Analog circuit soft fault diagnosis method fusing RAM (Random Access Memory) characteristics and Linformer

Abstract

The invention relates to the technical field of circuit fault diagnosis, in particular to an analog circuit soft fault diagnosis method fusing RAM (Random Access Memory) features and Linformer, which comprises the following steps: constructing filter circuit models in different element fault modes; extracting output response curve graphs corresponding to different element parameters and filter circuits; inputting the circuit output response to the 1D-CNN network; performing RAM processing on the output response curve to obtain a corresponding response curve 2D image; inputting the response curve 2D image into a 2D-CNN network; carrying out feature splicing on the output feature map of the 1D-CNN network and the output feature map of the 2D-CNN network, and then inputting the spliced features into a Linformer network; and carrying out feature fusion by a Linformer network, and inputting a result into a classifier to carry out fault classification. According to the method, the problems of difficulty in feature extraction of different types of soft fault diagnosis in an analog circuit and poor accuracy and efficiency compatibility in the existing method are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of circuit fault diagnosis, and particularly relates to a simulation circuit soft fault diagnosis method fusing RAM features and Linformer. BACKGROUND

[0002] Electronic systems are widely used in the fields of communication, control and signal processing, and the stability of analog circuits as a key component directly affects the overall performance of the system.

[0003] However, due to factors such as component aging, process deviation and environmental interference, soft faults are prone to occur in analog circuits during operation; such faults are usually characterized by weak but persistent performance deviation, which is difficult to diagnose and easy to ignore, and if not detected in time, it may cause serious consequences; therefore, building an efficient and intelligent soft fault diagnosis model has become a key problem in the operation and maintenance of analog circuits.

[0004] Traditional analog circuit fault diagnosis methods include PCA, IWOA, ELM, IGOA, KELM, IMODA and ADBN, etc., but these methods have the following shortcomings: 1. Feature extraction relies on domain experience, and the generalization ability is poor; 2. It is difficult to handle weak fault signals in high-dimensional noise background; 3. The process is complex, and the real-time performance is poor, which is difficult to meet the deployment requirements in embedded or edge computing environment.

[0005] In recent years, deep learning technology has gradually emerged in the application of fault diagnosis; including: Patches-CNN, DCAM combined with residual structure, CWT+DSCNN architecture, but these methods have the following shortcomings: 1. Data quality and diversity problem: fault features may be highly overlapped with normal signals, making it difficult to extract fault features; 2. Single multi-dimensional feature fusion: time series data and image data are processed separately, and it is difficult to fuse different types of features; 3. Model calculation efficiency problem: traditional deep learning models have high computational complexity when processing long sequence data. SUMMARY

[0006] In view of the shortcomings of the existing method, the present application solves the problem that the existing method fuses different types of features in analog circuits, and the compatibility of accuracy and efficiency is poor.

[0007] The technical solution adopted by the present application is: a simulation circuit soft fault diagnosis method fusing RAM features and Linformer includes the following steps: Step 1, build a filter circuit model under different element fault modes; extract different element parameters and the corresponding output response curve of the filter circuit; As a preferred embodiment of the present application, the filter circuit model includes: Sallen-key band-pass filter and four-op-amp double second-order high-pass filter.

[0008] As a preferred embodiment of the present application, the fault mode component is a capacitor and / or a resistor.

[0009] As a preferred embodiment of the present application, the fault mode is a deviation from the nominal value of the component.

[0010] Step two, input the output response curve graph into the 1D-CNN network; As a preferred embodiment of the present application, the 1D-CNN network includes: the input feature map sequentially passes through the first CBR module, the second CBR module, the third CBR module and the SEBlock module, and then is connected in residual with the input feature map, and then is input into the Relu function.

[0011] Step three, the output response curve graph is processed by RAM to obtain the corresponding response curve 2D image; the response curve 2D image is input into the 2D-CNN network; As a preferred embodiment of the present application, the RAM processing includes: standardizing the time one-dimensional sequence , constructing a sliding window to extract a local segment, and converting it into a matrix d with n rows and columns; extracting a central feature vector from the matrix ; calculating the cosine of the angle between different local feature vectors and the central feature vector ; combining the cosine of the angle in order to form an angle vector, and constructing a two-dimensional angle difference matrix ; normalizing the R to obtain a central relative matrix .

[0012] As a preferred embodiment of the present application, the 2D-CNN network includes: the first DConv2d, the first PConv2d, the first CBAM, the second DConv2d, the second PConv2d, and the second CBAM are cascaded, then flattened and input into the Linear layer.

[0013] Step four, the output feature map of the 1D-CNN network and the output feature map of the 2D-CNN network are input into the Linformer network after feature splicing for feature fusion; the fused feature map is input into the classifier for fault classification; As a preferred embodiment of the present application, the classifier includes: the first Linear layer and the second Linear layer.

[0014] As a preferred embodiment of the present application, the analog circuit soft fault diagnosis system fusing RAM features and Linformer comprises: a memory for storing instructions executable by a processor; and the processor is configured to execute the instructions to implement the analog circuit soft fault diagnosis method fusing RAM features and Linformer.

[0015] As a preferred embodiment of the present application, the computer readable medium storing computer program codes, when executed by a processor, implements the analog circuit soft fault diagnosis method fusing RAM features and Linformer.

[0016] Advantages of the present application: 1. The present application first introduces the RAM (Relative Angle Matrix) method into the analog circuit diagnosis task, converts the original time sequence signal into a color map, and takes it as a 2D-CNN input to mine spatial features. 2. The present application fuses the improved 1D ResNet and the SE module to extract time sequence features, and introduces the CBAM attention mechanism to improve the feature selection ability. 3. The Linformer is used to replace the traditional Transformer for 1D-2D feature fusion, which significantly reduces the computational complexity in a low-rank approximation manner, and realizes efficient global modeling. 4. Experiments are carried out on Sallen-Key band-pass filter and four-operational amplifier high-pass filter circuits, and the model respectively reaches the accuracy of 99.84% and 98.46%, showing superior accuracy and engineering usability. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is the analog circuit soft fault diagnosis model fusing RAM features and Linformer of the present application; Figure 2 is a schematic diagram of the 1D branch network structure of the present application; Figure 3 is a schematic diagram of the channel attention of the present application; Figure 4 is a schematic diagram of the spatial attention of the present application; Figure 5 is a schematic diagram of the Linformer model of the present application; Figure 6 is a Sallen-key band-pass filter circuit diagram of the present application; Figure 7 is an output response curve of the Sallen-key band-pass filter in different states of the present application; Figure 8 is a four-operational amplifier biquad high-pass filter of the present application; Figure 9 are the output response curves of the four operational amplifier biquad high-pass filter of the present application in different states; Figure 10 is the confusion matrix on the test set of the Sallen-key band-pass filter of the present application; Figure 11 is the confusion matrix on the test set of the four operational amplifier biquad high-pass filter of the present application; Figure 12 is the 2D image generated on two circuits using the RAM method. DETAILED DESCRIPTION

[0018] The present application will be further described below in conjunction with the accompanying drawings and examples, which are simplified schematic diagrams and only schematically illustrate the basic structure of the present application, and thus only show the components related to the present application.

[0019] As shown in Figure 1 , the analog circuit soft fault diagnosis method fusing the RAM feature and the Linformer includes the following steps: Step one, construct the filter circuit model under different element fault modes; extract the output response curves of different element parameters and filter circuits; The filter circuit model includes: Sallen-key band-pass filter and four operational amplifier biquad high-pass filter; As shown in Figure 6 , the Sallen-key band-pass filter circuit diagram, the circuit takes pulse signal V i as the excitation source, sets the signal amplitude to 5V, the pulse width to 10μs, the simulation running time to 80μs, and the maximum sampling interval to 0.15μs to obtain the output response of the circuit within a limited time window; in order to determine the fault sensitive elements, the sensitivity analysis of the circuit is carried out, and the results show that the parameter changes of resistor elements R1, R3 and capacitor elements C1, C2 have a significant impact on the circuit output; therefore, the four elements are selected as the soft fault injection objects in this embodiment; the specific setting is to introduce a ±30% offset on the basis of the nominal value of the element to simulate the slight performance degradation caused by factors such as aging, temperature or noise interference in the actual environment; on this basis, 9 different fault modes are designed, and 200 Monte Carlo analyses are carried out for each mode to enhance the distribution diversity of the samples and the generalization ability of the model; considering that manufacturing tolerances exist in the components of the analog circuit, the tolerance ranges of the resistor and capacitor elements are set to 5% and 10% respectively in this embodiment to more realistically simulate the volatility in actual operation. The fault mode settings are shown in Table 1: Table 1 Fault mode settings of Sallen-key band-pass filter

[0020] As Figure 7 The output response curve of the Sallen-key band-pass filter in different states, the abscissa is time, and the ordinate is the output voltage of U5; As Figure 8 The four-operational amplifier double second-order high-pass filter, when the number of resistors, capacitors and other nonlinear elements in the circuit increases, the circuit shows more complex characteristics; In PSpice, the circuit simulation model is built, the input is a 5V pulse signal, the pulse width is 10μs, the delay is 0.1μs, the period is 1ms, the output point V(OUT) of the operational amplifier (U4) in the circuit is selected as the optimal point to collect data, and the circuit running time is set to 300μs, the maximum interval sampling is 0.3μs; 13 kinds of early circuit faults are analyzed when the key components (R1, R2, R3, R4, C1, C2) deviate from the nominal value by 30%, each group of states is analyzed by 200 times of uniform distribution Monte Carlo, and the fault mode setting is shown in Table 2: Table 2 Fault mode setting of four-operational amplifier double second-order high-pass filter

[0021] As Figure 9 The output response curve of the four-operational amplifier double second-order high-pass filter in different states, the abscissa is time, and the ordinate is the output voltage of U4.

[0022] Step two, input the output response curve into the 1D-CNN network; As Figure 2 The 1D-CNN network is shown, the input feature map is sequentially connected with the input feature map after passing through the first CBR module, the second CBR module, the third CBR module and the SEBlock module, and then input into the Relu function; The 1D-CNN network is used to extract features from the original time series signal (response curve), to capture local patterns and model time-dependent relationships; 1D convolution is widely used in sequence classification, pattern recognition and other tasks due to its high efficiency in processing continuous data; In the first n layer, the convolution operation is represented as: (1) Where, and represent the first j output and the first i input feature map; represents the bias vector of the first j convolution kernel, and is the first j convolution kernel; and respectively represent the input and output feature maps of the first n layer, and the symbol represents the convolution operation.

[0023] To improve the training stability and convergence speed of the model, batch normalization (BN) is introduced after the convolution layer, which is calculated as follows: (2) where, and are the mean and variance of the current batch to prevent division by a small constant, and are trainable scaling and offset parameters; the BN operation keeps the input of each layer within a stable distribution range, which helps to speed up the training process and suppress overfitting; In addition, to further improve the feature learning ability of the deep network and alleviate the gradient vanishing problem, a residual connection mechanism is introduced in the 1D convolution module. The residual connection (skip connection) is a key feature of ResNet, which allows the network to learn the residual mapping (i.e., the difference between the input and the output) rather than learning the entire mapping. This approach effectively alleviates the gradient vanishing problem caused by deepening the network; it allows information to bypass the nonlinear transformation and pass directly from the input to the output, which can be mathematically expressed as: (3) where, x is the input of the structure, H ( x ) is the identity mapping function from input to output, F ( x ) is the fitting residual function; the residual block includes convolution layers, batch normalization layers, and nonlinear activation function layers; X is the input of the residual block, and the output Y can be expressed as: (4) where, is the nonlinear transformation function of the residual network, is the weight parameter, is the activation function.

[0024] Through the joint modeling of multiple layers of 1D convolution, BN normalization, ReLU activation, and residual connection, the continuous voltage signal of the analog circuit can be expressed in depth.

[0025] Step three, perform RAM processing on the output response curve to obtain the corresponding response curve 2D image; input the response curve 2D image into the 2D-CNN network; RAM method converts time series into two-dimensional color image, which can intuitively show the dynamic correlation between signals and provide spatial feature input for 2D-CNN; To further illustrate the application of RAM method in the present application, the following will be described in detail in combination with formulas and flowcharts: Collecting time one-dimensional sequence , wherein, represents the value of the i-th, i is the length of the time sequence; first, standardization is performed, and the formula is: m (5) (6) (7) After standardizing the sequence, a sliding window is constructed to extract local fragments; let the window length be d and the step length be h, and n local vectors are generated: in order to reduce the dimension, the standardized time sequence is converted into a matrix with rows and columns in the following manner, and the formula is: d n (8) Each column is regarded as a local feature vector of the time sequence, as follows: (9) The central feature vector can be obtained as follows: (10) The cosine value between different local feature vectors and the central feature vector is calculated, and the formula is: (11) , wherein, is the cosine angle between and and

[0026] All angles are combined in order to form an angle vector, and a two-dimensional angle difference matrix R is further constructed, and the formula is: (12) R By calculating the angle between each local time sequence and the center, and then evaluating the relative relationship between the angle matrix, the overall features of the time sequence are captured; Finally, the matrix is mapped to the interval [0, 255] through normalization, and each element is converted into a gray value, and then mapped to RGB image to obtain the center relative matrix, and the formula is: ​​​​ (13) The present invention uses the RAM method to generate a visual color image of a uniform size of 64×64 as the input of the 2D-CNN, revealing the small but critical differences in the soft fault signal in the two-dimensional space, significantly improving the feature perception ability of the subsequent network. The unique texture, color image features and local changes of the two-dimensional image reflect the fault mode and fault frequency of the circuit signal, so as to further understand and analyze the importance of these visual clues for circuit fault diagnosis. The circuit fault signal is converted into a 2D image by the RAM method. Figure 12 shown.

[0027] Two-dimensional convolutional neural networks (2D-CNNs), a typical deep learning architecture, are widely used in image processing and computer vision tasks, particularly for extracting spatial features from two-dimensional data. In this paper, 2D-CNNs are used to process fault feature images generated using the RAM (Relative Angle Matrix) method to effectively capture local spatial structure and pattern information, providing a more intuitive feature representation for circuit fault diagnosis. To further reduce model computational complexity and improve operational efficiency, a depthwise separable convolution (DSC) structure is introduced into 2D-CNNs as an alternative to the feature extraction module.

[0028] Depthwise Convolution: Each input channel is convolved independently without mixing channel information. The number of convolution kernels is the number of input channels. The formula is: (14) in, Y i It is i output channels, W i is the corresponding convolution kernel, X i Is the input i channels Pointwise Convolution: Use 1×1 convolution to fuse all channel outputs to achieve cross-channel information integration; (15) in, Y i are the outputs from the depthwise convolution, and the pointwise convolution combines them via a 1x1 convolution operation.

[0029] This method reduces the number of parameters of the standard convolution by Dk×Dk×C in ×C outThe reduction is: Dk x Dk x C in +1 x 1 x C in x C out ; wherein, Dk is the size of the convolution kernel, C in , C out are the input and output channel numbers respectively.

[0030] Compared with traditional convolution, depth separable convolution greatly reduces the computational complexity and parameter size of the model while ensuring the effect of feature extraction, making the model more suitable for deployment on resource-constrained edge devices.

[0031] CBAM (Convolutional Block Attention Module) is a lightweight attention mechanism designed to enhance the performance of convolutional neural networks by focusing on important features; the key components of CBAM are: For example Figure 3 Channel Attention (CA): helps the model focus on the most important channels by generating a channel attention map through global average pooling and two convolution layers, then multiplying the map with the input feature map to enhance important channels and suppress unimportant channels.

[0032] For example Figure 4 Spatial Attention (SA): this part focuses on important regions in the feature map; by calculating the mean and maximum values along the channel axis and concatenating the two maps, a spatial attention map is generated using a convolution layer; finally, the input feature map is multiplied by the spatial attention map to highlight important spatial regions.

[0033] CBAM can selectively highlight the most information-rich parts of the feature map to obtain superior fault diagnosis results; the channel attention mechanism is used to reconstruct the input feature map, which is then further refined by the spatial attention mechanism to produce the final feature map.

[0034] CBAM channel attention operation: first perform global average pooling and maximum pooling on the input feature map to obtain global features for each channel, then use a fully connected layer to convert the pooling results into channel descriptors, and then perform fusion operation to obtain channel attention weights through the Sigmoid function. Weight the original feature map to highlight important channels and suppress unimportant channels. The calculation of the channel attention map is given by the following formula: (16) (17) CBAM spatial attention operation: for the channel-enhanced feature map F ′, the spatial attention module continues to focus on the most important regions in the image along the spatial dimension; first, the channel-enhanced feature map FGlobal average pooling (GAP) and global maximum pooling (GMP) are performed to generate two spatial descriptor maps respectively, which can be used to represent the importance of each spatial position; the two spatial descriptors are spliced together along the channel dimension to form a unified spatial representation; after splicing, a 7x7 convolution operation is performed to capture local spatial dependencies; finally, a spatial attention map Ms is generated through a Sigmoid activation function, which represents the weight of each spatial position; finally, the channel-enhanced feature F' is weighted through the spatial attention map Ms to generate the final enhanced feature F''; then, the refined feature map can be used for subsequent layers of the convolutional neural network, so as to further enhance and reconstruct the feature.

[0035] (18) (19) The CBAM module is used after the application of the depth separable convolution; the depth separable convolution reduces the amount of calculation and the number of parameters, and the CBAM enhances the attention of the model to important features, so that the model can automatically focus on important channels and spatial regions without increasing significant calculation overhead.

[0036] Step four, the output feature map of the 1D-CNN network and the output feature map of the 2D-CNN network are spliced and input into the Linformer network; the output feature map of the Linformer network is input into the classifier for fault classification; In the simulation circuit soft fault diagnosis task, the information contained in the voltage signal has time and space dual characteristics; in order to fully fuse the multi-dimensional information extracted by the 1D and 2D convolution branches, the present application introduces a Linformer network as a feature fusion unit, which effectively models the long-distance dependency relationship between sequences, while avoiding the high computational overhead problem faced by traditional Transformers when processing long sequences; In the traditional Transformer model, the calculation of self-attention is based on a context mapping matrix P, and its calculation formula is as follows: (20) wherein, Q is the query matrix (Query), with a shape of nxd, K is the key matrix (Key), with a shape of nxd, dk is the dimension of the key; P is the context mapping matrix, which captures the dependency relationship between each element in the sequence, with a calculation complexity of O(n^2), wherein n is the length of the sequence; it is not suitable for large-scale feature maps or long sequence inputs.

[0037] As Figure 5, Linformer maps the original sequence length from n to a low-dimensional space k (k n) by imposing a low-rank projection on the key (K) and value (V) matrices, significantly reducing the attention matrix computation overhead; the core calculation is as follows: (21) (22) Wherein, E is a linear projection matrix of size k x n, F is a learnable linear projection matrix of size k x n, and are the low-dimensional representations of the keys and values, respectively; is used to project the value matrix V ; Linformer reduces the attention computation complexity from O (n^2) to O (n), greatly improving the inference efficiency.

[0038] In the model, the time sequence features F1D from the 1D convolution branch and the spatial features F2D from the 2D convolution branch are spliced and input into the Linformer module; the global dependency relationship between the channels and the space-time is modeled through the self-attention mechanism, and the fusion and reconstruction in the unified feature space are completed; the fused representation is further input into the fully connected layer and the Softmax classifier to complete the final fault type discrimination.

[0039] The detailed structure of the model is shown in Table 3: Table 3 Network structure parameters of the application

[0040] Experimental process: To verify the effectiveness of the model in the simulation circuit soft fault diagnosis, experiments were conducted on two typical circuit structures, and the experimental process consisted of five steps: data set construction, fault mode setting, data division, model training and fault diagnosis; when performing simulation analysis, PSpice in Capture CIS 24.1 was used to simulate the structure of the analog circuit; in the Python environment, the circuit fault diagnosis experiment was carried out based on the Pytorch framework.

[0041] To ensure the stability of the model training and the reliability of the evaluation results, the application standardized and scientifically divided the constructed fault data set; the specific steps are as follows: Firstly, the analog signal output by the circuit is subjected to standard normalization processing, and all data is uniformly mapped to the interval [0, 1] to eliminate the dimension effect and accelerate the model convergence speed; in this experiment, a self-defined data set class is adopted, which contains time sequence signals (1D) and image data (2D) from the analog circuit; 200 times of Monte Carlo analysis are performed for each fault state, so that circuit one has 1800 samples of 9 modes (1 normal mode and 8 fault modes); circuit two has 2600 samples of 13 modes (1 normal mode and 12 fault modes); in order to ensure the diversity and balance of the training data, the training set, the validation set and the test set are divided according to the ratio of 7:2:1; in order to improve the training effect and robustness of the model, the following training configuration is adopted for model construction and optimization: the training platform adopts Windows 11 system, NVIDIA RTX 4070 Laptop GPU, PyTorch 1.10 deep learning framework, and the loss function is cross entropy loss (Cross Entropy Loss); when training the model, the Adam gradient optimizer is selected to update the model parameters, the learning rate is set to 0.001, and the cosine annealing (CosineAnnealingLR) strategy is used for dynamic attenuation; the training process is performed for 100 epochs, and the batch size is 16.

[0042] The diagnosis experiment proves the diagnosis performance of the method of the application; the diagnosis accuracy is selected as the performance evaluation index of fault diagnosis, and the confusion matrix is used to represent the accuracy of fault diagnosis; in the confusion matrix, the row represents the predicted fault state, and the column represents the true fault state; the diagonal value represents the diagnosis accuracy of each fault state, such as Figure 10 、 Figure 11 The diagnosis results of circuit one and circuit two on the test set are shown in Table 4.

[0043] The application uses several different deep learning models to compare on the same data set of the four-operational-amplifier double-second-order high-pass filter circuit of circuit two, and evaluates the performance on the test set; as shown in Table 4, the diagnosis results of 10 experiments are shown in Table 4. Specifically, 1D-CNN, LSTM, WT-PCA-ELM and the model proposed in the application are compared.

[0044] Table 4 Comparison of accuracy of different models

[0045] By comparing the simulation results, the 1D-CNN is set to 3 layers with kernel sizes of 3, 5 and 7, and performs well in accuracy, reaching 88.46%; the application effectively extracts the local features of the input signal through the one-dimensional convolution layer, so it performs well in processing this task; the WT-PCA-ELM model combines wavelet transform (WT) and principal component analysis (PCA) to extract features, and classifies through extreme learning machine (ELM), with an accuracy of 91.77%; the effect of LSTM in this task is not as good as the previous two; the LSTM model structure is less adaptable to this data set, which may be due to the sparsity of long sequence data, making it difficult to effectively capture local features; it is shown that in terms of feature extraction, WT-PCA-ELM provides a relatively effective solution, while the application model combines 1D time sequence features and 2D spatial characteristics, and finally achieves the highest accuracy (98.46%), showing strong advantages.

[0046] Based on the experimental results, the application model performs best in the diagnosis task, with an accuracy of 98.46%, which is significantly improved compared to traditional 1D-CNN and LSTM models; further research can optimize the application model to improve its robustness and generalization ability on larger data sets.

[0047] Ablation experiment: To further verify the effectiveness of each module in the application model, the following three groups of ablation experiments are designed, which analyze from three dimensions of feature fusion strategy, model branch structure and attention mechanism introduction; The experiments are carried out on two data sets of circuit one and circuit two, under the same training and testing conditions, and each method is repeated 20 times independently.

[0048] The ablation results are shown in Table 5, where method A removes the CBAM module, and the output of the 2D convolution part no longer passes through the channel attention and spatial attention mechanisms, but is directly passed to the subsequent layer; method B removes the Linformer module, and uses a simple fully connected layer to fuse 1D and 2D features; method C evaluates the influence of removing the 2D convolution part (image data processing part) on the model performance; only using the 1D convolution part (time sequence data processing part); Finally, the application method; analyze its influence on model accuracy, training time and computational efficiency; Table 5 Comparison of the application method with existing methods

[0049] The experimental results of method A and method B show that the accuracy of the model on the test set decreases to different degrees after removing the CBAM and Linformer modules, indicating that the self-attention mechanism has certain effect in capturing global features, and the Linformer feature fusion is more advantageous than using a simple full connection layer; especially in processing time series data and global information fusion, the CBAM and Linformer modules have great contribution to the model; although the calculation efficiency and training time of the two are slightly improved, the performance decreases; the method C experiment shows that after removing the 2D convolution part, only using the 1D time series feature, the performance of the model for slightly complex circuits decreases significantly; proving the importance of 2D convolution in the task.

[0050] The present application proposes a DualCNN-Linformer-CBAM architecture around the efficient identification of soft faults in analog circuits; the model converts the time series signal into a two-dimensional feature map by introducing the RAM method, extracts 1D and 2D features in parallel using a dual-channel network, enhances the feature expression ability combined with the CBAM attention mechanism, and finally realizes efficient global feature fusion with the help of the Linformer module. A large number of experiments on two typical circuits of Sallen-Key band-pass filter and four-op-amp high-pass filter show that the proposed model achieves a diagnostic accuracy of 99.84% and 98.46% on the test set, respectively, which is significantly better than the existing 1D-CNN, LSTM and WT-PCA-ELM benchmark methods.

[0051] Based on the above ideal embodiments according to the present application, through the above description, relevant personnel can make various changes and modifications without deviating from the technical idea of the present application. The technical scope of the present application is not limited to the contents of the specification, and the technical scope must be determined according to the scope of the claims.

Claims

1. The analog circuit soft fault diagnosis method integrating RAM features and Linformer is characterized by: The following steps are involved: Step 1: Construct filter circuit models under different component failure modes; extract output response curves corresponding to different component parameters and filter circuits; Step 2: Input the output response curve graph into the 1D-CNN network; Step 3: Process the output response curve graph in RAM to obtain a corresponding 2D image of the response curve; input the 2D image of the response curve into the 2D-CNN network; Step 4: Concatenate the output feature maps of the 1D-CNN network and the 2D-CNN network and input them into the Linformer network for feature fusion; the fused feature maps are input into the classifier for fault classification.

2. The analog circuit soft fault diagnosis method integrating RAM features and Linformer according to claim 1 is characterized in that: The 1D-CNN network includes: the input feature map passes through the first CBR module, the second CBR module, the third CBR module and the SEBlock module in sequence, and then performs a residual connection with the input feature map, and then inputs the ReLU function.

3. The analog circuit soft fault diagnosis method integrating RAM features and Linformer according to claim 1 is characterized in that: RAM processing includes: For one-dimensional time series Normalize, extract local fragments using sliding windows, and convert them into matrices ; Pair Matrix Extract the central eigenvector; Calculate the cosine of the angle between different local eigenvectors and the central eigenvector ; Combine the cosine values ​​of the angle in sequence to form an angle vector and construct a two-dimensional angle difference matrix ; right R Perform normalized mapping to obtain the center relative matrix .

4. The analog circuit soft fault diagnosis method integrating RAM features and Linformer according to claim 1 is characterized in that: The 2D-CNN network includes: the first DConv2d, the first PConv2d, the first CBAM, the second DConv2d, the second PConv2d, and the second CBAM are cascaded, flattened, and then input into the Linear layer.

5. The analog circuit soft fault diagnosis method integrating RAM features and Linformer according to claim 1 is characterized in that: Classifiers include: The first Linear layer and the second Linear layer.

6. The analog circuit soft fault diagnosis method integrating RAM features and Linformer according to claim 1 is characterized in that: Filter circuit models include: Sallen-key bandpass filter and four-op-amp biquad high-pass filter.

7. The analog circuit soft fault diagnosis method integrating RAM features and Linformer according to claim 1 is characterized in that: The components of the failure mode are capacitors and / or resistors.

8. The analog circuit soft fault diagnosis method integrating RAM features and Linformer according to claim 1 is characterized in that: The failure mode is a deviation from the nominal value of the component.

9. The analog circuit soft fault diagnosis system integrating RAM features and Linformer is characterized by: include: a memory for storing instructions executable by the processor; A processor, configured to execute instructions to implement the analog circuit soft fault diagnosis method integrating RAM features and Linformer as described in any one of claims 1 to 8.

10. A computer-readable medium storing computer program code, characterized in that When the computer program code is executed by a processor, the computer program code implements the analog circuit soft fault diagnosis method integrating RAM features and Linformer as described in any one of claims 1 to 8.

Images