Deep learning based cable line fault location method

Abstract

The application discloses a cable line fault positioning method based on deep learning and belongs to the field of power cable fault detection and intelligent positioning. In order to solve the problem that the unstable positioning is caused by the mixed cable fault signals, load fluctuation and environmental noise, the application realizes the technical effects of narrowing the maintenance interval and prompting the retest channel through the multi-source state data preprocessing, pulse time-frequency feature extraction, deep model coding, dynamic physical graph constraint and positioning signal interval output.

Description

Technical Field

[0001] This invention relates to the field of power cable fault detection and intelligent location, and in particular to a cable line fault location method based on deep learning. Background Technology

[0002] Cable lines are subjected to long-term electrothermal stress, mechanical stress, and environmental factors in urban power distribution, industrial power supply, and substation outgoing lines. Problems such as partial discharge, insulation degradation, and joint abnormalities can easily develop gradually during operation. Current power maintenance and electrical testing typically combine partial discharge detection, acoustic-magnetic synchronous detection, sheath current monitoring, and manual experience to determine the fault location. However, the sampling frequency, phase reference, and noise background of different detection data are not consistent, and a single test data is difficult to reliably reflect the true location of the fault.

[0003] Under the combined effects of load current fluctuations, ambient temperature and humidity changes, and on-site electromagnetic noise, partial discharge pulses, reflected pulses, and sensor background noise are prone to aliasing, leading to deviations in the judgment of pulse arrival time, fault section, and distance to the fault point. Traditional methods also struggle to constrain cable length, joint location, propagation speed, attenuation coefficient, joint reflection characteristics, and sensor reliability into the positioning process, thus easily expanding the maintenance scope or requiring multiple retests.

[0004] Therefore, a deep learning-based cable line fault location method is needed to address the shortcomings of the existing technologies. Summary of the Invention

[0005] One objective of this invention is to propose a cable line fault location method based on deep learning. Addressing the problems in existing technologies where cable fault signals are mixed with load fluctuations and environmental noise, and where traditional fault location methods rely on human experience and single test data leading to unstable fault point identification, this invention proposes a technical solution that integrates multi-source state data preprocessing, pulse time-frequency feature extraction, deep learning encoding, dynamic physical graph constraints, and confidence interval location fusion. This invention achieves the technical effect of improving the consistency of fault section identification and narrowing the maintenance scope.

[0006] This invention provides a deep learning-based method for cable line fault location, comprising: S1. Acquire multi-source status data during cable line operation, perform time synchronization, abnormal sampling and elimination, power frequency phase alignment and amplitude normalization on the multi-source status data, generate synchronized status data, and generate interference tags based on load current change rate and environmental noise index. S2. Perform adaptive mode decomposition, pulse segmentation and time-frequency feature extraction on the partial discharge signal and acoustic-magnetic synchronization detection signal in the synchronization state data to generate candidate pulse events, incident pulse arrival time, reflected pulse arrival time, pulse segmentation confidence and time-frequency feature map. S3. Input the candidate pulse event into the trained time-domain feature encoder to generate a fault mode vector, and input the time-frequency feature map into the trained time-frequency feature encoder to generate a discharge mode vector. S4. Construct a dynamic physical graph using cable terminals, joints, cable segments, and sensors to determine sensor reliability. Adjust edge weights based on line parameters, sensor reliability, interference flags, and pulse segmentation confidence. Generate segment failure probability vectors through a graph attention network. S5. Input the fault mode vector, discharge mode vector, incident pulse arrival time, reflected pulse arrival time, and segmented fault probability vector into the positioning regression head to generate the fault type, fault segment, fault point distance, and positioning information interval. Then, output the maintenance interval or retest channel based on the comparison result between the interval length of the positioning information interval and the preset maintenance length threshold.

[0007] Optionally, S1 includes: The multi-source status data includes partial discharge high-frequency current signal, ultra-high frequency electromagnetic signal, acoustic-magnetic synchronous detection signal, sheath current, load current, ambient temperature and humidity, cable length, joint position, terminal position, and line impedance parameters. Data with different sampling rates are resampled according to a unified time base, and the power frequency zero-crossing time is used as the phase alignment base. The environmental noise index is obtained by weighting the root mean square value of the signal within the non-discharge pulse time window and the amplitude of the ultra-high frequency background. The interference flag is determined by whether the load current change rate exceeds a preset load change threshold and whether the environmental noise index exceeds a preset noise threshold.

[0008] Optionally, S2 includes: Adaptive mode decomposition is performed on the partial discharge high-frequency current signal, ultra-high frequency electromagnetic signal and acoustic-magnetic synchronous detection signal respectively, and the mode components consistent with the power frequency phase and sensor trigger time are retained. The retained modal components are input into the trained U-Net pulse segmentation network, which outputs candidate discharge pulse segments, candidate reflection pulse segments, incident pulse arrival time, reflection pulse arrival time, and pulse segmentation confidence. The candidate pulse events are formed from the candidate discharge pulse segments and candidate reflection pulse segments. Wavelet transform and Stockwell time-frequency transform are performed on the candidate discharge pulse segment and the candidate reflection pulse segment, respectively, to generate the time-frequency feature map; Furthermore, the trained U-Net pulse segmentation network, temporal feature encoder, time-frequency feature encoder, graph attention network, and localization regression head are trained end-to-end using historical samples with fault type labels, fault distance labels, pulse arrival time labels, pulse mask labels, and fault segmentation labels. The training loss includes pulse segmentation loss, fault type classification loss, fault distance regression loss, and segmented probability supervision loss for the output of the graph attention network. The training samples are stratified according to interference labels to maintain the sample ratio of load fluctuation scenarios and environmental noise scenarios.

[0009] Optionally, S3 includes: Candidate pulse events are divided into time-domain pulse sequences according to the power frequency cycle. One-dimensional convolutional layers are used to extract pulse rising edge, duration and decay envelope features, and Transformer self-attention layers are used to generate cross-cycle evolution features as the fault mode vector. The time-frequency feature map is divided into image blocks according to frequency band and time window, and the discharge mode vector is generated using the SwinTransformer window attention layer.

[0010] Optionally, S4 includes: Cable terminal nodes, joint nodes, cable segment nodes, and sensor nodes are used as nodes in the dynamic physical graph, and the detection edge from the sensor to the cable segment, the propagation edge of the adjacent cable segment, and the reflection edge at the joint are used as edges in the dynamic physical graph. The channel noise level is determined based on the noise statistics of the corresponding sensor channel within the non-discharge pulse time window. Initial edge weights are generated using cable length, medium propagation speed, frequency-dependent attenuation coefficient, connector reflection coefficient, sensor installation location, and channel noise level. Sensor reliability is determined based on historical effective sampling rate, channel noise level, and sensor installation location calibration coefficient; Based on the interference marker and its corresponding load current change rate, environmental noise index and pulse segmentation confidence, a side weight correction amount is generated, and the initial value of the side weight and the side weight correction amount are input into the graph attention network to obtain the association weight between each cable segment and the candidate pulse event and the segment fault probability vector. Furthermore, the edge weight correction includes detection edge correction, propagation edge correction, and reflection edge correction. The detection edge correction is determined based on sensor reliability, channel noise level, and pulse segmentation confidence. The propagation edge correction is determined based on medium propagation velocity, frequency-related attenuation coefficient, load fluctuation state in the interference marker, and current load current change rate. The reflection edge correction is determined based on connector reflection coefficient, connector position, and arrival time of the reflected pulse. The edge weight correction is used to perform weighted updates on the initial edge weight values.

[0011] Optionally, S5 includes: The fault mode vector, discharge mode vector, incident pulse arrival time, reflected pulse arrival time and segmented fault probability vector are input into the positioning regression head, and the positioning regression head outputs the fault type probability, fault segment identifier and initial value of fault point distance. The distance to the fault point is obtained by performing physical consistency correction based on the initial value of the distance to the fault point, the propagation speed, and the joint position. A location information interval is generated based on the distance to the fault point; When the interval length of the fixed position signal interval is less than or equal to the preset maintenance length threshold, the maintenance interval containing the distance to the fault point is output; when the interval length of the fixed position signal interval exceeds the preset maintenance length threshold, the sensor channel and candidate segment that need to be retested are output. Furthermore, the location information interval is generated as follows: the location residual distribution is generated based on the difference between the fault distance marked in the historical samples and the fault point distance after physical consistency correction; Calculate the segmented fault probability entropy based on the segmented fault probability vector; Generate a reliability reduction factor based on sensor reliability; The preset quantile value of the location residual distribution, the segmented fault probability entropy and the reliability reduction coefficient are input into the interval mapping function to obtain the inconsistency score, and the location information interval is generated based on the inconsistency score and the preset information level around the fault point distance after physical consistency correction. Furthermore, the retest channel is determined as follows: when the interval length of the fixed position signal interval exceeds the preset maintenance length threshold, the attention contribution of each sensor node is summarized according to the attention weights output or retained by the graph attention network on the detection edge from the sensor node to the cable segment. Sensor channels whose attention contribution is not less than a preset contribution threshold and whose channel noise level exceeds a preset channel noise threshold are marked as sensor channels that need to be retested. Cable segments with a segment failure probability not less than a preset segment probability threshold are identified as candidate segments.

[0012] The beneficial effects of this invention are: 1. By performing time synchronization, anomaly sampling and removal, power frequency phase alignment and amplitude normalization on multi-source state data, and generating interference markers based on load current change rate and environmental noise indicators, the impact of load fluctuations and on-site noise on subsequent pulse identification and location fusion can be reduced.

[0013] 2. By inputting candidate pulse events and time-frequency feature maps into the trained one-dimensional convolutional-transformer encoder and SwinTransformer encoder respectively, and combining them with the pulse arrival time output by U-Net, the characteristics of partial discharge and reflected pulses can be characterized from multiple dimensions such as temporal morphology, cross-cycle evolution and time-frequency pattern.

[0014] 3. By constructing a dynamic physical graph that includes terminals, connectors, cable segments, and sensors, and adjusting the edge weights based on line parameters, sensor reliability, interference markers, and pulse segmentation confidence, the deep learning output can be constrained by the cable's physical topology and propagation laws, thereby outputting the fault point distance, location information interval, maintenance interval, or channel requiring retesting. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a deep learning-based cable line fault location method.

[0016] Figure 2 This is a flowchart of step S4, dynamic physical graph construction and edge weight correction, of the present invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figure 1 A deep learning-based method for cable line fault location includes: S1. Acquire multi-source status data during cable line operation, perform time synchronization, abnormal sampling and elimination, power frequency phase alignment and amplitude normalization on the multi-source status data, generate synchronized status data, and generate interference tags based on load current change rate and environmental noise index. S2. Perform adaptive mode decomposition, pulse segmentation and time-frequency feature extraction on the partial discharge signal and acoustic-magnetic synchronization detection signal in the synchronization state data to generate candidate pulse events, incident pulse arrival time, reflected pulse arrival time, pulse segmentation confidence and time-frequency feature map. S3. Input the candidate pulse event into the trained time-domain feature encoder to generate a fault mode vector, and input the time-frequency feature map into the trained time-frequency feature encoder to generate a discharge mode vector. S4. Construct a dynamic physical graph using cable terminals, joints, cable segments, and sensors to determine sensor reliability. Adjust edge weights based on line parameters, sensor reliability, interference flags, and pulse segmentation confidence. Generate segment failure probability vectors through a graph attention network. S5. Input the fault mode vector, discharge mode vector, incident pulse arrival time, reflected pulse arrival time, and segmented fault probability vector into the positioning regression head to generate the fault type, fault segment, fault point distance, and positioning information interval. Then, output the maintenance interval or retest channel based on the comparison result between the interval length of the positioning information interval and the preset maintenance length threshold.

[0019] In this specific embodiment, S1 includes: The same data acquisition and calibration process is used to obtain multi-source status data during cable line operation. This multi-source status data includes partial discharge high-frequency current signals, ultra-high frequency electromagnetic signals, acoustic-magnetic synchronous detection signals, sheath current, load current, ambient temperature, ambient humidity, cable length, joint location, terminal location, and line impedance parameters. The partial discharge high-frequency current signal is acquired through a high-frequency current transformer and... The sampling rate is used to sample ultra-high frequency electromagnetic signals, which are acquired through an ultra-high frequency antenna and then... The sampling rate is used for sampling, and the acoustic-magnetic synchronous detection signal is acquired by simultaneous triggering of the acoustic sensor and the magnetic field sensor and then... The sampling rate is used to sample the sheath current and the load current. Sampling rate sampling, ambient temperature and ambient humidity The sampling rate is used to sample cable length, joint position, terminal position, and line impedance parameters as static parameters, which are then broadcast and written into the synchronization status data over time. The raw data from each channel are uniformly timestamped using the IEEE 1588 precision clock synchronization protocol. Perform time synchronization and construct a target time grid on a unified time base. and The partial discharge high-frequency current signal and the acoustic-magnetic synchronization detection signal are resampled to the target time grid using band-limited sinc interpolation. The ultra-high frequency electromagnetic signal is first subjected to anti-aliasing low-pass filtering and then decimated by a factor. After downsampling to 20MHz, the current and load current are aligned to the target time grid using cubic spline interpolation. The ambient temperature and ambient humidity are aligned to the target time grid using zero-order hold. Anomaly removal is achieved using a channel-based sliding window statistical method, with the sliding window length set to 1. 1 target time grid sampling point and window shift is For each sampling point, the median and absolute median difference of the sample in that channel are calculated within each sliding window, and an anomaly threshold is formed accordingly. Sampling points exceeding the anomaly threshold are judged as abnormal samples and replaced by linear interpolation of adjacent valid sampling points, thereby generating anomaly-free multi-channel time-aligned data. Power frequency phase alignment uses the power frequency component of the load current as a phase reference. First, bandpass filtering is performed on the load current to obtain the power frequency component with a center frequency of 50Hz, and its zero-crossing time sequence is detected. Then, at each zero-crossing time As a reference for power frequency phase alignment, the data of each channel is periodically sliced ​​and reindexed so that the starting point of each slice corresponds to the same power frequency phase zero point; Amplitude normalization is performed separately for each channel, using a "mean-removal and scaling by standard deviation" method to normalize the most recent amplitude of each channel. The synchronization data is standardized to zero mean and unit variance, and the cable length, joint position, terminal position and line impedance parameters are linearly scaled according to their engineering dimensions and then spliced ​​with the time sequence data of each channel to obtain the synchronization status data. An interference flag is generated based on the synchronization state data. The interference flag is determined by the load current change rate and the environmental noise index, wherein the load current change rate is at discrete time points after alignment. The calculation is based on the relative changes at adjacent times and compared with the preset load change threshold. The load fluctuation state is compared, and the environmental noise index is obtained by first selecting a non-discharge pulse time window in each power frequency cycle and then calculating the root mean square value of the signal and the ultra-high frequency background amplitude within the time window, and then weighting them. The non-discharge pulse time window is determined by the dual-channel threshold method, that is, selecting a time window in each power frequency cycle that satisfies the condition that "the peak-to-peak value of the partial discharge high-frequency current signal does not exceed its nearest..." The peak-to-peak value of the baseline is 1.2 times that of the ultra-high frequency electromagnetic signal, and the peak-to-peak value of the signal does not exceed its most recent peak value. A continuous time slice with a length of 1.2 times the baseline peak value is selected, and the slice is taken as the length of the continuous time slice. The segment is used as the non-discharge pulse time window, and the environmental noise index is compared with the preset noise threshold. The noise status is obtained through comparison, and the environmental noise index and the load current change rate are calculated using the following unified formula: ; in Indicates time Environmental noise index Indicates time The rate of change of load current, Represents the noise weighting coefficient and takes This represents the root mean square value of the partial discharge high-frequency current signal and the acousto-magnetic synchronization detection signal within the non-discharge pulse time window after alignment and normalization. This represents the average background amplitude of the ultra-high frequency electromagnetic signal within the non-discharge pulse time window after alignment and normalization. Indicates time The load current sampling value, Indicates the first Each target time grid moment, Indicates the time of the previous target time grid. Indicates taking the absolute value; Finally, the load fluctuation status and noise status are combined to generate an interference flag, which is then output along with the synchronization status data. The interference flag uses a binary tuple (...). ) indicates and express Otherwise, it is 0. express Otherwise, it is 0, thus providing a reproducible interference state input for subsequent steps such as pulse segmentation, edge weight correction, and localization fusion.

[0020] In this specific embodiment, S2 includes: Based on synchronization state data, event window truncation and adaptive mode decomposition are performed on partial discharge high-frequency current signals, ultra-high frequency electromagnetic signals, and acoustic-magnetic synchronization detection signals to obtain high signal-to-noise ratio mode components for pulse segmentation. The trigger time of each acquisition is used as the trigger point. Extract duration from center The event window is formed on a unified time grid with a length of The sequence and For the target time grid interval mentioned in step S1, each event window signal is decomposed into its target time grid interval using the complete ensemble empirical mode decomposition algorithm CEEMDAN. eigenmode functions The set number of the CEEMDAN is taken from the residual components. Furthermore, the standard deviation of the added white noise is taken as 0.2 times the standard deviation of the signal in the event window, and the residual energy is lower than the original signal energy. This serves as a stopping criterion to ensure that the decomposition results are reproducible; For each Calculate its Hilbert envelope and determine its consistency with the trigger time and power frequency phase under the trigger alignment coordinates. The trigger time consistency is determined by the "maximum value of the envelope time and the time of trigger time phase". The time difference does not exceed "As a criterion, the power frequency phase consistency is determined by the difference between the power frequency phase corresponding to the moment of maximum envelope value and the power frequency zero-crossing reference phase detected in step S1." As a criterion, modal components that satisfy the above two criteria are preserved and concatenated into the pulse segmentation network input tensor according to the signal source. ,in The channel number is represented by the concatenation of three modal components retained from the partial discharge high-frequency current signal, three modal components retained from the ultra-high frequency electromagnetic signal, and three modal components retained from the acoustic-magnetic synchronization detection signal according to the channel dimension. This indicates the number of sampling points for each event window; The input tensor The U-Net pulse segmentation network trained on the input obtains point-by-point classification probabilities and outputs candidate discharge pulse segments, candidate reflection pulse segments, arrival times of incident pulses, arrival times of reflected pulses, and pulse segmentation confidence scores. The U-Net pulse segmentation network is a one-dimensional U-Net structure comprising an encoder and a decoder. The encoder contains four downsampling modules, each consisting of two one-dimensional convolutional layers, using a kernel length of 7, a stride of 1, and padding of the same length. The number of output channels for the convolutions is as follows: Each convolutional layer is followed by batch normalization and ReLU activation, with max pooling at the end of each level to achieve a 2x downsampling. The decoder contains four upsampling modules, each using linear interpolation for a 2x upsampling and skipping connections with the symmetric coding layer features before passing through two layers of one-dimensional convolution, batch normalization, and ReLU activation. Finally, the output layer uses a one-dimensional convolution with a kernel length of 1 to map the channels into three types of pointwise probabilities. And index each sampling point Through Softmax normalization, where Represents the probability of the background class. Represents the probability of the incident pulse class. Indicates the probability of reflected pulse type; Candidate discharge pulse segments are selected by satisfying... The fragment is obtained by performing connected component aggregation on consecutive sampling points and the fragment duration is limited to 1. to Candidate reflected pulse segments are obtained by satisfying... The fragment is obtained by performing connected component aggregation on consecutive sampling points and the fragment duration is limited to 1. to ; Incident pulse arrival time Defined as within the candidate discharge pulse segment The time coordinate corresponding to the sampling point with the maximum value, and the arrival time of the reflected pulse. Defined as within the candidate reflected pulse segment The time coordinate corresponding to the sampling point with the maximum value, and the confidence level of pulse segmentation. Defined as the mean of the corresponding class probabilities within the incident pulse segment and the reflected pulse segment, and taking the minimum of the two to characterize the reliability of "event-level" segmentation; To satisfy in terms of time A set of candidate discharge pulse segments is paired with candidate reflection pulse segments to form candidate pulse events and ( The waveform of the event is output along with the waveform of the event segment. The candidate discharge pulse segment and the candidate reflection pulse segment are subjected to continuous wavelet transform and Stockwell time-frequency transform respectively to generate time-frequency feature maps. The continuous wavelet transform uses complex Morlet wavelets and the scale number is selected. It covers a frequency range of 0.1MHz to 10MHz, and the Stockwell time-frequency transform takes values ​​within the same frequency range. Each frequency sampling point is captured within the event window. Sampling points at equal intervals are used. After normalizing the two transform amplitude spectra, the time-frequency feature maps of each segment are spliced ​​together by channel. Then, the time-frequency feature maps of the incident segment and the reflected segment are superimposed by channel to form an event-level time-frequency feature map. The U-Net pulse segmentation network, as part of the end-to-end training system, is trained using historical samples. The training samples include pulse mask labels, pulse arrival time labels, fault type labels, fault distance labels, and fault segmentation labels, and are stratified by interference labels to maintain the sample ratio between load fluctuation scenarios and environmental noise scenarios. The total loss function for end-to-end training is defined as: ; in Indicates the total loss. The pulse segmentation loss is represented by a weighted average of the three types of pointwise cross-entropy and the Dice loss. Indicates the loss by fault type classification. This represents the distance to failure regression loss. This represents the probabilistic monitoring loss for segmented faults. Represent the weights of each loss and take This allows candidate pulse events, incident pulse arrival time, reflected pulse arrival time, pulse segmentation confidence, and time-frequency feature maps to remain stable under interference scenarios and be used for feature encoding and graph constraint localization in subsequent steps.

[0021] In this specific embodiment, S3 includes: For each candidate pulse event, fixed-length waveforms of the incident pulse segment and the reflected pulse segment are extracted and used as time-domain modeling objects, with the arrival time of the incident pulse as the modeling object. Extract duration from center The incident waveform segment arrives at the time of the reflected pulse. Extract duration from center The duration of the reflected waveform segment The sampling points within the time frame are resampled to a fixed length according to the unified time grid in step S1. Each sampling point is used to stack the incident waveform segment and the reflected waveform segment according to the channel dimension to form a single-event time-domain tensor; When dividing candidate pulse events into time-domain pulse sequences according to the power frequency period, the power frequency zero-crossing time sequence of S1 is used. As a periodic boundary, each candidate pulse event is assigned to the arrival time of its incident pulse. The power frequency cycle interval And select pulse segmentation confidence level within each power frequency cycle. The largest candidate pulse event is taken as the representative event of that period, and the most recent ones are continuously selected. The length of each power frequency cycle is If there are no candidate pulse events in a certain power frequency cycle of the time-domain pulse sequence, the position of that cycle is filled with all zero waveforms and the "empty event flag" is set synchronously to avoid the model misjudging empty input as weak discharge. The temporal feature encoder, trained using the temporal pulse sequence input, generates a fault mode vector. The temporal feature encoder is composed of a one-dimensional convolutional backbone network and a Transformer self-attention network connected in series. The one-dimensional convolutional backbone network independently extracts the pulse rising edge, duration, and decay envelope features from the single-event temporal tensor of each power frequency cycle, implemented using three layers of one-dimensional convolution. The first convolutional layer has a kernel length of 15, 64 output channels, and a stride of 1; the second convolutional layer has a kernel length of 9, 128 output channels, and a stride of 1; and the third convolutional layer has a kernel length of 5, 256 output channels, and a stride of 1. After each convolutional layer, batch normalization and ReLU activation functions are sequentially connected, and max pooling is performed after the second and third convolutional layers, with a pooling kernel length of 4 and a stride of 4. Subsequently, global average pooling is performed on the temporal dimension to obtain a 256-dimensional periodic feature vector corresponding to each power frequency cycle. This vector is then mapped to the empty event flag through a linear layer and added to form a periodic input token. A Transformer self-attention network is used to generate cross-cycle evolution features, and... It consists of stacked coding layers, with each coding layer having the following number of attention heads. The dimension of the latent vector is The feedforward network dimension is And employ learnable positional coding pairs The position of each period token is marked, and a learnable classification token is added to the beginning of the sequence, and the output of the classification token is taken as the fault mode vector. For the time-frequency feature map, the event-level time-frequency feature map is divided into image blocks according to frequency bands and time windows, and input into the trained time-frequency feature encoder to generate discharge mode vectors. The event-level time-frequency feature map uses a size of... Two-dimensional grid and For the number of frequency sampling points, The number of time sampling points is determined by stacking the time-frequency feature maps of the incident and reflected segments along the channel dimension. Channel input; The division method involves dividing the time-frequency feature map according to... Each frequency sampling point and Each time sampling point forms an image patch, which is then segmented non-overlapping to obtain an image patch sequence. Each image patch is linearly mapped to form a sequence with dimension 1. Patch embedding; The time-frequency feature encoder adopts a WinTransformer structure and the window size is [missing information]. The network is divided into four stages, and patching is used sequentially to achieve resolution downsampling. The number of Transformer blocks in the four stages is as follows: The number of heads receiving multi-head attention is as follows: In each Transformer block, LayerNorm and window attention mechanism with relative position bias are used, and finally global average pooling is performed on the last layer features to obtain a discharge mode vector with a dimension of 768. The output is represented in the following form: ; in Represents the fault mode vector. Represents the discharge mode vector. The mapping function represents the temporal feature encoder. The mapping function represents the time-frequency feature encoder. This represents the input time-domain pulse sequence after being segmented and rearranged according to the power frequency cycle. The input represents the event-level time-frequency feature map generated from candidate impulse events, and the... and Together with step S2 and The dynamic physical graph constraints for subsequent step S4 and the localization regression fusion for step S5 are output together.

[0022] In this specific embodiment, S4 includes: A dynamic physical graph is constructed based on the cable length, joint location, terminal location, and line impedance parameters given in S1. The edge weights are updated event by event for each candidate pulse event output in steps S2 and S3, and the segmented fault probability vector is inferred. The node set of the dynamic physical diagram consists of cable terminal nodes, joint nodes, cable segment nodes, and sensor nodes. The positions of the cable terminal nodes and joint nodes are written in their engineering coordinates along the cable length direction. The cable segment nodes are determined according to the rule that "adjacent terminals and joints form a segment", and each segment node is bound to the length of the segment, the frequency-related attenuation coefficient of the segment, and the joint reflection coefficient of the adjacent joints of the segment. The sensor nodes are bound to the sensor installation position and channel identifier. The set of edges in the dynamic physical graph includes detection edges from sensor nodes to cable segment nodes, propagation edges between adjacent cable segment nodes, and reflection edges between joint nodes and their adjacent cable segment nodes. Detection edges represent "the detectable association of the sensor channel with respect to the cable segment", propagation edges represent "the topological connectivity of pulses propagating between cable segments", and reflection edges represent "the topological relationship of pulses being reflected at the joint and coupled back to the adjacent segments". For each sensor channel, the channel noise level is statistically analyzed within the non-discharge pulse time window and recorded as the channel noise level. ,in The sensor channel index is defined as follows: the channel noise level is obtained by fusing the root mean square value and the normalized UHF background amplitude within this time window and is limited to a closed interval between 0 and 1. The higher the channel noise level, the stronger the background noise. The sensor reliability is determined and recorded as the sensor reliability based on the historical effective sampling rate, channel noise level, and sensor installation location calibration coefficient. The historical effective sampling rate is the ratio of the number of effective sampling points to the number of sampling points required for that channel within the most recent 24 hours, and is limited to a closed interval between 0 and 1. The sensor installation position calibration coefficient is given by a fixed rule: 1.00 for installation at the terminal, 0.95 for installation at the connector, and 0.90 for installation at the intermediate segment. The sensor reliability is weighted by 0.50 for the historical effective sampling rate and low noise level. The weighted summation, with a weight of 0.30 for the installation position calibration coefficient and a weight of 0.20 for the installation position calibration coefficient, is limited to a closed interval between 0 and 1. For the candidate pulse event, read the interference flag output in step S1 that is aligned with the arrival time of the incident pulse of the event. And read the pulse segmentation confidence score output in step S2. ,in Indicates the load fluctuation state and takes a value of 0 or Indicates a noise state and takes a value of 0 or This indicates the reliability of the event segmentation and is limited to a closed interval between 0 and 1; The initial edge weights are generated based on the cable length, medium propagation speed, frequency-dependent attenuation coefficient, connector reflection coefficient, sensor installation location, and channel noise level, and are recorded as initial edge weights. ,in For edge indexing in the dynamic physical graph, different determination rules are used for different types of edges. The initial value of the edge weight of the detection edge is calculated and limited to a closed interval of 0.01 to 1 according to the rule that "the shorter the distance from the sensor to the segment along the cable, the greater the edge weight; the smaller the segment attenuation coefficient, the greater the edge weight; the smaller the channel noise level, the greater the edge weight; and the larger the installation position calibration coefficient, the greater the edge weight". The initial value of the edge weight of the propagation edge is calculated and limited to a closed interval of 0.01 to 1 according to the rule that "the smaller the equivalent propagation loss of adjacent segments, the greater the edge weight". The initial value of the edge weight of the reflection edge is calculated and limited to a closed interval of 0.01 to 1 according to the rule that "the larger the absolute value of the joint reflection coefficient, the greater the edge weight". Generate a boundary weight correction value based on the initial boundary weight value and record it as the boundary weight correction value. The edge weight correction consists of three types: detection edge correction, propagation edge correction, and reflection edge correction, and its values ​​are limited to a certain range. Within the closed interval, the detection edge correction amount depends on the sensor reliability. Channel noise level and pulse segmentation confidence Determine and execute according to fixed rules, that is, first calculate the detection confidence score as "". and "arithmetic mean" and Then subtract 0.10 from the score. When the detection confidence score is not less than 0.80, the detection edge correction is +0.20. The detection edge correction is set to 0.00 when the detection confidence score is less than 0.60, and the detection edge correction is set to -0.20 when... If the value is not less than 0.80, subtract 0.10 from the above values. The propagation edge correction is determined based on the medium propagation velocity, frequency-related attenuation coefficient, load fluctuation state in the interference marker, and current load current change rate, and is executed according to a fixed rule, i.e., when... The propagation edge correction value is 0.00 when... The current load current change rate calculated by S1 is read and the propagation side correction is taken in three levels according to its ratio with the preset load change threshold. When the ratio is not less than 1.00, the propagation side correction is taken as -0.20. When the propagation edge correction is -0.10, and when the ratio is less than 0.50, the propagation edge correction is -0.05. When the frequency correlation attenuation coefficient of the segment corresponding to the propagation edge is not less than the preset attenuation threshold, the above value is further reduced by 0.05 to reflect the increased propagation uncertainty under high attenuation conditions. The reflection edge correction is determined based on the joint reflection coefficient, joint position, and arrival time of the reflected pulse, and is executed according to a fixed rule. That is, the difference between the arrival time of the incident pulse output by S2 and the arrival time of the reflected pulse is used as the two-way propagation time, and the two-way propagation distance is calculated by combining the medium propagation speed. Then, the two-way propagation distance is projected onto the set of possible joint positions on the cable geometry and the joint with the smallest distance error is selected as the consistent joint. When the minimum distance error is not greater than 15m, the reflection edge correction of the consistent joint is +0.20 and the reflection edge correction of the other joints is -0.10. When the minimum distance error is less than 15m but the absolute value of the joint reflection coefficient of the consistent joint is less than the preset reflection threshold, the reflection edge correction of the consistent joint is 0.00 to avoid the weak reflection being excessively amplified. The initial edge weights and the adjusted edge weights are fused to obtain the corrected edge weights, which are then used as the input edge weights for the graph attention network. The fusion method satisfies the following: ; in Representing an edge The revised right, Representing an edge The initial value of the edge weight, Representing an edge The edge weight correction amount Indicates the edge index in the dynamic physics graph; The corrected edge weights and node features are input into a graph attention network to generate a segmented fault probability vector. The node features are obtained by concatenating static physical features and event condition features. The static physical features include segment length, segment attenuation coefficient, reflection coefficient of adjacent joints, sensor installation location, and channel noise level. With sensor reliability The event condition features include fault mode vectors and discharge mode vectors, which are broadcast and spliced ​​to all nodes so that graph reasoning is modulated by the candidate pulse event; The graph attention network employs a two-layer graph attention structure. The first layer uses eight attention heads, maps node features to a 64-dimensional latent space, activates them using Leaky ReLU, and applies a dropout of 0.20. The second layer uses one attention head, maps the latent vectors to the fault logit of each cable segment node, and performs Softmax normalization on all cable segment nodes to obtain the segment fault probability vector. The attention coefficients are applied to each edge with adjusted edge weights. As a multiplicative gating factor, low-reliability channels, high-noise channels, load fluctuation propagation edges, and inconsistent reflection edges are suppressed in message passing. Finally, the correlation weight between each cable segment and the candidate pulse event, as well as the segment fault probability vector, are output and passed to step S5 for location regression and confidence interval generation.

[0023] In this specific embodiment, S5 includes: For the segmented fault probability vector and the attention weights output or retained by the graph attention network on the detection edges, combined with the fault mode vector With discharge mode vector And combined with the arrival time of the incident pulse With the arrival time of the reflected pulse The positioning regression head generates the fault type, fault section, fault point distance and positioning information interval, and further outputs the maintenance interval or retest channel. This implementation method classifies cables according to the location of their terminals and connectors. Each cable segment is indexed by segment index. The segmented fault probability vector output in step S4 is denoted as... And its first Each component Indicates the first The segmented failure probability and satisfy And define the fault section identifier as and Take The largest segment index; The localization regression head employs a multi-task fully connected structure and uses a single candidate impulse event as the inference unit. Its input vector is generated by... Normalization arrival time and Cascaded to obtain, among which This refers to the duration of the event window in S2. The shared trunk route for the location return head consists of three fully connected layers with layer widths as follows: Each layer uses ReLU activation, and dropout with a dropout rate of 0.20 is applied after the first and second layers. Fault type branches, fault segment branches, and fault distance branches are set based on the shared trunk. The fault type branch outputs... The fault type probability is used to represent four types of faults: partial discharge, connector abnormality, terminal abnormality, and external interference. The fault section branch output... The piecewise logit is divided into segments and normalized by Softmax to obtain the segmented probabilities, which are then compared with... Consistent alignment, fault distance branch outputs initial value of fault point distance. And the unit is meters and is limited to Within the interval, The cable length provided for step S1; To perform physical consistency correction, this implementation method specifies the medium propagation speed based on the line parameters. The two-way propagation distance obtained from the reflection constraint is calculated using the time difference of arrival. ,in Indicates the two-way transmission distance. Indicates the propagation speed of the medium. Indicates the arrival time of the reflected pulse. Indicates the arrival time of the incident pulse, and then in the topology... Limit to faulty section identifier The start and end position intervals of the corresponding segments are connected with Linear fusion is performed to obtain the distance to the fault point after physical consistency correction. The confidence level in the fusion is segmented by pulse. As a fusion weight, this allows high-confidence events to rely more on deep regression outputs and low-confidence events to rely more on reflection physical constraints; The distance of the fixed location information interval around the fault point To generate and quantify positioning uncertainty, the difference between the marked fault distance and the fault point distance after physical consistency correction in historical samples is first statistically analyzed offline to form a positioning residual distribution and solidify its confidence level. quantile As the basic residual scale, among which This indicates the distribution of positioning residuals at the confidence level. The preset quantile value below; Secondly, based on Calculate the probabilistic entropy of segmented faults and press it Normalized to a closed interval between 0 and 1, where The larger the value, the more dispersed the segmented judgments and the higher the uncertainty. Represents the natural logarithm; Again based on sensor reliability The reliability reduction coefficient is obtained by weighting and summing the data for each sensor channel using the detection side attention weights. and Limited to a closed interval between 0 and 1, where The smaller the value, the more the positioning relies on low-reliability or high-noise channels, thus requiring a larger confidence interval. Based on this, the interval mapping function will be used to... and Map to half-width intervals and generate positional confidence intervals: ; in This indicates the lower bound of the positional information interval. This indicates the upper bound of the positional information interval. This indicates the distance to the fault point after physical consistency correction. Indicates the half-width of the interval. This represents the preset quantile value of the location residual distribution. Represents the entropy gain coefficient and takes This represents the normalized segmented fault probability entropy. This represents the reliability reduction gain coefficient and takes... This represents the reliability reduction factor; The length of the positional information interval is defined as and the preset maintenance length threshold Comparison, when the interval length is less than or equal to The output includes And the boundary is The inspection range is specified, and the probability of the fault type and the fault section identifier are output simultaneously. When the interval length exceeds The system outputs the sensor channels and candidate segments that need to be retested. The retest channels are determined based on the attention weights output or retained on the detection edges from the sensor nodes to the cable segment nodes by the attention network in step S4. Specifically, for each sensor channel... The attention contribution is calculated by summing the attention weights of the detection edges connected to all segments, and the attention contribution is not less than a preset contribution threshold. And channel noise level Exceeding the preset channel noise threshold The sensor channels marked as those requiring retesting are selected based on the candidate segments. Determine and segment the probability of faults. Not less than the preset segmentation probability threshold The cable segments are output as candidate segments and output together with the sensor channels that need to be retested to guide on-site retesting and narrow down the scope of subsequent investigations.

[0024] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0025] This invention combines multi-source detection data, deep learning features, and cable physical topology constraints in the fault location process, forming a continuous data processing chain between partial discharge signals, reflected pulse arrival times, fault mode characteristics, and segmented fault probabilities. This enables the output of more specific fault sections and maintenance intervals to address the problem of unstable location in scenarios with load fluctuations and mixed environmental noise.

[0026] This invention employs a dynamic physical graph constraint module, which dynamically corrects edge weights using load current change rate, environmental noise index, and pulse segmentation confidence. It also generates non-consistent scores based on historical location residuals, segmented fault probability entropy, and sensor reliability, so that the fault point distance output simultaneously includes interval confidence results and retest channel prompts, facilitating subsequent maintenance decisions.

Claims

1. A cable line fault location method based on deep learning, characterized in that, include: S1. Acquire multi-source status data during cable line operation, perform time synchronization, abnormal sampling and elimination, power frequency phase alignment and amplitude normalization on the multi-source status data, generate synchronized status data, and generate interference tags based on load current change rate and environmental noise index. S2. Perform adaptive mode decomposition, pulse segmentation and time-frequency feature extraction on the partial discharge signal and acoustic-magnetic synchronization detection signal in the synchronization state data to generate candidate pulse events, incident pulse arrival time, reflected pulse arrival time, pulse segmentation confidence and time-frequency feature map. S3. Input the candidate pulse event into the trained time-domain feature encoder to generate a fault mode vector, and input the time-frequency feature map into the trained time-frequency feature encoder to generate a discharge mode vector. S4. Construct a dynamic physical graph using cable terminals, joints, cable segments, and sensors to determine sensor reliability. Adjust edge weights based on line parameters, sensor reliability, interference flags, and pulse segmentation confidence. Generate segment failure probability vectors through a graph attention network. S5. Input the fault mode vector, discharge mode vector, incident pulse arrival time, reflected pulse arrival time, and segmented fault probability vector into the positioning regression head to generate the fault type, fault segment, fault point distance, and positioning information interval. Then, output the maintenance interval or retest channel based on the comparison result between the interval length of the positioning information interval and the preset maintenance length threshold.

2. The cable line fault location method based on deep learning according to claim 1, characterized in that, S1 includes: The multi-source state data includes partial discharge high-frequency current signal, ultra-high frequency electromagnetic signal, acoustic-magnetic synchronous detection signal, sheath current, load current, ambient temperature and humidity, cable length, joint position, terminal position and line impedance parameters; data with different sampling rates are resampled according to a unified time reference, and the power frequency zero crossover time is used as the phase alignment reference; The environmental noise index is obtained by weighting the root mean square value of the signal within the non-discharge pulse time window and the amplitude of the ultra-high frequency background. The interference flag is determined by whether the load current change rate exceeds a preset load change threshold and whether the environmental noise index exceeds a preset noise threshold.

3. The cable line fault location method based on deep learning according to claim 1, characterized in that, S2 includes: Adaptive mode decomposition is performed on the partial discharge high-frequency current signal, ultra-high frequency electromagnetic signal and acoustic-magnetic synchronous detection signal respectively, and the mode components consistent with the power frequency phase and sensor trigger time are retained. The retained modal components are input into the trained U-Net pulse segmentation network, which outputs candidate discharge pulse segments, candidate reflection pulse segments, incident pulse arrival time, reflection pulse arrival time, and pulse segmentation confidence. The candidate pulse events are formed from the candidate discharge pulse segments and candidate reflection pulse segments. Wavelet transform and Stockwell time-frequency transform are performed on the candidate discharge pulse segment and the candidate reflection pulse segment, respectively, to generate the time-frequency feature map.

4. The cable line fault location method based on deep learning according to claim 1, characterized in that, S3 includes: dividing the candidate pulse event into a time-domain pulse sequence according to the power frequency cycle; extracting the pulse rising edge, duration, and attenuation envelope features using a one-dimensional convolutional layer; and generating cross-cycle evolution features as the fault mode vector using a Transformer self-attention layer; dividing the time-frequency feature map into image blocks according to the frequency band and time window; and generating the discharge mode vector using a SwingTransformer window attention layer.

5. The cable line fault location method based on deep learning according to claim 1, characterized in that, S4 includes: Cable terminal nodes, joint nodes, cable segment nodes, and sensor nodes are used as nodes in a dynamic physical graph. The detection edges from the sensor to the cable segment, the propagation edges of adjacent cable segments, and the reflection edges at the joint are used as edges in the dynamic physical graph. The channel noise level is determined based on the noise statistics of the corresponding sensor channel within the non-discharge pulse time window. Initial edge weights are generated using cable length, medium propagation velocity, frequency correlation attenuation coefficient, joint reflection coefficient, sensor installation location, and channel noise level. Sensor reliability is determined based on historical effective sampling rate, channel noise level, and sensor installation location calibration coefficient. Edge weight corrections are generated based on the interference marker and its corresponding load current change rate, environmental noise index, and pulse segmentation confidence level. The initial edge weights and edge weight corrections are input into a graph attention network to obtain the association weights between each cable segment and candidate pulse events, and the segment fault probability vector.

6. The cable line fault location method based on deep learning according to claim 1, characterized in that, S5 includes: inputting the fault mode vector, discharge mode vector, incident pulse arrival time, reflected pulse arrival time, and segmented fault probability vector into the positioning regression head; the positioning regression head outputs the fault type probability, fault segment identifier, and initial value of fault point distance; performing physical consistency correction based on the initial value of fault point distance, propagation velocity, and connector position to obtain the fault point distance; generating a positioning information interval based on the fault point distance; when the interval length of the positioning information interval is less than or equal to a preset maintenance length threshold, outputting a maintenance interval containing the fault point distance; when the interval length of the positioning information interval exceeds the preset maintenance length threshold, outputting the sensor channel and candidate segment that need to be retested.

7. The cable line fault location method based on deep learning according to claim 3, characterized in that, The trained U-Net pulse segmentation network, temporal feature encoder, time-frequency feature encoder, graph attention network, and localization regression head are trained end-to-end using historical samples with fault type labels, fault distance labels, pulse arrival time labels, pulse mask labels, and fault segmentation labels. The training loss includes pulse segmentation loss, fault type classification loss, fault distance regression loss, and segmented probability supervision loss for the output of the graph attention network. The training samples are stratified according to interference labels to maintain the sample ratio of load fluctuation scenarios and environmental noise scenarios.

8. The cable line fault location method based on deep learning according to claim 5, characterized in that, The edge weight correction in step S4 includes detection edge correction, propagation edge correction, and reflection edge correction. The detection edge correction is determined based on sensor reliability, channel noise level, and pulse segmentation confidence. The propagation edge correction is determined based on medium propagation velocity, frequency correlation attenuation coefficient, load fluctuation state in the interference marker, and current load current change rate. The reflection edge correction is determined based on connector reflection coefficient, connector position, and arrival time of the reflected pulse. The edge weight correction is used to perform weighted updates on the initial edge weight values.

9. The cable line fault location method based on deep learning according to claim 6, characterized in that, The location information interval in step S5 is generated as follows: a location residual distribution is generated based on the difference between the marked fault distance in historical samples and the fault point distance after physical consistency correction; the segmented fault probability entropy is calculated based on the segmented fault probability vector; a reliability reduction coefficient is generated based on the sensor reliability; the preset quantile value of the location residual distribution, the segmented fault probability entropy, and the reliability reduction coefficient are input into the interval mapping function to obtain the inconsistency score, and the location information interval is generated around the fault point distance after physical consistency correction based on the inconsistency score and the preset information level.

10. The cable line fault location method based on deep learning according to claim 5, characterized in that, The retest channel in step S5 is determined as follows: when the interval length of the fixed position signal interval exceeds the preset maintenance length threshold, the attention contribution of each sensor node is summarized according to the attention weight output or retained on the detection edge from the sensor node to the cable segment by the graph attention network; the sensor channel whose attention contribution is not less than the preset contribution threshold and whose channel noise level exceeds the preset channel noise threshold is marked as the sensor channel that needs to be retested. Cable segments with a segment failure probability not less than a preset segment probability threshold are identified as candidate segments.

Images