Analog circuit fault diagnosis method based on wavelet scattering and integrated learning

Abstract

The invention discloses an analog circuit fault diagnosis method based on wavelet scattering and integrated learning to solve the problem that fault diagnosis is difficult due to analog circuit fault response aliasing. Firstly, circuit response is divided into a plurality of frequency band subsets through wavelet scattering transformation, and Fisher discriminant analysis is utilized to enhance fault features of the subsets; secondly, each frequency band subset is sent to different extreme learning machines under Bagging integration, the classification accuracy of each fault mode in the frequency band subsets serves as the category weight of the extreme learning machine, then the output value of each extreme learning machine is weighted to obtain a fused output result, and the fault category is determined according to the fused output result; finally, simulation is conducted through two instance circuits, the diagnosis accuracy rate in the simulation result is 100%, it is indicated that the method has feasibility and effectiveness, and fault classification and positioning can be achieved.

Description

Technical Field

[0001] The present invention relates to the field of analog circuit fault diagnosis, to a frequency band subset feature extraction and enhancement method of analog circuit measurement point output signals, and to a pattern recognition technology, and in particular to an analog circuit fault diagnosis method based on wavelet scattering and ensemble learning. Background Art

[0002] In modern electronic systems, the integration of mixed analog and digital circuits is getting higher and higher. As an indispensable part of analog circuits, the fault diagnosis of analog circuits is complex and unpredictable due to the tolerance of components themselves and external interference. And with the increase in circuit integration, analog circuit diagnosis is becoming more and more difficult. Traditional analog circuit fault diagnosis methods such as fault dictionary method, probability statistics method and component parameter identification method can no longer meet actual needs. In response to this situation, domestic and foreign scholars have conducted research on new intelligent methods and proposed a large number of artificial intelligence-based methods, such as neural networks, support vector machines, wavelet transforms, and modal decomposition.

[0003] Intelligent diagnosis methods mainly include fault feature extraction and diagnostic device construction. The fault features of analog circuits are usually extracted from the responses of the circuit under test in the time domain and frequency domain. Appropriate feature extraction methods can provide the inherent information and essential information of the signal with minimal overlap. Generally, the feature extraction of the output response uses wavelet packet transform (WPT) and wavelet transform (WT) to obtain the decomposition coefficients of each frequency band and obtain the frequency band energy value, and combines dimensionality reduction methods such as principal component analysis to obtain fault features. However, this method ignores the recognition ability of different frequency bands for faults and does not make full use of limited fault information. At the same time, this feature extraction method cannot achieve good consistency with the pattern recognition model. The pattern recognition model cannot prescribe the right medicine for the extracted features, and the matching degree of the algorithm is not high, resulting in a low fault diagnosis rate. Summary of the invention

[0004] In view of the deficiencies of the prior art, the present invention proposes an analog circuit fault diagnosis method based on wavelet scattering and ensemble learning to solve the problem of low fault diagnosis accuracy in the prior art.

[0005] The present invention extracts and classifies the output signal based on the phenomenon that the output signal of the circuit is slightly different due to the offset of parameters of different components in the circuit, thereby solving the problem of low accuracy of fault diagnosis of analog circuits. The present invention only targets soft faults of resistors and capacitors in the circuit, and each fault only occurs from a single component, and does not include the special case where multiple components fail at the same time. A soft fault is a situation where the resistance value or capacitance value is offset by ±30% of the nominal value, such as a resistor with a nominal value of 100kΩ, and a soft fault of 70kΩ or 130kΩ, and the diagnosed circuit is collectively referred to as a measured circuit. If there is only one resistor and capacitor component that can be diagnosed in a measured circuit, then the circuit has three fault categories, namely, offset of 70% of the nominal value, no fault, and offset of 130% of the nominal value; if there are two resistor and capacitor components that can be diagnosed, then the circuit has five fault categories, among which no fault is also counted as one of the fault categories. Whether the resistor and capacitor components in the tested circuit can be diagnosed is determined by the sensitivity analysis in the simulation software. For example, if a time series test signal is input to the input end of the tested circuit, if the output response of the component under soft fault conditions and no fault conditions is no different, then the sensitivity of the component in the tested circuit cannot meet the basic conditions required for fault diagnosis, and the device cannot be diagnosed. Therefore, the output response of the circuit is the output of the output point of the tested circuit under each fault category collected when a time series test signal is input to the input end of the tested circuit, and the diagnosable device is determined, and the diagnosable device is artificially set with a soft fault. In order to ensure the effectiveness of the method of the present invention, the output response is divided into a training set response and a test set response in a ratio of 7:3. The training set response is used to train the model parameters in the method of the present invention, and the test set response is used to verify the final diagnostic effect of the method of the present invention. In the pre-training part of the method of the present invention, there is a secondary division of the training set response, and the secondary division will be described in detail in the specific implementation method; the genetic algorithm used for training model parameters is relatively classic and will not be elaborated in detail.

[0006] The training set response is divided into multiple frequency band subsets by wavelet scattering transform, and the fault characteristics of the frequency band subsets are enhanced by Fisher discriminant analysis. Secondly, each frequency band subset is sent to different extreme learning machines under bagging integration, where the extreme learning machine under bagging integration is mainly divided into pre-training extreme learning machine and classification extreme learning machine, and each frequency band subset corresponds to a pre-training extreme learning machine. The classification accuracy of each fault category in the frequency band subset is obtained in the pre-training extreme learning machine, and the classification accuracy is used as the category weight of each fault category in the frequency band subset. Then, the output value of each pre-training extreme learning machine is multiplied and summed with the category weight, and the weighted fusion result is sent to the classification extreme learning machine to obtain the diagnosis result, and the fault category is determined accordingly. In the present invention, this method combining pre-training and classification is collectively referred to as bagging subclass extreme learning machine.

[0007] The objective of the present invention is achieved through the following technical solutions:

[0008] A method for diagnosing analog circuit faults based on wavelet scattering and ensemble learning comprises the following steps:

[0009] Simulation data collection phase:

[0010] Step 1: determine the circuit under test, perform sensitivity analysis on the circuit under test in the simulation software, and determine the diagnosable components in the circuit under test;

[0011] Step 2: Set soft faults for each diagnosable device, and perform 100 Monte Carlo analyses on each fault type under the conditions of 5% resistance tolerance and 10% capacitance tolerance, and obtain 100 simulation samples for each fault type. The samples collected by simulation are used as the output response of the circuit under test, thereby obtaining a simulation data set;

[0012] Frequency band subset extraction stage:

[0013] Step 3: Divide all collected output responses into training set responses and test set responses in a ratio of 7:3 or 6:4. Perform wavelet scattering transform on each sample in the training set response to obtain a scattering coefficient matrix for each sample. The scattering coefficient matrix is ​​composed of convolution coefficients generated by different scattering channels, and different scattering channels represent the response characteristics within a specific frequency band. When verifying the final diagnostic effect, the test set response also uses wavelet scattering transform to obtain the scattering coefficient matrix of each test set response sample.

[0014] Step 4: Based on the relationship between the scattering coefficient matrix and the scattering channel, the two-dimensional scattering coefficient matrix obtained by transforming the training set response is split into multiple one-dimensional samples according to the scattering channel, and these one-dimensional samples are recombined into the same number of subsets according to the number of scattering channels. Each subset is composed of the convolution coefficients generated by the same scattering channel and represents the response characteristics within a specific frequency band. Therefore, they are collectively referred to as frequency band subsets. In this way, the division of frequency band subsets is completed. When verifying the final diagnostic effect of the test set response, the scattering coefficient matrix of the test set response is also divided into frequency band subsets using the method in this step.

[0015] Fault feature enhancement stage:

[0016] Step 5: Calculate the global scatter matrix, intra-class scatter matrix and inter-class scatter matrix of each frequency band subset, and use Fisher discriminant analysis to obtain the mapping matrix of each frequency band subset. The mapping matrix and the frequency band subset are multiplied to project the frequency band subset into a multidimensional space, so that the samples of the same fault category after mapping are clustered as much as possible, and the samples of different fault categories are separated as much as possible, thereby realizing the fault feature enhancement of each frequency band subset;

[0017] Feature fusion and fault diagnosis stage:

[0018] Step 6: After the fault feature enhancement, the frequency band subsets are respectively sent to the pre-trained extreme learning machine in the Bagging subclass extreme learning machine. The enhanced frequency band subsets are divided into a pre-training set and a pre-test set according to 7:3 in the pre-training stage. The pre-trained extreme learning machine in the Bagging subclass extreme learning machine is trained by the pre-training set, and the pre-test set is used to classify and verify the pre-training result. The verification result consists of the classification accuracy of each fault category, and the verification result is used as the category weight of the enhanced frequency band subset, thereby obtaining the category weight of each enhanced frequency band subset;

[0019] Step seven, multiply the output of the pre-trained extreme learning machine output layer in the Bagging subclass extreme learning machine by the corresponding category weight to achieve feature fusion, send the result of feature fusion to the classification extreme learning machine in the Bagging subclass extreme learning machine for training, and compare the output Y of the classification extreme learning machine with the label to obtain the diagnostic accuracy; during the training process, a genetic algorithm is used to optimize the parameters of the extreme learning machine, and the classification accuracy of the test set response is used as the final diagnostic effect to complete the fault diagnosis.

[0020] The present invention proposes innovations based on frequency band information and pattern recognition:

[0021] The present invention first extracts the output response frequency band subset of the circuit output point, enhances the fault characteristics of the frequency band subset through the Fisher criterion, and then establishes a diagnostic classification model based on the recognition ability of different fault modes under different frequency bands, thereby realizing accurate recognition of analog circuit faults.

[0022] Effects or advantages of the present invention:

[0023] The present invention establishes an analog circuit fault diagnosis model based on the differences in circuit output under different fault modes. Since the simulation is performed under the condition that the resistance tolerance in the measured circuit is 5% and the capacitance tolerance is 10%, it has a certain anti-interference ability. The simulation verification is carried out with two example circuits, Sallen-Key bandpass filter and four-op-amp biquadratic, and the diagnostic accuracy rate in the simulation results is 100%, indicating that the method is feasible and effective, and can realize the classification and location of faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is the framework diagram of wavelet scattering network;

[0025] Figure 2 It is a schematic diagram of frequency band subset and fault feature enhancement;

[0026] Figure 3 This is the structure diagram of the extreme learning machine of the Bagging subclass;

[0027] Figure 4 It is a block diagram of the analog circuit fault diagnosis method of wavelet scattering and ensemble learning in the embodiment;

[0028] Figure 5 This is the schematic diagram of the Sallen_key bandpass filter circuit;

[0029] Figure 6 This is the schematic diagram of the four-op-amp biquad circuit. DETAILED DESCRIPTION

[0030] The following is a further description of the invention in conjunction with the accompanying drawings and embodiments, but the invention is not limited thereto.

[0031] In order to better understand the present invention, the basic principles and related concepts of the present invention are briefly introduced below.

[0032] Wavelet dispersion band characteristics:

[0033] The wavelet scattering network was proposed by Professor Mallet in 2012. This method obtains the corresponding scattering coefficient through complex-valued wavelet transform, modular operation and average cascade operation. It has the characteristics of reducing data volume, model complexity and automatically extracting the advantages of relevant compact framework. In order to obtain signal features with both translation invariance and local deformation stability, the complex-valued Morlet wavelet, i.e., bandpass filter, is first convolved with the input response, and then the modular operation is used. In order to ensure the stability of the high-frequency coefficient, the scattering coefficient is obtained by the average operation, i.e., convolution low-pass filter. The above process can ensure the stability of local deformation, but high-frequency information will be lost, and the discriminative ability of diagnostic features will also be weakened. In order to restore the lost high-frequency information, after the convolution low-pass filter obtains the general features, the wavelet module of the previous layer is used to convolve the higher frequency filter to restore the lost high-frequency part. The pth channel of the wavelet scattering network is shown in formula (1):

[0034] S[j1,j2,...,jp]x=U[j1,j2,...,jp]x*φ (1)

[0035] Where j represents the scale coefficient, φ is the low-pass filter, ψ is the high-frequency wavelet, x is the training set response divided by the circuit output response, and the high-frequency part is expressed as:

[0036]

[0037] j1<Λ<j p (3)

[0038] At the same time, Mallet proved in the experiment that when the number of wavelet scattering network layers reaches three, the scattering energy will reach more than 99%. Too many network layers will increase the computational overhead, and the redundant scattering layer coefficients are close to 0, causing feature redundancy. Therefore, the present invention adopts a three-layer scattering network structure. Each layer of the network contains one or more scattering channels. The filter contained in the scattering channel convolves the input response, and the output after convolution is the scattering coefficient of the channel. The input response obtains the corresponding scattering coefficient in each scattering channel and forms a scattering coefficient matrix. The wavelet scattering network framework is as follows: Figure 1 As shown in Figure 2, since the input response will convolve high-pass filters with different numbers or scale coefficients on different scattering channels, each scattering channel coefficient represents the response characteristics within a specific frequency band.

[0039] Fault signature enhancement:

[0040] The training set responses are transformed by wavelet scattering, and the obtained scattering coefficient matrix consists of p scattering channels, which can be divided into p frequency band subsets. The g-th frequency band subset is represented by sg, 1≤g≤p. The discriminable fault features and tolerance interference are distributed in these frequency band information. In order to highlight the discriminable fault features in the frequency band, Fisher discriminant analysis is introduced.

[0041] For the g-th frequency band subset, the mapping matrix w is calculated to enhance the fault characteristics: g , so that the samples of the same type after mapping are gathered as much as possible, and the samples of different types are separated as much as possible. Define the global divergence matrix S gt , intra-class scatter matrix S gw and the inter-class scatter matrix S gb :

[0042] S gt =(s g -d g )(s g -d g ) T (4)

[0043]

[0044] where d g represents the sample mean of the g-th frequency band subset, l represents the number of fault categories, s gi represents the samples belonging to fault category i in the g-th frequency band subset, 1≤i≤l, d gi It represents the mean value of samples belonging to fault category i in the p-th frequency band subset, m i Represents the number of training samples belonging to fault category i in the frequency band subset. The mapping matrix w g It is composed of the closed-form solution of maximizing the Fisher criterion function. The mapping matrix and the frequency band subset are projected into the multidimensional space by the product method. The Fisher criterion function maximized by the g-th frequency band subset is shown in formula (7).

[0045]

[0046] The fault characteristics of each frequency band subset are enhanced by the projection of the mapping matrix, which suppresses the interference influence distributed in different frequency bands to a certain extent, and provides high-quality frequency band subsets for subsequent feature fusion and fault diagnosis. The frequency band subsets and fault characteristics are enhanced as shown in Figure 2 As shown, the projection of the mapping matrix is ​​shown in formula (8), where s g ′ represents the g-th frequency band subset after enhancement,

[0047] s g ′=w g s g (8).

[0048] Extreme Learning Machine:

[0049] The extreme learning machine (ELM) is mainly composed of an input layer, a hidden layer, and an output layer. It is essentially a single hidden layer feedforward neural network algorithm. The input layer and the hidden layer, as well as the hidden layer and the output layer are fully connected.

[0050] In the extreme learning machine subclass of Bagging, s g ′ corresponds to the g-th pre-trained extreme learning machine ELM g , m represents the number of training samples, s gm ′ represents the mth training sample, the number of features of each sample is n, and the sample label is T g To ensure the lowest output deviation, ELM g There is a weight a between the input layer and the hidden layer g , there is a bias b in the hidden layer g There is a weight β between the output layer and the hidden layer g , the hidden layer neuron output matrix H g The calculation equation is:

[0051]

[0052] In formula (9), η(·) is the activation function, h is the number of neurons in the hidden layer, and a gh 、b gh Represent the weight and bias of the hth neuron in the hidden layer. g The goal of training is to minimize the training error so that the output of the model after training is close to the actual output error E g Minimum. Its objective equation is:

[0053]

[0054] Where T g is the label, i.e. the expected output. The weight β between the output layer and the hidden layer g It can be obtained by solving the least squares:

[0055]

[0056] In formula (11), C is the penalty factor, ELM g The output is Y g , Y g =H g β g The output of the classification extreme learning machine is Y. ELM needs to optimize the penalty factor C and the number of hidden layer neurons h. The activation function η() uses the Sigmoid function.

[0057] Bagging subclass extreme learning machine:

[0058] In order to use the frequency band subset with the largest difference between different faults as the basis for fault feature extraction and effectively reduce the aliasing of analog circuit fault responses, a Bagging subclass extreme learning machine including feature fusion and fault classification is proposed. The specific steps of fusion and classification are as follows:

[0059] 1. The output response is divided into training set response and test set response in a ratio of 7:3. After the frequency band subsets are extracted and the fault features are enhanced, the same number of pre-trained ELMs are generated according to the number of enhanced frequency band subsets p. The corresponding labels of the inputs and samples of p ELMs are: where s g ′ represents the g-th enhanced frequency band subset, which is also the ELM g Input, T g Indicates g ’ in the corresponding label.

[0060] 2. The frequency band subset enhanced in step 1 is divided into pre-training set and pre-test set according to 7:3. The pre-training set and pre-test set are respectively used for 0.7s g ′ and 0.3s g ′ represents. The pre-trained extreme learning machine in the Bagging subclass extreme learning machine is trained through the pre-training set, and the pre-test set is used to classify and verify the pre-training results. The verification results are composed of the classification accuracy of each fault category, and the verification results are used as the category weights of the enhanced frequency band subsets, thereby obtaining the category weights z of each enhanced frequency band subset g .

[0061] 3. Output Y from the output layer of the pre-trained extreme learning machine in the p Bagging subclass extreme learning machines g The sum of the products of the corresponding output weights is used as the input X of the classification extreme learning machine in the Bagging subclass extreme learning machine, as shown in formula (12). The feature fusion is completed here, and the output Y of the classification extreme learning machine is compared with the label to obtain the diagnosis result. The structure of the Bagging subclass extreme learning machine is as follows Figure 3 As shown,

[0062]

[0063] Diagnostic model based on wavelet scattering network and ensemble learning:

[0064] In order to improve the fusion diagnosis performance of the Bagging subclass extreme learning machine, the present invention adopts the classic genetic algorithm (GA) to optimize the parameters of the ELM corresponding to the frequency band subset and the number of hidden layer neurons h and the penalty factor C in the classification ELM. GA is a representative classic method in evolutionary algorithms, mainly including selection, crossover and mutation. Combined with the Bagging subclass extreme learning machine, the diagnostic accuracy is used as the fitness value of the genetic algorithm. If the optimal fitness no longer changes, it means that the training is completed.

[0065] The analog circuit fault diagnosis framework proposed by the present invention is as follows: Figure 4 As shown in the figure, the whole framework is divided into three parts: frequency band subset extraction, fault feature enhancement, and feature fusion and fault diagnosis. In the first part, the scattering coefficient matrix is ​​obtained by convolution modulo operation through the wavelet scattering network, and multiple frequency band subsets are divided. In the second part, based on the frequency band subsets, the fault features of each frequency band subset are enhanced using the Fisher discriminant criterion. The enhanced frequency band subsets are sent to the Bagging subclass extreme learning machine in the third part, and the model is trained by GA. After training, it is sent to the test set to obtain the diagnosis results and complete the fault diagnosis.

[0066] Example:

[0067] like Figure 4 As shown, a method for fault diagnosis of analog circuits based on wavelet scattering and ensemble learning comprises the following steps:

[0068] Simulation data collection phase:

[0069] S1, determine the circuit under test, perform sensitivity analysis on the circuit under test in the simulation software, and determine the diagnosable components in the circuit under test under the test signal according to the analysis results;

[0070] S2, set soft faults for each diagnosable device, and perform 100 Monte Carlo analyses on each fault category under the conditions of 5% resistance tolerance and 10% capacitance tolerance, and obtain 100 simulation samples for each fault category. The samples collected by simulation are used as the output response x of the circuit under test, thereby obtaining a simulation data set;

[0071] Frequency band subset extraction stage:

[0072] S3, the response x obtained in S2 is divided into a training set response and a test set response in a ratio of 7:3, and each sample in the training set response is subjected to a wavelet scattering transform to obtain a scattering coefficient matrix for each sample;

[0073] S4, split the scattering coefficient matrix of each sample in S3 into multiple one-dimensional samples. The scattering coefficient matrix consists of p scattering channels and can be divided into p frequency band subsets. The gth frequency band subset is represented by s g , 1≤g≤p, to achieve the division of frequency band subsets. When the test set response verifies the final diagnostic effect, the scattering coefficient matrix of the test set response is also divided into frequency band subsets using the method in this step, and p frequency band subsets can be divided;

[0074] Fault feature enhancement stage:

[0075] S5, for s in S4 g , respectively define the global divergence matrix S gt , intra-class scatter matrix S gw And the inter-class divergence matrix Sgb:

[0076] S gt =(s g -d g )(s g -d g ) T (13)

[0077]

[0078] where d g represents the sample mean of the g-th frequency band subset, l represents the number of fault categories, s gi represents the samples belonging to fault category i in the g-th frequency band subset, 1≤i≤l, d gi It represents the mean value of samples belonging to fault category i in the p-th frequency band subset, m i represents the number of training samples belonging to fault category i in the frequency band subset, and the mapping matrix w g The closed-form solution of maximizing the Fisher criterion function is formed. The Fisher criterion function that maximizes the g-th frequency band subset is shown in formula (16):

[0079]

[0080] Formula (17)s g ′ represents the gth frequency band subset after enhancement.

[0081] s g ′=w g s g (17)

[0082] Feature fusion and fault diagnosis stage:

[0083] S6, the frequency band subset s after fault feature enhancement of S5 g′ are sent to the pre-training extreme learning machine in the Bagging subclass extreme learning machine. The enhanced frequency band subset is divided into a pre-training set and a pre-test set according to 7:3 in the pre-training stage. The pre-training set and the pre-test set are respectively used for 0.7s g ′ and 0.3s g ′ indicates that the pre-test set performs classification verification on the pre-training results. The verification results consist of the classification accuracy of each fault category, and the verification results are used as the category weights of the enhanced frequency band subsets, thereby obtaining the category weights z of each enhanced frequency band subset g ;

[0084] S7, output the ELM of each frequency band subset in S6 to Y g With weight z g The product is multiplied, and the product results are added to complete the feature fusion. The fused result is sent to the diagnosis ELM to obtain the final result, as shown in formula (18). The number of hidden layer neurons h and the penalty factor C in the ELM are optimized through the genetic algorithm. At this point, the fault diagnosis is completed.

[0085]

[0086] Sallen-Key circuit fault diagnosis and analysis:

[0087] Taking the Sallen-Key bandpass filter circuit fault diagnosis model as an example, this paper explains in detail how this method can achieve fault location and fault diagnosis. After the model is established, other experiments are carried out to verify the correctness of the model.

[0088] The simulation experiment was conducted on a personal computer with an Intel(R)Core(TM)i5-4210M CPU, using PSPICE16.6 for simulation analysis and obtaining simulation data sets, and MATLAB 2022a for data processing. First, the test signal source is set for the simulation circuit. The Sallen-key experiment selects a square wave signal source with a frequency of 5kHz, a duty cycle of 50%, and an amplitude of 1V. The transient analysis duration is 0.4ms and the sampling rate is 250kHz. The resistance tolerance in the circuit is 5%, the capacitance tolerance is 10%, and when the actual value of a component deviates by 30% of the nominal value, it can be determined that the component has failed. The Sallen-key circuit is as follows Figure 5 shown.

[0089] To conduct a simulation experiment, the steps are as follows:

[0090] 1) After drawing the circuit diagram in OrCAD Pspice, sensitivity analysis shows that R2, R3, R4, R5, C1, and C2 in the Sallen-key filter circuit are likely to affect the circuit output. They are set as diagnostic devices and soft faults are set. The fault settings are shown in Table 1, where NF means there is no fault in the circuit under test.

[0091] Table 1 Sallen-key circuit failure mode

[0092]

[0093] 2) Simulation parameter setting and data set division:

[0094] A total of 1300 raw data samples, i.e., circuit output responses, are generated by performing 100 Monte Carlo analyses on each fault category in Table 1, of which 910 samples are divided into training set responses and the remaining 390 samples are divided into test set responses;

[0095] 3) Wavelet scattering transform:

[0096] The training set response passes through the wavelet scattering network to obtain the wavelet scattering coefficient matrix of the three scattering layers. The scattering coefficient matrix consists of 6 scattering channels, each channel contains 32 scattering coefficients, and the three scattering layers contain channel 1, channel 2-4, channel 5-6, respectively. Each channel represents a frequency band coefficient with a bandwidth of 34.064kHz.

[0097] 4) Frequency band subset division:

[0098] The wavelet scattering coefficient matrix of all samples is divided according to the scattering channels. The scattering coefficient matrix consists of 6 scattering channels and is therefore divided into 6 frequency band subsets.

[0099] 5) Enhanced fault characteristics:

[0100] According to the Fisher discriminant criterion, the mapping matrix is ​​obtained by maximizing the criterion function, so that the similarity of samples of the same fault category increases after mapping, and the difference between different categories increases, thereby achieving the effect of enhancing fault characteristics;

[0101] 6) Feature Fusion:

[0102] Based on the p enhanced frequency band subsets, 30% of the pre-test set is divided, and the remaining training set is used as the pre-training set. The classification accuracy of the p ELMs for each fault category is obtained through pre-training and used as their respective category weights. The p ELM outputs are summed with the corresponding category weights, and the product sum is used as the input of the classification ELM. In order to compare the effect of fault feature enhancement, the category weights of each frequency band subset before and after enhancement are compared. The results before and after enhancement are shown in Tables 2 and 3. For easy observation, the mean value of the weight of each fault category under the frequency band subset (category mean), the weight mean of each frequency band subset under the fault category (subset mean) and the overall mean (lower right corner of the table) are added to the table respectively.

[0103] Table 2 Pre-enhancement band subset category weights

[0104]

[0105] Table 3 Enhanced band subset category weights

[0106]

[0107] 7) Fault diagnosis:

[0108] The fused results are sent to the classification ELM, and the number of hidden layer neurons and penalty factors in the model are optimized by combining the classic genetic algorithm, and finally the diagnosis results are obtained.

[0109] Four-op-amp dual-secondary circuit fault diagnosis and analysis:

[0110] The four-op-amp dual-secondary circuit selects a square wave signal source with a frequency of 2kHz, a duty cycle of 50%, and an amplitude of 1V. The fault offset and parameter tolerance settings are the same as those of the Sallen-key circuit. The four-op-amp dual-secondary circuit is as follows: Figure 6 The fault settings are shown in Table 4.

[0111] Table 4 Fault modes of four op amp dual secondary circuit

[0112]

[0113] The experimental results of this method are compared with those of other methods and are shown in Table 5.

[0114] Table 5 Comparison of diagnostic results

[0115]

[0116] This method uses a wavelet scattering network to extract response features from the time-frequency domain, and through the fusion processing of multiple frequency band subsets, it fully utilizes the effective features in each frequency band, and the diagnosis results are better than other literature methods. The fault feature enhancement process has been described in the previous article. After enhancing the fault features, the overall accuracy rate increased by 2.82% and 2.56% respectively, and the enhancement effect is obvious.

[0117] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for fault diagnosis of analog circuits based on wavelet scattering and ensemble learning, characterized in that: The steps include: Simulation data collection phase: Step 1: determine the circuit under test, perform sensitivity analysis on the circuit under test in the simulation software, and determine the diagnosable components in the circuit under test; Step 2: Set soft faults for each diagnosable device, and perform 100 Monte Carlo analyses on each fault type under the conditions of 5% resistance tolerance and 10% capacitance tolerance, and obtain 100 simulation samples for each fault type. The samples collected by simulation are used as the output response of the circuit under test, thereby obtaining a simulation data set; Frequency band subset extraction stage: Step 3: Divide all collected output responses into training set responses and test set responses in a ratio of 7:3 or 6:

4. Perform wavelet scattering transform on each sample in the training set response to obtain the scattering coefficient matrix of each sample. When verifying the final diagnostic effect, the test set response also uses wavelet scattering transform to obtain the scattering coefficient matrix of each test set response sample. Step 4: split the scattering coefficient matrix of each sample in the training set response into multiple one-dimensional samples, and reassemble these one-dimensional samples into the same number of subsets according to the number of scattering channels, each subset is composed of convolution coefficients generated by the same scattering channel. When verifying the final diagnostic effect, the scattering coefficient matrix of the test set response is also divided into frequency band subsets using the method in this step; Fault feature enhancement stage: Step 5: Calculate the global scatter matrix, intra-class scatter matrix, and inter-class scatter matrix of each frequency band subset, and maximize the Fisher discriminant criterion to obtain the mapping matrix of each frequency band subset. Project the frequency band subset into a multidimensional space by multiplying the mapping matrix and the frequency band subset, so that the samples of the same fault category after mapping are clustered as much as possible, and the samples of different fault categories are separated as much as possible, thereby realizing the fault feature enhancement of each frequency band subset; Feature fusion and fault diagnosis stage: Step 6: The enhanced frequency band sub-level is divided into a pre-training set and a pre-test set. The pre-training extreme learning machine in the Bagging sub-class extreme learning machine is trained by the pre-training set. The pre-test set is used to classify and verify the pre-training result. The verification result consists of the classification accuracy of each fault category, and the verification result is used as the category weight of the enhanced frequency band subset, thereby obtaining the category weight of each enhanced frequency band subset; Step seven, multiply the output of the pre-trained extreme learning machine output layer in the Bagging subclass extreme learning machine by the corresponding category weight to achieve feature fusion, and send the result of feature fusion to the classification extreme learning machine in the Bagging subclass extreme learning machine for training. During the training process, a genetic algorithm is used to optimize the parameters of the extreme learning machine, and the classification accuracy of the test set response is used as the final diagnostic effect to complete the fault diagnosis.

2. The analog circuit fault diagnosis method based on wavelet scattering and ensemble learning according to claim 1 is characterized in that: In step 2, the components selected in step 1 are limited to resistors and capacitors, and there are differences in output responses under fault conditions and non-fault conditions.

3. The analog circuit fault diagnosis method based on wavelet scattering and ensemble learning according to claim 1 is characterized in that: In step 3, the training set responses divided by the circuit output responses are transformed by wavelet scattering to obtain the scattering coefficient matrix of each training set response sample. The scattering coefficient matrix is ​​composed of the convolution results of the scattering channels. The pth channel of the wavelet scattering network is shown in formula (19): S[j1,j2,…,j p ]x=U[j1,j2,...,j p ]x*φ (19) Where j represents the scale coefficient, φ is the low-pass filter, ψ is the high-frequency wavelet, x is the training set response divided by the circuit output response, and the high-frequency part is expressed as: j1<Λ<j p (21).

4. The analog circuit fault diagnosis method based on wavelet scattering and ensemble learning according to claim 1 is characterized in that: In step 4, the scattering coefficient matrix obtained in step 4 is split into multiple one-dimensional samples. The scattering coefficient matrix is ​​composed of p scattering channels and can be divided into p frequency band subsets. The g-th frequency band subset is represented by s g , 1≤g≤p, realizing the division of frequency band subsets.

5. The analog circuit fault diagnosis method based on wavelet scattering and ensemble learning according to claim 1 is characterized in that: In step 5, the global divergence matrix S of each frequency band subset is calculated gt , intra-class scatter matrix S gw and the inter-class scatter matrix S gb , calculated as follows: S gt =(s g -d g )(s g -d g ) T (22) where d g represents the sample mean of the g-th frequency band subset, l represents the number of fault categories, s gi represents the samples belonging to fault category i in the g-th frequency band subset, 1≤i≤l, d gi It represents the mean value of samples belonging to fault category i in the p-th frequency band subset, m i represents the number of training samples belonging to fault category i in the frequency band subset, and the mapping matrix w g It is composed of the closed-form solution of maximizing the Fisher criterion function. The mapping matrix and the band subset are projected into the multidimensional space by the product method. The Fisher criterion function maximized by the g-th band subset is shown in formula (25). The enhanced g-th band subset is represented by s g ′ indicates that s g ′=w g s g , 6. The analog circuit fault diagnosis method based on wavelet scattering and ensemble learning according to claim 1 is characterized in that: In step six, s g ′ In the pre-training stage, the pre-training set and the pre-test set are divided into two parts according to the ratio of 7:

3. The pre-test set is used to classify and verify the pre-training results. The verification results are composed of the classification accuracy of each fault category, and the verification results are used as the category weights of the enhanced frequency band subsets. Thus, the category weights z of each enhanced frequency band subset are obtained. g .

7. The analog circuit fault diagnosis method based on wavelet scattering and ensemble learning according to claim 1 is characterized in that: In step 7, the output layer of the pre-trained extreme learning machine outputs Y g The sum of the products of the corresponding output weights is used as the input X of the classification extreme learning machine in the Bagging subclass extreme learning machine, as shown in formula (26). The feature fusion is completed here, and finally the final diagnosis is completed according to the principle of ELM.

8. The analog circuit fault diagnosis method based on wavelet scattering and ensemble learning described in any one of claims 1 to 7 is used for Sallen_Key bandpass filter or four-op-amp dual-quadratic circuit fault diagnosis.

Images