Power distribution network line fault positioning and detecting system

Abstract

The invention discloses a power distribution network line fault positioning detection system, and relates to the technical field of power distribution network fault detection. The system comprises a mixed information acquisition layer, a fault feature extraction layer, an intelligent diagnosis layer and a fault positioning layer. The mixed signal acquisition layer comprises a high-frequency transient wave recording unit, a power frequency measurement unit, a wireless pulse sensor and a distributed optical fiber temperature measurement unit; the fault feature extraction layer comprises a time-frequency analysis module, a preprocessing module and a three-dimensional feature vector module; the intelligent diagnosis layer comprises a convolutional attention network, a space-time diagram neural network and a transfer learning module; the fault positioning layer comprises a particle swarm module and a fuzzy reasoning module. According to the invention, data information of the cable is acquired through the mixed information acquisition layer, a video analysis window function is dynamically matched with signal characteristics, a time domain graph scale, a frequency domain resonance component and a space field intensity gradient are constructed, fault diagnosis and positioning are carried out by using the intelligent diagnosis layer, and the fault positioning detection efficiency of the power distribution network is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of fault detection, and in particular relates to a distribution network line fault location detection system. Background Art

[0002] Currently, distribution network fault location faces three major technical bottlenecks: (1) Insufficient time-frequency analysis accuracy: The traditional STFT method is limited by a fixed window function and is difficult to take into account both high-frequency transient and low-frequency steady-state characteristics. Although the wavelet transform has multi-resolution characteristics, the selection of basis functions relies on experience, resulting in incomplete fault feature extraction.

[0003] (2) Lack of topological association: Existing methods mostly use isolated node analysis and ignore the propagation path of fault current in the distribution network.

[0004] (3) Poor adaptability across power grids: Deep learning models need to be retrained in new topologies.

[0005] In view of the above problems, the present invention provides a distribution network line fault location detection system. Summary of the Invention

[0006] The purpose of the present invention is to provide a distribution network line fault location detection system, which collects cable data information through a hybrid information acquisition layer, dynamically matches the video analysis window function with the signal characteristics, constructs the time domain graph scale, frequency domain resonant component and spatial field intensity gradient, and uses the intelligent diagnosis layer to perform fault diagnosis and location, thereby solving the problems of insufficient time-frequency analysis accuracy and lack of topological association in existing systems.

[0007] To solve the above technical problems, the present invention is achieved through the following technical solutions: The present invention is a distribution network line fault location detection system, comprising a hybrid information acquisition layer, a fault feature extraction layer, an intelligent diagnosis layer and a fault location layer; The mixed signal acquisition layer includes a high-frequency transient recording unit (sampling rate ≥ 1MHz), a power frequency measurement unit, a wireless pulse sensor, and a distributed optical fiber temperature measurement unit; the high-frequency transient recording unit is used to record the microsecond transient process of the voltage / current traveling wave and capture characteristic parameters; the power frequency measurement unit is used to synchronously measure the effective value of the power frequency waveform set of the three-phase voltage / current, including the dynamic change process of the positive sequence, negative sequence, and zero sequence components; the wireless pulse sensor is used to capture the spatial electromagnetic field distortion; and the distributed optical fiber temperature measurement unit is used to measure the conductor temperature in real time; The fault feature extraction layer includes a time-frequency analysis module, a preprocessing module, and a three-dimensional feature vector module; the time-frequency analysis module is used to dynamically match the signal features through the video analysis window function; the preprocessing module is used to preprocess the mixed signal to eliminate common mode interference; the three-dimensional feature vector module is used to construct the time domain graph scale, frequency domain resonant component, and spatial field intensity gradient; The intelligent diagnosis layer includes a convolutional attention network, a spatiotemporal graph neural network, and a transfer learning module; the convolutional attention network is used to perform preliminary fault type screening; the spatiotemporal neural network is used to build line topology associations; and the transfer learning module is used to adapt to different distribution network structures. The fault location layer includes a particle swarm module and a fuzzy reasoning module; the particle swarm module is used to locate faults on the distribution line through a particle swarm algorithm; and the fuzzy reasoning module is used to process uncertain information.

[0008] As a preferred technical solution, the high-frequency transient recording unit includes a multi-channel synchronous sampling module, an adaptive filtering circuit, and a synchronous clock module. The multi-channel synchronous sampling module is equipped with an 8-channel 16-bit high-precision ADC for synchronous acquisition of voltage / current traveling waves and power frequency signals in three modes, with a sampling rate of ≥50MHz. The adaptive filtering circuit is used to dynamically reduce the noise of microsecond-level transient characteristics such as cable joint discharge and tree barrier flashover. The synchronous clock module is equipped with a GPS / passive dual-mode timing unit, with a time synchronization accuracy of <10ns. The wireless pulse sensor includes a distributed electromagnetic field sensor array, an adaptive power transmission module, and an edge computing module. Each tower in the distributed electromagnetic field sensor array deploys 3-6 UWB nodes, with a measurement bandwidth covering the 300MHz-3GHz frequency band. The adaptive power transmission module dynamically adjusts the transmission power by 0.110mW-10mW based on ambient noise to ensure signal integrity within a 20-meter radius. The edge computing module node has a built-in ARM Cortex-M7 processor, enabling localized raw data preprocessing. The distributed optical fiber temperature measurement unit includes a temperature measurement optical cable, a Raman scattering demodulation module and an OTDR positioning module; the temperature measurement optical cable is laid with anti-bending armored light along the line, with a spatial resolution of ≤1m; the Raman scattering demodulation module uses dual-wavelength laser pulses to measure the temperature of the optical cable, with a temperature measurement accuracy of ±0.5°C; the OTDR positioning module is used to combine optical time domain reflection technology to achieve meter-level positioning of temperature anomalies.

[0009] As a preferred technical solution, the specific workflow of the fault feature extraction layer is as follows: Step G01: Calculate the weighted exponent of the local signal-to-noise ratio and instantaneous frequency in real time, and drive the Gaussian window width to adaptively adjust according to the exponential law; the specific implementation process is as follows: Use a sliding window with a 50% overlap rate to frame the signal, calculate the ratio of signal energy to noise floor energy in each frame, obtain the phase derivative of the analytical signal through Hilbert transform, and construct a weighted function of SNR and frequency , dynamically adjust the Gaussian window based on the exponential output: ; Where SNR is the power ratio of signal to noise, is the instantaneous angular frequency, are the weighting coefficients of the signal-to-noise ratio term and the frequency term, respectively. is the standard deviation of the Gaussian window after dynamic adjustment, is the initial standard deviation reference value of the Gaussian window, k is the adjustment coefficient, is a composite weighting function, and e is a natural constant.

[0010] Step G02: Introduce the quantum genetic algorithm to optimize the function parameters and dynamically search for the optimal time-frequency resolution combination in the range of 0.1ms-10ms; Step G03: A standard Gaussian window is used in the low-frequency band (<1kHz), and a Morlet wavelet kernel is used in the high-frequency band (>10kHz). The transition region is smoothly switched using a Sigmoid function. Step G04: The logarithmic frequency axis and the linear time axis form a hyperbolic coordinate system; the hyperbolic coordinate system includes the time axis and the frequency axis; the time axis needs to maintain a linear sampling interval , the frequency axis uses log2 scale, with a minimum resolution of 0.1Hz; construct hyperbolic coordinates The non-uniform sampling network is used, and the time-frequency surface reconstruction is completed using cubic spline interpolation; Step G05: Refocusing the expanded energy to the real time-frequency trajectory by calculating the instantaneous frequency gradient field; Step G06: Decompose the three-dimensional time-frequency matrix into core tensors and factor matrices to extract cross-channel common features. The three-dimensional time-frequency matrix includes time, frequency, and sensor channels. Stack the multi-channel time-frequency matrices into a three-dimensional tensor (channel, frequency, time), and perform energy normalization on each channel. Use the Tucker decomposition model ;Where G is the core tensor, A is the channel factor matrix, B is the frequency factor matrix, and C is the time factor matrix; During feature extraction, the cross-channel coupling features are analyzed by slicing the core tensor G, the channel commonality pattern is extracted from the factor matrix A, and the energy threshold is set to screen the significant feature components; Step G07: Using adjacent non-fault line signals to establish an interference dictionary library, and perform noise suppression using compressed sensing theory; Step G08: Dynamically adjust the filter strength according to the local signal-to-noise ratio of the signal to retain valid fault features.

[0011] As a preferred technical solution, in step G03, the preset window width parameter matrix , load the time-frequency dictionary of the Morlet wavelet kernel, use a sliding window to calculate the in-band energy of the current analysis frame, estimate the noise floor through the Kalman filter, and define a dynamic signal-to-noise ratio threshold; When a transient pulse (du / dt>5kV / μs) is detected, high-frequency processing is performed to activate The narrow window has a time resolution of ; When the fundamental frequency fluctuation exceeds When the Wide window, frequency resolution reaches ; The Morlet wavelet kernel is designed with nonlinear frequency axis and a logarithmic-linear hybrid scale is constructed. Using linear division, Using logarithmic division, the center frequency is set to To ensure time-frequency balance, introduce bandwidth factor , which increases the frequency resolution of low-frequency area by 40% and the frequency resolution of high-frequency area by Temporal resolution is improved by 30%.

[0012] As a preferred technical solution, in step G05, the specific process of refocusing the expanded energy to the real time-frequency trajectory by calculating the instantaneous frequency gradient field is as follows: Step G051: Perform Sobel operator convolution on the generated time-frequency surface; the convolution kernel is designed as follows: horizontal kernel: [-1, 0, 1; -2, 0, 2; -1, 0, 1], detects vertical edges; vertical kernel: [-1, -2, -1; 0, 0, 0; 1, 2, 1], detects horizontal edges; perform horizontal and vertical convolution on the time-frequency surface matrix respectively to obtain the gradient component and ; then the magnitude of the gradient is: , the direction is: ; Step G052: Perform particle flow simulation on the time-frequency surface along the gradient direction. During the particle flow simulation, particles are generated in the high-energy region of the time-frequency surface (the first 10% of pixels in the amplitude). Particle properties include position, velocity, and mass. The particle motion equation is based on the Stokes drag model: , where gravity Along the gradient direction, the drag force The direction is opposite to the velocity. During the movement, the time step is set and the particle position is updated through Euler iteration. The particle stops when it reaches the local energy minimum or exceeds the time-frequency boundary. Step G053: Record the extreme points of the gradient amplitude along the particle trajectory as candidate points. After screening the candidate points, perform cubic spline interpolation to generate a continuous trajectory. When screening the candidate points, sort them in descending order of amplitude, retain the maximum point M, and eliminate adjacent points with a time-frequency overlap ratio with M greater than 0.5. Repeat the iteration until all candidate points have been processed. Step G054: Perform Gaussian energy redistribution with the trajectory as the center; redistribute the original time-frequency energy according to the Gaussian kernel weight with the trajectory point as the center, while maintaining the conservation of total energy, and perform bilinear interpolation on the redistributed time-frequency matrix to eliminate gridding artifacts.

[0013] As a preferred technical solution, in step G07, the interference dictionary is established using adjacent non-fault line signals, and the specific process of noise suppression is performed using compressed sensing theory as follows: Step G071: Collect signals from adjacent normal routes to construct a complete dictionary D; Step G072: Use K-SVD algorithm to train dictionary atoms; Step G073: Establish a multi-resolution dictionary hierarchy structure; Step G074: Send the signal to be processed , use ISTA algorithm for sparse representation; Step G075: Retain the main sparse components through hard threshold processing, reconstruct the signal and suppress the noise. The reconstructed signal is , adaptively adjust the regularization parameter .

[0014] As a preferred technical solution, the workflow of the intelligent diagnosis layer is as follows: Step Z01: Convolutional attention network performs initial fault screening; Step Z02: The spatiotemporal graph neural network performs topological modeling and outputs the propagation path of potential faults at each node; Step Z03: The transfer learning module performs domain adaptation training and dynamically adjusts the learning rate; In step Z01, parallel 1D-CNN branches are used to process signals of different frequency bands, and the receptive field is expanded through hollow convolution to capture long-range features. The inter-channel correlation matrix is ​​calculated for spatial attention to focus on key sensors. Gate control is used for spatial attention to enhance transient fault features. After feature fusion, the Softmax layer outputs the initial screening fault probability distribution. The specific implementation process is as follows: Design 3-5 parallel 1D-CNN branches, each equipped with a bandpass filter with a different cutoff frequency to preprocess the input signal. Each branch uses a convolution kernel of a specific scale (e.g., 64 / 128 / 256 points) to match the characteristic period of the corresponding frequency band. Introduce dilated convolution at the top network layer to expand the receptive field at an exponential rate (d=1, 2, 4, 8). Use residual connections to alleviate the vanishing gradient problem and preserve features at all levels. By calculating the covariance matrix of multi-channel signals, the principal component vector is obtained through singular value decomposition. The channel attention module (CBAM) is used to generate weight vectors to enhance key sensor features. LSTM units are used to construct time gating and calculate the attention score at each time point. Sigmoid gating is used to control the propagation strength of fault transient features. Adaptive average pooling is performed on feature maps of different scales to unify the dimensions, and a learnable 1×1 convolution is used to achieve cross-scale feature interaction. The fused features are reduced in dimension through a two-layer fully connected network, and finally the probability distribution of each fault type is output through Softmax.

[0015] In step Z02, for the spatial dimension, an initial adjacency graph is constructed based on the line impedance matrix, and Chebyshev polynomials are used to approximate graph filtering. For the temporal dimension, a time window is used to construct a spatiotemporal tensor structure, and dilated causal convolution is used to capture multi-cycle patterns. Adjacent line states are fused through a multi-head attention mechanism to output the potential fault propagation path for each node. The specific implementation process is as follows: Spatial dimension modeling: Calculate the electrical distance between nodes based on the line impedance matrix, generate a weighted adjacency matrix, and use a threshold method to sparsely process the adjacency matrix to retain strong connections; Time dimension modeling: Design sliding time windows with a 50% overlap rate, construct a three-dimensional space-time tensor (channel, frequency, time), and align data streams with different sampling rates using the dynamic time warping (DTW) algorithm; Spatial convolution implementation: Use K-order Chebyshev polynomials to approximate the graph Laplacian operator and design hierarchical aggregation strategies: local neighborhood, global topology, and cross-level feature fusion; Temporal convolution implementation: Use a causal convolution kernel with an exponentially growing expansion rate (d=1, 2, 4, ..., n), and introduce a gating mechanism to control the forgetting rate of historical information.

[0016] Construct query-key-value triples, calculate dynamic association weights between nodes, and use 4-8 heads to capture relationships in different semantic spaces in parallel. Use the gradient backpropagation algorithm to locate key impact paths and output a heat map of node-level failure probability and propagation direction. A dual-stream architecture of space and time enables topology-aware fault diagnosis. Spatial convolution captures the physical constraints of the power grid, while temporal convolution models the dynamic characteristics of fault evolution. For practical deployment, a curriculum learning strategy is recommended, training a static topology first and then gradually introducing dynamic graph complexity.

[0017] As a preferred technical solution, in step Z03, the shallow feature extractors of CAN and STGNN are fixed, and the MMD loss is used to minimize the difference in feature distribution between the source domain and the target domain. The latent space representations of different power grids are aligned through adversarial training. A replay buffer for new and old samples is set to balance historical memory, and the learning rate is dynamically adjusted. The specific implementation process is as follows: The convolution kernel parameters of the first five layers of the Convolutional Attention Network (CAN) are fixed; the weights of the graph convolution layers of the Spatio-Temporal Graph Neural Network (STGNN) remain unchanged; only the trainable flags of the last three fully connected layers are opened; gradient calculations of frozen layers are skipped during backpropagation; and a masking mechanism is used to block optimizer updates of frozen parameters. Calculate the MMD distance between the source domain and target domain features. The specific formula is as follows: ; Where, It is the maximum difference, which is used to measure the difference between two probability distributions X and Y. x and y represent the sample sets of the two distributions to be compared, n and m represent the number of samples of X and Y respectively. For kernel function mapping, the original data is mapped to the regenerated kernel Hilbert space. represents the norm operation in Hilbert space; H represents the Bronk constant.

[0018] Construct a discriminator network to distinguish the source domain of features, and use a gradient reversal layer to make the feature extractor generate domain-invariant representations; automatically adjust the learning rate based on the loss change rate. The specific formula is as follows: ; Where, and They represent the dynamically adjusted learning rate and the initial baseline learning rate at the tth iteration, e is the automatic logarithm base, is the attenuation coefficient hyperparameter, is the absolute value of the gradient of the loss function at the tth iteration; Through the exponential decay mechanism, when the gradient changes drastically ( When , the learning rate Automatically reduce the learning rate to slow down parameter updates and prevent new data from overwriting old knowledge. When the gradient is flat, maintain a high learning rate to continuously absorb new knowledge. This mechanism balances model stability and adaptability.

[0019] As a preferred technical solution, the particle swarm module positioning process is as follows: Step D1: The fault information uploaded by the intelligent diagnosis layer is used as the fault characteristic; Step D2: Encode the fault characteristic and establish a switching function; Step D3: Define the fault information encoding and generate the initial population of the Halton sequence. The Halton sequence can generate a uniformly distributed initial population. By sorting the fitness function values, the optimal population is selected as the initial solution, which helps improve the quality of the algorithm. Step D4: Use the particle swarm algorithm to solve and update the particle swarm position and velocity. When processing the two-dimensional Halton sequence, two prime numbers are used as the basic units. By continuously expanding these two basic prime numbers, a series of evenly distributed and non-repeating points are formed. The specific mathematical model formula is as follows: ; Where, , x is a prime number greater than or equal to 2, , is the Halton sequence function, is the generated two-dimensional uniform sequence; Step D5: Determine whether the number of iterations has been reached and output the location of the fault section; To improve the particle swarm algorithm's global search performance and avoid local extrema, an adaptive dynamic inertia weighting method based on fitness can be used. The adaptive weighting factor ω decreases with the number of iterations. A larger value at the beginning of an iteration enhances global search capabilities. As the iteration progresses, the particle swarm gradually approaches the optimal target, requiring a lower value for local search to improve optimization accuracy.

[0020] As a preferred technical solution, the fuzzy reasoning module workflow is as follows: Step M1: Determine the domain range of input variables (such as voltage fluctuation amplitude ΔU, frequency deviation Δf) and output variables (fault confidence); Step M2: Divide each variable into 3-5 fuzzy subsets (e.g., low / medium / high), and use Gaussian or triangular membership functions to describe the fuzzy intervals; Step M3: Calculate the degree to which the input data belongs to each fuzzy subset through the membership function; Step M4: Construct expert experience rules, automatically extract rule parameters from historical fault data using the FCM clustering algorithm, and set a credibility factor (0-1) for each rule to reflect the reliability of the rule in a specific scenario; Step M5: Calculate the matching degree of the premise part of each rule (take the minimum / product operation of the input membership degree); Step M6: truncate the activated rules to generate output fuzzy sets; Step M7: Use the maximum-minimum synthesis method to merge the output fuzzy sets of all rules; Step M8: Introduce evidence theory to deal with conflicting rules and calculate the confidence interval of each conclusion; Step M9: Use the centroid method to calculate the centroid position of the aggregated fuzzy set as the final output value, and use the maximum membership principle to select the most likely category for the classification task; Step M10: Output the accompanying confidence index to reflect the credibility of the inference result.

[0021] An adaptive optimization mechanism can also be set up: dynamically updating the fuzzy rule base through incremental FCM clustering; and adjusting membership function parameters based on feedback data.

[0022] The present invention has the following beneficial effects: (1) The present invention collects cable data information through a hybrid information acquisition layer, dynamically matches the video analysis window function with the signal characteristics, constructs the time domain graph scale, frequency domain resonant component and spatial field intensity gradient, and uses the intelligent diagnosis layer to perform fault diagnosis and location, thereby improving the efficiency of distribution network fault location detection; (2) The present invention realizes exponential adaptive adjustment of Gaussian window width by calculating the weighted exponent of the local signal-to-noise ratio and instantaneous frequency in real time, thus overcoming the limitations of traditional fixed window function in non-stationary signal analysis. At the same time, the standard Gaussian window is used in the low-frequency band to ensure energy concentration, and the Morlet wavelet kernel is switched in the high-frequency band to enhance the instantaneous frequency capture capability. The smooth transition of the kernel function is achieved through the Sigmoid function, avoiding the spectrum leakage caused by traditional hard switching, and improving the anti-interference capability and time-frequency resolution optimization capability. (3) The present invention is based on nonlinear time-frequency grid reconstruction in a hyperbolic coordinate system and combines tensor decomposition technology to extract cross-channel common features, significantly improving the joint analysis capability of multi-source heterogeneous signals, reducing actual signal reconstruction errors, and increasing the speed of time-frequency feature extraction; (4) The present invention achieves three-level diagnosis through feature cascading and decision fusion. The CAN module completes coarse-grained classification, the STGNN analyzes the fault propagation path, and the transfer learning module ensures the model's generalization capability across power grids. (5) The present invention quantifies uncertainty through fuzzy sets and realizes approximate reasoning using an interpretable rule system. It is particularly suitable for processing fuzzy information caused by sensor noise, topology changes, etc. in the distribution network, realizing the utilization of uncertain information and improving data utilization.

[0023] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0025] Figure 1 This is a schematic structural diagram of a distribution network line fault location and detection system according to the present invention; Figure 2 This is the workflow diagram of the fault feature extraction layer. DETAILED DESCRIPTION

[0026] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0027] In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0028] In order to make the purpose, technical solutions and advantages of this application more clear, the following Figure 1-2 It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application.

[0029] See also Figure 1 As shown, the present invention is a distribution network line fault location detection system, including a hybrid information acquisition layer, a fault feature extraction layer, an intelligent diagnosis layer and a fault location layer; The mixed signal acquisition layer includes a high-frequency transient recording unit (sampling rate ≥ 1MHz), a power frequency measurement unit, a wireless pulse sensor, and a distributed optical fiber temperature measurement unit. The high-frequency transient recording unit is used to record the microsecond transient process of the voltage / current traveling wave and capture characteristic parameters. The power frequency measurement unit is used to synchronously measure the effective value of the power frequency waveform set of the three-phase voltage / current, including the dynamic changes of the positive sequence, negative sequence, and zero sequence components. The wireless pulse sensor is used to capture the distortion of the spatial electromagnetic field. The distributed optical fiber temperature measurement unit is used to measure the conductor temperature in real time. The fault feature extraction layer includes a time-frequency analysis module, a preprocessing module, and a three-dimensional feature vector module. The time-frequency analysis module is used to dynamically match the signal features through the video analysis window function. The preprocessing module is used to preprocess the mixed signal to eliminate common-mode interference. The three-dimensional feature vector module is used to construct the time domain image scale, frequency domain resonant component, and spatial field intensity gradient. The intelligent diagnosis layer includes a convolutional attention network, a spatiotemporal graph neural network, and a transfer learning module. The convolutional attention network is used for initial fault type screening; the spatiotemporal neural network is used to build line topology associations; and the transfer learning module is used to adapt to different distribution network structures. The fault location layer includes a particle swarm module and a fuzzy reasoning module; the particle swarm module is used to locate faults on distribution lines through the particle swarm algorithm; the fuzzy reasoning module is used to process uncertain information.

[0030] The high-frequency transient recording unit includes a multi-channel synchronous sampling module, an adaptive filtering circuit, and a synchronous clock module. The multi-channel synchronous sampling module is equipped with an 8-channel 16-bit high-precision ADC for synchronous acquisition of voltage / current traveling waves and power frequency signals in three modes, with a sampling rate of ≥50MHz. The adaptive filtering circuit is used to dynamically reduce the noise of microsecond-level transient characteristics such as cable joint discharge and tree barrier flashover. The synchronous clock module is equipped with a GPS / passive dual-mode timing unit, with a time synchronization accuracy of <10ns. The wireless pulse sensor consists of a distributed electromagnetic field sensor array, an adaptive power transmitter module, and an edge computing module. Each tower in the distributed electromagnetic field sensor array deploys 3-6 UWB nodes, with a measurement bandwidth covering the 300MHz-3GHz frequency band. The adaptive power transmitter module dynamically adjusts the transmit power from 0.110mW to 10mW based on ambient noise, ensuring signal integrity within a 20-meter radius. The edge computing module's nodes have a built-in ARM Cortex-M7 processor, enabling on-site raw data preprocessing. The distributed fiber optic temperature measurement unit includes a temperature measurement optical cable, a Raman scattering demodulation module and an OTDR positioning module; anti-bending armored light is laid along the temperature measurement optical cable, with a spatial resolution of ≤1m; the Raman scattering demodulation module uses dual-wavelength laser pulses to measure the temperature of the optical cable, with a temperature measurement accuracy of ±0.5℃; the OTDR positioning module is used to combine optical time domain reflectometry technology to achieve meter-level positioning of temperature anomalies.

[0031] See also Figure 2 As shown in Figure 2, the specific workflow of the fault feature extraction layer is as follows: Step G01: Calculate the weighted exponent of the local signal-to-noise ratio and instantaneous frequency in real time, and drive the Gaussian window width to adaptively adjust according to the exponential law; the specific implementation process is as follows: Use a sliding window with a 50% overlap rate to frame the signal, calculate the ratio of signal energy to noise floor energy in each frame, obtain the phase derivative of the analytical signal through Hilbert transform, and construct a weighted function of SNR and frequency , dynamically adjust the Gaussian window based on the exponential output: ; Where SNR is the power ratio of signal to noise, is the instantaneous angular frequency, are the weighting coefficients of the signal-to-noise ratio term and the frequency term, respectively. is the standard deviation of the Gaussian window after dynamic adjustment, is the initial standard deviation reference value of the Gaussian window, k is the adjustment coefficient, is a composite weighting function, and e is a natural constant.

[0032] Step G02: Introduce the quantum genetic algorithm to optimize the function parameters and dynamically search for the optimal time-frequency resolution combination in the range of 0.1ms-10ms; Step G03: A standard Gaussian window is used in the low-frequency band (<1kHz), and a Morlet wavelet kernel is used in the high-frequency band (>10kHz). The transition region is smoothly switched using a Sigmoid function. Step G04: The logarithmic frequency axis and the linear time axis form a hyperbolic coordinate system; the hyperbolic coordinate system includes the time axis and the frequency axis; the time axis needs to maintain a linear sampling interval , the frequency axis uses log2 scale, with a minimum resolution of 0.1Hz; construct hyperbolic coordinates The non-uniform sampling network is used, and the time-frequency surface reconstruction is completed using cubic spline interpolation; Step G05: Refocusing the expanded energy to the real time-frequency trajectory by calculating the instantaneous frequency gradient field; Step G06: Decompose the three-dimensional time-frequency matrix into core tensors and factor matrices to extract cross-channel common features. The three-dimensional time-frequency matrix includes time, frequency, and sensor channels. Stack the multi-channel time-frequency matrices into a three-dimensional tensor (channel, frequency, time), and perform energy normalization on each channel. Use the Tucker decomposition model ;Where G is the core tensor, A is the channel factor matrix, B is the frequency factor matrix, and C is the time factor matrix; During feature extraction, the cross-channel coupling features are analyzed by slicing the core tensor G, the channel commonality pattern is extracted from the factor matrix A, and the energy threshold is set to screen the significant feature components; Step G07: Using adjacent non-fault line signals to establish an interference dictionary library, and perform noise suppression using compressed sensing theory; Step G08: Dynamically adjust the filter strength according to the local signal-to-noise ratio of the signal to retain valid fault features.

[0033] In step G03, the window width parameter matrix is ​​preset , load the time-frequency dictionary of the Morlet wavelet kernel, use a sliding window to calculate the in-band energy of the current analysis frame, estimate the noise floor through the Kalman filter, and define a dynamic signal-to-noise ratio threshold; When a transient pulse (du / dt>5kV / μs) is detected, high-frequency processing is performed to activate The narrow window has a time resolution of ; When the fundamental frequency fluctuation exceeds When the Wide window, frequency resolution reaches ; The Morlet wavelet kernel is designed with nonlinear frequency axis and a logarithmic-linear hybrid scale is constructed. Using linear division, Using logarithmic division, the center frequency is set to To ensure time-frequency balance, introduce bandwidth factor , which increases the frequency resolution of low-frequency area by 40% and the frequency resolution of high-frequency area by Temporal resolution is improved by 30%.

[0034] In step G05, the specific process of refocusing the expanded energy to the real time-frequency trajectory by calculating the instantaneous frequency gradient field is as follows: Step G051: Perform Sobel operator convolution on the generated time-frequency surface; the convolution kernel is designed as follows: horizontal kernel: [-1, 0, 1; -2, 0, 2; -1, 0, 1], detects vertical edges; vertical kernel: [-1, -2, -1; 0, 0, 0; 1, 2, 1], detects horizontal edges; perform horizontal and vertical convolution on the time-frequency surface matrix respectively to obtain the gradient component and ; then the magnitude of the gradient is: , the direction is: ; Step G052: Perform particle flow simulation on the time-frequency surface along the gradient direction. During the particle flow simulation, particles are generated in the high-energy region of the time-frequency surface (the first 10% of pixels in the amplitude). Particle properties include position, velocity, and mass. The particle motion equation is based on the Stokes drag model: , where gravity Along the gradient direction, the drag force The direction is opposite to the velocity. During the movement, the time step is set and the particle position is updated through Euler iteration. The particle stops when it reaches the local energy minimum or exceeds the time-frequency boundary. Step G053: Record the extreme points of the gradient amplitude along the particle trajectory as candidate points. After screening the candidate points, perform cubic spline interpolation to generate a continuous trajectory. When screening the candidate points, sort them in descending order of amplitude, retain the maximum point M, and eliminate adjacent points with a time-frequency overlap ratio with M greater than 0.5. Repeat the iteration until all candidate points have been processed. Step G054: Perform Gaussian energy redistribution with the trajectory as the center; redistribute the original time-frequency energy according to the Gaussian kernel weight with the trajectory point as the center, while maintaining the conservation of total energy, and perform bilinear interpolation on the redistributed time-frequency matrix to eliminate gridding artifacts.

[0035] In step G07, the interference dictionary is established using the adjacent non-fault line signals, and the specific process of noise suppression is performed using the compressed sensing theory as follows: Step G071: Collect signals from adjacent normal routes to construct a complete dictionary D; Step G072: Use K-SVD algorithm to train dictionary atoms; Step G073: Establish a multi-resolution dictionary hierarchy structure; Step G074: Send the signal to be processed , use ISTA algorithm for sparse representation; Step G075: Retain the main sparse components through hard threshold processing, reconstruct the signal and suppress the noise. The reconstructed signal is , adaptively adjust the regularization parameter .

[0036] The workflow of the intelligent diagnosis layer is as follows: Step Z01: Convolutional attention network performs initial fault screening; Step Z02: The spatiotemporal graph neural network performs topological modeling and outputs the propagation path of potential faults at each node; Step Z03: The transfer learning module performs domain adaptation training and dynamically adjusts the learning rate; In step Z01, parallel 1D-CNN branches are used to process signals in different frequency bands. Dilated convolution is used to expand the receptive field and capture long-range features. The inter-channel correlation matrix is ​​calculated for spatial attention to focus on key sensors. Gating is used to enhance transient fault features for spatial attention. After feature fusion, the Softmax layer outputs the initial fault probability distribution. The specific implementation process is as follows: Design 3-5 parallel 1D-CNN branches, each equipped with a bandpass filter with a different cutoff frequency to preprocess the input signal. Each branch uses a convolution kernel of a specific scale (e.g., 64 / 128 / 256 points) to match the characteristic period of the corresponding frequency band. Introduce dilated convolution at the top network layer to expand the receptive field at an exponential rate (d=1, 2, 4, 8). Use residual connections to alleviate the vanishing gradient problem and preserve features at all levels. By calculating the covariance matrix of multi-channel signals, the principal component vector is obtained through singular value decomposition. The channel attention module (CBAM) is used to generate weight vectors to enhance key sensor features. LSTM units are used to construct time gating and calculate the attention score at each time point. Sigmoid gating is used to control the propagation strength of fault transient features. Adaptive average pooling is performed on feature maps of different scales to unify the dimensions, and a learnable 1×1 convolution is used to achieve cross-scale feature interaction. The fused features are reduced in dimension through a two-layer fully connected network, and finally the probability distribution of each fault type is output through Softmax.

[0037] In step Z02, for the spatial dimension, an initial adjacency graph is constructed based on the line impedance matrix, and Chebyshev polynomials are used to approximate graph filtering. For the temporal dimension, a time window is used to construct a spatiotemporal tensor structure, and dilated causal convolution is used to capture multi-cycle patterns. Adjacent line states are fused through a multi-head attention mechanism to output the potential fault propagation path for each node. The specific implementation process is as follows: Spatial dimension modeling: Calculate the electrical distance between nodes based on the line impedance matrix, generate a weighted adjacency matrix, and use a threshold method to sparsely process the adjacency matrix to retain strong connections; Time dimension modeling: Design sliding time windows with a 50% overlap rate, construct a three-dimensional space-time tensor (channel, frequency, time), and align data streams with different sampling rates using the dynamic time warping (DTW) algorithm; Spatial convolution implementation: Use K-order Chebyshev polynomials to approximate the graph Laplacian operator and design hierarchical aggregation strategies: local neighborhood, global topology, and cross-level feature fusion; Temporal convolution implementation: Use a causal convolution kernel with an exponentially growing expansion rate (d=1, 2, 4, ..., n), and introduce a gating mechanism to control the forgetting rate of historical information.

[0038] Construct query-key-value triples, calculate dynamic association weights between nodes, and use 4-8 heads to capture relationships in different semantic spaces in parallel. Use the gradient backpropagation algorithm to locate key impact paths and output a heat map of node-level failure probability and propagation direction. A dual-stream architecture of space and time enables topology-aware fault diagnosis. Spatial convolution captures the physical constraints of the power grid, while temporal convolution models the dynamic characteristics of fault evolution. For practical deployment, a curriculum learning strategy is recommended, training a static topology first and then gradually introducing dynamic graph complexity.

[0039] In step Z03, the shallow feature extractors of CAN and STGNN are fixed, and the MMD loss is used to minimize the difference in feature distribution between the source and target domains. The latent space representations of different power grids are aligned through adversarial training. A replay buffer for new and old samples is set to balance historical memory, and the learning rate is dynamically adjusted. The specific implementation process is as follows: The convolution kernel parameters of the first five layers of the Convolutional Attention Network (CAN) are fixed; the weights of the graph convolution layers of the Spatio-Temporal Graph Neural Network (STGNN) remain unchanged; only the trainable flags of the last three fully connected layers are opened; gradient calculations of frozen layers are skipped during backpropagation; and a masking mechanism is used to block optimizer updates of frozen parameters. Calculate the MMD distance between the source domain and target domain features. The specific formula is as follows: ; Where, It is the maximum difference, which is used to measure the difference between two probability distributions X and Y. x and y represent the sample sets of the two distributions to be compared, n and m represent the number of samples of X and Y respectively. For kernel function mapping, the original data is mapped to the regenerated kernel Hilbert space. represents the norm operation in Hilbert space; H represents the Bronk constant.

[0040] Construct a discriminator network to distinguish the source domain of features, and use a gradient reversal layer to make the feature extractor generate domain-invariant representations; automatically adjust the learning rate based on the loss change rate. The specific formula is as follows: ; Where, and They represent the dynamically adjusted learning rate and the initial baseline learning rate at the tth iteration, e is the automatic logarithm base, is the attenuation coefficient hyperparameter, is the absolute value of the gradient of the loss function at the tth iteration; Through the exponential decay mechanism, when the gradient changes drastically ( When , the learning rate Automatically reduce the learning rate to slow down parameter updates and prevent new data from overwriting old knowledge. When the gradient is flat, maintain a high learning rate to continuously absorb new knowledge. This mechanism balances model stability and adaptability.

[0041] The particle swarm module positioning process is as follows: Step D1: The fault information uploaded by the intelligent diagnosis layer is used as the fault characteristic; Step D2: Encode the fault characteristic and establish a switching function; Step D3: Define the fault information encoding and generate the initial population of the Halton sequence. The Halton sequence can generate a uniformly distributed initial population. By sorting the fitness function values, the optimal population is selected as the initial solution, which helps improve the quality of the algorithm. Step D4: Use the particle swarm algorithm to solve and update the particle swarm position and velocity. When processing the two-dimensional Halton sequence, two prime numbers are used as the basic units. By continuously expanding these two basic prime numbers, a series of evenly distributed and non-repeating points are formed. The specific mathematical model formula is as follows: ; Where, , x is a prime number greater than or equal to 2, , is the Halton sequence function, is the generated two-dimensional uniform sequence; Step D5: Determine whether the number of iterations has been reached and output the location of the fault section; To improve the particle swarm algorithm's global search performance and avoid local extrema, an adaptive dynamic inertia weighting method based on fitness can be used. The adaptive weighting factor ω decreases with the number of iterations. A larger value at the beginning of an iteration enhances global search capabilities. As the iteration progresses, the particle swarm gradually approaches the optimal target, requiring a lower value for local search to improve optimization accuracy.

[0042] The fuzzy reasoning module workflow is as follows: Step M1: Determine the domain range of input variables (such as voltage fluctuation amplitude ΔU, frequency deviation Δf) and output variables (fault confidence); Step M2: Divide each variable into 3-5 fuzzy subsets (e.g., low / medium / high), and use Gaussian or triangular membership functions to describe the fuzzy intervals; Step M3: Calculate the degree to which the input data belongs to each fuzzy subset through the membership function; Step M4: Construct expert experience rules, automatically extract rule parameters from historical fault data using the FCM clustering algorithm, and set a credibility factor (0-1) for each rule to reflect the reliability of the rule in a specific scenario; Step M5: Calculate the matching degree of the premise part of each rule (take the minimum / product operation of the input membership degree); Step M6: truncate the activated rules to generate output fuzzy sets; Step M7: Use the maximum-minimum synthesis method to merge the output fuzzy sets of all rules; Step M8: Introduce evidence theory to deal with conflicting rules and calculate the confidence interval of each conclusion; Step M9: Use the centroid method to calculate the centroid position of the aggregated fuzzy set as the final output value, and use the maximum membership principle to select the most likely category for the classification task; Step M10: Output the accompanying confidence index to reflect the credibility of the inference result.

[0043] An adaptive optimization mechanism can also be set up: dynamically updating the fuzzy rule base through incremental FCM clustering; and adjusting membership function parameters based on feedback data.

[0044] It is worth noting that in the above system embodiment, the various units included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of the functional units are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of the present invention.

[0045] In addition, those skilled in the art will appreciate that all or part of the steps in the above-mentioned embodiments can be accomplished by instructing related hardware through a program, and the corresponding program can be stored in a computer-readable storage medium.

[0046] The preferred embodiments of the present invention disclosed above are intended only to help illustrate the present invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the present invention to the specific embodiments described. Obviously, many modifications and variations are possible based on the content of this specification. These embodiments are selected and described in detail in this specification to better explain the principles and practical applications of the present invention, thereby enabling those skilled in the art to better understand and utilize the present invention. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A distribution network line fault location detection system, comprising a hybrid information acquisition layer, a fault feature extraction layer, an intelligent diagnosis layer, and a fault location layer, characterized in that: The mixed signal acquisition layer includes a high-frequency transient recording unit, a power frequency measurement unit, a wireless pulse sensor, and a distributed optical fiber temperature measurement unit; the high-frequency transient recording unit is used to record the microsecond transient process of the voltage / current traveling wave and capture characteristic parameters; the power frequency measurement unit is used to synchronously measure the effective value of the power frequency waveform set of the three-phase voltage / current, including the dynamic change process of the positive sequence, negative sequence, and zero sequence components; the wireless pulse sensor is used to capture the distortion of the spatial electromagnetic field; and the distributed optical fiber temperature measurement unit is used to measure the conductor temperature in real time; The fault feature extraction layer includes a time-frequency analysis module, a preprocessing module and a three-dimensional feature vector module; the time-frequency analysis module is used to dynamically match the signal features through the video analysis window function; The preprocessing module is used to preprocess the mixed signal to eliminate common mode interference; the three-dimensional feature vector module is used to construct the time domain graph scale, frequency domain resonant component and spatial field intensity gradient; The intelligent diagnosis layer includes a convolutional attention network, a spatiotemporal graph neural network, and a transfer learning module; the convolutional attention network is used to perform preliminary fault type screening; the spatiotemporal neural network is used to build line topology associations; and the transfer learning module is used to adapt to different distribution network structures. The fault location layer includes a particle swarm module and a fuzzy reasoning module; the particle swarm module is used to locate faults on the distribution line through a particle swarm algorithm; and the fuzzy reasoning module is used to process uncertain information.

2. A distribution network line fault location detection system according to claim 1, characterized in that: The high-frequency transient recording unit includes a multi-channel synchronous sampling module, an adaptive filtering circuit, and a synchronous clock module; the multi-channel synchronous sampling module is equipped with an 8-channel 16-bit high-precision ADC for synchronous acquisition of voltage / current traveling waves and power frequency signals; the adaptive filtering circuit is used to dynamically reduce the noise of the microsecond-level transient characteristics of the cable; the synchronous clock module is equipped with a GPS / passive dual-mode timing unit; The wireless pulse sensor includes a distributed electromagnetic field sensor array, an adaptive power transmission module, and an edge computing module. Each tower in the distributed electromagnetic field sensor array deploys 3-6 UWB nodes, with a measurement bandwidth covering the 300MHz-3GHz frequency band. The adaptive power transmission module dynamically adjusts the transmission power by 0.1-10mW based on the ambient noise. The edge computing module node has a built-in ARM Cortex-M7 processor to achieve localized raw data preprocessing. The distributed optical fiber temperature measurement unit includes a temperature measurement optical cable, a Raman scattering demodulation module and an OTDR positioning module; anti-bending armored light is laid along the temperature measurement optical cable; the Raman scattering demodulation module uses dual-wavelength laser pulses to measure the optical cable temperature; and the OTDR positioning module is used to locate the temperature difference point.

3. A distribution network line fault location detection system according to claim 1, characterized in that: The specific workflow of the fault feature extraction layer is as follows: Step G01: Calculate the weighted exponent of the local signal-to-noise ratio and the instantaneous frequency in real time, and drive the Gaussian window width to adaptively adjust according to the exponential law; Step G02: Introduce the quantum genetic algorithm to optimize the function parameters and dynamically search for the optimal time-frequency resolution combination in the range of 0.1ms-10ms; Step G03: The standard Gaussian window is used in the low-frequency band, the Morlet wavelet kernel is switched to the high-frequency band, and the Sigmoid function is used for smooth switching in the transition region; Step G04: The logarithmic frequency axis and the linear time axis form a hyperbolic coordinate system; Step G05: Refocusing the expanded energy to the real time-frequency trajectory by calculating the instantaneous frequency gradient field; Step G06: Decompose the three-dimensional time-frequency matrix into core tensors and factor matrices to extract cross-channel common features; Step G07: Using adjacent non-fault line signals to establish an interference dictionary library, and perform noise suppression using compressed sensing theory; Step G08: Dynamically adjust the filtering strength according to the local signal-to-noise ratio of the signal.

4. A distribution network line fault location detection system according to claim 3, characterized in that: In step G03, the preset window width parameter matrix , a sliding window is used to calculate the in-band energy of the current analysis frame, the noise floor is estimated through the Kalman filter, and a dynamic signal-to-noise ratio threshold is defined; When a transient pulse is detected, high frequency processing is performed to activate The narrow window has a time resolution of ; When the fundamental frequency fluctuation exceeds When the Wide window, frequency resolution reaches .

5. A distribution network line fault location detection system according to claim 3, characterized in that: In step G05, the specific process of refocusing the expanded energy to the real time-frequency trajectory by calculating the instantaneous frequency gradient field is as follows: Step G051: performing Sobel operator convolution on the generated time-frequency surface; Step G052: performing particle flow simulation on the time-frequency surface along the gradient direction; Step G053: Record the extreme points of the gradient amplitude along the particle trajectory as candidate points, screen the candidate points, and perform cubic spline interpolation to generate a continuous trajectory; Step G054: Perform Gaussian energy redistribution with the trajectory as the center.

6. A distribution network line fault location detection system according to claim 3, characterized in that: In step G07, the interference dictionary is established using the adjacent non-fault line signals, and the specific process of noise suppression is performed using the compressed sensing theory as follows: Step G071: Collect signals from adjacent normal routes to construct a complete dictionary D; Step G072: Use K-SVD algorithm to train dictionary atoms; Step G073: Establish a multi-resolution dictionary hierarchy structure; Step G074: Perform sparse representation on the signal to be processed using the ISTA algorithm; Step G075: Retain the main sparse components through hard threshold processing, reconstruct the signal and perform noise suppression.

7. A distribution network line fault location detection system according to claim 1, characterized in that: The workflow of the intelligent diagnosis layer is as follows: Step Z01: Convolutional attention network performs initial fault screening; Step Z02: The spatiotemporal graph neural network performs topological modeling and outputs the propagation path of potential faults at each node; Step Z03: The transfer learning module performs domain adaptation training and dynamically adjusts the learning rate; In step Z01, parallel 1D-CNN branches are used to process signals of different frequency bands, and the receptive field is expanded through hollow convolution to capture long-range features. The inter-channel correlation matrix is ​​calculated for spatial attention to focus on key sensors. Gate control is used for spatial attention to enhance transient fault features. After feature fusion, the Softmax layer outputs the initial screening fault probability distribution. In step Z02, for the spatial dimension, an initial adjacency graph is constructed based on the line impedance matrix, and Chebyshev polynomial approximation graph filtering is used. For the temporal dimension, a time window is used to construct a spatiotemporal tensor structure, and dilated causal convolution is used to capture multi-cycle patterns. The states of adjacent lines are fused through a multi-head attention mechanism, and the potential fault propagation path of each node is output.

8. A distribution network line fault location detection system according to claim 1, characterized in that: In step Z03, the shallow feature extractors of CAN and STGNN are fixed, and the MMD loss is used to minimize the difference in feature distribution between the source domain and the target domain. The latent space representations of different power grids are aligned through adversarial training, a replay buffer for new and old samples is set to balance historical memory, and the learning rate is dynamically adjusted.

9. A distribution network line fault location detection system according to claim 1, characterized in that: The particle swarm module positioning process is as follows: Step D1: The fault information uploaded by the intelligent diagnosis layer is used as the fault characteristic; Step D2: Encode the fault characteristic and establish a switching function; Step D3: define the fault information encoding and generate the initial population of the Halton sequence; Step D4: Use the particle swarm algorithm to solve and update the particle swarm position and velocity; Step D5: Determine whether the number of iterations has been reached and output the location of the fault section.

10. A distribution network line fault location detection system according to claim 1, characterized in that: The fuzzy reasoning module workflow is as follows: Step M1: Determine the domain scope of input variables and output variables; Step M2: Divide each variable into 3-5 fuzzy subsets and use Gaussian or triangular membership functions to describe the fuzzy intervals; Step M3: Calculate the degree to which the input data belongs to each fuzzy subset through the membership function; Step M4: Construct expert experience rules and automatically extract rule parameters from historical fault data using the FCM clustering algorithm; Step M5: Calculate the matching degree of the premise part of each rule; Step M6: truncate the activated rules to generate output fuzzy sets; Step M7: Use the maximum-minimum synthesis method to merge the output fuzzy sets of all rules; Step M8: Introduce evidence theory to deal with conflicting rules and calculate the confidence interval of each conclusion; Step M9: Use the centroid method to calculate the centroid position of the aggregated fuzzy set as the final output value, and use the maximum membership principle to select the most likely category for the classification task; Step M10: Output the accompanying confidence index to reflect the credibility of the inference result.

Images