Lightweight method and system for real-time fault data of analog circuits based on depthwise separable convolution

Abstract

The present invention relates to the field of analog circuit fault data processing technology, specifically a method and system for lightweighting real-time fault data of analog circuits based on deep separable convolution, comprising: S1, determining the evaluation index of the lightweight model of real-time fault data of analog circuits, and calculating the evaluation index according to a standard convolutional network; S2, using a deep separable convolutional network to complete the construction of the lightweight model of real-time fault data of analog circuits; S3, obtaining real-time fault data of analog circuits through experiments and expanding the sample size through a data enhancement algorithm, and completing the optimization using the lightweight model of fault data. The present invention first determines the evaluation criteria for the degree of lightweighting of the model, improves the existing model, and performs global and local lightweight design based on deep separable convolution, obtains a lightweight model of fault data under the condition that the accuracy, complexity, and parameter quantity are balanced to meet actual needs, lightweights the collected fault data, and finally verifies the effect of the method through performance parameter comparison.

Description

Technical Field

[0001] The present invention belongs to the technical field of analog circuit fault data processing, and in particular relates to a method and system for lightweight quantization of analog circuit real-time fault data based on depthwise separable convolution. Background Art

[0002] Real-time monitoring requires the delivery of continuously updated data on systems or events. This type of monitoring provides information with zero or low latency, minimizing the delay between data collection and analysis. It enables rapid detection of anomalies and performance issues. Research in the field of real-time signal monitoring has long followed two paths: chip development, where advancements in process technology and integrated circuit design improve chip storage and computing resources to reduce latency and enhance data collection and processing efficiency. Algorithm research, on the other hand, aims to maximize efficiency in data processing by reducing computational overhead through improved algorithmic models. Currently, real-time monitoring scenarios primarily focus on terminals or mobile devices, with two deployment methods: online deployment, where sensors collect signals and transmit them to terminals. The terminals then send the data to a host computer, which processes and transmits the data back to the terminal. Terminals with low computing power and performance often act as "middlemen." However, high-reliability operation in online mode requires a smooth and stable network environment, which incurs significant communication costs. With 5G networks not yet fully covered, the speed of information transmission during real-time monitoring is severely constrained by the network environment. The second approach is offline deployment. In this scenario, the terminal and host computer are combined into one, and data received from sensors is processed directly on the terminal. However, the terminal's computing and storage resources are significantly limited. Furthermore, FPGAs, another commonly used terminal device, have less than 10MB of on-chip memory. Therefore, when deploying algorithms on terminal devices, computational and storage consumption are key considerations, in addition to accuracy. In summary, research on real-time monitoring from the perspective of lightweight algorithms is essential.

[0003] Mobile Net is a classic model for lightweight structural design. By employing depthwise separable convolutional modules, it achieves a classification accuracy of 70.6% on the ImageNet dataset, achieving prediction accuracy comparable to that of deep convolutional networks while reducing the number of parameters by 32 times and the model computational overhead by 27 times. By adding width and resolution factors as trainable network parameters, the network achieves lightweightness while maintaining accuracy. Building on Mobile Net's foundation, Mobile Net V2 was developed, combining it with a residual network architecture. Its standard convolutional module first expands the number of feature channels in the input data using 1×1 convolution kernels, then performs depthwise convolution with 3×3 kernels to extract single-channel feature information. Finally, it uses 1×1 convolution to reduce the channel dimensionality, and finally concatenates the output channels with the residual network module. Mobile Net V3 incorporates the feature attention mechanism proposed by SENet into Mobile Net V2, designs a novel nonlinear activation function, and uses neural network architecture search to determine the number of convolution kernels at each layer, further improving network performance. In addition to building such small and efficient neural networks, another way to obtain lightweight networks is to shrink and decompose neural networks. Compression methods include product quantization, vector quantization, and Huffman coding. Decomposition methods include factoring models across channels or filters to speed up pre-trained networks.

[0004] In order to adapt to real-time monitoring scenarios, while solving the noise problem, the computer computing power problem is considered to achieve the maximum balance between accuracy and real-time performance. This application proposes a lightweight diagnostic model method based on deep separable convolution, which offloads computing power while ensuring the accuracy of the fault diagnosis model, reduces latency, and is conducive to the integration, miniaturization, high-speed and real-time realization of accurate and reliable monitoring and diagnosis. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention provides a lightweight method for real-time fault data of analog circuits based on deep separable convolution. First, two evaluation criteria for measuring the degree of model lightweighting are determined. Then, improvements are made on the existing model. Based on the idea of ​​deep separable convolution, global and local lightweight designs are performed respectively. While ensuring a balance between accuracy, complexity and parameter quantity, a lightweight model of fault data that meets actual application needs is obtained. Then, the real-time collected fault data is lightweight processed. Finally, the effectiveness of this method is verified by comparing performance parameters.

[0006] To achieve the above objectives, the present invention discloses the following technical solutions:

[0007] A lightweight method for real-time fault data of analog circuits based on depthwise separable convolution, comprising:

[0008] S1: Determine the evaluation index of the lightweight model of real-time fault data of analog circuits and calculate the evaluation index based on the standard convolutional network;

[0009] The number of parameters and floating-point operations are used to evaluate the lightweight model of real-time acquisition of fault data of analog circuits. The three dimensions of the standard convolutional network feature map are set as: height, width and number of channels, which are combined as [H, W, C1]. The convolution kernel of the standard convolutional network is set as [K w ,K h′ ,C1], the number of filters is C2, the dimension of the feature map after the standard convolution network is [H, W, C2], and the analog circuit parameter quantity index and the analog circuit floating-point operation number evaluation index are calculated according to the standard convolution network;

[0010] S2: Use a deep separable convolutional network to build a lightweight model for real-time fault data of analog circuits;

[0011] S21: Based on the grouped convolution operation, the depth-wise separable convolutional network sets a filter for each channel of the input feature to perform sliding operation, reducing the number of parameters and regularizing; the standard convolutional network filter K with a size of W×W×M and a channel number of 1 is applied to the D f ×D f The input feature vector F of ×N is used to obtain the output simulation circuit fault data feature vector through the deep separable convolutional network.

[0012] S22: After filtering the input channel through the depthwise separable convolutional network, the features are combined using a pointwise convolutional network to generate a first analog circuit real-time fault data feature; the output vector of the depthwise separable convolutional network model is determined to be:

[0013]

[0014] Among them, G k,l,n The first analog circuit feature vector output by the depthwise separable convolutional network model; is the standard convolutional network filter; k is the width of the analog circuit feature vector; L is the length of the analog circuit feature vector; is the output analog circuit fault data feature vector; m is the number of convolution kernels; n is the output value of the number of convolution kernels;

[0015] S23: Design the overall network structure based on the reduction of the parameter quantity index and floating-point operation number index in step S1 using a depthwise separable convolutional network to improve the classification effect of the lightweight model of fault data;

[0016] S231: Design a first lightweight model DS1 and a second lightweight model DS2. First, set a depthwise separable convolutional network with a convolution kernel of 3×3, apply a single convolution filter to each input channel to perform lightweight filtering, and then use a point convolutional network with a convolution kernel of 1×1 to construct the real-time fault data features of the second analog circuit by calculating the linear combination of the input channels, so that the real-time fault data maintains the dimensionality of the input.

[0017] S232: Based on the first lightweight model DS1 and the second lightweight model DS2, the backbone of the fault data lightweight model is formed. The backbone is optimized by adopting a local lightweight design. The standard convolutional network and the depthwise separable convolutional network are alternately combined to achieve a balance between model lightweight and model accuracy. The first backbone structure DS_v1 and the second backbone structure DS_v2 of the two types of backbones are designed.

[0018] S3: Acquire real-time fault data of simulated circuits through experiments, expand the sample size through data enhancement algorithms, and use a lightweight model of fault data to complete optimization;

[0019] The sample size of the collected real-time fault data of the analog circuit is expanded by using a time series data enhancement algorithm; the obtained real-time fault data of the analog circuit is imported into the fault data lightweight model in step S2 for processing, thereby completing the lightweight optimization of the real-time fault data of the analog circuit.

[0020] Preferably, the analog circuit parameter index in step S1 is:

[0021] Params(SdCl)=C1K w K h C2;

[0022] Among them, Params (SdCl) is the parameter index of the analog circuit; C1 is the number of channels of the standard convolutional network; K w is the standard convolutional network filter width; K h is the height of the standard convolutional network filter; C2 is the number of filters; C1K w K h is the number of multiplication operations of a single convolutional layer.

[0023] Preferably, the floating-point operation number indicator of the analog circuit in step S1 is specifically:

[0024] The analog circuit floating-point operation index is used to measure the complexity of the algorithm. The computational load in the convolutional network is mainly borne by forward reasoning. The forward reasoning process is a multiplication-accumulation calculation at the mathematical level. Therefore, the floating-point operation number is mainly composed of the number of multiplication-addition operations. The floating-point operation number index is:

[0025] FLOPs(SdCl)=[C1K w Kh+(C1K w K h -1)]=(2C1K w K h -1)×C2WH;

[0026] Among them, FLOPs (SdCl) is the number of floating-point operations of analog circuits; C1K w K h -1 is the number of addition operations of a single convolutional layer; W is the standard convolutional network width; H is the standard convolutional network height.

[0027] Preferably, step S21 obtains the output simulation circuit fault data feature vector by a deep separable convolutional network Specifically:

[0028] The standard convolutional network filter K is used for the channel of the input feature vector F, and m groups of input feature vectors F are slid and cut with m standard convolutional network filters K respectively, and m groups of results are obtained by splicing to obtain the output analog circuit fault data feature vector for:

[0029]

[0030] Among them, K i,j,m is the standard convolutional network filter; F k+i-1,l+j-1,m is the input feature vector; i is the first position change of the convolution kernel; j is the second position change of the convolution kernel.

[0031] Preferably, the point convolution network in step S22 is a standard convolution network with convolution kernel The size is [1,1], which is equivalent to weighted combination of the feature map of the analog circuit fault data in the depth direction, which can reduce the dimension of the feature of the analog circuit fault data and increase the nonlinearity.

[0032] Preferably, the backbone part of the fault data lightweight model in step S232 has the feature extraction capability, and a 3×3 standard convolutional network is placed at the input end of the depthwise separable convolutional network, so as to extract the fault data features at the input end.

[0033] Preferably, the first trunk part structure DS_v1 and the second trunk part structure DS_v2 of the two types of trunk parts in step S232 are specifically:

[0034] The first backbone structure DS_v1 replaces the second lightweight model DS2 with the second deep convolutional neural network Res2Net based on the backbone of the classic convolutional network MobileNet;

[0035] The second backbone structure DS_v2 replaces the first lightweight model DS1 with the first deep convolutional neural network Res1Net based on the backbone of the classic convolutional network MobileNet.

[0036] Preferably, the time series data enhancement algorithm in step S3 has a unified API interface, which can control the probability of occurrence of each function implementation so that the randomly enhanced fault data has a desired distribution.

[0037] In a second aspect of the present invention, a model construction system for a method for lightweight quantization of real-time fault data of analog circuits based on the aforementioned depthwise separable convolution is proposed, which comprises: a depthwise separable convolutional network module, a pointwise convolutional network module, a residual network module, and a backbone module of a fault data lightweight model;

[0038] The depthwise separable convolutional network module uses two types of modules, depthwise convolution and pointwise convolution, instead of standard convolution to filter and combine the input fault data features respectively;

[0039] The point convolutional network module combines the input fault data features to generate new fault data features;

[0040] The residual network module is based on the AdaBN residual network algorithm for lightweight improvement, which can ensure accuracy while reducing the number of parameters and floating-point operations of the fault data lightweight model;

[0041] The backbone module of the fault data lightweight model includes a first lightweight model DS1 and a second lightweight model DS2; the overall network structure is designed on the basis of reducing the number of parameters and the number of floating-point operations of the fault data lightweight model to ensure the model classification effect and fully extract the fault data features near the input end.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] (1) Based on the computational and storage consumption issues of the data lightweight algorithm model, the present invention first proposes two evaluation criteria for measuring the degree of model lightweighting. Based on the existing model, the design improvement is carried out. Based on the idea of ​​deep separable convolution, the model is designed globally and locally lightweight respectively. Under the condition of ensuring the balance between accuracy, complexity and number of parameters, a fault data lightweight model that meets the actual application needs is obtained.

[0044] (2) The present invention collects and processes fault data at the physical level, uses the Tsaug data enhancement algorithm to expand the data volume, performs lightweight processing on the real-time collected data after data enhancement, and finally verifies the effectiveness of this method through performance parameter comparison. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a flow chart of the method for lightweight quantization of real-time fault data of analog circuits based on depthwise separable convolution of the present invention;

[0046] Figure 2 This is the standard convolution and depth-separable convolution structure diagram of the present invention;

[0047] Figure 3 This is a design diagram of the fault data lightweight model of the present invention;

[0048] Figure 4 This is a design diagram of the lightweight backbone structure of the present invention;

[0049] Figure 5 This is a design diagram of the lightweight overall network architecture of the present invention;

[0050] Figure 6 This is the structural diagram of the first backbone structure DS_v1backbone of the present invention;

[0051] Figure 7 This is a structural diagram of the second backbone structure DS_v2backbone of the present invention;

[0052] Figure 8 This is a structural design diagram of the partial lightweight overall model of the present invention;

[0053] Figure 9 This is a diagram for evaluating the comprehensive performance of the model lightweighting and accuracy of the present invention;

[0054] Figure 10 This is the confusion matrix structure diagram of the first main structure DS_v1 of the present invention;

[0055] Figure 11 This is the confusion matrix structure diagram of the second main structure DS_v2 of the present invention;

[0056] Figure 12 This is the original waveform diagram of the example signal of the present invention;

[0057] Figure 13 This is a waveform diagram of an example signal enhancement of the present invention;

[0058] Figure 14 This is the waveform diagram of the measured signal enhancement of the present invention;

[0059] Figures 15(a) and 15(b) are diagrams showing the ResNet_AdaBN_DS_v1 training process of the present invention;

[0060] Figures 16(a) and 16(b) are diagrams showing the ResNet_AdaBN_DS_v2 training process of the present invention;

[0061] Figure 17 This is the ResNet_AdaBN_DS_v1 dimensionality reduction visualization result diagram of the present invention;

[0062] Figure 18 This is the ResNet_AdaBN_DS_v2 dimensionality reduction visualization result diagram of the present invention;

[0063] Figure 19 This is a diagram of the classification index measurement results of the present invention;

[0064] Figure 20 This is a diagram evaluating the comprehensive performance of the model lightweighting and accuracy of the present invention. DETAILED DESCRIPTION

[0065] The exemplary embodiments, features, and aspects of the present invention will be described in detail below with reference to the accompanying drawings. The same reference numerals in the accompanying drawings represent elements with the same or similar functions. Although various aspects of the embodiments are shown in the accompanying drawings, the drawings are not necessarily drawn to scale unless otherwise indicated.

[0066] The embodiment of the present invention provides a lightweight method for real-time fault data of analog circuits based on depthwise separable convolution, such as Figure 1 As shown in the figure, the evaluation index of the lightweight model of real-time fault data of analog circuits is determined and calculated based on the standard convolutional network. The lightweight model of real-time fault data of analog circuits is constructed using a depthwise separable convolutional network. The real-time fault data of analog circuits is obtained through experiments and the sample size is expanded through a data enhancement algorithm. The lightweight model of fault data is used to complete the optimization. The method includes:

[0067] Step S1: Determine the evaluation index of the lightweight model of the analog circuit real-time fault data, and calculate the evaluation index according to the standard convolutional network.

[0068] The number of parameters and floating-point operations are used to evaluate the lightweight model of real-time acquisition of fault data of analog circuits. The three dimensions of the standard convolutional network feature map are set as: height, width and number of channels, which are combined as [H, W, C1]. The convolution kernel of the standard convolutional network is set as [K w ,K h ,C1], the number of filters is C2, the dimension of the feature map after the standard convolutional network is [H, W, C2], and the analog circuit parameter quantity index and the analog circuit floating-point operation number evaluation index are calculated according to the standard convolutional network.

[0069] The analog circuit parameter quantity index is specifically: the sum of the number of parameters involved in the algorithm operation. As an important indicator for evaluating the efficiency of the fault data lightweight model, the parameter quantity is strongly correlated with the model storage space. The larger the model scale, the larger the storage space required. The analog circuit parameter quantity index is:

[0070] Params(SdCl)=C1K w K h C2;

[0071] Among them, Params (SdCl) is the parameter index of the analog circuit; C1 is the number of channels of the standard convolutional network; K w is the standard convolutional network filter width; K h is the height of the standard convolutional network filter; C2 is the number of filters; C1K w K h is the number of multiplication operations of a single convolutional layer.

[0072] The floating-point operation count of analog circuits is used to measure algorithm complexity. The computational load in convolutional networks is mainly borne by forward reasoning, which is a mathematical multiplication-accumulation process. Therefore, the floating-point operation count is mainly composed of the number of multiplication-addition operations. The floating-point operation count index is:

[0073] FLOPs(SdCl)=[C1K w K h +(C1K w K h -1)]=(2C1K w K h -1)×C2WH;

[0074] Among them, FLOPs (SdCl) is the number of floating-point operations of analog circuits; C1K w K h -1 is the number of addition operations of a single convolutional layer; W is the standard convolutional network width; H is the standard convolutional network height.

[0075] The present invention reduces the number of parameters and floating-point operations without compromising the accuracy of the lightweight fault data model. It proposes a residual network based on adaptive batch normalization to balance the number of parameters and floating-point operations and calculate the model's floating-point operations. Table 1 shows the floating-point operations of the residual network model based on adaptive batch normalization. The number of multiplication operations is 215.17k, accounting for 48.69% of the floating-point operations in the entire computation graph; the number of addition operations is 212.67k, accounting for 48.13% of the floating-point operations in the entire computation graph. The forward reasoning process (accumulated multiplication and accumulation calculations) accounts for 96.82% of the floating-point operations. The total model computational complexity is 441,879, the total number of model parameters is 215,377, and the capacity is 841.32KB. This leaves much room for optimization in model computation and storage consumption.

[0076] Table 1 Number of floating-point operations of the residual network model based on adaptive batch normalization

[0077]

[0078] Step S2: Use a deep separable convolutional network to complete the construction of a lightweight model for real-time fault data of analog circuits.

[0079] Step S21: Based on the grouped convolution operation, the depthwise separable convolutional network sets a filter for each channel of the input feature to perform a sliding operation, reducing the number of parameters and regularizing.

[0080] like Figure 2 The standard convolution and depth-separable convolution structure of the present invention are shown; the standard convolution network filter K with a size of W×W×M and a channel number of 1 is applied to a network with a size of D f ×D f The input feature vector F of ×N is used to obtain the output simulation circuit fault data feature vector through the deep separable convolutional network. Specifically:

[0081] The standard convolutional network filter K is used for the channel of the input feature vector F, and n groups of input feature vectors F are slid together with m standard convolutional network filters K to obtain m groups of results and splice them to obtain the output analog circuit fault data feature vector. for:

[0082]

[0083] Among them, K i,j,m is the standard convolutional network filter; F k+i-1,l+j-1,m is the input feature vector; i is the first position change of the convolution kernel; j is the second position change of the convolution kernel; k is the width of the analog circuit feature vector.

[0084] Table 2 provides a comparison of the standard k×k convolution (Convolution 1d, Conv1d), k×k depthwise convolution (Dwise) and 1×1 point convolution network (Pointwise Convolution, Pwise). It is obvious that the standard convolution has more parameters than the depthwise convolution and point convolution networks.

[0085] Table 2 Comparison of standard 3x3 convolution, 3x3 depthwise convolution (Dwise) and 1x1 point convolution network (Pwise)

[0086]

[0087] This paper uses model parameter count (Params) as a key evaluation metric for model lightweighting. Parameter count is directly related to model storage space: larger models require more storage space, which translates to higher computing power requirements for the device. As shown in Table 2, compared to traditional convolutional layers (k×k Conv1d), effective depthwise separable convolution (k×k Dwise + 1×1 Pwise) can significantly reduce parameter count.

[0088] Step S22: After filtering the input channel through the depthwise separable convolutional network, the features are combined using the point convolutional network to generate the first analog circuit real-time fault data features; the point convolutional network is a standard convolutional network with convolution kernel The size is [1,1], which is equivalent to weighted combination of the feature map of the analog circuit fault data in the depth direction, which can reduce the dimension of the feature of the analog circuit fault data and increase the nonlinearity.

[0089] like Figure 3 The figure shows the design diagram of the fault data lightweight model of the present invention; the output vector of the deep separable convolutional network model is determined to be:

[0090]

[0091] Among them, G k,l,n The first analog circuit feature vector output by the depthwise separable convolutional network model; is the standard convolutional network filter; l is the length of the analog circuit feature vector; is the output analog circuit fault data feature vector; m is the number of convolution kernels; n is the output value of the number of convolution kernels.

[0092] Table 3 shows a comparison of the diagnostic results of the non-lightweight model Original and the fully lightweight model DS_full in this embodiment of the present invention. The model results show that the number of parameters is reduced to 20.51% of the original model, which is a relatively ideal result. However, the model accuracy is significantly reduced, indicating that the reference Mobile Net cannot achieve a balance between accuracy and lightweightness.

[0093] Table 3 Lightweight diagnosis results

[0094]

[0095] Step S23: Design the overall network structure based on the use of a depth-wise separable convolutional network to reduce the parameter amount index and floating-point operation number index in step S1, such as Figure 5 The figure shows the lightweight overall network architecture design diagram of the present invention; improving the classification effect of the lightweight model of fault data.

[0096] Step S231: Design a first lightweight model DS1 and a second lightweight model DS2; first, set the convolution kernel to a 3×3 depthwise separable convolutional network, apply a single convolution filter to each input channel to perform lightweight filtering, and then follow up with a point convolutional network with a 1×1 convolution kernel. By calculating the linear combination of the input channels, the real-time fault data features of the second analog circuit are constructed, so that the real-time fault data maintains the dimension at the time of input.

[0097] Step S232: The main part of the fault data lightweight model is formed according to the first lightweight model DS1 and the second lightweight model DS2, such as Figure 4 The figure shows the backbone structure design diagram of the lightweight backbone part of the present invention; the backbone part is optimized by local lightweight design, and the alternating combination of standard convolutional network and depthwise separable convolutional network is used to take into account both model lightweight and model accuracy.

[0098] The backbone of the fault data lightweight model has the feature extraction capability. A 3×3 standard convolutional network is placed at the input of the depthwise separable convolutional network, which can extract fault data features at the input.

[0099] like Figure 8 The figure shows the structural design of the partial lightweight overall model of the present invention. The partial lightweight design is used to optimize the backbone. The alternating combination of standard convolutional networks and depthwise separable convolutional networks is used to balance model lightweightness and model accuracy. Two types of backbone structures, the first backbone structure DS_v1 and the second backbone structure DS_v2, are designed. Specifically,

[0100] like Figure 6 The figure shows the structure diagram of the first backbone structure DS_v1backbone of the present invention; the first backbone structure DS_v1 replaces the second lightweight model DS2 with the second deep convolutional neural network Res2Net based on the backbone of the classic convolutional network MobileNet.

[0101] like Figure 7 The second backbone structure DS_v2backbone of the present invention is shown in the figure. The second backbone structure DS_v2 is based on the backbone of the classic convolutional network MobileNet, replacing the first lightweight model DS1 with the first deep convolutional neural network Res1Net. This embodiment of the present invention trains and tests the two types of networks. The results of the lightweight and accuracy index calculations are shown in Table 4:

[0102] Table 4 Comparison of local lightweight model scale and ACC index

[0103]

[0104] The first backbone structure, DS_v1, replaces all residual network modules 1 with the first lightweight model, DS1. The second backbone structure, DS_v2, replaces all residual network modules 2 with the second lightweight model, DS2. As shown in Table 4, the lightweight design reduces the size of the v1 model to 541.32KB, 64% of the initial model, and the v2 model to 478.32KB, 57% of the initial model. In terms of computational complexity, the number of floating-point operations (FPOs) for the v1 model decreased by 34.7% year-over-year, while that for the v2 model decreased by 43.5%.

[0105] Table 5. Indexes related to local lightweight accuracy

[0106]

[0107] As shown in Table 5, in terms of local lightweight accuracy-related indicators, the improved two types of networks can both achieve an accuracy rate of more than 86%. Among them, the accuracy of the first backbone partial structure DS_v1 model remains unchanged, while the accuracy of the second backbone partial structure DS_v2 loses 3%. It is believed that the accuracy loss sacrificed by the fault data lightweight model is within a reasonable range.

[0108] like Figure 9 The figure shows the comprehensive performance evaluation diagram of the model lightweight and accuracy of the present invention; the above three architectures are used to classify analog circuit fault signals, and the overall effect of the model is displayed by combining lightweight and accuracy indicators. The horizontal axis is the number of floating-point operations of the model, and the vertical axis is the model accuracy. The size of the circle is proportional to the number of model parameters; the first main part structure DS_v1 model ensures accuracy by means of a lower degree of lightweight, and the second main part structure DS_v2 improves the degree of model lightweight by sacrificing a smaller degree of accuracy. The two methods have their own characteristics.

[0109] The fault diagnosis results are provided in the form of confusion matrix, such as Figure 10 The confusion matrix structure diagram of the first main structure DS_v1 of the present invention is shown as follows; Figure 11 The figure shows the confusion matrix structure diagram of the second main structure DS_v2 of the present invention; the horizontal axis in the confusion matrix represents the true label x of the 17 types of fault types provided real , the vertical axis represents the predicted label x of 17 types of faults predict , when x real =x predict, indicating that the predicted labels are consistent with the true labels, and the values ​​on the matrix diagonal increase. The precision, recall, and F1 score were calculated based on the confusion matrix, as shown in Table 6. The average precision, recall, and F1 score for the 17 types of fault data for the DS_v1 model of the first backbone structure are 90%, 89%, and 89%, respectively.

[0110] Table 6 Diagnosis results of the first backbone structure DS_v1

[0111]

[0112] As shown in Table 7, the average accuracy, recall rate, and F1 score of the fault data classification model of the second main structure DS_v2 are 89%, 87%, and 87%, respectively. The accuracy performance of the first main structure DS_v1 is better than that of the second main structure DS_v2. Both models can demonstrate excellent diagnostic effects on the multi-index measurement of most fault types.

[0113] Table 7 Diagnosis results of the second trunk structure DS_V2

[0114]

[0115]

[0116] In summary, by locally lightweighting the model, we can reduce the reasoning and deployment capabilities in the real-time monitoring process and help maintain the model accuracy.

[0117] Step S3: Acquire real-time fault data of the simulated circuit through experiments and expand the sample size through data enhancement algorithm, and use the lightweight model of fault data to complete the optimization.

[0118] This embodiment of the present invention uses a Sallen-key circuit designed on an EDA platform as the device side, and a computer host as the host computer side. By collecting actual circuit output responses and uploading signals to the host computer in real time, a lightweight and improved fault diagnosis model and a CEEMDAN model based on soft screening are deployed on the host computer to complete feature extraction and fault diagnosis. Based on the RC array design, this embodiment can provide the same 17 fault types, as shown in Table 8. Within the upper and lower limits, the parameter sequence following a normal distribution is no longer collected through Monte Carlo simulation, but rather uses unique component fault parameter values.

[0119] Table 8 Actual simulation circuit fault types

[0120]

[0121]

[0122] like Figure 12The figure shows the original waveform of the example signal of the present invention, which is the real-time input fault data collected by the embodiment of the present invention. The Tsaug method, a time series data enhancement algorithm, is used to expand the sample size of the collected time series fault data. The Tsaug method can achieve cropping, amplification, timeline reversal, noise addition, time distortion, side shifting and superimposition of trends, such as Figure 13 The following is an example of signal enhancement waveform diagram of the present invention, which uses the Tsaug method to enhance the fault data. Figure 14 The waveform diagram of the measured signal enhancement of the present invention is shown in FIG. Figure 13 This method demonstrates the accuracy of the fault data augmentation process, demonstrating that the resulting fault data meets operational requirements. Fault data augmentation was performed on the actual waveforms collected, expanding the number of fault waveforms for each fault type to 300. During augmentation, the signal amplitude was randomly shifted by 10%-15% with a 10% probability, and randomly reversed with a 20% probability. Ten augmented waveforms are shown, taking the first 100 sampling points of each fault type as an example.

[0123] The obtained real-time collected fault data is imported into the fault data lightweight model in step S2 for processing and lightweight optimization.

[0124] As shown in Table 9, both the ResNet_AdaBN_DS_v1 and ResNet_AdaBN_DS_v2 fault data lightweight models performed well across all three lightweight evaluation criteria. The number of parameters and floating-point operations represent model scale. The ResNet_AdaBN_DS_v1 model saw a 35.7% year-on-year decrease in parameters and a 34.7% year-on-year decrease in floating-point operations. The ResNet_AdaBN_DS_v2 model saw a 43.1% year-on-year decrease in parameters and a 43.5% year-on-year decrease in floating-point operations. The training time ratio represents the model's offline computing speed, while the testing time ratio represents the model's online computing speed. A smaller ratio indicates a faster speed. The DS_v1 model, which has the first backbone structure, already exhibits good computing speed, but the DS_v2 model, which has the second backbone structure, outperforms it, achieving both twice the offline and online computing speeds.

[0125] Table 9 Lightweight evaluation results

[0126]

[0127] While measuring lightweightness, it is also necessary to ensure that the model achieves accurate classification results. The learning process is demonstrated by calculating the training accuracy and validation accuracy, as well as the training loss and validation loss. As shown in Figures 15(a) and 15(b), the ResNet_AdaBN_DS_v1 training process diagrams of the present invention are shown. Figure 15(a) shows the training accuracy and validation accuracy of the ResNet_AdaBN_DS_v1 training process; Figure 15(b) shows the training loss and validation loss of the ResNet_AdaBN_DS_v1 training process. As shown in Figures 16(a) and 16(b), the ResNet_AdaBN_DS_v2 training process diagram of the present invention is shown; Figure 16(a) shows the training accuracy (Training Accuracy) and validation accuracy (Validation Accuracy) of the ResNet_AdaBN_DS_v2 training process; Figure 16(b) shows the training loss (Training Loss) and validation loss (Validation Loss) of the ResNet_AdaBN_DS_v2 training process.

[0128] The accuracy of the training set continues to rise, and the loss value continues to decline, proving that underfitting has not occurred. The validation set also shows the same trend as the training set with smaller oscillations, proving that overfitting has not occurred. Finally, the accuracy on the validation set can reach more than 90%, proving that a relatively ideal classification result can be achieved after 100 rounds of training. Figure 17 The figure shows the dimensionality reduction visualization result of ResNet_AdaBN_DS_v1 of the present invention. The classification results of the test set after training are visualized by t-SNE dimensionality reduction. Figure 18 The figure shows the dimensionality reduction visualization result of ResNet_AdaBN_DS_v2 of the present invention. The classification results of the test set after training are visualized by t-SNE dimensionality reduction. Through dimensionality reduction visualization, the fault data of the test set are each given a prediction label after model prediction. The clustered fault data of different colors correspond to 17 types of fault types. The cluster sets are shown as follows: Figure 17 and Figure 18As shown in the figure, there is no aggregation of fault data between different categories, which proves that the classification effect is good from the visual and intuitive level. To further prove it, it is necessary to analyze the classification indicators. The classification effect is measured by five classification problem indicators: loss, precision, recall, F1 score, and accuracy. Figure 19 The figure shows the classification index measurement results of the present invention. The test results confirm that after lightweight improvements are carried out on the ResNet_AdaBN base model and the introduction of the depthwise separable convolutional model, not only the number of parameters, the number of floating-point operations, and the training speed ratio are greatly reduced, but the overall accuracy remains unchanged. The number of parameters of ResNet_AdaBN_DS_v1 decreased by 35.7%, while the accuracy rate decreased by 2%; the number of parameters of ResNet_AdaBN_DS_v2 decreased by 43.1% year-on-year, while the accuracy rate decreased by 4%. Among them, the accuracy of the ResNet_AdaBN_DS_v1 model is as high as 0.92, which means that the algorithm's judgment on fault classification is accurate with a level of 92%, and the error rate is extremely low.

[0129] Compared with the classic lightweight model MobileNetV1 and the small-scale model MobileNetV1_small, under the same experimental environment conditions, the lightweight and accuracy evaluation results are shown in Table 10:

[0130] Table 10 Algorithm performance parameter comparison

[0131]

[0132] Select the number of parameters, number of floating-point operations, and accuracy to plot the comprehensive performance of lightweight and accuracy. Figure 20 The figure shows the comprehensive performance evaluation diagram of the model lightweight and accuracy of the present invention. In the figure, the horizontal axis is the number of model floating-point operations, the vertical axis is the model accuracy, and the size of the circle is proportional to the number of model parameters. Compared with the original model, the first main part structure DS_v1 and the second main part structure DS_v2 only reduce the accuracy by 2% and 4%, respectively, and obtain 34.7% and 43.5% storage performance optimization; MobileNetV1 has better accuracy performance, sacrificing only 5% accuracy, but the computing consumption is 115.8% of the original model, and the storage consumption is 13.98 times that of the original model. MobileNetV1-small has the best computing and storage optimization performance, but the accuracy is reduced by 14%. In summary, through comparative test demonstration, it can be effectively proved that the fault data lightweight model proposed in this application can achieve a better balance in the two directions of accuracy and lightweight.

[0133] On the other hand, the present invention proposes a model construction system for a real-time fault data lightweight method for analog circuits based on deep separable convolution, including: a deep separable convolutional network module, a point convolutional network module, a residual network module and a backbone module of a fault data lightweight model; the deep separable convolutional network module uses two types of modules, deep convolution and point convolution, instead of standard convolution to filter and combine the input fault data features respectively; the point convolutional network module combines the input fault data features to generate new fault data features; the residual network module is lightweight-improved based on the residual network algorithm of AdaBN, and can ensure accuracy while reducing the number of parameters and floating-point operations of the fault data lightweight model; the backbone module of the fault data lightweight model includes a first lightweight model DS1 and a second lightweight model DS2; the overall network structure is designed on the basis of reducing the number of parameters and floating-point operations of the fault data lightweight model to ensure the model classification effect and to fully extract the fault data features near the input end.

[0134] The beneficial effects of the present invention are as follows: the present invention provides a lightweight method for real-time fault data of analog circuits based on deep separable convolution, which meets the low latency and low capacity requirements of real-time monitoring scenario diagnosis. First, based on the computational and storage consumption issues of the algorithm model, two evaluation criteria for measuring the degree of model lightweighting are proposed. Next, based on the existing model, design improvements are carried out. Based on the idea of ​​deep separable convolution, the model is lightweighted globally and locally. Multiple algorithm frameworks are designed and comparative experiments are carried out to determine the optimal model while ensuring a balance between accuracy, complexity, and parameter quantity. Real-time fault diagnosis hardware deployment verification is carried out: to collect fault data at the physical level, a resistor-capacitor array is designed for parameter drift fault injection, and the schematic design, PCB wiring, and plate making work are completed for the entire circuit. To address the problem of the small number of signals collected for each fault type, the Tsaug data enhancement algorithm is used to expand the waveform. Fault diagnosis tasks are carried out on the enhanced fault data set. Finally, the effectiveness of the lightweight method of the present invention is verified by comparing performance parameters, proving that the method meets actual needs in actual use.

[0135] The embodiments described above are merely descriptions of preferred implementations of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should fall within the scope of protection determined by the claims of the present invention.

Claims

1. A lightweight method for real-time fault data of analog circuits based on depthwise separable convolution, characterized in that: It includes: S1: Determine the evaluation index of the lightweight model of real-time fault data of analog circuits and calculate the evaluation index based on the standard convolutional network; The number of parameters and floating-point operations are used to evaluate the lightweight model of real-time acquisition of fault data of analog circuits. The three dimensions of the standard convolutional network feature map are set as: height, width and number of channels, which are combined as [H, W, C1]. The convolution kernel of the standard convolutional network is set as [K w ,K h ,C1], the number of filters is C2, the dimension of the feature map after the standard convolution network is [H, W, C2], and the analog circuit parameter quantity index and the analog circuit floating-point operation number evaluation index are calculated according to the standard convolution network; S2: Use a deep separable convolutional network to build a lightweight model for real-time fault data of analog circuits; S21: Based on the grouped convolution operation, the depth-wise separable convolutional network sets a filter for each channel of the input feature to perform sliding operation, reducing the number of parameters and regularizing; the standard convolutional network filter K with a size of W×W×M and a channel number of 1 is applied to the D f ×D f The input feature vector F of ×N is used to obtain the output simulation circuit fault data feature vector through the deep separable convolutional network. S22: After filtering the input channel through the depthwise separable convolutional network, the features are combined using a pointwise convolutional network to generate a first analog circuit real-time fault data feature; the output vector of the depthwise separable convolutional network model is determined to be: Among them, G k,l,n The first analog circuit feature vector output by the depthwise separable convolutional network model; is the standard convolutional network filter; k is the width of the analog circuit feature vector; l is the length of the analog circuit feature vector; is the output analog circuit fault data feature vector; m is the number of convolution kernels; n is the output value of the number of convolution kernels; S23: Design the overall network structure based on the reduction of the parameter quantity index and floating-point operation number index in step S1 using a depthwise separable convolutional network to improve the classification effect of the lightweight model of fault data; S231: Design a first lightweight model DS1 and a second lightweight model DS2. First, set a depthwise separable convolutional network with a convolution kernel of 3×3, apply a single convolution filter to each input channel to perform lightweight filtering, and then use a point convolutional network with a convolution kernel of 1×1 to construct the real-time fault data features of the second analog circuit by calculating the linear combination of the input channels, so that the real-time fault data maintains the dimensionality of the input. S232: Based on the first lightweight model DS1 and the second lightweight model DS2, the backbone of the fault data lightweight model is formed. The backbone is optimized by adopting a local lightweight design. The standard convolutional network and the depthwise separable convolutional network are alternately combined to achieve a balance between model lightweight and model accuracy. The first backbone structure DS_v1 and the second backbone structure DS_v2 of the two types of backbones are designed. S3: Acquire real-time fault data of simulated circuits through experiments, expand the sample size through data enhancement algorithms, and use a lightweight model of fault data to complete optimization; The sample size of the collected real-time fault data of the analog circuit is expanded by using a time series data enhancement algorithm; the obtained real-time fault data of the analog circuit is imported into the fault data lightweight model in step S2 for processing, thereby completing the lightweight optimization of the real-time fault data of the analog circuit.

2. The method for lightweight quantization of real-time fault data of analog circuits based on depthwise separable convolution according to claim 1 is characterized in that: The analog circuit parameter indicators in step S1 are: Params(SdCl)=C1K w K h C2; Among them, Params (SdCl) is the parameter index of the analog circuit; C1 is the number of channels of the standard convolutional network; K w is the standard convolutional network filter width; K h is the height of the standard convolutional network filter; C2 is the number of filters; C1K w K h is the number of multiplication operations of a single convolutional layer.

3. The method for lightweight quantization of real-time fault data of analog circuits based on depthwise separable convolution according to claim 1 is characterized in that: The floating-point operation times index of the analog circuit in step S1 is specifically: The analog circuit floating-point operation index is used to measure the complexity of the algorithm. The computational load in the convolutional network is mainly borne by forward reasoning. The forward reasoning process is a multiplication-accumulation calculation at the mathematical level. Therefore, the floating-point operation number is mainly composed of the number of multiplication-addition operations. The floating-point operation number index is: FLOPs(SdCl)=[C1K w K h +(C1K w K h -1)]=(2C1K w K h -1)×C2WH; Among them, FLOPs (SdCl) is the number of floating-point operations of analog circuits; G1K w K h -1 is the number of addition operations of a single convolutional layer; W is the standard convolutional network width; H is the standard convolutional network height.

4. The method for lightweight quantization of real-time fault data of analog circuits based on depthwise separable convolution according to claim 1 is characterized in that: Step S21 obtains the output analog circuit fault data feature vector through the deep separable convolutional network Specifically: The standard convolutional network filter K is used for the channel of the input feature vector F, and m groups of input feature vectors F are slid and cut with m standard convolutional network filters K respectively, and m groups of results are obtained by splicing to obtain the output analog circuit fault data feature vector for: Among them, K i,j,m is the standard convolutional network filter; F k+i-1,l+j-1,m is the input feature vector; i is the first position change of the convolution kernel; j is the second position change of the convolution kernel.

5. The method for lightweight quantization of real-time fault data of analog circuits based on depthwise separable convolution according to claim 1 is characterized in that: The point convolution network in step S22 is a standard convolution network with convolution kernel The size is [1,1], which is equivalent to weighted combination of the feature map of the analog circuit fault data in the depth direction, which can reduce the dimension of the feature of the analog circuit fault data and increase the nonlinearity.

6. The method for lightweight quantization of real-time fault data of analog circuits based on depthwise separable convolution according to claim 1 is characterized in that: The backbone of the fault data lightweight model in step S232 has the feature extraction capability. A 3×3 standard convolutional network is placed at the input end of the depthwise separable convolutional network, which can extract fault data features at the input end.

7. The method for lightweight quantization of real-time fault data of analog circuits based on depthwise separable convolution according to claim 1 is characterized in that: The first trunk part structure DS_v1 and the second trunk part structure DS_v2 of the two types of trunk parts in step S232 are specifically: The first backbone structure DS_v1 replaces the second lightweight model DS2 with the second deep convolutional neural network Res2Net based on the backbone of the classic convolutional network MobileNet; The second backbone structure DS_v2 replaces the first lightweight model DS1 with the first deep convolutional neural network Res1Net based on the backbone of the classic convolutional network Mobi leNet.

8. The method for lightweight quantization of real-time fault data of analog circuits based on depthwise separable convolution according to claim 1, characterized in that: The time series data enhancement algorithm in step S3 has a unified API interface, which can control the probability of occurrence of each function implementation so that the randomly enhanced fault data has the required distribution.

9. A model building system based on the lightweight method for real-time fault data of analog circuits based on depthwise separable convolution according to any one of claims 1 to 8, characterized in that: It includes a depthwise separable convolutional network module, a point convolutional network module, a residual network module, and a backbone module for a fault data lightweight model; The depthwise separable convolutional network module uses two types of modules, depthwise convolution and pointwise convolution, instead of standard convolution to filter and combine the input fault data features respectively; The point convolutional network module combines the input fault data features to generate new fault data features; The residual network module is lightweight and improved based on the residual network algorithm of AdaBN, ensuring accuracy while reducing the number of parameters and floating-point operations of the lightweight model of fault data; The backbone module of the fault data lightweight model includes a first lightweight model DS1 and a second lightweight model DS2; the overall network structure is designed on the basis of reducing the number of parameters and floating-point operations of the fault data lightweight model, and the fault data features are extracted near the input end.

Images