CN114355173B - A Fault Diagnosis Method of Analog Filter Circuit Based on Multi-input Residual Network - Google Patents

Abstract

The invention discloses a fault diagnosis method of an analog filter circuit based on a multi-input residual error network, which comprises the steps of firstly setting relevant parameters of Monte Carlo analysis, and carrying out simulation analysis on the analog filter circuit by using a Monte Carlo statistical analysis method to obtain voltage signals of the output end of the analog filter circuit under different fault states; the modal components of the voltage signals are obtained through empirical wavelet transformation, the modal components are respectively formed into a training set and a testing set, the training set is used for training the multi-input ResNet, the testing set is used for testing, and a multi-input ResNet model and a parameter Parameters corresponding to the highest testing precision are stored; and finally diagnosing unknown fault states of the analog filter circuit through the trained multi-input ResNet model and Parameters.

Description

A Fault Diagnosis Method of Analog Filter Circuit Based on Multi-input Residual Network

technical field

The invention belongs to the technical field of fault diagnosis of analog circuits, and more specifically relates to a fault diagnosis method of an analog filter circuit based on a multi-input residual network.

Background technique

With the rapid development of the electronic circuit industry, the degree of integration of electronic circuits continues to increase, the requirements for functionality and modularization are becoming more and more obvious, and the requirements for the reliability of electronic circuit systems are also getting higher and higher. According to statistics, nearly 80% of the circuits are digital circuits, but nearly 80% of the faults in the circuits are generated by analog circuits. Because of its good performance, the analog filter circuit is often used as a filter module in the analog part of the electronic circuit. If the analog filter circuit fails during use, it will inevitably affect the overall performance of the analog circuit and cause unavoidable losses. Therefore, it is very necessary to diagnose the fault of the analog filter circuit in time.

In recent years, a large number of analog filter circuit fault diagnosis methods based on the combination of feature extraction and classifiers have been proposed, but these methods need to extract, select or fuse features. Compared with neural network algorithms for adaptive feature extraction, this The method is more complicated. However, the traditional neural network-based analog filter circuit fault diagnosis method only processes the original signal, and does not fully consider the relationship between the filter performance and the original signal component when the circuit is faulty, which limits the improvement of the accuracy of analog filter circuit fault diagnosis.

Contents of the invention

The purpose of the present invention is to overcome the deficiencies in the prior art, provide a kind of analog filter circuit fault diagnosis method based on multi-input residual network, by combining empirical wavelet transform with multi-input residual network (Residual network, ResNet), realize Fault diagnosis of analog filter circuit improves the accuracy of fault diagnosis.

In order to realize the above-mentioned purpose of the invention, a kind of analog filter circuit fault diagnosis method based on multi-input residual network of the present invention is characterized in that, comprises the following steps:

(1), use the Monte Carlo analysis method to carry out simulation analysis on the analog filter circuit to be tested, and obtain the voltage signal v _ij (t) of k × n groups of sampling points at the output end of the circuit under k kinds of fault states, i=1,2 ,...,k,j=1,2,...,n, n represents the group number of voltage signals collected at the output end of the analog filter circuit to be tested;

(2), use the traditional empirical wavelet transform algorithm to decompose the voltage signal v _ij (t), and obtain m modal components, wherein the qth modal component is expressed as ewt _ijq (t), q=1,2 ,...,m, and then m modal components are composed of eigenvectors ewt _ij =[ewt _ij1 ; ewt _ij2 ;...;ewt _ijq (t);...;ewt _ijm ];

验证数据集

(3), randomly distribute n ewt _ij under k kinds of fault states according to the ratio of r ₁ :r ₂ to form training data sets respectively

Validation dataset

和验证集

(4), use the serial number i of the fault state corresponding to ewt _ij as a label, load the labels for the training data set and the verification data set respectively, and construct the training set of the multi-input residual network model

and validation set

(5), build a multi-input residual network, use train_set to train the multi-input residual network model, use validation_set to verify the model trained by train_set, and obtain the optimal multi-input residual network model model _max ;

(5.1), set the training rounds of the multi-input residual network model to N, and set the input sample size to batch_size in each round of training; it takes t iterations for the train_set to be completely input into the multi-input residual network, and each iteration is multi-input Input residual network input is recorded as train_set _p , p=1...t,

(5.2), train_set _p is input to the multi-input residual network, and the predicted label of the qth sample in train_set _p is predicted by the Softmax function to be the probability pro _qi of the loaded label i, q=1...batch_size, and the train_set _p in this iteration is calculated The training loss TL _mp obtained by training the multi-input residual network, m=1...N;

Among them, y _qi is equal to 0 or 1. When y _qi is equal to 1, it means that the label loaded by the qth sample in train_set _p is i, and when y _qi is equal to 0, it means that the label loaded by the qth sample in train_set _p is not i;

(5.3), utilize L _Pt to carry out backpropagation and update multi-input residual network;

和多输入残差网络模型model_m；(5.4), repeat steps (5.2)-(5.3) t times, until all the samples in the train_set are input to the input residual network to complete the current round of training, and then save the training loss after the completion of the current round of training

And multi-input residual network model model _m ;

(5.5), input the validation_set to the multi-input residual network model model _m saved in this round, predict the label value, and then calculate the verification accuracy Vacc _m ;

Among them, e _r is equal to 1 or 0. If the label of the rth sample in the validation_set predicted by the model model _m is the same as the label loaded by the sample in the validation_set, the value is 1, otherwise the value is 0;

(5.6), repeat steps (5.2)-(5.5) for N rounds to obtain N multi-input residual network models and verification accuracy, recorded as model=[model ₁ ...model _m ...model _N ], Vacc=[Vacc ₁ ... Vacc _m ... Vacc _N ];

(5.7), utilizing the index index _max where the maximum value is located in Vacc, obtain the optimal model model _max =model[index _max ] after N rounds of training, and as the failure prediction model of the analog filter circuit to be tested;

(6), according to the method described in step (1)-(2), obtain a group of eigenvectors of the analog filter circuit to be tested under the unknown fault state, then input to the fault prediction model, thereby obtaining the location of the analog filter circuit to be tested Fault state i.

The purpose of the invention of the present invention is achieved like this:

The present invention is based on the multi-input residual network fault diagnosis method for analog filter circuits. Firstly, the relevant parameters of Monte Carlo analysis are set, and the simulation analysis of the analog filter circuit is carried out through the Monte Carlo statistical analysis method, so as to obtain the output of the analog filter circuit under different fault states. Voltage signal; then use the empirical wavelet transform to obtain the modal components of the voltage signal, and form them into a training set and a test set respectively, use the training set for training multi-input ResNet, use the test set for testing, and save the corresponding The multi-input ResNet model and parameters Parameters; finally, the unknown fault state of the analog filter circuit is diagnosed through the trained multi-input ResNet model and Parameters.

The fault diagnosis method of the analog filter circuit based on the multi-input residual network of the present invention also has the following beneficial effects:

When a component in the analog filter circuit fails, the filtering performance of the circuit is affected. The characteristics of each modal component of the output voltage signal can be fully learned through the multi-input ResNet, without the need for feature extraction, selection and fusion methods, which not only reduces the Simulate the complexity of filter circuit fault diagnosis, and improve the accuracy of fault diagnosis.

Description of drawings

Fig. 1 is the flow chart of the present invention's analog filter circuit fault diagnosis method based on multi-input residual network;

Fig. 2 is the simulation schematic diagram for the second-order high-pass filter circuit of four operational amplifiers in the embodiment of the present invention;

Fig. 3 is the voltage signal at the output terminal of the second-order high-pass filter circuit of four operational amplifiers under normal state;

Fig. 4 is the modal component of the output terminal voltage signal under the normal state of the four-op-amp second-order high-pass filter circuit obtained from empirical wavelet transform decomposition;

Fig. 5 is the specific structure of multi-input ResNet in the present invention;

Fig. 6 is a graph of the training loss and verification accuracy of multi-input ResNet in the present invention as a function of the number of training rounds;

Figure 7 is a graph of the training loss and verification accuracy of the traditional single-input ResNet as a function of the number of training rounds.

Detailed ways

Specific embodiments of the present invention will be described below in conjunction with the accompanying drawings, so that those skilled in the art can better understand the present invention. It should be noted that in the following description, when detailed descriptions of known functions and designs may dilute the main content of the present invention, these descriptions will be omitted here.

Example

FIG. 1 is a flow chart of the method for diagnosing faults of an analog filter circuit based on a multi-input residual network in the present invention.

In this embodiment, as shown in Figure 1, a method for diagnosing faults of an analog filter circuit based on a multi-input residual network of the present invention comprises the following steps:

S1. Use the Monte Carlo analysis method to simulate and analyze the analog filter circuit to be tested, and obtain the voltage signal v _ij (t) of k × n sets of sampling points at the circuit output terminal under k fault states, where i=1,2,... , k, j=1, 2,..., n, n represents the group number of voltage signals collected at the output end of the analog filter circuit to be tested;

The specific method is as follows:

S1.1, setting the tolerance of resistance and capacitance in the analog filter circuit to be tested and the fault parameters in the i-th fault state;

结束时间/>

时间步长TSTEP为/>

运行次数为n，f表示待测模拟滤波电路输入端输入的信号的频率；S1.2. Set the start time of Monte Carlo analysis

end time/>

The time step TSTEP is />

The number of times of operation is n, and f represents the frequency of the signal input by the input terminal of the analog filter circuit to be tested;

S1.3. Under k kinds of fault states, the voltage signal at the output terminal of the analog filter circuit to be tested is used as the output variable analyzed by the Monte Carlo statistical method, thereby obtaining a voltage signal v _ij (t ),in

In this embodiment, the Monte Carlo statistical method is used to simulate and analyze the second-order high-pass filter circuit with four operational amplifiers to obtain a fault data set.

The simulation schematic diagram of the second-order high-pass filter circuit of four operational amplifiers is shown in Figure 2. A square wave signal with a frequency of 2K HZ, an amplitude of 5V, and a duty cycle of 50% is input from the input terminal of the circuit. The resistance and capacitance values in the second-order high-pass filter circuit are shown in Figure 3; after sensitivity analysis, the second-order high-pass filter circuit with four operational amplifiers mainly has 12 failure modes, which are R1↓, R1↑, R2↓, R2↑ , R3↓, R3↑, R4↓, R4↑, C1↓, C1↑, C2↓, C2↑, represented by F1～F12 respectively, F0 indicates that the circuit is in a normal state, Table 1 is the setting of specific faults in this embodiment . Set the tolerances of resistors and capacitors in the second-order high-pass filter circuit of four operational amplifiers to 5%, set the start time of Monte Carlo statistical analysis to 0.005s, end time to 0.01s, and set the time step to 6.25× 10 ^-6 s, the number of operations is 400, and the voltage signals v _ij (t) with 800 sampling points under 13 different fault states composed of F1～F12 and F0 are respectively obtained, i=1,...,13,j=1, 2,...,400;

fault number Fault type nominal value fault value F0 - - - F1 R1↓ 6.2kΩ 4.34kΩ F2 R1↑ 6.2kΩ 8.06kΩ F3 R2↓ 6.2kΩ 4.34kΩ F4 R2↑ 6.2kΩ 8.06kΩ F5 R3↓ 6.2kΩ 4.34kΩ F6 R3↑ 6.2kΩ 8.06kΩ F7 R4↓ 1.6kΩ 1.12kΩ F8 R4↑ 1.6kΩ 2.08kΩ F9 C1↓ 5nF 3.5kΩ F10 C1↑ 5nF 6.5kΩ F11 C2↓ 5nF 3.5kΩ F12 C2↑ 5nF 6.5kΩ

Table 1

S2. Use the traditional empirical wavelet transform to decompose the voltage signal v _ij (t) to obtain 6 modal components ewt _ijq (t), q=1,2,...,6, respectively divide 400 groups of modal components ewt _ijq (t) constitutes the eigenvector ewt _ij , and the modal components of the output terminal voltage signal of the four-op-amp second-order high-pass filter circuit under normal conditions obtained by traditional empirical wavelet transform decomposition are shown in Figure 4;

验证数据集/>

S3. The ewt _ij under 13 kinds of fault states are composed of training data sets according to the ratio of 3:1

validation dataset />

和验证集

S4. Use the serial number i of the fault state corresponding to ewt _ij as a label, load labels for the training data set and the verification data set respectively, and construct a training set for a multi-input residual network model

and validation set

S5. Building a multi-input residual network;

As shown in Figure 5, the multi-input residual network model includes an input layer, 6 residual block net-blocks with the same structure, a fully connected layer and an output layer;

Among them, each net-block includes 1 layer of convolution unit, 4 layers of residual units with the same structure and 1 layer of average pooling layer;

In the convolution unit, it first passes through a one-dimensional convolution layer with a kernel size of 5, an output channel of 16, and a step size of 2, then passes through a batch regularization layer and a Relu layer, and finally passes through a one-dimensional convolutional layer with a kernel size of 3 and a step size of 2. average pooling layer;

The residual unit first passes through a one-dimensional convolutional layer with a kernel size of 3, an input channel of channel _l = 2 ^3+l , l = 1, 2, 3, 4, an output channel of channel _l , and a step size of 2, Then go through the regularization layer and the Relu layer; then go through the one-dimensional convolution layer with a kernel size of 3, input and output channels of channel _l , and a step size of 1, and then go through the regularization layer and the Relu layer, where l represents the residual The number of layers corresponding to the unit;

Concatenate the output of 6 net-blocks together, through the fully connected layer to the output layer;

S6. Training a multi-input residual network;

Use train_set to train the multi-input residual network model, use validation_set to verify the model trained by train_set, and obtain the optimal multi-input residual network model model _max ;

S6.1. Set the number of training rounds of the multi-input residual network model to N, and set the input sample size to batch_size in each round of training; it takes t iterations for the train_set to be completely input into the multi-input residual network, and each iteration is multi-input Input residual network input is recorded as train_set _p , p=1...t,

S6.2. Train_set _p is input to the multi-input residual network, and the sofmax function is used to predict the probability pro _qi that the predicted label of the qth sample in train_set _p belongs to the loaded label i, q=1...batch_size, and calculate the train_set in this iteration The training loss TL _mp obtained by _p training the multi-input residual network, m=1...N;

Among them, y _qi is equal to 0 or 1. When y _qi is equal to 1, it means that the label loaded by the qth sample in train_set _p is i, and when y _qi is equal to 0, it means that the label loaded by the qth sample in train_set _p is not i;

S6.3, using L _Pt to perform backpropagation to update the multi-input residual network;

和多输入残差网络模型model_m；S6.4. Repeat steps S6.2-S6.3 for a total of t times until all the samples in the train_set are input to the residual network at most, thus completing the current round of training, and then save the training loss after the completion of the current round of training

And multi-input residual network model model _m ;

S6.5. Input the validation_set into the multi-input residual network model model _m saved in this round, predict the label value, and then calculate the verification accuracy Vacc _m ;

Among them, e _r is equal to 1 or 0. If the label of the rth sample in the validation_set predicted by the model model _m is the same as the label loaded by the sample in the validation_set, the value is 1, otherwise the value is 0;

S6.6. Repeat steps S6.2-S6.5 for a total of N rounds to obtain N multi-input residual network models and verification accuracy, recorded as model=[model ₁ ...model _m ...model _N ], Vacc=[Vacc ₁ ...Vacc _m ...Vacc _N ];

S6.7. Utilize the index index _max where the maximum value is located in Vacc to obtain the optimal model model _max =model[index _max ] after N rounds of training, and use it as the failure prediction model of the analog filter circuit to be tested;

Set the number of training rounds of the multi-input residual network to 100, and the input sample size for each round of training is set to 128. It takes 31 iterations for the train_set to be completely input to the multi-input residual network, and each iteration to the multi-input residual network The input is recorded as train_set _p , p=1...31; the training loss TL based on the multi-input residual network in this embodiment obtained by using train_set and validation_set, the variation curve of verification accuracy Vacc with the number of training rounds is shown in Figure 6, based on The variation curves of the training loss STL and verification accuracy SVacc of the traditional single-input residual network with the number of training rounds are shown in Figure 7. The maximum verification accuracy of the two residual network models after 100 rounds of verification is shown in Table 2.

Network Type Multi-Input Residual Networks Single Input Residual Network Maximum Validation Accuracy 99.46% 86.923%

Table 2

From Figure 6, Figure 7 and Table 2, it can be seen that compared with the traditional single-input residual network, the multi-input residual network has better convergence effect on training loss, higher verification accuracy, stronger generalization ability, and can learn more fully To simulate the fault characteristics of the filter circuit in different fault states, it is beneficial to improve the efficiency and accuracy of fault diagnosis;

S7. Obtain a set of eigenvectors of the analog filter circuit to be tested in an unknown fault state according to the method described in steps S1-S2, and then input them into the fault prediction model to obtain the fault state i of the analog filter circuit to be tested.

Although the illustrative specific embodiments of the present invention have been described above, so that those skilled in the art can understand the present invention, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, As long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concept of the present invention are included in the protection list.

Claims (3)

1. a kind of analog filtering circuit fault diagnosis method based on multi-input residual network, is characterized in that, comprises the following steps: (1), use the Monte Carlo analysis method to carry out simulation analysis on the analog filter circuit to be tested, and obtain the voltage signal v _ij (t) of k × n groups of sampling points at the output end of the circuit under k kinds of fault states, i=1,2 ,...,k,j=1,2,...,n, n represents the group number of voltage signals collected at the output end of the analog filter circuit to be tested; (2), use the traditional empirical wavelet transform algorithm to decompose the voltage signal v _ij (t), and obtain m modal components, wherein the qth modal component is expressed as ewt _ijq (t), q=1,2 ,...,m, and then m modal components are composed of eigenvectors ewt _ij =[ewt _ij1 ; ewt _ij2 ;...;ewt _ijq (t);...;ewt _ijm ]; (3), randomly distribute n ewt _ij under k kinds of fault states according to the ratio of r ₁ :r ₂ to form training data sets respectively

Validation dataset

(4), use the serial number i of the fault state corresponding to ewt _ij as a label, load the labels for the training data set and the verification data set respectively, and construct the training set of the multi-input residual network model

and validation set

(5), build a multi-input residual network, use train_set to train the multi-input residual network model, use validation_set to verify the model trained by train_set, and obtain the optimal multi-input residual network model model _max ; (5.1), set the training rounds of the multi-input residual network model to N, and set the input sample size to batch_size in each round of training; it takes t iterations for the train_set to be completely input into the multi-input residual network, and each iteration is multi-input Input residual network input is recorded as train_set _p , p=1...t,

(5.2), train_set _p is input to the multi-input residual network, and the predicted label of the qth sample in train_set _p is predicted by the Softmax function to be the probability pro _qi of the loaded label i, q=1...batch_size, and the train_set _p in this iteration is calculated The training loss TL _mp obtained by training the multi-input residual network, m=1...N;

Among them, y _qi is equal to 0 or 1. When y _qi is equal to 1, it means that the label loaded by the qth sample in train_set _p is i, and when y _qi is equal to 0, it means that the label loaded by the qth sample in train_set _p is not i; (5.3), utilize L _Pt to carry out backpropagation and update multi-input residual network; (5.4), repeat steps (5.2)-(5.3) t times, until all the samples in the train_set are input to the input residual network to complete the current round of training, and then save the training loss after the completion of the current round of training

And multi-input residual network model model _m ; (5.5), input the validation_set to the multi-input residual network model model _m saved in this round, predict the label value, and then calculate the verification accuracy Vacc _m ;

Among them, e _r is equal to 1 or 0. If the label of the rth sample in the validation_set predicted by the model model _m is the same as the label loaded by the sample in the validation_set, the value is 1, otherwise the value is 0; (5.6), repeat steps (5.2)-(5.5) for N rounds to obtain N multi-input residual network models and verification accuracy, recorded as model=[model ₁ ...model _m ...model _N ], Vacc=[Vacc ₁ ... Vacc _m ... Vacc _N ]; (5.7), utilizing the index index _max where the maximum value is located in Vacc, obtain the optimal model model _max =model[index _max ] after N rounds of training, and as the failure prediction model of the analog filter circuit to be tested; (6), according to the method described in step (1)-(2), obtain a group of eigenvectors of the analog filter circuit to be tested under the unknown fault state, then input to the fault prediction model, thereby obtaining the location of the analog filter circuit to be tested Fault state i. 2. the analog filter circuit fault diagnosis method based on the multi-input residual network according to claim 1, is characterized in that, in the described step (1), obtain k * n groups of sampling points of circuit output terminal under k kinds of fault states as The specific method of the voltage signal v _ij (t) of M is: (2.1), setting the tolerance of resistance and capacitance in the analog filter circuit to be tested and the fault parameter under the i kind of fault state; (2.2), set the start time of Monte Carlo analysis

end time/>

time step />

The number of runs is n; (2.3), under k kinds of fault states, the voltage signal at the output end of the analog filter circuit to be tested is used as the output variable analyzed by the Monte Carlo statistical method, thereby obtaining the voltage signal v _ij (t ),in

3. the analog filter circuit fault diagnosis method based on multi-input residual network according to claim 1, is characterized in that, described multi-input residual network model comprises input layer, m structural identical residual block, fully connected layer and the output layer; Among them, the outputs of m residual blocks are connected in series, and each residual block includes 1 layer of convolution unit, 4 layers of residual units with the same structure and 1 layer of average pooling layer; The convolution unit first passes through a one-dimensional convolution layer with a kernel size of 5, an output channel of 16, and a step size of 2, then passes through a batch regularization layer and a Relu layer, and finally passes through a one-dimensional convolution layer with a kernel size of 3 and a step size of 2. One-dimensional average pooling layer; The residual unit first passes through a one-dimensional convolutional layer with a kernel size of 3, an input channel of channel _l = 2 ^3+l , l = 1, 2, 3, 4, an output channel of channel _l , and a step size of 2, Then go through the regularization layer and the Relu layer; then go through the one-dimensional convolution layer with a kernel size of 3, input and output channels of channel _l , and a step size of 1, and then go through the regularization layer and the Relu layer, where l represents the residual The number of layers corresponding to the unit layer.

Images