A method and device for on-line monitoring of high voltage power cables

Abstract

The application discloses a kind of power grid high-voltage cable online monitoring method and device, belong to high-voltage cable state online monitoring and fault diagnosis technical field.Method includes: acquisition distributed vibration, temperature and sheath current signal and pre-processing;Adaptive space-time topology graph is constructed, and its edge weight is determined by space distance attenuation factor and real-time signal cross-correlation factor multiplication;Space-time deep fusion network is constructed, and space characteristics are extracted using physical residual guided enhanced graph convolution;After encoding time sequence characteristics by gate recurrent unit, by physical residual global statistics generation gate vector, time sequence characteristics are modulated, and fault category and fault location are output;Network is trained using composite loss function including classification loss, conditional positioning loss and physical residual consistency regularization loss;Forward calculation is carried out and alarm is given during online inference.The application embeds physical constraint into graph neural network architecture, improves diagnostic accuracy, robustness and explainability.

Description

Technical Field

[0001] This invention relates to the field of online monitoring and fault diagnosis technology for high-voltage cables, specifically to a method and device for online monitoring of high-voltage cables in power grids. Background Technology

[0002] As core equipment in urban power transmission and distribution systems, the operational reliability of high-voltage cables directly impacts power supply safety. Typical faults such as partial discharge, sheath grounding faults, and joint overheating, if not detected in time, can escalate into serious accidents like insulation breakdown. Utilizing distributed fiber optic sensing technology to acquire temperature and vibration information along the cable route, combined with sheath grounding current monitoring, to achieve online diagnosis of multiple physical quantities, has become an important technological direction for cable condition monitoring.

[0003] Currently, deep learning-based cable fault diagnosis methods mostly employ convolutional neural networks or recurrent neural networks to extract and classify features from sensor signal sequences. However, existing methods have shortcomings in the following aspects. First, most methods utilize only a single physical quantity or perform simple feature concatenation of various physical quantities, failing to effectively uncover the structured coupling relationships between multiple spatial locations along the optical fiber. Second, the graph structures used in existing graph neural networks largely rely on fixed physical distances or static cable topology, lacking the ability to capture real-time changes in the dynamic coupling of multiple physical quantities caused by faults. Third, data-driven models typically rely entirely on labeled data for training, requiring a high quantity and quality of labeled fault samples. Generalization performance is limited when labeled data is scarce, and the physical interpretability of the diagnostic process is insufficient. Furthermore, some methods exhibit inconsistencies in data processing windows or computational logic between model training and online inference phases, leading to performance degradation during deployment.

[0004] Therefore, how to effectively integrate the physical laws of cable heat conduction and signal attenuation into the deep learning network architecture while making full use of the spatial correlation structure of multiple physical quantities, and maintain consistency throughout the entire training and inference chain, so as to achieve high-precision, high-robustness and physically interpretable fault diagnosis and localization, is an urgent problem to be solved. Summary of the Invention

[0005] The purpose of this invention is to provide a method and device for online monitoring of high-voltage cables in power grids, which embeds physical constraints into a graph neural network architecture to improve diagnostic accuracy, robustness and interpretability.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for online monitoring of high-voltage power cables in a power grid, comprising: Step 1: Collect distributed vibration signals, distributed temperature signals, and sheath grounding current signals along the cable line and preprocess the collected signals to obtain normalized vibration signals, normalized temperature signals, and normalized sheath grounding current signals. Step 2: Construct an adaptive spatiotemporal topology graph; Step 3: Construct a physical information-guided spatiotemporal deep fusion network model, wherein the forward propagation process of the spatiotemporal deep fusion network model is as follows: By using differential prediction of cable thermal conduction and prediction of sheath current, the physical residuals of each fiber optic sensing node and current monitoring node are calculated to form a physical residual vector. The graph signal matrix and the physical residual vector are concatenated along the feature dimension to form an enhanced node feature matrix. Multi-layer physical enhancement graph convolution is performed on the adaptive spatiotemporal topology graph, and global pooling is performed on the graph convolution output to obtain the spatial global feature vector. The time series composed of spatial global feature vectors from multiple consecutive time points is input into a gated recurrent unit to obtain the time series feature vector at the current time point. A gating vector is generated based on the global statistics of the physical residual vector. The gating vector is then used to modulate the temporal feature vector dimension by dimension to obtain the physical guidance enhancement feature. Based on physical guidance enhancement features, the fault category probability is output through the classification head, and the normalized distance of the fault location is output through the positioning head. Step 4: Train the spatiotemporal deep fusion network model based on the training samples; Step 5: Use the trained spatiotemporal deep fusion network model to perform forward inference on the real-time acquired signals and output the fault category and fault location.

[0007] Furthermore, in step 2, an adaptive spatiotemporal topology graph is constructed, specifically as follows: Using multiple fiber optic sensing nodes and one current monitoring node distributed along the cable path as graph nodes, and using normalized vibration signals and normalized temperature signals to assign features to the corresponding nodes, a graph signal matrix is ​​formed. The spatial distance attenuation factor is calculated based on the spatial distance between nodes, and the real-time signal cross-correlation factor is calculated based on the normalized vibration signal or the normalized sheath grounding current signal within the causal time window. The edge weights are obtained by multiplying the spatial distance attenuation factor with the real-time signal cross-correlation factor. After sparsification, an adaptive adjacency weight matrix is ​​formed, which together with the graph signal matrix constitutes the adaptive spatiotemporal topology graph at the current moment.

[0008] Furthermore, the real-time signal cross-correlation factor is calculated as follows: When both nodes are fiber optic sensing nodes, the normalized vibration signal within the causal time window ending at the current time is acquired, and the absolute value of the Pearson correlation coefficient between the vibration signals of the two nodes is calculated as the cross-correlation factor. When one of the nodes is a current monitoring node, the normalized vibration signal and the normalized sheath grounding current signal within the causal time window are obtained, and the maximum absolute value of the time delay cross-correlation of the scanning with a preset time delay range is calculated as the cross-correlation factor. The time delay range is determined based on the total cable length, vibration propagation speed, and sampling interval.

[0009] Further, in step 3, a gating vector is generated based on the global statistics of the physical residual vector, specifically as follows: The spatial mean and spatial maximum of the absolute values ​​of each element in the physical residual vector are calculated. The two are then concatenated and passed through a fully connected layer and a Sigmoid activation function to obtain the gated vector.

[0010] Further, in step 4, the spatiotemporal deep fusion network model is trained based on the training samples, specifically as follows: Training samples are constructed based on historical data. Based on the training samples, a spatiotemporal deep fusion network model is trained using a composite loss function that includes classification loss, conditional localization loss for fault samples, and physical residual consistency regularization loss.

[0011] Furthermore, the physical residual consistency regularization loss characterizes the consistency between the predicted fault location of the fault sample and the centroid of the physical residual distribution. The physical residual consistency regularization loss is calculated on the fault sample, and the specific calculation process is as follows: The pseudo-location tag is calculated based on the absolute value of the physical residual of the fiber optic sensing node along the weighted centroid of the normalized distance coordinate. Then, the squared loss between the pseudo-location tag and the predicted fault location is calculated and multiplied by a boundary compensation factor that adaptively attenuates based on how close the pseudo-location tag is to the cable end.

[0012] Furthermore, step 5 also includes setting an alarm if the probability of a fault category exceeds a preset threshold.

[0013] This invention also provides an online monitoring device for high-voltage cables in power grids, applied to the aforementioned online monitoring method for high-voltage cables in power grids, comprising: a multi-source sensor acquisition front-end, an edge computing gateway, and a host computer monitoring platform connected in sequence, wherein: The multi-source sensing acquisition front end includes a distributed optical fiber sensing host and a sheath current transformer; the distributed optical fiber sensing host is connected to a single-mode sensing optical fiber laid along the cable to synchronously acquire distributed vibration signals and distributed temperature signals along the line; the sheath current transformer is fitted onto the cable sheath grounding wire to acquire the sheath grounding current signal. The edge computing gateway, as an industrial-grade embedded processing platform, is equipped with: The data preprocessing module is used to perform online filtering and normalization on the acquired signals; The graph construction and update module is used to generate an adaptive adjacency weight matrix and a graph signal matrix online based on the normalized vibration signal, normalized temperature signal and normalized sheath grounding current signal; The physical residual generation module is used to calculate the physical residual vector based on thermal conduction differential and current prediction. The physical augmentation spatial feature extraction module is used to perform physical kernel-modulated graph convolution and global pooling on the adaptive adjacency weight matrix on the augmentation node features formed by concatenating the graph signal matrix and the physical residual vector, and output a spatial global feature vector. The temporal coding module uses a gated recurrent unit to encode the spatial global feature vector sequence to obtain temporal feature vectors; The physical guidance fusion and diagnostic output module is used to generate a gated vector based on the global statistics of the physical residual vector, modulate the temporal feature vector, and output the fault category probability and the fault location normalized distance. The model management and storage unit is used to store the parameters and configuration information of the trained network model. The host computer monitoring platform is used to receive fault alarm information, fault type and fault location and display them visually.

[0014] In summary, the present invention has at least one of the following beneficial technical effects: First, an adaptive spatiotemporal topology graph construction method is proposed. By multiplying the spatial distance attenuation prior with the real-time signal cross-correlation factor to form edge weights, the graph structure can both respect the static constraints of the cable physical topology and dynamically reflect the spatiotemporal coupling changes of multiple physical quantities caused by faults. This design maintains a distance-dominated sparse graph under normal operating conditions, and automatically forms dense connections in abnormal regions when a fault occurs, enhancing the graph structure's fault identification capability and robustness.

[0015] Second, a physically enhanced graph convolution operator was designed in the spatial feature extraction module. The physical residual channel was incorporated into the node feature space, and a physical kernel modulation vector based on the longitudinal signal attenuation law of the cable was introduced into the graph convolution. This makes the feature transformation gain of different nodes related to their physical positions, which makes up for the shortcomings of traditional graph convolutional networks that ignore the differences in longitudinal physical attenuation of the cable and improves the physical consistency of feature representation.

[0016] Third, a physical-guided fusion mechanism was constructed, which uses the global statistics of physical residuals to generate gating vectors and modulates the temporal features dimension by dimension, making the network more sensitive to abnormal dimensions that deviate from the predictions of the physical model. Under limited fault labeling conditions, it can still effectively highlight fault features, thereby improving diagnostic accuracy and physical interpretability.

[0017] Fourth, a physical residual consistency regularization term is introduced into the loss function to constrain the predicted location of fault samples to be consistent with the distribution of the centroid of the physical residuals. This injects the prior knowledge of the simplified physical model into the training process in a regularized manner, reducing the reliance on a large number of precise location labels. At the same time, a boundary compensation factor is introduced to alleviate the regularization bias caused by the asymmetrical distribution of residuals in the cable endpoint area.

[0018] Fifth, a causal time window is uniformly used for cross-correlation calculation throughout the entire training and inference process. This eliminates the distribution shift caused by using future information in the offline training stage while only using historical information in the online inference stage, ensuring the consistency between the model's performance in real deployment scenarios and the offline evaluation results.

[0019] Sixth, the device of this invention achieves synchronous acquisition of vibration and temperature, two physical quantities, through a single fiber of a distributed optical fiber sensing host. Combined with the adaptive graph construction, physical enhancement graph convolution, and physical guided fusion modules integrated in the edge computing gateway, it forms a complete hardware closed loop from real-time perception of multiple physical quantities, online graph structure learning, physical constraint feature extraction to fusion diagnosis and visual alarm, meeting the requirements of real-time performance, reliability, and maintainability for actual online cable monitoring. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the device structure of the present invention; Figure 2 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0022] like Figure 1 As shown, this invention provides an online monitoring device for high-voltage cables in power grids, comprising: a multi-source sensor acquisition front-end, an edge computing gateway, and a host computer monitoring platform connected in sequence. The modules are connected via industrial Ethernet or fieldbus. These will be described in detail below: 1. Multi-source sensor acquisition front end The multi-source sensing acquisition front end includes a distributed fiber optic sensing host, a sheath current transformer, and corresponding signal conditioning units. The distributed fiber optic sensing host integrates a narrow-linewidth laser, an acousto-optic modulator, a circulator, a wavelength division multiplexer, a photodetector, and a data acquisition card, providing both distributed vibration and temperature sensing functions. The continuous coherent light emitted by the narrow-linewidth laser is acousto-optically modulated into probe light pulses, which are then injected through the circulator into a single-mode sensing fiber laid along the same path as the high-voltage cable. The returning backscattered Rayleigh light (with the same wavelength as the incident light, used for vibration sensing) and backscattered Raman light (anti-Stokes and Stokes components, wavelength shifted, used for temperature sensing) are separated by the wavelength division multiplexer and enter their respective photodetectors. Vibration signals are demodulated through coherent detection, and temperature signals are demodulated using the anti-Stokes / Stokes ratio. The sheath current transformer is mounted on the cable sheath grounding wire and senses and outputs a sheath circulating current signal. The signal conditioning unit performs analog-to-digital conversion and anti-aliasing filtering on the above signals, and sends the digital signals to the edge computing gateway via RS-485 or Ethernet interface.

[0023] 2. Edge computing gateway The edge computing gateway is an industrial-grade embedded processing platform, which internally deploys a data preprocessing module, a graph construction and update module, a physical residual generation module, a physical augmentation spatial feature extraction module, a temporal coding module, a physical guided fusion and diagnostic output module, and a model management and storage unit.

[0024] The input of the data preprocessing module is connected to the signal conditioning unit, receiving vibration, temperature and current digital signals. It has a built-in bandpass digital filter and sliding window normalization subunit to perform online filtering and standardization of the signals and output a normalized feature sequence.

[0025] The graph construction and update module receives a normalized feature sequence at its input end. It contains a circular buffer and a cross-correlation calculation engine. Based on the cross-correlation of signal delays between nodes, it generates an adaptive adjacency weight matrix online and outputs the graph signal matrix and adjacency matrix.

[0026] The physical residual generation module embeds a cable thermal conduction differential calculation unit and a sheath current steady-state prediction unit, reading normalized temperature and current data as well as preset cable thermal diffusion parameters. The residual vector between the physical prediction value and the measured value is calculated in real time, and the residual vector is sent to the spatial feature extraction module and the physical guidance fusion module respectively.

[0027] The physical enhancement spatial feature extraction module receives the graph signal matrix, adjacency matrix, and physical residual vector. Internally, it implements a two-layer graph convolution operator with physical kernel modulation (16 and 32 output channels respectively) to perform spatial aggregation of node features and output a spatial feature matrix. The spatial global feature vector is generated by concatenating the results after global average pooling and global max pooling. .

[0028] The temporal coding module consists of two stacked gated recurrent units (GRUs) with 64 hidden units, cascaded with the preceding spatial feature extraction module. Internally, it maintains a hidden state cache, updated after each frame of inference, and reads continuously. The spatial global feature vector sequence at each time step outputs a temporal feature vector fused with historical information. .

[0029] The physical-guided fusion and diagnostic output module includes a gated fusion subunit, a fault classification subunit, and a fault location subunit. The gated fusion subunit outputs the global residual strength from the physical residual generation module. and The gating signal source is generated by passing through a fully connected layer and a Sigmoid activation layer to produce a gating vector. For time series characteristics Dimensional modulation is performed to obtain physically guided enhancement features. This enhanced feature is fed in parallel into the fault classification subunit (two fully connected layers + Softmax, output...) 3D Fault Category Probability Vector ) and fault location subunit (two layers of fully connected + Sigmoid, outputting a normalized fault location scalar) The output of the fault classification subunit is connected to a comparator. When the maximum probability exceeds a set confidence threshold... An alarm signal is triggered at that time.

[0030] The edge computing gateway also has a model management and storage unit, which stores the trained model parameters and configuration information (including normalization parameters, filter coefficients, graph construction hyperparameters, etc.) and supports receiving model update instructions from the host computer to achieve online incremental updates.

[0031] 3. Host computer monitoring platform The host computer monitoring platform connects to the edge computing gateway via Ethernet, receives fault alarm information, fault type, location results and sensor health status, performs visual calibration of fault points and alarm prompts on the cable geographic information system interface, and records historical data for model iterative training.

[0032] Each device is connected sequentially according to the data flow: the three digital signals from the multi-source sensor acquisition front end are preprocessed and then flow into the graph construction and update module and the physical residual generation module, respectively; the graph signal matrix and adjacency matrix output by the graph construction and update module, together with the residual vector output by the physical residual generation module, enter the physical enhancement spatial feature extraction module; the spatial global feature vector output by the spatial feature extraction module enters the temporal coding module; the temporal feature vector output by the temporal coding module and the global residual intensity output by the physical residual generation module enter the physical guided fusion and diagnostic output module; finally, the diagnostic results are reported to the host computer. The device as a whole realizes a complete hardware closed loop from real-time perception of multiple physical quantities, online graph structure learning, physical constraint feature extraction to fusion diagnosis.

[0033] like Figure 2 As shown, the present invention also provides an online monitoring method for high-voltage power cables in power grids, applied to the aforementioned online monitoring device for high-voltage power cables in power grids, comprising: Step 1: Collect distributed vibration signals, distributed temperature signals, and sheath grounding current signals along the cable line and preprocess the collected signals to obtain normalized vibration signals, normalized temperature signals, and normalized sheath grounding current signals. Step 2: Construct an adaptive spatiotemporal topology graph; Step 3: Construct a spatiotemporal deep fusion network model guided by physical information; Step 4: Train the spatiotemporal deep fusion network model based on the training samples; Step 5: Use the trained spatiotemporal deep fusion network model to perform forward inference on the real-time acquired signals and output the fault category and fault location.

[0034] Next, the above steps will be explained in detail with reference to specific parameter settings: In step 1, distributed vibration signals, distributed temperature signals, and sheath grounding current signals along the cable route are collected. The collected signals are then preprocessed to obtain normalized vibration signals, normalized temperature signals, and normalized sheath grounding current signals, specifically: This step utilizes the collaborative operation of a distributed fiber optic sensing host and a current transformer to achieve synchronous sensing of multiple physical quantities related to the operating status of high-voltage cables. Specifically, it includes the following sub-steps: Step 101: Distributed vibration signal acquisition The distributed fiber optic sensing host integrates a narrow-linewidth laser, which emits a center wavelength of... ( Continuous coherent light (with a center wavelength in nm) is modulated by an acousto-optic modulator into a pulse width of [missing value]. ( The probe light pulse has a pulse width (in nanoseconds) and a pulse repetition frequency of . ( The pulse repetition frequency (in Hz) is used to detect the light pulse, which is injected into a single-mode sensing fiber laid along the same path as the high-voltage cable through a circulator. The backscattered Rayleigh light generated in the sensing fiber returns along the original path, passes through the circulator, and is converted into an electrical signal by a photodetector. Because Rayleigh scattered light is extremely sensitive to minute strains along the fiber, when abnormal events such as partial discharge or external force damage occur in the cable, the corresponding fiber segment experiences micro-strain changes, causing modulation of the phase and amplitude of the backscattered Rayleigh light. This scattered light signal is demodulated through coherent detection. The returned Rayleigh scattered light is heterodyne-interfered with the local reference light, and after photoelectric conversion, the phase change information is extracted through digital quadrature demodulation. Based on the principle of optical time-domain reflection, the phase change information is determined by the echo delay. ( To detect the time difference (in seconds) between the pulse emission time and the scattered light reception time, and the speed of light in the optical fiber. ( The relationship between the group velocity of light in the sensing fiber (in m / s) determines the spatial location of the anomalous event. ( (The distance from the beginning of the optical fiber to the abnormal event, in meters). ; After demodulation, a discretized vibration signal vector along the cable length is obtained. ( For the first Vibration signal vector at each sampling time, This represents the total number of sensor channels. express (a real vector space), where ( (The total number of sensing channels is determined by the spatial sampling resolution; dimensionless) ( (The index for the sampling time is dimensionless) represents the first sampling time. Each sampling time point. This vibration signal is used in subsequent steps to capture abnormal mechanical vibration characteristics of the cable body and accessories, providing dynamic behavioral information for fault detection.

[0035] Step 102: Distributed Temperature Signal Acquisition The distributed fiber optic sensing host also integrates a wavelength division multiplexer (WDM) to separate the anti-Stokes Raman scattered light component in the returned light, whose wavelength differs from the incident light. In this scheme, vibration sensing and temperature sensing share the same single-mode sensing fiber: vibration sensing utilizes Rayleigh scattered light with the same wavelength as the incident light, while temperature sensing utilizes anti-Stokes Raman scattered light with a wavelength shifted towards shorter wavelengths. The two are separated in the optical domain by the WDM. Although the Raman scattered light intensity in single-mode fiber is weaker than in multimode fiber, by increasing the incident light power, employing a high-sensitivity avalanche photodetector, and using multi-pulse accumulation averaging technology, the signal-to-noise ratio of temperature measurement can be ensured to meet engineering requirements. The intensity of the back-facing anti-Stokes Raman scattered light is highly sensitive to the fiber's body temperature; simultaneously, the Stokes Raman scattered light with a wavelength shifted towards longer wavelengths is not sensitive to temperature and is used as a reference signal. The anti-Stokes light signal output by the WDM is demodulated by ratio with the Stokes reference signal to eliminate light source power fluctuations and environmental interference, obtaining the absolute temperature value at each location along the fiber. Similarly, after spatial positioning based on the principle of optical temporal reflection, a distributed temperature signal vector is obtained. ( For the first Temperature signal vector at each sampling time, express (A real-valued vector space). This temperature signal reflects the operating temperature of each section of the cable and is used to detect abnormal local temperature rises caused by insulation defects or overloads.

[0036] Step 103: Acquisition of sheath grounding current signal A through-type current transformer is installed on the grounding lead of the metal sheath at the high-voltage cable terminal. The secondary output of this current transformer is based on its rated transformation ratio. ( The current transformer's rated turns ratio (dimensionless) is an AC voltage signal proportional to the primary side sheath grounding current. This voltage signal undergoes amplitude adjustment and anti-aliasing filtering by a signal conditioning circuit before being sent to a multi-channel synchronous acquisition unit. After sampling, the sheath grounding current scalar is obtained. ( For the first The sheath grounding current value at each sampling time. (Representing a real number), its physical meaning is the power frequency and high frequency stray current flowing through the metal sheath of the cable. This current signal is used to reflect the insulation integrity of the cable sheath and the status of the grounding system. An abnormal increase in the sheath current is usually associated with sheath damage or poor grounding.

[0037] Step 104: Multi-channel synchronous sampling Vibration simulation signal output by the distributed fiber optic sensor host ( For vibration simulation signals over time Changes in temperature (in V), temperature analog signal ( Temperature simulation signal over time The change (in V) and the analog signal of the sheath current after signal conditioning of the current transformer. ( The simulated signal of sheath current over time The changes (in V) are respectively connected to three independent analog input channels of the multi-channel synchronous acquisition unit. This multi-channel synchronous acquisition unit has a built-in GPS synchronization module, which outputs a high-precision second pulse signal and absolute time code. The multi-channel synchronous acquisition unit uses the second pulse signal as the unified trigger clock for the analog-to-digital converter, ensuring strict synchronization of the sampling times of each channel. Each channel operates at a sampling frequency... ( The sampling frequency (in Hz) is used for equally spaced sampling, which satisfies the Nyquist sampling theorem and covers the highest frequency component of the fault characteristics in each signal. After synchronous analog-to-digital conversion, the digital signal output from each channel is stamped with a uniform absolute timestamp, forming three strictly time-aligned digital signal streams, which are output to the preprocessing buffer in the edge computing node for subsequent multi-source data preprocessing.

[0038] Step 105: Multi-source data preprocessing includes: sequentially performing signal filtering, moving average denoising, and normalization on the three digital signal streams to eliminate noise interference and dimensional differences, forming a unified multi-source sample vector, specifically: 1. Bandpass filtering of vibration signals Vibration signal The fault typically contains low-frequency drift caused by environmental micro-vibrations and high-frequency noise caused by fiber shot noise. To preserve the frequency band components relevant to the fault characteristics, a digital bandpass filter is used. The lower limit of the filter's cutoff frequency... ( The lower and upper limits of the bandpass filter cutoff frequency (in Hz) are defined as follows: ( The upper limit of the bandpass filter cutoff frequency (in Hz) is determined based on the prior spectral characteristics of partial discharge and mechanical vibration. The filtering operation is implemented using a finite impulse response filter with linear phase characteristics, whose transfer function... ( Let be the transfer function of the digital filter. (For variables in the Z-transform domain) are: ; In the formula ( (where the order of the denominator polynomial of the filter is dimensionless) and ( (where is the order of the filter numerator polynomial, dimensionless) is the filter order. ( (These are the denominator coefficients of the filter, dimensionless) ( The filter numerator coefficients (dimensionless) are designed offline and embedded in the digital signal processing module of the edge computing node using the window function method or Chebyshev approximation method. The filtered vibration signal is obtained by processing with this transfer function. ( This is the filtered vibration signal vector. express (A real vector space). The output signal retains the frequency band energy closely related to cable faults and suppresses irrelevant interference.

[0039] 2. Bandpass filtering of sheath current signal Protective layer current signal Harmonic interference beyond the fundamental power frequency and high-frequency white noise also exist. A bandpass filter with the same structure as the vibration signal is used, but its cutoff frequency is designed separately based on the high-frequency components in the sheath current that reflect insulation defects and partial discharge. The filtering operation utilizes the filter coefficients embedded in the edge computing nodes. Perform convolution operation and output the filtered sheath current signal. ( This is the filtered sheath current signal. (Represents a real number).

[0040] 3. Temperature signal moving average noise reduction Temperature signal The rate of change of the temperature signal is much slower than that of vibration and current signals, and its noise mainly consists of slow temperature drift and random noise from the temperature measurement system itself. A moving average method is used for smoothing, and the denoised temperature signal is calculated. ( This is the temperature signal vector after denoising via moving average. express (dimensional real vector space) ; In the formula ( The sliding window length (unit: number of sampling points, dimensionless) represents the length of the temperature signal's moving average. This is based on the slow-changing characteristics of the temperature signal. Selected as the sampling frequency The number of sampling points within a corresponding several-minute time window. This operation suppresses high-frequency fluctuations in temperature measurement noise while preserving the long-term trend of temperature changes, providing a stable temperature reference for subsequent physical residual calculations.

[0041] 4. Signal normalization processing Different physical quantities have significantly different amplitude ranges and dimensions, and direct splicing will weaken the contribution of small-range signals in subsequent networks. To unify the input space, maximum-minimum normalization is performed on the three types of signals respectively.

[0042] Filtered vibration signal Normalization yields ( The normalized vibration signal vector, express (dimensional real vector space) ; In the formula ( (This represents the historical minimum value of the vibration signal, with units consistent with the original units of the vibration signal.) ( (These are the historical maximum values ​​of the vibration signal, with units consistent with the original units of the vibration signal.) They are the historical minimum and maximum values ​​of the vibration signal obtained during the offline statistics phase, and are stored in the configuration parameter table of the edge computing node.

[0043] Filtered sheath current signal Normalization yields ( This is the normalized sheath current signal. (Representing real numbers): ; In the formula ( (The minimum historical value of the sheath current signal, in A) and ( (The maximum historical value of the sheath current signal is in A) and (the minimum and maximum historical values ​​of the sheath current are statistically recorded).

[0044] For the denoised temperature signal Normalization yields ( The normalized temperature signal vector, express (dimensional real vector space) ; In the formula ( (The temperature signal represents the historical minimum value, in °C) and ( (The maximum historical temperature signal is represented by ℃). The minimum and maximum historical temperature signals are represented by ℃. All normalization parameters are determined using the full historical data before offline model training and remain unchanged during subsequent online runs to ensure consistent distribution of input features.

[0045] 5. Multi-source sample splicing The normalized signals are spliced ​​together in a fixed order to form a time sequence. Multi-source sample vectors ( For multi-source sample vectors, express (dimensional real vector space) ; In the formula, the symbol This indicates that vectors are concatenated vertically column-wise. Include A vibration normalized value, Include A normalized temperature value, This is the normalized value of a single sheath current. Sample vector. The dimension is .

[0046] This sample vector will be used as the signal matrix in step 2. The basic structure. Among them... and Corresponding to the characteristics of each fiber optic sensing node, The characteristics of the corresponding terminal current monitoring node are then determined. Step 1 concludes here, and the output is a normalized multi-source sample vector with timestamps. Then, it is fed into step 2 for adaptive graph construction.

[0047] In step 2, an adaptive spatiotemporal topology graph is constructed, specifically as follows: The multi-source sample vector output in step 1 Includes normalized vibration signals Normalized temperature signal and normalized sheath grounding current signal The goal of this step is to use this real-time data to construct an adaptive spatiotemporal topology graph that reflects the spatial physical relationships and real-time operating status along the cable route, providing a structured graph input for subsequent graph convolutional networks. This step specifically includes the following sub-steps: Step 201: Graph Node Definition and Feature Initialization Definition diagram ( For a moment Adaptive spatiotemporal topology graph, For a set of nodes, Let be the set of edges. The set of nodes (for the weight matrix) Total of ( (The total number of nodes in the graph is dimensionless.) node ( (Node index, dimensionless): corresponding to distributed fiber optic sensing Each spatial channel represents the physical location of the cable. The spatial coordinates of these nodes can be calibrated offline based on the fiber optic laying path and optical temporal reflectance resolution, and stored in the configuration table of the edge computing nodes. The feature vector of each node is composed of the normalized vibration and temperature signal values ​​at that location obtained in step 2, i.e., the node's... At any moment The feature vector is .

[0048] node This corresponds to the sheath grounding current monitoring point at the cable terminal. The physical location of this node is at the terminal grounding box at either the beginning or end of the cable line, and its spatial coordinates are also calibrated offline. Its eigenvector is taken from the normalized current value and padded with zeros to the same dimension as the fiber optic node. .

[0049] All node features are stacked row by row according to node number to form a graph signal matrix. ( The matrix is ​​a graph signal matrix (each row corresponds to a 2D feature vector of a node). This matrix will be used as the input to the spatial feature extraction module in step 4.

[0050] Step 202: Mechanism for Constructing Graph Edge Weights Edge set A fully connected form is used, meaning there is an undirected edge between any two distinct nodes, but the weight of the edge is not specified. ( For nodes With nodes Between moments The edge weights (dimensionless) are dynamically determined by the product of two factors: the spatial distance attenuation factor and the real-time signal cross-correlation factor. This design differs from existing technologies that use fixed physical distance maps or purely data-driven maps, and is the core improvement of this step.

[0051] (a) Spatial distance attenuation factor For nodes With nodes The physical distance along the cable path is calculated based on its offline calibrated spatial coordinates. ( For nodes With nodes The physical distance along the cable path between them, in meters. The spatial distance attenuation factor is defined as a Gaussian kernel form: ; In the formula ( The spatial scale parameter, namely the characteristic attenuation length of the physical correlation along the cable (in meters), is a spatial scale parameter. Its physical meaning is the characteristic attenuation length of the physical correlation along the cable. It can be set according to the cable type, laying method and the statistical value of the spatial influence range of historical faults. It is usually taken as one-tenth to one-fifth of the total cable length. The prior knowledge of the cable physical topology is encoded: the closer the two nodes are, the stronger their correlation in terms of heat conduction, electromagnetic coupling and mechanical vibration propagation, and therefore they should have a higher basic weight in the diagram.

[0052] (b) Real-time signal cross-correlation factor To capture signal coupling changes caused by faults during cable operation, a second factor is introduced. ( For nodes With nodes Between moments Real-time signal cross-correlation factor, value range (Dimensionless), this factor is calculated in real time based on the normalized signal output in step 2. To ensure consistency between training and online inference, the cross-correlation calculation uniformly adopts a causal time window, that is, only the data at the current time and historical time are used, and data at future time are not used.

[0053] When node and nodes All are fiber optic sensing nodes (i.e. At this time, the time-domain Pearson correlation coefficient between the two nodes is calculated using the normalized vibration signal. The current time is taken as... The endpoint is [missing information], and the time window length is [missing information]. ( The causal local window is the time window length for cross-correlation calculations (unit: number of sampling points, dimensionless). ,calculate: ; In the formula For a moment node Normalized vibration value at the location, ( For nodes At any moment The mean of vibrations within the causal window (dimensionless) and ( For nodes At any moment The mean vibrations within the causal window (dimensionless) represent the mean vibrations of the two nodes within the causal time window. The range of values ​​is The larger the absolute value of the cross-correlation factor, the more similar or inversely correlated the vibration modes of the two nodes are near the current moment. When partial discharge or mechanical vibration occurs at a point in the cable, the vibration signals of the fault point and its neighboring nodes will show strong correlation characteristics, while the correlation of nodes far from the fault point will be weak. Therefore, this cross-correlation factor can reflect the changes in spatial coupling strength caused by the fault in real time.

[0054] When one of the nodes is a current monitoring node ( or When the other node is a fiber optic sensing node, the correlation coefficient between vibrations cannot be directly used. In this case, the maximum value of the time-delay cross-correlation between vibration and current is used as the real-time correlation metric. A causal time window is then used. compute nodes Normalized vibration With normalized current The partially cross-correlated sequences between them, and the maximum absolute value is taken: ; In the formula ( (where is the time delay step size, in units of the number of sampling points, dimensionless) ( The maximum time-delay search step size (unit: number of sampling points, dimensionless) is determined by the total cable length. ( (Total cable length, in meters) and the speed of vibration propagation in the cable. ( (The speed of vibration wave propagation in the cable, in m / s) and sampling interval. ( The sampling time interval (in seconds) is jointly determined, and the calculation formula is as follows: ( (This indicates rounding up). This value reflects the temporal correlation strength between fiber vibration anomalies and sheath current anomalies, and can capture multi-physical quantity coupling phenomena caused by the same discharge source or fault point.

[0055] To unify the dimensions and combine it with the spatial distance factor, a real-time signal cross-correlation factor is defined. The range of values ​​is .

[0056] (c) Edge weight composition and sparsification Multiplying the two factors above, we obtain the edge weights: ; when When, set This multiplicative form ensures that the edge weights respect physical distance constraints while dynamically adjusting based on the correlation between real-time signals. Under normal cable operation, the vibration correlation between nodes is low. Generally small, weight matrix Primarily determined by distance attenuation, approximating the prior physical map; when a fault occurs, the distance between nodes in the fault region... The graph structure increases significantly, and high-weight connections appear in the corresponding local regions, guiding the graph convolutional network to focus on these anomalously coupled regions.

[0057] To reduce the computational cost of subsequent graph convolutions and filter out meaningless weak connections, a threshold sparsity operation is introduced. A weight threshold is set. ( (This is a weight sparsity threshold, dimensionless). Weights below this threshold will be reset to zero. ; threshold It can be set to a multiple of the median of all non-zero weights, or determined based on the autocorrelation baseline obtained from historical normal operation data.

[0058] The final weight matrix ( For a moment The adaptive adjacency weight matrix is ​​the time step (i.e., the time step). The adaptive spatiotemporal topology graph.

[0059] In step 3, a spatiotemporal deep fusion network model guided by physical information is constructed, specifically as follows: The network model received graph signal matrix constructed in this step and adaptive weight matrix The aim is to extract physically consistent spatiotemporal features from multi-source sensor data along the cable route and output the probability distribution of fault types and continuous estimates of fault locations. The model for this step consists of four cascaded core modules: a physical residual calculation module, a spatial feature extraction module, a temporal feature encoding module, and a physical information-guided fusion module. These will be described in detail below: 1. Physical Residual Calculation Module This module utilizes a simplified cable physical model, inputs the current moment's physical prediction value based on historical states, and compares it with the measured value to form physical residual features. These residuals are injected as guiding information into subsequent fusion processes, enabling the network to pay more attention to areas where the physical model's predictions are inaccurate (i.e., anomalous areas).

[0060] For each fiber optic sensing node The physical prediction of its temperature signal is based on the explicit differential discretization form of the one-dimensional heat conduction equation. The heat conduction equation retains its linear form under the minimum-maximum normalization transformation (normalization is a linear mapping, the heat conduction equation is a linear partial differential equation, and the normalized temperature value still satisfies the same form of the heat conduction equation; only the numerical scale of the thermal diffusivity is scaled proportionally). Therefore, the normalized temperature can be used directly. Perform the calculation. Let the current time be... Take node Previous moment Measured normalized temperature The normalized temperatures of adjacent nodes are used to calculate the predicted temperature at the current moment using the discretized heat conduction equation. ( For nodes At any moment (Normalized temperature prediction, dimensionless) ; In the formula ( is the thermal diffusivity, which is dimensionless; The equivalent thermal diffusivity of the cable is expressed in meters (m). 2 / s; The sampling time interval is expressed in seconds (s). (This represents the spatial resolution of fiber optic sensing, in meters), and the thermal diffusivity. Boundary nodes. and An adiabatic boundary condition approximation is used, that is, extrapolation is performed using the temperature gradient of adjacent internal nodes: for : ; for : ; For the sheath current node Its physical prediction is based on a quasi-steady-state model of the sheath current. Under normal operating conditions, the sheath current is mainly generated by capacitive coupling and is related to the cable operating voltage and sheath impedance. When the operating voltage is relatively stable, the sheath current changes slowly. Therefore, the measured value at the previous moment is used as a naive prediction for the current moment: ( For a moment (Normalized predicted sheath current, dimensionless).

[0061] The predicted values ​​are compared with the normalized measured values ​​obtained in step 2, and the physical residuals of each node are defined: For fiber optic nodes : ; For current nodes : ; Stack the residual values ​​of each node according to their node numbers to form a physical residual vector. ( For a moment The physical residual vector, For nodes The physical residual value (dimensionless) reflects the degree to which the actual cable condition deviates from the simplified physical model: in the normal operating range, the residual value is close to zero; in areas of abnormal heating, abnormal vibration, or insulation degradation, the physical model is violated, and the residual amplitude will increase significantly. This residual vector will serve as the additional node feature channel for the spatial feature extraction module and the gated signal source for the subsequent fusion module.

[0062] 2. Spatial Feature Extraction Module The adaptive graph constructed by this module in step 2 Graph convolution operations are performed on the network to extract spatially aware feature representations. To differentiate it from existing standard graph convolutional networks, this module proposes a physically enhanced graph convolution operator. Its core improvement lies in incorporating physical residual channels into the node feature space and performing depth-separable physical kernel modulation on the adjacency matrix.

[0063] First, construct the enhanced node feature matrix. ( To enhance the node feature matrix (each row corresponds to a 3D enhanced feature vector of a node), the original graph signal from step 3... (Including normalized vibrations and temperature) and physical residuals Concatenate along the feature dimension: ; That is, each node The input feature vector is ,in This refers to the physical residual.

[0064] Subsequently, a physically enhanced graph convolution operation is performed. This module is stacked. ( Let be the number of graph convolutional layers in the spatial feature extraction module (dimensionless). Let the be the . Layer input is ( For the first The input feature matrix of layer graph convolution, For the first The number of channels in the layer (dimensionless), the first layer input , Number of output channels in the first layer Number of output channels in the second layer ( (This refers to the final output channel number of the spatial feature extraction module, dimensionless).

[0065] The convolution operation for each layer is defined as follows: For the first layer... Each output channel of the layer : ; In the formula: ( The adjacency matrix is ​​the symmetric normalized adjacency matrix. for The degree matrix, diagonal elements () is the symmetric normalized adjacency matrix. for The degree matrix.

[0066] Indicates the first The first layer output The feature vectors of each channel.

[0067] ( For the first From the input channel in the layer To output channel The physical core modulation vector (PMC) is a physical core modulation vector corresponding to each node. This vector is not a traditional learnable weight matrix, but rather a modulation coefficient designed offline and dynamically corrected online based on the cable's physical characteristics. Its calculation formula is: ; in ( For nodes Normalized distance from the beginning of the cable path. (dimensionless) as nodes Normalized distance from the beginning of the cable path. ( (This refers to the characteristic length of cable signal attenuation, dimensionless, measured by normalized distance.) ( The baseline learnable parameters (dimensionless) for the physical kernel modulation vector and ( The attenuation learnable parameter (dimensionless) of the physical kernel modulation vector is a learnable scalar parameter. This design embeds the physical laws of signal attenuation along the cable into the channel mixing process of the graph convolution kernel—nodes near the terminal gain stronger feature transformation gain due to lower signal attenuation, while nodes farther from the terminal gain weaker transformation due to higher signal attenuation, enabling spatial feature extraction to express the longitudinal physical attenuation characteristics of the cable. Each input-output channel pairs... It has a set of independent learnable parameters and This enables differentiated mixing between channels. This is a significant improvement over traditional graph convolutional networks where all nodes share the same transformation matrix.

[0068] This represents element-wise multiplication. For the activation function, LeakyReLU is selected.

[0069] The above formula fully realizes cross-channel information mixing: each output channel All input channels The weighted summation result after graph aggregation and physical modulation, channel mixing is achieved through... Completed, physical modulation via Node-by-node differentiation is achieved, and graph aggregation is achieved through... Information propagates along the graph topology.

[0070] After two graph convolutions, the spatial feature matrix is ​​obtained. ( This is a spatial feature matrix, where each row corresponds to a node. (3D spatial feature embedding). Each row of this matrix corresponds to the spatial feature embedding of a node.

[0071] To reduce the number of parameters and extract global layout information, Perform global average pooling and global max pooling on the rows, and then concatenate the two to obtain the spatial global feature vector. ( For a moment (spatial global feature vectors) ; This vector summarizes the entire cable path at time [time]. Spatial characteristic distribution.

[0072] 3. Temporal Feature Encoding Module Spatial global feature vector Evolves over time. To capture the temporal dynamics of fault development, a gated loop unit is used to continuously... ( The length of the input sequence for time-series coding is given in units of sampling points (dimensionless). The spatial feature sequence at time points (time points) is encoded. to Spatial global feature vector sequence Input them in chronological order into a stacked two-layer array with a number of hidden units. ( Let be the number of hidden units in a gated loop unit (dimensionless). The gating mechanism within a gated loop unit consists of an update gate. ( To update the gate vector, control the degree of retention of historical states, and reset the gate. ( To reset the gate vector (controlling the degree to which historical states participate in candidate state calculations) and the candidate hidden state. The structure is defined by standard rules and will not be elaborated here. The hidden state of the last time step of the last layer is used as the output of the temporal coding, denoted as... ( For a moment (Time-series feature vector). This vector incorporates past... The spatial feature evolution information at each moment can characterize the accumulation process of faults and the temporal context of transient events.

[0073] 4. Physical Information Guidance and Fusion Module Standard deep learning methods typically directly incorporate temporal features The current method of feeding data into a fully connected layer for classification and localization treats all feature dimensions equally under both normal and abnormal conditions, failing to highlight abnormal features arising from the disruption of physical laws. This module proposes a physical information-guided gating fusion mechanism, utilizing physical residuals to construct gating vectors and dynamically modulating temporal features, making the network more sensitive to physically inconsistent regions.

[0074] First, calculate the global physical residual strength. Take the physical residual vector. The mean and maximum values ​​of the absolute values ​​of each element in space are then concatenated and linearly transformed to obtain the residual gating vector. ( For the residual gate vector, ): ; ; In the formula ( For a moment The global physical residual space average (dimensionless) and ( For a moment The global physical residual space maximum value (dimensionless) represents the residual mean and maximum value, respectively. ( (for learnable weight matrices of gating fusion) and ( (where the learnable bias vector is the gated fusion learnable parameter) This is the Sigmoid activation function. Each element takes the value of , representing the confidence strength of the corresponding feature dimension in an abnormal event.

[0075] Next, the temporal features are modulated using a gating vector to obtain physically guided enhanced features. ( (The enhanced temporal feature vector after physical guidance) ; This approach employs multiplicative gating instead of additive residuals. When the physical residuals are weak (in normal operation), each dimension of the gating vector tends towards the baseline value, and the features are scaled overall. When the physical residuals are significant, the anomaly-related dimensions in the gating vector are amplified, while irrelevant dimensions are suppressed, significantly enhancing the model's selectivity for physical anomaly dimensions. This guided fusion mechanism, where physical information directly impacts the feature representation level, is an innovative design not found in existing graph- or time-series-based cable fault diagnosis networks.

[0076] Finally, the enhanced features will be... Input is a parallel dual-head fully connected layer for output: Fault classification head: Consists of two fully connected layers and Softmax activation. The first layer will... Mapped to 128 dimensions, second layer mapped to dimension( (Total number of fault categories, including normal operating conditions, dimensionless) - Output ( For a moment The predicted fault category probability vector, where the sum of its components is 1.

[0077] Fault location head: Composed of two fully connected layers, the first layer will... Mapped to 64 dimensions, the second layer maps to 1 dimension, and the output scalar is activated by a Sigmoid function. ( For a moment The normalized distance to the predicted fault location (0 represents the cable head end, 1 represents the cable tail end) indicates the normalized distance from the predicted fault location to the cable head end. This location output is ignored by the conditional mechanism in the loss function when the classification head determines the condition to be normal.

[0078] At this point, the network forward propagation in step 3 is complete. Model output. and The input from step 4, together with the real labels, constructs a composite loss function, which is used to iteratively optimize the network parameters.

[0079] In step 4, the spatiotemporal deep fusion network model is trained based on the training samples, specifically as follows: Fault classification probability distribution output in step 3 and fault location scalar The results need to be compared with the true labels to calculate the optimization objective. This step proposes a physical information-guided composite loss function, which consists of a weighted average of classification loss, conditional localization loss, and physical residual consistency regularization. These will be described in detail below: 1. Multi-source data annotation format Before training the model, historical cable operation data needs to be labeled. Each training sample corresponds to a specific time point. The annotations include: Fault Category Labels ( For a moment The actual fault category label, 0 is normal, 1 to (for each fault type), among which Indicates normal operating status. to These represent different types of faults (such as partial discharge, sheath grounding fault, joint overheating, etc.). Fault location label ( For a moment The normalized distance to the actual fault location (represented by the normalized distance from the fault point to the cable head) indicates the normalized distance from the fault point to the cable head. (In normal condition) Set to invalid flag value This indicates that the sample is not included in the calculation of the localization loss.

[0080] 2. Fault Classification Losses The fault classification loss adopts a weighted multi-class cross-entropy loss function to address the common class imbalance problem in cable fault data.

[0081] Define the weight vectors for each category ( (a category weight vector), its elements ( For category The loss weights (dimensionless) are calculated and normalized based on the reciprocal of the number of samples in each class in the training set: ; In the formula ( For categories in the training set The number of samples (dimensionless) represents the number of classes in the training set. The number of samples. The introduction of class weights can effectively alleviate the training bias caused by the fact that the number of samples in normal operation is much larger than the number of samples in each fault category.

[0082] The single-sample classification loss is defined as: ; In the formula This is an indicator function (it takes the value 1 if the condition is true, and 0 otherwise). ( For the prediction vector The Each component, i.e., category The predicted probability is the prediction vector. The Each component.

[0083] 3. Location loss due to conditional failure The fault location loss is activated only when the true label of the sample is in an abnormal operating state. The location loss uses a smoothed L1 loss to balance the sensitivity of the mean square error loss to outliers and the non-differentiability of the absolute error loss at zero.

[0084] ; Define conditional activation function ( (Conditional activation indicator for localization loss, with a value of 0 or 1). ; The conditional localization loss for a single sample is: ; 4. Physical residual consistency regularization This is the core of the loss function. Its design principle is as follows: when an actual cable fault occurs, the measured temperature or vibration signal near the fault point will significantly deviate from the predicted value of the simplified physical model, i.e., the physical residual. Significant values ​​appear in the fault area. Therefore, the model predicts the fault location. It should be consistent with the high response region of the physical residual in space.

[0085] First, calculate the physical residual vector. Medium current node (index) Remove the component, retain only the component that is removed. The residual of each fiber node is denoted as . ( (This is the physical residual vector containing only fiber optic nodes). The normalized coordinates of each node along the cable path are: ( For nodes Normalized coordinates along the cable path, where 0 represents the beginning and 1 represents the end. .

[0086] Define residual-based pseudo-location labels ( For pseudo-location labels based on the centroids of the physical residual distribution, the normalized distance value is the weighted centroid of the absolute value of the physical residual along the spatial coordinates: ; In the formula ( To prevent division by zero, a dimensionless constant is used. When the cable is in normal condition, the residual is uniformly distributed at each node and has a small amplitude. The residual tends to be near the midpoint of the cable; when a fault occurs, the residual concentrates in the fault area, and the centroid naturally shifts to the vicinity of the actual fault location.

[0087] To mitigate the centroid shift problem caused by boundary truncation in residual distribution when a fault occurs near the cable endpoint (i.e., when the fault is located at the cable endpoint, the residual mass is only distributed on one side, causing the centroid to shift towards the midpoint), a boundary compensation factor is introduced. ( This is a boundary compensation factor, used to adjust the weight of the physical residual consistency loss near the endpoints (dimensionless). When the value is close to 0 or 1, it indicates that the fault may be located near the endpoint. In this case, the uncertainty of the physical residual consistency constraint increases, and the weight of this loss should be appropriately reduced. The boundary compensation factor is defined as: ; In the formula ( The width parameter of the boundary compensation factor controls the rate of weight reduction in the central region (dimensionless). When close to 0.5 (midpoint of cable) Close to 1, near the endpoint Approaching This reduces the regularization intensity in the boundary region.

[0088] The physical residual consistency regularization loss is defined as: ; The physical meaning of this regularization term is that when a cable fault occurs, the model output must not only match the manually labeled fault location, but also maintain a similar centroid to the distribution of the physical model residuals. Furthermore, this constraint is stronger in the middle region of the cable and appropriately relaxed in the endpoint region to tolerate boundary effects. This physical constraint allows the model to learn physically interpretable localization capabilities even with a limited number of fault label samples, reducing the reliance of purely data-driven methods on a large number of precise localization labels.

[0089] 5. Synthesis of total loss function The total loss function for a single sample is a weighted sum of the above three terms: ; In the formula ( The equilibrium coefficient for conditional location loss (dimensionless) and ( The balance coefficient (dimensionless) for the physical residual consistency regularization loss can be determined on the validation set via grid search. and .

[0090] For a batch of ( Let be the batch size, i.e., the number of samples contained in each batch (dimensionless). The batch loss is: .

[0091] In step 5, the trained spatiotemporal deep fusion network model is used to perform forward inference on the real-time acquired signals, outputting the fault category and fault location, specifically: This step details the entire process from data preparation to inference deployment, ensuring training stability, convergence, and tight coupling with the aforementioned modules. Specifically, it includes: Step 501: Training Data Organization The training data comes from historical cable operation records and must cover normal operating conditions and various known fault conditions. Each data record contains a raw multi-source sensor signal sequence (unprocessed digital signal streams of vibration, temperature, and sheath current). The training data is stored in its raw format, and preprocessing is dynamically performed during training to ensure that the data processing flow during the training phase is completely consistent with that during the online inference phase.

[0092] Let the total length of the original multi-channel signal sequence be... ( (This represents the total number of sampling points in the original multi-channel signal sequence, dimensionless). During training, for each potential training center time... It is necessary to obtain the original signal within its neighborhood: Bandpass filtering and moving average require and before The signal at a given moment; Cross-correlation calculation requires a causal window Normalized data within the range; Temporal feature encoding requires prior The spatial global feature vector calculated at each time step.

[0093] Therefore, the initialization time is defined. ( The time index is used to establish a valid starting time for training. Prior to this, it is only used to build the context and does not participate in loss calculation (it is dimensionless). During training, the time index is used. Each moment is used as the training center moment. For the initial... At any given moment, the time interval is used only for constructing the temporal context and initializing the filter state, and does not participate in the loss calculation.

[0094] Step 502: Network parameter initialization Learnable parameters of the spatial feature extraction module and Initialize using Xavier uniform distribution; The parameters of the gated recurrent units in the temporal feature encoding module are orthogonally initialized to enhance gradient propagation; In the physical information-guided fusion module and The distribution was initialized using a Gaussian distribution with a mean of zero and a standard deviation of 0.01. The fully connected layers of the classification head and positioning head are initialized using Xavier uniform distribution, and the bias is initialized to zero.

[0095] Step 503: Optimizer and Learning Rate Scheduling The AdamW optimizer was chosen for training, and its weight decay mechanism helps to suppress the overfitting risk caused by overparameterized graph convolutional layers. The initial learning rate was set to... ( (The initial learning rate is dimensionless), and the weight decay coefficient is set to... (Weight decay coefficient, dimensionless). The total number of training epochs is set to... ( (Maximum number of training rounds, dimensionless).

[0096] Learning rate scheduling employs a cosine annealing plus restart strategy: each ( The cosine annealing restart cycle is measured in training epochs (dimensionless). Cosine annealing is performed once per epoch, with the learning rate starting from [value missing]. Decrease according to the cosine curve to ( (Minimum learning rate, dimensionless), then restart to Continue to the next cycle. This scheduling method helps to escape local optima.

[0097] Step 504: Batch Construction and Training Loop Random sampling for each training batch A set of non-overlapping consecutive time interval samples, each sample corresponding to a central time. and before and after that moment The original signal context at each time step. For each sample, the forward propagation process is executed sequentially: The signal segments from each channel required for the current sample are extracted from the original signal buffer, and vibration bandpass filtering, current bandpass filtering, temperature moving average denoising, max-min normalization, and feature concatenation are performed sequentially to obtain the multi-source sample vector. .

[0098] Graph construction: The spatial distance attenuation factor is calculated using the offline calibrated node spatial coordinates, and a causal window is used. The normalized data within the matrix is ​​used to calculate the real-time signal cross-correlation factor. The two factors are multiplied and then sparsified to obtain the adaptive adjacency weight matrix. Simultaneously construct the graph signal matrix .

[0099] Network Forward: Physical residual calculation module: Read , and , Calculate the physical residual vector ; splicing enhances node features ; Spatial feature extraction module: with and As input, the vector is processed through two layers of physically augmented graph convolution and global pooling, and the output is a spatial global feature vector. ; Temporal feature encoding module: Take the first part Cached at this moment Sequence input gated loop unit, output ; Physically Guided Fusion Module: Computation and ,through Generate gate vectors with Sigmoid Modulation obtained ; Dual-head output: Classification head output Positioning head output .

[0100] Loss calculation: Reading the real labels and Calculate the classification loss sequentially. Conditional location loss Consistency regularization loss with physical residuals The weighted summation yields the total loss for a single sample. .

[0101] All within each batch After forward propagation of each sample is completed, calculate the batch average loss. The AdamW optimizer is based on Backpropagation and gradient updates are performed on all learnable parameters in the network. A complete training cycle (forward + backward + parameter update) is called a training step.

[0102] Step 505: Verification and Early Stop After each training epoch (i.e., after traversing all available samples in the training set), perform the complete forward propagation process on the independent validation set (same as step 504, but without backpropagation and parameter updates), and calculate the average total loss and the average absolute error of fault location on the validation set. If continuous ( If the validation loss does not decrease after a certain number of rounds (dimensionless), early stopping is triggered, and the model parameters that result in the lowest validation loss are restored as the final training result.

[0103] Step 506: Online Inference Deployment The trained model is exported in a unified format (such as TorchScript static graph or ONNX open neural network exchange format) and deployed on the edge computing node for cable monitoring.

[0104] During online inference, real-time data streams enter the system on a time-series basis. The edge computing nodes maintain the following cache structure: Signal ring buffer: length is It stores the original three-channel signals from recent moments for use in filtering and cross-correlation calculations; Spatial feature circular buffer: length is Store the most recent Spatial global feature vector at time n For use in timing coding; Hidden state of gated loop unit: two units with lengths of The vectors store the hidden states of the two GRU layers respectively. and It is updated after each frame of inference to ensure that timing information is continuously transmitted across frames.

[0105] Whenever a new sampling time arrive: Three-channel raw digital signal Write to the signal ring buffer; Perform online preprocessing: extract the required historical signal from the buffer, perform filtering, denoising, and normalization to obtain the desired signal. , and ; Execution graph construction: Extracting causal windows from the buffer Calculate the adaptive adjacency weight matrix using normalized data within the matrix. And graph signal matrix ; Perform forward propagation: Physical residual calculation obtained ; structure The result is obtained through two layers of graph convolution and pooling. Write to the spatial feature buffer; In the buffer The sequence input gated recurrent unit uses the hidden state saved in the previous frame as the initial state to calculate the output. At the same time, update the hidden state cache; Gated fusion and dual-head output are obtained and ; Fault diagnosis: If ( (the category index corresponding to the highest probability) and ( If the fault alarm confidence threshold (dimensionless, typical value 0.85) is set, a fault alarm will be triggered, and the normalized location value will be used. The actual cable mileage is mapped back to the fault location and pushed to the operation and maintenance personnel.

[0106] Embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0107] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0108] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0109] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0110] Contents not described in detail in this specification are prior art known to those skilled in the art. It is hereby indicated that the above description is intended to help those skilled in the art understand this invention, but does not limit the scope of protection of this invention. Any equivalent substitutions, modifications, improvements, or simplifications of the above descriptions that do not depart from the essential content of this invention fall within the scope of protection of this invention.

Claims

1. A method for online monitoring of high-voltage cables in a power grid, characterized in that, include: Step 1: Collect distributed vibration signals, distributed temperature signals, and sheath grounding current signals along the cable line and preprocess the collected signals to obtain normalized vibration signals, normalized temperature signals, and normalized sheath grounding current signals. Step 2: Construct an adaptive spatiotemporal topology graph; Step 3: Construct a physical information-guided spatiotemporal deep fusion network model, wherein the forward propagation process of the spatiotemporal deep fusion network model is as follows: By using differential prediction of cable thermal conduction and prediction of sheath current, the physical residuals of each fiber optic sensing node and current monitoring node are calculated to form a physical residual vector. The graph signal matrix and the physical residual vector are concatenated along the feature dimension to form an enhanced node feature matrix. Multi-layer physical enhancement graph convolution is performed on the adaptive spatiotemporal topology graph, and global pooling is performed on the graph convolution output to obtain the spatial global feature vector. The time series composed of spatial global feature vectors from multiple consecutive time points is input into a gated recurrent unit to obtain the time series feature vector at the current time point. A gating vector is generated based on the global statistics of the physical residual vector. The gating vector is then used to modulate the temporal feature vector dimension by dimension to obtain the physical guidance enhancement feature. Based on physical guidance enhancement features, the fault category probability is output through the classification head, and the normalized distance of the fault location is output through the positioning head. Step 4: Train the spatiotemporal deep fusion network model based on the training samples; Step 5: Use the trained spatiotemporal deep fusion network model to perform forward inference on the real-time acquired signals and output the fault category and fault location.

2. The online monitoring method for high-voltage power cables in a power grid according to claim 1, characterized in that, In step 2, an adaptive spatiotemporal topology graph is constructed, specifically as follows: Using multiple fiber optic sensing nodes and one current monitoring node distributed along the cable path as graph nodes, and using normalized vibration signals and normalized temperature signals to assign features to the corresponding nodes, a graph signal matrix is ​​formed. The spatial distance attenuation factor is calculated based on the spatial distance between nodes, and the real-time signal cross-correlation factor is calculated based on the normalized vibration signal or the normalized sheath grounding current signal within the causal time window. The edge weights are obtained by multiplying the spatial distance attenuation factor with the real-time signal cross-correlation factor. After sparsification, an adaptive adjacency weight matrix is ​​formed, which together with the graph signal matrix constitutes the adaptive spatiotemporal topology graph at the current moment.

3. The online monitoring method for high-voltage power cables according to claim 2, characterized in that, The calculation method for the real-time signal cross-correlation factor is as follows: When both nodes are fiber optic sensing nodes, the normalized vibration signal within the causal time window ending at the current time is acquired, and the absolute value of the Pearson correlation coefficient between the vibration signals of the two nodes is calculated as the cross-correlation factor. When one of the nodes is a current monitoring node, the normalized vibration signal and the normalized sheath grounding current signal within the causal time window are obtained, and the maximum absolute value of the time delay cross-correlation of the scanning with a preset time delay range is calculated as the cross-correlation factor. The time delay range is determined based on the total cable length, vibration propagation speed, and sampling interval.

4. The online monitoring method for high-voltage power cables in a power grid according to claim 3, characterized in that, In step 3, a gating vector is generated based on the global statistics of the physical residual vector, specifically as follows: The spatial mean and spatial maximum of the absolute values ​​of each element in the physical residual vector are calculated. The two are then concatenated and passed through a fully connected layer and a Sigmoid activation function to obtain the gated vector.

5. The online monitoring method for high-voltage power cables according to claim 4, characterized in that, In step 4, the spatiotemporal deep fusion network model is trained based on the training samples, specifically as follows: Training samples are constructed based on historical data. Based on the training samples, a spatiotemporal deep fusion network model is trained using a composite loss function that includes classification loss, conditional localization loss for fault samples, and physical residual consistency regularization loss.

6. The online monitoring method for high-voltage power cables according to claim 5, characterized in that, The physical residual consistency regularization loss characterizes the consistency between the predicted fault location of the fault sample and the centroid of the physical residual distribution. The physical residual consistency regularization loss is calculated on the fault sample, and the specific calculation process is as follows: The pseudo-location tag is calculated based on the absolute value of the physical residual of the fiber optic sensing node along the weighted centroid of the normalized distance coordinate. Then, the squared loss between the pseudo-location tag and the predicted fault location is calculated and multiplied by a boundary compensation factor that adaptively attenuates based on how close the pseudo-location tag is to the cable end.

7. The online monitoring method for high-voltage power cables in a power grid according to claim 6, characterized in that, Step 5 also includes setting an alarm if the probability of a fault category exceeds a preset threshold.

8. An online monitoring device for high-voltage power cables in a power grid, applied to the online monitoring method for high-voltage power cables in a power grid as described in any one of claims 1-7, characterized in that, include: The multi-source sensor acquisition front-end, edge computing gateway, and host computer monitoring platform are connected in sequence, wherein: The multi-source sensing acquisition front end includes a distributed optical fiber sensing host and a sheath current transformer; the distributed optical fiber sensing host is connected to a single-mode sensing optical fiber laid along the cable to synchronously acquire distributed vibration signals and distributed temperature signals along the line; the sheath current transformer is fitted onto the cable sheath grounding wire to acquire the sheath grounding current signal. The edge computing gateway, as an industrial-grade embedded processing platform, is equipped with: The data preprocessing module is used to perform online filtering and normalization on the acquired signals; The graph construction and update module is used to generate an adaptive adjacency weight matrix and a graph signal matrix online based on the normalized vibration signal, normalized temperature signal and normalized sheath grounding current signal; The physical residual generation module is used to calculate the physical residual vector based on thermal conduction differential and current prediction. The physical augmentation spatial feature extraction module is used to perform physical kernel-modulated graph convolution and global pooling on the adaptive adjacency weight matrix on the augmentation node features formed by concatenating the graph signal matrix and the physical residual vector, and output a spatial global feature vector. The temporal coding module uses a gated recurrent unit to encode the spatial global feature vector sequence to obtain temporal feature vectors; The physical guidance fusion and diagnostic output module is used to generate a gated vector based on the global statistics of the physical residual vector, modulate the temporal feature vector, and output the fault category probability and the fault location normalized distance. The model management and storage unit is used to store the parameters and configuration information of the trained network model. The host computer monitoring platform is used to receive fault alarm information, fault type and fault location and display them visually.

Images