestamp to complete the time window alignment.

[0073] Because the sampling rates of acoustic and hydraulic signals differ by orders of magnitude, forcibly aligning them using conventional interpolation resampling would destroy the original physical phase of the high-frequency sound waves. Therefore, this embodiment employs a time anchor alignment mechanism, specifically...

[0074] Each sampling point of the pressure transient signal sequence with a low sampling rate is defined as a discrete reference time anchor point; subsequently, according to the preset observation time window width (e.g., Using the absolute timestamp of each time anchor point as the center, the high-frequency data segments falling within the time window are addressed and extracted in the acoustic emission signal sequence to form the corresponding acoustic emission signal subsequence. Through this step, the two heterogeneous modes are accurately aligned in the underlying physical time window dimension, while completely preserving the transient phase information of the leaking shock wave.

[0075] S103. Perform continuous wavelet transform on the acoustic emission signal subsequence to generate an acoustic feature matrix, and perform temporal normalization and one-dimensional convolution mapping on the aligned pressure transient signal sequence to generate a hydraulic temporal feature tensor.

[0076] To map a one-dimensional time-domain acoustic signal to a higher-dimensional space that reflects the physical characteristics of frequency-varying dispersion, this embodiment employs continuous wavelet transform (CWT) to increase the dimensionality of the acoustic emission signal subsequence. When performing the CWT, the Morlet wavelet, which exhibits high sensitivity to transient stress shock waves, is selected as the mother function. The truncated acoustic emission signal subsequence... The specific transformation formula for scaling and translation is as follows:

[0077]

[0078] in, This represents the wavelet coefficients corresponding to scale a and translation position b after continuous wavelet transform; This represents the amplitude of the extracted acoustic emission signal subsequence at time t; t is a continuous time variable; a is a scaling factor (a>0), which controls the scaling of the wavelet function and is physically inversely proportional to the frequency of the acoustic emission signal; b is a time shift factor, which controls the shift of the wavelet function on the time axis and physically corresponds to the time position of the leakage signal. Indicates the selected Morlet wavelet mother function .

[0079] The wavelet coefficients under continuous scaling factor and continuous time shift factor are calculated using the above formula. Since this coefficient is a complex number, we need to solve for its complex modulus. The signal energy distribution values ​​are obtained; the scale factor a and the time shift factor b are discretized and sampled, with the length of the intercepted time window as the width (defined as the number of sampling points as W) and the set frequency analysis band scale range as the height (defined as the scale level as H), to generate a two-dimensional energy spectrum matrix of size H×W; then, using a preset pseudo-color mapping function (Colormap), the energy distribution values ​​in the two-dimensional energy spectrum matrix are mapped to the color space and converted into an RGB three-channel image tensor of size H×W×3; finally, the pixel values ​​of each channel in the RGB three-channel image tensor are normalized (for example, scaled to the [0,1] or [-1,1] interval), and the normalized three-channel image tensor is used as the acoustic feature matrix.

[0080] To address the interference of pressure baseline drift (such as gradual fluctuations caused by peak or trough water usage) during routine pipeline operation on feature extraction, Z-score normalization is first applied to the low-frequency pressure transient signal sequence. This normalizes each sampling point x in the original pressure transient signal sequence to a standard normal distribution with a mean μ of 0 and a standard deviation σ of 1. Subsequently, a pre-defined one-dimensional convolution operator is used to perform feature mapping on the normalized sequence. Preferably, a wide receptive field convolution kernel of size 1×5 or 1×7 is employed to capture water hammer wave features at different time scales. To meet the matrix dimension consistency requirements during subsequent cross-modal feature stitching, the number of output channels for the one-dimensional convolution operation is set to match the feature depth of the acoustic feature matrix (e.g., uniformly upscaling to 64 or 128 channels), thereby upscaling the one-dimensional pressure scalar time series into a one-dimensional hydraulic time series feature tensor with multiple feature channels.

[0081] S2. Input the acoustic feature matrix and the hydraulic time series feature tensor into two independent network branches of the pre-constructed two-stream neural network for feature extraction. Then, concatenate the extracted acoustic depth features and hydraulic transient features to obtain a fused feature vector.

[0082] In one specific embodiment, step S2 includes the following sub-steps:

[0083] S201. The acoustic feature matrix is ​​input into the two-dimensional convolutional branch of the two-stream neural network, and the acoustic depth features characterizing the frequency-varying attenuation law are extracted through multi-layer two-dimensional convolution and pooling operations. The hydraulic temporal feature tensor is input into the one-dimensional convolutional branch of the two-stream neural network, and the hydraulic transient features of different periods are captured through multi-scale convolution.

[0084] The two-dimensional convolutional branch employs a deep residual network with fully connected layers removed. An acoustic feature matrix of size H×W×3 is input to this branch; the network's initial shallow convolutional kernels capture the high-frequency edge pixel gradients representing the leakage shock wave in the acoustic feature matrix; the network's deep convolutional kernels extract the decay gradients of pixel energy as a function of the time scale factor (i.e., the frequency scale). Subsequently, a global average pooling (GAP) layer compresses the spatial dimension of the feature map, mapping the high-dimensional matrix to a length of H×W×3. A one-dimensional real column vector (e.g., a one-dimensional tensor of length 256) serves as an acoustic depth feature, which implicitly encodes the dispersion pattern of acoustic group velocity in the pipeline network under specific pipe diameters and materials at the data level.

[0085] Since the reflection period of hydraulic transient waves in the pipeline network varies with the distance from the leak point and the differences in topology, the one-dimensional convolutional branch adopts a multi-scale parallel one-dimensional convolutional architecture. The hydraulic temporal feature tensor generated in step S1 is synchronously input into three parallel one-dimensional convolutional layers with different kernel sizes (e.g., lengths of 3, 5, and 7) to extract high-frequency local abrupt changes and low-frequency global reflection periods from the hydraulic signal, respectively. Subsequently, the outputs of the three parallel branches are concatenated along the channel dimension, and dimensionality is reduced using a one-dimensional max pooling layer and a global average pooling layer, finally outputting a value of length [missing information]. A one-dimensional real column vector (e.g., a one-dimensional tensor of length 128) is used as a hydraulic transient feature.

[0086] S202. Calculate the self-attention cross-covariance weights of acoustic depth features and hydraulic transient features in the shared latent space.

[0087] Since acoustic and hydraulic modes differ in their ability to characterize leakage conditions under different operating conditions, this embodiment introduces a self-attention mechanism to assign modal weights.

[0088] First, the length is transformed through a linear mapping layer. The acoustic depth features and length are Hydraulic transient characteristics are mapped to have the same dimension The shared latent space yields the query vector Q and the key vector K (where... Secondly, calculate the outer product between the two and divide by the scaling factor. Then, after normalization along the column direction using the Softmax function, the cross-covariance weight matrix reflecting the correlation between the two types of physical signals is obtained. The calculation formula is as follows:

[0089]

[0090] The weight matrix A can adaptively evaluate the credibility of the physical interaction between the acoustic shock wave characteristics and the water hammer transient characteristics in each hidden layer dimension within the current analysis time window.

[0091] S204. Based on the cross-covariance weight, the acoustic depth features and hydraulic transient features are dynamically weighted and spliced ​​to generate a fused feature vector.

[0092] The global channel weights are extracted using the cross-covariance weight matrix A obtained in step S203, and then the acoustic attention vectors are obtained by dimensionality reduction through a linear projection layer. and hydraulic attention vector Subsequently, the original acoustic depth features were analyzed. and the aforementioned hydraulic transient characteristics Element-wise attention-weighted multiplication is performed, and the results are concatenated along the feature dimension to generate a total length of [length missing]. The one-dimensional real sequence tensor is used as the final output fused feature vector. The specific calculation formula for the weighted and splicing process is as follows:

[0093]

[0094] In the formula, For Hadama accumulation, This is a vector concatenation operation.

[0095] It is important to note that a joint L2-Margin loss function was configured for end-to-end parameter optimization during the pre-training of the dual-stream neural network. Let the two fused feature vectors extracted from the same pipeline sample under different operating conditions be... and The same type of leakage state is labeled y=1, and the different type is labeled y=0. m is a preset boundary threshold, and its loss function is defined as:

[0096]

[0097] The constraint of this loss function forces the two-stream neural network to maximize the Euclidean distance between heterogeneous feature clusters in the feature space when extracting features, so that the fused feature vector output by this embodiment has the physical property of supporting the large boundary partitioning of the subsequent classification network (such as support vector machine SVM) in terms of geometric topology.

[0098] S3. Input the fused feature vector into a pre-constructed support vector machine for state classification. When the state classification result is a leakage state, obtain the fused feature vectors corresponding to the upstream monitoring node and the downstream monitoring node on both sides of the leakage location, and calculate the time difference between the leakage signal propagating to the upstream monitoring node and the downstream monitoring node.

[0099] In one specific embodiment, step S3 includes the following sub-steps:

[0100] S301. Input the fused feature vector into the pre-constructed support vector machine, perform nonlinear spatial mapping through the radial basis kernel function, and output the state classification result based on the decision hyperplane.

[0101] Receive the fused feature vector output from step S2 Since the fused feature vector is constrained by the L2-Margin loss function during the training phase of the two-stream neural network, and the heterogeneous features already possess the maximized Euclidean distance in the underlying data space, this embodiment directly uses the radial basis function (RBF) to map it to a high-dimensional linearly separable space, and outputs the state classification result based on the decision hyperplane. The specific decision classification flow formula is as follows:

[0102]

[0103] In the formula, This is a symbolic function used to output discrete classification labels; The total number of support vectors retained by the support vector machine during the training phase; For the sequence index of support vectors, ; The i-th support vector is determined in advance during training; Let Lagrange multipliers be the values ​​corresponding to the i-th support vector; For the category label corresponding to the i-th support vector (e.g., set...), Characterize the leakage state. (representing the normal state) The bandwidth parameter of the radial basis kernel function controls the nonlinear range of the eigenmap. denoted by L2 norm (i.e., Euclidean distance) between feature vectors; b is the bias term of the decision classification hyperplane.

[0104] The state classification result is generated based on the calculation result output by the decision function: If If so, the current pipeline network is classified as being in a leaking state; if If so, the output pipeline is determined to be in a normal state.

[0105] S302. When the status classification result is a leakage status, the monitoring nodes are determined based on the pipeline topology correlation matrix, and the fused feature vectors corresponding to the upstream monitoring nodes and downstream monitoring nodes are obtained respectively.

[0106] When the output of step S301 triggers a leak status determination, the pre-preparation mechanism for the location data stream is executed. The geographic information system (GIS) topology association matrix of the pipeline network is retrieved, and the current monitoring node with the abnormal status is used as the center. The network is traced back bidirectionally along the physical topology to locate and mark the two hardware nodes closest to the leak location, which are designated as the upstream and downstream monitoring nodes, respectively.

[0107] For the identified upstream and downstream monitoring nodes, within the same absolute time window, the underlying data processing streams of steps S1 and S2 (including cross-modal phase alignment, continuous wavelet transform, and dual-stream feature extraction and splicing) are reused to independently extract the upstream fused feature vector containing the frequency-varying attenuation characteristics of the upstream acoustic waves. And downstream fused feature vectors containing downstream acoustic wave frequency attenuation characteristics. The two fused feature vectors obtained will be stored in a cache matrix, serving as the data source for subsequent SVR extraction of pipe nonlinear parameters.

[0108] S303. Extract the underlying acoustic emission signal sequences corresponding to the upstream and downstream monitoring nodes, and calculate the time difference based on the generalized cross-correlation-phase transformation algorithm.

[0109] To provide the underlying physical observations of time difference for subsequent regression models, this embodiment utilizes the underlying raw signals marked with high-precision time synchronization hardware for time delay estimation. Specifically, the acoustic emission signal sequences of the upstream and downstream monitoring nodes within this time window are extracted and denoted as follows: and .

[0110] The propagation delay of these two high-frequency signals is calculated using the Generalized Cross-Correlation-Phase Transform (GCC-PHAT) algorithm. The underlying data flow of the computation is as follows:

[0111] First of all, and Perform Fast Fourier Transform (FFT) on each signal to obtain the corresponding frequency domain signal. and Subsequently, the cross-power spectral density of the two components was calculated, and a phase transform weighting function (PHAT) was introduced for whitening to eliminate the interference of multipath reflections from the pipeline network and environmental reverberation on the phase characteristics. Then, the generalized cross-correlation function was obtained through inverse fast Fourier transform (IFFT). Its formula is:

[0112]

[0113] In the formula, for The complex conjugate; Indicates the modulus of a complex number; The time delay is a continuously sliding independent variable. Finally, a generalized cross-correlation function is searched. Extract the global maximum peak value on the curve and the corresponding horizontal coordinate time delay variable. , will The time difference between the propagation of the acoustic emission shock wave from the leak to the upstream and downstream monitoring nodes .

[0114] S4. Input the time difference, the fusion feature vector corresponding to the upstream monitoring node, and the fusion feature vector corresponding to the downstream monitoring node into the pre-constructed support vector regression model to perform regression calculation for nonlinear wave velocity dispersion compensation, and output the predicted leakage coordinates.

[0115] In one specific embodiment, step S4 includes the following sub-steps:

[0116] S401. Multi-dimensional data splicing is performed on the fusion feature vectors corresponding to the upstream monitoring nodes, the fusion feature vectors corresponding to the downstream monitoring nodes, and the time difference to construct a localization regression input matrix.

[0117] In traditional pipeline network physical models, the attenuation coefficient of high-frequency sound waves... The relationship between the frequency f and the nonlinear physical equation is satisfied. Since factors such as the aging degree of pipe materials and soil medium impedance are unmeasurable in actual pipeline networks, traditional methods are difficult to solve for the nonlinear parameters a and b of the pipe materials in real time.

[0118] To overcome this physical limitation, this embodiment implements a deep feature reuse mechanism. The upstream fused feature vector cached in step S302 is extracted. and downstream fusion feature vector This is compared with the initial time difference calculated in step S303. Perform data flattening and concatenation operations to generate a joint input vector. Its mathematical expression is:

[0119]

[0120] The joint input vector It integrates the propagation delay in the pure time dimension with the frequency band energy distribution in the spatial dimension, thus forming a data space that characterizes the nonlinear attenuation and dispersion of acoustic signals.

[0121] S402. Feed the joint input vector into the pre-built support vector regression model, and generate and output the predicted leakage coordinates after nonlinear wave velocity dispersion compensation by implicitly analyzing the wave velocity dispersion effect through the feature map activation value difference.

[0122] Joint input vector The input is a pre-trained, convergent Support Vector Regression (SVR) model based on a large amount of measured data from the pipeline network. During this process, the SVR model does not explicitly solve for the nonlinear parameters a and b, but instead utilizes the regression kernel function. Will Mapped to a high-dimensional feature space.

[0123] In a high-dimensional feature space, the SVR model calculates... and The difference in activation value attenuation across different frequency band feature maps implicitly fits the frequency-varying attenuation rate and group velocity dispersion law of sound waves propagating in the pipe. The specific nonlinear wave velocity dispersion compensation regression function is expressed as follows:

[0124]

[0125] In the formula, is the total number of support vectors in the support vector regression model; j is the index of the support vector. The j-th regression support vector determined during the training phase represents an anchor point sample with typical dispersion characteristics in the historical pipeline network data. and Let be the Lagrange multiplier corresponding to the j-th support vector, whose value is determined by the training phase. Determining the insensitive loss function; The regression kernel function for SVR (preferably the Gaussian radial basis function). This is the bias term of the regression model.

[0126] Through the kernel function mapping and hyperplane regression calculations described above, the support vector regression model transforms the complex frequency-varying attenuation and time-difference characteristics into one-dimensional continuous physical coordinates. The continuous output values ​​of the regression function are then extracted. This value, in a physical sense, represents the relative one-dimensional curved distance of the leakage source from the upstream monitoring node along the physical pipeline of the pipeline network.

[0127] To output the absolute spatial location suitable for actual engineering scheduling, the topological association matrix of the pipeline geographic information system (GIS) from step S302 is retrieved to obtain the three-dimensional absolute spatial coordinates of the upstream monitoring node and the three-dimensional trajectory equation of the pipeline space connecting the upstream and downstream monitoring nodes. Starting from the three-dimensional absolute spatial coordinates of the upstream monitoring node, a length of [missing value] is extracted along the three-dimensional trajectory equation of the pipeline space towards the downstream monitoring node. The curve segment is calculated in three dimensions, and a spatial location vector containing longitude, latitude and elevation information is generated. This vector is used as the final predicted leakage coordinates after nonlinear wave velocity dispersion compensation and is output. These coordinates directly map the physical leakage point location in the real world.

[0128] S5. Determine difficult case samples based on the residual between the predicted leakage coordinates and the actual leakage coordinates, and update the support vector machine and support vector regression model based on the difficult case samples; wherein, the difficult case samples include the acoustic emission signal sequence, pressure transient signal sequence and actual leakage coordinates corresponding to the residual.

[0129] In one specific embodiment, step S5 includes the following sub-steps:

[0130] S501. Obtain the actual leakage coordinates from the closed-loop feedback at the site, and calculate the spatial distance residual between the predicted leakage coordinates and the actual leakage coordinates.

[0131] Based on the predicted leak coordinates output in step S4, the pipeline maintenance unit dispatches manual inspections or conducts on-site excavation verification. Through an IoT interface or the enterprise work order system, it receives feedback results from on-site manual inspection equipment or the pipeline geographic information system (GIS). The feedback results include the absolute three-dimensional spatial coordinates confirming the existence of a leak, i.e., the actual leak coordinates. .

[0132] After obtaining the actual leak coordinates, the predicted leak coordinates output in step S4 are calculated based on a unified geodetic coordinate system (such as WGS84) or a pipeline relative coordinate system. Coordinates of the actual leak The three-dimensional Euclidean distance between them, or the geodesic distance along the pipeline trajectory, is defined as the positioning residual. :

[0133]

[0134] Furthermore, if the on-site feedback is a false alarm (i.e., a leak is predicted, but the pipeline network is confirmed to be operating normally), then logically the location residual of this identified event will be... Force the assignment to infinity (∞) or the preset maximum penalty value.

[0135] S502. Determine whether the localization residual exceeds the preset tolerance threshold. If it does, construct a difficult sample containing the original multimodal time series and perform the data entry process.

[0136] The positioning residual calculated in step S501 Compared with the pre-set engineering tolerance threshold (For example, set to 2.0 meters) for comparison.

[0137] like This indicates that the detection and positioning accuracy of the current model meets the requirements for on-site emergency repair. This event will be archived as a regular log and will not trigger the learning mechanism.

[0138] like (Including the aforementioned false alarm scenarios), this indicates that the nonlinear classification hyperplane or dispersion compensation regression curve of the current model has experienced a degradation in generalization ability under the current pipeline network operating conditions, and the data corresponding to this event is determined to be a difficult example sample. Immediately, the process backtracks to the underlying database, extracts the acoustic emission signal sequence and pressure transient signal sequence corresponding to the time of the deviation, and binds and encapsulates these two sequences with the actual leakage coordinates from the field feedback (as absolutely correct supervised learning labels) in a high-dimensional manner to construct a difficult example sample dataset, which is then stored in a dedicated dynamic difficult example database.

[0139] S503 monitors the sample size of the difficult example database and triggers model retraining when the set threshold is reached, dynamically updating the decision boundaries and regression weights of the underlying network and support vector machine.

[0140] The system monitors the cumulative total number of samples in the dynamic difficult example database in real time. When the cumulative total number of samples reaches a preset model update trigger threshold, the incremental learning process in the background is automatically suspended. Using newly added difficult example samples and combining them with historical core data, the two-stream neural network, support vector machine (SVM), and support vector regression (SVR) models are retrained.

[0141] First, the highly distorted original signals (acoustic emission signal sequence and pressure transient signal sequence) from the difficult example samples are used to fine-tune the convolution kernel parameters of the two-stream neural network, forcing it to extract acoustic depth features and hydraulic transient features that better characterize the current network attenuation. Second, based on the updated feature vectors, the Lagrange multipliers of the support vector machine (SVM) are recalculated. Adjust the maximum classification margin hyperplane that drifts due to changes in environmental noise; finally, use the actual leakage coordinates from the difficult sample. To achieve the regression objective, the kernel weights of the Support Vector Regression (SVR) model are re-optimized. and This allows for precise correction of the nonlinear wave velocity dispersion curve under the current aging degree of the pipe material.

[0142] After retraining is completed, the parameters of the online deployed model are updated, and the database of difficult examples that have participated in training is cleared or archived, thereby completing the self-evolutionary iteration of the entire intelligent algorithm closed loop.

[0143] like Figure 2 As shown, the present invention also provides a multimodal pipeline leak detection and localization device based on a CNN-SVM hybrid architecture. Based on the multimodal pipeline leak detection and localization method based on a CNN-SVM hybrid architecture described above, the device includes:

[0144] Preprocessing module 1 is used to collect acoustic emission signal sequences and pressure transient signal sequences from monitoring nodes in the pipeline network and add timestamps. Based on the timestamps, the acoustic emission signal sequences and pressure transient signal sequences are aligned with time windows and data dimension transformations are performed respectively to obtain acoustic feature matrices and hydraulic time series feature tensors.

[0145] Feature extraction module 2 is used to input the acoustic feature matrix and the hydraulic time series feature tensor into two independent network branches of a pre-constructed two-stream neural network for feature extraction, and to concatenate the extracted acoustic depth features and hydraulic transient features to obtain a fused feature vector;

[0146] The state classification module 3 is used to input the fused feature vector into a pre-constructed support vector machine for state classification. When the state classification result is a leakage state, the fused feature vectors corresponding to the upstream monitoring node and the downstream monitoring node on both sides of the leakage location are obtained respectively, and the time difference of the leakage signal propagating to the upstream monitoring node and the downstream monitoring node is calculated.

[0147] The coordinate prediction module 4 is used to input the time difference, the fusion feature vector corresponding to the upstream monitoring node, and the fusion feature vector corresponding to the downstream monitoring node into a pre-built support vector regression model to perform regression calculation for nonlinear wave velocity dispersion compensation, and output the predicted leakage coordinates.

[0148] The iterative update module 5 is used to determine difficult case samples based on the residual between the predicted leakage coordinates and the actual leakage coordinates, so as to update the support vector machine and support vector regression model based on the difficult case samples; wherein, the difficult case samples include the acoustic emission signal sequence, pressure transient signal sequence and actual leakage coordinates corresponding to the residual.

[0149] Each of the above modules is used to perform the corresponding steps in the multimodal pipeline leak detection and localization method based on the CNN-SVM hybrid architecture. The specific implementation method is as described in the above method embodiment, and will not be repeated here.

[0150] like Figure 3 As shown, the present invention also provides a computer device, which may be a server, and its internal structure may be as follows: Figure 3 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores all the data required for the process of the multimodal pipeline leak detection and localization method based on a CNN-SVM hybrid architecture. The network interface is used for communication with external terminals via a network connection. The computer program is executed by the processor to implement the multimodal pipeline leak detection and localization method based on a CNN-SVM hybrid architecture.

[0151] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer equipment on which the present application is applied.

[0152] An embodiment of this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the above-described multimodal pipeline leak detection and location methods based on a CNN-SVM hybrid architecture.

[0153] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by hardware related to computer program instructions. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the above method embodiments. Any references to memory, storage, databases, or other media provided in this application and used in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM), such as dynamic RAM (used as main storage) or static RAM (commonly used as cache memory). By way of illustration and not limitation, RAM has various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), and Rambus DRAM (RDRAM).

[0154] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0155] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A multimodal pipeline leak detection and localization method based on a CNN-SVM hybrid architecture, characterized in that, include: S1. Collect the acoustic emission signal sequence and pressure transient signal sequence of the monitoring nodes in the pipeline network and add timestamps. Based on the timestamps, align the acoustic emission signal sequence and pressure transient signal sequence with time windows and perform data dimension transformation respectively to obtain the acoustic feature matrix and hydraulic time series feature tensor. S2. Input the acoustic feature matrix and the hydraulic time series feature tensor into two independent network branches of a pre-constructed two-stream neural network for feature extraction. Then, concatenate the extracted acoustic depth features and hydraulic transient features to obtain a fused feature vector. S3. Input the fused feature vector into a pre-constructed support vector machine for state classification. When the state classification result is a leakage state, obtain the fused feature vectors corresponding to the upstream monitoring node and the downstream monitoring node on both sides of the leakage location, and calculate the time difference between the leakage signal propagating to the upstream monitoring node and the downstream monitoring node. S4. Input the time difference, the fusion feature vector corresponding to the upstream monitoring node, and the fusion feature vector corresponding to the downstream monitoring node into the pre-constructed support vector regression model to perform regression calculation for nonlinear wave velocity dispersion compensation, and output the predicted leakage coordinates. S5. Determine difficult case samples based on the residual between the predicted leakage coordinates and the actual leakage coordinates, and update the support vector machine and support vector regression model based on the difficult case samples; wherein, the difficult case samples include the acoustic emission signal sequence, pressure transient signal sequence and actual leakage coordinates corresponding to the residual.

2. The multimodal pipeline leak detection and localization method based on a CNN-SVM hybrid architecture according to claim 1, characterized in that, S1 specifically includes: The acoustic emission signal sequence and the pressure transient signal sequence are acquired synchronously, and an absolute timestamp is injected into the acoustic emission signal sequence and the pressure transient signal sequence using a high-precision time synchronization module; Each sampling point of the pressure transient signal sequence is defined as a discrete reference time anchor point; based on the preset observation time window width, with the absolute timestamp of each reference time anchor point as the center, the acoustic emission signal subsequence falling within the corresponding time window is extracted from the acoustic emission signal sequence to complete the time window alignment. The acoustic emission signal subsequence is processed by continuous wavelet transform to generate a two-dimensional energy spectrum matrix, which is then mapped into an image tensor and normalized to obtain the acoustic feature matrix. The aligned pressure transient signal sequence is normalized and then feature-mapped using a one-dimensional convolution operator to generate a hydraulic time series feature tensor.

3. The multimodal pipeline leak detection and localization method based on a CNN-SVM hybrid architecture according to claim 1, characterized in that, S2 specifically includes: The acoustic feature matrix is ​​input into the two-dimensional convolutional branch of the two-stream neural network, and acoustic depth features are extracted through multi-layer two-dimensional convolution and global average pooling layer operations. The hydraulic time series feature tensor is input into the parallel one-dimensional convolutional branch of the two-stream neural network. High-frequency local mutation points and low-frequency global reflection periods in the hydraulic time series feature tensor are extracted through multiple parallel one-dimensional convolutional layers. After being concatenated in the channel dimension, the hydraulic transient features are output after dimensionality reduction. The acoustic depth features and the hydraulic transient features are mapped to a shared latent space through a linear mapping layer to obtain the query vector and the key vector, respectively. Calculate the outer product between the query vector and the key vector, and obtain the cross-covariance weight matrix after scaling and normalization. Use the cross-covariance weight matrix to extract the global channel weights, and obtain the acoustic attention vector and hydraulic attention vector by dimensionality reduction through a linear projection layer. The acoustic depth feature is weighted element-wise using the acoustic attention vector and the hydraulic transient feature is weighted using the hydraulic attention vector, and then concatenated along the feature dimension to generate a fused feature vector.

4. The multimodal pipeline leak detection and localization method based on a CNN-SVM hybrid architecture according to claim 1, characterized in that, S3 specifically includes: The fused feature vector is input into a pre-constructed support vector machine, spatial mapping is performed through a radial basis kernel function, and the state classification result is output based on the decision hyperplane; When the state classification result is a leak state, the topology association matrix of the pipeline geographic information system is retrieved, and bidirectional backtracking is performed along the physical topology of the pipeline with the current monitoring node as the center. The two nodes closest to the leak location are marked as the upstream monitoring node and the downstream monitoring node, respectively. Data from the upstream and downstream monitoring nodes within the same absolute time window are extracted respectively, and the time window alignment, data dimension transformation, and feature extraction and splicing processes are reused to generate fused feature vectors for the upstream and downstream monitoring nodes respectively. Extract the acoustic emission signal sequences corresponding to the upstream and downstream monitoring nodes, and perform fast Fourier transform on each to obtain the corresponding frequency domain signals; The cross-power spectral density of the frequency domain signal is calculated, and a phase transform weighting function is introduced for whitening. The generalized cross-correlation function is obtained by inverse fast Fourier transform. The time delay independent variable corresponding to the global maximum peak value on the generalized cross-correlation function curve is extracted and used as the time difference for the leakage signal to propagate to the upstream and downstream monitoring nodes.

5. The multimodal pipeline leak detection and localization method based on a CNN-SVM hybrid architecture according to claim 4, characterized in that, The decision classification flow formula for the decision hyperplane is: In the formula, This is a symbolic function used to output discrete classification labels; The total number of support vectors retained by the support vector machine during the training phase; For the sequence index of support vectors, ; The i-th support vector is determined in advance during training; Let Lagrange multipliers be the values ​​corresponding to the i-th support vector; The category label corresponding to the i-th support vector; The bandwidth parameter of the radial basis kernel function controls the nonlinear range of the eigenmap. denoted by L2 norm between feature vectors; b is the bias term of the decision classification hyperplane; The formula for the generalized cross-correlation function is: In the formula, for The complex conjugate; Indicates the modulus of a complex number; For a continuously sliding time delay independent variable, search for a generalized cross-correlation function. Extract the global maximum peak value on the curve and the corresponding horizontal coordinate time delay variable. The time difference between the propagation of the acoustic emission shock wave from the leak to the upstream and downstream monitoring nodes .

6. The multimodal pipeline leak detection and localization method based on a CNN-SVM hybrid architecture according to claim 1, characterized in that, S4 specifically includes: The fusion feature vectors corresponding to the upstream monitoring nodes, the fusion feature vectors corresponding to the downstream monitoring nodes, and the time difference are flattened and stitched together to generate a joint input vector. The joint input vector is input into a pre-constructed support vector regression model. The joint input vector is mapped to a high-dimensional feature space using a regression kernel function. The continuous output value of the support vector regression model is extracted. The continuous output value is used as the relative one-dimensional curve distance of the leakage source from the upstream monitoring node along the physical pipeline of the pipeline network. Retrieve the topological correlation matrix of the pipeline network geographic information system to obtain the three-dimensional spatial absolute coordinates of the upstream monitoring node and the three-dimensional spatial trajectory equation of the pipeline network connecting the upstream monitoring node and the downstream monitoring node; Starting from the absolute coordinates in the three-dimensional space, a curve segment with a length equal to the relative one-dimensional curve distance is intercepted along the three-dimensional trajectory equation of the pipeline space towards the downstream monitoring node; The endpoint of the curve segment is calculated using three-dimensional coordinates to obtain a spatial location vector containing longitude, latitude, and elevation information. This vector is then used as the predicted leakage coordinates and output.

7. The multimodal pipeline leak detection and localization method based on a CNN-SVM hybrid architecture according to claim 6, characterized in that, The support vector regression model uses a nonlinear wave speed dispersion compensation regression function to output continuous values. : In the formula, is the total number of support vectors in the support vector regression model; j is the index of the support vector. This refers to the j-th regression support vector determined during the training phase. and Let Lagrange multipliers be the values ​​corresponding to the j-th support vector. Here is the regression kernel function for SVR; This is the bias term of the regression model.

8. The multimodal pipeline leak detection and localization method based on a CNN-SVM hybrid architecture according to claim 1, characterized in that, S5 specifically includes: Obtain the actual leakage coordinates from the field feedback, and calculate the spatial distance residual between the predicted leakage coordinates and the actual leakage coordinates as the positioning residual; if the field feedback result is a false alarm, assign the positioning residual a preset maximum penalty value; Determine whether the positioning residual exceeds a preset tolerance threshold. If the positioning residual exceeds the tolerance threshold, extract the corresponding acoustic emission signal sequence and pressure transient signal sequence, and bind and encapsulate the acoustic emission signal sequence and pressure transient signal sequence with the actual leakage coordinates to construct a difficult case sample and store it in the difficult case database. The database of difficult examples is monitored. When the total number of accumulated samples reaches the preset model update trigger threshold, the acoustic emission signal sequence and pressure transient signal sequence in the difficult example samples are used to fine-tune the convolution kernel parameters of the two-stream neural network in reverse. Based on the updated feature vector after parameter fine-tuning, the Lagrange multipliers of the support vector machine are recalculated to adjust the classification hyperplane; Using the actual leakage coordinates in the difficult sample as the regression target, the kernel function weights and Lagrange multipliers of the support vector regression model are re-optimized.

9. A multimodal pipeline leak detection and localization device based on a CNN-SVM hybrid architecture, based on any one of claims 1 to 8, characterized in that, The device includes: The preprocessing module is used to collect the acoustic emission signal sequence and pressure transient signal sequence of the monitoring nodes in the pipeline network and add timestamps. Based on the timestamps, the acoustic emission signal sequence and pressure transient signal sequence are aligned with time windows and the data dimensions are transformed respectively to obtain the acoustic feature matrix and hydraulic time series feature tensor. The feature extraction module is used to input the acoustic feature matrix and the hydraulic time series feature tensor into two independent network branches of a pre-constructed two-stream neural network for feature extraction, and to concatenate the extracted acoustic depth features and hydraulic transient features to obtain a fused feature vector. The state classification module is used to input the fused feature vector into a pre-constructed support vector machine for state classification. When the state classification result is a leakage state, the module obtains the fused feature vectors corresponding to the upstream monitoring nodes and downstream monitoring nodes on both sides of the leakage location, and calculates the time difference between the leakage signal propagating to the upstream monitoring nodes and downstream monitoring nodes. The coordinate prediction module is used to input the time difference, the fused feature vector corresponding to the upstream monitoring node, and the fused feature vector corresponding to the downstream monitoring node into a pre-constructed support vector regression model to perform regression calculations for nonlinear wave velocity dispersion compensation, and output the predicted leakage coordinates. An iterative update module is used to determine difficult case samples based on the residual between the predicted leakage coordinates and the actual leakage coordinates, so as to update the support vector machine and support vector regression model based on the difficult case samples; wherein, the difficult case samples include the acoustic emission signal sequence, pressure transient signal sequence and actual leakage coordinates corresponding to the residual.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

Images