Power System Fault Diagnosis Method Based on MP-Convformer Parallel Network
The power system transient signal fault diagnosis method using MP-Convformer parallel network extracts local and global features of power system transient signals by utilizing wavelet denoising, inverted residual module and ConvFormer module, which solves the problem of difficult identification of composite disturbance signals and realizes accurate fault diagnosis and detection.
Patent Information
- Application Number
- CN202410786665.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-18
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2044-06-18
AI Technical Summary
Existing signal processing methods are difficult to effectively handle transient signals with complex disturbances in power systems. In particular, due to the rapid change of signals, strong noise interference, and coupling between different transient signal characteristics, it is difficult to accurately identify the fault type.
A fault diagnosis method for transient signals in power systems based on MP-Convformer parallel network is adopted. The method involves wavelet denoising preprocessing, extraction of local features by inverted residual module, extraction of global features by ConvFormer module, and fault classification by feature fusion module and Softmax classifier.
It enables accurate identification and detection of transient signal faults in power systems, reduces noise interference, improves fault feature representation capabilities, and has good stability and generalization ability.
Smart Images

Figure CN118410415B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of power system transient signal fault identification and detection, and specifically to a power system transient signal fault diagnosis method and system based on an MP-Convformer parallel network. Background Art
[0002] The analysis of power quality disturbance signals has become an important task in power system status monitoring and maintenance. Due to the widespread use of distributed power generation systems such as large-scale photovoltaic power generation, hydropower generation, and various nonlinear loads such as new energy vehicle charging piles and industrial drives, the power grid signals are becoming more and more complex. These distributed power generation systems and nonlinear loads connected to the power grid will input disturbance signals to the power grid, resulting in power quality degradation such as waveform distortion of the grid voltage and current, which will pose a serious threat to the safe and stable operation of the power system and the surrounding electrical environment. Therefore, it is of great significance to identify and classify transient disturbance signals in the power system.
[0003] Nowadays, transient signal disturbances in power grids are not all single basic disturbances. Most of them are composite power quality disturbances composed of a mixture of basic disturbances of different disturbance types, different disturbance intensities, and different start and end times. There may be overlaps and complex crossovers between the time-frequency domain features of these composite power quality disturbance characteristics, which is also the difficulty faced by the current power quality disturbance identification.
[0004] In addition, transient signals have the characteristics of short duration, gradual attenuation over time, fast change speed, and unclear periodicity, which causes transient signals to be affected by attenuation and distortion of lines and equipment during transmission, causing the waveform and amplitude of the signal to change. Attenuation and distortion cause the loss of characteristic information of the signal, increasing the difficulty of signal identification and detection.
[0005] With the development of computer simulation technology, many methods for extracting power quality disturbance signal features based on physical characteristics have emerged, such as fast Fourier transform (FFT), wavelet transform (WT) and Hilbert-Huang transform (HHT). However, these traditional methods have some shortcomings that are difficult to overcome. For example, FFT is only applicable to stationary signals, and signal acquisition must satisfy the sampling theorem, otherwise the analysis results will show spectrum aliasing effects; WT has limitations in the selection of wavelet basis, and there is no wavelet basis that can be applied to all situations; HHT can handle nonlinear and non-stationary signals well, and has better universality than the first two, but there are problems such as modal confusion.
[0006] In view of the problems of fast changing speed, strong noise interference, and coupling between different transient signal features in power systems, it is difficult to obtain the global context information of transient signals by simply extracting single-scale features and local feature information of transient signals, which makes it difficult to detect faults in time and accurately identify fault types. Due to the complexity of transient signal disturbance types and disturbance feature types, traditional signal processing methods are difficult to deal with such complex disturbances in power systems. Summary of the invention
[0007] The present application provides a method and system for diagnosing transient signal faults in a power system based on an MP-Convformer parallel network, so as to solve the technical problem that existing signal processing methods are difficult to process composite disturbance information.
[0008] The present application provides a method for diagnosing transient signal faults in a power system based on an MP-Convformer parallel network, which includes:
[0009] Collect data on transient signals under abnormal conditions;
[0010] Preprocess the transient signal using wavelet denoising and obtain the preprocessed signal;
[0011] The convolution operation in the inverted residual module is used to upsample the preprocessed signal and extract local feature information;
[0012] The channel attention mechanism embedded in the inverted residual module is used to integrate the feature information of different channels;
[0013] The ConvFormer module is used to extract global feature information at different scales under different paths;
[0014] Based on the feature fusion module, local feature information and global feature information are fused;
[0015] The classifier based on Softmax function is used to classify the faults of transient signals in power system.
[0016] Optionally, the step of preprocessing the transient signal by using wavelet denoising and obtaining the preprocessed signal comprises the following steps:
[0017] Based on orthogonal wavelet decomposition, the noisy transient signal is locally decomposed in the time domain and frequency domain. After orthogonal wavelet decomposition, the wavelet coefficients with larger amplitudes are denoised signals, while the coefficients with smaller amplitudes are noise signals.
[0018] Determine whether the wavelet coefficient is greater than the preset threshold. If so, keep the wavelet coefficient. If not, replace the wavelet coefficient with the predicted variable threshold. The formula for processing the wavelet coefficient using the soft threshold function is:
[0019] σ λ s ( d ) = { [ sgn ( d ) ( | d | − λ ) ] , | d | > λ 0 , | d | ≤ λ } ,
[0020] and Respectively represent the wavelet transform coefficients before and after denoising, is the sign function, λ is the preset threshold, and s represents the number of wavelet decomposition layers;
[0021] The processed wavelet coefficients are restored using inverse wavelet transform.
[0022] Optionally, in the step of determining whether the wavelet coefficient is greater than a preset threshold, the predicted variable threshold is calculated according to the following formula:
[0023] ,
[0024] crit = [ log ( n ) / log 2 ] 1 . 5 n ,
[0025] ,
[0026] Where n is the length of the transient signal x, is the threshold of the universal fixed threshold rule, eta is the universal fixed threshold rule, is the threshold of the unbiased likelihood estimation rule, crit is the unbiased likelihood estimation rule, and λ is the predicted variable threshold.
[0027] Optionally, in the step of restoring the processed wavelet coefficients by using inverse wavelet transform, wavelet reconstruction is used to reduce noise interference, and the wavelet reconstruction formula is:
[0028] ,
[0029] ,
[0030] ,
[0031] Among them, j is the number of wavelet decomposition layers, k is the number of discrete sampling points, and represents the low-pass filter bank coefficients, and represents the high-pass filter bank coefficients, and are the scale coefficient and the wavelet coefficient after filtering respectively, and m represents the filter index parameter.
[0032] Optionally, the step of upsampling the preprocessed signal and extracting local feature information by using the convolution operation in the inverse residual module comprises the following steps:
[0033] Use depth-wise separable convolution with a kernel size of 3 to upsample and extract local feature information;
[0034] The dimension of the preprocessed signal is changed using a point convolution operation with a kernel size of 1.
[0035] Optionally, the step of integrating feature information of different channels by using a channel attention mechanism embedded in the inverted residual module comprises the following steps:
[0036] The feature information of different channels is integrated using the channel attention mechanism, which is embedded between two standard convolutional layers with a convolution kernel size of 1X1;
[0037] Use point convolution operation with kernel size 1 to restore the dimension of preprocessed signal data;
[0038] The preprocessed signal after convolution is restored using depthwise separable convolution with a kernel size of 3.
[0039] Optionally, in extracting global feature information at different scales under different paths using a ConvFormer module, the ConvFormer module includes a multi-layer convolution module and a multi-head attention convolution module;
[0040] The multi-layer convolution module includes three convolution layers with a convolution kernel size of 3 stacked in sequence to extract deeper features of the preprocessed signal;
[0041] The multi-head attention convolution module includes three convolution layers with convolution kernels of different scales, a multi-head attention mechanism, a normalization layer and a feedforward network layer.
[0042] Optionally, the multi-head attention convolution module performs the following steps:
[0043] Three convolution layers with convolution kernels of different scales are used to extract fault features of preprocessed signals of different scales. The convolution kernel sizes of the three convolution layers are 3, 5, and 7 respectively.
[0044] Divide the input preprocessed signal into three vectors of the same dimension, which are a query vector q, a key vector k and a value vector v;
[0045] Encapsulate the three vectors into three matrices, namely query Q, value K, and value vector V;
[0046] Calculate the dot product operation between the matrix Q and the value K to get the similarity between the two;
[0047] Normalization is performed through Softmax operation;
[0048] The performance of the self-attention layer is improved by weighting the normalized weights and the value vector V to give the attention layer a different representation subspace. The attention mechanism calculation formula is as follows:
[0049] ,
[0050] Among them, Q represents query, K represents value, and V represents value vector. Represents the scaling factor.
[0051] Optionally, in the step of fusing local feature information and global feature information based on the feature fusion module, the calculation formula of the feature fusion module is:
[0052] A i = Concat ( [ L i , G i , 0 , G i , 1 ,..., G i , n ] ) ,
[0053] Among them, G in represents the output of the multi-head attention convolution module, L i Represents the output of the multi-layer convolution module, A i represents the aggregated features, i represents the i-th category feature, and n represents the n-th parallel Convformer module output;
[0054] In the step of classifying the fault of the transient signal of the power system by the classifier based on the Softmax function, the calculation formula of the Softmax function is:
[0055] ,
[0056] in, represents the score of the fth category, and k represents the total number of categories.
[0057] Correspondingly, the present application also provides a method and system for diagnosing transient signal faults in a power system based on an MP-Convformer parallel network, which includes a memory and a processor, the memory being used to store executable program codes; the processor being connected to the memory, and running a computer program corresponding to the executable program codes by reading the executable program codes, so as to execute any of the steps in the above-mentioned method for diagnosing transient signal faults in a power system based on an MP-Convformer parallel network.
[0058] The present application provides a method and system for diagnosing transient signal faults in a power system based on an MP-Convformer parallel network. The method and system for diagnosing transient signal faults in a power system based on an MP-Convformer parallel network determine the selection of different types of wavelet transform thresholds in the preprocessing stage by comparing the relative sizes of the measurement indicators of a fixed threshold rule and an unbiased likelihood estimation rule, thereby removing noise information to the greatest extent and retaining fault signal characteristics.
[0059] The inverted residual module uses depthwise separable convolution instead of ordinary convolution operations, which can effectively reduce the number of model parameters. At the same time, in order to focus on the features on important channels, a channel attention mechanism is added to the structure, which can effectively improve the feature representation ability of the model and make the model more lightweight.
[0060] The multi-head attention convolution structure in the ConvFormer module can be used to model different fine-grained features, enrich the fault feature representation of transient signals, and combine CNN with the multi-head attention mechanism in a complementary way, taking into account both the local and global features of transient signals, thereby achieving a more accurate understanding and judgment of classification tasks. BRIEF DESCRIPTION OF THE DRAWINGS
[0061] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.
[0062] Figure 1 It is a flow chart of a method for diagnosing transient signal faults in a power system based on an MP-Convformer parallel network provided by the present application;
[0063] Figure 2 It is a flow chart of step S200 in the method for diagnosing transient signal faults in a power system based on an MP-Convformer parallel network provided by the present application;
[0064] Figure 3 It is a flowchart of step S300 and step S400 in the method for diagnosing transient signal faults in a power system based on an MP-Convformer parallel network provided by the present application;
[0065] Figure 4 It is a flow chart of step S500 in the method for diagnosing transient signal faults in a power system based on an MP-Convformer parallel network provided in the present application. DETAILED DESCRIPTION
[0066] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present application. In addition, it should be understood that the specific implementation methods described herein are only used to illustrate and explain the present application and are not used to limit the present application. In the present application, unless otherwise stated, the directional words used, such as "up", "down", "left", and "right", generally refer to the up, down, left, and right of the device in actual use or working state, specifically the drawing direction in the accompanying drawings.
[0067] The present application provides a method and system for diagnosing transient signal faults in a power system based on an MP-Convformer parallel network, which are described in detail below. It should be noted that the description order of the following embodiments is not intended to limit the preferred order of the embodiments of the present application. In the following embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.
[0068] See also Figure 1-Figure 4 The present application provides a method for diagnosing transient signal faults in a power system based on an MP-Convformer parallel network. The method combines the advantages of CNN in extracting local features and self-attention mechanism in capturing global dependencies in a complementary manner, and simultaneously extracts local features at different scales of transient signals and global features on different paths; then the extracted local features and global features are fused, which further enriches the representation of fault features and realizes the identification and detection of transient signal faults in the power system.
[0069] See also Figure 1-Figure 4 , a power system transient signal fault diagnosis method based on MP-Convformer parallel network specifically includes the following steps:
[0070] S100, collecting data of transient signals under abnormal conditions;
[0071] S200, preprocessing the transient signal by using wavelet denoising, and obtaining the preprocessed signal; in this application, the data samples are randomly divided into a training set, a validation set, and a test set in a ratio of 7:2:1;
[0072] Wavelet decomposition can effectively separate signals from noise, so this application uses the most widely used and fastest nonlinear wavelet transform threshold denoising method. The denoising principle is: first, collect the transient measured signal of the power system; secondly, the orthogonal wavelet decomposition locally decomposes the measured signal containing noise in the time domain and frequency domain. After the orthogonal wavelet decomposition, the wavelet coefficients with relatively large amplitudes are mostly denoised signals, while the coefficients with smaller amplitudes are generally noise signals; then find a reasonable threshold, retain the wavelet coefficients greater than the threshold, and specially process the wavelet coefficients less than the threshold; finally, restore the processed wavelet coefficients to the denoised signal through the inverse wavelet transform.
[0073] Step S200 specifically includes the following steps:
[0074] S210, locally decomposing the noisy transient signal in the time domain and the frequency domain based on orthogonal wavelet decomposition, wherein the wavelet coefficients with relatively large amplitudes after orthogonal wavelet decomposition are denoised signals, and the coefficients with relatively small amplitudes are noise signals;
[0075] S220, determine whether the wavelet coefficient is greater than a preset threshold, if so, retain the wavelet coefficient, if not, replace the wavelet coefficient with the predicted variable threshold, wherein the formula for processing the wavelet coefficient using the soft threshold function is:
[0076] σ λ s ( d ) = { [ sgn ( d ) ( | d | − λ ) ] , | d | > λ 0 , | d | ≤ λ } ,
[0077] and Respectively represent the wavelet transform coefficients before and after denoising, is the sign function, λ is the preset threshold, and s represents the number of wavelet decomposition layers;
[0078] The key to the wavelet transform threshold denoising method lies in the quantization processing and selection of the threshold. Commonly used threshold functions include soft threshold function and hard threshold function. In order to ensure the extraction accuracy of the noise signal, this application uses the soft threshold function to process the wavelet coefficients.
[0079] S230, restoring the processed wavelet coefficients by using inverse wavelet transform;
[0080] In step S220, the predicted variable threshold is calculated according to the following formula:
[0081] ,
[0082] crit = [ log ( n ) / log 2 ] 1 . 5 n ,
[0083] ,
[0084] Where n is the length of the transient signal x, is the threshold of the universal fixed threshold rule, eta is the universal fixed threshold rule, is the threshold of the unbiased likelihood estimation rule, crit is the unbiased likelihood estimation rule, and λ is the predicted variable threshold.
[0085] In the threshold processing function, the selection of the threshold λ will directly affect the denoising effect. Generally, when the signal-to-noise ratio is large, the unbiased likelihood estimation rule is usually used, and when the signal-to-noise ratio is small, the fixed threshold rule is usually used. Therefore, the heuristic threshold rule adopted in this application comprehensively utilizes the fixed threshold rule and the unbiased likelihood estimation rule to adaptively select the variable threshold for optimal prediction.
[0086] In step S230, wavelet reconstruction is used to reduce noise interference, which can effectively reduce noise interference. The wavelet reconstruction formula is:
[0087] ,
[0088] ,
[0089] ,
[0090] Among them, j is the number of wavelet decomposition layers, k is the number of discrete sampling points, and represents the low-pass filter bank coefficients, and represents the high-pass filter bank coefficients, and are the scale coefficient and the wavelet coefficient after filtering respectively, and m represents the filter index parameter.
[0091] By comparing the relative sizes of the metrics of the fixed threshold rule and the unbiased likelihood estimation rule, it is decided to select different types of wavelet transform thresholds in the preprocessing stage, so as to remove noise information to the greatest extent and retain the fault signal characteristics.
[0092] S300, upsampling the preprocessed signal and extracting local feature information by using the convolution operation in the inverse residual module;
[0093] S400, using the channel attention mechanism embedded in the inverted residual module to integrate feature information of different channels;
[0094] In view of the characteristics of fast changing speed and short duration of transient signal faults in power systems, this application designs an inverted residual module, which is mainly composed of two convolution blocks with different convolution kernels and a channel attention mechanism using residual connection, so that the original feature information can be effectively reused and the loss of useful feature information can be reduced.
[0095] Step S300 specifically includes the following steps:
[0096] S310, performing upsampling and extracting local feature information using a depthwise separable convolution with a convolution kernel size of 3;
[0097] Replacing standard convolutions with depthwise separable convolutions with a kernel size of 3 can reduce the number of model parameters.
[0098] S320, changing the dimension of the preprocessed signal by using a point convolution operation with a convolution kernel size of 1;
[0099] Step S400 specifically includes the following steps:
[0100] S410, integrating feature information of different channels using a channel attention mechanism, wherein the channel attention mechanism is embedded between two standard convolution layers with a convolution kernel size of 1×1;
[0101] By adding a channel attention mechanism between two standard convolutional layers with a convolution kernel size of 1X1, the role of specific channels can be selectively enhanced or weakened to extract more discriminative features.
[0102] S420, restoring the dimension of the preprocessed signal data by using a point convolution operation with a convolution kernel size of 1;
[0103] S430, using a depthwise separable convolution with a convolution kernel size of 3 to restore the preprocessed signal after the convolution processing;
[0104] In addition, in order to prevent the gradient explosion problem, the ReLU6 activation function is used instead of the ordinary ReLU activation function in this module.
[0105] The inverted residual module uses depthwise separable convolution instead of ordinary convolution operations, which can effectively reduce the number of model parameters. At the same time, in order to focus on the features on important channels, a channel attention mechanism is added to the structure, which can effectively improve the feature representation ability of the model and make the model more lightweight.
[0106] S500, using the ConvFormer module to extract global feature information at different scales under different paths;
[0107] In order to fully extract the fault features of transient signals, this application designs a parallel ConvFormer feature extraction module, which mainly consists of two parts: a multi-layer convolution module and a multi-head attention convolution module.
[0108] The multi-layer convolution module includes three convolution layers with a convolution kernel size of 3 stacked in sequence to extract deeper features of the preprocessed signal, thereby increasing the depth of the network to learn more abstract and advanced feature representations.
[0109] Due to the close contextual relationship between transient signals, ordinary convolution operations are difficult to fully capture the dependencies between transient signals. The multi-head attention convolution module includes three convolution layers with convolution kernels of different scales, a multi-head attention mechanism, a normalization layer, and a feedforward network layer. The multi-head attention mechanism can be used to assign different weights to transient signals at different positions in the input sequence to better capture the correlation between transient signals.
[0110] The multi-head attention convolution module includes a convolution layer (Conv) with three convolution kernels of different scales, a multi-head attention mechanism (Factotized MHSA), a normalization layer (Layer Norm) and a feed-forward network layer (FFN).
[0111] The multi-head attention convolution structure in the ConvFormer module can be used to model different fine-grained features, enrich the fault feature representation of transient signals, and combine CNN with the multi-head attention mechanism in a complementary way, taking into account both the local and global features of transient signals, thereby achieving a more accurate understanding and judgment of classification tasks.
[0112] The steps of the multi-head attention convolution module include:
[0113] S510, using three convolution layers with convolution kernels of different scales to extract fault features of preprocessed signals of different scales, where the convolution kernel sizes of the three convolution layers are 3, 5, and 7 respectively;
[0114] S520, dividing the input preprocessed signal into three vectors with the same dimensions, the vectors being a query vector q, a key vector k and a value vector v;
[0115] S530, encapsulate the three vectors into three matrices, namely query Q, value K, and value vector V;
[0116] S540, calculating the dot product operation between the matrix Q and the value K to obtain the similarity between the two;
[0117] S550, performing normalization processing through a Softmax operation;
[0118] S560, by weighted summing the normalized weights and the value vector V, the attention layer is given a different representation subspace to improve the performance of the self-attention layer. The attention mechanism calculation formula is as follows:
[0119] ,
[0120] Among them, Q represents query, K represents value, and V represents value vector. represents the scaling factor;
[0121] The above attention mechanism can be used to achieve sequence data modeling from fine to coarse, from coarse to fine and across scales, thereby improving the computational efficiency of the module.
[0122] Compared with traditional CNN, the ConvFormer module can utilize the multi-head attention mechanism in the structure to better handle long-distance dependencies and capture contextual information in fault signals, which helps to extract the global features of transient signal faults. At the same time, the combination of various types of transient signal fault features enriches the fault feature representation, thereby providing a more accurate understanding and judgment of the classification task.
[0123] S600, fusing local feature information and global feature information based on a feature fusion module;
[0124] In step S600, the calculation formula of the feature fusion module is:
[0125] A i = Concat ( [ L i , G i , 0 , G i , 1 ,..., G i , n ] ) ,
[0126] Among them, G in represents the output of the multi-head attention convolution module, L i Represents the output of the multi-layer convolution module, A i Represents the aggregated features, i represents the i-th category feature, and n represents the output of the n-th parallel Convformer module.
[0127] The feature fusion module can effectively fuse the global and local features of transient signals extracted from different paths in the previous stage through the Concat operation, which helps to make full use of the fault information of transient signals.
[0128] S700, classifying the fault of the transient signal of the power system by a classifier based on the Softmax function;
[0129] In step S700, the calculation formula of the Softmax function is:
[0130] ,
[0131] in, represents the score of the fth category, and k represents the total number of categories.
[0132] A Softmax classifier is constructed based on the Softmax function. The Softmax classifier first divides the input features into multiple mutually exclusive categories, and then maps the input features into a vector space. By applying the Softmax function to normalize these features, the probability value of each category is obtained, and then the category is determined and classified according to the size of the probability.
[0133] The method for diagnosing transient signal faults in power systems based on the MP-Convformer parallel network in the present application can effectively diagnose transient signal faults in power systems, and has good stability, strong robustness to noise, and certain generalization ability.
[0134] The advantage of this method is first reflected in the fact that the heuristic threshold selection can reduce the loss of fault feature information and has a good noise removal effect. Then it is reflected in the fact that the algorithm extracts features of different scales through multi-scale feature convolution operations and fuses these features together to make the fault signal feature information more accurate. Finally, the algorithm makes the extracted fault data information more complete by fusing the local features of the multi-layer convolution module and the global features extracted by the multi-head attention convolution module. Therefore, compared with other algorithms that only rely on a single feature or a single local feature, this algorithm can more effectively diagnose transient signal faults in power systems.
[0135] The present application also discloses a power system transient signal fault diagnosis system based on an MP-Convformer parallel network, which includes a memory and a processor, the memory being used to store executable program code; the processor is connected to the memory, and runs a computer program corresponding to the executable program code by reading the executable program code to execute any step in the above-mentioned power system transient signal fault diagnosis method based on an MP-Convformer parallel network.
[0136] The above is a detailed introduction to the method and system for diagnosing transient signal faults in power systems based on an MP-Convformer parallel network provided by the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea. At the same time, for general technical personnel in this field, according to the idea of the present application, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.
Claims
1. A method for fault diagnosis of transient signals in power systems based on MP-Convformer parallel networks, characterized in that: include: Collect data on transient signals under abnormal conditions; Preprocess the transient signal using wavelet denoising and obtain the preprocessed signal; The convolution operation in the inverted residual module is used to upsample the preprocessed signal and extract local feature information; The channel attention mechanism embedded in the inverted residual module is used to integrate the feature information of different channels; The ConvFormer module is used to extract global feature information at different scales under different paths; Based on the feature fusion module, local feature information and global feature information are fused; The classifier based on Softmax function is used to classify the faults of transient signals of power system; In the process of extracting global feature information at different scales under different paths using a ConvFormer module, the ConvFormer module includes a multi-layer convolution module and a multi-head attention convolution module; The multi-layer convolution module includes three convolution layers with a convolution kernel size of 3 stacked in sequence to extract deeper features of the preprocessed signal; The multi-head attention convolution module includes three convolution layers with convolution kernels of different scales, a multi-head attention mechanism, a normalization layer and a feedforward network layer; The step of integrating feature information of different channels by using the channel attention mechanism embedded in the inverted residual module includes the following steps: The feature information of different channels is integrated using the channel attention mechanism, which is embedded between two standard convolutional layers with a convolution kernel size of 1*1; Use point convolution operation with kernel size 1 to restore the dimension of preprocessed signal data; Use depth-wise separable convolution with a kernel size of 3 to restore the preprocessed signal after convolution processing; In the step of fusing local feature information and global feature information based on the feature fusion module, the feature fusion module uses the following formula for fusion: , Among them, G in represents the output of the multi-head attention convolution module, L i Represents the output of the multi-layer convolution module, A i represents the aggregated features, i represents the i-th category feature, and n represents the n-th parallel Convformer module output; The step of preprocessing the transient signal by using wavelet denoising and obtaining the preprocessed signal comprises the following steps: Based on orthogonal wavelet decomposition, the noisy transient signal is locally decomposed in the time domain and frequency domain. After orthogonal wavelet decomposition, the wavelet coefficients with larger amplitudes are denoised signals, while the coefficients with smaller amplitudes are noise signals. Determine whether the wavelet coefficient is greater than the preset threshold. If so, keep the wavelet coefficient. If not, replace the wavelet coefficient with the predicted variable threshold. The formula for processing the wavelet coefficient using the soft threshold function is: , and Respectively represent the wavelet transform coefficients before and after denoising, is the symbolic function, is the preset threshold, s represents the number of wavelet decomposition layers; Use inverse wavelet transform to restore the processed wavelet coefficients; In the step of determining whether the wavelet coefficient is greater than a preset threshold, the predicted variable threshold is calculated according to the following formula: , , , Where n is the length of the transient signal x, is the threshold of the universal fixed threshold rule, eta is the universal fixed threshold rule, is the threshold of the unbiased likelihood estimation rule, crit is the unbiased likelihood estimation rule, is the variable threshold for prediction; In the step of restoring the processed wavelet coefficients by using inverse wavelet transform, wavelet reconstruction is used to reduce noise interference. The wavelet reconstruction formula is: , , , Among them, j is the number of wavelet decomposition layers, k is the number of discrete sampling points, and represents the low-pass filter bank coefficients, and represents the high-pass filter bank coefficients, and are the scale coefficient and the wavelet coefficient after filtering respectively, and m represents the filter index parameter.
2. The method for fault diagnosis of transient signals in power systems based on MP-Convformer parallel networks according to claim 1 is characterized in that: The step of upsampling the preprocessed signal and extracting local feature information by using the convolution operation in the inverse residual module comprises the following steps: Use depth-wise separable convolution with a kernel size of 3 to upsample and extract local feature information; The dimension of the preprocessed signal is changed using a point convolution operation with a kernel size of 1.
3. The method for fault diagnosis of transient signals in power system based on MP-Convformer parallel network according to claim 1, characterized in that: The multi-head attention convolution module performs the following steps: Three convolution layers with convolution kernels of different scales are used to extract fault features of preprocessed signals of different scales. The convolution kernel sizes of the three convolution layers are 3, 5, and 7 respectively. The input preprocessed signal is divided into three vectors of the same dimension, namely the query vector q, the key vector k and the value vector v; Encapsulate the three vectors into three matrices, namely query Q, value K, and value vector V; Calculate the dot product operation between the matrix Q and the value K to get the similarity between the two; Normalization is performed through Softmax operation; The performance of the self-attention layer is improved by weighting the normalized weights and the value vector V to give the attention layer a different representation subspace. The attention mechanism calculation formula is as follows: , Among them, Q represents query, K represents value, and V represents value vector. Represents the scaling factor.
4. The method for fault diagnosis of transient signals in power systems based on MP-Convformer parallel networks according to claim 1 is characterized in that: In the step of classifying the fault of the transient signal of the power system by the classifier based on the Softmax function, the Softmax function is: , in, represents the score of the fth category, and k represents the total number of categories.
5. A power system transient signal fault diagnosis system based on MP-Convformer parallel network, characterized in that: include: A memory for storing executable program codes; as well as A processor, connected to the memory, runs a computer program corresponding to the executable program code by reading the executable program code to execute the steps in the power system transient signal fault diagnosis method based on the MP-Convformer parallel network as claimed in any one of claims 1 to 4.
