High-voltage cable partial discharge monitoring and early warning method and system based on multi-dimensional data perception

Abstract

The application relates to the technical field of partial discharge detection, in particular to a high-voltage cable partial discharge monitoring and early warning method and system based on multi-dimensional data sensing, a preprocessing module pre-processes partial discharge detection data including ultrasonic signal data and electric signal data, removes environmental noise and carries out signal feature screening to generate an initial signal feature set; the initial signal feature set is input into a type fusion module, ultrasonic features and electric signal features of the signal features are spliced and dimensionally reduced based on a partial discharge fault type, and an optimal feature set is output; the optimal feature set is respectively input into a deep learning network to determine whether a partial discharge fault exists and the type of the partial discharge fault. The preprocessing module and the type fusion module carry out two times of dimension reduction processing on the dimensions, which not only avoids the dimension explosion problem of a detection model, but also extracts multi-dimensional features to avoid the problem that a single signal is easily disturbed, and is beneficial to improving the accuracy of partial discharge fault detection.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of partial discharge detection, and particularly relates to a high-voltage cable partial discharge monitoring and early warning method and system based on multi-dimensional data perception. BACKGROUND

[0002] High-voltage cable partial discharge (partial discharge) is a microscopic breakdown phenomenon caused by the concentration of partial electric field inside the insulation, although the instantaneous energy is small, but it hides great harm. The electric, thermal and chemical effects produced by partial discharge will continuously erode the insulation material, accelerate the aging process, cause the insulation performance to gradually deteriorate, and eventually may cause insulation breakdown, resulting in cable short circuit, large area power failure, and even fire and explosion caused by arc, threatening the safety of power grid equipment and personnel. Detecting high-voltage cable partial discharge is the key to ensuring the stable operation of the power system, early detection of fault defects, so as to take timely repair or replacement measures to avoid the expansion of faults, which can effectively reduce the economic loss caused by sudden power failure and ensure the safe and reliable power supply of the power grid.

[0003] The existing partial discharge detection technology is mainly divided into ultrasonic detection method and electromagnetic signal detection method, which respectively refer to ultrasonic signal characteristics and electric signal characteristics. The advantage of this is to reduce the feature extraction dimension of the judgment model, improve the training speed of the model, reduce the training difficulty, avoid the dimension explosion of the model, and the disadvantage is that both kinds of signals are easily disturbed by environmental noise, especially in high-voltage power transmission environment, mechanical vibration and electromagnetic interference are everywhere in the detection space. In actual detection, it is difficult to completely filter out the interference of environmental noise on a single signal, resulting in inaccurate model judgment. SUMMARY

[0004] The application provides a high-voltage cable partial discharge monitoring and early warning method and system based on multi-dimensional data perception to solve or partially solve the problems raised in the background.

[0005] The application provides a high-voltage cable partial discharge monitoring and early warning method based on multi-dimensional data perception, which comprises the following steps:

[0006] S100: The sensor group acquires real-time partial discharge detection data of the detection area, including ultrasonic signal data and electric signal data;

[0007] S200: The preprocessing module pre-processes the partial discharge detection data, removes environmental noise and performs signal feature screening to generate an initial signal feature set;

[0008] S300: The initial signal feature set is input into the type fusion module, the signal features are spliced and dimensionally reduced based on the partial discharge fault type, and an optimal feature set is output;

[0009] S400: input the optimal feature set into the deep learning network respectively, and determine whether there is a partial discharge fault and the type of the partial discharge fault.

[0010] Preferably, in the step S200, the specific method of the preprocessing module for preprocessing the partial discharge detection data at least includes noise suppression, time-frequency domain conversion, outlier processing, and data normalization, and further includes time-based segmentation and alignment of ultrasonic signals and electrical signals.

[0011] Preferably, in the step S200, the method of the preprocessing module for generating the initial signal feature set is as follows:

[0012] The dependence of the quantized signal features on the fault labels is quantified by using mutual information MI, and the quantization formula is as follows:

[0013]

[0014]

[0015]

[0016] wherein x is a certain signal feature, X is the total set of features, y is a certain fault label, Y is the total set of fault labels, P(x) is the marginal probability distribution of feature x, and P(x, y) is the joint probability distribution of feature x and label y.

[0017] The first K1 ultrasonic signal features and the first K2 electrical signal features with the maximum MI values are obtained respectively to generate the initial signal feature set.

[0018] Preferably, in the step S300, a parallel XGBoost module consistent with the number of partial discharge fault types is set in the type fusion module, each module corresponds to one type of partial discharge, and the XGBoost is used to reduce the dimension of features for the corresponding fault type in a supervised mode, and redundant features are recursively removed.

[0019] Preferably, specifically, the training method of the XGBoost module is as follows:

[0020] S301: The ultrasonic features and electrical signal features in the signal feature training sample processed by the preprocessing module are spliced and combined to generate an initial feature set for each sample, and the samples are divided into a training set and a validation set;

[0021] S302: The parameters of the XGBoost module are initialized, and the rejection ratio r and the performance decline threshold τ are set;

[0022] S303: Iterative training is performed to remove features, and the iterative method is as follows:

[0023] Based on the current feature set XGBoost classifier on the training set, and calculate the contribution of each feature to the model decision and the normalized importance during the training process

[0024]

[0025]

[0026] wherein, is the set of decision tree nodes using feature x for splitting, is the reduction of the loss function after splitting node t, and M is the current feature set collection;

[0027] eliminate features with low importance scores by a preset proportion r;

[0028] S304: Calculate the performance of the current model using the validation set, and determine whether the termination condition is met

[0029]

[0030] wherein, and are the current and historical optimal F1 scores of the model, if, output the optimal feature subset, if not, jump to step S303 and continue training XGBoost using the training set after eliminating features.

[0031] Preferably, in step S301, the ultrasonic features and electrical signal features in the training sample are combined in the form of a preset numerical arrangement mode with electrical signal features first and ultrasonic features last. The missing values are filled with the median or mean value. The ratio of the training set and the validation set can be set to 8:2.

[0032] Preferably, in step S400, the deep learning network includes a variational autoencoder VAE and a classification neural network. The latent features output by the last layer of the encoder of the variational autoencoder VAE are input to the input layer of the classification neural network, and the weight parameters of the encoder are used as the initial weights of the subsequent classification neural network.

[0033] Preferably, the variational autoencoder VAE includes an encoder and a decoder. The encoder is a 3-layer fully connected network, and outputs latent abstract features. The decoder is also a 3-layer fully connected network, and uses the latent features to reconstruct the original features. The gradients of the loss function with respect to the parameters of the encoder and the decoder are calculated, and the parameters are updated in reverse using an optimizer to maintain the effectiveness of the latent features.

[0034] The optimizer is an Adam optimizer, and the loss function includes a reconstruction error MSE and a KL divergence.

[0035] Preferably, the classification neural network comprises an input layer, two hidden layers, and an output layer, the input layer is connected to the VAE encoder, the hidden layer activation function is selected as the ReLU function, and the output layer is provided with a softmax classifier.

[0036] The latent features output by the last layer of the VAE encoder are input into the classification neural network to obtain a label probability distribution corresponding to each optimal feature subset, where j is the label type number, and takes values of 1, 2, 3,..., J, J is the number of labels, and the label type with the maximum probability value is selected from the label probability distribution as the result output.

[0037] The application also provides a high-voltage cable partial discharge monitoring and early warning system based on multi-dimensional data perception, comprising a sensor group, a preprocessing module, a type fusion module, and a deep learning network, and the functions of each module are as follows:

[0038] The sensor group acquires real-time partial discharge detection data in a detection area.

[0039] The preprocessing module pre-processes the partial discharge detection data, removes environmental noise, and performs signal feature screening to generate an initial signal feature set.

[0040] The type fusion module performs ultrasonic feature and electrical signal feature splicing and dimensionality reduction processing on the signal features based on the partial discharge fault type to output an optimal feature set.

[0041] The deep learning network inputs the optimal feature set to determine whether a partial discharge fault exists and the type of the partial discharge fault.

[0042] Compared with the prior art, the application has the following beneficial effects:

[0043] (1) The partial discharge fault determination method of the application extracts ultrasonic signal features and electrical signal features from the partial discharge detection data, and performs two times of dimensionality reduction processing through the preprocessing module and the type fusion module, which not only avoids the dimensionality explosion problem of the detection model, but also extracts multi-dimensional features to avoid the problem that a single signal is easily disturbed, thereby improving the accuracy of partial discharge fault detection.

[0044] (2) The application uses mutual information MI to quantify the dependence relationship between the signal features and the fault labels, performs preliminary dimensionality reduction on the signal features, and uses statistical methods to preliminarily screen the signal features, thereby reducing the difficulty of dimensionality reduction of the type fusion module and the training difficulty of the deep learning network, and improving the overall response speed of the model.

[0045] ​(3) Based on the simplification of partial discharge fault types, the application performs secondary dimension reduction on signal features based on XGBoost model fault type, removes the least important batch of features by using the recursive feature elimination mode of XGBoost, captures the nonlinear relationship between features, extracts the most critical feature subset, strengthens the expression ability of signal features on the basis of reducing dimensions and maintaining model size, and is beneficial to subsequent accurate identification.

[0046] (4) The application extracts high-order abstract features in signal features by using variational autoencoder VAE, which is beneficial to mining the internal abstract relationship between ultrasonic signal features and electrical signal features, provides judgment basis and initial weight for subsequent classification neural network, and improves the learning efficiency and judgment accuracy of the classification neural network. BRIEF DESCRIPTION OF DRAWINGS

[0047] The application will be further described below in conjunction with the drawings and examples.

[0048] Fig. 1 For the method flow of the application,

[0049] Fig. 2 For the system composition of the application,

[0050] Fig. 3 For the partial discharge data processing implementation of the application. DETAILED DESCRIPTION

[0051] As some terms are used in the description and claims to refer to certain components. Those skilled in the art should understand that hardware manufacturers may use different names to refer to the same component. The description and claims of this specification do not distinguish components by name, but by the functional difference between components. As mentioned throughout the description and claims, "including" is an open term, which should be interpreted as "including but not limited to". "Approximately" means within an acceptable error range, and those skilled in the art can solve technical problems within a certain error range, and basically achieve the technical effect.

[0052] In the description of the application, it should be understood that the terms "upper", "lower", "front", "back", "left", "right", "horizontal" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the application.

[0053] In this application, unless specifically defined and limited otherwise, the terms "mounting", "connected", "connection", "fixed", and the like, should be construed broadly and do not necessarily mean fixedly connected, but can also mean detachably connected, or integrally connected; can be mechanical connection, can also be electrical connection; can be directly connected, can also be indirectly connected through an intermediate medium, and can be internal communication of two elements. For those skilled in the art, the specific meanings of the above terms in this application can be understood according to the specific circumstances.

[0054] Embodiment 1

[0055] As Figs. 1-2 shown, the application provides a high-voltage cable partial discharge monitoring and early warning method based on multi-dimensional data perception, which includes the following steps:

[0056] S100: The sensor group acquires real-time partial discharge detection data of the detection area, including ultrasonic signal data and electric signal data;

[0057] S200: The preprocessing module pre-processes the partial discharge detection data, removes environmental noise and performs signal feature screening to generate an initial signal feature set;

[0058] S300: The initial signal feature set is input into the type fusion module, and the signal features are spliced and dimensionally reduced based on the partial discharge fault type, and the optimal feature set is output;

[0059] S400: The optimal feature set is input into the deep learning network respectively, to determine whether there is a partial discharge fault and the type of the partial discharge fault.

[0060] Specifically, in step S100, the sensor group at least includes an ultrasonic detection sensor for detecting ultrasonic waves and a very high frequency sensor for detecting partial discharge electric signals.

[0061] Specifically, in step S200, the preprocessing module processes the ultrasonic signal data and the electric signal data respectively to suppress interference noise, retain signal features, and standardize signal format.

[0062] The preprocessing method of the ultrasonic signal includes data cleaning and baseline correction, noise suppression, time-frequency domain conversion, outlier processing, data normalization, etc. Specifically, the baseline drift of the ultrasonic signal is mainly caused by sensor zero drift or low-frequency vibration, and high-order high-pass filtering is used to suppress sensitive low frequencies. The wavelet packet transform algorithm is used to suppress high-frequency noise, which supports higher frequency signal decomposition than wavelet transform. The time-frequency characteristics of the ultrasonic signal are calculated, and the short-time Fourier transform (STFT) and / or continuous wavelet transform (CWT) are used to enhance the extraction of video features. The power spectral density (PSD) of the ultrasonic signal is calculated as a time-frequency auxiliary feature. The "amplitude mutation detection" (such as the difference between adjacent sampling points exceeding a threshold) is used to identify typical outliers caused by unexpected situations such as sudden vibration. Z-Score normalization or other algorithms are used to normalize the signal features.

[0063] The preprocessing method of the electrical signal is similar to that of the ultrasonic signal, and also includes baseline correction, targeted noise suppression, time-frequency domain feature conversion, outlier processing, data normalization, etc. Noise suppression is the core step of partial discharge signal preprocessing, including power frequency interference suppression, high-frequency electromagnetic interference and white noise suppression. Wavelet transform and other algorithms can be used to effectively suppress noise.

[0064] The preprocessing method of the electrical signal and the ultrasonic signal has the same process but different technical details. The former is a pulsed electrical signal, and the latter is an acoustic vibration signal. The specific algorithms used in each step of the above processing method (such as wavelet transform) are conventional technical means in the field, and will not be described here.

[0065] Preferably, the preprocessing method of the partial discharge detection data by the preprocessing module further includes time-based segmentation and alignment of the ultrasonic signal and the electrical signal to synchronize their times.

[0066] The preprocessing module performs noise processing and feature extraction on the partial discharge detection data, including ultrasonic signal features and electrical signal features. Signal features include time domain features, frequency domain features, etc. For example, the vibration peak (time domain) and center frequency (frequency domain) in the ultrasonic signal feature, and the pulse peak and pulse half-width (time domain) in the electrical signal feature. If all signal features are used for subsequent deep learning network learning, it is easy to cause dimension explosion, affecting recognition accuracy and speed. The preprocessing module needs to preliminarily screen the signal features.

[0067] The specific method of the preprocessing module for screening signal features is as follows:

[0068] The mutual information MI is used to quantify the dependence of the signal features on the fault label, and the quantification formula is as follows:

[0069]

[0070]

[0071]

[0072] Where x is a certain signal feature, X is the feature set, y is a certain fault label, Y is the fault label set, P(x) is the marginal probability distribution of feature x, and P(x,y) is the joint probability distribution of feature x and label y.

[0073] The top K1 ultrasound signal features and the top K2 electrical signal features with the largest MI values ​​are obtained respectively to generate an initial signal feature set.

[0074] The preprocessing module uses statistical methods to initially screen signal features, which reduces the dimensionality reduction difficulty of the subsequent type fusion module and the training difficulty of the deep learning network, thus improving the overall response speed of the model.

[0075] The general classification of partial discharge fault types is as follows:

[0076] Corona discharge is a localized discharge that occurs in a gaseous medium at the tip, edge, or irregular surface of a conductor due to a highly concentrated electric field (the electric field strength exceeds the gas breakdown field strength). It is classified as "discharge on the surface of a conductor in a gas".

[0077] Surface discharge is a partial discharge that occurs along the interface between different media (such as the interface between solid insulation and gas / liquid), caused by the distortion of the electric field at the interface (electric field lines tend to concentrate when they are perpendicular to the interface).

[0078] Internal discharge occurs in the gaps, bubbles, or impurities inside solid / liquid insulation. It is caused by defects inside the insulation (such as air gaps, cracks, inclusions) that lead to local electric field concentration, which exceeds the breakdown field strength of the medium inside the defect.

[0079] Floating potential discharge occurs when a metal component in a high-voltage device that is insulated from the main electrode (such as a loose metal sheet or an ungrounded shielding layer) becomes charged due to electromagnetic induction, forming a "floating potential" and creating an electric field with the surrounding conductor (or ground), resulting in a discharge in the gap.

[0080] This application can use the above-mentioned 4-category fault types, or it can be finely adjusted and added or subtracted based on the actual scenario. However, the number of fault types is relatively concise overall. Based on the concise number of types, multiple parallel XGBoost modules (consistent with the number of partial discharge fault types) are set in the type fusion module. Each module corresponds to a partial discharge type. The supervised mode is used to perform feature dimensionality reduction on XGBoost for the corresponding fault type, identify important features, and recursively remove redundant features.

[0081] Specifically, the training method for the XGBoost module is as follows:

[0082] S301: The ultrasound features and electrical signal features in the signal feature training sample processed by the preprocessing module are spliced and combined to generate an initial feature set for each sample, and the samples are divided into a training set and a validation set;

[0083] S302: Initialize the XGBoost module parameters, set the rejection ratio r and the performance decline threshold τ;

[0084] S303: Iterative training to eliminate features, the iterative method is as follows:

[0085] Based on the current feature set (the feature set of the kth iteration), train the XGBoost classifier on the training set, and calculate the contribution of each feature to the model decision and the normalized importance during the training process

[0086]

[0087]

[0088] wherein, is the set of decision tree nodes that use feature x for splitting, is the reduction of the loss function after node t is split, and M is the current feature set;

[0089] Reject features with low importance scores according to the preset ratio r;

[0090] S304: Calculate the current model performance using the validation set to determine whether the termination condition is met

[0091]

[0092] wherein, and are the current and historical optimal F1 scores of the model, if the condition is met, output the optimal feature subset, if the condition is not met, jump to step S303 and continue training XGBoost using the training set after removing the features.

[0093] XGBoost-RFE based on XGBoost is a combination method of "iterative feature selection + model evaluation", the core logic is: use XGBoost model to evaluate feature importance, remove the least important batch of features each time, then retrain the model with the remaining features, until the model performance declines to an unacceptable level, its advantage lies in that it can not only accurately evaluate the value of features by using the capturing ability of XGBoost for nonlinear relationships, but also remove redundant features through recursion, finally retaining the most critical feature subset for model performance.

[0094] It is worth noting that for each type of arcing fault corresponding to the XGBoost module, when labeling the training samples, only the fault or not of the specific fault type needs to be labeled, so that each XGBoost module can perform targeted feature screening and dimension reduction for specific faults.

[0095] In step S301, the typical form of the combination of the ultrasonic features and the electrical signal features in the training sample can be a preset numerical arrangement mode in which the electrical signal features are in front and the ultrasonic features are behind. The missing values are filled with the median or mean value. The ratio of the training set and the validation set can be set to 8:2.

[0096] Specifically, in step S400, the deep learning network includes a variational autoencoder VAE and a classification neural network. The variational autoencoder VAE is used to mine high-order abstract relationships in the ultrasonic signal features and the electrical signal features, and provides pre-training weights for the classification neural network.

[0097] Specifically, the variational autoencoder VAE includes an encoder and a decoder. The encoder is a 3-layer fully connected network, and outputs latent abstract features. The decoder is also a 3-layer fully connected network, reconstructs the original features using the latent features, and propagates back to the encoder to maintain the effectiveness of the latent features. The latent features output by the last layer of the encoder are input to the input layer of the classification neural network, and the weight parameters of the encoder are used as the initial weights of the subsequent classification neural network.

[0098] Specifically, the training method of the variational autoencoder VAE is as follows:

[0099] The encoder maps the input feature x to the latent distribution in the neural network:

[0100] ,

[0101] wherein, is the mean vector and the standard deviation vector of the latent distribution, K is the dimension of the latent space, and K is less than the dimension D of the feature x, is the encoder neural network, is the mean network weight and bias parameter, is the variance network parameter;

[0102] Reparameterization sampling:

[0103]

[0104] is a noise vector subject to a normal distribution;

[0105] The decoder reconstructs the input from the latent variable z

[0106]

[0107] where g is a decoder neural network, are weights and biases;

[0108] calculating a loss function L

[0109]

[0110]

[0111]

[0112] where, is a reconstruction error MSE, is a KL divergence for constraining the latent distribution, and is the d-th component of input feature and reconstructed feature, is the k-th component of , and β is a KL divergence weight;

[0113] calculating the gradient of the total loss L with respect to the encoder, decoder parameters (θ, φ), , , updating the parameters using the Adam optimizer by backpropagation;

[0114] stopping the training iteration when the total loss L of the validation set no longer decreases for consecutive iterations or the reconstruction error MSE is less than a preset threshold.

[0115] The encoder of the VAE maps the input feature to a low-dimensional latent space, learning a compressed abstract representation of the input data. The decoder reconstructs the original input feature from the sample in the latent space, learning the mapping relationship from the latent feature to the original feature. The goal is to recover the details of the input, and the latent feature is optimized through backpropagation.

[0116] Specifically, the classification neural network includes an input layer, two hidden layers, and an output layer. The input layer is connected to the VAE encoder, the hidden layer activation function is selected as the ReLU function, and the output layer is set as a softmax classifier. The input layer receives the VAE latent feature, the hidden layer performs multi-level abstract extraction on the latent feature, and the output layer outputs the fault-free or fault type label and the probability corresponding to the label. During training, the Adam optimizer is used, and the cross-entropy loss function is used as the loss function.

[0117] The ultrasonic signal data and the electric signal data in the partial discharge detection data are denoised and screened by the preprocessing module to generate an initial signal feature set. The initial signal feature set is input into an XGBoost module corresponding to each partial discharge fault type in a type fusion module to output an optimal feature subset C-i corresponding to each partial discharge fault type, where i is 1, 2, 3,..., I, I is the number of partial discharge fault types in the application, and each optimal feature subset is input into a variational autoencoder VAE to extract high-order abstract features. The latent features output by the last layer of the VAE encoder are input into a classification neural network to obtain a label probability distribution corresponding to each optimal feature subset. }, where j is a label type number and takes a value of 1, 2, 3,..., J, J is the number of labels, and the label type with the maximum probability value is selected from the label probability distribution as the result output.

[0118] Since a softmax classifier is used, the sum of the probabilities of all labels in the label probability distribution corresponding to each optimal feature subset is 1, that is, , which has a normalization effect. Therefore, the label type with the maximum probability value is selected from the label probability distribution as the result output. Generally, the optimal feature subset corresponding to the selected label is also output by the XGBoost module optimized for the corresponding fault type. If the two do not match, the corresponding partial discharge detection data is reanalyzed and identified, or the partial discharge detection data is discarded as a special interference signal.

[0119] The XGBoost module is used in the application to optimize the partial discharge fault types, to perform targeted screening and dimensionality reduction on the signal features, and to extract high-order abstract correlations in the screened signal features using the variational autoencoder VAE to enhance the expression ability of the signal features, which is conducive to improving the response rate and response speed of the partial discharge faults.

[0120] Embodiment 2

[0121] Based on the high-voltage cable partial discharge monitoring and early warning method based on multi-dimensional data perception in embodiment 1, the application further provides a high-voltage cable partial discharge monitoring and early warning system based on multi-dimensional data perception, which includes a sensor group, a preprocessing module, a type fusion module, and a deep learning network. The functions of each module are as follows:

[0122] The sensor group acquires real-time partial discharge detection data in a detection area.

[0123] The preprocessing module preprocesses the partial discharge detection data, removes environmental noise, and performs signal feature screening to generate an initial signal feature set.

[0124] ​​The type fusion module performs splicing and dimension reduction processing on the ultrasonic features and electrical signal features of the signal features based on the partial discharge fault type, and outputs an optimal feature set;

[0125] The deep learning network inputs the optimal feature set, and judges whether there is a partial discharge fault and a partial discharge fault type.

[0126] The embodiments of the application are described in detail above in combination with the drawings, but the application is not limited to the above embodiments, and various changes can be made within the knowledge of those skilled in the art without departing from the purpose of the application.

Claims

1. A method for monitoring and early warning of partial discharge in high-voltage cables based on multi-dimensional data perception, characterized in that, Includes the following steps: S100: The sensor group acquires real-time partial discharge detection data of the detection area, including ultrasonic signal data and electrical signal data; S200: The preprocessing module preprocesses the partial discharge detection data, removes environmental noise, and performs signal feature filtering to generate an initial signal feature set. S300: Input the initial signal feature set into the type fusion module, perform ultrasonic feature and electrical signal feature splicing and dimensionality reduction processing on the signal features based on the partial discharge fault type, and output the optimal feature set; S400: Input the optimal feature set into the deep learning network to determine whether there is a partial discharge fault and the type of partial discharge fault; In step S200, the preprocessing module uses mutual information (MI) to quantize the dependency relationship between signal features and fault labels, and uses statistical methods to perform preliminary screening of signal features. In step S300, a parallel XGBoost module with the same number of partial discharge fault types is set in the type fusion module. Each module corresponds to a partial discharge type. The supervised mode is used to perform feature dimensionality reduction on XGBoost for the corresponding fault type and recursively remove redundant features. For each type of partial discharge fault, when labeling the training samples of the XGBoost module, it is only necessary to label whether the fault is present or not for the specific fault type, so that each XGBoost module can perform targeted feature selection and dimensionality reduction for specific faults. In step S400, the deep learning network includes a variational autoencoder (VAE) plus a classification neural network. The latent features output from the last layer of the encoder of the variational autoencoder (VAE) are input to the input layer of the classification neural network, and the weight parameters of the encoder are used as the initial weights of the subsequent classification neural network. The classifier of the classification neural network is a softmax classifier. The initial signal feature set is input into the type fusion module. Each XGBoost module outputs the optimal feature subset Ci, where i takes values ​​1, 2, 3, ..., I, and I is the number of partial discharge fault types. Each optimal feature subset is input into a variational autoencoder (VAE) for high-order abstract feature extraction. The latent features output from the last layer of the VAE encoder are then input into a classification neural network to obtain the label probability distribution corresponding to each optimal feature subset. }, j is the tag type number, taking values ​​1, 2, 3, ... J, where J is the number of tags, derived from the tag probability distribution { The label type with the highest probability value is selected as the output result.

2. The high-voltage cable partial discharge monitoring and early warning method based on multi-dimensional data perception according to claim 1, characterized in that: In step S200, the specific methods for the preprocessing module to preprocess the partial discharge detection data include at least noise suppression, time-frequency domain conversion, outlier processing, and data normalization, as well as time-based segmentation and alignment of the ultrasonic signal and electrical signal.

3. The high-voltage cable partial discharge monitoring and early warning method based on multi-dimensional data perception according to claim 1, characterized in that: In step S200, the preprocessing module generates the initial signal feature set using the following method: The dependency between signal features and fault labels is quantized using mutual information MI (Multi-Information) quantization. The quantization formula is as follows: Where x is a certain signal feature, X is the feature set, y is a certain fault label, Y is the fault label set, P(x) is the marginal probability distribution of feature x, and P(x,y) is the joint probability distribution of feature x and label y. The top K1 ultrasound signal features and the top K2 electrical signal features with the largest MI values ​​are obtained respectively to generate an initial signal feature set.

4. The high-voltage cable partial discharge monitoring and early warning method based on multi-dimensional data perception according to claim 1, characterized in that: Specifically, the training method for the XGBoost module is as follows: S301: The ultrasound features and electrical signal features in the signal feature training samples after preprocessing are spliced ​​and combined to generate an initial feature set for each sample, and the samples are divided into training set and validation set. S302: Initialize XGBoost module parameters, set the elimination ratio r and the performance degradation threshold τ; S303: Iterative training to remove features, the iterative method is as follows: Based on the current feature set An XGBoost classifier is trained on the training set, and during training, the contribution and normalization importance of each feature to the model's decision are calculated. in, This is the set of decision tree nodes that are split using feature x. Let M be the amount of reduction in the loss function after node t splits, and M be the current feature set. Features with low importance scores are removed according to a preset ratio r; S304: Calculate the current model performance using the validation set and determine if the termination condition is met. in, and The current and historical best F1 scores of the model are given. If they are satisfied, the best feature subset is output. If not, the process jumps to step S303 and continues to train XGBoost using the training set after removing features.

5. The high-voltage cable partial discharge monitoring and early warning method based on multi-dimensional data perception according to claim 4, characterized in that: In step S301, the ultrasound features and electrical signal features in the training samples are combined in a preset numerical arrangement pattern with electrical signal features first and ultrasound features second. Missing values ​​are filled with the median or mean. The ratio of the training set to the validation set can be set to 8:

2.

6. The high-voltage cable partial discharge monitoring and early warning method based on multi-dimensional data perception according to claim 1, characterized in that: The variational autoencoder (VAE) includes an encoder and a decoder. The encoder is a 3-layer fully connected network that outputs latent abstract features. The decoder is also a 3-layer fully connected network that reconstructs the original features using the latent features. The gradient of the loss function with respect to the encoder and decoder parameters is calculated. An optimizer is used to update the parameters in reverse order to maintain the effectiveness of the latent features. The optimizer is the Adam optimizer, and the loss function includes the reconstruction error MSE and the KL divergence.

7. The high-voltage cable partial discharge monitoring and early warning method based on multi-dimensional data perception according to claim 1, characterized in that: The classification neural network includes an input layer, two hidden layers, and an output layer. The input layer is connected to a VAE encoder, the hidden layer activation function is the ReLU function, and the output layer is set to a softmax classifier.

8. A high-voltage cable partial discharge monitoring and early warning system based on multi-dimensional data perception, characterized in that, The method for monitoring and early warning of partial discharge in high-voltage cables based on multi-dimensional data perception as described in claim 1 includes: a sensor group, a preprocessing module, a type fusion module, and a deep learning network. The functions of each module are as follows: The sensor array acquires real-time partial discharge detection data of the detection area; The preprocessing module preprocesses the partial discharge detection data, removes environmental noise, and filters signal features to generate an initial set of signal features. The type fusion module performs splicing and dimensionality reduction processing on the signal features based on the partial discharge fault type, combining ultrasonic features and electrical signal features, and outputs the optimal feature set. A deep learning network is input with the optimal feature set to determine whether a partial discharge fault exists and the type of partial discharge fault.

Images