Analog circuit continuous fault diagnosis method for strong noise scene
By constructing a multi-branch multi-scale convolutional network (MBSCNN) combined with global average pooling analog circuit continuous fault diagnosis model, the fault diagnosis problem under strong noise and continuous change fault conditions is solved, and high-precision fault recognition performance is achieved.
Patent Information
- Application Number
- CN202510209348.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-25
- Publication Date
- 2025-06-13
AI Technical Summary
Under the conditions of strong noise and continuous change faults, it is difficult to accurately identify the fault diagnosis of analog circuits, and it is difficult for traditional feature extraction methods to extract the inherent fault characteristics.
A multi-branch multi-scale convolutional network (MBSCNN) combined with global average pooling is used to construct a continuous fault diagnosis model for simulated circuits. The sample set is obtained through PSPICE software simulation, including noise-free and noisy samples, and the initial simulation circuit continuous fault diagnosis model is trained.
In the highly noise scenario, MBSCNN can automatically learn and fuse rich and complementary features, improving the accuracy and accuracy of analog circuit fault diagnosis, which is significantly better than other typical methods.
Smart Images
Figure CN120142898A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of fault detection, and particularly to a continuous fault diagnosis method for analog circuits facing strong noise scenarios. Background Art
[0002] Electronic devices play a crucial role in fields such as aerospace, transportation, medical treatment, and communication. However, although there are only less than 20% analog circuits in electronic devices, analog circuit faults account for 80% of circuit faults. Analog circuit faults are mainly divided into hard faults and soft faults. Hard faults indicate that the circuit is completely ineffective due to component short circuits or open circuits, and thus are easy to diagnose. Soft faults indicate a decline in circuit performance caused by component parameter values deviating from the tolerance range. Due to the non-linearity of the circuit and the tolerance of components, the fault states of analog circuits overlap with each other, making it difficult to accurately identify soft faults.
[0003] For an actual analog circuit, on the one hand, it will be affected by external environmental interference, and on the other hand, the noise interference caused by the tolerance of its own components or component parameter faults will make the output of the system deviate from the pre-expected stable state, corresponding to a new state. Therefore, it is very important and challenging to have high-precision fault diagnosis ability in background noise. Secondly, the workload of analog circuits changes from time to time, and the relevant signal characteristics also change accordingly. The degree of deviation of parameter values is not fixed and single, and will actually continuously change within the range of 10% to 50%. Since it is unrealistic to collect and label sufficient training data, traditional feature extraction is difficult to truly extract the inherent fault features, because the fault features are completely submerged in strong noise and are distributed in different feature intervals under different fault degree conditions. Summary of the Invention
[0004] In order to overcome the deficiencies of the prior art, the purpose of the present invention is to provide a continuous fault diagnosis method for analog circuits facing strong noise scenarios, and the present invention solves the problem that there are deficiencies in the prior art's fault diagnosis methods for strong noise and continuously changing fault conditions.
[0005] To achieve the above purpose, the present invention provides the following solution:
[0006] A continuous fault diagnosis method for analog circuits facing strong noise scenarios, comprising:
[0007] Obtain data of the analog circuit to be measured;
[0008] Input the data of the analog circuit to be measured into the continuous fault diagnosis model of the analog circuit to obtain a diagnosis result;
[0009] The construction method of the continuous fault diagnosis model of the analog circuit is:
[0010] Use PSPICE software to simulate the experimental circuit to obtain a sample set, which includes: noise-free original samples and noisy samples;
[0011] Input the sample set into the initial multi-branch multi-scale convolutional network to obtain the final fused features;
[0012] Replace the fully connected layer in the initial multi-branch multi-scale convolutional network with global average pooling, and connect the classification layer to obtain the final multi-branch multi-scale convolutional network;
[0013] Obtain an initial analog circuit continuous fault diagnosis model according to the final multi-branch multi-scale convolutional network and the final fused features;
[0014] Use the noisy samples to train the initial analog circuit continuous fault diagnosis model to obtain a constructed analog circuit continuous fault diagnosis model.
[0015] Preferably, the use of PSPICE software to simulate the experimental circuit to obtain a sample set includes:
[0016] Use PSPICE software to determine the experimental circuit and select the circuit output terminal as the test point to obtain the circuit sine superposition response signal;
[0017] Perform Monte-Carlo analysis on the circuit sine superposition response signal and set parameter value deviations for each faulty component to obtain a simulation data set, where the degree of parameter value deviation is 10%-50%;
[0018] Add noise to the simulation data set to obtain a sample set.
[0019] Preferably, the inputting the sample set into the initial multi-branch multi-scale convolutional network to obtain the final fused features includes:
[0020] Input the sample set into the equal mapping branch and the noise reduction branch of the initial multi-branch multi-scale convolutional network respectively to obtain a first feature set and a second feature set;
[0021] Use the feature fusion layer of the initial multi-branch multi-scale convolutional network to fuse the first feature set and the second feature set to obtain the final fused features.
[0022] Preferably, the working process of the equal mapping branch is:
[0023] Use multiple parallel convolutional layers to simultaneously learn the multi-scale features of the sample set to obtain the outputs of each convolutional layer;
[0024] Concatenate the outputs of each convolutional layer to obtain a first feature set.
[0025] Preferably, the expressions of the outputs of the respective convolutional layers are as follows:
[0026]
[0027] where i is the layer index, ω i and b i are the convolution kernel and bias of the i-th convolutional layer respectively, and β(·) represents the functional transformation of the activation function (ReLU), batch normalization (BN), and dropout.
[0028] Preferably, the working process of the noise reduction branch is as follows:
[0029] Introduce a Gaussian filter;
[0030] Combine the Gaussian filter with the CNN and obtain the signal components with high SNR from the sample set;
[0031] Replace the 5 standard convolutional layers in the noise reduction branch with 5 MSC modules, and obtain the second feature set according to the signal components with high SNR.
[0032] Preferably, the expression of the first feature set is as follows:
[0033] y 1 = IMB(X) = MSC(X) = C s (O);
[0034] where X is the sample set, and O is the concatenated vector of the outputs of the respective convolutional layers.
[0035] Preferably, the expression of the second feature sample set is as follows:
[0036] y 2 = DB(X) = MSC ( MSC ( MSC ( MSC(Xg))));
[0037] where Xg is the sample processed by the Gaussian filter.
[0038] Preferably, the calculation expression of the diagnosis result is as follows:
[0039]
[0040] where z i is the original vector value of the i-th classification in the classification layer.
[0041] The present invention discloses the following technical effects:
[0042] The present invention provides a method for continuous fault diagnosis of analog circuits for strong noise scenarios, including: obtaining data of the analog circuit to be measured; inputting the data of the analog circuit to be measured into the continuous fault diagnosis model of the analog circuit to obtain a diagnosis result; the construction method of the continuous fault diagnosis model of the analog circuit is: using PSPICE software to simulate the experimental circuit to obtain a sample set, the sample set includes: noise-free original samples and noisy samples; inputting the sample set into the initial multi-branch multi-scale convolutional network to obtain the final fused features; using global average pooling to replace the fully connected layer in the initial multi-branch multi-scale convolutional network, connecting the classification layer to obtain the final multi-branch multi-scale convolutional network; obtaining the initial continuous fault diagnosis model of the analog circuit according to the final multi-branch multi-scale convolutional network and the final fused features; using the noisy samples to train the initial continuous fault diagnosis model of the analog circuit to obtain the constructed continuous fault diagnosis model of the analog circuit. The present invention can intelligently diagnose component faults from analog circuit signals under strong noise and continuously changing fault conditions. MBSCNN combines multi-branch and multi-scale learning, can automatically learn and fuse rich and complementary features from multiple signal components and time scales, efficiently and accurately screen out the devices where faults occur, improve the detection accuracy. Experiments show that compared with other typical methods, the method proposed in this paper shows better fault recognition performance. Description of the Drawings
[0043] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.
[0044] Figure 1 It is a flowchart of a method for continuous fault diagnosis of analog circuits for strong noise scenarios provided by an embodiment of the present invention;
[0045] Figure 2 It is a schematic diagram of the Leap Frog filter circuit provided by an embodiment of the present invention;
[0046] Figure 3 It is a schematic diagram of the multi-branch multi-scale convolutional network (MBSCNN) model provided by an embodiment of the present invention;
[0047] Figure 4 It is a schematic diagram of multi-scale learning provided by an embodiment of the present invention;
[0048] Figure 5 It is a schematic diagram of deep multi-scale learning provided by an embodiment of the present invention. Detailed Embodiments
[0049] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.
[0050] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0051] As Figure 1 shown, the present invention provides a method for continuous fault diagnosis of analog circuits for strong noise scenarios, including:
[0052] Step 100: Obtain the data of the analog circuit to be measured;
[0053] Step 200: Input the data of the analog circuit to be measured into the continuous fault diagnosis model of the analog circuit to obtain a diagnosis result;
[0054] The construction method of the continuous fault diagnosis model of the analog circuit is as follows:
[0055] Step 201: Use PSPICE software to simulate the experimental circuit to obtain a sample set, and the sample set includes: noise-free original samples and noisy samples;
[0056] Step 202: Input the sample set into the initial multi-branch multi-scale convolutional network to obtain the final fused features;
[0057] Step 203: Replace the fully connected layer in the initial multi-branch multi-scale convolutional network with global average pooling, and connect the classification layer to obtain the final multi-branch multi-scale convolutional network;
[0058] Step 204: Obtain the initial continuous fault diagnosis model of the analog circuit according to the final multi-branch multi-scale convolutional network and the final fused features;
[0059] Step 205: Use the noisy samples to train the initial continuous fault diagnosis model of the analog circuit to obtain the constructed continuous fault diagnosis model of the analog circuit.
[0060] Specifically, use PSPICE software to simulate the experimental circuit, and the simulation circuit diagram is as Figure 2 shown, where V in represents the power supply excitation of the circuit, V out represents the output signal of the circuit and is also the signal acquisition node, and two V a represent that the two nodes are connected in the circuit (Vb Similarly, taking the resistor parameter (R1, 10k) as an example, the resistance value of resistor R1 is 10 kiloohms, and taking the capacitor parameter (C1, 10nF) as an example, the capacitance value of capacitor C1 is 10 nanofarads. Preprocess the simulation data to construct a noise-free original sample (Ⅰ) and a noisy sample (Ⅱ); for multi-branch learning, divide the original signal sample data into two branch channels, including an identity mapping branch (IMB) and a denoising branch (DB), and extract features of the samples through the two channels; for the feature extraction methods of the two divided branches, the identity mapping branch adopts multi-scale learning, and the denoising branch adopts deep multi-scale learning; use a feature fusion layer to fuse multiple features, including the fusion of multi-scale features of the identity mapping branch and the fusion of multi-signal features learned by the two branches; adopt global average pooling to replace the common fully connected layer in CNN, connect the classification layer, and use the Softmax function to give the result; use the sample data to train the fault diagnosis model to obtain the best diagnosis model for continuous faults of analog circuits in a strong noise scenario, so as to obtain the diagnosis result of continuous faults of analog circuits in a strong noise scenario.
[0061] Further, the obtaining of the sample set by simulating the experimental circuit using PSPICE software includes:
[0062] Use PSPICE software to determine the experimental circuit and select the circuit output terminal as the test point to obtain the circuit sine superposition response signal;
[0063] Perform Monte-Carlo analysis based on the circuit sine superposition response signal and set parameter value deviations for each faulty component to obtain a simulation data set, where the degree of the parameter value deviation is 10% - 50%;
[0064] Add noise to the simulation data set to obtain the sample set.
[0065] Specifically, construct a Leap Frog filter circuit in PSPICE software, as shown in the circuit diagram Figure 2 shown, select the circuit output terminal as the test point to obtain the circuit sine superposition response signal, and through Monte-Carlo analysis, set the parameter value deviation degree of each faulty component to continuously change from 10% to 50%; the preprocessing is to add a certain magnitude of noise to the original simulation data to simulate real noise interference, thereby generating multiple groups of noise-free sample sets (Ⅰ) and noisy sample sets (Ⅱ) for each fault category;
[0066] The noise magnitude is the signal-to-noise ratio (SNR dB ), which refers to the ratio of the signal power to the noise power in the sample, and the unit is dB. Its calculation formula is shown as follows:
[0067]
[0068] Among them, P signal refers to the signal power, and P noise refers to the noise power.
[0069] Furthermore, inputting the sample set into the initial multi-branch multi-scale convolutional network to obtain the final fused feature includes:
[0070] Inputting the sample set into the identity mapping branch and the noise reduction branch of the initial multi-branch multi-scale convolutional network respectively to obtain a first feature set and a second feature set;
[0071] Using the feature fusion layer of the initial multi-branch multi-scale convolutional network to fuse the first feature set and the second feature set to obtain the final fused feature.
[0072] In this solution, the two branch channels are as Figure 3 shown. For the feature extraction methods of the two divided branches, the identity mapping branch adopts multi-scale learning, and the noise reduction branch adopts deep multi-scale learning. Finally, n classifications (C1, C2, ···, Cn) are output.
[0073] Even further, the working process of the identity mapping branch is as follows:
[0074] Using multiple parallel convolutional layers to simultaneously learn the multi-scale features of the sample set to obtain the outputs of each convolutional layer;
[0075] Concatenating the outputs of each convolutional layer to obtain a first feature set.
[0076] Specifically, the identity mapping branch (IMB) refers to learning global fault feature information from the original signal. IMB is a wide-spectrum range analysis. Given a signal sample X = [x 1 , x 2 , ···, x N , where N is the size of X;
[0077] As Figure 4 shown, the core of the multi-scale learning (Multiscale Learning) is to learn complementary long-term and short-term features in the original signal from different time scales. First, the multi-scale convolutional module (Multiscale Convolutional Module, MSC) uses h parallel convolutional layers to simultaneously learn the multi-scale features of the input signal X. The output of each convolutional layer is expressed as:
[0078]
[0079] where i is the layer index, ω i and b iare the convolution kernel and bias of the i-th convolution layer, respectively, ω i has a size of 1×2 i- 1 k, where k is an integer, and β(·) represents the functional transformation of the activation function (ReLU), batch normalization (BN), and dropout;
[0080] Using the output features o of each convolution layer i are concatenated into a feature vector O = [o 1 , o 2 , ···, o h ;
[0081] Using O to enter the feature fusion layer C s , effectively fusing complementary features of different time scales, then the output of MSC, which is also the output of IMB, is:
[0082] y 1 = IMB(X) = MSC(X) = C s (O);
[0083] Furthermore, the working process of the noise reduction branch is as follows:
[0084] Introduce a Gaussian filter;
[0085] Combine the Gaussian filter with CNN and obtain the signal component with high SNR from the sample set;
[0086] Use 5 MSC modules to replace 5 standard convolution layers in the noise reduction branch, and obtain the second feature set according to the signal component with high SNR.
[0087] Specifically, the noise reduction branch (DB) refers to obtaining the signal component with high SNR from the original signal, thereby enhancing the learning and discrimination ability of CNN for fault features, introducing a Gaussian filter, and combining it with CNN to improve the anti-noise ability of the network;
[0088] The Gaussian filter is g (convolution kernel), and the formula for the one-dimensional discrete Gaussian function f[j] is:
[0089]
[0090] where j is an integer and δ is the standard deviation of j (δ = 1);
[0091] Then the length of the filter g is selected as 5 and can be expressed by the following formula:
[0092] g = [f[-2], f[-1], f[0], f[1], f[2]];
[0093] Process the signal sample X with the Gaussian filter g to obtain X g, can be expressed by the following formula:
[0094]
[0095] Where represents the convolution operation;
[0096] As Figure 5 shown, in the Deep Multiscale Learning (DMSN), 5 standard convolutional layers in the DB are replaced by 5 MSC modules, and the output of the DB is:
[0097] y 2 = DB(X) = MSC ( MSC ( MSC(MSC ( MSC(Xg))))) ;
[0098] In this solution, a feature fusion layer is used to fuse multiple features, including the identity mapping branch multi-scale feature fusion and the fusion of multi-signal features learned by two branches. Specifically:
[0099] The fusion of multi-signal features learned by the two branches is to first concatenate the multi-feature learned y 1 , y 2 into a feature vector y c , which can be expressed by the following formula:
[0100] y c = [y 1 , y 2 ;
[0101] Using y c and the fusion layer C b of multi-signal features, the final feature y learned from X can be expressed by the following formula:
[0102] y = C b (y c ) = C b [y 1 , y 2 ;
[0103] In this solution, global average pooling is used to replace the common fully connected layer in the CNN, connect the classification layer, and use the Softmax function to give the result. Specifically:
[0104] For the analog circuit fault diagnosis task, the Softmax function is used in the classification layer to give the result. Assuming that the input sample has n categories, the output probability of category j is Q i , and the diagnostic output is the fault label Q i corresponding to the maximum, and the calculation is as follows:
[0105]
[0106] Among them, z i is the original vector value of the i-th classification in the classification layer, and
[0107] In this solution, the fault diagnosis model is trained using sample data to obtain the best diagnosis model for continuous faults in analog circuits facing strong noise scenarios, so as to obtain the diagnosis results of continuous faults in analog circuits facing strong noise scenarios. Specifically:
[0108] The fault diagnosis model is trained using noise sample data to obtain the best diagnosis model for continuous faults in analog circuits facing strong noise scenarios. The noisy sample set is input into the diagnosis model to realize the identification and classification of floating soft faults in analog circuits under potential strong noise conditions.
[0109] In this specification, each embodiment is described in a progressive manner. The key point of each embodiment is to illustrate the differences from other embodiments. The same or similar parts among the embodiments can be referred to each other.
[0110] In this article, specific examples are used to elaborate on the principles and implementation manners of the present invention. The descriptions of the above embodiments are only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation to the present invention.
Claims
1. A method for continuous fault diagnosis of analog circuits in strong noise scenarios, characterized in that: include: Obtaining the analog circuit data to be tested; Inputting the analog circuit data to be tested into an analog circuit continuous fault diagnosis model to obtain a diagnosis result; The construction method of the analog circuit continuous fault diagnosis model is: The experimental circuit is simulated by using PSPICE software to obtain a sample set, wherein the sample set includes: original samples without noise and samples with noise; Inputting the sample set into an initial multi-branch multi-scale convolutional network to obtain a final fusion feature; Using global average pooling to replace the fully connected layer in the initial multi-branch multi-scale convolutional network, connecting the classification layer, and obtaining the final multi-branch multi-scale convolutional network; Obtaining an initial analog circuit continuous fault diagnosis model according to the final multi-branch multi-scale convolutional network and the final fusion features; The initial analog circuit continuous fault diagnosis model is trained using the noisy samples to obtain a constructed analog circuit continuous fault diagnosis model.
2. The method for continuous fault diagnosis of analog circuits in strong noise scenarios according to claim 1, characterized in that: The method of using PSPICE software to simulate the experimental circuit to obtain a sample set includes: Use PSPICE software to determine the experimental circuit and select the circuit output end as the test point to obtain the circuit sinusoidal superposition response signal; Performing Monte-Carlo analysis on the circuit sinusoidal superposition response signal and setting parameter value deviation for each faulty component to obtain a simulation data set, wherein the degree of parameter value deviation is 10%-50%; Noise is added to the simulation data set to obtain a sample set.
3. The method for continuous fault diagnosis of analog circuits in strong noise scenarios according to claim 1, characterized in that: The step of inputting the sample set into an initial multi-branch multi-scale convolutional network to obtain the final fusion features includes: Inputting the sample set into the equal mapping branch and the denoising branch of the initial multi-branch multi-scale convolutional network respectively to obtain a first feature set and a second feature set; The first feature set and the second feature set are fused using the feature fusion layer of the initial multi-branch multi-scale convolutional network to obtain a final fused feature.
4. The method for continuous fault diagnosis of analog circuits in strong noise scenarios according to claim 3, characterized in that: The working process of the equal mapping branch is as follows: Use multiple parallel convolutional layers to simultaneously learn the multi-scale features of the sample set and obtain the output of each convolutional layer; The outputs of the convolutional layers are concatenated to obtain a first feature set.
5. The method for continuous fault diagnosis of analog circuits in strong noise scenarios according to claim 4, characterized in that: The output expression of each convolutional layer is: Where i is the layer index, ω i and b i are the convolution kernel and bias of the i-th convolutional layer, respectively. β(·) represents the function transformation of activation function (ReLU), batch normalization (BN), and dropout.
6. The method for continuous fault diagnosis of analog circuits in strong noise scenarios according to claim 3, characterized in that: The working process of the noise reduction branch is as follows: Introducing Gaussian filter; Combining the Gaussian filter with the CNN and obtaining a signal component with a high SNR from a sample set; Five standard convolutional layers in the noise reduction branch are replaced by five MSC modules, and a second feature set is obtained according to the signal component with high SNR.
7. The method for continuous fault diagnosis of analog circuits in strong noise scenarios according to claim 6, characterized in that: The expression of the first feature set is: y1=IMB(X)=MSC(X)=C s (O); Among them, X is the sample set, and O is the concatenated vector of the output of each convolutional layer.
8. The method for continuous fault diagnosis of analog circuits in strong noise scenarios according to claim 6, characterized in that: The expression of the second feature sample set is: y2=DB(X)=MSC ( MSC ( MSC(MSC ( MSC(Xg) ) ))); Among them, Xg is the sample processed by Gaussian filter.
9. The method for continuous fault diagnosis of analog circuits in strong noise scenarios according to claim 6, characterized in that: The calculation expression of the diagnosis result is: Among them, z i is the original vector value of category i in the classification layer.