A high-precision electronic circuit fault diagnosis method

Abstract

The present application relates to the technical field of electronic circuit fault diagnosis, in particular to a high-precision electronic circuit fault diagnosis method, which comprises the following steps: S1, setting different sizes of intercepting boxes for the multi-dimensional output voltage data of the circuit to intercept the data; S2, encoding the data intercepted in step S1 through GADF, expanding the data, and finally forming fault data containing two-dimensional feature images of fault characteristics; the extracted tiny features are highlighted through a stacking operation; the reverse Ghost module can expand the influence range of important features in important channels, and the tiny features previously extracted from the feature images are further highlighted through RPGMCBAM, the obtained diagnosis model can effectively extract the fault characteristics in the fault signals, under the assistance of CSM and RPGMCBAM, the tiny features are effectively extracted, the precision of the diagnosis model is improved, and the diagnosis model has excellent diagnosis performance and robustness.

Description

Technical Field

[0001] This invention relates to the field of electronic circuit fault diagnosis, specifically a high-precision electronic circuit fault diagnosis method. Background Technology

[0002] Currently, power electronics technology has been widely used in new energy, biomedicine, railway transportation, Internet of Things and other fields. In the field of smart grids, power storage, high voltage DC transmission technology, flexible AC transmission technology and other aspects are all related to power electronics technology. Therefore, the safe use of power electronic devices is even more important.

[0003] Currently, in the field of electronic circuit fault diagnosis, existing technologies use known switching circuit models and measurement data to establish circuit state judgment models. The difference between actual measured values ​​and predicted values ​​is used to determine whether a fault has occurred. For example, the online fault diagnosis method for three-phase inverters based on model prediction proposed in patent application number "CN201710500276.X" is particularly suitable for fault diagnosis of complex power electronic circuits with closed-loop control.

[0004] In the prior art, such as the diagnostic method proposed in patent application number "CN201710500276.X", the features in the fault are not distinguished, so the diagnostic model treats important features and minor features equally. When the sample data volume is large, it cannot separate important features and minor features, resulting in insufficient accuracy of the model diagnosis. Summary of the Invention

[0005] The purpose of this invention is to provide a high-precision method for diagnosing electronic circuit faults, so as to solve the problems mentioned in the background art.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A high-precision electronic circuit fault diagnosis method includes the following steps:

[0008] S1. Set different sized capture frames for the multidimensional output voltage data of the circuit to capture the data;

[0009] S2. Encode the data extracted in step S1 using GADF, expand the data, and finally form fault data containing a two-dimensional feature image of fault characteristics.

[0010] S3. The fault data formed in step S3 is divided into training data and test data in a 7:3 ratio. The fault feature images in the training data are used as input data for the CSM-RPGMCBAM-ResNet34 model to train the model and finally establish an electronic circuit fault diagnosis model.

[0011] S4. Input the test data into the established electronic circuit fault diagnosis model to perform the final fault diagnosis.

[0012] Preferably, in step S1, during the two-dimensional feature image production process, a resampling method is used to generate feature maps in order to increase the number of two-dimensional feature maps generated.

[0013] Preferably, in step S3, the method for constructing the CSM-RPGMCBAM-ResNet34 model includes the following steps:

[0014] S31. First, the three-channel two-dimensional feature image is divided into channel one, channel two and channel three. Channel one is designated as the primary operation channel, and channels two and three are designated as secondary operation channels. The primary operation channel and the secondary operation channel are convolved through convolution to achieve three channels respectively. Then, the convolved primary operation channel is mixed with the channel to complete the CSM operation.

[0015] S32. Stack the mixed main operation channels and then splice them with the secondary operation channels as the input of RPGMCBAM;

[0016] S33. Using ResNet34 as the carrier for the input of CSM in step S31 and RPGMCBAM in step S32, a CSM-RPGMCBAM-ResNet34 diagnostic model is constructed by deploying CSM and RPGMCBAM on top of the ResNet34 network.

[0017] Preferably, in step S32, the RPGMCBAM performs preliminary feature extraction on the output of the CSM by adding a reverse Ghost module. The extracted feature tensor is then input into the CBAM. After the CBAM further highlights the important features in the feature tensor, it is input into the topic structure of the convolutional neural network to achieve the final purpose of feature extraction and fault diagnosis.

[0018] Preferably, the reverse Ghost module replaces the DFC attention in the original Ghost convolutional module with a reverse pruning operation, thereby further highlighting the important features in the input Ghost convolutional module and suppressing the secondary features.

[0019] Preferably, the pruning operation first converts the feature tensor into a weight tensor, which is composed of numbers between [0, 1]. The weight tensor is reversed by subtracting the weight tensor from the unit tensor. At this time, the original important features are suppressed and the secondary features are highlighted. Then, appropriate pruning parameters are set to prune the highlighted secondary features. At this time, the weight of the important features becomes 0. Finally, the weight matrix is ​​reversed again, so that the important features are highlighted to the maximum extent and the secondary features are suppressed, generating the final feature weight tensor.

[0020] Preferably, in step S4, after the established electronic circuit fault diagnosis model reaches stability, the model weights are saved, and then the test data is input into the model with saved weights for final fault diagnosis.

[0021] The beneficial effects of this invention are:

[0022] This invention extracts multiple output voltage data from the original fault signal and then performs feature encoding using GADF to transform the original multidimensional fault signal into a two-dimensional feature image. Compared to directly diagnosing the original fault signal, GADF encoding can further characterize the implicit features in the fault signal and simultaneously express multidimensional voltage signals on a single feature image, preserving the temporal and spatial information of the fault signal to the greatest extent. Furthermore, CSM extracts minute features from the original feature image and emphasizes these features through stacking operations. The reverse Ghost module expands the influence range of important features in important channels, and RPGMCPA further highlights the pre-extracted minute features in the feature image. The resulting diagnostic model can effectively extract fault features from the fault signal. With the support of CSM and RPGMCPA, minute features are effectively extracted, improving the accuracy of the diagnostic model and giving it excellent diagnostic performance and robustness.

[0023] This invention employs CSM-RPGMCBAM to pay more attention to minute features in previously overlooked feature channels, thereby making full use of fault features, continuously expanding the influence range of important features in important channels, and increasing the distinction between important and secondary features by selecting appropriate pruning parameters, thus improving the robustness of the diagnostic model. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1This is a flowchart illustrating a preferred embodiment of the present invention;

[0026] Figure 2 This is a schematic diagram of CSM operation in this invention;

[0027] Figure 3 This is a schematic diagram of the channel mixing operation in this invention;

[0028] Figure 4 This is a schematic diagram of the structure of CBAM in the prior art;

[0029] Figure 5 yes Figure 4 A schematic diagram of the spatial attention module;

[0030] Figure 6 yes Figure 4 A schematic diagram of the structure of the mid-channel attention module;

[0031] Figure 7 This is a structural diagram of the Ghost convolutional module in the existing technology;

[0032] Figure 8 This is a schematic diagram of the reverse Ghost convolution structure in this invention;

[0033] Figure 9 This is a schematic diagram of the structure of RPGMCBAM in this invention. Detailed Implementation

[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] A high-precision electronic circuit fault diagnosis method includes three parts: fault data generation, diagnostic model construction, and fault diagnosis.

[0036] The fault data generation method is as follows:

[0037] First, different sizes of capture frames are set for the multidimensional output voltage data of the circuit to capture the data. Then, GADF is used to convert the captured multidimensional electronic circuit voltage data into a two-dimensional feature image containing fault characteristics, thereby generating power circuit fault data.

[0038] In the process of feature image generation, in order to obtain a larger number of feature images for model training, the overlapping sampling method is used to generate feature maps. The overlapping sampling method means that the data frame will always include the data in the previous data frame when it moves, thereby increasing the number of feature maps generated. The overlapping sampling method is also used in Long Short-Term Memory Networks to increase the training samples of the model.

[0039] GADF (Gramian Angular Difference Fields): Gram angular difference fields are a method for encoding feature images of time-series signals, capable of representing the spatial features in time-series signals.

[0040] The diagnostic model is built as follows:

[0041] From the perspective of feature image channels, a CSM-RPGMCBAM structure is constructed. A schematic diagram of CSM operation is shown below. Figure 2 As shown, the three-channel two-dimensional feature image is first divided into channels. Channel 1 is designated as the primary operation channel, and channels 2 and 3 are designated as secondary operation channels. The primary and secondary operation channels are then convolved to achieve three channels. Finally, the convolved primary operation channels are subjected to a channel blending operation, as illustrated in the diagram below. Figure 3 As shown, the information from each channel is mixed to achieve the purpose of information extraction. Figure 2 and Figure 3 In the diagram, number 1 represents channel one, and numbers 2 and 3 represent channels two and three, respectively.

[0042] CSM (Channel Segmentation Mechanism) is a feature highlighting method for features containing fault characteristics. For a three-channel feature image, one channel is selected as the primary operation channel and the rest are secondary operation channels. Different feature extraction methods are applied to each channel and then the channels are stacked to highlight the important features in the feature image.

[0043] Finally, the main operation channels after mixing are stacked and then spliced ​​with the secondary operation channels as the input of RPGMCBAM.

[0044] like Figures 4-9 RPGMCBAM is a feature highlighting method improved on CBAM. It adds a reverse Ghost module to perform preliminary feature extraction on the output of CSM. The extracted feature tensor is then input into CBAM. After CBAM highlights the important features in the feature tensor again, it is input into the topic structure of the convolutional neural network to achieve the final feature extraction and fault diagnosis.

[0045] The reverse Ghost module is an improvement on the original Ghost convolution module. By replacing the DFC (Dynamic Feature Calibration) attention in the original Ghost convolution module with a reverse pruning operation, the important features in the input Ghost convolution module are further highlighted and the minor features are suppressed.

[0046] The pruning operation first converts the feature tensor into a weight tensor, which is composed of numbers between [0,1]. The weight tensor is reversed by subtracting the weight tensor from the unit tensor. At this point, the important features are suppressed and the secondary features are highlighted. Then, appropriate pruning parameters are set to prune the highlighted secondary features. Ideally, the weights of the important features become 0. Finally, the weight matrix is ​​reversed again to maximize the highlighting of the important features and suppress the secondary features, generating the final feature weight tensor.

[0047] Using ResNet34 as the carrier of CSM and RPGMCBAM, a CSM-RPGMCBAM-ResNet34 diagnostic model is constructed by deploying CSM and RPGMCBAM on top of the ResNet34 network.

[0048] like Figure 4 CBAM (Convolutional Block Attention Module) is a simple and effective attention module for convolutional neural networks. Given any intermediate feature map in a convolutional neural network, CBAM can inject the attention map along two independent dimensions: the channel and the space of the feature map. Then, the attention is multiplied by the input feature map to adaptively refine the input feature map.

[0049] like Figure 5 CAM (Channel Attention Module): Channel attention mechanism is a mechanism that highlights channel information by applying different attention to channels. Each channel in the feature image is treated as a feature detector.

[0050] like Figure 6 SAM (SpatialAttentionModule): Spatial attention mechanism, which is a spatial attention feature map generated by analyzing the relationships within the feature image space, focusing on the location of effective information on the feature image.

[0051] ResNet34 (Residual Networks) convolutional neural network: It is characterized by its ease of optimization and the ability to improve accuracy by increasing its depth. Its internal residual blocks use skip connections to alleviate the gradient vanishing problem caused by increasing the depth in deep neural networks. ResNet34 is a relatively simple ResNet structure, consisting of 34 convolutional layers, including 18 residual blocks.

[0052] GM (GhostModule): The Ghost module is a method of generating more feature maps through inexpensive operations.

[0053] The fault diagnosis method is as follows:

[0054] The data is encoded using GADF to form fault data, and the training and testing data are divided in a 7:3 ratio. The training data is then input into the constructed CSM-RPGMCBAM-ResNet34 diagnostic model for training. When training the model, the fault feature images are first segmented into channels using CSM to extract weak features from the channels. Then, RPGMCBAM is used to highlight important weak features, making the differences in fault features between different faults more obvious, enabling the model to perform the classification task of different faults.

[0055] Once the model reaches stability, the model weights are saved. Then, the test data is input into the model with the saved weights for final fault diagnosis.

[0056] Compared with related technologies, the high-precision electronic circuit fault diagnosis method provided by this invention has the following beneficial effects:

[0057] This invention extracts multiple output voltage data from the original fault signal and then performs feature encoding using GADF to transform the original multidimensional fault signal into a two-dimensional feature image. Compared to directly diagnosing the original fault signal, GADF encoding can further characterize the implicit features in the fault signal and simultaneously express multidimensional voltage signals on a single feature image, preserving the temporal and spatial information of the fault signal to the greatest extent. Furthermore, CSM extracts minute features from the original feature image and emphasizes these features through stacking operations. The reverse Ghost module expands the influence range of important features in important channels, and RPGMCPA further highlights the pre-extracted minute features in the feature image. The resulting diagnostic model can effectively extract fault features from the fault signal. With the support of CSM and RPGMCPA, minute features are effectively extracted, improving the accuracy of the diagnostic model and giving it excellent diagnostic performance and robustness.

[0058] This invention employs CSM-RPGMCBAM to pay more attention to minute features in previously overlooked feature channels, thereby making full use of fault features, continuously expanding the influence range of important features in important channels, and increasing the distinction between important and secondary features by selecting appropriate pruning parameters, thus improving the robustness of the diagnostic model.

[0059] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A high-precision method for diagnosing electronic circuit faults, characterized in that, Includes the following steps: S1. Set different sized capture frames for the multidimensional output voltage data of the circuit to capture the data; S2. Encode the data extracted in step S1 using GADF, expand the data, and finally form fault data containing a two-dimensional feature image of fault characteristics. S3. The fault data formed in step S3 is divided into training data and test data in a 7:3 ratio. The fault feature images in the training data are used as input data for the CSM-RPGMCBAM-ResNet34 model to train the model and finally establish an electronic circuit fault diagnosis model. S4. Input the test data into the established electronic circuit fault diagnosis model to perform the final fault diagnosis; In step S3, the method for constructing the CSM-RPGMCBAM-ResNet34 model includes the following steps: S31. First, the three-channel two-dimensional feature image is divided into channel one, channel two and channel three. Channel one is designated as the primary operation channel, and channels two and three are designated as secondary operation channels. The primary operation channel and the secondary operation channel are convolved through convolution to achieve three channels respectively. Then, the convolved primary operation channel is mixed with the channel to complete the CSM operation. S32. Stack the mixed main operation channels and then splice them with the secondary operation channels as the input of RPGMCBAM; S33. Using ResNet34 as the carrier for the input of CSM in step S31 and RPGMCBAM in step S32, a CSM-RPGMCBAM-ResNet34 diagnostic model is constructed by deploying CSM and RPGMCBAM on top of the ResNet34 network. In step S32, the RPGMCBAM performs preliminary feature extraction on the output of the CSM by adding a reverse Ghost module. The extracted feature tensor is then input into the CBAM. After the CBAM further highlights the important features in the feature tensor, it is input into the topic structure of the convolutional neural network to achieve the final purpose of feature extraction and fault diagnosis.

2. The high-precision electronic circuit fault diagnosis method according to claim 1, characterized in that, In step S1, during the two-dimensional feature image production process, a resampling method is used to generate feature maps in order to increase the number of two-dimensional feature maps generated.

3. The high-precision electronic circuit fault diagnosis method according to claim 1, characterized in that, The reverse Ghost module replaces the DFC attention in the original Ghost convolution module with a reverse pruning operation, which further highlights the important features in the input Ghost convolution module and suppresses the minor features.

4. The high-precision electronic circuit fault diagnosis method according to claim 3, characterized in that, The pruning operation first converts the feature tensor into a weight tensor, which is composed of numbers between [0, 1]. The weight tensor is reversed by subtracting the weight tensor from the unit tensor. At this point, the important features are suppressed and the secondary features are highlighted. Then, appropriate pruning parameters are set to prune the highlighted secondary features. The weight of the important features becomes 0. Finally, the weight matrix is ​​reversed again, so that the important features are highlighted to the maximum extent and the secondary features are suppressed, generating the final feature weight tensor.

5. The high-precision electronic circuit fault diagnosis method according to claim 4, characterized in that, In step S4, after the established electronic circuit fault diagnosis model reaches stability, the model weights are saved, and then the test data is input into the model with saved weights for final fault diagnosis.

Images