An underground pipeline impact event identification method, device, equipment and medium

Abstract

The application discloses a kind of underground pipeline impact event identification method, device, equipment and medium, it is related to underground pipeline safety monitoring technical field, comprising: initial piezoelectric signal of target underground pipeline is collected using piezoelectric sensing array;The initial piezoelectric signal is preprocessed by preset composite algorithm, to generate piezoelectric signal after processing;Using multidimensional feature extraction algorithm, multidimensional feature extraction and fusion operation are carried out to the piezoelectric signal after processing, corresponding fusion feature is obtained;Impact event of the target underground pipeline is identified and classified based on the fusion feature by preset identification model;If the identification result represents that the impact event exists, corresponding target impact event report is generated, and the impact point of the impact event is positioned using preset positioning algorithm, to complete the identification of target underground pipeline impact event.

Description

Technical Field

[0001] This invention relates to the field of underground pipeline safety monitoring technology, and in particular to a method, device, equipment and medium for identifying impact events in underground pipelines. Background Technology

[0002] With the increasing complexity of urban underground pipeline systems, accidental impacts on underground gas pipelines caused by third-party construction and geological activities have become a major safety threat. Traditional monitoring methods mostly use distributed optical fibers or vibration sensors, but these methods have drawbacks such as low sensitivity, poor positioning accuracy, and inability to quantify impact energy.

[0003] In conclusion, improving the safety early warning capabilities of underground gas pipelines is an urgent problem to be solved. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a method, apparatus, equipment, and medium for identifying impact events in underground pipelines, which can improve the safety early warning capability of underground gas pipelines. The specific solution is as follows: Firstly, this application provides a method for identifying impact events in underground pipelines, including: The initial piezoelectric signal of the target underground pipeline is acquired using a piezoelectric sensor array; the piezoelectric sensor array includes piezoelectric sensors arranged at preset positions on the surface of the target underground pipeline. The initial piezoelectric signal is preprocessed using a preset composite algorithm to generate a processed piezoelectric signal; wherein the preset composite algorithm is an algorithm constructed based on adaptive bionic filtering technology, joint denoising algorithm and time-varying blind source separation algorithm; A multidimensional feature extraction algorithm is used to perform multidimensional feature extraction and fusion operations on the processed piezoelectric signal to obtain the corresponding fused features. Based on the fused features, the impact events of the target underground pipeline are identified and classified using a preset recognition model; If the identification result indicates the existence of the impact event, a corresponding target impact event report is generated, and the impact point of the impact event is located using a preset positioning algorithm to complete the identification of the target underground pipeline impact event.

[0005] Optionally, the step of preprocessing the initial piezoelectric signal using a preset composite algorithm to generate a processed piezoelectric signal includes: A preset filter bank is constructed, and a psychoacoustic masking model is introduced; wherein, the preset filter bank is a set of filters that mimic the cochlear basilar membrane; The masking threshold curve is determined by the psychoacoustic masking model in the preset filter group; Based on the instantaneous power spectrum of the initial piezoelectric signal, the corresponding masking threshold in the masking threshold curve is determined, and the background noise in the initial piezoelectric signal that is less than the masking threshold is suppressed to obtain the biomimetic piezoelectric signal; Multi-scale wavelet coefficients are generated by performing multi-level wavelet decomposition on the biomimetic piezoelectric signal. The multi-scale wavelet coefficients are divided into blocks to obtain the block-divided wavelet coefficients. An overcomplete dictionary is determined using a pre-defined dictionary learning algorithm to adapt to the characteristics of pipeline impact. The block-based wavelet coefficients are sparsely reconstructed on the overcomplete dictionary using the orthogonal matching pursuit algorithm to obtain the reconstructed coefficients. Based on the local statistical characteristics of the multi-scale wavelet coefficients, a preset adaptive threshold is determined, and a preset coefficient shrinkage operation is performed on the reconstructed coefficients according to the preset adaptive threshold to obtain the denoised piezoelectric signal. The denoised piezoelectric signal is divided into preset sliding time windows, and blind source separation is performed within each preset sliding time window using a preset JADE algorithm to obtain the separated piezoelectric signal. The components of the separated piezoelectric signals in adjacent windows are tracked, and the changes in the time-varying mixing matrix in the separated piezoelectric signals are estimated using a Kalman filter to extract the source signal related to the impact event as the processed piezoelectric signal.

[0006] Optionally, the step of using a multidimensional feature extraction algorithm to perform multidimensional feature extraction and fusion operations on the processed piezoelectric signal to obtain the corresponding fused features includes: Multi-scale morphological features of the processed piezoelectric signal are extracted based on a preset set of multi-scale structural elements; the multi-scale morphological features are the features of the local geometric structure of the impact waveform. The processed piezoelectric signal is input into a preset depth dual-stream convolutional neural network to generate deep features containing the joint time-frequency distribution characteristics of the piezoelectric signal; The frequency and wavenumber spectrum are determined based on the processed piezoelectric signal, and the dispersion curve of the target underground pipeline is extracted from the frequency and wavenumber spectrum through peak detection operation. Based on the extracted dispersion curve, the equivalent physical parameters of the preset pipeline and soil combined system are inverted to determine the array propagation characteristics of the processed piezoelectric signal; the array propagation characteristics are those that reflect the propagation characteristics of shock waves. By using a preset feature weight allocation network, the fusion weights of the multi-scale morphological features, the depth features, and the array propagation features are determined respectively. The multi-scale morphological features, the depth features, and the array propagation features are weighted and fused according to the fusion weights to generate fused features.

[0007] Optionally, the step of extracting the multi-scale morphological features of the processed piezoelectric signal based on a preset multi-scale structuring element set includes: Based on the statistical characteristics of the processed piezoelectric signal and the physical parameters of the target underground pipeline, the scale sequence of the corresponding structural elements is determined; wherein, the scale sequence of the structural elements includes linear structural elements, circular structural elements and rectangular structural elements, the statistical characteristics include the standard deviation, kurtosis and instantaneous amplitude distribution of the signal, and the physical parameters include the elastic modulus of the pipe, the wall thickness and the estimated stress wave propagation velocity. Morphological gradient calculations are performed using the processed piezoelectric signal of the linear structural element to obtain impact signal characteristics; The processed piezoelectric signal of the circular structural element is used to perform a top cap operation to obtain waveform contour features; Iterative expansion calculations are performed using the processed piezoelectric signal of the rectangular structural element to extract the local maximum region of the target signal; Based on the impact signal characteristics, the waveform contour characteristics, and the local maxima region of the target signal, multi-scale morphological features are constructed.

[0008] Optionally, the step of inputting the processed piezoelectric signal into a preset depth dual-stream convolutional neural network to generate deep features containing the joint time-frequency distribution characteristics of the piezoelectric signal includes: The temporal flow features of the processed piezoelectric signal are extracted using the temporal flow branch of a pre-defined deep dual-stream convolutional neural network. Frequency domain flow features in the processed piezoelectric signal are extracted using the frequency domain flow branch of a deep two-stream convolutional neural network. The attention weights of the temporal flow features and the frequency flow features are determined by the bidirectional attention fusion layer of the preset deep dual-stream convolutional neural network. The attention weights are used to perform weighted fusion of the temporal flow features and the frequency flow features to obtain the deep features.

[0009] Optionally, the step of identifying and classifying the impact events of the target underground pipeline based on the fused features using a preset recognition model includes: Pre-train the basic classification model to obtain the trained classification model; The trained classification model is transferred to the target domain using a preset hierarchical domain adaptation strategy to obtain a transferred classification model, which is then used as a preset recognition model; the target domain is the domain for classifying impact events on underground pipelines. Based on the fused features, the impact events of the target underground pipeline are identified and classified using a preset recognition model.

[0010] Optionally, if the identification result indicates the existence of the impact event, then generating a corresponding target impact event report includes: If the identification result indicates the existence of the impact event, the classification result of the impact event is obtained, a corresponding target impact event report is generated based on the classification result, and a preset uncertainty quantization operation is performed on the classification result based on the Bayesian deep learning framework to obtain the corresponding quantization score, so that the user terminal can manually review the target impact event report corresponding to the classification result with the quantization score less than the preset confidence threshold.

[0011] Secondly, this application provides an impact event identification device for underground pipelines, comprising: A signal acquisition module is used to acquire the initial piezoelectric signal of the target underground pipeline using a piezoelectric sensor array; the piezoelectric sensor array includes various piezoelectric sensors arranged on the surface of the target underground pipeline according to preset positions; The signal generation module is used to preprocess the initial piezoelectric signal using a preset composite algorithm to generate a processed piezoelectric signal; wherein the preset composite algorithm is an algorithm constructed based on adaptive bionic filtering technology, joint denoising algorithm and time-varying blind source separation algorithm; The feature fusion module is used to perform multi-dimensional feature extraction and fusion operations on the processed piezoelectric signal using a multi-dimensional feature extraction algorithm to obtain the corresponding fused features. The event classification module is used to identify and classify the impact events of the target underground pipeline based on the fused features and a preset recognition model. The impact point localization module is used to generate a corresponding target impact event report if the identification result indicates the existence of the impact event, and to locate the impact point of the impact event using a preset localization algorithm, so as to complete the identification of the target underground pipeline impact event.

[0012] Thirdly, this application provides an electronic device, comprising: Memory, used to store computer programs; A processor is used to execute the computer program to implement the aforementioned method for identifying impact events in underground pipelines.

[0013] Fourthly, this application provides a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned method for identifying impact events in underground pipelines.

[0014] In summary, this application first utilizes a piezoelectric sensor array to acquire initial piezoelectric signals from a target underground pipeline. The piezoelectric sensor array includes piezoelectric sensors arranged at preset positions on the surface of the target underground pipeline. The initial piezoelectric signals are preprocessed using a preset composite algorithm to generate processed piezoelectric signals. The preset composite algorithm is based on an adaptive biomimetic filtering algorithm, a joint denoising algorithm, and time-varying blind source separation. A multidimensional feature extraction algorithm is used to extract and fuse multidimensional features from the processed piezoelectric signals to obtain corresponding fused features. Based on the fused features, a preset recognition model is used to identify and classify impact events in the target underground pipeline. If the identification result indicates the existence of an impact event, a corresponding target impact event report is generated, and a preset positioning algorithm is used to locate the impact point of the impact event, thereby completing the identification of impact events in the target underground pipeline. As can be seen, this application first deploys a piezoelectric sensor array at preset positions on the pipeline surface to acquire initial piezoelectric signals. Next, the signal is preprocessed using a composite algorithm that integrates adaptive bionic filtering, wavelet-dictionary learning for joint denoising, and time-varying blind source separation to obtain the filtered signal. Subsequently, a multi-dimensional feature extraction algorithm is used to extract and fuse features from the filtered signal, forming fused features. Based on these features, a pre-defined recognition model is used to identify and classify pipeline impact events. If an impact event is identified, an event hypothesis report is generated, and a pre-defined positioning algorithm is used to determine the impact point location, thereby achieving complete monitoring of pipeline impact loads. In this way, high-density, adaptive vibration sensing of the gas pipeline surface is achieved using a piezoelectric array. The composite preprocessing algorithm, consisting of bionic filtering, joint denoising, and blind source separation, significantly improves the signal-to-noise ratio of the impact signal in complex noise environments. Based on multi-dimensional feature extraction and fusion technology, a fused feature vector capable of comprehensively characterizing the spatiotemporal features of impact events is constructed. Furthermore, a recognition model based on a transfer learning framework is used to achieve high-precision impact classification, and sub-meter-level accuracy impact positioning is achieved by combining intelligent wave velocity correction and compressed sensing theory. Ultimately, a complete monitoring system will be formed, encompassing signal perception, feature extraction, event recognition, and precise positioning, enabling intelligent and highly reliable monitoring and early warning of impact loads on underground gas pipelines. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0016] Figure 1 This application discloses a flowchart of a method for identifying impact events in underground pipelines. Figure 2 This is a schematic diagram of the structure of an impact event identification device for underground pipelines disclosed in this application; Figure 3 This is a structural diagram of an electronic device disclosed in this application. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Currently, with the increasing complexity of urban underground pipeline systems, accidental impacts on underground gas pipelines caused by third-party construction and geological activities have become a major safety threat. Traditional monitoring methods mostly employ distributed optical fibers or vibration sensors, but these suffer from drawbacks such as low sensitivity, poor positioning accuracy, and inability to quantify impact energy. To address these technical problems, this application discloses a method, device, equipment, and medium for identifying impact events in underground pipelines, which can improve the safety early warning capabilities of underground gas pipelines.

[0019] See Figure 1 As shown in the figure, an embodiment of the present invention discloses a method for identifying impact events in underground pipelines, including: Step S11: Acquire the initial piezoelectric signal of the target underground pipeline using a piezoelectric sensor array; the piezoelectric sensor array includes various piezoelectric sensors arranged on the surface of the target underground pipeline according to preset positions.

[0020] In this embodiment, a PZT (piezoelectric ceramics) sheet is used as a high-sensitivity core sensing unit, embedded or laminated into a broadband sensing substrate of a flexible PVDF (Polyvinylidene Fluoride) piezoelectric thin film sensor, to obtain a rigid-flexible composite piezoelectric sensing unit. Based on the risk assessment results along the pipeline, a non-uniform pitch is generated, and the piezoelectric sensing units are arranged spirally along the pipeline axis to form a piezoelectric sensing array. Specifically, multi-source information such as pipeline burial depth, geological conditions, historical accident data, frequency of overhead activity, and pipeline criticality is collected and input into a risk quantification model to calculate a continuous risk distribution curve along the pipeline mileage; based on this curve, the pipeline is divided into high-risk, medium-risk, and low-risk sections. Different spiral arrangement parameters are adopted for different risk sections: in high-risk sections, a tight spiral arrangement is used with a pitch of three to five meters and a spiral angle of forty-five degrees to improve spatial sampling density and impact positioning accuracy; in medium-risk sections, a medium-pitch arrangement is used with a pitch of eight to twelve meters and a spiral angle of thirty degrees to balance monitoring performance and deployment cost; in low-risk sections, a loose pitch arrangement is used with a pitch of fifteen to twenty meters and a spiral angle of fifteen degrees to achieve wide-area coverage. Based on the above-generated parameter scheme, a spiral array is deployed along the pipeline axis. During deployment, the pipeline outer wall treatment technology is first used to ensure that the bonding area is clean and flat. Then, an adaptive bonding device is used to install the composite piezoelectric sensing unit point by point along a preset spiral trajectory. In this way, the bonding pressure and angle can be automatically adjusted according to the pipeline diameter and curvature to ensure that a uniform and firm acoustic coupling is formed between the sensing unit and the pipe wall. The output of each sensing unit is connected to the local acquisition module via corrosion-resistant, tensile-resistant flexible cables or wireless nodes, ultimately forming a three-dimensional piezoelectric sensor array network that covers the entire pipeline, has a non-uniform spatial resolution distribution, and combines high sensitivity with wide bandwidth response. By combining rigid-flexible composite design at the material level with intelligent risk-driven layout at the system level, and utilizing the synergistic optimization of microstructure and macro topology, optimal monitoring efficiency for pipeline impact loads is achieved with limited sensing resources. This ensures both precise perception of high-risk areas and economical and effective coverage of the entire pipeline.

[0021] Step S12: The initial piezoelectric signal is preprocessed using a preset composite algorithm to generate a processed piezoelectric signal; wherein the preset composite algorithm is an algorithm constructed based on adaptive bionic filtering technology, joint denoising algorithm and time-varying blind source separation algorithm.

[0022] In this embodiment, a composite algorithm combining adaptive bionic filtering, wavelet-dictionary learning for joint denoising, and time-varying blind source separation is used to preprocess the piezoelectric signal. First, a preset filter group is constructed, and a psychoacoustic masking model is introduced. The preset filter group is a set of filters mimicking the cochlear basilar membrane. A masking threshold curve is determined using the psychoacoustic masking model in the preset filter group. Based on the instantaneous power spectrum of the initial piezoelectric signal, the corresponding masking threshold in the masking threshold curve is determined to suppress background noise in the initial piezoelectric signal that is less than the masking threshold, thus obtaining a bionically processed piezoelectric signal. The bionically processed piezoelectric signal is then subjected to multi-level wavelet decomposition to generate multi-scale... Multi-scale wavelet coefficients are divided into blocks to obtain block-wise wavelet coefficients. An overcomplete dictionary is determined using a preset dictionary learning algorithm to adapt to the characteristics of pipeline impact. The block-wise wavelet coefficients are sparsely reconstructed on the overcomplete dictionary using an orthogonal matching pursuit algorithm to obtain reconstructed coefficients. A preset adaptive threshold is determined based on the local statistical characteristics of the multi-scale wavelet coefficients. A preset coefficient shrinkage operation is performed on the reconstructed coefficients according to the preset adaptive threshold to obtain a denoised piezoelectric signal. The denoised piezoelectric signal is divided into preset sliding time windows, and blind source separation is performed within each preset sliding time window using a preset JADE (Joint Approximate Diagonalization of Eigenmatrices) algorithm to obtain a separated piezoelectric signal. Component tracking is performed on the separated piezoelectric signals in adjacent windows, and the change of the time-varying mixing matrix in the separated piezoelectric signal is estimated using a Kalman filter to extract the source signal related to the impact event as the processed piezoelectric signal. Specifically, simulating the frequency analysis mechanism of the basilar membrane of the cochlea in the human auditory system, a set of bandpass filters with non-uniformly distributed center frequencies and dynamically changing bandwidths was designed and implemented. The filter bank was constructed based on an equivalent rectangular bandwidth model, using a smaller number of filters but a wider bandwidth in the low-frequency region to efficiently capture the energy-concentrated low-frequency components in impulse signals. As the frequency increases, the number of filters gradually increases, while the bandwidth correspondingly narrows to improve the frequency resolution of high-frequency details. The filter bank was implemented using an infinite impulse response or higher-order finite impulse response structure. Its center frequency and bandwidth parameters were pre-calculated according to the target analysis frequency band and fixed in the signal processing unit, forming the basis for multi-resolution analysis of signals across the entire frequency band. Based on the filter bank analysis, a psychoacoustic masking model was introduced to intelligently distinguish and suppress maskable background noise. This process analyzed the instantaneous signals output by each filter channel in real time and calculated their short-time power spectra. For each moment, based on the frequency and intensity of the current dominant signal component, and according to the critical band masking characteristics in psychoacoustics, the masking threshold curve near each frequency point was dynamically calculated.The masking threshold curve describes the sound components below a certain threshold that will be ignored by the human ear in the current auditory context. Applying this auditory model to vibration signal processing, it is assumed that background noise vibration components below this masking threshold contribute negligibly to the detection and identification of impact events and should be suppressed.

[0023] Subsequently, adaptive noise suppression based on a masking threshold is performed. The power spectrum of the signal output from each filter channel is compared with the real-time masking threshold at the corresponding frequency. Frequency components with power spectrum values ​​below the masking threshold are identified as background noise that can be effectively masked, and their energy is significantly attenuated or zeroed through frequency domain gain adjustment or time domain filtering techniques; while components with power spectrum values ​​above the masking threshold are considered valid signals and are preserved or enhanced. This process is performed in parallel across all channels of the entire filter bank, and finally, the signals processed by all channels are resynthesized to output a biomimetic piezoelectric signal.

[0024] Next, the biomimetic-processed piezoelectric signal undergoes multi-level wavelet decomposition to obtain multi-scale wavelet coefficients. Simultaneously, an overcomplete dictionary adapted to the pipeline impact characteristics is trained using the K-SVD (K-Singular Value Decomposition) dictionary learning algorithm. An orthogonal matching pursuit algorithm is then employed to sparsely reconstruct the noisy wavelet coefficients on the overcomplete dictionary. An adaptive threshold is set based on the local statistical characteristics of the multi-scale wavelet coefficients to shrink the coefficients, resulting in a denoised multi-channel array piezoelectric signal. Specifically, wavelet basis functions with tight support and good regularity are selected to decompose the signal layer by layer to a predetermined depth, obtaining approximate coefficients and detail coefficients at each layer. This process reveals the multi-resolution representation of the signal at different time scales and frequency resolutions. The approximate coefficients from higher-level decompositions characterize the macroscopic trend and low-frequency main body of the signal, while the detail coefficients at each layer capture the local abrupt changes and transient characteristics of the signal at the corresponding scale, particularly the short-term energy burst mode exhibited by the impact event.

[0025] Then, a training dataset is constructed using a large number of piezoelectric signal samples from typical pipeline impact events. The K-SVD dictionary learning algorithm can be used to train this dataset. This algorithm progressively optimizes an overcomplete dictionary through alternating iterative atomic updates and sparse coding. In the atomic update stage, each atom in the dictionary is updated column-by-column to more sparsely represent the current error signal. In the sparse coding stage, a greedy pursuit algorithm is used to find the sparsest representation of each training sample under the current dictionary. After sufficient iteration, an overcomplete dictionary specifically adapted to the vibration morphology of pipeline impact is finally obtained, whose atoms can effectively match various potential waveform structures of the impact signal. The coefficients at each scale obtained from multi-level wavelet decomposition are considered as signals to be processed, and the orthogonal matching pursuit algorithm is used to find the optimal sparse linear combination of these coefficients on the overcomplete dictionary. This algorithm iteratively selects the dictionary atom most relevant to the current residual and uses the selected atom set to orthogonally project the signal to update the residual until a preset sparsity or residual condition is met. This process effectively separates and reconstructs the impact signal components highly correlated with dictionary atoms from noisy wavelet coefficients, while suppressing mismatched random noise. Considering the differences in statistical characteristics of wavelet coefficients at different scales and spatial locations, an adaptive thresholding method based on the statistical information of neighboring coefficients is adopted. For each wavelet coefficient, the statistics of the coefficients within its local neighborhood window, such as the standard deviation or the median absolute value, are calculated, and the shrinkage threshold of the coefficient is dynamically set accordingly. Then, a soft thresholding function is applied to shrink all wavelet coefficients, significantly attenuating or zeroing coefficients with absolute values ​​below the threshold, while retaining or slightly attenuating coefficients above the threshold, resulting in a denoised multi-channel array piezoelectric signal with a significantly improved signal-to-noise ratio, i.e., the denoised piezoelectric signal.

[0026] Finally, the denoised multi-channel array piezoelectric signal is divided into preset sliding time windows. Within each window, the JADE algorithm based on fourth-order cumulants is used for blind source separation. The multi-channel signals within each time window constitute an observation data matrix, which is composed of sampling points from all sensors in the array within the same time period. Considering the transient characteristics of impact events and the non-stationarity of noise, a certain proportion of overlap is usually set between sliding windows to ensure signal continuity and event integrity, and to avoid the impact event being interrupted at the window boundaries. The JADE algorithm calculates the fourth-order cumulant matrix of the observed signal and uses joint diagonalization to find an optimal separation matrix. This process does not rely on the Gaussianity assumption of the signal, but utilizes the higher-order statistical properties of the signal to effectively separate statistically independent source signal components. In the pipeline monitoring scenario, the separated independent components correspond to different types of vibration sources, such as real pipeline impact events, continuous environmental background vibration, equipment operating noise, or sensor inherent noise. To address the possibility that the impact source and propagation path may change slowly over time, a component tracking and time-varying hybrid matrix estimation mechanism is introduced. Cross-window component correspondences can be established by calculating the eigenvector correlations between the independent components separated in the current time window and the components in the previous time window. This correlation analysis is typically based on eigenvectors such as the time-domain waveform, spectral characteristics, or statistical moments of the components. By finding the component pairs with the highest correlation, the trajectory tracking of components from the same physical source across different windows can be achieved. Based on this, a Kalman filter is used to estimate and update the implicit time-varying mixing matrix online. The state change of the mixing matrix is ​​modeled as a dynamic process, and the Kalman filter recursively updates the estimate of the mixing matrix using the signal observed in the current window and the predicted source signal. This process can smoothly track the slow drift or abrupt changes in the mixing relationship over time, such as changes in transmission characteristics caused by changes in soil moisture, pipeline stress state, or sensor performance drift. Based on the tracking results and the time-varying mixing matrix estimation, the independent source signal most relevant to the impact event is identified and extracted from the separated components of each window. This is typically determined by analyzing the time-domain characteristics of components such as impulsivity and transientity, the frequency-domain characteristics of matching with known impact spectra, and the degree of conformity with prior knowledge of pipeline impact. Components identified as impact events will be retained and may be reconstructed back into a sensor domain representation through inverse transformation, or directly used as signals for subsequent processing, i.e., processed piezoelectric signals. Other source components identified as noise or interference will be suppressed or discarded.

[0027] Step S13: Using a multidimensional feature extraction algorithm, perform multidimensional feature extraction and fusion operations on the processed piezoelectric signal to obtain the corresponding fused features.

[0028] In this embodiment, multi-scale morphological features of the processed piezoelectric signal are extracted based on a preset multi-scale structuring element set. These multi-scale morphological features are characteristics of the local geometric structure of the impact waveform. The processed piezoelectric signal is input into a preset deep dual-stream convolutional neural network to generate depth features containing the time-frequency joint distribution characteristics of the piezoelectric signal. The frequency and wavenumber spectra are determined based on the processed piezoelectric signal, and the dispersion curve of the target underground pipeline is extracted from the frequency and wavenumber spectra through peak detection. Based on the extracted dispersion curve, the equivalent physical parameters of the preset pipeline-soil combined system are inverted to determine the array propagation characteristics of the processed piezoelectric signal. These array propagation characteristics reflect the propagation characteristics of the shock wave. A preset feature weighting network is used to determine the fusion weights of the multi-scale morphological features, the depth features, and the array propagation features. The multi-scale morphological features, the depth features, and the array propagation features are weighted and fused according to the fusion weights to generate fused features. Specifically, a multi-scale structuring element set is first used to perform mathematical morphological operations on the processed piezoelectric signal to extract multi-scale morphological features reflecting the local geometric structure of the impact waveform. The processed piezoelectric signal is then input into a deep two-stream convolutional neural network. A bidirectional attention fusion layer adaptively weights and fuses the feature outputs of the two streams to obtain deep features containing the joint time-frequency distribution characteristics of the piezoelectric signal. Next, based on the processed piezoelectric signal, the frequency-wavenumber spectrum is calculated, and the dispersion curve of the pipeline is extracted using peak detection. A spectral clustering algorithm is used to separate different propagation modes. Based on the extracted dispersion curve, the equivalent physical parameters of the pipeline-soil system are inverted to obtain array propagation features reflecting the propagation characteristics of the shock wave. Finally, a meta-learning-based feature weight allocation network is used to dynamically generate fusion weights for multi-scale morphological features, deep features, and array propagation features. Feature-level adaptive weighted fusion is then performed to obtain the fused features.

[0029] Understandably, to obtain multi-scale morphological features, it is necessary to determine the scale sequence of corresponding structural elements based on the statistical characteristics of the processed piezoelectric signal and the physical parameters of the target underground pipeline. The scale sequence of structural elements includes linear, circular, and rectangular structural elements. The statistical characteristics include the standard deviation, kurtosis, and instantaneous amplitude distribution of the signal. The physical parameters include the elastic modulus, wall thickness, and estimated stress wave propagation velocity of the pipe material. Morphological gradient calculations are performed using the processed piezoelectric signal of the linear structural elements to obtain impact signal features. Top-hat calculations are performed using the processed piezoelectric signal of the circular structural elements to obtain waveform contour features. Iterative dilation calculations are performed using the processed piezoelectric signal of the rectangular structural elements to extract local maxima regions of the target signal. Based on the impact signal features, the waveform contour features, and the local maxima regions of the target signal, multi-scale morphological features are constructed. Specifically, statistical properties include the signal's standard deviation, kurtosis, and instantaneous amplitude distribution, used to estimate the severity of local signal changes; pipe physical parameters include the pipe's elastic modulus, wall thickness, and estimated stress wave propagation velocity, which affect the broadening and morphology of the impact vibration in the time domain waveform. By integrating signal statistics and physical parameters into a scale mapping model, a non-uniformly distributed sequence of scale values ​​is generated, ensuring that the selected structural element size matches the signal structure at different time scales, from rapid transients to slow evolutions. A linear structural element, aligned with the time axis, is used to perform morphological gradient operations on the signal. This operation calculates the difference between the signal after expansion and erosion by the element, yielding the gradient signal. The gradient signal highlights points where the signal amplitude changes rapidly, with its peak directly corresponding to the steep rising edge at the start and the falling edge at the end of the impact event, thus extracting gradient features characterizing the timing and sharpness of the impact pulse. A circular structural element is used to perform a top-hat transform on the signal. This operation subtracts the opening result from the original signal. Due to the symmetry of the circular structure, the opening operation smooths out details and glitches smaller than the size of the circular "probe" in the signal, while the top-hat transform preserves these smoothed-out components, especially those waveform contours that appear as local bulges or depressions. These contour features can reflect the envelope shape and decay mode of the impact oscillation. Morphological reconstruction using rectangular structuring elements is used to extract the local maxima regions of the signal. This method starts from a labeled signal and, within the constrained signal range, iteratively performs dilation and intersection operations to gradually grow and ultimately reconstruct the local maxima connected regions of the signal. This process can accurately identify the peak position of each impact pulse in the signal and its time range of influence, while suppressing spurious extrema caused by noise. The responses of the above three operations on multi-scale sequences are systematically organized.For each preset scale, the gradient response energy generated by linear elements, the top-hat transformation energy generated by circular elements, and the coverage or intensity of the maximum region extracted by rectangular elements are calculated and recorded to form a three-dimensional data unit. These data units from all scales are arranged in scale order to form a scale-morphological response matrix. The rows of this matrix correspond to different analysis scales, and the columns correspond to different morphological operation types. The matrix element values ​​quantify the characteristic intensity of the signal at that scale through specific morphological operations. The scale-morphological response matrix is ​​then characterized. By calculating the statistics of the matrix's row and column vectors, the overall and local patterns of the matrix are extracted, such as analyzing the variation of feature intensity with scale and the correlation between the results of different morphological operations. These statistics and patterns derived from the matrix constitute the final multi-scale morphological feature vector used to describe the signal morphology.

[0030] It is important to understand that, in order to obtain deep features, the temporal flow branch of a pre-defined deep dual-stream convolutional neural network is used to extract temporal flow features from the processed piezoelectric signal; the frequency flow branch of the same network is used to extract frequency flow features; and the bidirectional attention fusion layer of the pre-defined deep dual-stream convolutional neural network determines the attention weights for the temporal flow features and the frequency flow features, respectively. These attention weights are then used to perform a weighted fusion of the temporal flow features and the frequency flow features to obtain the deep features. Specifically, the temporal flow branch of the deep dual-stream convolutional neural network is used to extract temporal flow features from the filtered piezoelectric signal; specifically, in the improved structure of the temporal flow branch, a one-dimensional residual network with porous convolution and an adaptive receptive field is employed. The network input is a pre-processed piezoelectric signal time series. The first layer uses a large-width convolutional kernel for initial feature mapping to capture long-range correlations in the signal. The subsequent core module consists of multiple stacked "porous residual blocks". Within each residual block, three dilated convolutional paths with different dilation rates are set in parallel, for example, dilation rates of 1, 2, and 4. This allows the network to simultaneously perceive local details, mid-range dependencies, and long-range trends of the signal within a single layer without increasing network depth or the number of parameters. The output of each path is dynamically weighted and fused through a learnable gated attention mechanism. The gating weights are calculated from the local statistical characteristics of the signal segment, ensuring that the network adaptively focuses on the most relevant scale features of the current signal segment. A deformable convolutional layer is also integrated after each residual block, whose kernel sampling position can be slightly adaptively shifted according to the shape of the input signal, thereby more accurately aligning with the abrupt change points of non-stationary impact signals. The final output of this branch is a high-dimensional time-domain feature tensor after multi-layer nonlinear transformation and feature compression, encoding the fine structure of the signal in the time dimension, the temporal relationship of events, and the dynamic evolution pattern. In the improved structure of the frequency domain flow branch, a time-frequency joint analysis architecture based on complex neural networks is designed. The original time-domain signal is first fed into a time-frequency transform layer with learnable parameters. This layer is not a fixed short-time Fourier transform, but rather consists of a set of trainable one-dimensional convolutional kernels. It projects the time-domain signal into an optimal time-frequency representation space in a data-driven manner, and its output is a complex time-frequency spectrum. Subsequently, this complex time-frequency spectrum is input into a complex convolutional neural network for processing. The convolutional layers, batch normalization layers, and activation functions in the network all support complex number operations, enabling simultaneous modeling of the amplitude and phase information of the time-frequency spectrum. This is crucial for capturing the frequency domain energy distribution and phase coherence of impact signals.The network also incorporates an asymmetric pooling layer along the frequency axis. During pooling operations, a larger pooling kernel is used along the frequency dimension to significantly compress frequency resolution, while a smaller pooling kernel is used along the time dimension to preserve the original resolution. This design aims to reduce redundancy in the frequency dimension while maintaining the accuracy of impact event temporal localization, allowing the network to focus more on key frequency bands related to the impact. Furthermore, a frequency band importance attention module is embedded in the branch. This module automatically calculates the importance weights of different frequency sub-bands for the current classification task and applies channel weighting to the feature map accordingly, suppressing irrelevant noise bands and enhancing the feature response of key frequency bands. The final output of this branch is a deeply encoded frequency domain feature tensor, characterizing the energy distribution pattern, frequency band coupling relationships, and transient spectral changes of the signal in the time-frequency plane. A bidirectional attention fusion layer is used to weightedly fuse time-domain and frequency-domain features. This bidirectional attention fusion layer employs an innovative cross-modal gating attention and feature recombination mechanism. First, attention weights are generated for both temporal and frequency-domain features. The frequency-domain feature can be used as the query, with the temporal-domain feature as the key and value, to calculate the attention weight of the frequency-domain feature on each position (time point) of the temporal-domain feature. Conversely, the time-domain feature can be used as the query to calculate the attention weight of the temporal-domain feature on each channel of the frequency-domain feature. These two weight matrices reveal the importance distribution of temporal and frequency features from each other's perspectives. Then, the calculated attention weight matrices are applied to the original temporal and frequency-domain feature tensors, respectively, to obtain feature representations weighted by the other's perspective. Subsequently, a learnable fusion gating vector is introduced, generated from the concatenated weighted features by a small neural network. This gating vector determines the proportion of temporal-dominant and frequency-dominant representations in the final fused feature, allowing the network to dynamically adjust the fusion strategy based on the specific input signal characteristics. Finally, the two sets of attention-modulated and gating features are integrated into a unified, complementary deep feature vector through weighted summation or convolution after concatenation.

[0031] In this way, the time-domain flow, through the synergy of porous convolution and deformable convolution, achieves accurate adaptive extraction of multi-scale time-domain patterns of non-stationary impact signals; the frequency-domain flow, through a learnable complex neural network, achieves end-to-end optimized modeling of the time-frequency representation of the signal and its phase information; the bidirectional attention fusion layer goes beyond simple feature concatenation or addition, achieving deep interaction and semantic alignment between time-frequency features through a cross-attention mechanism, and using a gating mechanism to achieve adaptive fusion strategy, thereby obtaining deep features that can comprehensively and robustly characterize the essence of pipeline impact events.

[0032] Furthermore, to obtain array propagation characteristics reflecting the propagation properties of shock waves, a sparse reconstruction algorithm based on compressed sensing is employed to overcome the spatial resolution limitations imposed by the limited number of sensors. Treating the sensor spatial locations as non-uniform sampling points, the algorithm utilizes the sparse prior in the frequency domain of the signal and solves a constrained optimization problem to reconstruct the complete frequency-wavenumber spectrum with high accuracy. The reconstructed spectrum clearly reveals the distribution of vibration energy with frequency and spatial wavenumber, where energy peaks correspond to the dominant propagation modes. From the frequency-wavenumber spectrum, an adaptive peak detection and tracking algorithm is used to extract the dispersion curve of the pipeline. This algorithm slides along the frequency axis, and on each frequency slice, a peak detection method based on local contrast is used to identify the wavenumber corresponding to the energy peak. Simultaneously, utilizing the continuity and smoothness prior of the dispersion curve, a dynamic programming algorithm is introduced to connect candidate peak points at different frequencies to obtain the optimal path, automatically generating multiple complete dispersion curves corresponding to different orders of guided wave modes such as longitudinal waves, transverse waves, and flexural waves.

[0033] An improved spectral clustering algorithm is used to separate the modes of the extracted dispersion curves. Traditional methods struggle to distinguish dispersion curves of different propagation modes because they may intersect or be close in frequency-wavenumber space. This approach constructs a similarity map containing all candidate dispersion points. The similarity is based not only on the Euclidean distance between points but also on the shape characteristics of the local spectral energy at each point, such as local curvature and energy gradient direction. Spectral clustering is then performed on this basis, more robustly aggregating points belonging to the same physical propagation mode but potentially spatially discontinuous, and accurately separating intertwined mode curves.

[0034] Finally, based on the separated dispersion curves of each order, the equivalent physical parameters of the pipeline-soil system are inverted and encoded as propagation feature vectors. The inversion process is achieved by solving a physics-driven optimization problem. A parameterized waveguide model is established, incorporating parameters such as the pipeline's elastic modulus, density, wall thickness, and diameter, and the soil's equivalent stiffness and damping. This model can forward calculate the theoretical dispersion curve. Using a hybrid optimization strategy combining genetic algorithms and local search, a set of physical parameters is sought that minimizes the overall difference between the theoretical dispersion curve calculated by the parameterized model and the corresponding order curve extracted from measurements. Ultimately, the key equivalent physical parameters obtained from the inversion, such as the equivalent phase velocity of the bending wave and the attenuation coefficient at a specific frequency, are combined with the statistical descriptors of the dispersion curves of each mode to form an array propagation feature vector that comprehensively reflects the propagation characteristics of the shock wave in a specific pipeline-soil system. This feature is directly related to the state of the propagation medium and is of great value for impact location and pipeline health assessment.

[0035] Step S14: Based on the fused features, identify and classify the impact events of the target underground pipeline using a preset recognition model.

[0036] In this embodiment, a basic classification model is pre-trained to obtain a trained classification model. A pre-defined hierarchical domain adaptation strategy is used to transfer the trained classification model to a target domain, resulting in a transferred classification model, which is then used as a pre-defined recognition model. The target domain is the area for classifying impact events on underground pipelines. Based on the fused features, the pre-defined recognition model identifies and classifies the impact events on the target underground pipeline. Specifically, a basic classification model is pre-trained on a heterogeneous dataset. A hierarchical domain adaptation strategy is used to transfer the trained basic classification model to the target domain, resulting in a transferred classification model. The model classification result and the classification conclusion based on physical rules are fused at the decision level using Dempster-Shafer (DS) evidence theory to output an impact event classification result with uncertainty measurement.

[0037] It's important to understand that classification conclusions based on physical rules refer to judgments derived from the automatic reasoning and diagnosis of impact events through a rule base encoded within the system, derived from mechanical principles and engineering experience. The core knowledge of these rules originates from stress wave propagation theory in solid mechanics, pipeline structural dynamics, and extensive historical accident analysis and field experience. This knowledge is systematically encoded into computable logical judgments or quantitative models. For example: the energy attenuation-distance constraint rule assesses data consistency based on the energy attenuation model of stress wave propagation; the wave velocity-material matching rule verifies or challenges impact type hypotheses by comparing measured wave velocities with theoretical wave velocity ranges; the time-frequency characteristic-impact source mapping rule uses quantified time-frequency fingerprints of typical impact sources for feature matching; and the operating condition context consistency rule integrates pipeline operating status and external environmental information to make logically reasonable judgments. These rules collectively constitute a computable, independent reasoning engine, providing physically interpretable classification hypotheses and confidence assignments for decision fusion. A structured impact event report is automatically generated, typically containing key characterization information such as the event's time, signal waveform, and energy intensity, providing a basis for subsequent handling.

[0038] In addition, while the transfer model performs classification, the physics rule engine scans and evaluates the same set of input features in parallel, including but not limited to signal energy, time difference of arrival, dispersion characteristics, and time-frequency statistics. Each triggered rule outputs one or more propositions about the possible categories of the impact event, such as "it may be a hydraulic excavator impact," and assigns a basic confidence score based on the rule's determinism and its fit with the current data. The rule engine ultimately outputs a formalized set of classification hypotheses that integrates the inference results of multiple rules, along with their corresponding confidence scores.

[0039] Step S15: If the identification result indicates the existence of the impact event, a corresponding target impact event report is generated, and the impact point of the impact event is located using a preset positioning algorithm to complete the identification of the target underground pipeline impact event.

[0040] In this embodiment, if the identification result indicates the existence of the impact event, the classification result of the impact event is obtained, a corresponding target impact event report is generated based on the classification result, and a preset uncertainty quantification operation is performed on the classification result based on a Bayesian deep learning framework to obtain a corresponding quantization score. This allows the user to manually review the target impact event reports corresponding to classification results with quantization scores lower than a preset confidence threshold. Specifically, based on the Bayesian deep learning framework, uncertainty quantification is performed on the classification results of the transferred classification model, and classification results lower than a preset confidence level are manually reviewed. When an unknown impact pattern is detected, a dynamic category expansion mechanism based on metric learning is initiated to perform incremental learning and obtain an incremental classification model. When it is determined that the currently collected signal features meet the preset "impact event" criteria, such as external force damage behaviors like excavation, drilling, or heavy object crushing, an emergency response mechanism is immediately activated.

[0041] Next, a positioning algorithm combining intelligent wave velocity correction and compressed sensing theory is used to locate the impact point. First, a mapping model between pipeline-soil environmental parameters and local wave velocity is constructed. This model uses a feedforward neural network as its core, with inputs being multi-dimensional feature vectors describing the local state of the pipeline, including: pipeline material and diameter, burial depth, soil type and density inferred from historical geological exploration data or real-time sensor feedback, estimated soil moisture, and ambient temperature. The network output is the predicted equivalent propagation velocity of the pipeline segment under different types of stress waves. In the initial deployment phase, a small number of test impact signals from known locations are used to invert the initial wave velocity distribution through array signal inversion, and this is used as supervisory data to train the network. During long-term operation, the wave velocity prediction model is continuously fine-tuned and updated online using the results of each high-confidence positioning and its corresponding array signal, enabling it to adapt to the slow wave velocity drift caused by seasonal soil changes and water content fluctuations, thereby obtaining an intelligent dynamic wave velocity field that varies with time and space. High-precision sparse positioning is then solved based on the compressed sensing framework. After obtaining the dynamic wave velocity field, the impact source localization problem is transformed into a sparse signal reconstruction problem. The entire surface of the pipeline that may be impacted is discretized into a dense grid of potential source points. The theoretical arrival time from each potential source point to each sensor can be accurately calculated from the currently estimated dynamic wave velocity field. The actual signal arrival times measured by all sensors constitute an observation vector. The relationship between the observation vector and the potential source point grid is described by a time delay dictionary matrix constructed based on wave velocity and geometric path. Since each impact event is inherently spatially sparse, the localization problem is equivalent to finding the sparsest source activation vector under this overcomplete time delay dictionary, such that the error between the theoretical arrival time and the observed time is minimized. To solve this underdetermined inverse problem, a compressed sensing algorithm with weighted L1 norm minimization is used. A spatially adaptive weighting strategy is introduced, where the weights not only consider the geometric relationship between the grid points and the sensors, but are also negatively correlated with the local velocity uncertainty estimate output by the first-stage intelligent wave velocity field model. In areas with high wave velocity uncertainty, such as abrupt changes in soil properties, the weight of corresponding grid points in the optimization problem is reduced to avoid mislocation due to unreliable wave velocity information. Simultaneously, prior knowledge, such as the likelihood that impact events may occur in specific areas above pipelines, further constrains the spatial distribution of the solution. By solving this optimization problem, a highly sparse impact energy distribution map on the grid can be directly reconstructed, and its energy peak point is determined as the precise location of the impact event, i.e., the impact point. Through this series of automated closed-loop processes from identification and recording to precise location, the threatened locations of underground pipelines can be quickly pinpointed, providing timely and accurate data support for pipeline protection decisions.

[0042] As described above, in this embodiment, a piezoelectric sensor array is deployed at preset positions on the pipeline surface to collect initial piezoelectric signals. Then, the signal is preprocessed using a composite algorithm that integrates adaptive bionic filtering, wavelet-dictionary learning joint denoising, and time-varying blind source separation to obtain a filtered signal. Subsequently, a multi-dimensional feature extraction algorithm is used to extract and fuse features from the filtered signal, forming fused features. Based on these features, a preset recognition model is used to identify and classify pipeline impact events. If an impact event is identified, an event hypothesis report is generated, and a preset positioning algorithm is used to determine the impact point location, thereby achieving complete monitoring of pipeline impact loads. In this way, high-density, adaptive vibration sensing of the gas pipeline surface is achieved through the piezoelectric array. The composite preprocessing algorithm consisting of bionic filtering, joint denoising, and blind source separation significantly improves the signal-to-noise ratio of the impact signal in complex noise environments. Based on multi-dimensional feature extraction and fusion technology, a fused feature vector capable of comprehensively characterizing the spatiotemporal features of impact events is constructed. Further utilizing the recognition model of the transfer learning framework, high-precision impact classification is achieved, and combined with intelligent wave velocity correction and compressed sensing theory, sub-meter-level impact positioning accuracy is realized. Ultimately, a complete monitoring system is formed, encompassing signal perception, feature extraction, event recognition, and precise positioning, enabling intelligent and highly reliable monitoring and early warning of impact loads on underground gas pipelines.

[0043] See Figure 2 As shown, an embodiment of the present invention discloses an impact event identification device for underground pipelines, comprising: The signal acquisition module 11 is used to acquire the initial piezoelectric signal of the target underground pipeline using a piezoelectric sensor array; the piezoelectric sensor array includes piezoelectric sensors arranged on the surface of the target underground pipeline according to preset positions; The signal generation module 12 is used to preprocess the initial piezoelectric signal using a preset composite algorithm to generate a processed piezoelectric signal; wherein the preset composite algorithm is an algorithm constructed based on adaptive bionic filtering technology, joint denoising algorithm and time-varying blind source separation algorithm; Feature fusion module 13 is used to perform multi-dimensional feature extraction and fusion operations on the processed piezoelectric signal using a multi-dimensional feature extraction algorithm to obtain the corresponding fused features; Event classification module 14 is used to identify and classify impact events of the target underground pipeline based on the fused features and a preset recognition model; The impact point location module 15 is used to generate a corresponding target impact event report if the identification result indicates the existence of the impact event, and to locate the impact point of the impact event using a preset location algorithm, so as to complete the identification of the target underground pipeline impact event.

[0044] As described above, firstly, this application deploys a piezoelectric sensor array at predetermined positions on the pipeline surface to collect initial piezoelectric signals. Next, the signal is preprocessed using a composite algorithm that integrates adaptive bionic filtering, wavelet-dictionary learning joint denoising, and time-varying blind source separation to obtain a filtered signal. Subsequently, a multi-dimensional feature extraction algorithm is used to extract and fuse features from the filtered signal, forming fused features. Based on these features, a preset recognition model is used to identify and classify pipeline impact events. If an impact event is identified, an event hypothesis report is generated, and a preset positioning algorithm is used to determine the impact point location, thereby achieving complete monitoring of pipeline impact loads. In this way, high-density, adaptive vibration sensing of the gas pipeline surface is achieved through the piezoelectric array. The composite preprocessing algorithm, consisting of bionic filtering, joint denoising, and blind source separation, significantly improves the signal-to-noise ratio of the impact signal in complex noise environments. Based on multi-dimensional feature extraction and fusion technology, a fused feature vector capable of comprehensively characterizing the spatiotemporal features of impact events is constructed. Further utilizing the recognition model of the transfer learning framework, high-precision impact classification is achieved, and combined with intelligent wave velocity correction and compressed sensing theory, sub-meter-level impact positioning accuracy is realized. Ultimately, a complete monitoring system is formed, encompassing signal perception, feature extraction, event recognition, and precise positioning, enabling intelligent and highly reliable monitoring and early warning of impact loads on underground gas pipelines.

[0045] In some specific embodiments, the signal generation module 12 may specifically include: The model introduction unit is used to construct a preset filter group and introduce a psychoacoustic masking model; wherein, the preset filter group is a set of filters that mimic the cochlear basilar membrane; The curve determination unit is used to determine the masking threshold curve through the psychoacoustic masking model in the preset filter group; The first signal acquisition unit is used to determine the corresponding masking threshold in the masking threshold curve based on the instantaneous power spectrum of the initial piezoelectric signal, suppress background noise in the initial piezoelectric signal that is less than the masking threshold, and obtain the piezoelectric signal after biomimetic processing. The coefficient generation unit is used to perform multi-level wavelet decomposition on the biomimetic piezoelectric signal to generate multi-scale wavelet coefficients. The first coefficient acquisition unit is used to divide the multi-scale wavelet coefficients into blocks to obtain the block-divided wavelet coefficients. The dictionary determination unit is used to determine an overcomplete dictionary to adapt to the characteristics of pipeline impact using a preset dictionary learning algorithm; The second coefficient acquisition unit is used to sparsely reconstruct the block-based wavelet coefficients on the overcomplete dictionary using the orthogonal matching pursuit algorithm to obtain the reconstructed coefficients. The second signal acquisition unit is used to determine a preset adaptive threshold based on the local statistical characteristics of the multi-scale wavelet coefficients, and to perform a preset coefficient shrinkage operation on the reconstructed coefficients according to the preset adaptive threshold to obtain a denoised piezoelectric signal. The third signal acquisition unit is used to divide the denoised piezoelectric signal into a preset sliding time window, and perform blind source separation within each preset sliding time window using a preset JADE algorithm to obtain the separated piezoelectric signal. The fourth signal acquisition unit is used to perform component tracking on the separated piezoelectric signals of adjacent windows, and to use a Kalman filter to estimate the change of the time-varying mixing matrix in the separated piezoelectric signals, so as to extract the source signal related to the impact event as the processed piezoelectric signal.

[0046] In some specific implementations, the feature fusion module 13 may specifically include: The feature extraction unit is used to extract multi-scale morphological features of the processed piezoelectric signal based on a preset multi-scale structuring element set; the multi-scale morphological features are features of the local geometric structure of the impact waveform; The feature generation unit is used to input the processed piezoelectric signal into a preset depth dual-stream convolutional neural network to generate deep features containing the time-frequency joint distribution characteristics of the piezoelectric signal. The curve extraction unit is used to determine the frequency and wavenumber spectrum based on the processed piezoelectric signal, and extract the dispersion curve of the target underground pipeline from the frequency and wavenumber spectrum through peak detection operation; The feature determination unit is used to invert the equivalent physical parameters of the preset pipeline and soil combined system based on the extracted dispersion curve, so as to determine the array propagation characteristics of the processed piezoelectric signal; the array propagation characteristics are features that reflect the propagation characteristics of shock waves; The weight determination unit is used to determine the fusion weights of the multi-scale morphological features, the depth features, and the array propagation features respectively through a preset feature weight allocation network; The feature fusion unit is used to perform weighted fusion of the multi-scale morphological features, the depth features, and the array propagation features according to the fusion weights to generate fused features.

[0047] In some specific implementations, the feature extraction unit may specifically include: The scale sequence determination subunit is used to determine the scale sequence of the corresponding structural elements based on the statistical characteristics of the processed piezoelectric signal and the physical parameters of the target underground pipeline. The scale sequence of the structural elements includes linear structural elements, circular structural elements, and rectangular structural elements. The statistical characteristics include the standard deviation, kurtosis, and instantaneous amplitude distribution of the signal. The physical parameters include the elastic modulus, wall thickness, and estimated stress wave propagation velocity of the pipe material. The impact signal feature acquisition subunit is used to perform morphological gradient calculation using the processed piezoelectric signal of the linear structural element to obtain impact signal features. The waveform contour feature acquisition subunit is used to perform top-hat operation using the processed piezoelectric signal of the circular structural element to obtain waveform contour features; The maximum region extraction subunit is used to perform iterative expansion operations using the processed piezoelectric signal of the rectangular structural element to extract the local maximum region of the target signal; The feature construction subunit is used to construct multi-scale morphological features based on the impact signal features, the waveform contour features, and the local maximum region of the target signal.

[0048] In some specific implementations, the feature generation unit may specifically include: The first feature extraction subunit is used to extract the temporal flow features in the processed piezoelectric signal by utilizing the temporal flow branch of a preset deep dual-stream convolutional neural network. The second feature extraction subunit is used to extract frequency domain flow features from the processed piezoelectric signal using the frequency domain flow branch of the deep two-stream convolutional neural network. The attention weight determination subunit is used to determine the attention weights of the temporal flow features and the frequency flow features respectively through the bidirectional attention fusion layer of the preset deep dual-stream convolutional neural network. The deep feature acquisition subunit is used to perform weighted fusion of the temporal flow features and the frequency flow features using the attention weights to obtain the deep features.

[0049] In some specific implementations, the event classification module 14 may specifically include: The model training unit is used to pre-train the basic classification model to obtain the trained classification model. The model transfer unit is used to transfer the trained classification model to the target domain using a preset hierarchical domain adaptation strategy, thereby obtaining a transferred classification model, and using the transferred classification model as a preset recognition model; the target domain is the domain for classifying impact events of underground pipelines. The event classification unit is used to identify and classify the impact events of the target underground pipeline based on the fused features and a preset recognition model.

[0050] In some specific embodiments, the impact point positioning module 15 may specifically include: The event quantization unit is used to obtain the classification result of the impact event if the identification result indicates the existence of the impact event, generate a corresponding target impact event report based on the classification result, and perform a preset uncertainty quantization operation on the classification result based on the Bayesian deep learning framework to obtain a corresponding quantization score, so that the user terminal can manually review the target impact event report corresponding to the classification result with the quantization score less than a preset confidence threshold.

[0051] Furthermore, embodiments of this application also disclose an electronic device, Figure 3 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application.

[0052] Figure 3 This is a schematic diagram of the structure of an electronic device 20 provided in an embodiment of this application. Specifically, the electronic device 20 may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the underground pipeline impact event identification method disclosed in any of the foregoing embodiments. Alternatively, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0053] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0054] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0055] The operating system 221 is used to manage and control the various hardware devices on the electronic device 20 and the computer program 222, which may be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program capable of performing the underground pipeline impact event identification method executed by the electronic device 20 as disclosed in any of the foregoing embodiments, the computer program 222 may further include computer programs capable of performing other specific tasks.

[0056] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned method for identifying impact events in underground pipelines. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0057] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0058] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0059] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0060] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0061] The technical solutions provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for identifying impact events in underground pipelines, characterized in that, include: The initial piezoelectric signal of the target underground pipeline was acquired using a piezoelectric sensor array; The piezoelectric sensing array includes individual piezoelectric sensors arranged at preset positions on the surface of the target underground pipeline; The initial piezoelectric signal is preprocessed using a preset composite algorithm to generate a processed piezoelectric signal; wherein the preset composite algorithm is an algorithm constructed based on adaptive bionic filtering technology, joint denoising algorithm and time-varying blind source separation algorithm; A multidimensional feature extraction algorithm is used to perform multidimensional feature extraction and fusion operations on the processed piezoelectric signal to obtain the corresponding fused features. Based on the fused features, the impact events of the target underground pipeline are identified and classified using a preset recognition model; If the identification result indicates the existence of the impact event, a corresponding target impact event report is generated, and the impact point of the impact event is located using a preset positioning algorithm to complete the identification of the target underground pipeline impact event.

2. The method for identifying impact events in underground pipelines according to claim 1, characterized in that, The step of preprocessing the initial piezoelectric signal using a preset composite algorithm to generate a processed piezoelectric signal includes: A preset filter bank is constructed, and a psychoacoustic masking model is introduced; wherein, the preset filter bank is a set of filters that mimic the cochlear basilar membrane; The masking threshold curve is determined by the psychoacoustic masking model in the preset filter group; Based on the instantaneous power spectrum of the initial piezoelectric signal, the corresponding masking threshold in the masking threshold curve is determined, and the background noise in the initial piezoelectric signal that is less than the masking threshold is suppressed to obtain the biomimetic piezoelectric signal; Multi-scale wavelet coefficients are generated by performing multi-level wavelet decomposition on the biomimetic piezoelectric signal. The multi-scale wavelet coefficients are divided into blocks to obtain the block-divided wavelet coefficients. An overcomplete dictionary is determined using a pre-defined dictionary learning algorithm to adapt to the characteristics of pipeline impact. The block-based wavelet coefficients are sparsely reconstructed on the overcomplete dictionary using the orthogonal matching pursuit algorithm to obtain the reconstructed coefficients. Based on the local statistical characteristics of the multi-scale wavelet coefficients, a preset adaptive threshold is determined, and a preset coefficient shrinkage operation is performed on the reconstructed coefficients according to the preset adaptive threshold to obtain the denoised piezoelectric signal. The denoised piezoelectric signal is divided into preset sliding time windows, and blind source separation is performed within each preset sliding time window using a preset JADE algorithm to obtain the separated piezoelectric signal. The components of the separated piezoelectric signals in adjacent windows are tracked, and the changes in the time-varying mixing matrix in the separated piezoelectric signals are estimated using a Kalman filter to extract the source signal related to the impact event as the processed piezoelectric signal.

3. The method for identifying impact events in underground pipelines according to claim 1, characterized in that, The step of using a multidimensional feature extraction algorithm to perform multidimensional feature extraction and fusion operations on the processed piezoelectric signal to obtain the corresponding fused features includes: Multi-scale morphological features of the processed piezoelectric signal are extracted based on a preset set of multi-scale structural elements; the multi-scale morphological features are the features of the local geometric structure of the impact waveform. The processed piezoelectric signal is input into a preset depth dual-stream convolutional neural network to generate deep features containing the joint time-frequency distribution characteristics of the piezoelectric signal; The frequency and wavenumber spectrum are determined based on the processed piezoelectric signal, and the dispersion curve of the target underground pipeline is extracted from the frequency and wavenumber spectrum through peak detection operation. Based on the extracted dispersion curve, the equivalent physical parameters of the preset pipeline and soil combined system are inverted to determine the array propagation characteristics of the processed piezoelectric signal; the array propagation characteristics are those that reflect the propagation characteristics of shock waves. By using a preset feature weight allocation network, the fusion weights of the multi-scale morphological features, the depth features, and the array propagation features are determined respectively. The multi-scale morphological features, the depth features, and the array propagation features are weighted and fused according to the fusion weights to generate fused features.

4. The method for identifying impact events in underground pipelines according to claim 3, characterized in that, The extraction of multi-scale morphological features of the processed piezoelectric signal based on a preset multi-scale structuring element set includes: Based on the statistical characteristics of the processed piezoelectric signal and the physical parameters of the target underground pipeline, the scale sequence of the corresponding structural elements is determined; wherein, the scale sequence of the structural elements includes linear structural elements, circular structural elements and rectangular structural elements, the statistical characteristics include the standard deviation, kurtosis and instantaneous amplitude distribution of the signal, and the physical parameters include the elastic modulus of the pipe, the wall thickness and the estimated stress wave propagation velocity. Morphological gradient calculations are performed using the processed piezoelectric signal of the linear structural element to obtain impact signal characteristics; The processed piezoelectric signal of the circular structural element is used to perform a top cap operation to obtain waveform contour features; Iterative expansion calculations are performed using the processed piezoelectric signal of the rectangular structural element to extract the local maximum region of the target signal; Based on the impact signal characteristics, the waveform contour characteristics, and the local maxima region of the target signal, multi-scale morphological features are constructed.

5. The method for identifying impact events in underground pipelines according to claim 3, characterized in that, The step of inputting the processed piezoelectric signal into a preset depth dual-stream convolutional neural network to generate deep features containing the joint time-frequency distribution characteristics of the piezoelectric signal includes: The temporal flow features of the processed piezoelectric signal are extracted using the temporal flow branch of a pre-defined deep dual-stream convolutional neural network. Frequency domain flow features in the processed piezoelectric signal are extracted using the frequency domain flow branch of a deep two-stream convolutional neural network. The attention weights of the temporal flow features and the frequency flow features are determined by the bidirectional attention fusion layer of the preset deep dual-stream convolutional neural network. The attention weights are used to perform weighted fusion of the temporal flow features and the frequency flow features to obtain the deep features.

6. The method for identifying impact events in underground pipelines according to claim 1, characterized in that, The process of identifying and classifying impact events on the target underground pipeline based on the fused features using a preset recognition model includes: Pre-train the basic classification model to obtain the trained classification model; The trained classification model is transferred to the target domain using a preset hierarchical domain adaptation strategy to obtain a transferred classification model, which is then used as a preset recognition model; the target domain is the domain for classifying impact events on underground pipelines. Based on the fused features, the impact events of the target underground pipeline are identified and classified using a preset recognition model.

7. The method for identifying impact events in underground pipelines according to any one of claims 1 to 6, characterized in that, If the identification result indicates the existence of the impact event, a corresponding target impact event report is generated, including: If the identification result indicates the existence of the impact event, the classification result of the impact event is obtained, a corresponding target impact event report is generated based on the classification result, and a preset uncertainty quantization operation is performed on the classification result based on the Bayesian deep learning framework to obtain the corresponding quantization score, so that the user terminal can manually review the target impact event report corresponding to the classification result with the quantization score less than the preset confidence threshold.

8. A device for identifying impact events in underground pipelines, characterized in that, include: The signal acquisition module is used to acquire the initial piezoelectric signal of the target underground pipeline using a piezoelectric sensor array; The piezoelectric sensing array includes individual piezoelectric sensors arranged at preset positions on the surface of the target underground pipeline; The signal generation module is used to preprocess the initial piezoelectric signal using a preset composite algorithm to generate a processed piezoelectric signal; wherein the preset composite algorithm is an algorithm constructed based on adaptive bionic filtering technology, joint denoising algorithm and time-varying blind source separation algorithm; The feature fusion module is used to perform multi-dimensional feature extraction and fusion operations on the processed piezoelectric signal using a multi-dimensional feature extraction algorithm to obtain the corresponding fused features. The event classification module is used to identify and classify the impact events of the target underground pipeline based on the fused features and a preset recognition model. The impact point localization module is used to generate a corresponding target impact event report if the identification result indicates the existence of the impact event, and to locate the impact point of the impact event using a preset localization algorithm, so as to complete the identification of the target underground pipeline impact event.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the impact event identification method for underground pipelines as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Used to store a computer program; wherein, when the computer program is executed by a processor, it implements the method for identifying impact events in underground pipelines as described in any one of claims 1 to 7.

Images