An underground cable identification method and system

Abstract

The application provides an underground cable identification method and system, and belongs to the technical field of underground pipe network detection and electromagnetic measurement. K The application comprises the following steps: synchronously acquiring physical channel signals in a space at an offset position directly above a target cable, constructing a three-dimensional magnetic field gradient tensor, and solving the gradient tensor to obtain a transverse attenuation gradient feature; a tensor fusion is performed on one intrinsic modal component generated by each physical channel to construct a high-latitude space-time joint observation matrix, and independent source signal components are mapped from the space-time joint observation matrix; a pure target intrinsic active signal is acquired from the output source signal components; a three-dimensional heterogeneous feature flow matrix is obtained by splicing the transverse attenuation gradient feature, the standard deviation of the instantaneous phase, and the harmonic distortion rate of the spectrum according to time steps; deep feature extraction and classification are performed on the continuously obtained three-dimensional heterogeneous feature flow matrix to acquire a confidence probability that a detection object in a current sliding time window belongs to the target cable.

Description

Technical Field

[0001] This invention relates to the field of underground pipeline detection and electromagnetic measurement technology, and in particular to a method and system for identifying underground cables. Background Technology

[0002] Underground power cables and communication networks play an indispensable role in the deep spatial development of modern cities. In past engineering practices, cable detection technology based on the principle of active electromagnetic induction has been widely used. This technology mainly relies on handheld single-axis or dual-axis receiving coils on the ground to capture alternating magnetic fields of the same frequency and use peak-valley methods or simple absolute phase comparisons to achieve rough location and identification of pipelines. However, due to the high-frequency drift of the physical space reference plane caused by heavy reliance on manual dynamic scanning, and the strong mutual inductance coupling effect in high-density parallel pipeline networks, traditional single-node detection architectures face an unsolvable mathematical underdeterminacy state of blind source separation under the superposition of multi-phase field sources, which easily leads to false peak misjudgments. Furthermore, static, low-dimensional classification models lack the ability to smooth the temporal progression of transient high-energy electromagnetic interference, severely hindering the further promotion and application of accurate identification technology for complex utility tunnels: 1. In environments where multiple cables run in close parallel proximity, strong electromagnetic inductance coupling exists between the cables, leading to significant proximity effects and a high likelihood of false alarms; 2. Spatial feature extraction heavily relies on spatial reference errors caused by manual scanning; 3. Ground receivers typically face a high degree of temporal and frequency overlap between target signals, co-frequency induced interference, and environmental background noise. Existing detection equipment is limited by single-channel signal acquisition, and directly using blind source separation algorithms faces an underdetermined physical bottleneck where the number of source signals exceeds the number of observation channels, making it impossible to completely decouple and separate interference mathematically; 4. Existing intelligent pipeline identification models mostly employ simple fixed threshold comparisons or static classifiers such as SVM, lacking a comprehensive mechanism for examining the evolutionary patterns of features over time.

[0003] Intelligent identification technology based on spatial gradient tensor and spatiotemporal joint blind source separation is an effective means to overcome the bottleneck of accurate detection in high-density underground pipe networks. Compared with traditional single-channel equipment that relies on manual dynamic displacement and simple decision-making systems based on static thresholds, this intelligent identification system achieves a qualitative breakthrough in the extraction of underlying physical features by means of orthogonal multi-coil sensing arrays, establishing an absolutely rigid spatial reference system; its signal decoupling dimension is significantly improved, eliminating the risk of false peaks caused by strong mutual inductance coupling from the underlying mathematical logic, while possessing dynamic tracking and strong anti-interference ability of multi-dimensional heterogeneous features in space, time and frequency based on temporal convolutional networks (TCN).

[0004] Therefore, this paper proposes an underground cable identification method and system. By applying spatial gradient tensor and spatiotemporal joint blind source separation techniques to harsh engineering scenarios such as densely interwoven underground pipeline construction and complex electromagnetic environment exploration, the system can achieve accurate and unique identification of target cables, which has important application significance. Summary of the Invention

[0005] In view of this, the present invention proposes an underground cable identification method and system based on hardware array synchronous sampling to obtain spatial gradient, frequency domain adaptive virtual dimensionality upscaling, construction of a spatiotemporal joint observation matrix to remove crosstalk, extraction of multidimensional heterogeneous fingerprint sequences, and output of final confidence through a temporal convolutional network.

[0006] On one hand, the present invention provides a method for identifying underground cables, comprising the following steps: S1: Configure an orthogonal multi-coil sensing array to synchronously acquire the three-dimensional magnetic induction intensity in the offset position space near the center point directly above the target cable. M The physical channel signal is used to construct a three-dimensional magnetic field gradient tensor based on the three-dimensional magnetic induction intensity. G Solve the gradient tensor in real time G The normalized lateral decay gradient features are obtained. F grad ; S2: Yes M The signals of each physical channel are decomposed separately, so that each physical channel signal is decomposed in the frequency domain into... K One intrinsic mode component; S3: Yes M physical paths and the generation of each path K Tensor fusion of the intrinsic modal components is performed to construct a high-dimensional spatiotemporal joint observation matrix. X ST joint spatiotemporal observation matrix X ST A fast independent component analysis algorithm based on maximizing negative entropy is employed to analyze the spatiotemporal joint observation matrix. X ST The process maps out independent source signal components, and extracts the pure intrinsic active signal of the target from the output source signal components. s ( t ); S4: From the reconstructed target intrinsic active signal s ( t After that, the lateral decay gradient feature from step S1 is inherited. F grad ; for the target's intrinsic active signals s ( t Perform a Hilbert transform to extract the instantaneous phase, and calculate the standard deviation of the instantaneous phase within a set discrete sliding window. F phase ; for the target's intrinsic active signals s ( t Perform a Fast Fourier Transform to calculate the harmonic distortion rate of the spectrum. FHD ; lateral decay gradient features F grad Standard deviation of instantaneous phase F phase Harmonic distortion rate of the spectrum F HD The three features are concatenated according to the time step to obtain a three-dimensional heterogeneous feature flow matrix that evolves continuously over time. F ( t ); S5: Continuously obtained L Three-dimensional heterogeneous feature flow matrix F ( t Perform data serialization construction based on a sliding time window, with a construction dimension of 3× L Temporal input tensor X seq Translate the time series into the tensor X seq Input the temporal convolutional network (TCN) for deep feature extraction and classification, and obtain the confidence probability that the detected object belongs to the target cable within the current sliding time window.

[0007] Based on the above technical solutions, preferably, the orthogonal multi-coil sensing array mentioned in step S1 includes... M Each triaxial magnetic field sensor node is rigidly fixed in a cross-shaped or star-shaped orthogonal topology, and the three-dimensional spatial relative spacing Δ between the triaxial magnetic field sensor nodes is [missing information]. x , △ y , △ z It is calibrated and solidified as a constant.

[0008] Based on the above technical solutions, preferably, the method described in step S1 for constructing a three-dimensional magnetic field gradient tensor based on three-dimensional magnetic induction intensity is preferred. G The ratio of the magnetic field amplitude at the peak point directly above the target cable to that at the lateral offset point is calculated to obtain the normalized lateral attenuation gradient feature. F grad , is the definition of the magnetic field vector Vector components along the three orthogonal directions of the spatial rectangular coordinate system The full-order spatial partial derivatives are used as the three-dimensional magnetic field gradient tensor. G Solving the three-dimensional magnetic field gradient tensor G The trace or principal eigenvalue, and the three-dimensional magnetic field gradient tensor G The trace or principal eigenvalue is mapped to a normalized lateral decay gradient feature. F grad .

[0009] Preferably, step S2 involves using envelope entropy as the evaluation index and employing a heuristic algorithm to dynamically optimize the variational mode decomposition (VMD) parameters; then using the optimized VMD parameters to... M The physical channel signals are decomposed separately.

[0010] Preferably, the step S3 involves... M physical paths and the generation of each path K Tensor fusion of the intrinsic modal components is performed to construct a high-dimensional spatiotemporal joint observation matrix. X ST It is the process of performing a joint space-time blind source separation operation to synchronously output the three-dimensional magnetic induction intensity from an orthogonal multi-coil sensing array. M physical channel signal, and K The number of effective linearly independent observation channels is expanded by matrix concatenation of the intrinsic mode components. M × K One, to ensure the full-rank property of the observation matrix.

[0011] Preferably, the spatiotemporal joint observation matrix described in step S3 X ST A fast independent component analysis algorithm based on maximizing negative entropy is employed to analyze the spatiotemporal joint observation matrix. X ST The process maps out independent source signal components, and extracts the pure intrinsic active signal of the target from the output source signal components. s ( t (This refers to the construction of a joint spatiotemporal observation matrix.) X ST Subsequently, after mean-neutralization and whitening preprocessing, eigenvalue decomposition is performed to compress the search space of the subsequent unmixing matrix into the orthogonal matrix category. Then, blind source extraction is performed based on the central limit theory, and negative entropy is selected as the objective function to measure the non-Gaussianity of the separated signal. The orthogonal unmixing matrix is ​​updated using the Newton-Raphson iteration method to maximize the objective function, thereby maximizing the spatiotemporal joint observation matrix. X ST Mapped to independent source signal components, and combined with the pre-defined prior coded pulse or modulation envelope characteristics of the target cable transmitter, correlation matching is performed on the independent source signal components to obtain the pure intrinsic active signal of the target after eliminating co-frequency induced interference. s ( t ).

[0012] Preferably, step S4 involves calculating the standard deviation of the instantaneous phase within a set discrete sliding window. F phase ; for the target's intrinsic active signals s ( tPerform a Fast Fourier Transform to calculate the harmonic distortion rate of the spectrum. F HD This means setting the number of discrete sampling points within the discrete sliding window to be... W , respectively obtain the first w The instantaneous phase value at each discrete sampling point and the average instantaneous phase value within the discrete sliding window. w The sequence number of the discrete sampling point. w =1, 2, ..., W The summation is performed by taking the squares of the differences between the instantaneous phase values ​​at all sampling points and the average instantaneous phase value within the discrete sliding window, and then dividing the sum by the number of discrete sampling points. W The back square root is used as the standard deviation of the instantaneous phase. F phase ; for the target's intrinsic active signals s ( t After performing a Fast Fourier Transform, let A 1 represents the amplitude of the target fundamental wave in the frequency domain amplitude spectrum. A n Indicates the first n The amplitude of the second-highest frequency harmonic is calculated by summing the squares of the amplitudes of the second-highest frequency harmonics and taking the square root, then dividing by the amplitude of the target fundamental frequency. A 1. Obtain the harmonic distortion rate of the spectrum. F HD .

[0013] Preferably, the step S5 involves converting the temporal input tensor... X seq Deep feature extraction and classification are performed on the input Temporal Convolutional Network (TCN). This involves configuring multi-level residual connection modules within the TCN, each containing dilated causal convolutional layers. These dilated causal convolutional layers are used to compute the hidden state at each time step, only considering the state at that time step. t The input data at historical moments are subjected to dilated causal convolution operations. A dilation factor is introduced into the convolution kernel to expand the historical receptive field of temporal features. The dilation factor is set to increase exponentially. The residual connection module includes two parallel forward computation paths. One is the main path, which is used to perform dilated causal convolution, weight normalization and nonlinear activation processing. The other is the side path, which is used to directly map the input tensor of the residual connection module to the output end and add it element-wise with the output of the main path.

[0014] Preferably, in step S5, obtaining the confidence probability that the detected object within the current sliding time window belongs to the target cable involves configuring a global average pooling layer and a fully connected layer. The global average pooling layer is placed at the end of the multi-level residual connection module to compress the high-dimensional temporal feature map extracted by the multi-level residual connection module in the time dimension, generating a one-dimensional deep feature vector.H out The data is then fed into a fully connected layer to perform classification probability mapping and output confidence probabilities. P Pre-set a high confidence threshold P high and low confidence threshold P low When the confidence probability P ≥ High confidence threshold P high When the current cable is the target cable, output a confirmation feedback signal; when the confidence probability is... P ≤low confidence threshold P low When the current cable is determined to be a nearby interfering cable, the confidence probability is... P At high confidence threshold P high and low confidence threshold P low When the two are in between, the judgment state of the previous sliding time window is maintained, and the system waits for the input of newly collected timing characteristics for further judgment.

[0015] On the other hand, the present invention also provides an underground cable identification system for implementing the above-mentioned method, comprising: The lateral attenuation gradient feature acquisition unit, through a pre-configured orthogonal multi-coil sensing array, synchronously acquires the three-dimensional magnetic induction intensity in the offset position space near the center point directly above the target cable. M Physical channel signals are used to construct a three-dimensional magnetic field gradient tensor. G Solve the three-dimensional magnetic field gradient tensor G Obtain the normalized lateral decay gradient features F grad ; The target intrinsic active signal acquisition unit is used for... M The signals of each physical channel are decomposed separately, so that each physical channel signal is decomposed in the frequency domain into... K Each intrinsic mode component M physical channel signal and K Tensor fusion of the intrinsic modal components is performed to construct a high-dimensional spatiotemporal joint observation matrix. X ST joint spatiotemporal observation matrix X ST A fast independent component analysis algorithm based on maximizing negative entropy is employed to analyze the spatiotemporal joint observation matrix. X ST The process maps out independent source signal components, and extracts the pure intrinsic active signal of the target from the output source signal components. s ( t ); A three-dimensional heterogeneous feature flow matrix generation unit is used to generate intrinsic active signals of the target. s ( t Perform a Hilbert transform to extract the instantaneous phase, and calculate the standard deviation of the instantaneous phase within a set discrete sliding window. F phase ; for the target's intrinsic active signals s ( t Perform a Fast Fourier Transform to calculate the harmonic distortion rate of the spectrum. F HD The lateral decay gradient feature is obtained from the output of the lateral decay gradient feature acquisition unit. F grad Standard deviation of instantaneous phase F phase Harmonic distortion rate of the spectrum F HD The three features are concatenated according to the time step to obtain a three-dimensional heterogeneous feature flow matrix that evolves continuously over time. F ( t ); The target cable determination unit is used to obtain the three-dimensional heterogeneous feature flow matrix output by the three-dimensional heterogeneous feature flow matrix generation unit. F ( t ), and will obtain continuously L Three-dimensional heterogeneous feature flow matrix F ( t Perform data serialization construction based on a sliding time window, with a construction dimension of 3× L Temporal input tensor X seq Translate the time series into the tensor X seq Input the temporal convolutional network (TCN) for deep feature extraction and classification, and obtain the confidence probability that the detected object belongs to the target cable within the current sliding time window.

[0016] The underground cable identification method and system provided by this invention have the following advantages compared with the prior art: 1. This invention, by designing an orthogonal multi-coil sensing array that requires no dynamic displacement and combining it with a multi-dimensional spatial finite difference algorithm to construct a three-dimensional magnetic field gradient tensor in real time, completely eliminates spatial reference errors caused by manual scanning at the physical sensing level. This array architecture can establish a rigid spatial physical reference system within the same microsecond-level sampling period, directly locking the spatial physical curvature of the target cable according to the inverse attenuation model. This physical mechanism effectively eliminates data jumps caused by probe pose offset and inconsistent ground clearance, providing the entire recognition system with a highly objective and stable spatial dimension data stream.

[0017] 2. In the underlying signal demixing and denoising stage, the technical solution of this invention breaks through the underdetermined mathematical limit of single-channel blind source separation, effectively decoupling co-frequency interference caused by strong mutual inductance coupling in complex pipe networks. The system introduces adaptive variational mode decomposition driven by minimizing envelope entropy, which finely deconstructs the limited observation signals obtained from physical space in the frequency domain, extending them into high-dimensional virtual observation components. This core cross-dimensional reconstruction operation forces the underdetermined mathematical state of a single receiving node into a well-determined or overdetermined state with full column rank. Based on this, combined with a fast independent component analysis algorithm based on maximizing negative entropy, the system can completely separate environmental noise and adjacent line induced crosstalk from a deep mathematical and statistical level, reconstructing the target intrinsic source signal with high fidelity, and solving the problem of current amplitude overshoot and false alarm caused by mutual inductance effect in high-density parallel pipe corridors.

[0018] 3. To address the technical limitations of single-physical-dimensional criteria failing under complex electromagnetic conditions, this invention constructs a multi-dimensional heterogeneous feature matrix encompassing space, time, and frequency, significantly broadening the physical boundaries of target identification. The system performs Hilbert analysis and Fast Fourier Transform on the reconstructed pure source signal, extracting the instantaneous phase variance and high-frequency harmonic distortion rate, and combines this with the amplitude lateral attenuation gradient output by the front-end array to construct a feature flow matrix that continuously evolves over time. This feature flow deeply quantifies the inherent objective differences between active driving signals and passive sensing signals in macroscopic spatial topology, microscopic temporal jitter, and frequency domain energy nonlinear distortion, enabling the system to maintain highly accurate cable uniqueness identification capabilities even when facing absolute phase drift caused by high impedance mismatch at the far end or long-distance distributed capacitance.

[0019] 4. At the top-level intelligent classification decision-making level, this invention achieves an architectural evolution from static single-frame judgment to dynamic temporal tracking, significantly enhancing the system's robustness against transient electromagnetic pulses. The system introduces a temporal convolutional network with a dilated causal convolutional topology and residual connection modules, acquiring a long-span temporal context receptive field through exponentially increasing expansion factors. When transient high-energy electromagnetic interference caused by strong electrical sparks, large motor start-ups and shutdowns, etc., occurs in the industrial environment, causing abrupt changes in local feature data, this dynamic network can rely on the stable evolution trajectory within the historical time window to perform temporal smoothing and noise reduction on abnormal fluctuations. Simultaneously, combined with the hysteresis comparison judgment strategy configured at the end, the system effectively avoids frequent jumps in classification instructions under critical electromagnetic states, ensuring the reliability of equipment identification and the safety of on-site operations in harsh physical environments. Attached Figure Description

[0020] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of the architecture and process of an underground cable identification method and system according to the present invention. Detailed Implementation

[0022] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0023] Traditional single-node detection architectures face an unsolvable mathematical underdeterminacy in blind source separation under the superposition of multiple coherent field sources, which easily leads to false peak misjudgments. Furthermore, static low-dimensional classification models lack the ability to smooth the temporal progression of transient high-energy electromagnetic interference, exhibiting the following drawbacks: 1. In environments where multiple cables run in close parallel proximity, strong electromagnetic inductance coupling exists between the cables, resulting in significant proximity effects and a high likelihood of false alarms; 2. Spatial feature extraction heavily relies on spatial reference errors caused by manual scanning; 3. Ground receivers typically face a high degree of temporal and frequency overlap between target signals, co-frequency induced interference, and environmental background noise. Existing detection equipment is limited by single-channel signal acquisition; directly using blind source separation algorithms would face an underdeterminacy in physical terms, where the number of source signals exceeds the number of observation channels, making it impossible to completely decouple and separate interference mathematically; 4. Existing intelligent pipeline identification models often employ simple fixed threshold comparisons or static classifiers such as SVM, lacking a comprehensive mechanism for examining the evolution of features over time.

[0024] In view of this, such as Figure 1 As shown, in one aspect, the present invention provides a method for identifying underground cables, comprising the following steps: S1: Configure an orthogonal multi-coil sensing array to synchronously acquire the three-dimensional magnetic induction intensity in the offset position space near the center point directly above the target cable. M The physical channel signal is used to construct a three-dimensional magnetic field gradient tensor based on the three-dimensional magnetic induction intensity. G Solve the gradient tensor in real time G The normalized lateral decay gradient features are obtained. F grad .

[0025] This embodiment employs an orthogonal multi-coil array architecture that eliminates the need for dynamic displacement in its underlying physical sensing method. The active alternating current injected into the target cable... I The alternating magnetic induction intensity generated in the surrounding space B , and radial distance r Inversely proportional, its fundamental spatial equation satisfies: ,in Given the vacuum permeability, if the cable burial depth is set to... h The lateral offset of the detection plane is x Then the radial distance of this space is To accurately capture this attenuation characteristic, the orthogonal multi-coil sensing array provided here includes... M Each triaxial magnetic field sensor node is rigidly fixed in a cross-shaped or star-shaped orthogonal topology, and the three-dimensional spatial relative spacing Δ between the triaxial magnetic field sensor nodes is [missing information]. x , △ y , △ z The calibration is performed and the result is fixed as a constant. In one specific embodiment, the three-dimensional spatial relative spacing Δ between the nodes of the triaxial magnetic field sensor is... x , △ y , △ z The preferred value range is 5 cm to 30 cm. The numerical boundary of the relative spacing in three-dimensional space is determined by cross-optimization of spatial finite difference truncation error and sensor hardware signal-to-noise ratio: the lower limit of 5 cm is limited by the resolution and inherent background noise of the physical magnetic field sensor, and it is necessary to ensure that the difference in magnetic field strength generated by lateral offset can be effectively picked up and is not affected by the cross magnetic field of adjacent sensor devices; the upper limit of 30 cm is limited by the mathematical truncation error of Taylor expansion and the mechanical rigidity of the array structure, and it is necessary to ensure that under the normal burial depth of the target cable, the discrete difference quotient between the two points can approximate the ideal continuous spatial derivative with high precision, thereby ensuring that the spatial tensor matrix can objectively and realistically reflect the microscopic physical curvature of the magnetic field and avoid introducing nonlinear distortion based on finite difference approximation calculation.

[0026] In the specific implementation of the detection, the detector only needs to be placed statically on the ground to be measured or perform a single linear push-broom along the pipeline route. The sensing array can synchronously and without phase difference capture the three-dimensional magnetic induction intensity components in the space of the center point directly above the target cable and its surrounding lateral and longitudinal offset positions within the same microsecond-level sampling clock cycle. Define the magnetic field vector. The vector components along the three orthogonal directions of the spatial rectangular coordinate system are: .

[0027] In acquiring the synchronous output of an orthogonal multi-coil sensing array MAfter receiving the physical channel signal, the spatial rate of change of the magnetic field is calculated in real time using a spatial finite difference algorithm. x Taking the partial derivative of the magnetic field normal gradient along the axial direction as an example, its discrete difference calculation formula can be expressed as: ,in express z Magnetic field component along the axial direction x Spatial gradient in the axial direction, Indicates the horizontal coordinate x place z axial magnetic field component measurement values Indicates the horizontal coordinate x The location was offset. Distance z The measured values ​​of the axial magnetic field components represent the physical calibration spacing of the sensor nodes. Based on this multidimensional spatial difference principle, the system combines data from the central reference node and each orthogonal offset node to construct a three-dimensional magnetic field gradient tensor. G This tensor is mathematically represented as a 3×3 Jacobian matrix, containing the magnetic field vector. Vector components along the three orthogonal directions of the spatial rectangular coordinate system The full-order space partial derivatives, This tensor matrix, utilizing multi-node difference operations, can characterize the curvature and topological morphology of the superimposed electromagnetic field underground at a microscopic scale. This physical tensor construction helps distinguish the spatial morphological differences between the target source and the interference source. As the active loop directly driven by the transmitter, the target cable has a highly concentrated magnetic field radiation source; therefore, near the area directly above the target, the eigenvalues ​​of this tensor matrix exhibit a property consistent with 1 / r The attenuation model exhibits narrow-peak gradient characteristics. In contrast, passive induction signals on nearby non-target cables often show a certain degree of divergence due to secondary scattering from surrounding multi-conductors, resulting in a relatively flat magnetic field distribution curve or a distorted broad peak.

[0028] Real-time solution of the three-dimensional magnetic field gradient tensor G The trace or principal eigenvalue, and the three-dimensional magnetic field gradient tensor G The trace or principal eigenvalue is mapped to a normalized lateral decay gradient feature. F grad Its characteristic definition can be calculated as the attenuation rate at a specific lateral offset, i.e. ,in To represent the lateral offset distance of the probe as The amplitude of the magnetic field detected at that location, This represents the magnetic field amplitude detected at the peak point, typically directly above the pipeline. Through this hardware-based spatial transient topology mapping, a set of spatial dimensional feature data streams with objective physical support can be output, providing fundamental data support for subsequent joint blind source separation and multidimensional map identification.

[0029] S2: Yes M The signals of each physical channel are decomposed separately, so that each physical channel signal is decomposed in the frequency domain into... K Each intrinsic mode component.

[0030] Specifically, the content involves constructing a virtual observation channel based on adaptive frequency domain decomposition, using envelope entropy as the evaluation index, and employing a heuristic algorithm to dynamically optimize the variational mode decomposition (VMD) parameters. The optimized VMD parameters are then used to... M The physical channel signals are decomposed separately.

[0031] The output of the orthogonal multi-coil sensing array M The physical observation channel signals are mixed with non-stationary co-frequency induced interference and environmental noise. To expand the observation dimension, the system performs frequency domain decomposition processing on the signals of each physical observation channel to generate multiple virtual frequency domain observation components, thereby providing additional observation dimensions for subsequent joint observation matrix construction. In a preferred embodiment, the frequency domain decomposition can be implemented using the variational mode decomposition method. This method can transform the original observation signal, i.e. M The physical channel signal is decomposed into several intrinsic mode components with finite bandwidth to reduce mode aliasing and improve the distinguishability of frequency domain components. Its constrained variational model can be expressed as: ,in Represents the decomposed first k Each intrinsic mode component k =1,2,……, K , Indicates the first k The center frequencies corresponding to each intrinsic mode component Represents the Dirac impulse function. Indicates time t Find the partial derivative. j The model uses the imaginary unit (*) and the convolution operator (*) to deconstruct complex signals in the frequency domain by finding the optimal frequency band distribution for each mode. To adapt to the dynamic and non-stationary electromagnetic environment of underground soil, an adaptive parameter optimization mechanism is introduced before frequency domain decomposition to dynamically determine the construction parameters of the virtual frequency domain observation channels. This adaptive parameter optimization mechanism can use evaluation indicators reflecting signal sparsity, periodicity, or energy concentration to assess the decomposition effect under different parameter combinations.

[0032] In a preferred embodiment, envelope entropy can be used as an evaluation index for decomposition performance. Envelope entropy reflects the sparsity and periodicity of the decomposition results at the information science level. When a certain modal component contains a relatively pure periodic target signal with minimal noise aliasing, its evaluation result tends to be better. The following explanation uses envelope entropy as an optional evaluation index: After this adaptive decomposition operation, each physical space signal is analyzed in the frequency domain as... K The system has eigenmode components with different center frequencies. Through this dynamic optimization mechanism based on envelope entropy feedback, the system can... M The one-dimensional physical channel signal is mapped and extended in the frequency domain to generate a total M × K The virtual frequency domain observation components of the path. This parameterized virtual dimensionality-up operation at the frequency domain level creates data conditions for subsequently constructing a high-dimensional joint observation space and satisfying the channel number constraint for blind source separation.

[0033] S3: Yes M physical paths and the generation of each path K Tensor fusion of the intrinsic modal components is performed to construct a high-dimensional spatiotemporal joint observation matrix. X ST joint spatiotemporal observation matrix X ST A fast independent component analysis algorithm based on maximizing negative entropy is employed to analyze the spatiotemporal joint observation matrix. X ST The process maps out independent source signal components, and extracts the pure intrinsic active signal of the target from the output source signal components. s ( t ).

[0034] After completing the variational mode decomposition based on parameter adaptive optimization, the underlying layer obtains the data from... M physical channel extension generation M × K The virtual frequency domain observation components of the path. Based on the mathematical prerequisites for solving Independent Component Analysis (ICA), the number of effective linearly independent observation channels in the system must be greater than or equal to the number of unknown independent source signals. To satisfy the full-rank constraint of the blind source unmixing matrix, the system performs a space-time joint blind source separation ST-ICA operation, concatenating the physical space signal synchronously output by the orthogonal multi-coil array with the previously adaptively decomposed eigenmode functions with finite bandwidth. M The physical channel of the path and the generation of each physical channel K Tensor fusion of the intrinsic modal components is performed to construct a high-dimensional spatiotemporal joint observation matrix. X ST The matrix structure is represented as: In the space-time joint observation matrixX ST In the left-hand side, the dimension of the item is ( M × K )× N , M The total number of sensor nodes in the physical sensing array; K This represents the number of intrinsic mode components obtained after optimized variational mode decomposition (A-VMD) for each physical channel. N This represents the total number of discrete sampling points within the current analysis time window. Indicates the first m The first physical space channel decomposed into the first k A sequence of intrinsic mode components, m =1,2,……, M This joint reconstruction in the spatial and frequency domains directly alters the mathematical solvability boundary of the blind source separation process. This is achieved by expanding the number of effective linearly independent observation channels of the system to... M × K The system state changes from an underdetermined state to a well-determined or overdetermined state, satisfying the basic constraint of the independent component analysis algorithm on the number of channels and ensuring the full rank characteristic of the observation matrix.

[0035] joint observation matrix in space and time X ST After construction, the FastICA algorithm based on maximizing negative entropy is invoked to perform iterative demixing of independent source signals. The processing flow first involves the spatiotemporal joint observation matrix... X ST Mean centering and whitening preprocessing are implemented, and the correlation between the joint observation components is eliminated by eigenvalue decomposition, so that the covariance matrix of the mixed signal is transformed into an identity matrix, thereby compressing the search space of the subsequent unmixing matrix into the orthogonal matrix category.

[0036] Subsequently, blind source extraction is carried out based on the central limit theorem: due to the linear mixing of multiple independent non-Gaussian physical signals, their probability density distribution must tend towards a Gaussian distribution; conversely, if a specific projection vector can be found to maximize the non-Gaussianity of the separated signal, it is equivalent to stripping the independent original physical source components from the mixed electromagnetic field. Negative entropy is selected as the objective function to measure the non-Gaussianity of the separated signal, and its approximate calculation expression can be expressed as: ,in This represents the calculated negative entropy measure. y This represents the independent components of the single-channel signal separated by the demixing matrix projection. Let be a standard Gaussian random variable with zero mean and unit variance. EG represents the expected value, and is the selected non-quadratic nonlinear smoothing function. In this embodiment, the derivative form of the hyperbolic secant function or the Gaussian function is used to enhance the algorithm's robustness to impulse noise. The orthogonal unmixing matrix is ​​updated using Newton's iteration method to maximize the objective function, thereby maximizing the spatiotemporal joint observation matrix. X ST Mapped to independent source signal components, and combined with the pre-defined prior coded pulse or modulation envelope characteristics of the target cable transmitter, correlation matching is performed on the independent source signal components to obtain the pure intrinsic active signal of the target after eliminating co-frequency induced interference. s ( t ).

[0037] S4: From the reconstructed target intrinsic active signal s ( t After that, the lateral decay gradient feature from step S1 is inherited. F grad ; for the target's intrinsic active signals s ( t Perform a Hilbert transform to extract the instantaneous phase, and calculate the standard deviation of the instantaneous phase within a set discrete sliding window. F phase ; for the target's intrinsic active signals s ( t Perform a Fast Fourier Transform to calculate the harmonic distortion rate of the spectrum. F HD ; lateral decay gradient features F grad Standard deviation of instantaneous phase F phase Harmonic distortion rate of the spectrum F HD The three features are concatenated according to the time step to obtain a three-dimensional heterogeneous feature flow matrix that evolves continuously over time. F ( t ).

[0038] After the system reconstructs the target intrinsic source signal through demixing using a joint observation matrix, a multidimensional current fingerprint feature system is further constructed based on electromagnetic propagation and coupling mechanisms. It should be noted that these multidimensional features are not arbitrarily selected superpositions of conventional features, but rather orthogonal feature domains derived from different physical mechanisms: spatial dimension features characterize the spatial topological distribution of the field source, temporal dimension features characterize the difference in phase stability between the active and inductive loops, and frequency dimension features characterize the distortion law of the energy spectrum distribution under electromagnetic coupling conditions. These features together constitute an identification feature system with a clear physical origin.

[0039] In the spatial dimension, the normalized lateral decay gradient feature obtained earlier is directly inherited and synchronously invoked.F grad This spatial characteristic quantity, from the perspective of macroscopic geometric topology, locks in the physical property that the target magnetic field conforms to the radial inverse attenuation distribution, and constitutes the first dimension of the basic data of the characteristic flow matrix.

[0040] In the time dimension, the system extracts the instantaneous phase stability of the source signal as the second-dimensional time fingerprint. The phase of the active signal flowing in the target cable is controlled by the transmitter's clock phase-locked loop, exhibiting high temporal consistency and stability; while the passive induced signal on the adjacent cable, during long-distance underground transmission, is affected by the combined effects of pipeline-to-ground distributed capacitance, insulation dielectric loss, and complex mutual inductance nonlinear time delay, resulting in high-frequency jitter and irregular drift in its microscopic instantaneous phase. The reconstructed one-dimensional time-domain signal, i.e., the target intrinsic active signal, is then analyzed using the Hilbert transform. s ( t The Hilbert transform is then analyzed. The continuous-time mathematical formula for the Hilbert transform can be expressed as: , τ As the variable of integration, in this integral expression, Indicates the intrinsic active signal of the target s ( t The imaginary part obtained after the Hilbert transform is used to construct an analytic signal in the complex domain. And thereby extract the instantaneous phase function of the signal. To quantify the degree of micro-jitter in this phase, a discrete-time sliding window is set, and the standard deviation of the instantaneous phase within this time window is calculated. F phase : The number of discrete sampling points within the set discrete sliding window is W , respectively obtain the first w Instantaneous phase value at each discrete sampling point and the instantaneous phase average within the discrete sliding window , w The sequence number of the discrete sampling point. w =1, 2, ..., W The summation is performed by taking the squares of the differences between the instantaneous phase values ​​at all sampling points and the average instantaneous phase value within the discrete sliding window, and then dividing the sum by the number of discrete sampling points. W The standard deviation of the instantaneous phase can be obtained by taking the square root. F phase The difference in evolutionary stability between active current and induced current on a microscopic time scale was objectively quantified.

[0041] In the frequency dimension, harmonic distortion rate is extracted as the third-dimensional frequency fingerprint. From the perspective of electromagnetic field frequency domain coupling mechanism analysis, the mutual inductance coupling effect between underground parallel pipelines is equivalent to a complex nonlinear low-pass or band-pass filter in the frequency domain. There is a physical difference in the energy transmission attenuation ratio of the fundamental wave and each high-frequency harmonic between actively applied signals and passively induced signals. For the source signal within the current time window, i.e., the target intrinsic active signal... s ( t Perform a Fast Fourier Transform (FFT) to transform the signal to the frequency domain to obtain the discrete amplitude spectrum of the signal, and calculate the high-frequency harmonic distortion rate accordingly. F HD : ,in A 1 represents the amplitude of the target fundamental wave in the frequency domain amplitude spectrum. A n Indicates the first n The amplitude of the second high frequency harmonics, H This is the preset highest harmonic order for analysis.

[0042] After quantifying the physical properties in the three dimensions mentioned above, these three types of heterogeneous features are concatenated as column vectors according to the time step to construct a three-dimensional heterogeneous feature flow matrix that evolves continuously over time. By constructing a multidimensional physical feature map, the spatial curvature, temporal jitter, and frequency domain distortion of the signal are deeply algebraically fused, providing a complete decision input matrix for subsequent time-series intelligent classification networks.

[0043] S5: Continuously obtained L Three-dimensional heterogeneous feature flow matrix F ( t Perform data serialization construction based on a sliding time window, with a construction dimension of 3× L Temporal input tensor X seq Translate the time series into the tensor X seq Input the temporal convolutional network (TCN) for deep feature extraction and classification, and obtain the confidence probability that the detected object belongs to the target cable within the current sliding time window.

[0044] Obtaining the 3D heterogeneous feature flow matrix F ( t Afterwards, data serialization based on a sliding time window is performed to generate the two-dimensional input tensor of the Temporal Convolutional Network (TCN). The time observation window and corresponding sliding step size are set to continuously acquire the three-dimensional heterogeneous feature flow matrix. F ( t The corresponding tensors are concatenated to construct a 3× dimension. L Temporal input tensor X seqThis temporal input tensor continuously carries, at the physical information level, the joint dynamic evolution trajectory of the target magnetic field over a macroscopic time span, encompassing spatial gradient decay, microscopic phase jitter, and frequency distortion characteristics. The constructed temporal features are then input into the tensor. X seq A temporal convolutional network is fed into the system for deep feature mining. The core component of this network topology is a dilated causal convolutional layer. The causal convolution mechanism restricts the network to perform convolution operations only on the input data of the current time step and its historical moments when calculating the hidden state at each time step. This aligns with the objective premise that the feature sequence of the online detection system does not possess the ability to predict future information. To expand the model's historical receptive field for temporal features within a limited network depth, a dilation factor is introduced into the convolutional kernel. d For the input feature sequence and the spatial size is k The discrete operation expression for a one-dimensional convolution filter with 0 and dilated causal convolution can be expressed as: , This represents the node value of the output feature tensor after dilated convolution at a specific time step. In a one-dimensional convolution kernel, the first... i One weight parameter, This indicates that the input tensor is strided. d Historical time-step feature data is sampled at intervals. An expansion factor is then applied to the stacked hidden layers. d The sequence is set to increase exponentially, for example, according to the sequence 1, 2, 4, 8, so that the Temporal Convolutional Network (TCN) can extract deep features of long-term temporal evolution patterns while maintaining high feature temporal resolution.

[0045] To overcome gradient decay caused by multi-layer network stacking and improve the robustness of feature mapping, a residual connection module is introduced in the Temporal Convolutional Network (TCN). The residual module contains two parallel forward computation paths: the main path performs the dilated causal convolution, weight normalization, and non-linear activation processing described above; the bypass connection directly and identically maps the module's input tensor to the output, adding it element-wise with the output of the main path. The mathematical model of this residual connection process can be expressed as: ,in Indicates the first l The combined output feature tensor of each residual module This represents the input feature tensor of the previous hidden layer. This represents a composite nonlinear residual mapping function that includes multiple layers of dilated convolution and normalization operations in the main path; This is the set of learnable weight parameters within the residual module. For non-linear activation functions, the ReLU function is typically chosen. Through this residual topology, the model can smoothly filter isolated feature mutations caused by sudden impulse interference during network forward propagation, preserving low-frequency temporal features that reflect the physical essence of the target. After deep feature extraction through multi-level residual modules, the network's end uses a global average pooling layer to compress the high-dimensional temporal feature map in time dimension, and then feeds it into a fully connected layer to perform the final classification probability mapping. The Sigmoid function is used to output the confidence probability that the detected object belongs to the target cable within the current sliding time window; its non-linear mapping formula is expressed as: In the above formula, P This represents the output confidence probability, and its value range is normalized to 0 to 1. This represents the weight matrix parameters of the fully connected layer; This is a one-dimensional deep feature vector output after the global pooling operation; These are the bias parameters for the classification and decision layer. To prevent frequent jumps in recognition results when in a critical electromagnetic state, the system's decision module is configured with a hysteresis comparison decision strategy: a high confidence threshold is pre-set. P high and low confidence threshold P low When the confidence probability P ≥ High confidence threshold P high When the current cable is the target cable, output a confirmation feedback signal; when the confidence probability is... P ≤low confidence threshold P low When the current cable is determined to be a nearby interfering cable, the confidence probability is... P At high confidence threshold P high and low confidence threshold P low When the two conditions are in between, the judgment state of the previous sliding time window is maintained, waiting for the input of newly collected time-series characteristics for further judgment. This dynamic decision-making mechanism combines the output of deep networks with engineering logic strategies, completing the entire technical implementation process from the perception of the underlying physical data array to the intelligent hierarchical classification at the top level.

[0046] In one specific embodiment, a pre-set high confidence threshold P high The value ranges from 0.85 to 0.95, representing the low confidence threshold. P low The value ranges from 0.15 to 0.30, and both must satisfy the hysteresis safety margin constraint. P high - P low ≥0.6The aforementioned threshold range is set based on data-driven optimization to strictly control safety risks in industrial settings: the model's operating point is shifted towards the extremely low false positive rate (FPR) region using the Receiver Operating Characteristic (ROC) curve, and a high confidence threshold is derived in reverse to ensure high reliability and "zero false alarms" in target confirmation commands under complex electromagnetic environments. Simultaneously, a low confidence threshold is established by forcibly setting a wide hysteresis range greater than or equal to 0.6, effectively eliminating frequent reversals of classification commands and high-frequency oscillations of underlying hardware control relays under critical electromagnetic conditions, such as when transient high-energy electromagnetic pulses penetrate. This dual-threshold hysteresis mechanism endows the dynamic decision-making algorithm with extremely high fault tolerance and anti-interference / anti-shake capabilities from an engineering implementation perspective.

[0047] On the other hand, the present invention also provides an underground cable identification system for implementing the above-mentioned method, comprising: The lateral attenuation gradient feature acquisition unit, through a pre-configured orthogonal multi-coil sensing array, synchronously acquires the three-dimensional magnetic induction intensity in the offset position space near the center point directly above the target cable. M Physical channel signals are used to construct a three-dimensional magnetic field gradient tensor. G Solve the three-dimensional magnetic field gradient tensor G Obtain the normalized lateral decay gradient features F grad ; The target intrinsic active signal acquisition unit is used for... M The signals of each physical channel are decomposed separately, so that each physical channel signal is decomposed in the frequency domain into... K Each intrinsic mode component M physical channel signal and K Tensor fusion of the intrinsic modal components is performed to construct a high-dimensional spatiotemporal joint observation matrix. X ST joint spatiotemporal observation matrix X ST A fast independent component analysis algorithm based on maximizing negative entropy is employed to analyze the spatiotemporal joint observation matrix. X ST The process maps out independent source signal components, and extracts the pure intrinsic active signal of the target from the output source signal components. s ( t ); A three-dimensional heterogeneous feature flow matrix generation unit is used to generate intrinsic active signals of the target. s ( t Perform a Hilbert transform to extract the instantaneous phase, and calculate the standard deviation of the instantaneous phase within a set discrete sliding window. F phase ; for the target's intrinsic active signals s ( tPerform a Fast Fourier Transform to calculate the harmonic distortion rate of the spectrum. F HD The lateral decay gradient feature is obtained from the output of the lateral decay gradient feature acquisition unit. F grad Standard deviation of instantaneous phase F phase Harmonic distortion rate of the spectrum F HD The three features are concatenated according to the time step to obtain a three-dimensional heterogeneous feature flow matrix that evolves continuously over time. F ( t ); The target cable determination unit is used to obtain the three-dimensional heterogeneous feature flow matrix output by the three-dimensional heterogeneous feature flow matrix generation unit. F ( t ), and will obtain continuously L Three-dimensional heterogeneous feature flow matrix F ( t Perform data serialization construction based on a sliding time window, with a construction dimension of 3× L Temporal input tensor X seq Translate the time series into the tensor X seq Input the temporal convolutional network (TCN) for deep feature extraction and classification, and obtain the confidence probability that the detected object belongs to the target cable within the current sliding time window.

[0048] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying underground cables, characterized in that, Includes the following steps: S1: Configure an orthogonal multi-coil sensing array to synchronously acquire the three-dimensional magnetic induction intensity in the offset position space near the center point directly above the target cable. M The physical channel signal is used to construct a three-dimensional magnetic field gradient tensor based on the three-dimensional magnetic induction intensity. G Solve the gradient tensor in real time G The normalized lateral decay gradient features are obtained. F grad ; S2: Yes M The signals of each physical channel are decomposed separately, so that each physical channel signal is decomposed in the frequency domain into... K One intrinsic mode component; S3: Yes M physical paths and the generation of each path K Tensor fusion of the intrinsic modal components is performed to construct a high-dimensional spatiotemporal joint observation matrix. X ST joint spatiotemporal observation matrix X ST A fast independent component analysis algorithm based on maximizing negative entropy is employed to analyze the spatiotemporal joint observation matrix. X ST The process maps out independent source signal components, and extracts the pure intrinsic active signal of the target from the output source signal components. s ( t ); S4: From the reconstructed target intrinsic active signal s ( t After that, the lateral decay gradient feature from step S1 is inherited. F grad ; for the target's intrinsic active signals s ( t Perform a Hilbert transform to extract the instantaneous phase, and calculate the standard deviation of the instantaneous phase within a set discrete sliding window. F phase ; for the target's intrinsic active signals s ( t Perform a Fast Fourier Transform to calculate the harmonic distortion rate of the spectrum. F HD ; lateral decay gradient features F grad Standard deviation of instantaneous phase F phase Harmonic distortion rate of the spectrum F HD The three features are concatenated according to the time step to obtain a three-dimensional heterogeneous feature flow matrix that evolves continuously over time. F ( t ); S5: Continuously obtained L Three-dimensional heterogeneous feature flow matrix F ( t Perform data serialization construction based on a sliding time window, with a construction dimension of 3× L Temporal input tensor X seq Translate the time series into the tensor X seq Input the temporal convolutional network (TCN) for deep feature extraction and classification, and obtain the confidence probability that the detected object belongs to the target cable within the current sliding time window.

2. The method for identifying underground cables according to claim 1, characterized in that, The orthogonal multi-coil sensing array mentioned in step S1 includes M Each triaxial magnetic field sensor node is rigidly fixed in a cross-shaped or star-shaped orthogonal topology, and the three-dimensional spatial relative spacing Δ between the triaxial magnetic field sensor nodes is [missing information]. x , △ y , △ z It is calibrated and solidified as a constant.

3. The method for identifying underground cables according to claim 1, characterized in that, The construction of the three-dimensional magnetic field gradient tensor based on the three-dimensional magnetic induction intensity described in step S1 G The ratio of the magnetic field amplitude at the peak point directly above the target cable to that at the lateral offset point is calculated to obtain the normalized lateral attenuation gradient feature. F grad , is the definition of the magnetic field vector Vector components along the three orthogonal directions of the spatial rectangular coordinate system The full-order spatial partial derivatives are used as the three-dimensional magnetic field gradient tensor. G Solving for the three-dimensional magnetic field gradient tensor G The trace or principal eigenvalue, and the three-dimensional magnetic field gradient tensor G The trace or principal eigenvalue is mapped to a normalized lateral decay gradient feature. F grad .

4. The method for identifying underground cables according to claim 3, characterized in that, Step S2 involves using envelope entropy as the evaluation metric and employing a heuristic algorithm to dynamically optimize the variational mode decomposition (VMD) parameters; then using the optimized VMD parameters to... M The physical channel signals are decomposed separately.

5. The method for identifying underground cables according to claim 4, characterized in that, In step S3, the pull M physical paths and the generation of each path K Tensor fusion of the intrinsic modal components is performed to construct a high-dimensional spatiotemporal joint observation matrix. X ST It is the process of performing a joint space-time blind source separation operation to synchronously output the three-dimensional magnetic induction intensity from an orthogonal multi-coil sensing array. M physical channel signal, and K The number of effective linearly independent observation channels is expanded by matrix concatenation of the intrinsic mode components. M × K One, to ensure the full-rank property of the observation matrix.

6. The method for identifying underground cables according to claim 5, characterized in that, The joint space-time observation matrix mentioned in step S3 X ST A fast independent component analysis algorithm based on maximizing negative entropy is employed to analyze the spatiotemporal joint observation matrix. X ST The process maps out independent source signal components, and extracts the pure intrinsic active signal of the target from the output source signal components. s ( t (This refers to the construction of a joint spatiotemporal observation matrix.) X ST Subsequently, after mean-neutralization and whitening preprocessing, eigenvalue decomposition is performed to compress the search space of the subsequent unmixing matrix into the orthogonal matrix category. Then, blind source extraction is performed based on the central limit theory, and negative entropy is selected as the objective function to measure the non-Gaussianity of the separated signal. The orthogonal unmixing matrix is ​​updated using the Newton-Raphson iteration method to maximize the objective function, thereby maximizing the spatiotemporal joint observation matrix. X ST Mapped to independent source signal components, and combined with the pre-defined prior coded pulse or modulation envelope characteristics of the target cable transmitter, correlation matching is performed on the independent source signal components to obtain the pure intrinsic active signal of the target after eliminating co-frequency induced interference. s ( t ).

7. The method for identifying underground cables according to claim 6, characterized in that, Step S4 describes calculating the standard deviation of the instantaneous phase within a set discrete sliding window. F phase ; for the target's intrinsic active signals s ( t Perform a Fast Fourier Transform to calculate the harmonic distortion rate of the spectrum. F HD This means setting the number of discrete sampling points within the discrete sliding window to be... W , respectively obtain the first w The instantaneous phase value at each discrete sampling point and the average instantaneous phase value within the discrete sliding window. w The sequence number of the discrete sampling point. w =1, 2, ..., W The summation is performed by taking the squares of the differences between the instantaneous phase values ​​at all sampling points and the average instantaneous phase value within the discrete sliding window, and then dividing the sum by the number of discrete sampling points. W The back square root is used as the standard deviation of the instantaneous phase. F phase ; for the target's intrinsic active signals s ( t After performing a Fast Fourier Transform, let A 1 represents the amplitude of the target fundamental wave in the frequency domain amplitude spectrum. A n Indicates the first n The amplitude of the second-highest frequency harmonic is calculated by summing the squares of the amplitudes of the second-highest frequency harmonics and taking the square root, then dividing by the amplitude of the target fundamental frequency. A 1. Obtain the harmonic distortion rate of the spectrum. F HD .

8. The method for identifying underground cables according to claim 7, characterized in that, The step S5 involves converting the temporal input tensor... X seq Deep feature extraction and classification are performed on the input Temporal Convolutional Network (TCN). This involves configuring multi-level residual connection modules within the TCN, each containing dilated causal convolutional layers. These dilated causal convolutional layers are used to compute the hidden state at each time step, only considering the state at that time step. t The input data at historical moments are subjected to dilated causal convolution operations. A dilation factor is introduced into the convolution kernel to expand the historical receptive field of temporal features. The dilation factor is set to increase exponentially. The residual connection module includes two parallel forward computation paths. One is the main path, which is used to perform dilated causal convolution, weight normalization and nonlinear activation processing. The other is the side path, which is used to directly map the input tensor of the residual connection module to the output end and add it element-wise with the output of the main path.

9. The method for identifying underground cables according to claim 8, characterized in that, Step S5, which involves obtaining the confidence probability that the detected object within the current sliding time window belongs to the target cable, involves configuring a global average pooling layer and a fully connected layer. The global average pooling layer is placed at the end of the multi-level residual connection module and is used to compress the high-dimensional temporal feature map extracted by the multi-level residual connection module in the time dimension, generating a one-dimensional deep feature vector. H out The data is then fed into a fully connected layer to perform classification probability mapping and output confidence probabilities. P ; Preset high confidence threshold P high and low confidence threshold P low When the confidence probability P ≥ High confidence threshold P high When the current cable is the target cable, output a confirmation feedback signal; when the confidence probability is... P ≤low confidence threshold P low When the current cable is determined to be a nearby interfering cable, the confidence probability is... P At high confidence threshold P high and low confidence threshold P low When the two are in between, the judgment state of the previous sliding time window is maintained, and the system waits for the input of newly collected timing characteristics for further judgment.

10. An underground cable identification system for implementing the method as described in any one of claims 1-9, characterized in that, include: The lateral attenuation gradient feature acquisition unit, through a pre-configured orthogonal multi-coil sensing array, synchronously acquires the three-dimensional magnetic induction intensity in the offset position space near the center point directly above the target cable. M Physical channel signals are used to construct a three-dimensional magnetic field gradient tensor. G Solve the three-dimensional magnetic field gradient tensor G Obtain the normalized lateral decay gradient features F grad ; The target intrinsic active signal acquisition unit is used for... M The signals of each physical channel are decomposed separately, so that each physical channel signal is decomposed in the frequency domain into... K Each intrinsic mode component M physical channel signal and K Tensor fusion of the intrinsic modal components is performed to construct a high-dimensional spatiotemporal joint observation matrix. X ST joint spatiotemporal observation matrix X ST A fast independent component analysis algorithm based on maximizing negative entropy is employed to analyze the spatiotemporal joint observation matrix. X ST The process maps out independent source signal components, and extracts the pure intrinsic active signal of the target from the output source signal components. s ( t ); A three-dimensional heterogeneous feature flow matrix generation unit is used to generate intrinsic active signals of the target. s ( t Perform a Hilbert transform to extract the instantaneous phase, and calculate the standard deviation of the instantaneous phase within a set discrete sliding window. F phase ; for the target's intrinsic active signals s ( t Perform a Fast Fourier Transform to calculate the harmonic distortion rate of the spectrum. F HD The lateral decay gradient feature is obtained from the output of the lateral decay gradient feature acquisition unit. F grad Standard deviation of instantaneous phase F phase Harmonic distortion rate of the spectrum F HD The three features are concatenated according to the time step to obtain a three-dimensional heterogeneous feature flow matrix that evolves continuously over time. F ( t ); The target cable determination unit is used to obtain the three-dimensional heterogeneous feature flow matrix output by the three-dimensional heterogeneous feature flow matrix generation unit. F ( t ), and will be obtained continuously L Three-dimensional heterogeneous feature flow matrix F ( t Perform data serialization construction based on a sliding time window, with a construction dimension of 3× L Temporal input tensor X seq Translate the time series into the tensor X seq Input the temporal convolutional network (TCN) for deep feature extraction and classification, and obtain the confidence probability that the detected object belongs to the target cable within the current sliding time window.

Images