Fault diagnosis method of AC filter

Abstract

The invention discloses a fault diagnosis method for an AC filter, and relates to the technical field of fault detection. The method comprises the following steps: inputting transient signal data collected by sensors at a plurality of positions of the AC filter into a trained fault diagnosis model to obtain a fault classification result; the fault diagnosis model comprises a wavelet packet decomposition module used for performing wavelet packet decomposition on transient signal data and superposing decomposition results to obtain a three-dimensional wavelet packet coefficient tensor; the multi-attention mechanism fusion module is used for carrying out feature fusion according to the three-dimensional wavelet packet coefficient tensor on the basis of a plurality of attention mechanisms to obtain a first fusion feature; the multi-branch network is used for carrying out feature extraction on the first fusion features based on multiple target requirements and fusing extraction results to obtain second fusion features; and the classification module is used for outputting a fault classification result according to the second fusion feature. Therefore, fault diagnosis of the AC filter is carried out by using the acquired data of the plurality of position sensors, and the accuracy and reliability of diagnosis can be improved.

Description

Technical Field

[0001] The invention relates to the technical field of fault diagnosis, and in particular to a fault diagnosis method for an AC filter. Background Art

[0002] In the DC transmission system, due to the limitations of DC transmission technology, the converter will generate a large number of harmonics and consume a lot of reactive power when converting current. In order not to burden the power grid, each converter station will automatically put into the corresponding AC filter group on the AC side according to the transmission power of the DC transmission system during operation, provide the reactive power required by the converter and eliminate harmonics to improve the power quality of the power grid. Due to the flexibility of DC transmission, the transmission power needs to be adjusted many times a day, so the AC filter will also be frequently switched, making the switching frequency of the circuit breaker connected to it much higher than that of other line circuit breakers in the station, which undoubtedly puts pressure on the safe operation of the AC filter group and greatly tests the performance of the AC filter circuit breaker. If the AC filter has capacitor failure, poor terminal quality, current transformer failure, secondary equipment failure and other problems, the control of the converter will be unstable, which will interfere with the communication system and sometimes cause local resonant overvoltage in the power grid. Therefore, it is very necessary to diagnose the fault of the AC filter.

[0003] With the development of power systems, fault diagnosis technology is also constantly improving. However, traditional fault diagnosis methods often rely on a single data source or a single fault feature, which may not be able to accurately and efficiently diagnose faults in a complex and changeable power system environment, especially when the fault is complex and hidden, a single diagnostic method is even more difficult to do. Therefore, it is necessary to explore new fault diagnosis methods to improve the accuracy and efficiency of diagnosis. Summary of the invention

[0004] The present invention aims to solve one of the technical problems in the related art at least to a certain extent. To this end, the purpose of the present invention is to provide a fault diagnosis method for an AC filter to improve the accuracy and reliability of the fault diagnosis of the AC filter.

[0005] An embodiment of the present invention provides a fault diagnosis method for an AC filter, wherein sensors are provided at C positions of the AC filter, where C is an integer greater than 1, and the method comprises: obtaining transient signal data under abnormal conditions collected by sensors at the C positions; inputting the C transient signal data into a trained fault diagnosis model to obtain a fault classification result of the AC filter; wherein the fault diagnosis model comprises a wavelet packet decomposition module, a multi-attention mechanism fusion module, a multi-branch network and a classification module, wherein the wavelet packet decomposition module is used to perform wavelet packet decomposition on the C transient signal data respectively to obtain C two-dimensional wavelet packet coefficient matrices, and superimpose the C two-dimensional wavelet packet coefficient matrices to obtain a three-dimensional wavelet packet coefficient tensor; the multi-attention mechanism fusion module is used to perform feature fusion on the C transient signal data according to the three-dimensional wavelet packet coefficient tensor based on multiple attention mechanisms to obtain a first fusion feature; the multi-branch network is used to perform feature extraction on the first fusion feature respectively based on multiple target requirements to obtain multiple effective features, and fuse the multiple effective features to obtain a second fusion feature; the classification module is used to output a fault classification result according to the second fusion feature.

[0006] According to one embodiment of the present invention, the wavelet packet decomposition module performs two-layer wavelet packet decomposition on the transient signal data using DB1 as the wavelet basis, and arranges the coefficients in the last layer in ascending order of frequency subbands to obtain a two-dimensional wavelet packet coefficient matrix corresponding to the transient signal data.

[0007] According to one embodiment of the present invention, the multiple attention mechanisms include a position attention mechanism and a channel attention mechanism; based on the multiple attention mechanisms, feature fusion is performed on C transient signal data according to the three-dimensional wavelet packet coefficient tensor to obtain a first fusion feature, including: using a 1×1 convolution layer to process the three-dimensional wavelet packet coefficient tensor to obtain a first feature matrix; based on the position attention mechanism, feature extraction is performed on the first feature matrix to obtain position features; based on the channel attention mechanism, feature extraction is performed on the first feature matrix to obtain channel features; the position features and the channel features are summed to obtain the first fusion features.

[0008] According to one embodiment of the present invention, the feature extraction of the first feature matrix based on the position attention mechanism to obtain the position feature includes: processing the first feature matrix using a 1×1 convolution layer to obtain a second feature matrix; performing a shape transformation operation on the second feature matrix to obtain a third feature matrix; transposing the third feature matrix to obtain a fourth feature matrix; performing a matrix multiplication operation on the fourth feature matrix and the third feature matrix, and applying a SoftMax function to process the matrix multiplication result to obtain a fifth feature matrix; transposing the fifth feature matrix to obtain a sixth feature matrix; performing a matrix multiplication operation on the sixth feature matrix and the third feature matrix, and performing a shape transformation operation on the matrix multiplication result to obtain a seventh feature matrix; summing the seventh feature matrix and the first feature matrix, and processing the summation result using a 1×1 convolution layer to obtain the position feature.

[0009] According to one embodiment of the present invention, the feature extraction of the first feature matrix based on the channel attention mechanism to obtain channel features includes: performing a shape transformation operation on the first feature matrix to obtain an eighth feature matrix; performing a transposition operation on the eighth feature matrix to obtain a ninth feature matrix; performing a matrix multiplication operation on the ninth feature matrix and the eighth feature matrix, and applying a SoftMax function to process the matrix multiplication result to obtain a tenth feature matrix; performing a matrix multiplication operation on the tenth feature matrix and the eighth feature matrix, and performing a shape transformation operation on the matrix multiplication result to obtain an eleventh feature matrix; summing the eleventh feature matrix and the first feature matrix, and processing the summation result using a 1×1 convolutional layer to obtain the channel features.

[0010] According to one embodiment of the present invention, the multiple target requirements include denoising requirements, high-frequency feature extraction requirements and low-frequency feature extraction requirements; based on the multiple target requirements, feature extraction is performed on the first fused feature respectively to obtain multiple valid features, including: denoising the first fused feature based on the denoising requirement to obtain a first valid feature; extracting the high-frequency features of the first fused feature based on the high-frequency feature extraction requirement to obtain a second valid feature; extracting the low-frequency features of the first fused feature based on the low-frequency feature extraction requirement to obtain a third valid feature.

[0011] According to one embodiment of the present invention, the denoising process is performed on the first fusion feature based on the denoising requirement to obtain the first effective feature, including: performing global average pooling processing, Conv (8, 1, 1) convolution processing, standardization processing, ReLU function processing, Conv (8, 1, 1) convolution processing, soft threshold processing and maximum pooling processing on the first fusion feature in sequence to obtain the first effective feature.

[0012] According to one embodiment of the present invention, the high-frequency features of the first fused features are extracted based on the high-frequency feature extraction requirements to obtain the second effective features, including: performing Conv (32, 6, 1) convolution processing, standardization processing, ReLU function processing, Conv (16, 6, 1) convolution processing, standardization processing, ReLU function processing, maximum pooling processing, Conv (10, 6, 1) convolution processing, standardization processing, ReLU function processing, Conv (10, 8, 2) convolution processing, standardization processing, ReLU function processing and maximum pooling processing on the first fused features in sequence to obtain the second effective features.

[0013] According to one embodiment of the present invention, the low-frequency features of the first fused features are extracted based on the low-frequency feature extraction requirement to obtain a third effective feature, including: performing Conv (16, 18, 2) convolution processing, standardization processing, ReLU function processing, Conv (10, 10, 2) convolution processing, standardization processing, ReLU function processing and maximum pooling processing on the first fused features in sequence to obtain the third effective feature.

[0014] According to one embodiment of the present invention, the multiple valid features are fused to obtain a second fused feature, including: summing the multiple valid features, and processing the summation result through a global average pooling layer and a fully connected layer in sequence to obtain the second fused feature.

[0015] The fault diagnosis method of the AC filter of the embodiment of the present invention first obtains the transient signal data collected by the sensor at C positions under abnormal conditions; then the C transient signal data are input into the trained fault diagnosis model, and the wavelet packet decomposition module in the model performs wavelet packet decomposition on the C transient signal data respectively to obtain C two-dimensional wavelet packet coefficient matrices, and the C two-dimensional wavelet packet coefficient matrices are superimposed to obtain a three-dimensional wavelet packet coefficient tensor, and then the multi-attention mechanism fusion module in the model is used to perform feature fusion on the C transient signal data according to the three-dimensional wavelet packet coefficient tensor based on multiple attention mechanisms to obtain a first fusion feature, and then the multi-branch network in the model is used to extract features from the first fusion feature based on multiple target requirements to obtain multiple effective features, and the multiple effective features are fused to obtain a second fusion feature, and finally the classification module in the model outputs the fault classification result of the AC filter according to the second fusion feature. Therefore, the multi-source information fusion strategy is used to enable the model to process data from different data sources with different characteristics, enhance the adaptability of the model to different types of faults, improve the feature extraction ability of the model, and improve the accuracy and reliability of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a structural schematic diagram of a fault diagnosis model according to an embodiment of the present invention; Figure 2 is a schematic diagram of a wavelet packet decomposition process according to an embodiment of the present invention; Figure 3 is a schematic diagram of the structure of a multi-attention mechanism fusion module according to an embodiment of the present invention; Figure 4 It is a schematic diagram of the structure of a multi-branch network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0017] Since information from a single source is difficult to fully reflect the status of the equipment, the present invention proposes a fault diagnosis method for an AC filter. This method integrates sensor data collected from different locations, fully utilizes information from various data sources, overcomes the one-sidedness, susceptibility to interference and diagnostic limitations of single-source information, and can more comprehensively understand the operating status of the AC filter, thereby improving the accuracy and reliability of AC filter fault diagnosis and providing strong support for the safe and stable operation of the power system.

[0018] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0019] The following describes a fault diagnosis method for an AC filter according to an embodiment of the present invention with reference to the accompanying drawings.

[0020] In an embodiment of the present invention, sensors are provided at C positions of the AC filter, where C is an integer greater than 1. The sensors at the C positions can be selected as required, and may specifically include multiple of the following sensors: The input side sensor set at the input end of the AC filter can be set close to the power supply or load to monitor the input voltage, current, etc. to evaluate the power quality, and detect input harmonics to evaluate the filter effect, etc.; the output side sensor set at the output end of the AC filter can be set close to the load to monitor the output voltage, current, etc. to ensure the normal operation of the load, and detect output harmonics to evaluate the filtering effect; the internal sensor set at the key components (such as inductors and capacitors) inside the AC filter is used to monitor the component status to prevent failures, and to detect component temperature, current, etc. in real time to prevent overload; the grounding loop sensor set in the grounding loop of the AC filter is used to monitor Measure the ground current to ensure that the grounding system is normal, and detect leakage current to prevent electrical accidents; install environmental sensors around the AC filter to monitor ambient temperature and humidity to ensure that the filter operates under suitable conditions, and detect dust, corrosive gases, etc. to prevent environmental damage; install cooling system sensors on the radiator or cooling system of the AC filter to monitor the cooling system temperature to prevent overheating, and detect the flow of cooling medium to ensure cooling efficiency; install protection device sensors at the protection devices of the AC filter (such as circuit breakers and fuses) to monitor the status of the protection devices to ensure their normal operation, and detect the action of the protection devices to respond to faults in time.

[0021] The fault diagnosis method of the AC filter includes the following steps S11-S12: S11, obtaining transient signal data under abnormal conditions collected by sensors at C positions.

[0022] The transient signal data may refer to a signal that changes significantly in a short period of time.

[0023] Exemplarily, when the data collected by any sensor is abnormal (such as the aforementioned significant change, which can be determined based on whether the data change rate is greater than the corresponding threshold), the signal data collected by the sensors at C positions can be obtained. For example, when the signal data collected by the input side sensor is abnormal, the signal data collected by other sensors at different positions (such as the output side sensor, internal sensor, protection device sensor, etc.) can be obtained while the signal data of the input sensor is obtained, and these signal data are all recorded as transient signal data for subsequent AC filter fault diagnosis.

[0024] S12, inputting C transient signal data into the trained fault diagnosis model to obtain the fault classification result of the AC filter.

[0025] Among them, see Figure 1 The fault diagnosis model includes a wavelet packet decomposition module, a multi-attention mechanism fusion module, a multi-branch network and a classification module. The wavelet packet decomposition module is used to perform wavelet packet decomposition on C transient signal data respectively to obtain C two-dimensional wavelet packet coefficient matrices, and superimpose the C two-dimensional wavelet packet coefficient matrices to obtain a three-dimensional wavelet packet coefficient tensor; the multi-attention mechanism fusion module is used to perform feature fusion on C transient signal data according to the three-dimensional wavelet packet coefficient tensor based on multiple attention mechanisms to obtain a first fusion feature; the multi-branch network is used to extract features from the first fusion feature based on multiple target requirements to obtain multiple valid features, and fuse the multiple valid features to obtain a second fusion feature; the classification module is used to output the fault classification result according to the second fusion feature.

[0026] Specifically, Figure 1 As shown, the multi-source information fusion fault diagnosis model designed by the present invention is mainly divided into three parts: the first part converts the composite fault samples collected from multiple data sources into a three-dimensional wavelet coefficient tensor; the second part designs a feature fusion module based on multiple attention mechanisms, such as including a position attention mechanism and a channel attention mechanism. The position attention mechanism can deeply mine the internal time-frequency information of the source, and the channel attention mechanism can integrate the differences in multi-source information to achieve comprehensive information fusion; the third part designs a three-branch network to perform multi-scale feature extraction and denoising, extract effective information, and finally realize fault classification through a classifier.

[0027] For example, when training a fault diagnosis model, the sample data is randomly divided into a training set, a validation set, and a test set according to a preset ratio (such as 7:2:1). The training set is used for model training, the validation set is used for tuning and selecting the model, and the test set is used for final performance evaluation.

[0028] The fault diagnosis method of the AC filter utilizes a multi-source information fusion strategy to enable the model to process data from different data sources with different characteristics, thereby enhancing the model's adaptability to different types of faults, improving the model's feature extraction capability, and improving the accuracy and reliability of fault diagnosis.

[0029] In some embodiments of the present invention, the wavelet packet decomposition module performs two-layer wavelet packet decomposition on the transient signal data using DB1 as the wavelet basis, and arranges the coefficients in the last layer in ascending order of the frequency subband to obtain a two-dimensional wavelet packet coefficient matrix corresponding to the transient signal data.

[0030] Specifically, the Mallat fast discrete wavelet transform is based on the theory of multi-resolution analysis. It decomposes the signal into sub-signals in different frequency bands, processes and analyzes them separately, and implements discrete approximation analysis of the signal at different scales, thereby revealing the local and global characteristics of the signal. However, the Mallat fast discrete wavelet transform cannot effectively segment the high-frequency band containing rich fault modulation information. For this reason, the present invention proposes to use wavelet packet decomposition to recursively decompose the low-frequency and high-frequency parts of the signal at the same time, providing a finer frequency band division in the entire analysis band. For a length of Input signal data , its wavelet packet decomposition process is as follows (1): (1) in, represents the nth wavelet packet coefficient of the jth sub-band (i.e., frequency sub-band) of the lth layer decomposition, Indicates Layer decomposition The wavelet packet coefficient sequence of sub-bands, , and They represent low-pass filter and high-pass filter respectively, satisfying the following formula (2): (2) in, express The nth coefficient in positive order, express Arranged in reverse order coefficients.

[0031] The sequence length is Wavelet packet nodes The corresponding sub-band energy calculation formula is as follows (3): (3) in, Represents a wavelet packet node The corresponding sub-band energy.

[0032] The two-layer wavelet packet decomposition process is as follows Figure 2 As shown, firstly, the signal data collected by sensors at multiple positions are decomposed by wavelet packet using DB1 as the wavelet basis. In order to obtain the final two-dimensional wavelet packet coefficient matrix, the wavelet packet coefficients in the last layer In ascending order of subbands. In particular, the wavelet packet coefficients in a subband are regarded as a row of a two-dimensional wavelet packet coefficient matrix, and the size of the matrix is , is the number of points of the original discrete signal data, that is, the signal data length mentioned above. When the number of sensors is C, C two-dimensional wavelet packet coefficient matrices are superimposed to form a shape of The three-dimensional tensor of can be obtained by taking the three-dimensional wavelet packet coefficient tensor ,in for .

[0033] By performing wavelet packet decomposition on the signal data, not only can the signal data be filtered and denoised, but the signal data can also be cleverly transformed into a two-dimensional time-frequency feature space. More uniquely, by superimposing C two-dimensional wavelet packet coefficient matrices as described above, the information data from multiple signal sources can be further integrated to obtain a three-dimensional wavelet packet coefficient tensor. By constructing a three-dimensional matrix (i.e., a three-dimensional wavelet packet coefficient tensor), the data is processed in an up-dimensional manner, thereby mining richer and deeper feature information, which helps to improve the accuracy of fault diagnosis.

[0034] In some embodiments of the present invention, multiple attention mechanisms include a position attention mechanism and a channel attention mechanism; based on the multiple attention mechanisms, feature fusion is performed on C transient signal data according to the three-dimensional wavelet packet coefficient tensor to obtain a first fused feature, including: using a 1×1 convolution layer to process the three-dimensional wavelet packet coefficient tensor to obtain a first feature matrix; based on the position attention mechanism, feature extraction is performed on the first feature matrix to obtain position features; based on the channel attention mechanism, feature extraction is performed on the first feature matrix to obtain channel features; the position features and the channel features are summed to obtain the first fused features.

[0035] Specifically, due to the limitation of sensor characteristics, it is difficult for a single data source to fully monitor the working state of the AC filter. Therefore, multiple sensor channels are used to collect fault samples with multiple data sources as the original data set. Based on this, a multi-attention mechanism fusion module is proposed. The multi-attention mechanism can be composed of at least two of the position attention mechanism, the channel attention mechanism, and the space attention mechanism. This embodiment is illustrated by taking the position attention mechanism and the channel attention mechanism as an example. The above-mentioned three-dimensional wavelet packet coefficient tensor is simultaneously input into the multi-attention mechanism fusion module composed of the position attention mechanism and the channel attention mechanism, so as to deeply mine the time-frequency information inside the source through the position attention mechanism to obtain the position feature, which is helpful to capture the subtle changes of the fault feature in the time domain and the frequency domain; the channel attention mechanism is used to integrate the differences of the information data of multiple sources to obtain the channel feature, and the position feature and the channel feature are summed to obtain the first fusion feature, so as to realize the comprehensive fusion of information data, avoid the information loss or misleading that may be caused by a single data source, and thus improve the reliability of fault diagnosis.

[0036] Exemplarily, feature extraction is performed on the first feature matrix based on the position attention mechanism to obtain position features, including: using a 1×1 convolution layer to process the first feature matrix to obtain a second feature matrix; performing a shape transformation (i.e., Reshape) operation on the second feature matrix to obtain a third feature matrix; performing a transposition (i.e., Transpose) operation on the third feature matrix to obtain a fourth feature matrix; performing a matrix multiplication operation on the fourth feature matrix and the third feature matrix, and applying a SoftMax function to process the matrix multiplication result to obtain a fifth feature matrix; performing a Transpose operation on the fifth feature matrix to obtain a sixth feature matrix; performing a matrix multiplication operation on the sixth feature matrix and the third feature matrix, and performing a Reshape operation on the matrix multiplication result to obtain a seventh feature matrix; summing the seventh feature matrix and the first feature matrix, and processing the summation result using a 1×1 convolution layer to obtain position features.

[0037] Exemplarily, based on the channel attention mechanism, feature extraction is performed on the first feature matrix to obtain channel features, including: performing a Reshape operation on the first feature matrix to obtain an eighth feature matrix; performing a Transpose operation on the eighth feature matrix to obtain a ninth feature matrix; performing a matrix multiplication operation on the ninth feature matrix and the eighth feature matrix, and applying the SoftMax function to process the matrix multiplication result to obtain a tenth feature matrix; performing a matrix multiplication operation on the tenth feature matrix and the eighth feature matrix, and performing a Reshape operation on the matrix multiplication result to obtain an eleventh feature matrix; summing the eleventh feature matrix and the first feature matrix, and processing the summation result using a 1×1 convolutional layer to obtain channel features.

[0038] Specifically, Figure 3 As shown in Figure 2, the multi-attention mechanism fusion module can include a position attention unit and a channel attention unit. The input ends of the two units can share a 1×1 convolution layer or set a 1×1 convolution layer respectively ( Figure 3 As shown in this example), the three-dimensional wavelet packet coefficient tensor is processed to obtain the first feature matrix. The position attention unit can deeply explore the time-frequency information within each source. It can not only keenly capture the subtle changes in the signal data in the time and frequency dimensions, but also accurately analyze these changes, thereby revealing the core features hidden deep in the complex data. The channel attention unit plays a key role in integrating the differences in information data between multiple sources. By deeply comparing and analyzing information data from different sources, it can accurately extract the complementary information and unique contributions between the sources, thereby achieving comprehensive integration and optimization of information.

[0039] Specifically, the position attention unit first uses a 1×1 convolution layer to extract features from the first feature matrix to obtain the second feature matrix. This process aims to more effectively capture the position features in the input data. Then, the second feature matrix is ​​reshaped (to change the shape of the matrix) to obtain the third feature matrix; the third feature matrix is ​​transposed (transposed) to obtain the fourth feature matrix. Then, the third feature matrix is ​​multiplied by the transposed feature matrix (i.e., the fourth feature matrix), and the SoftMax function is applied to the result to output The feature matrix (i.e., the fifth feature matrix) not only realizes the effective integration of features, but also cleverly eliminates the channel dimension , further simplifying the data processing process. In addition, in order to ensure smoother gradient propagation of the network and effectively prevent the problem of too small feature response, the position attention unit also cleverly uses residual connection, first converting the fifth feature matrix into the seventh feature matrix with the same dimension as the input (i.e., the first feature matrix), and then summing the seventh feature matrix with the input, which not only improves the stability and robustness of the network, but also significantly enhances the expressiveness and generalization ability of the fault diagnosis model.

[0040] In order to focus more on the channel dimension and reduce the structural complexity, the channel attention unit first reshapes the first feature matrix to obtain the eighth feature matrix, then transposes the eighth feature matrix to obtain the ninth feature matrix, multiplies it with the eighth feature matrix, and then outputs it through the SoftMax function. This process effectively removes the position dimension and Similarly, the channel attention unit also sets a residual connection.

[0041] Finally, the outputs of the channel attention unit and the position attention unit are summed, and the result is regarded as the multi-attention mechanism fusion feature, that is, the first fusion feature. This fusion feature not only deeply mines the time-frequency information in each data source, but also fully considers the correlation between multiple data sources, realizing the deep integration and efficient utilization of information data. Through this design, not only the efficiency and accuracy of information data processing are improved, but also the generalization ability and robustness of the fault diagnosis model are enhanced. At the same time, the unique fusion strategy and precise extraction ability of the multi-attention mechanism fusion module make the multi-attention mechanism fusion module show a wide range of application prospects in signal data processing, data mining and multi-source information integration.

[0042] In some embodiments of the present invention, the multiple target requirements include at least two of a denoising requirement, a high-frequency feature extraction requirement, and a low-frequency feature extraction requirement. The present invention is described by taking the three as an example.

[0043] Among them, feature extraction is performed on the first fused feature based on multiple target requirements to obtain multiple effective features, including: denoising the first fused feature based on the denoising requirement to obtain the first effective feature; extracting the high-frequency feature of the first fused feature based on the high-frequency feature extraction requirement to obtain the second effective feature; extracting the low-frequency feature of the first fused feature based on the low-frequency feature extraction requirement to obtain the third effective feature.

[0044] For example, Figure 4 As shown, based on the denoising requirement, the first fusion feature is denoised to obtain the first effective feature, including: performing global average pooling processing, Conv (8, 1, 1) convolution processing, standardization processing, ReLU function processing, Conv (8, 1, 1) convolution processing, soft threshold processing and maximum pooling processing on the first fusion feature in sequence to obtain the first effective feature.

[0045] The 8 in Conv(8,1,1) indicates the number of output channels of the convolution process, the 1 in the middle indicates that the convolution kernel size of the convolution process is 1×1, and the other 1 indicates that the convolution stride of the convolution process is 1. The meanings of the numbers in Conv(32,6,1), Conv(16,6,1), Conv(10,6,1), Conv(10,8,2), Conv(16,18,2), Conv(10,10,2), etc. are similar to those in Conv(8,1,1). The stride of the maximum pooling process can be 2 and / or the window size can be 2×2.

[0046] For example, Figure 4 As shown, based on the high-frequency feature extraction requirement, the high-frequency features of the first fused features are extracted to obtain the second effective features, including: performing Conv (32, 6, 1) convolution processing, standardization processing, ReLU function processing, Conv (16, 6, 1) convolution processing, standardization processing, ReLU function processing, maximum pooling processing, Conv (10, 6, 1) convolution processing, standardization processing, ReLU function processing, Conv (10, 8, 2) convolution processing, standardization processing, ReLU function processing and maximum pooling processing on the first fused features in sequence to obtain the second effective features.

[0047] For example, Figure 4 As shown, based on the low-frequency feature extraction requirement, the low-frequency features of the first fusion features are extracted to obtain the third effective features, including: performing Conv (16, 18, 2) convolution processing, standardization processing, ReLU function processing, Conv (10, 10, 2) convolution processing, standardization processing, ReLU function processing and maximum pooling processing on the first fusion features in sequence to obtain the third effective features.

[0048] Specifically, after fusing the original features of multi-source information data, in order to more effectively mine and utilize the key information in these data, a crucial step is to remove redundant information that is irrelevant to the target task (i.e., fault diagnosis), and accurately extract effective features that are of significant value for subsequent analysis. In view of this, the present invention designs an innovative three-branch network, whose architecture is as follows Figure 4 As shown in Figure 2, the architecture is designed to process the input feature data (i.e., the first fusion feature) comprehensively and carefully through three parallel and independent processing branches.

[0049] Specifically, see Figure 2 , the denoising branch focuses on denoising. It uses the soft threshold activation layer to perform soft threshold processing to effectively filter out the noise components in the first fusion feature, laying the foundation for subsequent feature extraction. Soft threshold processing designs a nonlinear mapping to discard low threshold coefficients whose absolute values ​​are less than the critical threshold, thereby reducing noise interference. Soft threshold function The definition is as follows: (4) in, Represents the input value, represents a non-negative threshold, represents the symbolic function, Represents the maximum value function.

[0050] ReLu function It is a nonlinear activation function commonly used in CNN (Convolutional Neural Network), and is defined as follows: (5) in, represents the maximum value function. According to equations (4) and (5), we can get and The relationship between: (6) In order to more conveniently implement adaptive filtering and denoising within the network, the present invention regards the threshold as a trainable parameter and embeds it into the network structure. The optimal threshold is self-learned from the data through the gradient descent back propagation algorithm. Combined with formula (6): The function constructs a nonlinear activation layer and places it after the convolutional layer to perform sparse noise reduction on the information data of each channel.

[0051] The high-frequency branch is dedicated to the extraction of high-frequency features. Figure 4The convolutional layers and pooling layers shown can keenly capture and enhance high-frequency features that reflect rapid changes or detailed information, which is crucial for identifying and analyzing certain subtle fault modes.

[0052] The low-frequency branch focuses on the extraction of low-frequency features. It focuses on revealing the global trends and long-term changes in signal data. Through appropriate smoothing and feature aggregation strategies, it extracts low-frequency features that have an important impact on the overall system state.

[0053] These three branches not only work independently, but also achieve the complementarity and fusion of multi-scale features through the subsequent clever fusion mechanism, thereby constructing a comprehensive and rich fault-sensitive feature set.

[0054] Multi-scale feature extraction through a three-branch network can enable the fault diagnosis model to capture the performance of fault features at different scales, while removing noise, thereby improving the generalization ability of the fault diagnosis model when dealing with complex faults.

[0055] When training the fault diagnosis model, the back propagation algorithm of the Adam optimizer is used. The process can be written as: (7) in, , , Represents the parameter set of each of the three branches, represents the cross entropy loss.

[0056] In some embodiments of the present invention, multiple valid features are fused to obtain a second fused feature, including: summing the multiple valid features, and processing the summation result through a global average pooling layer and a fully connected layer in sequence to obtain the second fused feature.

[0057] Specifically, in the forward propagation process, the features extracted by the three branches are first fused by linear summation of the following formula (8): (8) in, , , Respectively represent the features extracted by the three branches, namely the first effective feature, the second effective feature and the third effective feature, Indicates the summation result.

[0058] Then, the global average pooling layer is used to retain the semantic information of the channel dimension and enhance the generalization ability of the model; the fully connected layer is used to integrate all the information of the features.

[0059] The purpose of the fully connected layer is to represent the input features as deeper features, which can map the input features into a vector space. The fully connected layer connects all neurons in the previous layer to each neuron in the current layer. Its calculation expression is as follows: (9) in, and Respectively represent The weight and bias of the jth neuron in the fully connected layer, represents the activation function, Indicates The output of the fully connected layer, Indicates The input of a fully connected layer.

[0060] In some embodiments of the present invention, the classification module may use a softmax classifier. The softmax classifier is constructed based on the softmax function, which is used to normalize the second fusion feature and classify by obtaining the probability of each category. The softmax function is defined as follows: (10) in, Indicates The characteristics of the categories, Represents the total number of categories, and e represents a natural constant. The softmax function converts the input features into probabilities so that the sum of the probabilities of all categories is 1.

[0061] In summary, the fault diagnosis method of the AC filter of the embodiment of the present invention can effectively diagnose the fault state of the AC filter, and has good stability, strong robustness to noise, and certain generalization ability. Since different mechanical structures, working conditions or fault types usually excite fault pulses with different intensities and frequencies, this may limit the effective feature representation ability of a single fault signal. For this reason, the present invention adopts a multi-source information fusion strategy to enable the fault diagnosis model to process signal data from different data sources with different characteristics, enhance the adaptability of the fault diagnosis model to different types of faults, improve the feature extraction ability of the fault diagnosis model, and thus improve the accuracy of fault diagnosis.

[0062] It should be noted that the logic and / or steps represented in the flowchart or described in other ways herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in combination with these instruction execution systems, devices or apparatuses. For the purpose of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in combination with these instruction execution systems, devices or apparatuses. More specific examples (non-exhaustive list) of computer-readable media include the following: an electrical connection portion with one or more wirings (electronic device), a portable computer disk box (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium and then editing, interpreting or processing in other suitable ways if necessary, and then stored in a computer memory.

[0063] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0064] 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.

[0065] In the description of the present invention, it is to be understood that the terms “center”, “longitudinal”, “lateral”, “length”, “width”, “thickness”, “up”, “down”, “front”, “back”, “left”, “right”, “vertical”, “horizontal”, “top”, “bottom”, “inside”, “outside”, “clockwise”, “counterclockwise”, “axial”, “radial”, “circumferential”, etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as limiting the present invention.

[0066] 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.

[0067] In the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements, unless otherwise clearly defined. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0068] In the present invention, unless otherwise clearly specified and limited, a first feature being "above" or "below" a second feature may mean that the first and second features are in direct contact, or the first and second features are in indirect contact through an intermediate medium. Moreover, a first feature being "above", "above" or "above" a second feature may mean that the first feature is directly above or obliquely above the second feature, or simply means that the first feature is higher in level than the second feature. A first feature being "below", "below" or "below" a second feature may mean that the first feature is directly below or obliquely below the second feature, or simply means that the first feature is lower in level than the second feature.

[0069] 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 fault diagnosis method for an AC filter, characterized in that: Sensors are provided at C positions of the AC filter, where C is an integer greater than 1, and the method comprises: Acquire transient signal data collected by sensors at C locations under abnormal conditions; Inputting the C transient signal data into a trained fault diagnosis model to obtain a fault classification result of the AC filter; The fault diagnosis model includes a wavelet packet decomposition module, a multi-attention mechanism fusion module, a multi-branch network and a classification module. The wavelet packet decomposition module is used to perform wavelet packet decomposition on C transient signal data respectively to obtain C two-dimensional wavelet packet coefficient matrices, and superimpose the C two-dimensional wavelet packet coefficient matrices to obtain a three-dimensional wavelet packet coefficient tensor; the multi-attention mechanism fusion module is used to perform feature fusion on C transient signal data according to the three-dimensional wavelet packet coefficient tensor based on multiple attention mechanisms to obtain a first fusion feature; the multi-branch network is used to perform feature extraction on the first fusion feature based on multiple target requirements respectively to obtain multiple effective features, and fuse the multiple effective features to obtain a second fusion feature; the classification module is used to output a fault classification result according to the second fusion feature.

2. The fault diagnosis method of the AC filter according to claim 1, characterized in that: The wavelet packet decomposition module performs two-layer wavelet packet decomposition on the transient signal data using DB1 as the wavelet basis, and arranges the coefficients in the last layer in ascending order of frequency subbands to obtain a two-dimensional wavelet packet coefficient matrix corresponding to the transient signal data.

3. The fault diagnosis method of the AC filter according to claim 1, characterized in that: The multiple attention mechanisms include a position attention mechanism and a channel attention mechanism; based on the multiple attention mechanisms, feature fusion is performed on C transient signal data according to the three-dimensional wavelet packet coefficient tensor to obtain a first fusion feature, including: Processing the three-dimensional wavelet packet coefficient tensor using a 1×1 convolution layer to obtain a first feature matrix; Performing feature extraction on the first feature matrix based on the position attention mechanism to obtain position features; Performing feature extraction on the first feature matrix based on the channel attention mechanism to obtain channel features; The position feature and the channel feature are summed to obtain the first fusion feature.

4. The fault diagnosis method of the AC filter according to claim 3 is characterized in that: The extracting features from the first feature matrix based on the position attention mechanism to obtain position features includes: Processing the first feature matrix using a 1×1 convolutional layer to obtain a second feature matrix; Performing a shape transformation operation on the second characteristic matrix to obtain a third characteristic matrix; Performing a transposition operation on the third characteristic matrix to obtain a fourth characteristic matrix; Performing a matrix multiplication operation on the fourth characteristic matrix and the third characteristic matrix, and applying a SoftMax function to process the matrix multiplication result to obtain a fifth characteristic matrix; Performing a transposition operation on the fifth characteristic matrix to obtain a sixth characteristic matrix; Performing a matrix multiplication operation on the sixth characteristic matrix and the third characteristic matrix, and performing a shape transformation operation on the result of the matrix multiplication operation to obtain a seventh characteristic matrix; The seventh feature matrix is ​​summed with the first feature matrix, and the summation result is processed by using a 1×1 convolution layer to obtain the position feature.

5. The fault diagnosis method of the AC filter according to claim 3, characterized in that: The extracting features from the first feature matrix based on the channel attention mechanism to obtain channel features includes: Performing a shape transformation operation on the first characteristic matrix to obtain an eighth characteristic matrix; Performing a transposition operation on the eighth characteristic matrix to obtain a ninth characteristic matrix; Performing a matrix multiplication operation on the ninth characteristic matrix and the eighth characteristic matrix, and applying a SoftMax function to process the matrix multiplication operation result to obtain a tenth characteristic matrix; Performing a matrix multiplication operation on the tenth characteristic matrix and the eighth characteristic matrix, and performing a shape transformation operation on the result of the matrix multiplication operation to obtain an eleventh characteristic matrix; The eleventh feature matrix is ​​summed with the first feature matrix, and the summation result is processed by using a 1×1 convolution layer to obtain the channel feature.

6. The fault diagnosis method of the AC filter according to claim 1, characterized in that: The multiple target requirements include denoising requirements, high-frequency feature extraction requirements, and low-frequency feature extraction requirements; Based on multiple target requirements, feature extraction is performed on the first fusion feature to obtain multiple effective features, including: Performing denoising processing on the first fused feature based on the denoising requirement to obtain a first effective feature; Extracting high-frequency features of the first fused features based on the high-frequency feature extraction requirement to obtain second effective features; Based on the low-frequency feature extraction requirement, the low-frequency feature of the first fusion feature is extracted to obtain a third effective feature.

7. The fault diagnosis method of the AC filter according to claim 6, characterized in that: The performing denoising on the first fused feature based on the denoising requirement to obtain a first effective feature includes: The first fusion feature is sequentially subjected to global average pooling processing, Conv (8, 1, 1) convolution processing, standardization processing, ReLU function processing, Conv (8, 1, 1) convolution processing, soft threshold processing and maximum pooling processing to obtain the first effective feature.

8. The fault diagnosis method of the AC filter according to claim 6, characterized in that: The extracting the high-frequency feature of the first fused feature based on the high-frequency feature extraction requirement to obtain the second effective feature includes: The first fusion feature is sequentially subjected to Conv (32, 6, 1) convolution processing, standardization processing, ReLU function processing, Conv (16, 6, 1) convolution processing, standardization processing, ReLU function processing, maximum pooling processing, Conv (10, 6, 1) convolution processing, standardization processing, ReLU function processing, Conv (10, 8, 2) convolution processing, standardization processing, ReLU function processing and maximum pooling processing to obtain the second valid feature.

9. The fault diagnosis method of the AC filter according to claim 6, characterized in that: The step of extracting the low-frequency feature of the first fused feature based on the low-frequency feature extraction requirement to obtain a third effective feature includes: The first fusion feature is sequentially subjected to Conv (16, 18, 2) convolution processing, standardization processing, ReLU function processing, Conv (10, 10, 2) convolution processing, standardization processing, ReLU function processing and maximum pooling processing to obtain the third effective feature.

10. The fault diagnosis method of the AC filter according to claim 1, characterized in that: The multiple valid features are fused to obtain a second fused feature, including: The multiple valid features are summed, and the summed result is processed by a global average pooling layer and a fully connected layer in sequence to obtain the second fusion feature.

Images