Analog circuit fault diagnosis method based on improved TQWT and DBN

Abstract

The invention discloses an analog circuit fault diagnosis method based on improved tunable orthogonal wavelet transform (TQWT) and a deep belief network (DBN), and aims to improve the efficiency and accuracy of fault diagnosis. The method comprises the following steps of: 1, refining circuit signal data by adopting a TQWT algorithm, and more effectively capturing the local characteristics of the signal by utilizing the characteristics of the TQWT, especially when the complex circuit signal with larger noise interference is processed; in order to optimize the performance of the TQWT algorithm, a cuckoo search algorithm is adopted to dynamically adjust the Q factor of the TQWT, and it is ensured that the optimal Q factor is selected in a specific application scene, so that the fault feature separation capability is improved. And after the data are refined, inputting the refined data into an Inception module to carry out feature extraction. The Inception module can effectively capture key features in circuit signals due to unique parallel processing capability and multi-scale feature extraction advantages, so that the recognition capability of the model on various fault modes is remarkably enhanced. After feature extraction is completed, a multi-core support vector machine (MKL-SVM) is finally constructed for fault classification. The MKL-SVM can improve the accuracy and robustness of the classification model by integrating a plurality of kernel functions, and can adapt to different types of fault classification requirements.

Description

Technical Field

[0001] The invention belongs to the field of analog circuit fault diagnosis, and in particular relates to an analog circuit fault diagnosis method based on improved TQWT and DBN network. Background Art

[0002] Electronic circuit systems are widely used in aerospace, military and industrial control, and the stability of circuit systems is crucial. Circuits can be divided into analog circuits and digital circuits: the failure frequency of digital circuits is very small, and the fault diagnosis technology has been established, so the stability of the circuit system mainly depends on the stability of the analog circuit. Analog circuit faults are divided into hard faults and soft faults. Hard faults can be attributed to circuit short circuits or open circuits. Soft faults refer to analog circuit components that are affected by the surrounding environment (such as temperature, humidity or pressure), and their parameters are outside the allowable range. In this case, the function of the components is not completely lost, and they can still work for a period of time. The parameter range of soft faults usually appears symmetrically on both sides of the nominal value. After the soft fault occurs, if it is not handled in time, it will turn into a hard fault. Therefore, soft fault diagnosis of analog circuits is one of the effective means to ensure the normal operation of analog circuits.

[0003] When performing analog circuit fault diagnosis, the electronic circuits encountered often have multi-level structures or are packaged in a module, which leads to a very limited number of nodes where voltage can be directly measured during the troubleshooting process. Therefore, the available test data is very limited, which means that the amount of information available for fault diagnosis is insufficient, resulting in uncertainty and ambiguity in fault location. On the one hand, the problem of insufficient information must be solved in the process of fault diagnosis. Due to the limited number of measurable nodes, the fault diagnosis algorithm needs to rely on limited available information to infer the state of the entire circuit. This may lead to misjudgment or omission of certain potential faults. Therefore, how to extract as much fault information as possible from limited measurement data is an important research direction. On the other hand, the problem of nonlinearity of analog circuits also increases the difficulty of fault diagnosis. Nonlinearity means that when processing signals, the relationship between the input and output of the circuit is nonlinear. These problems complicate fault location because even small deviations can lead to significant changes in circuit performance. To address these two aspects, research work needs to explore more advanced algorithms and models. For example, by training models to identify and classify fault modes, artificial intelligence and machine learning techniques are used to deal with nonlinear problems and tolerance issues. In addition, integrating more sensors and measurement technologies is an effective way to improve the accuracy of fault diagnosis.

[0004] In recent years, data-driven fault diagnosis methods have developed rapidly. The method of machine learning and deep learning is used to establish a data-driven fault diagnosis model for fault diagnosis. This method has high accuracy and does not require the establishment of complex physical models. It has become one of the popular methods in the field of analog circuit fault diagnosis. The analog circuit fault diagnosis method based on improved TQWT and DBN network proposed in the present invention is a data-driven method. Summary of the invention

[0005] In view of the problem that analog circuits in electronic systems are prone to failures that may cause the entire device to fail, the present invention provides a fault diagnosis method for locating analog circuit component faults. Feature extraction is performed through voltage data of analog circuit test points, which is input into a classifier for fault classification, thereby achieving component diagnosis of analog circuit soft faults.

[0006] The present invention discloses an analog circuit fault diagnosis method based on improved TQWT and DBN network, which specifically includes: Step 1: Use Multisim 14.0 to simulate the circuit and form the initial data set with the test point data; Step 2: Using an improved Tunable Q-factor Wavelet Transform (TQWT) algorithm to process the analog circuit data in step 1, forming a multi-branch data set for each signal; Step 3: Use the principal component analysis (PCA) method to extract features from the multi-branch data in step 2 to obtain multi-scale features of the signal; Step 4: After dividing the multi-scale data set obtained in step 3, the training set is input into the deep belief network (DBN), the fault diagnosis network is trained, and a mass spectrometer fault prediction model is established; Step 5: Input the test set into the fault prediction model based on the DBN network established in step 4 to obtain and output the fault diagnosis result.

[0007] Furthermore, the method used to solve the problem of small amount of data in the process of constructing the initial data set in step 1 is Monte Carlo analysis.

[0008] Furthermore, the method for improving TQWT in step 2 is Cuckoo Search Algorithm (CSA). CSA is used to optimize the Q factor of TQWT, and the specific steps are as follows: (1) Population initialization: randomly generate the initial positions of m bird nests ; (2) Calculate the fitness function value of each bird's nest and record the optimal solution with the largest current fitness function value. The value of the fitness function here is: fitness\, =\, 1 / \sqrt {\frac {1} {n}\sum ^{n}_{i=1} {\left [ {{f'}_{i}-{f}_{i}} \right ]^{2}}} In the formula, is the function value at a certain point, for The value after TQWT and inverse TQWT transformation, n is the number of acquisition points of each data in the signal data set; (3) Use the cuckoo nesting location update formula to update the bird's nest location. The formula is as follows: In the formula, is the nest position of the ith nest in the tth generation, is the dot product symbol, is the step size of the cuckoo position update, is a randomly generated search path whose step length follows the following Levy distribution: Levy~u=t -λ (1<λ≤3) (4) Compare the fitness function values ​​of the current bird nest and the previous generation of bird nest locations, and retain the location with a higher fitness function value; (5) Generate a random number r∈[0, 1] that follows a uniform distribution and compare it with the probability Pa of the host bird discovering an alien egg: If r>Pa, the position of the bird's nest is randomly updated; if r≤Pa, the position of the bird's nest remains unchanged; (6) If the end condition is not met, return to (2); (7) Output the global optimal position.

[0009] Furthermore, for the data set processed using the improved TQWT in step 2, the processing process is divided into two parts: data decomposition and single-branch reconstruction: Data decomposition stage: The CSA-improved TQWT algorithm will adaptively select the optimal Q value to decompose the signal according to the signal characteristics, and obtain J+1 branch signals. J is the number of layers of TQWT decomposition of the signal. The calculation formula is as follows: J=\frac {lg\left [ {N / 4\left ( {Q+1} \right )} \right ]} {lg\left ( {Q+1} \right ) / \left ( {Q+1-2 / r} \right )} Where, J is rounded to 0, and N is the length of the input signal of TQWT; Single-branch reconstruction stage: The CSA-improved TQWT algorithm performs single-branch reconstruction on each branch data; single-branch reconstruction is to set the values ​​of the remaining branches except the target branch to 0 and perform inverse TQWT transformation.

[0010] Furthermore, for the PCA technology described in step 3, the main idea of ​​feature extraction is dimensionality reduction. Assuming there are m pieces of n-dimensional data, the specific steps for feature extraction are as follows: (1) Organize the original data into an n-row and m-column matrix X; (2) Transform the matrix X and perform zero-mean processing on each row of X. Zero-mean processing means subtracting the average value of each row from the value of the row. (3) Find the covariance matrix C of matrix X. The calculation formula is: (4) Find the eigenvalues ​​and eigenvectors of the covariance matrix C; (5) Arrange the eigenvectors in rows from top to bottom according to their eigenvalues, and take the first k rows to form the matrix P, where k is the dimension of the data after dimensionality reduction; (6) The final k-dimensional data after dimensionality reduction is: .

[0011] Furthermore, for the DBN-based fault diagnosis network described in step three, a DBN network with three hidden layers is used, that is, its structure is composed of three restricted Boltzmann machines.

[0012] Beneficial effect: The present invention is based on the TQWT algorithm and the DBN network, uses the voltage data of the analog circuit test points, and performs fault location and prediction based on the signal time-frequency detail characteristics, thereby realizing the diagnosis and location of analog circuit faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 is a flow chart of the method of the present invention; Figure 2 Schematic diagram of signal decomposition and reconstruction of TQWT algorithm; Figure 3 It is the structural diagram of DBN network; Figure 4is the confusion matrix for fault diagnosis. Specific implementation methods

[0014] The present invention will be further described in detail below with reference to the accompanying drawings and specific examples. 1. Fault diagnosis method like Figure 1 As shown, Step 1: Use Multisim 14.0 to simulate the circuit and form the initial data set with the test point data; Step 2: Using an improved Tunable Q-factor Wavelet Transform (TQWT) algorithm to process the analog circuit data in step 1, forming a multi-branch data set for each signal; Step 3: Use the principal component analysis (PCA) method to extract features from the multi-branch data in step 2 to obtain multi-scale features of the signal; Step 4: After dividing the multi-scale data set obtained in step 3, the training set is input into the deep belief network (DBN), the fault diagnosis network is trained, and a mass spectrometer fault prediction model is established; Step 5: Input the test set into the fault prediction model based on the DBN network established in step 4 to obtain and output the fault diagnosis result.

[0015] Furthermore, the method used to solve the problem of small amount of data in the process of constructing the initial data set in step 1 is Monte Carlo analysis.

[0016] Furthermore, the method for improving TQWT in step 2 is Cuckoo Search Algorithm (CSA). CSA is used to optimize the Q factor of TQWT, and the specific steps are as follows: (1) Population initialization: randomly generate the initial positions of m bird nests ; (2) Calculate the fitness function value of each bird's nest and record the optimal solution with the largest current fitness function value. The value of the fitness function here is: fitness\, =\, 1 / \sqrt {\frac {1} {n}\sum ^{n}_{i=1} {\left [ {{f'}_{i}-{f}_{i}} \right ]^{2}}} In the formula, is the function value at a certain point, for The value after TQWT and inverse TQWT transformation, n is the number of acquisition points of each data in the signal data set; (3) Use the cuckoo nesting location update formula to update the bird's nest location. The formula is as follows: In the formula, is the nest position of the ith nest in the tth generation, is the dot product symbol, is the step size of the cuckoo position update, is a randomly generated search path whose step length follows the following Levy distribution: Levy~u=t -λ (1<λ≤3) (4) Compare the fitness function values ​​of the current bird nest and the previous generation of bird nest locations, and retain the location with a higher fitness function value; (5) Generate a random number r∈[0, 1] that follows a uniform distribution and compare it with the probability Pa of the host bird discovering an alien egg: If r>Pa, the position of the bird's nest is randomly updated; if r≤Pa, the position of the bird's nest remains unchanged; (6) If the end condition is not met, return to (2); (7) Output the global optimal position.

[0017] Furthermore, for the data set processed using the improved TQWT in step 2, the processing process is divided into two parts: data decomposition and single-branch reconstruction: Data decomposition stage: The CSA-improved TQWT algorithm will adaptively select the optimal Q value to decompose the signal according to the signal characteristics, and obtain J+1 branch signals. J is the number of layers of TQWT decomposition of the signal. The calculation formula is as follows: J=\frac {lg\left [ {N / 4\left ( {Q+1} \right )} \right ]} {lg\left ( {Q+1} \right ) / \left ( {Q+1-2 / r} \right )} Where, J is rounded to 0, and N is the length of the input signal of TQWT; Single-branch reconstruction stage: The CSA-improved TQWT algorithm performs single-branch reconstruction on each branch data; single-branch reconstruction is to set the values ​​of the remaining branches except the target branch to 0 and perform inverse TQWT transformation.

[0018] Furthermore, for the PCA technology described in step 3, the main idea of ​​feature extraction is dimensionality reduction. Assuming there are m pieces of n-dimensional data, the specific steps for feature extraction are as follows: (1) Organize the original data into an n-row and m-column matrix X; (2) Transform the matrix X and perform zero-mean processing on each row of X. Zero-mean processing means subtracting the average value of each row from the value of the row. (3) Find the covariance matrix C of matrix X. The calculation formula is: (4) Find the eigenvalues ​​and eigenvectors of the covariance matrix C; (5) Arrange the eigenvectors in rows from top to bottom according to their eigenvalues, and take the first k rows to form the matrix P, where k is the dimension of the data after dimensionality reduction; (6) The final k-dimensional data after dimensionality reduction is: .

[0019] Furthermore, for the DBN-based fault diagnosis network described in step three, a DBN network with three hidden layers is used, that is, its structure is composed of three restricted Boltzmann machines. Specific implementation steps

[0020] The present invention is further described in detail below through specific implementations in conjunction with the accompanying drawings. The present application can be implemented in a variety of different forms and is not limited to the implementation described in this embodiment. The purpose of providing the following specific implementations is to facilitate a clearer and more thorough understanding of the content of the present invention.

[0021] Reference Figure 1 As shown, the flowchart of the analog circuit fault diagnosis method of the present invention is completed by 6 steps. Step 1 uses Multisim 14.0 to perform circuit simulation on the test circuit, and the circuits used are Sallen-key bandpass filter circuit and four-op-amp dual second-order high-pass filter circuit to collect the test point data of the analog circuit, and the test point voltage data constitute an initial data set; Step 2 uses the adjustable Q factor wavelet transform algorithm optimized by the cuckoo search algorithm to decompose and reconstruct the initial data set to form a multi-branch data set for each signal; Step 3 uses the principal component analysis method to extract features from the multi-branch data of step 2 to obtain multi-scale features of the signal; Step 4 inputs the training set into the deep belief network after dividing the feature data set, trains the fault diagnosis network, and establishes a mass spectrometer fault prediction model; Step 5 inputs the test set into the fault prediction model based on the DBN network established in step 4, and obtains and outputs the fault diagnosis result.

[0022] Step 1: The voltage data of the output point is used as the initial data set. In order to solve the problem of insufficient data volume, Monte Carlo analysis is used to generate 200 sets of data for each fault. The fault types of the Sallen-key bandpass filter circuit and the four-op-amp biquad high-pass filter circuit are shown in Table 1 and Table 2: Table 1 Fault types of Sallen-key bandpass filter circuit Fault type Fault Tags Nominal value Fault scope C1↑ 1 5nF [6nF,7.5nF] C1↓ 2 5nF [2.5nF,4nF] C2↑ 3 5nF [6nF,7.5nF] C2↓ 4 5nF [2.5nF,4nF] R2↑ 5 3KΩ [3.6 KΩ,4.5 KΩ] R2↓ 6 3KΩ [1.5 KΩ,2.4 KΩ] R3↑ 7 2KΩ [2.4 KΩ,3 KΩ] R3↓ 8 2KΩ [1 KΩ,1.6 KΩ] R4↑ 9 4KΩ [4.8 KΩ,6 KΩ] R4↓ 10 4KΩ [2 KΩ,3.2 KΩ] R5↑ 11 4KΩ [4.8 KΩ,6 KΩ] R5↓ 12 4KΩ [2 KΩ,3.2 KΩ] NF 13 / / Table 2 Fault types of four-op-amp biquad high-pass filter circuit Fault type Fault Tags Nominal value Fault scope C1↑ 1 5nF [6nF,7.5nF] C1↓ 2 5nF [2.5nF,4nF] C2↑ 3 5nF [6nF,7.5nF] C2↓ 4 5nF [2.5nF,4nF] R1↑ 5 6.2KΩ [7.44 KΩ,9.3 KΩ] R1↓ 6 6.2KΩ [3.1 KΩ,4.96 KΩ] R2↑ 7 6.2KΩ [7.44 KΩ,9.3 KΩ] R2↓ 8 6.2KΩ [3.1 KΩ,4.96 KΩ] R3↑ 9 6.2KΩ [7.44 KΩ,9.3 KΩ] R3↓ 10 6.2KΩ [3.1 KΩ,4.96 KΩ] NF 11 / / In step 2, CSA is used to optimize the Q factor of TQWT, and the CSA-improved TQWT algorithm is used for data decomposition and single-branch reconstruction. The J of the two experimental circuits are 10 and 14, and the Q are 1.37 and 3.27, respectively.

[0023] In step 3, in order to reduce the redundancy between data, the principal component analysis method is used to reduce the dimension of the data to form a time-frequency domain feature data set.

[0024] Input the time-frequency domain feature data set in step 3 into the DBN network to obtain the fault diagnosis result. Figure 4 As shown, the accuracy of the Sallen-key bandpass filter circuit and the four-op-amp biquad high-pass filter circuit are 100% and 98.92% respectively.

Claims

1. The analog circuit fault diagnosis method based on improved TQWT and DBN network is characterized by: It is achieved by the following steps: Step 1: Use Multisim 14.0 to simulate the circuit and form the initial data set with the test point data; Step 2: Using an improved Tunable Q-factor Wavelet Transform (TQWT) algorithm to process the analog circuit data in step 1, forming a multi-branch data set for each signal; Step 3: Use the principal component analysis (PCA) method to extract features from the multi-branch data in step 2 to obtain multi-scale features of the signal; Step 4: After dividing the multi-scale data set obtained in step 3, the training set is input into the Deep Belief Network (DBN), the fault diagnosis network is trained, and a mass spectrometer fault prediction model is established; Step 5: Input the test set into the fault prediction model based on the DBN network established in step 4 to obtain and output the fault diagnosis result.

2. The analog circuit fault diagnosis method based on improved TQWT and DBN network according to claim 1 is characterized in that: The method used to solve the problem of small data volume in the process of constructing the initial data set in step 1 is Monte Carlo analysis.

3. The analog circuit fault diagnosis method based on improved TQWT and DBN network according to claim 1 is characterized in that: The method for improving TQWT in step 2 is Cuckoo Search Algorithm (CSA). CSA is used to optimize the Q factor of TQWT. The specific steps are as follows: (1) Population initialization: randomly generate the initial positions of m bird nests ; (2) Calculate the fitness function value of each bird's nest and record the optimal solution with the largest current fitness function value. The value of the fitness function here is: In the formula, is the function value at a certain point, for The value after TQWT and inverse TQWT transformation, n is the number of acquisition points of each data in the signal data set; (3) Use the cuckoo nesting location update formula to update the bird's nest location. The formula is as follows: In the formula, is the nest position of the ith nest in the tth generation, is the dot product symbol, is the step size of the cuckoo position update, is a randomly generated search path whose step length follows the following Levy distribution: Levy~u=t -λ (1<λ≤3) (4) Compare the fitness function values ​​of the current bird nest and the previous generation of bird nest locations, and retain the location with a higher fitness function value; (5) Generate a random number r∈[0, 1] that follows a uniform distribution and compare it with the probability Pa of the host bird discovering an alien egg: If r>Pa, the position of the bird's nest is randomly updated; if r≤Pa, the position of the bird's nest remains unchanged; (6) If the end condition is not met, return to (2); (7) Output the global optimal position.

4. The analog circuit fault diagnosis method based on improved TQWT and DBN network according to claim 1 is characterized in that: For the data set processing using the improved TQWT described in step 2, the processing process is divided into two parts: data decomposition and single-branch reconstruction: Data decomposition stage: The CSA-improved TQWT algorithm will adaptively select the optimal Q value to decompose the signal according to the signal characteristics, and obtain J+1 branch signals. J is the number of layers of TQWT decomposition of the signal. The calculation formula is as follows: Where, J is rounded to 0, and N is the length of the input signal of TQWT; Single-branch reconstruction stage: The CSA-improved TQWT algorithm performs single-branch reconstruction on each branch data; single-branch reconstruction is to set the values ​​of the remaining branches except the target branch to 0 and perform inverse TQWT transformation.

5. The analog circuit fault diagnosis method based on improved TQWT and DBN network according to claim 1 is characterized in that: For the PCA technology described in step 3, the main idea of ​​feature extraction is dimensionality reduction. Assuming there are m pieces of n-dimensional data, the specific steps for feature extraction are as follows: (1) Organize the original data into an n-row and m-column matrix X; (2) Transform the matrix X and perform zero-mean processing on each row of X. Zero-mean processing means subtracting the average value of each row from the value of the row. (3) Find the covariance matrix C of matrix X. The calculation formula is: (4) Find the eigenvalues ​​and eigenvectors of the covariance matrix C; (5) Arrange the eigenvectors in rows from top to bottom according to their eigenvalues, and take the first k rows to form the matrix P, where k is the dimension of the data after dimensionality reduction; (6) The final k-dimensional data after dimensionality reduction is:

6. The analog circuit fault diagnosis method based on improved TQWT and DBN network according to claim 1 is characterized in that: For the DBN-based fault diagnosis network described in step three, a DBN network with three hidden layers is used, that is, its structure is composed of three restricted Boltzmann machines.

Images