Power cable fault type diagnosis method and system based on deep learning

Abstract

The application provides a kind of power cable fault type diagnosis method and system based on deep learning, it is related to deep learning technical field, obtains the original light time domain scattering signal data set collected by distributed sensing device, carries out scattering spectrum feature demodulation processing, obtains scattering frequency shift feature sequence and scattering spectrum broadening feature sequence;Call fault feature space-time correlation coding network to carry out position coding injection and multi-head self-attention transformation processing to two kinds of feature sequences, generate cable state deep time series dependent representation vector containing frequency shift and broadening long-range time series correlation mode global context mapping relationship;Call fault type hierarchical diagnosis network to carry out multilevel fault feature layer-by-layer diagnosis processing, generate fault type attribution probability distribution vector and fault boundary space sampling point position identification sequence;Accordingly generate cable fault diagnosis report data set and send to operation and maintenance management terminal.The application realizes the probabilistic diagnosis of cable fault type and the accurate positioning of fault boundary.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and more specifically, to a method and system for diagnosing power cable fault types based on deep learning. Background Technology

[0002] During long-term operation, power cables are subjected to the coupled effects of multiple factors, including electric field stress, thermal stress, mechanical forces, and environmental corrosion. This makes them highly susceptible to faults such as insulation aging, partial discharge, short circuits, and even wire breaks at weak points like cable joints. Cable joints, due to their complex structure and concentrated electric field, have historically been high-risk areas for these faults. Accurate and rapid diagnosis of cable fault types and precise location of fault areas are of significant engineering importance for shortening power outage time and reducing operation and maintenance costs.

[0003] Currently, cable online monitoring methods based on distributed fiber optic sensing technology have been widely applied, among which optical time-domain scattering (OTS) technology has attracted much attention due to its ability to continuously acquire scattered signals along the cable axis. Existing technologies typically process OTS signals using time-domain waveform analysis, directly comparing amplitude thresholds or matching waveform similarity to determine the approximate location and severity of faults. However, such methods only utilize the amplitude information of the scattered signal in the time domain, completely ignoring the frequency dimension characteristics inherent in the scattered signal, particularly the deep spectral information reflecting changes in the cable's microscopic physical state, such as the scattering frequency shift and scattering spectrum broadening parameters. This results in a severe deficiency in the ability to identify early, subtle faults and their types.

[0004] On the other hand, existing technologies mostly employ single architectures such as convolutional neural networks or recurrent neural networks to classify processed signal features. However, when dealing with long-sequence features composed of a large number of spatial sampling points distributed along the cable axis, these models struggle to effectively capture the complex spatiotemporal relationships between different spatial locations and different acquisition times. In particular, they cannot establish a global context mapping relationship between long-range temporal correlation patterns of frequency shift features and long-range temporal correlation patterns of broadened features, which significantly limits the diagnostic accuracy of fault types and the location accuracy of fault boundaries. Furthermore, existing diagnostic methods typically only output a single fault classification result, lacking a quantitative expression of the probability of fault type attribution and a precise definition of fault boundary regions, making it difficult to meet the actual needs of refined operation and maintenance management. Summary of the Invention

[0005] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, the present invention provides a method for diagnosing power cable fault types based on deep learning, the method comprising: Acquire the raw optical time-domain scattering signal data set collected by the distributed sensing device deployed at the cable joint. The raw optical time-domain scattering signal data set includes the location markers of spatial sampling points distributed along the cable axis and the scattering radio frequency shift time-domain response curve corresponding to each spatial sampling point. The original optical temporal scattering signal data set is subjected to scattering spectrum feature demodulation processing to obtain the scattering frequency shift feature sequence corresponding to the spatial sampling point position identifiers distributed along the cable axis and the scattering spectrum broadening feature sequence corresponding to each spatial sampling point position identifier. The scattering frequency shift feature sequence includes the scattering frequency shift of each spatial sampling point position identifier at different acquisition times, and the scattering spectrum broadening feature sequence includes the scattering spectrum half-width parameter of each spatial sampling point position identifier at different acquisition times. The pre-built fault feature spatiotemporal correlation coding network is invoked to perform position coding injection and multi-head self-attention transformation processing on the scattered frequency shift feature sequence and the scattering spectrum broadening feature sequence, generating a deep temporal dependency representation vector of cable state corresponding to the position identifier of each spatial sampling point. The deep temporal dependency representation vector of cable state contains the global context mapping relationship between the frequency shift long-range temporal correlation mode feature and the broadening long-range temporal correlation mode feature. The pre-built fault type hierarchical diagnosis network is invoked to perform multi-level fault feature layer-by-layer diagnosis processing on the deep temporal dependency representation vector of the cable state, generating the fault type attribution probability distribution vector and fault boundary spatial sampling point location identifier sequence corresponding to each spatial sampling point location identifier; Based on the fault type attribution probability distribution vector and the fault boundary space sampling point location identifier sequence, a cable fault diagnosis report data set is generated, and the cable fault diagnosis report data set is sent to the cable operation and maintenance management terminal.

[0006] Furthermore, this invention also provides a deep learning-based power cable fault type diagnosis system, comprising: A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to perform the aforementioned deep learning-based power cable fault type diagnosis method by executing the machine-executable instructions.

[0007] Based on the above, by acquiring the original optical temporal scattering signal data set containing spatial sampling point location identifiers and scattering frequency shift time-domain response curves collected by distributed sensing devices, and performing scattering spectrum feature demodulation processing on it, the scattering frequency shift feature sequences and scattering spectrum broadening feature sequences corresponding to each spatial sampling point along the cable axis are extracted. On this basis, a pre-constructed fault feature spatiotemporal correlation coding network is invoked to perform position coding injection and multi-head self-attention transformation processing on the scattering frequency shift feature sequences and scattering spectrum broadening feature sequences, generating a deep temporal dependence representation vector of the cable state containing the global context mapping relationship between frequency shift long-range temporal correlation mode features and broadening long-range temporal correlation mode features, so that the relationships between different spatial sampling points and different acquisition times are understood. The complex spatiotemporal dependencies between moments are systematically captured and encoded in a unified high-dimensional representation space. Then, a pre-built fault type hierarchical diagnostic network is invoked to perform multi-level fault feature layer-by-layer diagnostic processing on this deep temporal dependency representation vector. This generates a fault type attribution probability distribution vector and a fault boundary spatial sampling point location identifier sequence for each spatial sampling point. This gives the fault type diagnosis results a probability-quantified credibility measure, while the fault boundary location results have an accurate spatial sequence expression. Finally, based on the above diagnostic results, a cable fault diagnosis report dataset is generated and merged and sent to the operation and maintenance management terminal, which significantly improves the overall performance of cable fault diagnosis in terms of type identification accuracy, boundary location accuracy, and operation and maintenance decision support capabilities. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the execution flow of the deep learning-based power cable fault type diagnosis method provided by the present invention.

[0009] Figure 2 This is a logical schematic diagram of the power cable fault type diagnosis method based on deep learning provided by the present invention.

[0010] Figure 3 This is a schematic diagram of exemplary hardware and software components of the deep learning-based power cable fault type diagnosis system provided by the present invention. Detailed Implementation

[0011] Figure 1 This is a flowchart illustrating a deep learning-based power cable fault type diagnosis method according to an embodiment of the present invention. The following is a summary of the process. Figure 2 A detailed introduction is provided. This embodiment applies the entire technical solution to a scenario of fault type diagnosis for an underground urban cable line. The cable line is several kilometers long, and distributed fiber optic sensing devices are deployed at multiple cable joints. The distributed fiber optic sensing devices operate based on the Brillouin time-domain scattering principle, setting 80 spatial sampling points numbered P01 to P80 at fixed spatial intervals along the cable axis to continuously collect raw optical time-domain scattering signal data.

[0012] Step S110: Obtain the original optical time-domain scattering signal data set collected by the distributed sensing device deployed at the cable joint. The original optical time-domain scattering signal data set includes the location identifiers of spatial sampling points distributed along the cable axis and the scattering radio frequency shift time-domain response curve corresponding to each spatial sampling point.

[0013] The distributed fiber optic sensing device incorporates a tunable laser source to inject pulsed probe light into the single-mode sensing fiber, receiving the backscattered Brillouin light returning along the fiber. After coherent detection and analog-to-digital conversion of the returned scattered light, the original optical time-domain scattering signal data set is obtained. This data set is a two-dimensional data structure, with row indices corresponding to 80 spatial sampling point location identifiers P01 to P80. Each spatial sampling point location identifier uniquely corresponds to a physical location coordinate along the cable axis. Each row stores one scattered radio frequency shift time-domain response curve, which is a one-dimensional array arranged in chronological order of acquisition time. Array elements represent scattered radio frequency shifts in megahertz, and the array length is equal to the number of time-domain sampling points in a single acquisition cycle, denoted as N. The time axes of all 80 scattered radio frequency shift time-domain response curves are aligned using a timing synchronization pulse. The original optical time-domain scattering signal data set is transmitted back to the cable condition analysis server in real time via industrial Ethernet.

[0014] Step S120: Perform scattering spectrum feature demodulation processing on the original optical time-domain scattering signal data set to obtain the scattering frequency shift feature sequence corresponding to the spatial sampling point position identifiers distributed along the cable axis and the scattering spectrum broadening feature sequence corresponding to each spatial sampling point position identifier. The scattering frequency shift feature sequence includes the scattering frequency shift of each spatial sampling point position identifier at different acquisition times, and the scattering spectrum broadening feature sequence includes the scattering spectrum half-width parameter of each spatial sampling point position identifier at different acquisition times.

[0015] The original optical time-domain scattering signal data set obtained in step S110 is subjected to scattering spectrum feature demodulation processing, and two key feature parameters, frequency shift and broadening, are extracted from each scattering frequency shift time-domain response curve.

[0016] Step S121: Extract the scattering radio-shift time-domain response curve corresponding to the spatial sampling point location identifier from the original optical time-domain scattering signal data set, perform time-domain segmentation processing on the scattering radio-shift time-domain response curve, and divide each scattering radio-shift time-domain response curve into multiple time-domain analysis window segments according to the acquisition time sequence.

[0017] Extract the scattering radio-shift time-domain response curve corresponding to each spatial sampling point location identifier from the original optical temporal scattering signal dataset. Perform time-domain segmentation processing on each scattering radio-shift time-domain response curve using a fixed-length sliding window method. Set the window length of the time-domain analysis window segment to W, and the sliding step size between adjacent time-domain analysis window segments to S, where W and S are in units of the number of time-domain sampling points. Starting from the initial sampling point of the scattering radio-shift time-domain response curve, extract a time-domain analysis window segment of length W every S time-domain sampling points until the end of the scattering radio-shift time-domain response curve is reached. The total number of time-domain analysis window segments corresponding to each spatial sampling point location identifier is denoted as K, where K = floor((NW) / S) + 1. Perform this operation on spatial sampling point location identifiers P01 to P80, resulting in K time-domain analysis window segments for each spatial sampling point location identifier.

[0018] Step S122: Perform Lorentz line fitting on the time-domain response curve of the scattered radio frequency shift within each time-domain analysis window segment, and extract the center frequency offset of the fitted Lorentz curve within each time-domain analysis window segment as the scattered radio frequency shift corresponding to that time-domain analysis window segment.

[0019] The scattered radio shift time-domain response curves within each time-domain analysis window segment generated in step S121 are subjected to Lorentz line fitting processing using the Levenberg-Marquardt nonlinear least squares fitting algorithm. The mathematical form of the Lorentz function is: Where A is the peak amplitude, f0 is the center frequency offset, and G is the full width at half maximum (FWHM) parameter. The fitting process uses the sampled data points of the divergent radio frequency shift time-domain response curve within the time-domain analysis window as observations. The values ​​of the three parameters to be fitted, A, f0, and G, are iteratively adjusted to minimize the sum of squared residuals. The sum of squared residuals is equal to the sum of the squares of the differences between the observed values ​​and L(f). After the fitting converges, the center frequency offset f0 of the fitted Lorentz curve corresponding to the time-domain analysis window is extracted as the divergent radio frequency shift corresponding to that time-domain analysis window.

[0020] Step S123: Arrange the scattered radio shift quantities of each spatial sampling point location identifier in all time domain analysis window segments according to the acquisition time sequence to form the scattered radio shift feature sequence corresponding to the spatial sampling point location identifier.

[0021] For each spatial sampling point location identifier, the scattered radio frequency shifts f0 corresponding to the K time-domain analysis window segments extracted in step S122 are arranged sequentially according to the starting acquisition time of the time-domain analysis window segments, forming a one-dimensional array of length K. This one-dimensional array is the scattered radio frequency shift feature sequence corresponding to the spatial sampling point location identifier. The k-th element in the scattered radio frequency shift feature sequence corresponds to the scattered radio frequency shift of the spatial sampling point location identifier in the k-th time-domain analysis window segment. This arrangement operation is performed on all spatial sampling point location identifiers P01 to P80, resulting in a total of 80 scattered radio frequency shift feature sequences.

[0022] Step S124: When performing Lorentz line fitting on the scattering frequency shift time-domain response curve within each time-domain analysis window segment, simultaneously extract the full width at half maximum (FWHM) parameter of the fitted Lorentz curve as the FWHM parameter of the scattering spectrum corresponding to that time-domain analysis window segment.

[0023] In step S122, during the Lorentz line fitting process of the scattering radio shift time-domain response curve within each time-domain analysis window segment, the full width at half maximum (FWHM) parameter G in the Lorentz function after fitting convergence is simultaneously extracted. This FWHM parameter G is used as the FWHM parameter of the scattering spectrum corresponding to that time-domain analysis window segment. This FWHM parameter characterizes the spectral broadening of the Brillouin scattering spectrum.

[0024] Step S125: Arrange the scattering spectrum half-width parameters of each spatial sampling point location identifier in all time domain analysis window segments according to the acquisition time order, to form the scattering spectrum broadening feature sequence corresponding to the spatial sampling point location identifier.

[0025] For each spatial sampling point location identifier, the scattering spectrum half-width parameters G corresponding to the K time-domain analysis window segments extracted in step S124 are arranged sequentially according to the starting acquisition time of the time-domain analysis window segments, forming a one-dimensional array of length K. This one-dimensional array is the scattering spectrum broadening feature sequence corresponding to the spatial sampling point location identifier. The k-th element in the scattering spectrum broadening feature sequence corresponds to the scattering spectrum half-width parameter of the spatial sampling point location identifier in the k-th time-domain analysis window segment. This arrangement operation is performed on all spatial sampling point location identifiers P01 to P80, resulting in a total of 80 scattering spectrum broadening feature sequences.

[0026] Step S126: The scattering frequency shift feature sequence and scattering spectrum broadening feature sequence corresponding to each spatial sampling point location identifier are spliced ​​along the feature dimension to generate a scattering spectrum dual-channel feature matrix corresponding to each spatial sampling point location identifier. The scattering spectrum dual-channel feature matrix is ​​then post-processed to obtain the final output scattering frequency shift feature sequence and scattering spectrum broadening feature sequence.

[0027] For each spatial sampling point location identifier, the scattered radio frequency shift feature sequence generated in step S123 and the scattering spectrum broadening feature sequence generated in step S125 are concatenated along the feature dimension. The concatenation operation involves stacking the scattered radio frequency shift feature sequence and the scattering spectrum broadening feature sequence of length K along the second dimension to generate a scattering spectrum dual-channel feature matrix of shape 2*K. The first row of the scattering spectrum dual-channel feature matrix stores the scattered radio frequency shift feature sequence of the spatial sampling point location identifier, and the second row stores the scattering spectrum broadening feature sequence of the spatial sampling point location identifier. Then, post-processing operations are performed on the scattering spectrum dual-channel feature matrix, including temporal consistency alignment, abnormal amplitude constraint processing, and normalization processing.

[0028] Step S1261: Perform time-series consistency alignment processing on the scattered radio frequency shift feature sequence and the scattering spectrum broadening feature sequence respectively, and match the timestamps of each time-domain analysis window segment in the scattered radio frequency shift feature sequence with the timestamps of each time-domain analysis window segment in the scattering spectrum broadening feature sequence.

[0029] Each time-domain analysis window segment in the scattered radio shift feature sequence corresponds to a timestamp, and each time-domain analysis window segment in the scattering spectrum broadening feature sequence also corresponds to a timestamp. Both timestamps are taken from the starting acquisition time of the same time-domain analysis window segment. The time sequence consistency alignment process involves verifying and matching the timestamp of the k-th element of the scattered radio shift feature sequence with that of the k-th element of the scattering spectrum broadening feature sequence to confirm that they belong to the same time-domain analysis window segment. After the matching verification is successful, the order of the sequence elements remains unchanged.

[0030] Step S1262: Perform abnormal amplitude constraint processing on the scattered frequency shift feature sequence and the scattering spectrum broadening feature sequence after time-series consistency alignment processing respectively. Replace the abnormal scattered frequency shift in the scattered frequency shift feature sequence that exceeds the preset reasonable frequency shift range with the scattered frequency shift interpolation result of the time-adjacent time domain analysis window segment. Replace the abnormal scattering spectrum half-width parameter in the scattering spectrum broadening feature sequence that exceeds the preset reasonable broadening range with the scattering spectrum half-width parameter interpolation result of the time-adjacent time domain analysis window segment.

[0031] A preset reasonable frequency shift interval is set as [fmin, fmax], and a preset reasonable broadening interval is set as [wmin, wmax]. For each element in the scattering frequency shift feature sequence, if the scattering frequency shift f0 corresponding to a certain element exceeds the preset reasonable frequency shift interval [fmin, fmax], then the scattering frequency shift is determined to be an abnormal scattering frequency shift. The replacement method for abnormal scattering frequency shifts is as follows: take the scattering frequency shifts corresponding to the two adjacent time-domain analysis window segments before and after the time-domain analysis window segment containing the abnormal scattering frequency shift, calculate the arithmetic mean of the two as the interpolation result, and use this interpolation result to replace the abnormal scattering frequency shift. Similarly, for each element in the scattering spectrum broadening feature sequence, if the scattering spectrum half-width parameter G corresponding to a certain element exceeds the preset reasonable broadening interval [wmin, wmax], then the scattering spectrum half-width parameter is determined to be an abnormal scattering spectrum half-width parameter, and the abnormal scattering spectrum half-width parameter is replaced using the same time-adjacent interpolation method.

[0032] Step S1263: Normalize the scattered radio frequency shift characteristic sequence and the scattering spectrum broadening characteristic sequence after abnormal amplitude constraint processing, respectively. Calculate the mean and standard deviation of the frequency shift of the scattered radio frequency shift characteristic sequence, subtract the mean from the scattered radio frequency shift and divide by the standard deviation. Calculate the mean and standard deviation of the broadening of the scattering spectrum broadening characteristic sequence, subtract the mean from the scattering spectrum half-width parameter and divide by the standard deviation.

[0033] The scattered radio frequency shift feature sequence after the abnormal amplitude constraint processing in step S1262 is normalized. First, the arithmetic mean Mf and standard deviation Sf of the K scattered radio frequency shifts in the feature sequence are calculated. Then, Z-Score normalization is performed on each scattered radio frequency shift f0 in the feature sequence. The normalized scattered radio frequency shift f0norm = (f0 - Mf) / Sf, resulting in the normalized scattered radio frequency shift feature sequence. Similarly, the scattering spectrum broadening feature sequence after the abnormal amplitude constraint processing is normalized. The arithmetic mean Mw and standard deviation Sw of the K scattering spectrum half-width parameters G are calculated. Z-Score normalization is performed on each scattering spectrum half-width parameter G. The normalized scattering spectrum half-width parameter Gnorm = (G - Mw) / Sw, resulting in the normalized scattering spectrum broadening feature sequence. The normalized scattered radio frequency shift feature sequence and the normalized scattering spectrum broadening feature sequence are used as the final output scattered radio frequency shift feature sequence and scattering spectrum broadening feature sequence.

[0034] Step S130: Call the pre-built fault feature spatiotemporal correlation coding network to perform position coding injection and multi-head self-attention transformation processing on the scattering frequency shift feature sequence and the scattering spectrum broadening feature sequence to generate a deep temporal dependency representation vector of the cable state corresponding to the position identifier of each spatial sampling point. The deep temporal dependency representation vector of the cable state contains the global context mapping relationship between the frequency shift long-range temporal correlation mode feature and the broadening long-range temporal correlation mode feature.

[0035] The final output of the scattered radio shift feature sequence and the scattering spectrum broadening feature sequence generated in step S1263 are input into the pre-constructed fault feature spatiotemporal correlation coding network. Through position coding injection and multi-head self-attention transformation, deep dependency representations across time and space are extracted.

[0036] Step S131: Obtain a pre-constructed fault feature spatiotemporal correlation coding network, wherein the fault feature spatiotemporal correlation coding network includes a cascaded position coding injection layer, a first multi-head self-attention transformation layer, a feedforward fully connected transformation layer, and a second multi-head self-attention transformation layer.

[0037] The pre-built fault feature spatiotemporal correlation encoding network employs a deep neural network based on the Transformer encoder architecture. The location encoding injection layer is responsible for injecting location-aware information into the temporal features. The first multi-head self-attention transformation layer contains H1 parallel scaled dot-product attention heads, each using an independent query weight matrix, key weight matrix, and value weight matrix to perform self-attention computation on the input feature sequence. The feedforward fully connected transformation layer contains two fully connected network layers: the first fully connected weight matrix has a dimension of D*Df, and the second fully connected weight matrix has a dimension of Df*D. A non-linear activation function is used between the two layers, where D is the dimension of the model's hidden layers, and Df is the dimension of the feedforward network's intermediate layers. The second multi-head self-attention transformation layer contains H2 parallel scaled dot-product attention heads for attention interaction across spatial sampling points. The above layers are cascaded in the order of location encoding injection layer, first multi-head self-attention transformation layer, feedforward fully connected transformation layer, and second multi-head self-attention transformation layer.

[0038] Step S132: Input the preprocessed scattered radio shift feature sequence and scattering spectrum broadening feature sequence into the position coding injection layer of the fault feature spatiotemporal correlation coding network, inject a sinusoidal position coding vector into the scattered radio shift amount corresponding to each time domain analysis window segment in the scattered radio shift feature sequence, and inject a cosine position coding vector into the scattering spectrum half-width parameter corresponding to each time domain analysis window segment in the scattering spectrum broadening feature sequence.

[0039] The scattered radio frequency shift feature sequence and the scattering spectrum broadening feature sequence, which are finally output from step S1263, are input into the position coding injection layer of the fault feature spatiotemporal correlation coding network. For the scattered radio frequency shift quantity corresponding to the k-th time domain analysis window segment in the scattered radio frequency shift feature sequence, a sinusoidal position coding vector of dimension D is generated. The element value of the 2i-th dimension in this sinusoidal position coding vector is sin(k / 10000^(2i / D)), and the element value of the 2i+1-th dimension is sin(k / 10000^(2i+1 / D)), where i is an integer index from 0 to D / 2-1. For the half-width at half-maximum (WHM) parameter of the scattering spectrum corresponding to the k-th time-domain analysis window segment in the broadened scattering spectrum feature sequence, a cosine position encoding vector of dimension D is generated. The element value of the 2i-th dimension of this cosine position encoding vector is cos(k / 10000^(2i / D)), and the element value of the (2i+1)-th dimension is cos(k / 10000^(2i+1 / D)). The scattering frequency shift is added element-wise to the sine position encoding vector, and the half-width at half-maximum (WHM) parameter of the scattering spectrum is added element-wise to the cosine position encoding vector to complete the position encoding injection. The scattering frequency shift feature sequence after injecting the sine position encoding vector is denoted as Xf, and the scattering spectrum broadened feature sequence after injecting the cosine position encoding vector is denoted as Xw. Both Xf and Xw have the shape K*D.

[0040] Step S133: Concatenate the scattering frequency shift feature sequence after injecting the sinusoidal position encoding vector and the scattering spectrum broadening feature sequence after injecting the cosine position encoding vector along the feature dimension to generate a joint scattering spectrum feature sequence with injected position-aware information.

[0041] The scattering frequency shift feature sequence Xf after the position encoding injection in step S132 and the scattering spectrum broadening feature sequence Xw are spliced ​​in the feature dimension direction, that is, the two matrices are merged side by side along the second dimension to generate a scattering spectrum joint feature sequence with shape K*(2*D) injected with position-aware information, denoted as Xj.

[0042] Step S134: Input the scattering spectrum joint feature sequence with injected position-aware information into the first multi-head self-attention transformation layer. Through multiple parallel scaling dot product attention heads in the first multi-head self-attention transformation layer, perform attention weight calculation on the joint feature vectors corresponding to any two time-domain analysis window segments in the scattering spectrum joint feature sequence to generate the first multi-head self-attention output feature sequence.

[0043] The scattering spectrum joint feature sequence Xj generated in step S133, containing injected position-aware information, is input into the first multi-head self-attention transformation layer. The first multi-head self-attention transformation layer contains H1 parallel attention heads. For each attention head h, using its independent query weight matrix Wqh, key weight matrix Wkh, and value weight matrix Wvh, Xj is linearly projected to generate a query matrix Qh = Xj × Wqh, a key matrix Kh = Xj × Wkh, and a value matrix Vh = Xj × Wvh. Then, scaled dot product attention is calculated, and the attention calculation result... Where Dk is the dimension of the key vector, KhT is the transpose of the key matrix Kh, and softmax is the flexible maximum transfer function. The outputs Ah of H1 attention heads are concatenated along the feature dimension and then multiplied by an output projection weight matrix Wo to obtain the first multi-head self-attention output feature sequence Xa1.

[0044] Step S135: Perform residual connection processing on the first multi-head self-attention output feature sequence and the scattering spectrum joint feature sequence of the injected position-aware information, and perform layer normalization processing on the residual connection processing result to obtain the first residual normalized feature sequence.

[0045] The first multi-head self-attention output feature sequence Xa1 generated in step S134 and the scattering spectrum joint feature sequence Xj generated in step S133 with injected position-aware information are subjected to residual connection processing, and the residual connection result is Xr1 = Xa1 + Xj. Then, layer normalization processing is performed on Xr1. The layer normalization process is to calculate the mean Mu1 and variance Var1 of Xr1 in the feature dimension, and the normalized output is... , where ε is a very small positive number, and γ is a learnable scaling parameter. The learnable offset parameters are used to obtain the first residual normalized feature sequence Xn1.

[0046] Step S136: Input the first residual normalized feature sequence into the feedforward fully connected transform layer, and perform nonlinear feature mapping processing on the first residual normalized feature sequence through the first fully connected weight matrix and nonlinear activation function in the feedforward fully connected transform layer to generate the feedforward transform feature sequence.

[0047] The first residual normalized feature sequence Xn1 generated in step S135 is input into the feedforward fully connected transform layer. The first fully connected weight matrix of the feedforward fully connected transform layer is Wf1, with a dimension of (2*D)*Df, and the second fully connected weight matrix is ​​Wf2, with a dimension of Df*(2*D). Xn1 undergoes two linear mappings and one nonlinear activation: first, Xf1 = Xn1 × Wf1 is calculated; then, the nonlinear activation function ReLU is applied to Xf1 to obtain Xf1a = ReLU(Xf1); finally, Xf2 = Xf1a × Wf2 is calculated to obtain the feedforward transform feature sequence Xff.

[0048] Step S137: Perform residual concatenation processing on the feedforward transform feature sequence and the first residual normalized feature sequence, and perform layer normalization processing on the residual concatenation processing result to obtain the second residual normalized feature sequence.

[0049] The feedforward transform feature sequence Xff generated in step S136 and the first residual normalized feature sequence Xn1 generated in step S135 are subjected to residual concatenation, resulting in Xr2 = Xff + Xn1. Then, layer normalization is performed on Xr2, and the mean Mu2 and variance Var2 of Xr2 in the feature dimension are calculated. The normalized output is... The second residual normalized feature sequence Xn2 is obtained.

[0050] Step S138: Input the second residual normalized feature sequence into the second multi-head self-attention transformation layer, perform cross-spatial sampling point attention interaction processing, and output the deep temporal dependency representation vector of the cable state.

[0051] The second residual normalized feature sequence Xn2 generated in step S137 is input into the second multi-head self-attention transformation layer. The second multi-head self-attention transformation layer contains H2 parallel scaled dot product attention heads, which are used to perform cross-spatial sampling point attention interaction processing on the feature vectors corresponding to the location identifiers of different spatial sampling points.

[0052] Step S1381: The second multi-head self-attention output feature sequence is generated by performing cross-spatial sampling point attention weight calculation on the feature vectors corresponding to different spatial sampling point position identifiers in the second residual normalized feature sequence through multiple parallel scaling dot product attention heads in the second multi-head self-attention transformation layer.

[0053] In the second multi-head self-attention transformation layer, the feature vectors corresponding to the location identifiers of different spatial sampling points are input simultaneously. For each attention head h, the feature vectors of all spatial sampling points are linearly projected using the query weight matrix Wqh2, the key weight matrix Wkh2, and the value weight matrix Wvh2 to calculate the attention weights across spatial sampling points. The attention calculation method is similar to step S134, but the query matrix, key matrix, and value matrix are concatenated from the feature vectors corresponding to all 80 spatial sampling point location identifiers, enabling the feature vector of each spatial sampling point location identifier to interact with the feature vectors of all other spatial sampling point location identifiers. The output feature sequence Xa2 of the second multi-head self-attention is then calculated.

[0054] Step S1382: Perform residual connection processing on the second multi-head self-attention output feature sequence and the second residual normalized feature sequence, and perform layer normalization processing on the residual connection processing result to obtain the third residual normalized feature sequence.

[0055] The second multi-head self-attention output feature sequence Xa2 generated in step S1381 and the second residual normalized feature sequence Xn2 generated in step S137 are subjected to residual concatenation, resulting in Xr3 = Xa2 + Xn2. Then, layer normalization is performed on Xr3, calculating the mean Mu3 and variance Var3 of Xr3 along the feature dimension. The normalized output is... The third residual normalized feature sequence Xn3 is obtained.

[0056] Step S1383: Output the third residual normalized feature sequence as the deep temporal dependency representation vector of the cable state. The feature vector corresponding to the spatial sampling point position identifier in the deep temporal dependency representation vector of the cable state contains the global context mapping relationship with the long-range temporal correlation mode features of scattering frequency shift and the long-range temporal correlation mode features of scattering spectrum broadening within the entire spatial sampling point range.

[0057] The third residual normalized feature sequence Xn3 generated in step S1382 is directly output as the deep temporal dependency representation vector of the cable state, denoted as Vdeep. Each spatial sampling point location identifier in Vdeep corresponds to a feature vector. This feature vector integrates the long-range temporal correlation pattern features of the scattering frequency shift of the spatial sampling point location identifier itself, the long-range temporal correlation pattern features of the scattering spectrum broadening, and the cross-spatial attention interaction information between the spatial sampling point location identifier and all other spatial sampling point location identifiers on the entire cable line.

[0058] Step S140: Call the pre-built fault type hierarchical diagnosis network to perform multi-level fault feature layer-by-layer diagnosis processing on the deep temporal dependency representation vector of the cable state, and generate the fault type attribution probability distribution vector and fault boundary spatial sampling point location identifier sequence corresponding to each spatial sampling point location identifier.

[0059] The deep temporal dependency representation vector Vdeep of the cable state output in step S1383 is input into the pre-constructed fault type hierarchical diagnosis network to perform layer-by-layer diagnosis processing from fault existence determination to coarse-grained diagnosis of fault categories and then to fine-grained diagnosis of fault subcategories.

[0060] Step S141: Obtain a pre-constructed fault type hierarchical diagnosis network, which includes a cascaded fault existence binary classification diagnosis layer, a coarse-grained fault category diagnosis layer, and a fine-grained fault subcategory diagnosis layer.

[0061] The pre-constructed fault type hierarchical diagnostic network is a three-layer cascaded structure. The fault existence binary classification diagnostic layer determines whether a fault exists at each spatial sampling point location marker and outputs a binary classification probability. The coarse-grained fault category diagnostic layer categorizes faults into three main categories: partial discharge, overheating, and mechanical damage. The fine-grained fault subcategory diagnostic layer further subdivides each fault category into specific fault subcategories, such as dividing partial discharge into internal air gap discharge, surface discharge, and corona discharge. The three layers are cascaded sequentially; the output of the previous layer determines whether the next layer performs further diagnostics on the spatial sampling point location marker.

[0062] Step S142: Input the deep temporal dependency representation vector of the cable state into the fault existence binary classification diagnostic layer, and perform linear mapping processing on the feature vector corresponding to the location identifier of each spatial sampling point through the first fully connected weight matrix in the fault existence binary classification diagnostic layer to generate a first linear mapping feature vector; perform nonlinear activation processing on the first linear mapping feature vector, and input the nonlinear activation processing result into the fault existence classifier in the fault existence binary classification diagnostic layer to generate a fault existence binary classification probability value corresponding to the location identifier of each spatial sampling point.

[0063] The feature vector corresponding to the spatial sampling point location identifier in the deep temporal dependency representation vector Vdeep of the cable state is input into the fault presence binary classification diagnostic layer. The first fully connected weight matrix of the fault presence binary classification diagnostic layer is denoted as Wc1, which performs a linear mapping on the feature vector, resulting in the first linearly mapped feature vector Z1 = Vdeep × Wc1. A non-linear activation function ReLU is applied to Z1, yielding Z1a = ReLU(Z1). Z1a is then input into the fault presence classifier, which consists of a fully connected layer with an output dimension of 2 and a softmax transfer function, outputting a two-dimensional vector [p0, p1], where p0 represents the probability of no fault and p1 represents the probability of a fault, with p0 + p1 = 1. p1 is the fault presence binary classification probability value corresponding to the spatial sampling point location identifier.

[0064] Step S143: Compare the binary classification probability value of the fault existence with the preset fault existence judgment threshold, mark the spatial sampling point location identifiers whose binary classification probability value of the fault existence exceeds the preset fault existence judgment threshold as candidate fault spatial sampling point location identifiers, and extract the feature vectors in the deep temporal dependency representation vector of the cable state corresponding to all candidate fault spatial sampling point location identifiers to form a candidate fault feature vector set.

[0065] A preset fault existence determination threshold is set as Thr. The fault existence binary classification probability value p1 corresponding to each spatial sampling point location identifier generated in step S142 is compared with Thr. If p1 > Thr, the spatial sampling point location identifier is marked as a candidate fault spatial sampling point location identifier. After traversing all 80 spatial sampling point location identifiers, the feature vectors in Vdeep corresponding to all candidate fault spatial sampling point location identifiers are extracted to form a candidate fault feature vector set Vcand.

[0066] Step S144: Based on the candidate fault feature vector set, perform fault type classification diagnosis through the coarse-grained diagnosis layer of the fault category and the fine-grained diagnosis layer of the fault subcategory, and generate a fault boundary space sampling point location identifier sequence.

[0067] Using the candidate fault feature vector set Vcand generated in step S143, classification and diagnosis are performed sequentially through the coarse-grained diagnosis layer of the fault category and the fine-grained diagnosis layer of the fault subcategory, and spatial boundary clustering is performed on the diagnosis results.

[0068] Step S1441: Input the set of candidate fault feature vectors into the coarse-grained fault category diagnosis layer. Perform linear mapping processing on each candidate fault feature vector through the second fully connected weight matrix in the coarse-grained fault category diagnosis layer to generate a second linear mapping feature vector. Perform nonlinear activation processing on the second linear mapping feature vector and input the nonlinear activation processing result into the fault category classifier in the coarse-grained fault category diagnosis layer to generate a fault category assignment probability distribution vector corresponding to the spatial sampling point location identifier of each candidate fault. The fault category assignment probability distribution vector includes the probability values ​​of partial discharge faults, overheating faults, and mechanical damage faults.

[0069] The set of candidate fault feature vectors, Vcand, is input into the coarse-grained fault category diagnostic layer. The second fully connected weight matrix of this layer, denoted as Wc2, performs a linear mapping on each candidate fault feature vector, resulting in the second linearly mapped feature vector Z2 = Vcand × Wc2. A non-linear activation function, ReLU, is applied to Z2, yielding Z2a = ReLU(Z2). Z2a is then input into the fault category classifier, which consists of a fully connected layer with an output dimension of 3 and a softmax layer. The classifier outputs a three-dimensional vector [pc1, pc2, pc3], where pc1 represents the probability value of partial discharge faults, pc2 represents the probability value of overheating faults, and pc3 represents the probability value of mechanical damage faults; the sum of these three probabilities equals 1. This three-dimensional vector represents the probability distribution vector of the fault category corresponding to the spatial sampling point location identifier of the candidate fault.

[0070] Step S1442: Input the set of candidate fault feature vectors into the fine-grained diagnosis layer for fault subclasses, perform linear mapping processing on each candidate fault feature vector through the third fully connected weight matrix in the fine-grained diagnosis layer for fault subclasses to generate a third linear mapping feature vector; perform nonlinear activation processing on the third linear mapping feature vector, and input the nonlinear activation processing result into the fault subclass classifier in the fine-grained diagnosis layer for fault subclasses to generate a fault subclass classification probability distribution vector corresponding to the location identifier of each candidate fault spatial sampling point.

[0071] The set of candidate fault feature vectors, Vcand, is simultaneously input into the fine-grained fault subclass diagnostic layer. The third fully connected weight matrix of the fine-grained fault subclass diagnostic layer is denoted as Wc3, which performs a linear mapping on each candidate fault feature vector, resulting in the third linearly mapped feature vector Z3 = Vcand × Wc3. A non-linear activation function, ReLU, is applied to Z3, yielding Z3a = ReLU(Z3). Z3a is then input into the fault subclass classifier, which consists of a fully connected layer with an output dimension of M and a softmax layer, where M is the total number of fault subclasses. The classifier outputs an M-dimensional vector [ps1, ps2, ..., psM], which represents the probability distribution vector of the fault subclass classification corresponding to the spatial sampling point location identifier of the candidate fault.

[0072] Step S1443: Perform probability distribution concatenation processing on the fault category probability distribution vector and the fault sub-category probability distribution vector to generate a fault type probability distribution vector corresponding to the location identifier of each candidate fault spatial sampling point. The fault type probability distribution vector contains the probability score of each fault sub-category.

[0073] For each candidate fault spatial sampling point location identifier, the fault category attribution probability distribution vector [pc1, pc2, pc3] generated in step S1441 and the fault sub-category attribution probability distribution vector [ps1, ps2, ..., psM] generated in step S1442 are concatenated along the vector dimension to generate a fault type attribution probability distribution vector Pfinal of length 3+M. The first three elements of Pfinal are the fault category probability values, and the following M elements are the probability scores of each fault sub-category.

[0074] Step S1444: Based on the spatial continuity of all candidate fault spatial sampling point location identifiers, perform spatial boundary clustering processing on adjacent candidate fault spatial sampling point location identifiers with the same fault type affiliation to generate a fault boundary spatial sampling point location identifier sequence. The fault boundary spatial sampling point location identifier sequence includes the starting spatial sampling point location identifier and the ending spatial sampling point location identifier of each fault segment.

[0075] Extract the spatial continuity information of all candidate fault spatial sampling point location identifiers. For each candidate fault spatial sampling point location identifier, take the fault sub-category corresponding to the maximum probability value in the fault type attribution probability distribution vector Pfinal generated in step S1443 as its fault type attribution. Merge candidate fault spatial sampling point location identifiers that are spatially adjacent and have the same fault type attribution into one fault segment. For each fault segment, record its starting and ending spatial sampling point location identifiers to form a fault boundary spatial sampling point location identifier sequence. Each record in this fault boundary spatial sampling point location identifier sequence represents a boundary information pair for a fault segment.

[0076] Step S150: Generate a cable fault diagnosis report data set based on the fault type attribution probability distribution vector and the fault boundary space sampling point location identifier sequence, and send the cable fault diagnosis report data set to the cable operation and maintenance management terminal.

[0077] Using the fault type attribution probability distribution vector Pfinal generated in step S1443 and the fault boundary space sampling point location identifier sequence generated in step S1444, a structured cable fault diagnosis report data set is generated.

[0078] Step S151: Extract the fault sub-category corresponding to the maximum probability value of each candidate fault spatial sampling point location identifier from the fault type attribution probability distribution vector as the dominant fault type label of the candidate fault spatial sampling point location identifier.

[0079] For each candidate fault spatial sampling point location identifier, find the fault sub-category name corresponding to the dimension with the largest probability score in its fault type attribution probability distribution vector Pfinal, and use the fault sub-category name as the dominant fault type label for the candidate fault spatial sampling point location identifier.

[0080] Step S152: Extract the starting spatial sampling point location identifier and the ending spatial sampling point location identifier of each fault segment from the fault boundary spatial sampling point location identifier sequence, and calculate the difference between the starting spatial sampling point location identifier and the ending spatial sampling point location identifier as the number of sampling points for the spatial span of the fault segment.

[0081] Extract the starting and ending spatial sampling point location identifiers for each fault segment from the fault boundary spatial sampling point location identifier sequence. Calculate the absolute value of the difference between these two numbers; this difference plus 1 represents the number of spatial sampling points contained within that fault segment, denoted as Ns.

[0082] Step S153: Calculate the actual physical length of each fault segment based on the number of sampling points for the spatial span of the fault segment and the physical distance parameters between adjacent spatial sampling points.

[0083] The physical distance parameter between adjacent spatial sampling points is determined by the spatial resolution of the distributed fiber optic sensing device, denoted as Ds. The actual physical length of each fault segment is Lf = Ns × Ds.

[0084] Step S154: Summarize the dominant fault type markers of all candidate fault spatial sampling point location identifiers in each fault segment, and perform statistical determination on the dominant fault type markers that occur most frequently in the fault segment to obtain the segment-wide unified fault type marker for that fault segment.

[0085] For each fault segment, extract the location identifiers of all candidate fault spatial sampling points within the fault segment, and assign a dominant fault type label to each. Count the frequency of each dominant fault type label within the fault segment, and determine the dominant fault type label with the highest frequency as the unified fault type label for that fault segment.

[0086] Step S155: Extract the probability score corresponding to the unified fault type label of the segment from the probability distribution vector of the fault type attribution corresponding to the location identifier of all candidate fault spatial sampling points in each fault segment, and calculate the average value of the probability scores as the fault confidence score of the fault segment.

[0087] For each faulty section, extract the fault type attribution probability distribution vector Pfinal corresponding to the location identifiers of all candidate fault spatial sampling points within the faulty section. Extract the probability score corresponding to the unified fault type label of the section from each Pfinal. Calculate the arithmetic mean of all extracted probability scores to obtain the fault confidence score Cf for the faulty section.

[0088] Step S156: Based on the scattered radio displacement corresponding to the location identifiers of all candidate fault spatial sampling points in each fault section, calculate the gradient of the scattered radio displacement along the spatial sampling points in the fault section, and mark the location identifiers of spatial sampling points whose scattered radio displacement gradient exceeds the preset gradient threshold as fault core point location identifiers.

[0089] For each fault segment, extract the scattered radio displacement sequence corresponding to the location identifiers of all candidate fault spatial sampling points within that segment. Calculate the difference in scattered radio displacement between adjacent spatial sampling point location identifiers, and divide this difference by the physical distance parameter Ds between adjacent spatial sampling points to obtain the gradient of the scattered radio displacement along the spatial sampling points. Mark the location identifiers of spatial sampling points whose scattered radio displacement gradient exceeds a preset gradient threshold as fault core point location identifiers.

[0090] Step S157: Extract the half-width at half-maximum (WHM) parameter of the scattering spectrum corresponding to the location identifier of all candidate fault spatial sampling points in each fault section, calculate the fluctuation amplitude of the WHM parameter of the scattering spectrum along the spatial sampling points in the fault section, compare the fluctuation amplitude of the WHM parameter of the scattering spectrum with the standard fluctuation amplitude database of fault type, and obtain the auxiliary assessment parameter of fault severity.

[0091] For each fault segment, the half-width at half-maximum (WHM) of the scattering spectrum corresponding to the location identifiers of all candidate fault spatial sampling points within that segment is extracted. The standard deviation of the WHM of the scattering spectrum within the fault segment is calculated as the fluctuation amplitude along the spatial sampling points. The calculated fluctuation amplitude is compared with the standard fluctuation amplitude range corresponding to the fault type in the pre-built fault type standard fluctuation amplitude database, and the auxiliary assessment parameter for fault severity is determined based on the range of the fluctuation amplitude.

[0092] Step S158: The starting spatial sampling point location identifier, ending spatial sampling point location identifier, actual physical length, unified fault type label of the section, fault confidence score, fault core point location identifier and fault severity auxiliary assessment parameters of each fault section are processed in a structured combination to generate a single fault section diagnostic record data unit.

[0093] For each faulty section, the starting spatial sampling point location identifier, ending spatial sampling point location identifier, actual physical length Lf, section unified fault type marker, fault confidence score Cf, fault core point location identifier, and auxiliary assessment parameters for fault severity are structurally combined by field to generate a single faulty section diagnostic record data unit. This single faulty section diagnostic record data unit is stored in key-value pair format, where the key is the field name and the value is the corresponding parameter.

[0094] Step S159: Summarize the single fault section diagnostic record data units corresponding to all fault sections to generate a fault section diagnostic record data set; perform time association binding processing on the acquisition timestamp of the fault section diagnostic record data set and the original optical time domain scattering signal data set to generate a cable fault diagnosis report data set carrying timestamp markers. The cable fault diagnosis report data set includes the fault section diagnostic record data set and the corresponding acquisition timestamp markers.

[0095] All single-fault section diagnostic record data units generated in step S158 are summarized into a list to form a fault section diagnostic record data set. The fault section diagnostic record data set is then time-linked and bound to the acquisition timestamp of the original optical time-domain scattering signal data set; that is, a timestamp field is appended to the data set, the value of which is the acquisition start time of the original optical time-domain scattering signal data set. After binding the timestamp, a cable fault diagnosis report data set is generated. The cable fault diagnosis report data set is then sent to the cable operation and maintenance management terminal via a dedicated power communication network.

[0096] Step S210: Obtain the set of historical raw optical time-domain scattering signal data collected by the distributed sensing device deployed at the cable joint. The set of historical raw optical time-domain scattering signal data includes the location identifiers of historical spatial sampling points distributed along the cable axis and the historical scattering radio frequency shift time-domain response curve corresponding to each historical spatial sampling point.

[0097] Retrieve the historical raw optical time-domain scattering signal data set collected by the distributed optical fiber sensing device over the past few months from the historical database of the cable condition analysis server. The format of the historical raw optical time-domain scattering signal data set is completely consistent with the raw optical time-domain scattering signal data set in step S110, including the location identifier of the historical spatial sampling point and the historical scattering radio frequency shift time-domain response curve corresponding to each historical spatial sampling point.

[0098] Step S220: Perform scattering spectrum feature demodulation processing on the historical original optical time-domain scattering signal data set to obtain the historical scattering frequency shift feature sequence corresponding to the historical spatial sampling point location identifiers distributed along the cable axis and the historical scattering spectrum broadening feature sequence corresponding to each historical spatial sampling point location identifier.

[0099] Using the same scattering spectrum feature demodulation processing procedure as step S120, the set of historical original optical time-domain scattering signal data is processed to obtain historical scattering frequency shift feature sequence and historical scattering spectrum broadening feature sequence.

[0100] Step S230: Obtain the artificial fault type labeling data set corresponding to the historical original optical temporal scattering signal data set. The artificial fault type labeling data set includes the real fault type label corresponding to the location identifier of each historical spatial sampling point and the sequence of location identifiers of the real fault boundary space sampling points.

[0101] Fault type labeling data confirmed by manual offline inspection is retrieved from the cable operation and maintenance record database. Each historical spatial sampling point location identifier in the manual fault type labeling dataset corresponds to a real fault type label, with the label format being a combination of fault category and fault subcategory encoding. It also includes a sequence of real fault boundary spatial sampling point location identifiers labeled by maintenance personnel based on fault location results.

[0102] Step S240: Input the historical scattering frequency shift feature sequence and the historical scattering spectrum broadening feature sequence into the fault feature spatiotemporal correlation coding network to be trained. Perform forward propagation calculation processing through the position coding injection layer, the first multi-head self-attention transformation layer, the feedforward fully connected transformation layer and the second multi-head self-attention transformation layer in the fault feature spatiotemporal correlation coding network to be trained, and generate the deep temporal dependence representation vector of the cable state corresponding to the position identifier of each historical spatial sampling point.

[0103] The historical scattering frequency shift feature sequence and historical scattering spectrum broadening feature sequence generated in step S220 are used as training inputs to the fault feature spatiotemporal correlation coding network to be trained. The architecture of the fault feature spatiotemporal correlation coding network to be trained is completely consistent with the pre-constructed fault feature spatiotemporal correlation coding network described in step S131, and the network weight parameters are randomly initialized. The forward propagation calculation process is performed in complete consistency with steps S132 to S1383 to generate the deep temporal dependency representation vector of the cable state for training.

[0104] Step S250: Input the deep temporal dependency representation vector of the cable state for training into the fault type hierarchical diagnosis network to be trained. Perform forward propagation calculation processing through the fault existence binary classification diagnosis layer, the fault category coarse-grained diagnosis layer, and the fault sub-category fine-grained diagnosis layer in the fault type hierarchical diagnosis network to be trained, and generate the training fault type attribution probability distribution vector and the training fault boundary space sampling point location identifier sequence corresponding to the location identifier of each historical spatial sampling point.

[0105] The deep temporal dependency representation vector of cable state for training generated in step S240 is input into the fault type hierarchical diagnostic network to be trained. The architecture of the fault type hierarchical diagnostic network to be trained is completely consistent with the pre-built fault type hierarchical diagnostic network described in step S141. The forward propagation computation process, which is completely consistent with steps S142 to S1444, is performed to generate the fault type attribution probability distribution vector for training and the fault boundary space sampling point location identifier sequence for training.

[0106] Step S260: Calculate the fault type classification cross-entropy loss value based on the training fault type attribution probability distribution vector and the real fault type label, and calculate the fault boundary regression loss value based on the training fault boundary space sampling point location identifier sequence and the real fault boundary space sampling point location identifier sequence.

[0107] The training fault type attribution probability distribution vector generated in step S250 is compared with the real fault type labels obtained in step S230, and the fault type classification cross-entropy loss value Lcls = -Σ(real label log predicted probability) is calculated. The training fault boundary space sampling point location identifier sequence generated in step S250 is compared with the real fault boundary space sampling point location identifier sequence, and the fault boundary regression loss value Lreg = Σ|predicted boundary minus real boundary| is calculated using the L1 norm.

[0108] Step S270: Perform a weighted summation of the fault type classification cross-entropy loss value and the fault boundary regression loss value to generate a joint training loss value.

[0109] The joint training loss value Ltotal = λ1 × Lcls + λ2 × Lreg, where λ1 and λ2 are preset loss weight coefficients.

[0110] Step S280: Calculate the gradient values ​​of all network weight parameters in the fault feature spatiotemporal correlation coding network and the fault type hierarchical diagnosis network to be trained based on the joint training loss value, and update all network weight parameters along the gradient descent direction using the gradient descent optimization algorithm.

[0111] The backpropagation algorithm is used to calculate the partial derivatives of all network weight parameters in the fault feature spatiotemporal correlation coding network and the fault type hierarchical diagnostic network to be trained, obtaining the gradient value. The Adam gradient descent optimization algorithm is then used to update all network weight parameters along the gradient descent direction. The learning rate of the Adam optimizer is set to α, the first-order moment decay coefficient is β1, and the second-order moment decay coefficient is β2.

[0112] Step S290: Repeat the step of inputting the historical scattering frequency shift feature sequence and the historical scattering spectrum broadening feature sequence into the fault feature spatiotemporal correlation coding network to be trained, and updating all network weight parameters along the gradient descent direction using the gradient descent optimization algorithm, until the joint training loss value converges to below the preset loss threshold; save the network weight parameters of the fault feature spatiotemporal correlation coding network after the joint training loss value converges as a pre-constructed fault feature spatiotemporal correlation coding network, and save the network weight parameters of the fault type hierarchical diagnosis network after the joint training loss value converges as a pre-constructed fault type hierarchical diagnosis network.

[0113] Repeat steps S240 to S280, calculating the joint training loss value on the validation set after each iteration. Training is stopped when the joint training loss value on the validation set no longer decreases after multiple iterations and falls below a preset loss threshold. The network weight parameters of the fault feature spatiotemporal correlation coding network corresponding to the current training round are saved as a pre-constructed fault feature spatiotemporal correlation coding network, and the network weight parameters of the fault type hierarchical diagnosis network corresponding to the current training round are saved as a pre-constructed fault type hierarchical diagnosis network.

[0114] For example, the method may further include: step S310: obtaining the feature vector corresponding to the spatial sampling point position identifier in the deep temporal dependency representation vector of the cable state, and generating the cross-spatial sampling point attention weight matrix in the second multi-head self-attention transformation layer of the fault feature spatiotemporal correlation coding network.

[0115] During the generation of the second multi-head self-attention output feature sequence in step S1381, each attention head in the second multi-head self-attention transformation layer generates a cross-spatial sampling point attention weight matrix. The row and column indices of this cross-spatial sampling point attention weight matrix are both 80 spatial sampling point location identifiers. The element value in the i-th row and j-th column of the matrix represents the attention weight value of the i-th spatial sampling point location identifier on the j-th spatial sampling point location identifier. The cross-spatial sampling point attention weight matrices of all attention heads are obtained, and the average value is calculated along the attention head dimension to obtain the average cross-spatial sampling point attention weight matrix.

[0116] Step S320: Extract the attention weight value between each spatial sampling point location identifier and all other spatial sampling point location identifiers from the cross-spatial sampling point attention weight matrix, sort the associated spatial sampling points of each spatial sampling point location identifier in descending order of attention weight value, and generate the attention association sorting sequence corresponding to each spatial sampling point location identifier.

[0117] For each spatial sampling point location identifier, extract the row vector corresponding to that spatial sampling point location identifier from the average cross-spatial sampling point attention weight matrix obtained in step S310. This row vector contains the attention weight values ​​between that spatial sampling point location identifier and all other spatial sampling point location identifiers. Sort the attention weight values ​​in this row vector in descending order. Each element in the sorting result contains the associated spatial sampling point location identifier number and the corresponding attention weight value. This sorting result is the attention association sorting sequence corresponding to that spatial sampling point location identifier.

[0118] Step S330: Extract a preset number of associated spatial sampling point location identifiers from the attention association sorting sequence, with the attention weight values ​​ranking first. Combine the extracted associated spatial sampling point location identifiers and their corresponding attention weight values ​​to form a strong attention association subgraph corresponding to each spatial sampling point location identifier.

[0119] Set a preset number of nodes, R. For each spatial sampling point location identifier, extract the top R associated spatial sampling point location identifiers from its attention association ranking sequence. Use the extracted R associated spatial sampling point location identifiers as nodes, and the attention weight values ​​between this spatial sampling point location identifier and each associated spatial sampling point location identifier as edge weights, to form a strong attention association subgraph centered on this spatial sampling point location identifier.

[0120] Step S340: Construct a cable full-segment attention association topology based on the strong attention association subgraphs corresponding to all spatial sampling point location identifiers. The cable full-segment attention association topology uses spatial sampling point location identifiers as nodes, strong attention association relationships as edges, and attention weight values ​​as edge weights.

[0121] The strong attention association subgraphs of all 80 spatial sampling point location identifiers generated in step S330 are merged to form a full-segment cable attention association topology. The full-segment cable attention association topology is stored in the form of an adjacency list, where each node stores its spatial sampling point location identifier number and each edge stores its edge weight, i.e., the attention weight value.

[0122] Step S350: Perform community detection clustering on the attention association topology of the entire cable segment, identify the cluster of spatial sampling points with the largest sum of attention weight values ​​as cable fault sensitive areas, and identify the locations of all spatial sampling points contained in the cable fault sensitive areas to form a fault sensitive area spatial sampling point set.

[0123] The Leuven community detection algorithm is used to perform community detection and clustering on the attention association topology of the entire cable segment. The Leuven algorithm divides nodes into communities by iteratively optimizing the modularity. After clustering, the sum of attention weights of all edges within each community is calculated, and the community with the largest sum of attention weights is identified as a cable fault-sensitive area. The location identifiers of all spatial sampling points contained within the cable fault-sensitive area are extracted to form a set of spatial sampling points for the fault-sensitive area.

[0124] Step S360: Perform frequency shift time-series fluctuation pattern mining processing on the scattered radio frequency shift feature sequence corresponding to the location identifier of each spatial sampling point in the set of spatial sampling points in the fault-sensitive area, and extract the local extreme point occurrence pattern of scattered radio frequency shift in the continuous time domain analysis window segment as the frequency shift local extreme value pattern feature.

[0125] For each spatial sampling point in the fault-sensitive area's spatial sampling point set, its divergent frequency shift feature sequence is extracted. Local extrema are detected within these divergent frequency shift feature sequences. A local extremum is determined by whether its divergent frequency shift is greater than that of its immediate neighbors (i.e., a local maximum) or less than that of its immediate neighbors (i.e., a local minimum). The mean and variance of the time intervals between adjacent local extrema, as well as the mean and variance of the amplitudes of the local extrema, are statistically analyzed. These statistics are combined to form the frequency shift local extremum pattern feature.

[0126] Step S370: Perform broadening temporal mutation mode mining processing on the scattering spectrum broadening feature sequence corresponding to the location identifier of each spatial sampling point in the set of spatial sampling points in the fault sensitive area, and extract the number of mutations and mutation amplitudes of the scattering spectrum half-width parameter exceeding the preset mutation threshold between adjacent time domain analysis window segments as broadening temporal mutation mode features.

[0127] For each spatial sampling point in the set of spatial sampling points in the fault-sensitive area, its scattering spectrum broadening feature sequence is extracted. The absolute value of the difference between the half-width at half-maximum (WHM) parameters of the scattering spectrum between adjacent time-domain analysis window segments is calculated, and adjacent window pairs whose absolute difference exceeds a preset mutation threshold are marked as mutation events. The total number of mutation events in the entire scattering spectrum broadening feature sequence and the mean of the absolute difference of all mutation events, i.e., the mean mutation amplitude, are counted. The number of mutations and the mean mutation amplitude are combined to form the broadening time-series mutation mode feature.

[0128] Step S380: Perform mode association combination processing on the frequency shift local extremum mode features and the broadened temporal abrupt change mode features to generate a joint temporal mode feature vector for the fault-sensitive region.

[0129] The frequency shift local extremum mode feature vector generated in step S360 and the broadened temporal mutation mode feature vector generated in step S370 are concatenated in the feature dimension to generate the joint temporal mode feature vector of the fault-sensitive region.

[0130] Step S390: Perform similarity comparison processing between the joint temporal pattern feature vector of the fault-sensitive region and the standard temporal pattern feature vector of each fault type in the pre-constructed fault type temporal pattern fingerprint database to generate fault-sensitive region temporal pattern matching fault type results.

[0131] A pre-built temporal pattern fingerprint database of fault types stores the standard temporal pattern feature vectors exhibited by each fault type historically. The cosine similarity between the joint temporal pattern feature vector of the fault-sensitive region and the standard temporal pattern feature vector of each fault type in the fingerprint database is calculated. The fault type with the highest cosine similarity is taken as the fault type matching result for the temporal pattern of the fault-sensitive region.

[0132] Step S3100: Perform cross-validation processing on the fault type result of the time-series pattern matching of the fault sensitive area and the dominant fault type marker of the spatial sampling point location identifier in the fault sensitive area in the fault type attribution probability distribution vector to generate a double confirmation diagnosis result of the fault type.

[0133] The fault type result of the time-series pattern matching of the fault-sensitive area generated in step S390 is compared with the dominant fault type label of the spatial sampling point location identifier in the corresponding fault-sensitive area generated in step S151. If they match, the fault type is marked as a high-confidence fault type; if they do not match, it is marked as a fault type requiring manual review. A double-confirmation diagnosis result for the fault type is generated.

[0134] Step S410: Obtain the first layer attention weight matrix generated by the first multi-head self-attention transformation layer in the fault feature spatiotemporal correlation coding network when calculating the scattering spectrum joint feature sequence of injected position sensing information. The first layer attention weight matrix contains the temporal attention weight values ​​between each temporal analysis window segment and all other temporal analysis window segments.

[0135] During the generation of the first multi-head self-attention output feature sequence in step S134, each attention head in the first multi-head self-attention transformation layer generates a temporal attention weight matrix. The row and column indices of this temporal attention weight matrix are both numbers of K temporal analysis window segments. The element value in the i-th row and j-th column of the matrix represents the temporal attention weight value of the i-th temporal analysis window segment on the j-th temporal analysis window segment. The temporal attention weight matrices of all attention heads are obtained, and the average value is calculated along the attention head dimension to obtain the first-layer attention weight matrix.

[0136] Step S420: Extract the distribution curve of the temporal attention weight value corresponding to the spatial sampling point position identifier along the temporal analysis window segment dimension from the first layer attention weight matrix, perform peak detection processing on the distribution curve, and identify the temporal analysis window segment corresponding to the peak value of the attention weight value in the distribution curve as the key temporal response window.

[0137] For each spatial sampling point location identifier, the self-attention weight distribution curve corresponding to that spatial sampling point location identifier is extracted from the first-layer attention weight matrix; that is, the curve showing the change of elements on the diagonal of the matrix along the temporal analysis window segment dimension. Peak detection processing is performed on this distribution curve, using the criteria of the first derivative crossing zero and the second derivative being negative to identify the peak position. The temporal analysis window segment corresponding to the peak position is marked as the key temporal response window.

[0138] Step S430: Arrange all key timing response windows corresponding to each spatial sampling point location identifier in order of acquisition time to generate a sequence of key timing response windows corresponding to each spatial sampling point location identifier.

[0139] For each spatial sampling point location identifier, all its key time-series response windows are arranged in ascending order according to the starting acquisition time of the time-domain analysis window segment, forming a key time-series response window sequence.

[0140] Step S440: Calculate the time interval between adjacent key timing response windows in the key timing response window sequence to obtain the interval duration sequence between key timing response windows. Perform periodic detection processing on the interval duration sequence between key timing response windows and extract the repetitive periodic pattern of the interval duration as the fault feature timing periodic characterization parameter.

[0141] The time interval between any two adjacent critical timing response windows in the critical timing response window sequence is calculated. The time interval is equal to the start acquisition time of the subsequent window minus the start acquisition time of the preceding window. All the obtained time intervals constitute the interval duration sequence between critical timing response windows. A Fast Fourier Transform is performed on the interval duration sequence between critical timing response windows to identify the maximum peak frequency (excluding the DC component) in the spectrum. The reciprocal of this peak frequency is the repetition periodic pattern of the interval duration, which serves as a parameter for characterizing the timing periodicity of fault features.

[0142] Step S450: Perform statistical feature extraction on the scattered frequency shift of the scattered frequency shift feature sequence within each key time series response window, and calculate the mean and standard deviation of the scattered frequency shift within the key time series response window as the statistical features of frequency shift in the key window.

[0143] For each spatial sampling point location identifier, extract the scattered frequency shift feature sequence segment corresponding to the key temporal response window of that spatial sampling point location identifier. Calculate the arithmetic mean and standard deviation of the scattered frequency shift within this segment, which are used as the key window frequency shift statistical features of that key temporal response window.

[0144] Step S460: Perform statistical feature extraction on the scattering spectrum half-width parameter of the scattering spectrum broadening feature sequence within each key time-series response window, and calculate the mean and standard deviation of the scattering spectrum half-width parameter within the key time-series response window as the statistical features of key window broadening.

[0145] For each spatial sampling point location identifier, extract the scattering spectrum broadening feature sequence segment corresponding to the key temporal response window of that spatial sampling point location identifier. Calculate the arithmetic mean and standard deviation of the half-width at half-maximum (FWHM) parameter of the scattering spectrum within this segment, which are used as the key window broadening statistical features of that key temporal response window.

[0146] Step S470: Perform feature fusion processing on the fault feature time-series periodic characterization parameters, the key window frequency shift statistical features, and the key window widening statistical features to generate a self-attention interpretable fault feature vector corresponding to the location identifier of each spatial sampling point.

[0147] The fault feature time-series periodic characterization parameters generated in step S440, the key window frequency shift statistical features generated in step S450, and the key window widening statistical features generated in step S460 are concatenated along the feature dimension to generate a self-attention interpretable fault feature vector.

[0148] Step S480: Input the self-attention interpretable fault feature vector into the pre-constructed fault type interpretable decision tree model, and generate an interpretable fault type classification path corresponding to the location identifier of each spatial sampling point by comparing the feature thresholds of each decision node in the fault type interpretable decision tree model.

[0149] The pre-built fault type interpretability decision tree model adopts a classification decision tree structure based on the CART algorithm. The self-attention interpretable fault feature vector is input into the root node of this decision tree model. At each decision node, the value of a specified feature dimension in the self-attention interpretable fault feature vector is compared with the feature threshold of the decision node. Based on the comparison result, the model proceeds to the left or right child node, until a leaf node is reached. The complete path from the root node to the leaf node constitutes the interpretable fault type classification path.

[0150] Step S490: Extract the feature type and feature threshold comparison direction used by each decision node from the explainable fault type classification path, and generate a fault diagnosis logic explanation text sequence corresponding to the fault type classification result.

[0151] Iterate through each decision node along the explainable fault type classification path, extracting the feature dimension name, feature threshold, and comparison direction (greater than or less than equal to) used for splitting at that decision node. Arrange this information in the decision order to generate a fault diagnosis logical explanation text sequence. This sequence describes the logical reasoning process of fault type classification in the form of a natural language template.

[0152] Step S4100: Associate the fault diagnosis logic explanation text sequence with the fault type attribution probability distribution vector to generate a fault type diagnosis result data set with interpretable diagnosis logic.

[0153] The fault diagnosis logic explanation text sequence generated in step S490 is associated and bound with the fault type attribution probability distribution vector generated in step S1443, and stored accordingly according to the spatial sampling point location identifier number to generate a fault type diagnosis result data set with interpretable diagnosis logic.

[0154] Step S510: Obtain the first linear mapping feature vector of the deep temporal dependency representation vector of the cable state before it is processed by the fault existence binary classification diagnosis layer in the fault type hierarchical diagnosis network. The first linear mapping feature vector is the intermediate feature representation of the feature vector corresponding to the spatial sampling point position identifier after linear mapping by the first fully connected weight matrix.

[0155] In step S142, the deep temporal dependency representation vector Vdeep of the cable state is linearly mapped by the first fully connected weight matrix Wc1 to obtain the first linear mapping feature vector Z1. Z1 is truncated and saved before the activation function processing. The first linear mapping feature vector Z1 is then obtained.

[0156] Step S520: Obtain the second linear mapping feature vector of the deep temporal dependency representation vector of the cable state before it is processed by the coarse-grained diagnosis layer of the fault category in the fault type hierarchical diagnosis network. The second linear mapping feature vector is the intermediate feature representation of the candidate fault feature vector after linear mapping by the second fully connected weight matrix.

[0157] In step S1441, the candidate fault feature vector Vcand is linearly mapped by the second fully connected weight matrix Wc2 to obtain the second linearly mapped feature vector Z2. Z2 is truncated and saved before the activation function processing. This second linearly mapped feature vector Z2 is then obtained.

[0158] Step S530: Perform principal component analysis on the first linear mapping feature vector to reduce its dimensionality, extract the first preset number of principal component components of the first linear mapping feature vector, and construct the first dimensionality-reduced visualization feature vector from the first preset number of principal component components.

[0159] Principal component analysis (PCA) is used to reduce the dimensionality of the first linear mapping eigenvector Z1 obtained in step S510. The covariance matrix of Z1 is calculated, and eigenvalue decomposition is performed on the covariance matrix. The eigenvalues ​​are sorted from largest to smallest, and the eigenvectors corresponding to the two largest eigenvalues ​​are selected. Z1 is projected onto the two-dimensional subspace spanned by these two eigenvectors to obtain the first two-dimensional reduced-dimensional visualization eigenvector.

[0160] Step S540: Perform principal component analysis on the second linear mapping feature vector to reduce its dimensionality, extract the first preset number of principal component components of the second linear mapping feature vector, and construct the second dimensionality-reduced visualization feature vector from the first preset number of principal component components.

[0161] Perform the same principal component analysis dimensionality reduction process as in step S530 on the second linear mapping feature vector Z2 obtained in step S520 to obtain a two-dimensional second dimensionality-reduced visual feature vector.

[0162] Step S550: Map the first dimension-reduced visualization feature vector corresponding to the location identifier of all spatial sampling points to the two-dimensional visualization space. Perform color encoding on the data points in the two-dimensional visualization space according to the binary classification probability value of the fault existence corresponding to the location identifier of each spatial sampling point to generate visualization image data of fault existence feature distribution.

[0163] Map the first dimensionality-reduced visualization feature vectors corresponding to the location identifiers of all 80 spatial sampling points to coordinate points in a two-dimensional visualization space. Map the binary classification probability value p1 of the fault presence corresponding to each spatial sampling point location identifier to a color; the closer p1 is to 1, the closer the color is to red; the closer p1 is to 0, the closer the color is to blue. Plot and color all data points in the two-dimensional coordinate system to generate a visualization image of the fault presence feature distribution.

[0164] Step S560: Map the second dimension-reduced visualization feature vector corresponding to the location identifier of all candidate fault space sampling points to the two-dimensional visualization space. Based on the fault category with the highest probability value in the fault category probability distribution vector corresponding to the location identifier of each candidate fault space sampling point, color-encode the data points in the two-dimensional visualization space to generate fault category feature distribution visualization image data.

[0165] Map the second-dimensionally reduced visualization feature vector corresponding to the candidate fault spatial sampling point location identifiers selected in step S143 onto coordinate points in the two-dimensional visualization space. Map the fault category with the highest probability value in the fault category probability distribution vector of each candidate fault spatial sampling point location identifier to a different color; for example, partial discharge is mapped to yellow, overheating to orange, and mechanical damage to brown. Plot all candidate fault data points in the two-dimensional coordinate system and color them according to fault category to generate a visualization image data of fault category feature distribution.

[0166] Step S570: Perform boundary shape analysis on the spatial clusters formed by data points whose binary classification probability values ​​of fault existence in the visualization image data of fault existence feature distribution exceed the preset fault existence judgment threshold, and extract the contour envelope of the spatial clusters and the density distribution of data points within the clusters as the geometric morphological features of the fault clusters.

[0167] In the visualized image data of fault existence feature distribution, data points with p1 greater than Thr are selected and identified as faulty data points. A density-based spatial clustering algorithm is used to cluster these faulty data points, identifying several spatial clusters. For each spatial cluster, its convex hull contour envelope is calculated, which is the smallest convex polygon containing all data points within the cluster. Simultaneously, the density distribution of data points within the cluster is calculated, i.e., the number of data points per unit area. The vertex coordinates of the convex hull contour envelope and the density distribution values ​​are used as the geometric morphological features of the fault cluster.

[0168] Step S580: Perform regional overlap analysis on the data point distribution areas corresponding to different fault categories in the visualization image data of fault category feature distribution, and calculate the overlap area ratio between the data point distribution areas of different fault categories as a parameter for measuring the separability of fault category features.

[0169] In the visualized image data of fault category feature distribution, the distribution area of ​​data points corresponding to each fault category is extracted. The overlapping area between the distribution areas corresponding to any two fault categories is calculated; the overlapping area is equal to the area of ​​the intersection of the two areas. The overlapping area is divided by the area of ​​the union of the two areas to obtain the overlap area ratio. The average of the overlap area ratio ratios of all fault category pairs is used as the parameter for measuring the separability of fault category features.

[0170] Step S590: Based on the geometric morphological features of the fault clusters and the separability measurement parameters of the fault category features, perform a comprehensive evaluation of the feature representation quality of the fault feature spatiotemporal correlation coding network and the fault type hierarchical diagnosis network, and generate fault feature representation quality evaluation report data.

[0171] The quality of feature representation is evaluated by combining the geometric morphological characteristics of fault clusters and the separability metrics of fault category features. Evaluation indicators include: the compactness of fault clusters (small spacing between data points within a cluster), the separability of fault clusters (large spacing between different fault clusters), and the separability of fault category features (small overlap area). These indicators are combined to form the data for the fault feature representation quality evaluation report.

[0172] Step S5100: Package the fault existence feature distribution visualization image data, the fault category feature distribution visualization image data, and the fault feature representation quality assessment report data to generate a fault feature visualization analysis data package, and send the fault feature visualization analysis data package to the cable operation and maintenance management terminal.

[0173] The fault presence feature distribution visualization image data generated in step S550, the fault category feature distribution visualization image data generated in step S560, and the fault feature representation quality assessment report data generated in step S590 are packaged and processed to generate a fault feature visualization analysis data package. This fault feature visualization analysis data package is then sent to the cable operation and maintenance management terminal via a dedicated power communication network for operation and maintenance personnel to perform visualization analysis.

[0174] Step S610: Obtain the probability score of each fault subclass corresponding to the position identifier of each candidate fault spatial sampling point in the fault subclass attribution probability distribution vector output by the fine-grained diagnosis layer of the fault subclass in the fault type hierarchical diagnosis network, and extract the top three fault subclasses with the highest probability scores to form a candidate fault subclass triplet.

[0175] For each candidate fault spatial sampling point location identifier generated in step S1442, the probability distribution vector corresponding to the fault subclass is used. The M probability scores are sorted from largest to smallest, and the names of the top three fault subclasses and their corresponding probability scores are extracted to form a candidate fault subclass triplet.

[0176] Step S620: Obtain the starting spatial sampling point location identifier and the ending spatial sampling point location identifier of each fault segment in the fault boundary spatial sampling point location identifier sequence, and determine all candidate fault spatial sampling point location identifiers contained in each fault segment.

[0177] Extract the boundary information of each fault segment from the fault boundary space sampling point location identifier sequence generated in step S1444, and determine the list of all candidate fault space sampling point location identifiers contained in each fault segment.

[0178] Step S630: Summarize and statistically process the triplet of candidate fault subclass corresponding to the location identifier of all candidate fault spatial sampling points in each fault section, count the total number of times each fault subclass appears in the fault section, determine the fault subclass with the highest total number of occurrences as the main fault subclass of the fault section, and determine the fault subclass with the second highest total number of occurrences as the auxiliary fault subclass of the fault section.

[0179] For each fault segment, count the total number of occurrences of each fault subclass in the triplet of candidate fault subclasses identified by all candidate fault spatial sampling point locations within the fault segment. The fault subclass with the highest total occurrence is determined as the primary fault subclass of the fault segment, and the fault subclass with the second highest total occurrence is determined as the secondary fault subclass of the fault segment.

[0180] Step S640: Calculate the average of all probability scores of the main fault subclass within the fault section as the average confidence level of the main fault subclass, and calculate the average of all probability scores of the auxiliary fault subclass within the fault section as the average confidence level of the auxiliary fault subclass.

[0181] For the primary fault subclass, extract the probability scores of the primary fault subclass corresponding to the location identifiers of all candidate fault spatial sampling points containing the primary fault subclass within the fault segment, calculate the arithmetic mean, and obtain the average confidence score of the primary fault subclass. Similarly, calculate the average confidence score of the secondary fault subclass.

[0182] Step S650: Calculate the difference between the average confidence level of the main fault subclass and the average confidence level of the auxiliary fault subclass as the fault subclass confidence interval, compare the fault subclass confidence interval with a preset confidence interval threshold, and mark the fault segment with a fault subclass confidence interval less than the preset confidence interval threshold as a composite fault suspected segment.

[0183] The difference between the average confidence score of the primary fault subclass and the average confidence score of the secondary fault subclass is calculated and used as the fault subclass confidence interval. The fault subclass confidence interval is compared with a preset confidence interval threshold. If the fault subclass confidence interval is less than the preset confidence interval threshold, it indicates that the confidence difference between the primary and secondary fault subclasses is not significant, and the fault segment is marked as a suspected composite fault segment.

[0184] Step S660: Obtain the scattered radio frequency shift feature sequence and the scattering spectrum broadening feature sequence corresponding to the spatial sampling point location identifier of each candidate fault in the suspected composite fault segment. Input the scattered radio frequency shift feature sequence and the scattering spectrum broadening feature sequence into the pre-constructed independent diagnostic sub-network set of fault subclasses. The independent diagnostic sub-network set of fault subclasses contains an independent binary classification diagnostic sub-network corresponding to each fault subclass.

[0185] In the pre-constructed set of independent diagnostic sub-networks for each fault subclass, each fault subclass corresponds to an independent binary classification diagnostic sub-network. This binary classification diagnostic sub-network is a small convolutional neural network specifically trained to determine whether a fault belongs to that specific fault subclass. The scattering frequency shift feature sequence and scattering spectrum broadening feature sequence corresponding to the spatial sampling point location identifier of each candidate fault in the suspected composite fault segment are respectively input into the independent binary classification diagnostic sub-network of each fault subclass.

[0186] Step S670: Perform independent fault feature diagnosis processing on the scattered radio frequency shift feature sequence and the scattering spectrum broadening feature sequence through the independent binary classification diagnostic sub-network corresponding to each fault subclass, and generate an independent fault existence probability value corresponding to each fault subclass.

[0187] Each independent binary classification diagnostic subnetwork performs convolution, pooling, and fully connected processing on its input scattering radio shift feature sequence and scattering spectrum broadening feature sequence. Finally, it outputs a probability value through a sigmoid function, representing the probability that the location identifier of the candidate fault spatial sampling point belongs to that fault subclass.

[0188] Step S680: Perform joint probability analysis on the independent fault existence probability values ​​corresponding to the main fault subclass and the auxiliary fault subclass within the suspected composite fault section, and calculate the joint probability value of the simultaneous existence of the main fault subclass and the auxiliary fault subclass, as well as the conditional probability value of their individual existence.

[0189] The joint probability of both the primary and secondary fault subclasses existing simultaneously is equal to the product of the independent fault probability of the primary fault subclass and the independent fault probability of the secondary fault subclass. The conditional probability of the primary fault subclass existing alone is equal to the difference between the independent fault probability of the primary fault subclass and the independent fault probability of the secondary fault subclass. The conditional probability of the secondary fault subclass existing alone is equal to the difference between the independent fault probability of the secondary fault subclass and the independent fault probability of the primary fault subclass.

[0190] Step S690: Based on the joint probability value and the conditional probability value, perform composite fault type determination processing on the suspected composite fault segment, determine the suspected composite fault segment with the joint probability value exceeding the preset joint probability threshold as the real composite fault segment, and generate the real composite fault segment mark and the corresponding composite fault subclass combination.

[0191] The joint probability value is compared with a preset joint probability threshold. If the joint probability value exceeds the preset joint probability threshold, the suspected composite fault segment is determined to be a real composite fault segment. A real composite fault segment marker is generated, and the composite fault subclass is a combination of the main fault subclass name and the auxiliary fault subclass name.

[0192] Step S6100: Combine the actual composite fault section markers and corresponding composite fault subclasses with the fault type attribution probability distribution vector to generate a cable fault diagnosis report data set containing composite fault diagnosis information.

[0193] The actual composite fault section markers and composite fault subclasses generated in step S690 are merged and updated into the fault type attribution probability distribution vector generated in step S1443. Composite fault markers are added to the spatial sampling point location identifiers within the actual composite fault sections. Step S150 is re-executed to generate a cable fault diagnosis report dataset, which contains composite fault diagnosis information.

[0194] Step S710: Perform temporal phase space reconstruction processing on the frequency shift amount corresponding to the spatial sampling point position identifier in the scattered frequency shift feature sequence along the temporal analysis window segment dimension, expand the phase space trajectory of the scattered frequency shift feature sequence with a preset embedding dimension and a preset delay time, and generate a set of frequency shift phase space orbit points corresponding to the spatial sampling point position identifier.

[0195] Temporal phase space reconstruction is performed on the scattered radio frequency shift feature sequence identified by each spatial sampling point. A preset embedding dimension of E and a preset delay time of Tau are set. Points are taken from the scattered radio frequency shift feature sequence at intervals of Tau to construct an E-dimensional phase space vector. The m-th phase space vector is [f0(m), f0(m+Tau), f0(m+2*Tau), ..., f0(m+(E-1)*Tau)]. All phase space vectors constitute the set of frequency-shifted phase space orbital points.

[0196] Step S720: Perform recursive graph construction processing on the frequency shift phase space orbit point set, calculate the Euclidean distance between any two phase points in the phase space orbit, mark the phase point pairs whose Euclidean distance is less than a preset distance threshold as recursive points, and generate frequency shift recursion. Figure 2 Value matrix.

[0197] Calculate the Euclidean distance between any two phase space vectors in the frequency-shifted phase space orbital point set. Mark phase point pairs whose Euclidean distance is less than a preset distance threshold Eps as recursive points. Construct a square matrix whose dimension equals the total number of phase space vectors. The element in the i-th row and j-th column of the square matrix is ​​set to 1 if it is a recursive point, and 0 otherwise. This square matrix is ​​the frequency-shifted recursive vector. Figure 2 Value matrix.

[0198] Step S730: Perform temporal phase space reconstruction processing on the broadening parameters corresponding to the spatial sampling point position identifier in the broadening feature sequence of the scattering spectrum along the temporal analysis window segment dimension, expand the phase space trajectory of the broadening feature sequence of the scattering spectrum with a preset embedding dimension and a preset delay time, and generate the broadened phase space orbit point set corresponding to each spatial sampling point position identifier.

[0199] Using the same temporal phase space reconstruction processing method as in step S710, the phase space trajectory of the broadened scattering spectrum feature sequence is expanded to generate a broadened phase space orbit point set.

[0200] Step S740: Perform recursive graph construction processing on the expanded phase space orbit point set, calculate the Euclidean distance between any two phase points in the phase space orbit, mark the phase point pairs whose Euclidean distance is less than a preset distance threshold as recursive points, and generate the expanded recursive graph. Figure 2 Value matrix.

[0201] Using the same recursive graph construction processing method as in step S720, the set of orbital points in the broadened phase space is processed to generate a broadened recursive graph. Figure 2 Value matrix.

[0202] Step S750: Recursively apply the frequency shift. Figure 2 Value matrix and the broadened recursion Figure 2 The value matrix is ​​input into a pre-constructed recursive graph cross-correlation analysis network, and the frequency shift is recursively processed through the first convolutional feature extraction layer of the recursive graph cross-correlation analysis network. Figure 2 The value matrix is ​​subjected to recursive texture feature extraction processing to generate a frequency-shifted recursive texture feature map; the second convolutional feature extraction layer of the recursive graph cross-correlation analysis network is used to expand the recursive texture feature map. Figure 2 The value matrix is ​​subjected to recursive texture feature extraction processing to generate a widened recursive texture feature map.

[0203] The pre-constructed recursive graph cross-correlation analysis network includes a first convolutional feature extraction layer and a second convolutional feature extraction layer, both employing the same residual convolutional network structure. The first convolutional feature extraction layer receives frequency-shifted recursion. Figure 2The value matrix is ​​used as input, and recursive texture feature extraction is performed through three convolutional layers. Each convolutional layer has a 3x3 kernel size, with the number of kernels increasing sequentially. Batch normalization and ReLU activation are inserted between convolutional layers. The first convolutional feature extraction layer outputs a frequency-shifted recursive texture feature map. The second convolutional feature extraction layer receives the stretched recursive texture map. Figure 2 The value matrix is ​​used as input and processed through the same network structure to output a broadened recursive texture feature map.

[0204] Step S760: Input the frequency-shifted recursive texture feature map and the widened recursive texture feature map into the cross-bilinear fusion layer of the recursive graph cross-correlation analysis network, and perform an outer product operation on the local texture descriptor at each spatial position in the frequency-shifted recursive texture feature map and the local texture descriptor at the corresponding spatial position in the widened recursive texture feature map to generate a recursive texture cross-correlation feature matrix.

[0205] Align the frequency-shifted recurrent texture feature map and the widened recurrent texture feature map position by position in space. For each spatial position, extract the local texture descriptor vector u from the frequency-shifted recurrent texture feature map and the local texture descriptor vector v from the widened recurrent texture feature map at that position. Calculate the outer product of u and v to obtain a matrix. Summate and pool the outer product matrices of all spatial positions, and then perform signed square root normalization to obtain the recurrent texture cross-association feature matrix.

[0206] Step S770: Input the recursive texture cross-association feature matrix into the recursive pattern classification layer of the recursive graph cross-association analysis network, perform global average pooling and fully connected classification on the recursive texture cross-association feature matrix, and generate the recursive pattern fault type attribution probability vector corresponding to the location identifier of each spatial sampling point.

[0207] The recursive pattern classification layer first performs global average pooling on the recursive texture cross-correlation feature matrix to obtain a one-dimensional feature vector. Then, this one-dimensional feature vector is linearly mapped through a fully connected classification layer. The output dimension of the fully connected classification layer is the total number of fault subclasses M. After softmax normalization, a recursive pattern fault type attribution probability vector is generated.

[0208] Step S780: Perform a diagnostic consistency comparison process between the fault type corresponding to the maximum probability value in the recursive mode fault type attribution probability vector and the dominant fault type label corresponding to the maximum probability value in the fault type attribution probability distribution vector. The fault type with consistent diagnostic results is determined as the target fault type, and the fault type with inconsistent diagnostic results is labeled as the fault type to be reviewed.

[0209] For each spatial sampling point location identifier, extract the fault type corresponding to the maximum probability value in the recursive pattern fault type attribution probability vector, and extract the dominant fault type label corresponding to the maximum probability value in the fault type attribution probability distribution vector. Compare whether the two are consistent. If they are consistent, determine the fault type as the target fault type; if they are inconsistent, mark the fault type as the fault type to be reviewed.

[0210] Step S790: Associate and combine the target fault type and the fault type to be reviewed with the corresponding spatial sampling point location identifier to generate a fault type review and diagnosis result data set based on recursive graph cross-correlation analysis.

[0211] Each spatial sampling point location identifier is associated with its corresponding target fault type or fault type to be verified, and arranged in order of spatial sampling point location identifier number to generate a fault type verification and diagnosis result data set based on recursive graph cross-correlation analysis. This fault type verification and diagnosis result data set, together with the cable fault diagnosis report data set, is sent to the cable operation and maintenance management terminal.

[0212] In one exemplary embodiment, a deep learning-based power cable fault type diagnosis system is provided, which can be a terminal, server, etc., and its internal structure diagram can be as follows. Figure 3 As shown, it includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, near-field communication, or other technologies. When the computer program is executed by the processor, it implements a deep learning-based method for diagnosing power cable fault types. The display unit is used to form a visually visible image and can be a display screen, projection device, or virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device can be a touch layer covering the display screen, or a button, trackball, or touchpad set on the casing of the deep learning-based power cable fault type diagnosis system, or an external keyboard, touchpad, or mouse, etc.

[0213] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A method for diagnosing power cable fault types based on deep learning, characterized in that, The method includes: Acquire the raw optical time-domain scattering signal data set collected by the distributed sensing device deployed at the cable joint. The raw optical time-domain scattering signal data set includes the location markers of spatial sampling points distributed along the cable axis and the scattering radio frequency shift time-domain response curve corresponding to each spatial sampling point. The original optical temporal scattering signal data set is subjected to scattering spectrum feature demodulation processing to obtain the scattering frequency shift feature sequence corresponding to the spatial sampling point position identifiers distributed along the cable axis and the scattering spectrum broadening feature sequence corresponding to each spatial sampling point position identifier. The scattering frequency shift feature sequence includes the scattering frequency shift of each spatial sampling point position identifier at different acquisition times, and the scattering spectrum broadening feature sequence includes the scattering spectrum half-width parameter of each spatial sampling point position identifier at different acquisition times. The pre-built fault feature spatiotemporal correlation coding network is invoked to perform position coding injection and multi-head self-attention transformation processing on the scattered frequency shift feature sequence and the scattering spectrum broadening feature sequence, generating a deep temporal dependency representation vector of cable state corresponding to the position identifier of each spatial sampling point. The deep temporal dependency representation vector of cable state contains the global context mapping relationship between the frequency shift long-range temporal correlation mode feature and the broadening long-range temporal correlation mode feature. The pre-built fault type hierarchical diagnosis network is invoked to perform multi-level fault feature layer-by-layer diagnosis processing on the deep temporal dependency representation vector of the cable state, generating the fault type attribution probability distribution vector and fault boundary spatial sampling point location identifier sequence corresponding to each spatial sampling point location identifier; Based on the fault type attribution probability distribution vector and the fault boundary space sampling point location identifier sequence, a cable fault diagnosis report data set is generated, and the cable fault diagnosis report data set is sent to the cable operation and maintenance management terminal.

2. The deep learning-based power cable fault type diagnosis method according to claim 1, characterized in that, The process of demodulating the scattering spectrum features of the original optical temporal scattering signal data set yields a scattering frequency shift feature sequence corresponding to the spatial sampling point location identifiers distributed along the cable axis and a scattering spectrum broadening feature sequence corresponding to each spatial sampling point location identifier, including: Extract the scattering radio-shift time-domain response curve corresponding to the spatial sampling point location identifier from the original optical time-domain scattering signal data set, perform time-domain segmentation processing on the scattering radio-shift time-domain response curve, and divide each scattering radio-shift time-domain response curve into multiple time-domain analysis window segments according to the acquisition time sequence. Lorentz line fitting is performed on the time-domain response curve of the scattered radio frequency shift within each time-domain analysis window segment, and the center frequency offset of the fitted Lorentz curve within each time-domain analysis window segment is extracted as the scattered radio frequency shift corresponding to that time-domain analysis window segment. Arrange the scattered radio shift of each spatial sampling point location identifier in all time domain analysis window segments according to the acquisition time sequence to form the scattered radio shift feature sequence corresponding to the spatial sampling point location identifier; When performing Lorentz line fitting on the scattering frequency shift time-domain response curve within each time-domain analysis window segment, the full width at half maximum (FWHM) parameter of the fitted Lorentz curve is simultaneously extracted as the FWHM parameter of the scattering spectrum corresponding to that time-domain analysis window segment. Arrange the scattering spectrum half-width parameters of each spatial sampling point location identifier in all time domain analysis window segments according to the acquisition time sequence to form the scattering spectrum broadening feature sequence corresponding to the spatial sampling point location identifier; The scattering frequency shift feature sequence and scattering spectrum broadening feature sequence corresponding to each spatial sampling point location identifier are spliced ​​along the feature dimension to generate a scattering spectrum dual-channel feature matrix corresponding to each spatial sampling point location identifier. The scattering spectrum dual-channel feature matrix is ​​then post-processed to obtain the final output scattering frequency shift feature sequence and scattering spectrum broadening feature sequence.

3. The deep learning-based power cable fault type diagnosis method according to claim 2, characterized in that, The post-processing of the dual-channel feature matrix of the scattering spectrum to obtain the final output scattering frequency shift feature sequence and scattering spectrum broadening feature sequence includes: Temporal consistency alignment processing was performed on the scattered radio frequency shift feature sequence and the scattering spectrum broadening feature sequence respectively, and the timestamps of each time domain analysis window segment in the scattered radio frequency shift feature sequence were matched with the timestamps of each time domain analysis window segment in the scattering spectrum broadening feature sequence. Abnormal amplitude constraint processing is performed on the scattered frequency shift feature sequence and the scattering spectrum broadening feature sequence after time-series consistency alignment processing. Abnormal scattered frequency shift in the scattered frequency shift feature sequence that exceeds the preset reasonable frequency shift range is replaced with the scattered frequency shift interpolation result of the time-adjacent time domain analysis window segment. Abnormal scattering spectrum half-width parameter in the scattering spectrum broadening feature sequence that exceeds the preset reasonable broadening range is replaced with the scattering spectrum half-width parameter interpolation result of the time-adjacent time domain analysis window segment. Normalize the scattered radio frequency shift characteristic sequence and the scattering spectrum broadening characteristic sequence after abnormal amplitude constraint processing respectively. Calculate the mean and standard deviation of the frequency shift of the scattered radio frequency shift characteristic sequence. Subtract the mean from the scattered radio frequency shift and divide by the standard deviation. Calculate the mean and standard deviation of the broadening of the scattering spectrum broadening characteristic sequence. Subtract the mean from the half-width at half-maximum (WHM) parameter of the scattering spectrum and divide by the standard deviation. The preprocessed scattered radio frequency shift feature sequence and the scattering spectrum broadening feature sequence are used as the final output scattered radio frequency shift feature sequence and scattering spectrum broadening feature sequence.

4. The deep learning-based power cable fault type diagnosis method according to claim 1, characterized in that, The pre-constructed fault feature spatiotemporal correlation coding network is invoked to perform position encoding injection and multi-head self-attention transformation processing on the scattered radio frequency shift feature sequence and the scattered spectrum broadening feature sequence, generating a deep temporal dependency representation vector of the cable state corresponding to the position identifier of each spatial sampling point, including: A pre-constructed fault feature spatiotemporal correlation coding network is obtained, wherein the fault feature spatiotemporal correlation coding network includes a cascaded position coding injection layer, a first multi-head self-attention transformation layer, a feedforward fully connected transformation layer and a second multi-head self-attention transformation layer; The preprocessed scattered radio shift feature sequence and scattering spectrum broadening feature sequence are input into the position coding injection layer of the fault feature spatiotemporal correlation coding network. A sinusoidal position coding vector is injected into the scattered radio shift amount corresponding to each time domain analysis window segment in the scattered radio shift feature sequence, and a cosine position coding vector is injected into the scattering spectrum half-width parameter corresponding to each time domain analysis window segment in the scattering spectrum broadening feature sequence. The scattering frequency shift feature sequence after injecting the sinusoidal position coding vector and the scattering spectrum broadening feature sequence after injecting the cosine position coding vector are concatenated along the feature dimension to generate a joint scattering spectrum feature sequence with injected position-aware information. The scattering spectrum joint feature sequence with injected position-aware information is input into the first multi-head self-attention transformation layer. Multiple parallel scaling dot product attention heads in the first multi-head self-attention transformation layer perform attention weight calculation on the joint feature vectors corresponding to any two time-domain analysis window segments in the scattering spectrum joint feature sequence to generate the first multi-head self-attention output feature sequence. The first multi-head self-attention output feature sequence and the scattering spectrum joint feature sequence of the injected position-aware information are subjected to residual connection processing, and the residual connection processing result is subjected to layer normalization processing to obtain the first residual normalized feature sequence. The first residual normalized feature sequence is input into the feedforward fully connected transform layer. The first residual normalized feature sequence is then subjected to nonlinear feature mapping processing through the first fully connected weight matrix and nonlinear activation function in the feedforward fully connected transform layer to generate the feedforward transform feature sequence. The feedforward transform feature sequence is joined with the first residual normalized feature sequence, and the result of the residual join is subjected to layer normalization to obtain the second residual normalized feature sequence. The second residual normalized feature sequence is input into the second multi-head self-attention transformation layer to perform cross-spatial sampling point attention interaction processing, and the deep temporal dependency representation vector of the cable state is output.

5. The deep learning-based power cable fault type diagnosis method according to claim 4, characterized in that, The step of inputting the second residual normalized feature sequence into the second multi-head self-attention transformation layer for cross-spatial sampling point attention interaction processing and outputting the deep temporal dependency representation vector of the cable state includes: The second multi-head self-attention output feature sequence is generated by performing cross-spatial sampling point attention weight calculation on the feature vectors corresponding to different spatial sampling point position identifiers in the second residual normalized feature sequence through multiple parallel scaling dot product attention heads in the second multi-head self-attention transformation layer. The second multi-head self-attention output feature sequence and the second residual normalized feature sequence are subjected to residual concatenation processing, and the residual concatenation processing result is subjected to layer normalization processing to obtain the third residual normalized feature sequence. The third residual normalized feature sequence is output as the deep temporal dependency representation vector of the cable state. The feature vector corresponding to the spatial sampling point position identifier in the deep temporal dependency representation vector of the cable state contains the global context mapping relationship with the long-range temporal correlation mode features of scattering frequency shift and the long-range temporal correlation mode features of scattering spectrum broadening within the entire spatial sampling point range.

6. The deep learning-based power cable fault type diagnosis method according to claim 1, characterized in that, The pre-built fault type hierarchical diagnostic network is invoked to perform multi-level fault feature layer-by-layer diagnostic processing on the deep temporal dependency representation vector of the cable state, generating a fault type attribution probability distribution vector and a fault boundary spatial sampling point location identifier sequence corresponding to each spatial sampling point location identifier, including: Obtain a pre-constructed hierarchical fault type diagnosis network, which includes a cascaded fault existence binary classification diagnosis layer, a coarse-grained fault category diagnosis layer, and a fine-grained fault subcategory diagnosis layer. The deep temporal dependency representation vector of the cable state is input into the fault existence binary classification diagnosis layer. The first fully connected weight matrix in the fault existence binary classification diagnosis layer is used to perform linear mapping processing on the feature vector corresponding to the spatial sampling point location identifier to generate the first linear mapping feature vector. The first linear mapping feature vector is subjected to nonlinear activation processing, and the result of the nonlinear activation processing is input into the fault existence classifier in the fault existence binary classification diagnostic layer to generate the fault existence binary classification probability value corresponding to the location identifier of each spatial sampling point. The binary classification probability value of the fault existence is compared with the preset fault existence judgment threshold. The spatial sampling point location identifiers whose binary classification probability value of the fault existence exceeds the preset fault existence judgment threshold are marked as candidate fault spatial sampling point location identifiers. The feature vectors in the deep temporal dependency representation vector of the cable state corresponding to all candidate fault spatial sampling point location identifiers are extracted to form a candidate fault feature vector set. Based on the candidate fault feature vector set, fault type classification diagnosis is performed through the coarse-grained diagnosis layer of the fault category and the fine-grained diagnosis layer of the fault subcategory, and a fault boundary space sampling point location identifier sequence is generated.

7. The deep learning-based power cable fault type diagnosis method according to claim 6, characterized in that, Based on the candidate fault feature vector set, the fault type classification and diagnosis are performed through the coarse-grained diagnosis layer for the major fault categories and the fine-grained diagnosis layer for the subcategories of faults, and a fault boundary space sampling point location identifier sequence is generated, including: The set of candidate fault feature vectors is input into the coarse-grained diagnosis layer of the fault category. The second fully connected weight matrix in the coarse-grained diagnosis layer of the fault category is used to perform linear mapping on each candidate fault feature vector to generate a second linear mapping feature vector. The second linear mapping feature vector is subjected to nonlinear activation processing, and the nonlinear activation processing result is input into the fault class classifier in the fault class coarse-grained diagnostic layer to generate a fault class classification probability distribution vector corresponding to the location identifier of each candidate fault spatial sampling point. The fault class classification probability distribution vector includes the probability values ​​of partial discharge faults, overheating faults, and mechanical damage faults. The set of candidate fault feature vectors is input into the fine-grained diagnosis layer of the fault subclass. The third fully connected weight matrix in the fine-grained diagnosis layer of the fault subclass performs linear mapping processing on each candidate fault feature vector to generate a third linear mapping feature vector. The third linear mapping feature vector is subjected to nonlinear activation processing, and the result of the nonlinear activation processing is input into the fault subclass classifier in the fault subclass fine-grained diagnosis layer to generate the fault subclass belonging probability distribution vector corresponding to the location identifier of each candidate fault spatial sampling point. The probability distribution vector of the fault category is concatenated with the probability distribution vector of the fault sub-category to generate the fault type probability distribution vector corresponding to the location identifier of each candidate fault spatial sampling point. The fault type probability distribution vector contains the probability score of each fault sub-category. Based on the spatial continuity of all candidate fault spatial sampling point location identifiers, spatial boundary clustering is performed on adjacent candidate fault spatial sampling point location identifiers with the same fault type affiliation to generate a fault boundary spatial sampling point location identifier sequence. The fault boundary spatial sampling point location identifier sequence includes the starting spatial sampling point location identifier and the ending spatial sampling point location identifier of each fault segment.

8. The deep learning-based power cable fault type diagnosis method according to claim 1, characterized in that, The process of generating a cable fault diagnosis report data set based on the fault type attribution probability distribution vector and the fault boundary space sampling point location identifier sequence includes: Extract the fault sub-category corresponding to the maximum probability value of each candidate fault spatial sampling point location identifier from the fault type attribution probability distribution vector as the dominant fault type label of the candidate fault spatial sampling point location identifier; Extract the starting spatial sampling point location identifier and the ending spatial sampling point location identifier of each fault segment from the fault boundary spatial sampling point location identifier sequence, and calculate the difference between the starting spatial sampling point location identifier and the ending spatial sampling point location identifier as the number of sampling points for the spatial span of the fault segment. The actual physical length of each fault segment is calculated based on the number of sampling points for the spatial span of the fault segment and the physical distance parameters between adjacent spatial sampling points. The dominant fault type markers of all candidate fault spatial sampling point location identifiers in each fault segment are summarized, and the dominant fault type markers with the highest frequency in the fault segment are statistically determined to obtain the segment-wide unified fault type marker for that fault segment. Extract the probability score corresponding to the unified fault type label of the segment from the probability distribution vector of the fault type attribution corresponding to the location identifier of all candidate fault spatial sampling points in each fault segment, and calculate the average value of the probability scores as the fault confidence score of the fault segment. Based on the scattered radio displacement corresponding to the location identifiers of all candidate fault spatial sampling points in each fault section, calculate the gradient of the scattered radio displacement along the spatial sampling points in the fault section, and mark the location identifiers of spatial sampling points whose scattered radio displacement gradient exceeds the preset gradient threshold as fault core point location identifiers. Extract the half-width at half-maximum (WHM) parameter of the scattering spectrum corresponding to the location identifier of all candidate fault spatial sampling points in each fault section, calculate the fluctuation amplitude of the WHM parameter of the scattering spectrum along the spatial sampling points in the fault section, compare the fluctuation amplitude of the WHM parameter of the scattering spectrum with the standard fluctuation amplitude database of fault type, and obtain the auxiliary assessment parameter of fault severity. The starting spatial sampling point location identifier, ending spatial sampling point location identifier, actual physical length, unified fault type label of the section, fault confidence score, fault core point location identifier, and auxiliary assessment parameters of fault severity for each fault section are structured and combined to generate a single fault section diagnostic record data unit. Summarize the single fault section diagnostic record data units corresponding to all fault sections to generate a fault section diagnostic record data set. The fault section diagnostic record data set and the original optical time-domain scattering signal data set are time-associated and bound together to generate a cable fault diagnosis report data set carrying a timestamp. The cable fault diagnosis report data set includes the fault section diagnostic record data set and the corresponding acquisition timestamp.

9. The deep learning-based power cable fault type diagnosis method according to claim 1, characterized in that, The method further includes: Acquire a set of historical raw optical time-domain scattering signal data collected by a distributed sensing device deployed at the cable joint. The set of historical raw optical time-domain scattering signal data includes the location identifiers of historical spatial sampling points distributed along the cable axis and the historical scattering radio frequency shift time-domain response curve corresponding to each historical spatial sampling point. The historical raw optical temporal scattering signal data set is subjected to scattering spectrum feature demodulation processing to obtain the historical scattering frequency shift feature sequence corresponding to the historical spatial sampling point location identifiers distributed along the cable axis and the historical scattering spectrum broadening feature sequence corresponding to each historical spatial sampling point location identifier; Obtain a set of artificial fault type labels corresponding to the set of historical raw optical temporal scattering signal data. The set of artificial fault type labels includes the real fault type label and the real fault boundary space sampling point location identifier sequence corresponding to each historical spatial sampling point location identifier. The historical scattering frequency shift feature sequence and the historical scattering spectrum broadening feature sequence are input into the fault feature spatiotemporal correlation coding network to be trained. Forward propagation calculation is performed through the position coding injection layer, the first multi-head self-attention transformation layer, the feedforward fully connected transformation layer and the second multi-head self-attention transformation layer in the fault feature spatiotemporal correlation coding network to be trained, generating the deep temporal dependence representation vector of the cable state corresponding to the position identifier of each historical spatial sampling point. The deep temporal dependency representation vector of the cable state used for training is input into the fault type hierarchical diagnosis network to be trained. The fault existence binary classification diagnosis layer, the coarse-grained diagnosis layer of the fault category, and the fine-grained diagnosis layer of the fault subcategory in the fault type hierarchical diagnosis network to be trained are used for forward propagation calculation to generate the training fault type attribution probability distribution vector and the training fault boundary space sampling point location identifier sequence corresponding to each historical space sampling point location identifier. The fault type classification cross-entropy loss value is calculated based on the fault type attribution probability distribution vector used in training and the real fault type label, and the fault boundary regression loss value is calculated based on the fault boundary space sampling point location identifier sequence used in training and the real fault boundary space sampling point location identifier sequence. The cross-entropy loss value for fault type classification and the fault boundary regression loss value are weighted and summed to generate a joint training loss value. The gradient values ​​of all network weight parameters in the fault feature spatiotemporal correlation coding network and the fault type hierarchical diagnosis network to be trained are calculated based on the joint training loss value. The gradient descent optimization algorithm is then used to update all network weight parameters along the gradient descent direction. Repeat the step of inputting the historical scattering radio frequency shift feature sequence and the historical scattering spectrum broadening feature sequence into the fault feature spatiotemporal correlation coding network to be trained, and updating all network weight parameters along the gradient descent direction using the gradient descent optimization algorithm, until the joint training loss value converges to below the preset loss threshold. The network weight parameters of the fault feature spatiotemporal correlation coding network after the joint training loss value converges are saved as a pre-constructed fault feature spatiotemporal correlation coding network, and the network weight parameters of the fault type hierarchical diagnosis network after the joint training loss value converges are saved as a pre-constructed fault type hierarchical diagnosis network.

10. A power cable fault type diagnosis system based on deep learning, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the deep learning-based power cable fault type diagnosis method according to any one of claims 1 to 9 by executing the machine-executable instructions.

Images