GIS partial discharge signal fault diagnosis method and system

Abstract

The invention discloses a GIS partial discharge signal fault diagnosis method and system. The method comprises the following steps: acquiring multi-modal data of a GIS partial discharge signal; wavelet transform is adopted to carry out noise reduction processing on the multi-modal data to obtain multi-modal data after noise reduction, and a continuous domain wavelet threshold noise reduction function is adopted in the wavelet transform; the multi-modal data after noise reduction is input to a junction association time domain convolution auto-encoder network, a reconstruction sequence of the multi-modal data is obtained, and the junction association time domain convolution auto-encoder network comprises a junction processing module, an encoder and a decoder which are connected in sequence. The junction processing module is used for constructing a junction signal adjacent to the multi-modal data in time after noise reduction, the encoder is used for converting the junction signal into a compressed representation, and the decoder is used for reconstructing the compressed representation into a reconstruction sequence of the multi-modal data; performing fault diagnosis on the GIS partial discharge signal based on the reconstruction sequence; according to the invention, the accuracy of GIS partial discharge signal monitoring and early warning probability is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment monitoring, and in particular to a GIS partial discharge signal fault diagnosis method and system. Background Art

[0002] Gas Insulated Substation (GIS) is one of the important equipment in the power system. It has the advantages of long service life and low maintenance, and has been widely used in many high-voltage hub substations at home and abroad. Insulation defect is a common defect of GIS equipment. The internal uncleanness of GIS, such as the residual free conductive particles (such as metal debris or metal particles) in GIS equipment, or accidental collision during transportation and installation, and poor quality of insulation parts, can cause internal electrical phenomena. This is a relatively common defect.

[0003] Traditional power outage preventive test detection methods mainly include insulation resistance test, dielectric loss and capacitance test and other detection methods. However, since the voltage applied during the power outage preventive test is only a few thousand volts, which is far less than the equipment operating voltage of hundreds of kilovolts, it is difficult to detect the latent insulation defects of the equipment. At the same time, the equipment is affected by factors such as electrical, magnetic, thermal and mechanical stress during operation, and the insulation state will deteriorate, and eventually the insulation function will be lost, resulting in accidents. The consequences are much more serious than those of separated open equipment, and the fault repair is more complicated and takes longer to repair. Therefore, the operation detection of GIS is very important. Not only does it need to conduct regular preventive tests carefully, but it should also develop GIS online monitoring technology to promptly detect various possible insulation abnormalities or fault signs for early intervention.

[0004] Due to the unique fully enclosed structure of GIS, it is difficult to measure its internal discharge signal externally. It is urgent to study new discharge sensing methods outside the box and promote online intelligent diagnosis technology. There are many advantages to using live detection and online monitoring of the insulation status of GIS equipment: there is no need to shut down the equipment, which reduces the probability of misoperation caused by equipment switching operations; the voltage is higher when the equipment is running, the detection of defects is more sensitive, and the reliability of the detection results is higher; the insulation status of the equipment can be monitored in real time. Therefore, it is possible to avoid GIS failures by timely small-scale maintenance before the failure occurs in a predictive manner.

[0005] In the related art, the patent application document with publication number CN115586406A proposed the use of volumetric convolutional neural networks to diagnose GIS local discharge faults. However, this solution only performs fault diagnosis on a single signal source of the discharge signal, and the fault diagnosis result is not accurate. In addition, this solution uses a deep volumetric convolutional neural network 3D CNN, which needs to expand the time series to the spatial series in the process of identifying the discharge signal. This will cause greater computational complexity and greater delay for online monitoring of GIS discharge signal faults.

[0006] In the literature "Study on GIS Partial Discharge Signal Denoising Method Based on Empirical Wavelet and Wavelet Transform, Qin Jinfei et al., High Voltage Electrical Appliances", it is proposed to study the denoising of GIS partial discharge signals based on the improved empirical wavelet transform (EWT) combined with the wavelet transform algorithm, select the effective components of the partial discharge signal according to the appropriate threshold and reconstruct the signal. The wavelet transform used in this scheme is improved on the basis of the empirical wavelet transform. The signal is decomposed into high-frequency and low-frequency modal functions through the empirical wavelet transform, and then the heuristic denoising is used for denoising. The denoised modal functions are divided according to the kurtosis value, and finally the EWT decomposition layer with small kurtosis is removed according to the appropriate threshold to reconstruct the signal and achieve denoising. The characteristic of heuristic denoising is that they usually determine the threshold based on the statistical characteristics of the signal or other heuristic rules. If the heuristic rule is not selected properly, it will lead to the risk of signal distortion or information loss. Summary of the invention

[0007] The technical problem to be solved by the present invention is how to improve the accuracy of GIS partial discharge signal monitoring and early warning probability.

[0008] The present invention solves the above technical problems by the following technical means:

[0009] The present invention proposes a GIS partial discharge signal fault diagnosis method, which comprises:

[0010] Collecting multimodal data of GIS partial discharge signals, wherein the multimodal data includes magnetic pulse data and acoustic pulse data;

[0011] Performing denoising on the multimodal data by using wavelet transform to obtain denoised multimodal data, wherein the wavelet transform uses a continuous domain wavelet threshold denoising function;

[0012] Inputting the denoised multimodal data into a knot-associated time-domain convolutional autoencoder network to obtain a reconstructed sequence of the multimodal data, wherein the knot-associated time-domain convolutional autoencoder network comprises a knot processing module, an encoder and a decoder connected in sequence, the knot processing module is used to construct a temporally adjacent joint signal of the denoised multimodal data, the encoder is used to convert the joint signal into a compressed representation, and the decoder is used to reconstruct the compressed representation into a reconstructed sequence of the multimodal data;

[0013] Fault diagnosis of GIS partial discharge signals is performed based on the reconstruction sequence.

[0014] Furthermore, the multimodal data of collecting GIS partial discharge signals includes:

[0015] Based on the pre-built GIS discharge defect structure approximate model, the high voltage electricity is adapted to simulate partial discharge faults and generate partial discharge signals;

[0016] Acoustic emission sensors and UHF sensors are used to collect acoustic pulse data and magnetic pulse data of partial discharge signals respectively.

[0017] Furthermore, the GIS discharge defect structure approximation model includes a basin-type insulation structure, and the defect models are fixed in sequence on the surface of the basin-type insulation structure close to the central conductor to form channels with different spacings; the metal particles at both ends of the channel are in close contact with the high potential terminal and the grounded central conductor respectively to form a discharge channel.

[0018] Furthermore, the adopting wavelet transform to perform noise reduction processing on the multimodal data to obtain the noise-reduced multimodal data includes:

[0019] Using wavelet transform to calculate the wavelet coefficients of the magnetic pulse data and the acoustic pulse data at each scale respectively;

[0020] Based on the constructed continuous domain wavelet threshold denoising function, the wavelet coefficients of each modal data are denoised respectively, and the denoised waveform scale coefficients corresponding to each modal data are obtained as the denoised multimodal data.

[0021] Furthermore, the formula of the continuous domain wavelet threshold denoising function is expressed as:

[0022]

[0023] Where d represents the acoustic pulse data s or magnetic pulse data c, W d (j, k) represents the coefficient of acoustic pulse data or magnetic pulse data at each scale, It represents the waveform scale coefficient of the acoustic pulse data or magnetic pulse data after denoising, γ represents the threshold of the wavelet coefficient, m represents the adjustment coefficient of the denoising threshold function, and sign() represents the sign function.

[0024] Furthermore, after performing denoising processing on the multimodal data by using wavelet transform to obtain denoised multimodal data, the method further includes:

[0025] Performing standardization processing on the denoised multimodal data to obtain standardized multimodal data;

[0026] Accordingly, the denoised multimodal data is input into the knot-correlated time-domain convolutional autoencoder network to obtain a reconstructed sequence of the multimodal data, specifically:

[0027] The standardized multimodal data is input into a knot-correlated time-domain convolutional autoencoder network to obtain a reconstructed sequence of the multimodal data.

[0028] Further, the encoder includes an input convolution layer Conv_1, a first TCN block, a convolution layer Conv_2, an average pooling layer and a GRU network connected in sequence;

[0029] The connection signal is used as the input of the input convolution layer Conv_1. The first TCN block captures the short-term and long-term patterns of the sequence of input connection signals. The convolution layer Conv_2 is used to reduce the dimension of the feature map. The average pooling layer is used to downsample the sequence along the time axis to compress the original input connection signal into a compact encoding representation sequence. The GRU network is used to extract the long-term dependencies in the encoding representation sequence to obtain a compressed representation of the connection signal.

[0030] Furthermore, the decoder includes an upsampling layer, a second TCN block and an output convolution layer Conv_3 connected in sequence, the upsampling layer is used to restore the compressed representation of the input connection signal to the length of the original input connection signal, the second TCN block is used to restore the characteristics of the compressed representation of the input connection signal, and the output convolution layer Conv_3 is used to restore the dimension of the compressed representation of the connection signal, and output a reconstructed sequence of multimodal data.

[0031] Furthermore, the fault diagnosis of GIS partial discharge signal based on the reconstruction sequence includes:

[0032] Calculating a reconstruction error based on the reconstructed sequence and the original sequence, and sliding a time window on the reconstruction error to smooth occasional noise and organize continuous error values ​​into an error matrix;

[0033] The error matrix is ​​reconstructed into a one-dimensional vector, and the one-dimensional vector is estimated, and the Mahalanobis distance used to measure the degree of deviation between the data point and the mean is calculated at each time point;

[0034] When the Mahalanobis distance is greater than a set fault threshold, determining that the GIS partial discharge signal is faulty;

[0035] When the Mahalanobis distance is less than or equal to the set fault threshold, it is determined that the GIS operates normally.

[0036] In addition, the present invention also proposes a GIS partial discharge signal fault diagnosis system, the system comprising:

[0037] An acquisition module, used for acquiring multimodal data of GIS partial discharge signals, wherein the multimodal data includes magnetic pulse data and acoustic pulse data;

[0038] A denoising module, configured to perform denoising on the multimodal data using wavelet transform to obtain denoised multimodal data, wherein the wavelet transform uses a continuous domain wavelet threshold denoising function;

[0039] A signal reconstruction module, used for inputting the denoised multimodal data into a knot-associated time-domain convolutional autoencoder network to obtain a reconstructed sequence of the multimodal data, wherein the knot-associated time-domain convolutional autoencoder network comprises a knot processing module, an encoder and a decoder connected in sequence, the knot processing module is used for constructing a temporally adjacent joint signal of the denoised multimodal data, the encoder is used for converting the joint signal into a compressed representation, and the decoder is used for reconstructing the compressed representation into a reconstructed sequence of the multimodal data;

[0040] A fault diagnosis module is used to perform fault diagnosis of GIS partial discharge signals based on the reconstruction sequence.

[0041] The advantages of the present invention are:

[0042] (1) The present invention first collects multimodal data of GIS local discharge signals. Different from the single signal source detection of traditional local discharge signals, the acoustic pulse data and the magnetic pulse data are connected. When the wavelet transform is used to perform noise reduction on the multimodal data, a continuous domain wavelet threshold noise reduction function is used. This helps to avoid the signal reconstruction oscillation caused by the discontinuity of traditional soft functions or hard functions, and helps to reduce misjudgment in the recognition process. In addition, an autoencoder structure is introduced into the knot-correlated time domain convolutional autoencoder network. By automatically learning the compressed representation and reconstruction of the data, the most representative features are learned from the original discharge signal without the need for additional manual feature engineering. This can more comprehensively capture the potential information and patterns in the input signal and more effectively capture the abstract features of the signal, thereby improving the performance and accuracy of discharge signal recognition, making the recognition of discharge signals more accurate and reliable.

[0043] (2) The present invention adopts a wavelet denoising method with an adjustable threshold. By adjusting the parameters, the constant deviation between the original signal and the reconstructed signal is reduced, so as to better adapt to noise-contaminated signals with different signal-to-noise ratios and achieve the optimal filtering effect. The correlation and fault characteristics of the connection signal can be better extracted during the network training process, thereby improving the diagnostic reliability of the connection time domain convolutional autoencoder network for fault signals.

[0044] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a flow chart of a GIS partial discharge signal fault diagnosis method proposed in one embodiment of the present invention;

[0046] Figure 2 It is a complete flow chart of a GIS partial discharge signal fault diagnosis method proposed in one embodiment of the present invention;

[0047] Figure 3 It is a structural schematic diagram of an approximate model of a GIS discharge defect structure constructed in one embodiment of the present invention;

[0048] Figure 4 is a schematic diagram of the structure of a knot-correlated time-domain convolutional autoencoder network constructed in one embodiment of the present invention;

[0049] Figure 5 It is a structural schematic diagram of a GIS partial discharge signal fault diagnosis system proposed in one embodiment of the present invention. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described in combination with the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0051] like Figure 1 to Figure 2 As shown, the first embodiment of the present invention proposes a GIS partial discharge signal fault diagnosis method, the method comprising the following steps:

[0052] S10, collecting multimodal data of GIS partial discharge signals, wherein the multimodal data includes magnetic pulse data and acoustic pulse data;

[0053] S20, performing denoising processing on the multimodal data by using wavelet transform to obtain denoised multimodal data, wherein the wavelet transform uses a continuous domain wavelet threshold denoising function;

[0054] This embodiment is different from the traditional single signal source detection of partial discharge signals. The acoustic pulse data and the magnetic pulse data are connected. When the wavelet transform is used to perform noise reduction on the multimodal data, a continuous domain wavelet threshold noise reduction function is used. This helps to avoid signal reconstruction oscillations caused by discontinuities of traditional soft functions or hard functions, and helps to reduce misjudgments in the recognition process.

[0055] S30, inputting the denoised multimodal data into a knot-associated time-domain convolutional autoencoder network to obtain a reconstructed sequence of the multimodal data, wherein the knot-associated time-domain convolutional autoencoder network comprises a knot processing module, an encoder and a decoder connected in sequence, the knot processing module is used to construct a temporally adjacent joint signal of the denoised multimodal data, the encoder is used to convert the joint signal into a compressed representation, and the decoder is used to reconstruct the compressed representation into a reconstructed sequence of the multimodal data;

[0056] S40, performing fault diagnosis of GIS partial discharge signals based on the reconstructed sequence.

[0057] It should be noted that existing studies generally use an acoustic emission or ultra-high frequency signal to identify GIS discharge signals, while GIS discharge signals are accompanied by multiple signals such as infrared, voltage, current, optical signals, etc. Judging the generation of a discharge signal by a single signal may limit the accuracy and reliability of fault diagnosis, and fail to fully understand the essence of the discharge phenomenon, which in turn affects the accurate assessment and prediction of the equipment status. This embodiment can connect multiple signals and extract features from the junction signal based on the feature extraction function of the neural network, which enables the model to learn more general feature representations, so that it can effectively identify signals under different conditions, and provide more diverse and rich feature representations, thereby improving the recognition accuracy and generalization ability of the model for discharge signals.

[0058] This embodiment introduces an autoencoder structure and a decoder structure in a knot-correlated time-domain convolutional autoencoder network, which has a deep time series learning capability, captures the key features of the input sequence through the encoder, and reconstructs these sequences as accurately as possible through the decoder. By automatically learning the compressed representation and reconstruction of the data, the most representative features are learned from the original discharge signal without the need for additional manual feature engineering, which can more comprehensively capture the potential information and patterns in the input signal and more effectively capture the abstract features of the signal, thereby improving the performance and accuracy of discharge signal recognition, making the recognition of discharge signals more accurate and reliable.

[0059] As a further preferred technical solution, in step S10, collecting multimodal data of GIS partial discharge signals includes the following steps:

[0060] S11, based on the pre-built GIS discharge defect structure approximate model, adapt the high voltage electric simulation partial discharge fault to generate partial discharge signal;

[0061] S12. Use an acoustic emission sensor and a UHF sensor to respectively collect acoustic pulse data and magnetic pulse data of partial discharge signals.

[0062] Specifically, Figure 3 As shown, the GIS discharge defect structure approximate model includes a basin-type insulation structure, and the defect models are fixed in sequence on the surface of the basin-type insulation structure close to the central conductor to form channels with different spacings; the metal particles at both ends of the channel are in close contact with the high potential terminal and the grounded central conductor respectively to form a discharge channel.

[0063] During use, the application adapts high voltage electricity to simulate the occurrence of partial discharge (PD) faults and uses acoustic emission sensors and ultra-high frequency sensors to collect acoustic pulse data and magnetic pulse data of discharge signals, and then inputs the signals into the processing terminal PC with a high-speed data acquisition card through a preamplifier.

[0064] This embodiment selects copper beads of different diameters and changes the interval of their placement distance, so it has strong operability, and can simulate metal particle accumulation defects for different environments and materials to obtain defect conditions of different degrees, thereby enhancing the network fault identification capability.

[0065] As a further preferred technical solution, the step S20: using wavelet transform to perform noise reduction processing on the multimodal data to obtain noise-reduced multimodal data includes the following steps:

[0066] S21, using wavelet transform to respectively calculate the wavelet coefficients of the magnetic pulse data and the acoustic pulse data at each scale;

[0067] Specifically, considering the short duration of partial discharge pulses, Db wavelet is selected as the wavelet basis, and the coefficient W of the acoustic pulse data at each scale in the Db wavelet transform is calculated. s (j, k) and the coefficient W of the magnetic pulse data at each scale c (j,k):

[0068]

[0069] Wherein, variable N is the number of sample points of the partial discharge signal, n is the time point of the partial discharge signal, and x is s (n) and xc (n) is the sample value of the original acoustic pulse data and magnetic pulse data, h j,k-n is the decomposition filter coefficient, j and k are the decomposition scale and translation parameters of the discrete wavelet transform, and · is the dot product symbol.

[0070] S22. Based on the constructed continuous domain wavelet threshold denoising function, the wavelet coefficients of each modal data are subjected to denoising respectively, and the denoised waveform scale coefficients corresponding to each modal data are obtained as the denoised multimodal data.

[0071] When filtering the collected magnetic-acoustic pulse signal data, this embodiment constructs a wavelet threshold denoising function that is continuous in the entire real number domain, which helps to avoid signal reconstruction oscillations caused by discontinuities of traditional soft functions or hard functions and helps to reduce misjudgments in the recognition process.

[0072] As a further preferred technical solution, the formula of the continuous domain wavelet threshold denoising function is expressed as:

[0073]

[0074] Where W s (j, k) represents the coefficient of the acoustic pulse data at each scale, W c (j,k) represents the coefficient of magnetic pulse data at each scale, It represents the waveform scale coefficient after the noise reduction of the acoustic pulse data. It represents the waveform scale coefficient of the magnetic pulse data after denoising, γ represents the threshold of the wavelet coefficient, m represents the adjustment coefficient of the denoising threshold function, and sign() represents the sign function.

[0075] Furthermore, the calculation formula of the threshold γ of each layer of wavelet coefficients is:

[0076]

[0077] Where the variable σ is the noise variance VAR.

[0078] This embodiment reconstructs the scale coefficient after noise reduction to obtain a good discharge signal magnetic acoustic-acoustic pulse data set.

[0079] Specifically, when m ≥ 1, as W s (j,k) or W c The increase of (j,k), With W s The constant deviation between (j,k) decreases or With W cThe constant deviation between (j, k) is reduced, and the accuracy of signal reconstruction is improved; when m is between 6 and 10, the denoising effect of signals with low signal-to-noise ratio is better; when m is between 3 and 5, the denoising effect of signals with medium and high signal-to-noise ratio is better. By adjusting m, the optimal threshold function for contaminated signals with different signal-to-noise ratios can be obtained.

[0080] This embodiment reduces the constant deviation between the original signal and the reconstructed signal by adjusting the parameters, so as to better adapt to noise-contaminated signals with different signal-to-noise ratios. The adaptability and flexibility of the wavelet denoising method with adjustable thresholds enable it to adjust the noise contamination of the magnetic-acoustic input pulse signal separately to achieve the optimal filtering effect, so that the correlation and fault characteristics of the connection signal can be better extracted during the network training process, thereby improving the diagnostic reliability of the network for fault signals. Moreover, the threshold function set can more effectively retain the original characteristics of the signal due to its continuity and parameter adjustability. Compared with heuristic threshold denoising, it can reduce the risk of signal distortion or information loss caused by improper selection of heuristic rules. By adjusting the adjustment coefficient of the denoising threshold function, the parameters are adjusted for the acoustic emission signal or the ultra-high frequency signal to select the optimal denoising threshold.

[0081] As a further preferred technical solution, in the step S20: after the multimodal data is subjected to denoising by wavelet transform to obtain the denoised multimodal data, the method further comprises the following steps:

[0082] Performing standardization processing on the denoised multimodal data to obtain standardized multimodal data;

[0083] Accordingly, the step S30: inputting the denoised multimodal data into the knot-correlated time-domain convolutional autoencoder network to obtain a reconstructed sequence of the multimodal data is specifically:

[0084] The standardized multimodal data is input into a knot-correlated time-domain convolutional autoencoder network to obtain a reconstructed sequence of the multimodal data.

[0085] It should be noted that, in this embodiment, two groups of discrete signal data x after noise reduction are processed. θ (n) Normalize it so that it is distributed in the activation area of ​​the network to avoid gradient disappearance.

[0086]

[0087] In the formula, mean is the mean operation, std is the variance operation, and when θ=1, x θ (n) represents the magnetic pulse data after noise reduction. When θ=2, x θ (n) represents the sound pulse data after noise reduction.

[0088] As a further preferred technical solution, Figure 4 As shown, the knot processing module is specifically used to: splice the standardized signals x1, x2, the dual-channel signals according to the sampling time, and obtain a three-dimensional knot signal matrix (N, L, C) adjacent in time with a single pulse period L, where C is the feature dimension, the value is 3, representing the signal frequency, amplitude and phase characteristics, and N is the number of samples.

[0089] As a further preferred technical solution, Figure 4 As shown, the encoder includes an input convolution layer Conv_1, a first TCN block, a convolution layer Conv_2, an average pooling layer and a GRU network connected in sequence;

[0090] The connection signal is used as the input of the input convolution layer Conv_1. The first TCN block captures the short-term and long-term patterns of the sequence of input connection signals. The convolution layer Conv_2 is used to reduce the dimension of the feature map. The average pooling layer is used to downsample the sequence along the time axis to compress the original input connection signal into a compact encoding representation sequence. The GRU network is used to extract the long-term dependencies in the encoding representation sequence to obtain a compressed representation of the connection signal.

[0091] Specifically, in this embodiment, the encoder accepts all the input knot signal sequences (L, C), represents the two-dimensional connection matrix of each sample, passes it through the first TCN block with 32 filters to capture the short-term and long-term patterns of the input sequence, applies a 1×1 convolution layer Conv_1 with 4 filters to reduce the dimension of the feature map, and uses an average pooling layer of size 4 to downsample the sequence along the time axis, compressing the original input C[n] into a compact encoding representation g[n]=enc(C[n]), where g: C[n], i.e. Connect[n], is the nth input connection signal sequence, enc() is the encoding processing function, is the c-dimensional field of real numbers.

[0092] Then, the time series and feature channels of each sample (N, L, C2) after encoding are flattened into a single, continuous long vector while maintaining the order of the samples. C2 is the dimension of the encoded signal feature. Subsequently, these vectors of all samples are reshaped (i.e., the tensor and matrix dimensions are reshaped) into a new dimension to form a new tensor. Finally, the dimension is adjusted to obtain the fused feature (1, n, l·C2).

[0093] The GRU (Gated Recurrent Unit) neural network is then used to extract long-term dependencies in the sequence, enhancing the model's ability to understand and utilize contextual information in the sequence. The GRU operation formula is as follows:

[0094]

[0095] In the formula, x t Enter information for the current moment, h t is the hidden state at the current time t, h t-1 is the hidden state output at the previous moment t-1, is the candidate hidden state, z t is the update gate, r t is the reset gate, W is the weight matrix, σ is the sigmoid activation function, tanh is the hyperbolic tangent activation function, and * represents the multiplication operation.

[0096] As a further preferred technical solution, the decoder includes an upsampling layer, a second TCN block and an output convolution layer Conv_3 connected in sequence, the upsampling layer is used to restore the compressed representation of the input connection signal to the length of the original input connection signal, the second TCN block is used to restore the characteristics of the compressed representation of the input connection signal, and the output convolution layer Conv_3 is used to restore the dimension of the compressed representation of the connection signal and output a reconstructed sequence of multimodal data.

[0097] Specifically, this embodiment uses the sample-and-hold interpolation method to restore the length of the original sequence through the upsampling layer. The sampled sequence is processed by the second TCN block, which also has 32 filters and has a structure similar to the encoder but with independent weights, to restore the features of the input sequence. Finally, a 1×1 convolutional layer with 2 filters and the same dimension as the original is used to restore the dimension to the original state and obtain the reconstructed sequence.

[0098] It should be noted here that TCN (Temporal Convolutional Network) is a neural network architecture designed specifically for processing time series data. It is a deep learning model, commonly used for sequence modeling tasks, such as time series prediction, speech recognition, natural language processing, etc. In the discharge signal detection of GIS, TCN effectively extracts valuable features from the complex discharge signal time series through a deep convolutional network structure, such as the repeatability of the discharge pattern and the change of the discharge intensity. However, in the discharge signal recognition, the traditional TCN usually requires manual design and selection of features in the process of labeling the discharge signal data set, and often has problems such as large data volume, high dimensionality, and label imbalance. The first TCN block and the second TCN block designed in this embodiment can optimize the calculation process through their one-dimensional convolution structure, especially by using dilated convolution, while maintaining a low computational cost while covering a wide range of input sequences. This efficient computing feature makes the TCN block very suitable for GIS online fault detection that requires fast response.

[0099] Therefore, the ConnectRelevance Temporal Convolutional Networks-Auto-encoder (CRTCN-AE) designed in this embodiment introduces an autoencoder structure, which learns the most representative features from the original discharge signal by automatically learning the compressed representation and reconstruction of the data without the need for additional manual feature engineering, and more effectively captures the abstract features of the signal, thereby improving the performance and accuracy of discharge signal recognition. Specifically, CRTCN-AE converts the discharge signal into a compressed representation through the encoder part, and then reconstructs the compressed representation into the original signal in the decoder part. In this process, the model can more comprehensively capture the potential information and patterns in the input signal, making the recognition of the discharge signal more accurate and reliable. In contrast, the traditional TCN simply extracts features through operations such as convolution and pooling, and may not be able to fully explore the deep information in the signal. The application of CRTCN-AE in discharge signal recognition highlights the advantages of the autoencoder, and improves the accuracy and robustness of signal recognition by more effectively learning and representing the features of the signal, and has better practicality and application prospects.

[0100] As a further preferred technical solution, the step S40: performing fault diagnosis of GIS partial discharge signals based on the reconstructed sequence comprises the following steps:

[0101] S41, calculating a reconstruction error based on the reconstructed sequence and the original sequence, and sliding a time window on the reconstruction error to smooth occasional noise and organize continuous error values ​​into an error matrix;

[0102] Specifically, this embodiment calculates the reconstruction error e[n] based on the reconstructed sequence and the original sequence, and slides a window of length δ on the reconstruction error to smooth occasional noise and organize continuous error values ​​into an error matrix E[n].

[0103] S42, reconstructing the error matrix into a one-dimensional vector, estimating the one-dimensional vector, and calculating the Mahalanobis distance for measuring the degree of deviation between the data point and the mean at each time point;

[0104] Specifically, this embodiment analyzes the error matrix E[n] to reconstruct the error into a one-dimensional vector E′[n], estimates E′[n] and calculates the Mahalanobis distance M[n] for each time point to measure the degree of deviation between the data point and the mean:

[0105]

[0106] Where μ is the mean of E′[n], ∑ is the covariance matrix of E′[n], and T is the transposed symbol.

[0107] S43, when the Mahalanobis distance is greater than a set fault threshold, determining that a GIS partial discharge signal fault occurs;

[0108] S44: When the Mahalanobis distance is less than or equal to a set fault threshold, it is determined that the GIS operates normally.

[0109] It should be noted that, in this embodiment, the calculated Mahalanobis distance M[n] and the given abnormal threshold M τ , compare and identify data points that are significantly different from the normal pattern, and finally output the abnormal mark a[n] for each time series point to identify the abnormal pattern of magnetic-acoustic pulses in the discharge signal.

[0110] In this embodiment, the abnormal behavior of GIS local discharge is identified by analyzing the reconstruction error and applying the local outlier factor anomaly detection algorithm, thereby identifying the abnormal pattern in the acoustic emission of the discharge signal without the need for labels, which can improve the accuracy of the predicted GIS local discharge signal monitoring and early warning probability.

[0111] In addition, if Figure 5 As shown, the second embodiment of the present invention further proposes a GIS partial discharge signal fault diagnosis system, characterized in that the system comprises:

[0112] An acquisition module 10 is used to acquire multimodal data of GIS partial discharge signals, wherein the multimodal data includes magnetic pulse data and acoustic pulse data;

[0113] A denoising module 20, configured to perform denoising on the multimodal data using wavelet transform to obtain denoised multimodal data, wherein the wavelet transform uses a continuous domain wavelet threshold denoising function;

[0114] A signal reconstruction module 30 is used to input the denoised multimodal data into a knot-associated time-domain convolutional autoencoder network to obtain a reconstructed sequence of the multimodal data, wherein the knot-associated time-domain convolutional autoencoder network includes a knot processing module, an encoder and a decoder connected in sequence, the knot processing module is used to construct a temporally adjacent joint signal of the denoised multimodal data, the encoder is used to convert the joint signal into a compressed representation, and the decoder is used to reconstruct the compressed representation into a reconstructed sequence of the multimodal data;

[0115] The fault diagnosis module 40 is used to perform fault diagnosis of GIS partial discharge signals based on the reconstruction sequence.

[0116] As a further preferred technical solution, the system also pre-builds a GIS discharge defect structure approximate model for adapting high-voltage electrical simulation of partial discharge faults and generating partial discharge signals. The GIS discharge defect structure approximate model includes a basin-type insulation structure, and the defect model is sequentially fixed on the surface of the basin-type insulation structure close to the central conductor to form channels with different spacings; the metal particles at both ends of the channel are in close contact with the high-potential terminal and the grounded central conductor, respectively, to form a discharge channel. The acoustic pulse data and magnetic pulse data of the partial discharge signal are collected by using an acoustic emission sensor and a UHF sensor, and then the signal is input into a processing terminal PC with a high-speed data acquisition card through a preamplifier.

[0117] As a further preferred technical solution, the noise reduction module 20 specifically includes:

[0118] A wavelet decomposition unit, used for respectively calculating the wavelet coefficients of the magnetic pulse data and the acoustic pulse data at each scale by using wavelet transform;

[0119] The denoising unit is used to perform denoising on the wavelet coefficients of each modal data based on the constructed continuous domain wavelet threshold denoising function, and obtain the denoised waveform scale coefficient corresponding to each modal data as the denoised multimodal data.

[0120] As a further preferred technical solution, the system further includes:

[0121] A normalization module is used to perform normalization processing on the denoised multimodal data to obtain standardized multimodal data;

[0122] Correspondingly, the denoising module 20 is specifically configured to input the standardized multimodal data into a knot-correlated time-domain convolutional autoencoder network to obtain a reconstructed sequence of the multimodal data.

[0123] As a further preferred technical solution, the encoder includes an input convolution layer Conv_1, a first TCN block, a convolution layer Conv_2, an average pooling layer and a GRU network connected in sequence;

[0124] The connection signal is used as the input of the input convolution layer Conv_1. The first TCN block captures the short-term and long-term patterns of the sequence of input connection signals. The convolution layer Conv_2 is used to reduce the dimension of the feature map. The average pooling layer is used to downsample the sequence along the time axis to compress the original input connection signal into a compact encoding representation sequence. The GRU network is used to extract the long-term dependencies in the encoding representation sequence to obtain a compressed representation of the connection signal.

[0125] As a further preferred technical solution, the decoder includes an upsampling layer, a second TCN block and an output convolution layer Conv_3 connected in sequence, the upsampling layer is used to restore the compressed representation of the input connection signal to the length of the original input connection signal, the second TCN block is used to restore the characteristics of the compressed representation of the input connection signal, and the output convolution layer Conv_3 is used to restore the dimension of the compressed representation of the connection signal and output a reconstructed sequence of multimodal data.

[0126] As a further preferred technical solution, the fault diagnosis module 40 specifically includes:

[0127] an error matrix construction unit, for calculating a reconstruction error based on the reconstruction sequence, and sliding a time window on the reconstruction error to smooth occasional noise and organize continuous error values ​​into an error matrix;

[0128] A Mahalanobis distance calculation unit is used to reconstruct the error matrix into a one-dimensional vector, estimate the one-dimensional vector, and calculate the Mahalanobis distance at each time point for measuring the degree of deviation between the data point and the mean;

[0129] The GIS state determination unit is used to determine that the GIS partial discharge signal is faulty when the Mahalanobis distance is greater than a set fault threshold; and to determine that the GIS is operating normally when the Mahalanobis distance is less than or equal to the set fault threshold.

[0130] It should be noted that other embodiments or implementation methods of the GIS partial discharge signal fault diagnosis system of the present invention can refer to the above-mentioned method embodiments, which will not be repeated here.

[0131] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0132] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0133] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A GIS partial discharge signal fault diagnosis method, characterized in that: The method comprises: Collecting multimodal data of GIS partial discharge signals, wherein the multimodal data includes magnetic pulse data and acoustic pulse data; Performing denoising on the multimodal data by using wavelet transform to obtain denoised multimodal data, wherein the wavelet transform uses a continuous domain wavelet threshold denoising function; Inputting the denoised multimodal data into a knot-associated time-domain convolutional autoencoder network to obtain a reconstructed sequence of the multimodal data, wherein the knot-associated time-domain convolutional autoencoder network comprises a knot processing module, an encoder and a decoder connected in sequence, the knot processing module is used to construct a temporally adjacent joint signal of the denoised multimodal data, the encoder is used to convert the joint signal into a compressed representation, and the decoder is used to reconstruct the compressed representation into a reconstructed sequence of the multimodal data; Fault diagnosis of GIS partial discharge signals is performed based on the reconstruction sequence.

2. The GIS partial discharge signal fault diagnosis method according to claim 1, characterized in that: The multi-modal data of collecting GIS partial discharge signals includes: Based on the pre-built GIS discharge defect structure approximate model, the high voltage electricity is adapted to simulate partial discharge faults and generate partial discharge signals; Acoustic emission sensors and UHF sensors are used to collect acoustic pulse data and magnetic pulse data of partial discharge signals respectively.

3. The GIS partial discharge signal fault diagnosis method according to claim 2, characterized in that: The GIS discharge defect structure approximate model includes a basin-type insulation structure, and the defect models are fixed in sequence on the surface of the basin-type insulation structure close to the central conductor to form channels with different spacings; the metal particles at both ends of the channel are in close contact with the high potential terminal and the grounded central conductor respectively to form a discharge channel.

4. The GIS partial discharge signal fault diagnosis method according to claim 1, characterized in that: The adopting wavelet transform to perform noise reduction processing on the multimodal data to obtain the noise-reduced multimodal data includes: Using wavelet transform to calculate the wavelet coefficients of the magnetic pulse data and the acoustic pulse data at each scale respectively; Based on the constructed continuous domain wavelet threshold denoising function, the wavelet coefficients of each modal data are denoised respectively, and the denoised waveform scale coefficients corresponding to each modal data are obtained as the denoised multimodal data.

5. The GIS partial discharge signal fault diagnosis method according to claim 4, characterized in that: The formula of the continuous domain wavelet threshold denoising function is expressed as: Where d represents the acoustic pulse data s or magnetic pulse data c, W d (j, k) represents the coefficient of acoustic pulse data or magnetic pulse data at each scale, It represents the waveform scale coefficient of the acoustic pulse data or magnetic pulse data after denoising, γ represents the threshold of the wavelet coefficient, m represents the adjustment coefficient of the denoising threshold function, and sign() represents the sign function.

6. The GIS partial discharge signal fault diagnosis method according to claim 1, characterized in that: After the multimodal data is subjected to denoising by wavelet transform to obtain denoised multimodal data, the method further includes: Performing standardization processing on the denoised multimodal data to obtain standardized multimodal data; Accordingly, the denoised multimodal data is input into the knot-correlated time-domain convolutional autoencoder network to obtain a reconstructed sequence of the multimodal data, specifically: The standardized multimodal data is input into a knot-correlated time-domain convolutional autoencoder network to obtain a reconstructed sequence of the multimodal data.

7. The GIS partial discharge signal fault diagnosis method according to claim 1, characterized in that: The encoder includes an input convolution layer Conv_1, a first TCN block, a convolution layer Conv_2, an average pooling layer and a GRU network connected in sequence; The connection signal is used as the input of the input convolution layer Conv_1. The first TCN block captures the short-term and long-term patterns of the sequence of input connection signals. The convolution layer Conv_2 is used to reduce the dimension of the feature map. The average pooling layer is used to downsample the sequence along the time axis to compress the original input connection signal into a compact encoding representation sequence. The GRU network is used to extract the long-term dependencies in the encoding representation sequence to obtain a compressed representation of the connection signal.

8. The GIS partial discharge signal fault diagnosis method according to claim 1, characterized in that: The decoder includes an upsampling layer, a second TCN block and an output convolution layer Conv_3 connected in sequence, the upsampling layer is used to restore the compressed representation of the input connection signal to the length of the original input connection signal, the second TCN block is used to restore the characteristics of the compressed representation of the input connection signal, and the output convolution layer Conv_3 is used to restore the dimension of the compressed representation of the connection signal and output a reconstructed sequence of multimodal data.

9. The GIS partial discharge signal fault diagnosis method according to claim 1, characterized in that: The fault diagnosis of GIS partial discharge signal based on the reconstruction sequence includes: Calculating a reconstruction error based on the reconstructed sequence and the original sequence, and sliding a time window on the reconstruction error to smooth occasional noise and organize continuous error values ​​into an error matrix; The error matrix is ​​reconstructed into a one-dimensional vector, and the one-dimensional vector is estimated, and the Mahalanobis distance used to measure the degree of deviation between the data point and the mean at each time point is calculated; When the Mahalanobis distance is greater than a set fault threshold, determining that the GIS partial discharge signal is faulty; When the Mahalanobis distance is less than or equal to the set fault threshold, it is determined that the GIS operates normally.

10. A GIS partial discharge signal fault diagnosis system, characterized in that: The system comprises: An acquisition module, used for acquiring multimodal data of GIS partial discharge signals, wherein the multimodal data includes magnetic pulse data and acoustic pulse data; A denoising module, configured to perform denoising on the multimodal data using wavelet transform to obtain denoised multimodal data, wherein the wavelet transform uses a continuous domain wavelet threshold denoising function; A signal reconstruction module, used for inputting the denoised multimodal data into a knot-associated time-domain convolutional autoencoder network to obtain a reconstructed sequence of the multimodal data, wherein the knot-associated time-domain convolutional autoencoder network comprises a knot processing module, an encoder and a decoder connected in sequence, the knot processing module is used for constructing a temporally adjacent joint signal of the denoised multimodal data, the encoder is used for converting the joint signal into a compressed representation, and the decoder is used for reconstructing the compressed representation into a reconstructed sequence of the multimodal data; A fault diagnosis module is used to perform fault diagnosis of GIS partial discharge signals based on the reconstruction sequence.

Images