High-voltage cable partial discharge on-line monitoring method and device
By embedding distributed optical fiber sensing units into the sheath layer of high-voltage cables and combining wavelet transform and 1D-CNN models, the problem of insufficient feature extraction in traditional monitoring methods is solved, enabling accurate identification and trend prediction of partial discharge in high-voltage cables, and improving the accuracy and reliability of monitoring.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- WUHAN BILLION TECH DEV CO LTD
- Filing Date
- 2026-01-22
- Publication Date
- 2026-05-05
AI Technical Summary
Traditional methods for monitoring partial discharge in high-voltage cables cannot effectively extract abstract feature combinations that are highly correlated with the discharge type from the signal, and it is difficult to establish complex nonlinear classification boundaries, resulting in inaccurate monitoring results.
A distributed optical fiber sensing unit is embedded in the high-voltage cable sheath layer. By combining wavelet transform algorithm and one-dimensional convolutional neural network (1D-CNN) model, a partial discharge classification model is constructed through wavelet denoising and feature extraction. The discharge trend is then predicted using spatial clustering and autoregressive integral moving average model.
It achieves deep and automated pattern recognition of partial discharge characteristics, improves the ability to distinguish discharge patterns and the accuracy of trend prediction, and provides high-confidence discharge type identification and risk assessment.
Smart Images
Figure CN121978476A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of online monitoring engineering for high-voltage cables, specifically to a method and device for online monitoring of partial discharge in high-voltage cables. Background Technology
[0002] High-voltage cables are key equipment in urban power grids and long-distance power transmission, and the reliability of their insulation condition directly affects the safe and stable operation of the power supply system. Partial discharge is an important phenomenon characterizing cable insulation degradation, and its activity characteristics are closely related to the type and severity of insulation defects. Therefore, effective monitoring of partial discharge in cables is an important means of achieving condition-based maintenance and preventing faults.
[0003] Traditional methods for monitoring partial discharge in high-voltage cables employ a high-frequency current transformer coupling method. This method involves installing a high-frequency current transformer at the cable grounding wire or joint to couple the high-frequency current signal generated by partial discharge, and then analyzing its amplitude, frequency, and other characteristics using the pulse current method to determine the severity of the discharge. However, this method has inherent bottlenecks in feature extraction and pattern recognition. It lacks the ability to automatically learn the deep features of the original signal, cannot effectively extract abstract feature combinations highly correlated with the discharge type, and struggles to establish complex nonlinear classification boundaries, leading to inaccurate online monitoring results for partial discharge in high-voltage cables. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention provides an online monitoring method for partial discharge in high-voltage cables, solving the problem of inaccurate online monitoring results for partial discharge in high-voltage cables.
[0005] To achieve the above objectives, the present invention provides the following technical solution: an online monitoring method for partial discharge in high-voltage cables, comprising the following steps: Step S1: Embed a distributed optical fiber sensing unit in the high-pressure cable sheath layer, and convert the physical field change data generated when a partial discharge event occurs into an initial partial discharge optical signal sequence through the distributed optical fiber sensing unit. Step S2: Denoise the initial partial discharge optical signal sequence using wavelet transform algorithm to obtain a denoised partial discharge optical signal sequence, and extract features from the denoised partial discharge optical signal sequence to obtain a partial discharge feature vector; Step S3: Based on the one-dimensional convolutional neural network 1D-CNN model, construct a partial discharge classification model, input the partial discharge feature vector into the partial discharge classification model, and output the partial discharge classification result and the spatiotemporal information of the partial discharge event; Step S4: Based on the partial discharge classification results and the spatiotemporal information of partial discharge events, the DBSCAN spatial clustering algorithm is used to perform spatiotemporal clustering of partial discharge events to generate the spatiotemporal evolution path of partial discharge. Step S5: Based on the autoregressive integral moving average model, construct a discharge development trend prediction model, input the characteristic parameters of the spatiotemporal evolution path into the discharge development trend prediction model, output the discharge development trend prediction result, calculate the partial discharge risk value based on the discharge development trend prediction result, and perform graded early warning for partial discharge events based on the partial discharge risk value.
[0006] Preferably, the step of embedding a distributed optical fiber sensing unit in the high-pressure cable sheath layer includes: In the manufacturing process of the sheath layer of high-voltage cables, distributed optical fiber sensing units are directly embedded, and close physical contact and signal coupling with the cable's metal shielding layer are achieved through conductive adhesive layers. The distributed optical fiber sensing unit adopts a four-layer composite structure design from the outside to the inside: the outermost layer is a nickel-plated fiber braided shielding layer; the middle layer is a polyimide electro-erosion resistant coating; a thermally conductive silicone rubber layer is provided between the outermost and middle layers; and the innermost layer is a standard single-mode optical fiber core.
[0007] Preferably, converting the physical field change data generated during a partial discharge event into an initial partial discharge optical signal sequence includes: A nanosecond-level laser pulse is emitted into an optical fiber, and its backscattered signal is detected. The phase change Δφ and intensity change Δφ of the Rayleigh scattered light are then extracted. Brillouin radio frequency shift change ; Each emitted laser pulse is marked with an absolute timestamp using a high-precision clock. When in fiber optic location When a scattering event is detected, record the absolute timestamp of the received scattering signal. Light from emission to position The round-trip time to the receiving end is ; Finally, the aforementioned Rayleigh scattered light phase change Δφ and Rayleigh scattered light intensity change Δφ are calculated. Brillouin radio frequency shift change By associating the signal with the corresponding spatiotemporal coordinates (t, x), an initial partial discharge optical signal sequence with spatiotemporal coordinates is constructed. .
[0008] Preferably, the initial partial discharge optical signal sequence is denoised using a wavelet transform algorithm to obtain a denoised partial discharge optical signal sequence, including: First, the initial partial discharge signal sequence... By spatial coordinates Group, each Corresponding to a three-dimensional time series Wavelet decomposition and denoising are performed separately on each time series to obtain wavelet coefficients. ; Next, thresholding is performed on the detail coefficients of each layer obtained from the decomposition. An adaptive threshold quantization is performed using an unbiased risk estimation threshold method. The threshold of this method... Determined by the following formula:
[0009] in, The noise standard deviation is determined by the first level detail coefficient. The median estimate, The signal length; Subsequently, a soft thresholding function is applied to process the coefficients:
[0010] in, After soft thresholding, the first Layer reconstruction detail coefficients; After wavelet decomposition, the th The original detail factor of the layer; It is a symbolic function; Finally, wavelet reconstruction is performed to extract the detail coefficients after thresholding. Approximation coefficients retained Perform inverse wavelet transform to reconstruct the denoised partial discharge optical signal sequence. .
[0011] Preferably, feature extraction of the noise-reduced partial discharge optical signal sequence to obtain a partial discharge feature vector includes: First, extract the peak intensity of Rayleigh scattering. This feature comes directly from the noise reduction. , equal to the measured intensity during discharge Reference strength when there is no discharge The difference, reference strength Calibrated during initialization; peak value It is the pulse duration. Inside, The maximum value; Secondly, extract the phase change peak value. , directly from The extraction method involves detecting the peak offset of the phase pulse waveform. ; Furthermore, extract the pulse duration. ; Then, extract the Rayleigh scattering pulse energy. Also stemming from noise reduction The extraction method involves calculating the integral energy of the pulse waveform over its duration.
[0012] in, It is the change in the intensity of the noise-reduced Rayleigh scattering light; Finally, the Brillouin frequency shift offset is extracted. The direct source is the change in frequency shift of the noise-reduced Brillouin scattering light. During extraction, the peak value of the frequency-shifted pulse synchronized with the partial discharge event is captured, and the calculation formula is as follows:
[0013] in, It is the peak value of the Brillouin frequency shift. The frequency shift change of the noise-reduced Brillouin scattering light; Finally, a corresponding partial discharge feature vector is generated for each partial discharge event. .
[0014] Preferably, step S3 includes: First, the input data is preprocessed, and the five-dimensional feature vector output from step S2 is reconstructed into a tensor format suitable for network processing. The data is converted into a preset data structure and then standardized to obtain preprocessed input data. ; Subsequently, we enter the core feature extraction stage, where deep feature learning is performed sequentially through two residual modules; The first residual module processing flow, input features First, local feature patterns are extracted through convolution operations:
[0015] in, It is the feature map output by the first convolutional layer; These are the convolution kernel weights, responsible for capturing local correlations between features; This is a bias term used to adjust the output distribution; It is the convolution operator; Next, batch normalization is performed on the convolutional output to stabilize the training process; Finally, feature fusion is achieved through residual connections:
[0016] in, This is the final output of the first residual module; A composite function representing convolution and batch normalization; ( () is a linear rectification activation function; The original input to the residual module; The second residual module's processing flow uses the output of the first module as the input of the second module, achieving deep feature abstraction:
[0017] in, It is the input for the second module; Finally, the classification decision is made through a fully connected layer and a softmax function, and the output of the second residual module is used. After flattening, the data is fed into a fully connected layer to obtain the raw scores for each category; Then, the original scores are converted into a probability distribution using the Softmax function; The category corresponding to the highest probability is used as the prediction result. This probability value also serves as the confidence level. When confidence level When the result is determined to be of high confidence, it proceeds directly to the next step. process.
[0018] Preferably, the DBSCAN spatial clustering algorithm is used to perform spatiotemporal clustering of partial discharge events, generating spatiotemporal evolution paths of partial discharges, including: Using the three-dimensional coordinates of the discharge event An improved spatial clustering algorithm, DBSCAN, is used for spatiotemporal clustering:
[0019] in, Indicates the first Each discharge cluster has a spatial neighborhood radius of 2m to ensure geographical relevance, and a time window of 10s to ensure event continuity. It is the k-th point in the cluster. These are the times when the signal arrives at points i and j, respectively; Represents the spatial coordinates of two partial discharge events; and Indicates the type label of the discharge event; Finally, based on the clustering results, three key characteristic parameters of each discharge cluster were calculated, including spatial distribution density. Average discharge frequency optical signal disturbance amplitude ; Next, by analyzing the spatiotemporal distribution patterns of events within the cluster, the propagation direction vector is calculated. and evolutionary acceleration ; Ultimately, this information from all discharge clusters together constitutes the complete spatiotemporal evolution path G of partial discharge.
[0020] Preferably, the three-dimensional coordinates of the discharge event The determination includes: For each discharge event, its three-dimensional coordinates have been determined by step S1, and its spatial location is determined by fiber optic coordinates. Provide elevation information. The data is derived from cable laying data; the planar coordinates are determined by combining the cable route diagram, and finally the three-dimensional coordinates of the discharge event are obtained. .
[0021] Preferably, a discharge development trend prediction model is constructed based on an autoregressive integral moving average model. The characteristic parameters of the spatiotemporal evolution path are input into the discharge development trend prediction model, and the output discharge development trend prediction results include: First, classify by discharge type Grouping the discharge cluster set G into subsets of clusters of the same type yields clusters of the same type. Divide the timeline into equidistant windows. For each window Aggregate all cluster features it covers:
[0022] in, For window The set of covered clusters, ; It is the average discharge density in the i-th time window; The average discharge frequency of the i-th time window; the average discharge amplitude of the i-th time window. ; Arrange all windows in chronological order. To form a multivariate time series ; The model is improved using Support Vector Regression (SVR), which combines the linear fitting ability of ARIMA with the nonlinear mapping ability of SVR:
[0023] in, The discharge development trend prediction model at a certain time point Predicted values; ARIMA ( ) is an ARIMA component; SVR ( , ) is an SVR component; For ARIMA components at time The predicted residuals; For a moment The feature gradient vector; The discharge development trend prediction model training adopts a two-stage optimization strategy. First, the optimal parameter combination (p, d, q) of ARIMA is determined by the AIC criterion. Then, grid search is used to optimize the penalty parameters of SVR. and kernel parameters ; After training is completed, predict the future simultaneously. Multidimensional feature values at each time point :
[0024] in, It is a sequence of the spatial distribution density of discharge events over a predicted future period of time. It is a sequence of the average discharge frequency of discharge events over a predicted future period of time. It is a sequence of the amplitude of optical signal disturbance of the predicted discharge event over a period of time. It predicts the step size.
[0025] A high-voltage cable partial discharge online monitoring device includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method.
[0026] This invention provides an online monitoring method for partial discharge in high-voltage cables, involving machine learning and deep learning technologies, which has the following beneficial effects: (1) By introducing a one-dimensional convolutional neural network to construct a partial discharge classification model, deep and automated pattern recognition of discharge feature vectors is realized. Compared with the traditional method that relies on manual extraction of thresholds or simple classifiers, it can automatically learn and extract deeper and more abstract feature representations from the original, sequential feature data, and has a stronger ability to distinguish complex and nonlinear discharge patterns.
[0027] (2) By introducing a residual network structure into 1D-CNN, the problem of gradient vanishing or gradient exploding in the training of deep neural networks is effectively solved. The shortcut connections in the residual structure allow gradients to propagate directly in the backpropagation direction, making it possible to build deeper networks without worrying about training degradation.
[0028] (3) By introducing support vector regression and ARIMA model, a hybrid prediction model is formed. ARIMA model is mainly good at handling linear relationships, while the development process of partial discharge often contains complex nonlinear dynamic characteristics. SVR can efficiently capture nonlinear patterns in the data that ARIMA cannot describe through kernel function techniques. This hybrid model makes full use of the respective advantages of ARIMA in linear trend prediction and SVR in nonlinear fitting. Compared with the single model, it significantly improves the overall accuracy and robustness of the prediction of discharge development trend. Attached Figure Description
[0029] Figure 1 This is a flowchart of an online monitoring method for partial discharge in high-voltage cables proposed in this invention.
[0030] Figure 2 This invention provides a hierarchical diagram of the partial discharge feature vector obtained from an online monitoring method for partial discharge in high-voltage cables.
[0031] Figure 3 This is a hierarchical diagram of graded early warning obtained in the online monitoring method for partial discharge of high-voltage cables proposed in this invention. Detailed Implementation
[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0033] Please see Figures 1-3 This invention provides a technical solution: an online monitoring method for partial discharge in high-voltage cables. Specifically, the following online monitoring method for partial discharge in high-voltage cables is provided; please refer to [link / reference]. Figure 1 The method includes the following steps: Step S1: Embed a distributed optical fiber sensing unit in the high-pressure cable sheath layer. Through the distributed optical fiber sensing unit, the physical field change data generated when a partial discharge event occurs is converted into an initial partial discharge optical signal sequence.
[0034] This step aims to construct a distributed optical fiber sensing unit integrated with the cable body to synchronously acquire multi-physics field change signals caused by partial discharge and output raw data with spatiotemporal coordinates. The specific implementation is as follows: First, during the manufacturing process of the high-voltage cable sheath, distributed optical fiber sensing units are directly embedded, and a conductive adhesive layer is used to achieve close physical contact and signal coupling with the cable's metal shielding layer. This ensures that the sensing unit forms a tight physical connection with the cable's main insulation system, laying a structural foundation for accurately sensing partial discharge activities.
[0035] This distributed optical fiber sensing unit adopts a four-layer composite structure design from the outside to the inside: the outermost layer is a nickel-plated fiber braided shielding layer, which uses its low surface resistivity to capture electromagnetic disturbance signals; the middle layer is a polyimide electro-erosion resistant coating, which is used to protect internal components from electrochemical corrosion; a thermally conductive silicone rubber layer is provided between the outermost and middle layers, which is responsible for efficiently conducting the heat generated by partial discharge; the innermost layer is a standard single-mode optical fiber core, which directly senses the changes in optical wave parameters caused by changes in external electric, thermal and force fields.
[0036] Based on the above structure, when partial discharge occurs inside the high-voltage cable, the distributed optical fiber sensing unit converts the physical field change data into a signal through the following physical mechanism: The electromagnetic waves generated by the discharge create an induced electric field in the shielding layer, which causes micro-strain in the optical fiber through electrostriction, thereby causing a phase shift in the transmitted light wave. The heat from the discharge point is transferred to the optical fiber through the thermally conductive layer, changing the fiber density and refractive index, causing a shift in the characteristic frequency of the Brillouin scattered light. The stress wave generated by the discharge excites instantaneous micro-bending deformation of the optical fiber, modulating the intensity of Rayleigh scattered light. ; To acquire these signals and assign them spatial properties, phase-sensitive optical time-domain reflectometry (Φ-OTDR) and Brillouin optical time-domain reflectometry (BOTDR) techniques are used for synchronous demodulation. A high-precision synchronization clock (synchronization accuracy ≤1ns) triggers the laser pulse emission of both techniques, ensuring that the Rayleigh scattering signal and the Brillouin scattering signal are aligned in the time dimension, thus achieving synchronous acquisition and demodulation of multiple parameters. Nanosecond-level laser pulses are emitted into an optical fiber, and the backscattered signal is detected. Phase-sensitive optical time-domain reflectometry (Φ-OTDR) is used to extract the phase change Δφ and intensity change Δφ of the Rayleigh scattered light. BOTDR technology is used to extract the Brillouin radio frequency shift variation. .
[0037] Each emitted laser pulse is marked with an absolute timestamp using a high-precision clock. (UTC time); when in fiber optic location When a scattering event is detected, record the absolute timestamp of the received scattering signal. Light from emission to position The round-trip time to the receiving end is .
[0038] According to the principle of optical time-domain reflectometry, the speed of light in an optical fiber for:
[0039] in, The speed of light in a vacuum (approximately) ), The effective group refractive index of the optical fiber (typical value 1.468). Since the round trip distance of light is Therefore, according to the distance formula The fiber optic coordinates of the partial discharge event are obtained by refining the formula as follows:
[0040] in, These are the fiber optic coordinates of partial discharge events, from which the round-trip time can be used to determine their location. Precisely locate the event to fiber optic coordinates (Positioning accuracy can reach ±1 meter).
[0041] Finally, the aforementioned optical phase change, Brillouin scattering frequency shift change, and Rayleigh scattering intensity change are correlated with the corresponding spatiotemporal coordinates (t, x) to construct an initial partial discharge optical signal sequence with spatiotemporal coordinates. The mathematical formula is as follows:
[0042] in, It is the initial partial discharge optical signal sequence, representing a complete set of optical parameter changes measured at time t and position x; It is the phase change of the Rayleigh scattered light; It is the change in Brillouin divergence radio frequency shift; The change in Rayleigh scattered light intensity; t is a UTC timestamp accurate to microseconds, representing the time of partial discharge occurrence; These are the fiber optic coordinates of a partial discharge event, representing the location where the partial discharge occurs.
[0043] It should be noted that, It is a signal sequence (macroscopic sequence), a set concept, serving as the basic data unit for subsequent processing; These are three independent physical "variables" (microscopic parameters) that constitute the signal. They are quantitative descriptions of a specific physical phenomenon (phase, frequency, intensity) deviating from its reference state.
[0044] Step S1 provides the raw data foundation for the entire monitoring system. Its output multiphysics signal sequence (including time, position and optical parameters) is the direct input for subsequent signal analysis (S2) and feature extraction. The introduction of spatiotemporal coordinates lays the foundation for localization and path reconstruction in S4.
[0045] Step S2: Denoise the initial partial discharge optical signal sequence using wavelet transform algorithm to obtain a denoised partial discharge optical signal sequence, and extract features from the denoised partial discharge optical signal sequence to obtain a partial discharge feature vector.
[0046] This step aims to process the initial partial discharge optical signal sequence output in step S1. The noise is denoised, and key features that characterize the nature of partial discharge events are extracted to provide input for subsequent pattern recognition.
[0047] Because the original signal contains strong environmental electromagnetic interference, cable vibration noise, and system thermal noise, it must be effectively filtered out. Wavelet transform is used as the core noise reduction algorithm. Its advantage lies in its ability to analyze the signal simultaneously in the time and frequency domains, making it very suitable for processing non-stationary abrupt signals such as partial discharge. The specific process is as follows: First, the initial partial discharge signal sequence... By spatial coordinates Group, each Corresponding to a three-dimensional time series Wavelet decomposition and denoising are performed separately for each time series to ensure that the spatiotemporal attributes are not lost, and wavelet coefficients are obtained. :
[0048] in, As a scale factor, The translation factor is... These are wavelet basis functions. In practical discrete processing, a 5-level multi-resolution analysis (Mallat algorithm) is performed to decompose the signal into approximate coefficients of different frequency bands. With detail coefficient ,in This represents the number of decomposition layers.
[0049] Next, thresholding is performed on the detail coefficients of each layer obtained from the decomposition. (Mainly includes high-frequency components of noise and partial discharge pulses), an adaptive threshold quantization is performed using an unbiased risk estimation thresholding method. The threshold of this method... Determined by the following formula:
[0050] in, The noise standard deviation is typically determined by the first level detail factor. Median estimate ( ), This is the signal length.
[0051] Subsequently, a soft thresholding function is applied to process the coefficients:
[0052] in, After soft thresholding, the first The reconstruction detail coefficients of the layer (are the noise reduction results, used for subsequent wavelet reconstruction). After wavelet decomposition, the th The original detail coefficients of the layer (containing high-frequency abrupt changes and noise in the partial discharge signal); It is a sign function, its function is to preserve... The polarity (positive / negative) ensures that the phase characteristics of the signal are not lost.
[0053] Finally, wavelet reconstruction is performed to extract the detail coefficients after thresholding. Approximation coefficients retained ( Perform inverse wavelet transform (to obtain the maximum decomposition level) to reconstruct the denoised partial discharge optical signal sequence. = .
[0054] Based on the completed signal denoising, the denoised partial discharge optical signal sequence will be further processed. Key feature extraction was conducted to capture typical patterns of partial discharge events. Specifically, focusing on Rayleigh scattering, phase modulation, Brillouin scattering, and temporal characteristics, the following five categories of core features directly traceable to physical quantities were extracted: First, extract the peak intensity of Rayleigh scattering. This feature comes directly from the noise reduction. , equal to the measured intensity during discharge Reference strength when there is no discharge The difference, reference strength Calibrated during initialization; peak value It is the pulse duration. Inside, The maximum value, i.e. , This is the pulse duration, which directly reflects the maximum modulation intensity of the stress wave on the optical fiber; Secondly, extract the phase change peak value. , directly from The physical essence of the phase change peak is the fiber phase modulation caused by the partial discharge electric field disturbance through the electrostriction effect (the electric field acting on the fiber material to induce micro-strain). This characteristic is extremely sensitive to the initial electric field strength of the discharge. The extraction method involves detecting the peak offset of the phase pulse waveform. This characteristic is extremely sensitive to the initial electric field strength of the discharge.
[0055] Furthermore, extract the pulse duration. From the signal with the highest signal-to-noise ratio after noise reduction (usually 100%) Extracted from the pulse waveform, specifically by calculating the peak amplitude of the pulse leading edge. Towards the peak value The time interval that has elapsed.
[0056] Then, extract the Rayleigh scattering pulse energy. Also stemming from noise reduction The total mechanical energy released during the duration of a partial discharge event is characterized by calculating the integral energy of the pulse waveform over its duration.
[0057] in, It is the change in the intensity of Rayleigh scattering light after noise reduction.
[0058] Finally, the Brillouin frequency shift offset is extracted. The direct source is the change in frequency shift of the noise-reduced Brillouin scattering light. This reflects the instantaneous temperature rise caused by the thermal effect of partial discharge. The peak value of the frequency-shifted pulse synchronized with the partial discharge event is captured during extraction, and the calculation formula is:
[0059] in, It is the peak value of the Brillouin frequency shift. The frequency shift change of the noise-reduced Brillouin scattering light; Finally, a partial discharge feature vector is output for each partial discharge event. :
[0060] This vector integrates the intensity characteristics of Rayleigh scattering, the electric field response of phase modulation, the energy characteristics of Rayleigh scattering, the temperature response of Brillouin scattering, and the temporal characteristics of the event. Each element has a clear physical origin, providing an interpretable and high-quality data foundation for subsequent intelligent recognition.
[0061] Step S3: Based on the one-dimensional convolutional neural network 1D-CNN model, construct a partial discharge classification model, input the partial discharge feature vector into the partial discharge classification model, and output the partial discharge classification result and the spatiotemporal information of the partial discharge event.
[0062] To train a partial discharge classification model, historical data needs to be collected, primarily from three sources: First, typical partial discharge types (such as corona discharge, surface discharge, etc.) are simulated in a controlled laboratory environment, and a benchmark dataset is established by synchronously recording the data with the optical sensing system in this paper using standard detection equipment. Secondly, long-term operational data is collected from operating cables with deployed monitoring systems, covering different voltage levels, cable types, and operating environments. Finally, fault case data confirmed by on-site disassembly were collected, and the correspondence between the data before and after the fault was established. After processing all the raw data in S1 and S2 as described in the process, feature vectors were extracted, and the domain expert team performed type labeling according to strict standards, ultimately constructing a labeled dataset of partial discharge. Partial discharge Partial discharge is the number of samples; partial discharge Partial discharge is labeled as the discharge type. This dataset provides a solid foundation for model training in step S3.
[0063] The core task of this step is to establish an intelligent mapping relationship from feature vectors to discharge types. By performing deep pattern recognition on the feature vectors output by S2, accurate classification of discharge patterns is achieved. A one-dimensional convolutional neural network (1D-ResCNN) model based on residual network enhancement is adopted. This design effectively solves the gradient vanishing problem in deep network training by introducing a shortcut connection mechanism, and significantly improves the ability of the partial discharge classification model to represent complex discharge features.
[0064] First, the input data is preprocessed. The five-dimensional feature vector output from step S2 is reconstructed into a tensor format suitable for network processing. Specifically, the feature vector... Convert to The data structure is then subjected to Z-score standardization:
[0065] in, and These represent the mean and standard deviation of each feature dimension in the historical training dataset, respectively. This standardization process ensures the uniformity of the feature dimensions, laying a numerical foundation for subsequent model training.
[0066] Subsequently, the data enters the core feature extraction stage. Deep feature learning will be performed sequentially through two residual modules.
[0067] Each residual module performs feature transformation through specific mathematical operations. The specific process is as follows: Processing flow of the first residual module: Input features First, local feature patterns are extracted through convolution operations:
[0068] in, It is the feature map output by the first convolutional layer; These are the convolution kernel weights, responsible for capturing local correlations between features; This is a bias term used to adjust the output distribution; It is the convolution operator. Convolution operations use a sliding window approach to perform local perception along the feature dimension, extracting discriminative feature combinations.
[0069] Next, batch normalization is performed on the convolutional output to stabilize the training process.
[0070] in, This is the output after batch normalization. and The mean and variance of the current batch of data. and For learnable scaling and translation parameters, It is a very small constant, the purpose of which is to prevent the denominator from being zero and to ensure computational stability; batch normalization effectively alleviates the internal covariate offset problem, enabling the partial discharge classification model to use a larger learning rate and accelerate convergence.
[0071] Finally, feature fusion is achieved through residual connections:
[0072] in, This is the final output of the first residual module; A composite function representing convolution and batch normalization; ( () is a linear rectification activation function; The original input of the residual module; this shortcut connection structure ensures that gradients can be directly backpropagated, solving the gradient vanishing problem in deep network training and enabling the network to effectively learn residual mappings.
[0073] The second residual module's processing flow: The output of the first module is used as the input of the second module to achieve deep feature abstraction.
[0074] in, It is the input to the second module, repeating the same computational process, but using a separate set of parameters: ; ; ; The second residual module further refines features based on the first module, learns more complex feature combination patterns, and enhances the feature representation capability of the partial discharge classification model.
[0075] During the model training phase, based on the collected historical case library... The parameters are optimized using the cross-entropy loss function. Defined as:
[0076] in This represents the total number of discharge types (such as corona discharge, surface discharge, etc.). For the sample Corresponding category The true label is obtained using one-hot encoding (if the sample belongs to category). ,but ,otherwise ); The total number of samples, For the sample Category The predicted probability; parameters are updated using the Adam optimizer, with an initial learning rate set to... And an early stopping strategy is adopted to prevent overfitting.
[0077] Finally, the classification decision is made through a fully connected layer and a softmax function. The output of the second residual module is then processed. After flattening, the data is fed into a fully connected layer to obtain the raw scores for each category:
[0078] in, This represents the output vector of the fully connected layer. It is a vector of length K (the number of classes), where each element... The original predicted score for category c. A higher score indicates that the model believes the sample is more likely to belong to that category. This represents the weight matrix of the fully connected layer, used to linearly map the flattened features to the class scores; This indicates that the output Output_2 of the second residual module is flattened. This represents the bias vector of the fully connected layer. It is a learnable parameter of length K used to adjust the score baseline for each class.
[0079] Then, the original scores are converted into a probability distribution using the Softmax function:
[0080] in, Indicates the given input When the sample belongs to class c, it is the predicted probability, which is a conditional probability with a value in the range [0,1]. The sum of the probabilities of all classes is 1. e represents the natural constant (approximately equal to 2.71828), which is used in the exponential function to convert the original score into a positive number and amplify the score difference. This represents the raw score of category c (from the output Z of the fully connected layer); This represents the raw score for category k, where k ranges from 1 to K; K is the number of items.
[0081] The category corresponding to the highest probability is used as the prediction result. This probability value also serves as the confidence level. When confidence level When the result is determined to be of high confidence, it proceeds directly to the next step. Process; when When an indeterminate discharge type is identified, automatic data re-acquisition (increasing the sampling frequency in that spatiotemporal region) is triggered, and the data is re-processed. If the confidence level is still below 0.85 after three consecutive re-collections, the suspected fault type will be output and pushed to maintenance personnel for manual review.
[0082] After the partial discharge classification model is identified, the classification results are matched and bound with the original spatiotemporal coordinates in step S1 using timestamps. Each feature vector corresponds to a unique acquisition time window. A one-to-one correspondence is established between the high-precision timestamp and the original optical signal sequence in S1, thereby ensuring that each identified discharge event has complete spatiotemporal attributes and corresponding spatiotemporal coordinate information (position x and time t from S1).
[0083] This step constructs a classification model based on a one-dimensional convolutional neural network (1D-CNN) and a residual network (ResNet). First, the feature vectors from step S2 are preprocessed using standardization. Then, deep feature learning is performed through a residual module (containing convolution, batch normalization, and shortcut connections). Finally, the Softmax function is used to output the discharge type (e.g., corona discharge, surface discharge) and its corresponding confidence score and spatiotemporal coordinates. Model training relies on a historical case library, and parameters are optimized using a cross-entropy loss function. The discharge type identification results provide key label information for the spatiotemporal clustering and evolutionary analysis in step S4.
[0084] Step S4: Based on the partial discharge classification results and the spatiotemporal information of partial discharge events, the DBSCAN spatial clustering algorithm is used to perform spatiotemporal clustering of partial discharge events to generate the spatiotemporal evolution path of partial discharge.
[0085] In this step, the spatial coordinates of the partial discharge event are directly taken from the results calculated and output in step S1 using Φ-OTDR and BOTDR techniques. Specifically, for each discharge event, its three-dimensional coordinates have been determined in step S1: the spatial position is determined by fiber optic coordinates. (through the round-trip time difference of light) and formula The calculated positioning accuracy is... (meters) are given; elevation information The data is derived from cable laying data; the planar coordinates are determined by combining the cable route diagram, and finally the three-dimensional coordinates of the discharge event are obtained. .
[0086] Using the three-dimensional coordinates of the discharge event An improved spatial clustering algorithm, DBSCAN, is used for spatiotemporal clustering:
[0087] in, Indicates the first Each discharge cluster has a spatial neighborhood radius of 2m to ensure geographical relevance, and a time window of 10s to ensure event continuity. It is the k-th point in the cluster. These are the times when the signal arrives at points i and j, respectively; Represents the spatial coordinates of two partial discharge events; and Indicates the type label of the discharge event. = The requirement is that the discharge types of the two events must be the same to ensure that the clustering only merges discharge activities of the same type and avoids confusion of dissimilar events; the clustering output organizes discrete discharge events into a set of discharge clusters with spatiotemporal correlation.
[0088] Finally, based on the clustering results, the three key feature parameters of each discharge cluster are calculated as follows: For spatial distribution density , ,in The length of the discharge cluster along the cable axis (in meters) is the projected length along the axis by linearly fitting the three-dimensional coordinates of all discharge events within the cluster. The density index represents the total number of discharge events within the cluster and directly reflects the spatial concentration of discharge activity. For the average discharge frequency ,in The time span of events within the cluster; For the amplitude of optical signal disturbance , ; Next, by analyzing the spatiotemporal distribution patterns of events within the cluster, the propagation direction vector and evolutionary acceleration are calculated: The formula for calculating the propagation direction vector is:
[0089] in, Indicates the first The propagation direction vector of each discharge cluster is a vector that has both magnitude (representing the propagation speed) and direction (representing the propagation direction along the cable). This indicates the change in the spatial position of the discharge cluster; This represents the time interval corresponding to the aforementioned spatial position changes; it is obtained by linearly fitting the sequence of spatial position changes of events within the cluster over time, describing the movement trajectory of the centroid or leading edge of the discharge cluster. For the spatial position change of the discharge cluster The calculation formula is:
[0090] in, , The centroid coordinates at the start and end times of the discharge cluster are respectively obtained by least-squares linear fitting of the spatiotemporal coordinates of events within the cluster. The formula for calculating evolutionary acceleration is: ; ; in, Indicates the first The evolution acceleration of each discharge cluster characterizes the trend of its propagation speed; This represents the change in the propagation speed of the discharge cluster. , The propagation velocities at the start and end times of the discharge cluster are respectively obtained by linear fitting of the velocity sequence; A positive value indicates accelerated propagation, while a negative value indicates decelerated propagation. Ultimately, this information from all discharge clusters collectively constitutes the complete spatiotemporal evolution path G of partial discharge:
[0091] This spatiotemporal evolution path fully describes the spatiotemporal distribution characteristics of discharge activity, including spatial distribution density. The average discharge frequency reflects the spatial concentration of the discharge. Characterizing activity over time, The amplitude of the optical signal disturbance reflects the discharge intensity, and M represents the number of discharge clusters. These three characteristics will serve as the core input parameters for S5 risk assessment, providing quantitative evidence for risk evaluation from the three dimensions of spatial distribution, temporal frequency, and intensity, respectively.
[0092] This step combines the discharge type information and spatiotemporal coordinates of S3 with the DBSCAN clustering algorithm to group spatiotemporally adjacent events. Subsequently, for each cluster, evolution parameters such as spatial distribution density, average discharge frequency, and optical signal perturbation amplitude are calculated, providing the most critical input data—the spatiotemporal evolution path of partial discharge—for the next step of trend prediction and risk assessment.
[0093] Step S5: Based on the autoregressive integral moving average model, construct a discharge development trend prediction model, input the characteristic parameters of the spatiotemporal evolution path into the discharge development trend prediction model, output the discharge development trend prediction result, calculate the partial discharge risk value based on the discharge development trend prediction result, and perform graded early warning for partial discharge events based on the partial discharge risk value.
[0094] This step involves extracting the feature parameters of the spatiotemporal evolution path of the partial discharge provided by S4; and constructing a discharge development trend prediction model based on the ARIMA model to achieve accurate prediction of the discharge development trend, as follows: This step, based on the spatiotemporal evolution path G of partial discharge provided by S4, constructs a multivariate time series dataset and then uses the ARIMA-SVR hybrid model to achieve trend prediction. The specific process is as follows: First, classify by discharge type Grouping the discharge cluster set G into subsets of clusters of the same type yields clusters of the same type. Set time window Hours, dividing the timeline into equidistant windows For each window Aggregate all cluster features it covers:
[0095] in, For window The set of covered clusters, ; It is the average discharge density in the i-th time window; The average discharge frequency of the i-th time window; the average discharge amplitude of the i-th time window. ; Arrange all windows in chronological order. To form a multivariate time series :
[0096] Where t is the time section index and Z is the number of time points, this multidimensional sequence comprehensively reflects the spatial distribution, temporal frequency and intensity evolution characteristics of discharge activity.
[0097] To address the limitations of traditional ARIMA models in multivariate time series forecasting, Support Vector Regression (SVR) is used to improve the model. The core improvement lies in combining the linear fitting ability of ARIMA with the nonlinear mapping ability of Support Vector Regression (SVR).
[0098] in, At a certain point in time The predicted value is the ultimate result that the discharge development trend prediction model aims to obtain; ARIMA ( () is an ARIMA component that takes multivariate time series as input. It is responsible for capturing the linear components (such as long-term trends and seasonal cycles); SVR ( , () is an SVR component that takes the residuals and feature gradients of the ARIMA component as input and is responsible for capturing nonlinear components and complex patterns that ARIMA cannot handle. For ARIMA components at time The prediction residual (i.e., the error between the predicted value and the actual value). For a moment The feature gradient vector reflects the rate of change of the feature parameters and represents the speed and direction of change of the sequence.
[0099] The ARIMA components are explained in detail as follows: They use autoregression, differencing, and moving averages to model and predict linear dependencies in time series data, with the goal of finding... The best linear expression:
[0100] Where p is the order of the autoregressive term (AR), representing the order of the past regression term. The values at each time point are used to predict the current value; d is the order of the difference (I), which is applied to the original data. The second difference is used to transform it into a stationary sequence (mean and variance do not change over time); q is the order of the moving average (MA). This indicates that the model considers past... The prediction error at each time point is used to improve the current prediction; It is a shift operator (or lag operator); It is the first The autoregressive coefficient measures the past... The degree of influence of each value on the current value; It is the first The moving average coefficient. Measured over the past [number] [period]. The degree of influence of each prediction error on the current value.
[0101] The SVR component is specifically designed to handle residual sequences. and feature gradient Nonlinear modes in:
[0102] in, It is the number of support vectors, that is, the number of training samples that play a key role in model building; These are Lagrange multipliers, introduced when solving the SVR optimization problem. Each training sample corresponds to a pair of multipliers, which determine which samples become "support vectors". It is a kernel function; It is the bias term (intercept), which is the bias of the final regression hyperplane.
[0103] The discharge development trend prediction model training adopts a two-stage optimization strategy. First, the optimal parameter combination (p, d, q) of ARIMA is determined by the AIC criterion. Then, grid search is used to optimize the penalty parameters of SVR. and kernel parameters .
[0104] After training, the model can simultaneously predict the future. Multidimensional feature values at each time point :
[0105] in, It is a sequence of the spatial distribution density of discharge events over a predicted future period of time. It is a sequence of the average discharge frequency of discharge events over a predicted future period of time. It is a sequence of the amplitude of optical signal disturbance of the predicted discharge event over a period of time. It is the prediction step size, that is, how many future time points the model should predict.
[0106] Based on the prediction results of the above model, a dynamic risk assessment function is constructed:
[0107] in, In the future The comprehensive discharge risk value at any given time is weighted based on expert experience and historical data analysis, with spatial distribution density weighting. Average discharge frequency weighting Optical signal perturbation amplitude weighting Predicted value weights .
[0108] Mini-maximum normalization is achieved by dividing by the historical maximum value. It is the future Spatial distribution density predicted at any time It is the future The average discharge frequency predicted at any given time; It is the future Predicted optical signal perturbation amplitude These are the historical maximum values of each feature parameter, used for normalization. It represents the acceleration of the change in characteristic parameters, and the output of this formula is a comprehensive risk value. The value range is [0,1].
[0109] Then, based on the calculated comprehensive risk value This triggers a tiered early warning mechanism: For normal operation (R<0.3), the system maintains the regular monitoring frequency and generates daily operation reports; for the level of concern (0.3≤R<0.6), the data collection frequency is automatically increased to 3 times the original, and the system generates a special analysis report every 4 hours; for the warning level (0.6≤R<0.8), the maintenance plan is prepared, requiring the maintenance team to complete on-site diagnosis and formulate a handling plan within 72 hours; for the alarm level (R≥0.8), the audible and visual alarms are triggered immediately, the emergency response procedure is initiated, intervention measures are required within 24 hours, and the situation is reported to management.
[0110] This technical solution provides a method and device for online monitoring of partial discharge in high-voltage cables based on the fusion of distributed fiber optic sensing and intelligent algorithms. The core of the solution lies in constructing a complete technical closed loop from physical perception to intelligent decision-making, achieving accurate perception, intelligent identification, precise location, trend prediction, and proactive early warning of partial discharge behavior. By innovatively embedding a specially designed fiber optic sensing unit within the cable sheath, it achieves synchronous and distributed perception of the electrical, thermal, and mechanical multi-physical field information accompanying partial discharge, providing comprehensive and spatiotemporally labeled raw data for subsequent analysis. Utilizing advanced signal processing techniques such as wavelet transform, it effectively filters out strong electromagnetic noise in the field and extracts time-frequency domain feature vectors that characterize the nature of the discharge, laying a solid foundation for pattern recognition. The system is based on quality data; a classification model built on 1D-CNN and residual networks has powerful deep feature learning capabilities, enabling high-precision and high-confidence automatic identification of discharge types, achieving intelligent diagnosis from signal to pattern; through DBSCAN clustering analysis, the system reveals the spatiotemporal evolution path and patterns of discharge clusters, achieving a leap from single point to panoramic view; a hybrid prediction model combining ARIMA and support vector regression is used to accurately predict discharge development trends, and risk values are calculated based on multi-dimensional feature fusion to trigger graded early warnings, realizing predictive maintenance from post-event alarm to pre-event warning.
[0111] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, the phrase "comprising an element defined as..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0112] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their likenesses.
Claims
1. A method for online monitoring of partial discharge in high-voltage cables, characterized in that, Includes the following steps: Step S1: Embed a distributed optical fiber sensing unit in the high-pressure cable sheath layer, and convert the physical field change data generated when a partial discharge event occurs into an initial partial discharge optical signal sequence through the distributed optical fiber sensing unit. Step S2: Denoise the initial partial discharge optical signal sequence using wavelet transform algorithm to obtain a denoised partial discharge optical signal sequence, and extract features from the denoised partial discharge optical signal sequence to obtain a partial discharge feature vector; Step S3: Based on the one-dimensional convolutional neural network 1D-CNN model, construct a partial discharge classification model, input the partial discharge feature vector into the partial discharge classification model, and output the partial discharge classification result and the spatiotemporal information of the partial discharge event; Step S4: Based on the partial discharge classification results and the spatiotemporal information of partial discharge events, the DBSCAN spatial clustering algorithm is used to perform spatiotemporal clustering of partial discharge events to generate the spatiotemporal evolution path of partial discharge. Step S5: Based on the autoregressive integral moving average model, construct a discharge development trend prediction model, input the characteristic parameters of the spatiotemporal evolution path into the discharge development trend prediction model, output the discharge development trend prediction result, calculate the partial discharge risk value based on the discharge development trend prediction result, and perform graded early warning for partial discharge events based on the partial discharge risk value.
2. The online monitoring method for partial discharge of high-voltage cables according to claim 1, characterized in that, The method of embedding a distributed optical fiber sensing unit in the high-pressure cable sheath layer includes: In the manufacturing process of the sheath layer of high-voltage cables, distributed optical fiber sensing units are directly embedded, and close physical contact and signal coupling with the cable's metal shielding layer are achieved through conductive adhesive layers. The distributed optical fiber sensing unit adopts a four-layer composite structure design from the outside to the inside: the outermost layer is a nickel-plated fiber braided shielding layer; the middle layer is a polyimide electro-erosion resistant coating; a thermally conductive silicone rubber layer is provided between the outermost and middle layers; and the innermost layer is a standard single-mode optical fiber core.
3. The online monitoring method for partial discharge of high-voltage cables according to claim 2, characterized in that, The physical field change data generated during a partial discharge event is converted into an initial partial discharge optical signal sequence, including: A nanosecond-level laser pulse is emitted into an optical fiber, and its backscattered signal is detected. The phase change Δφ and intensity change Δφ of the Rayleigh scattered light are then extracted. Brillouin radio frequency shift change ; Each emitted laser pulse is marked with an absolute timestamp using a high-precision clock. When in fiber optic location When a scattering event is detected, record the absolute timestamp of the received scattering signal. Light from emission to position The round-trip time to the receiving end is ; Finally, the aforementioned Rayleigh scattered light phase change Δφ and Rayleigh scattered light intensity change Δφ are calculated. Brillouin radio frequency shift change By associating the signal with the corresponding spatiotemporal coordinates (t, x), an initial partial discharge optical signal sequence with spatiotemporal coordinates is constructed. .
4. The online monitoring method for partial discharge of high-voltage cables according to claim 3, characterized in that, The initial partial discharge optical signal sequence is denoised using a wavelet transform algorithm to obtain a denoised partial discharge optical signal sequence, including: First, the initial partial discharge signal sequence... By spatial coordinates Group, each Corresponding to a three-dimensional time series Wavelet decomposition and denoising are performed separately on each time series to obtain wavelet coefficients. ; Next, thresholding is performed on the detail coefficients of each layer obtained from the decomposition. An adaptive threshold quantization is performed using an unbiased risk estimation threshold method. The threshold of this method... Determined by the following formula: ； in, The noise standard deviation is determined by the first level detail coefficient. The median estimate, The signal length; Subsequently, a soft thresholding function is applied to process the coefficients: ； in, After soft thresholding, the first Layer reconstruction detail coefficients; After wavelet decomposition, the th The original detail factor of the layer; It is a symbolic function; Finally, wavelet reconstruction is performed to extract the detail coefficients after thresholding. Approximation coefficients retained Perform inverse wavelet transform to reconstruct the denoised partial discharge optical signal sequence. .
5. The online monitoring method for partial discharge of high-voltage cables according to claim 4, characterized in that, Feature extraction is performed on the noise-reduced partial discharge optical signal sequence to obtain a partial discharge feature vector, including: First, extract the peak intensity of Rayleigh scattering. This feature comes directly from the noise reduction. , equal to the measured intensity during discharge Reference strength when there is no discharge The difference, reference strength Calibrated during initialization; peak value It is the pulse duration. Inside, The maximum value; Secondly, extract the phase change peak value. , directly from The extraction method involves detecting the peak offset of the phase pulse waveform. ; Furthermore, extract the pulse duration. ; Then, extract the Rayleigh scattering pulse energy. Also stemming from noise reduction The extraction method involves calculating the integral energy of the pulse waveform over its duration. ； in, It is the change in the intensity of the noise-reduced Rayleigh scattering light; Finally, the Brillouin frequency shift offset is extracted. The direct source is the change in frequency shift of the noise-reduced Brillouin scattering light. During extraction, the peak value of the frequency-shifted pulse synchronized with the partial discharge event is captured, and the calculation formula is as follows: ； in, It is the peak value of the Brillouin frequency shift. The frequency shift change of the noise-reduced Brillouin scattering light; Finally, a corresponding partial discharge feature vector is generated for each partial discharge event. .
6. The online monitoring method for partial discharge of high-voltage cables according to claim 5, characterized in that, Step S3 includes: First, the input data is preprocessed, and the five-dimensional feature vector output from step S2 is reconstructed into a tensor format suitable for network processing. The data is converted into a preset data structure and then standardized to obtain preprocessed input data. ; Subsequently, we enter the core feature extraction stage, where deep feature learning is performed sequentially through two residual modules; The first residual module processing flow, input features First, local feature patterns are extracted through convolution operations: ； in, It is the feature map output by the first convolutional layer; These are the convolution kernel weights, responsible for capturing local correlations between features; This is a bias term used to adjust the output distribution; It is the convolution operator; Next, batch normalization is performed on the convolutional output to stabilize the training process; Finally, feature fusion is achieved through residual connections: ； in, This is the final output of the first residual module; A composite function representing convolution and batch normalization; ( () is a linear rectification activation function; The original input to the residual module; The second residual module's processing flow uses the output of the first module as the input of the second module, achieving deep feature abstraction: ； in, It is the input for the second module; Finally, the classification decision is made through a fully connected layer and a softmax function, and the output of the second residual module is used. After flattening, the data is fed into a fully connected layer to obtain the raw scores for each category; Then, the original scores are converted into a probability distribution using the Softmax function; The category corresponding to the highest probability is used as the prediction result. This probability value also serves as the confidence level. When confidence level When the result is determined to be of high confidence, it proceeds directly to the next step. process.
7. The online monitoring method for partial discharge of high-voltage cables according to claim 6, characterized in that, The DBSCAN spatial clustering algorithm is used to perform spatiotemporal clustering of partial discharge events, generating spatiotemporal evolution paths of partial discharges, including: Using the three-dimensional coordinates of the discharge event An improved spatial clustering algorithm, DBSCAN, is used for spatiotemporal clustering: ； in, Indicates the first Each discharge cluster has a spatial neighborhood radius of 2m to ensure geographical relevance, and a time window of 10s to ensure event continuity. It is the k-th point in the cluster. These are the times when the signal arrives at points i and j, respectively; Represents the spatial coordinates of two partial discharge events; and Indicates the type label of the discharge event; Finally, based on the clustering results, three key characteristic parameters of each discharge cluster were calculated, including spatial distribution density. Average discharge frequency optical signal disturbance amplitude ; Next, by analyzing the spatiotemporal distribution patterns of events within the cluster, the propagation direction vector is calculated. and evolutionary acceleration ; Ultimately, this information from all discharge clusters together constitutes the complete spatiotemporal evolution path G of partial discharge.
8. The online monitoring method for partial discharge of high-voltage cables according to claim 7, characterized in that, The three-dimensional coordinates of the discharge event The determination includes: For each discharge event, its three-dimensional coordinates have been determined by step S1, and its spatial location is determined by fiber optic coordinates. Provide elevation information. The data is derived from cable laying data; the planar coordinates are determined by combining the cable route diagram, and finally the three-dimensional coordinates of the discharge event are obtained. .
9. The online monitoring method for partial discharge of high-voltage cables according to claim 8, characterized in that, Based on the autoregressive integral moving average model, a discharge development trend prediction model is constructed. The characteristic parameters of the spatiotemporal evolution path are input into the discharge development trend prediction model, and the output is the discharge development trend prediction result, including: First, classify by discharge type Grouping the discharge cluster set G into subsets of clusters of the same type yields clusters of the same type. Divide the timeline into equidistant windows. For each window Aggregate all cluster features it covers: ； in, For window The set of covered clusters, ; It is the average discharge density in the i-th time window; The average discharge frequency of the i-th time window; the average discharge amplitude of the i-th time window. ; Arrange all windows in chronological order. To form a multivariate time series ; The model is improved using Support Vector Regression (SVR), which combines the linear fitting ability of ARIMA with the nonlinear mapping ability of SVR: ； in, The discharge development trend prediction model at a certain time point Predicted values; ARIMA ( ) is an ARIMA component; SVR ( , ) is an SVR component; For ARIMA components at time The predicted residuals; For a moment The feature gradient vector; The discharge development trend prediction model training adopts a two-stage optimization strategy. First, the optimal parameter combination (p, d, q) of ARIMA is determined by the AIC criterion. Then, grid search is used to optimize the penalty parameters of SVR. and kernel parameters ; After training is completed, predict the future simultaneously. Multidimensional feature values at each time point : ； in, It is a sequence of the spatial distribution density of discharge events over a predicted future period of time. It is a sequence of the average discharge frequency of discharge events over a predicted future period of time. It is a sequence of the amplitude of optical signal disturbance of the predicted discharge event over a period of time. It predicts the step size.
10. A high-voltage cable partial discharge online monitoring device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1-9.