Simulation circuit fault signal processing method based on empirical wavelet transform decomposition

Abstract

The invention discloses an analog circuit fault signal processing method based on empirical wavelet transform decomposition, and belongs to the field of analog circuit fault signal processing. An existing analog circuit signal has the problems of nonlinearity and serious fault feature confusion. An analog circuit fault signal processing method based on empirical wavelet transform decomposition comprises the following steps: carrying out adaptive wavelet transform decomposition on a circuit fault signal, calculating an initial demarcation point obtained by a scale space method to obtain a mutual information value of a component, and redividing a frequency spectrum to obtain a new decomposition signal component; performing feature information extraction on the signal reconstructed by selecting the optimal component by using the kurtosis index; and the characteristic parameters are classified by using K-means clustering analysis, so that the diagnosis of the soft fault of the analog circuit is realized. The method reduces the calculation complexity of the whole algorithm, improves the efficiency of the algorithm, and improves the recognition accuracy.

Description

Technical Field

[0001] The invention relates to a method for processing analog circuit fault signals, and in particular to a method for processing analog circuit fault signals based on empirical wavelet transform decomposition. Background Art

[0002] Since the 1960s, analog circuit fault diagnosis has been widely concerned by the academic community. After half a century of research and exploration, it has gradually developed into a relatively complete theoretical system, and together with network analysis and network synthesis, it has formed the main body of network theory. With the rapid development of science and technology, there are more and more electronic devices in daily life. The circuit complexity and integration scale of these electronic devices are increasing, making the difficulty of equipment system testing and diagnosis continue to grow. Research on fault diagnosis technology of electronic systems and equipment, especially research on fault diagnosis of analog circuits and complex electronic systems, is an effective way to break through the bottleneck of equipment testing and diagnosis. During the long-term operation of circuit systems and equipment, due to factors such as electromagnetic interference, component aging and manufacturing errors, it is likely to have incalculable faults, especially in miniaturized and ultra-large-scale integrated circuits. Any minor fault may lead to system performance degradation, shutdown, and even irreparable property losses and casualties. Therefore, research on fault diagnosis theory and methods for electronic systems and complex equipment is a common demand in the current electronic industry and social economy.

[0003] According to relevant data, the proportion of mixed-signal circuits in electronic devices has exceeded 60%, while the proportion of analog circuits is less than 20%, but the probability of analog circuit failure is as high as 80%, far exceeding the probability of digital circuit failure. In recent years, the development of electronic technology and semiconductor technology, as well as the progress of chip manufacturing technology, have promoted the continuous increase in the integration of digital and analog circuits. With the continuous growth of mixed-signal circuits, it not only brings about the improvement of electronic product performance, but also high circuit production and diagnosis costs. In mixed-signal integrated circuits, analog circuits only account for 5% of the integrated chip area, but the diagnosis time can account for 80% to 90% of the entire chip test time, and the diagnosis cost is as high as 95% of the total test cost. Therefore, it is urgent to study and improve the theory and technology of analog circuit fault diagnosis.

[0004] Analog circuit failure modes mainly include soft failures (parametric failures) and hard failures (catastrophic failures). Structural damage caused by extreme changes such as short circuits or open circuits of certain components in the circuit network is usually called hard failures. Once such failures occur, the system will be completely paralyzed or fail, and the losses caused will be irreversible. Soft failures refer to the deviation of circuit component parameters from their nominal values ​​due to environmental influences or increased working time. Although in most cases the circuit system will not fail completely, it will cause system performance to deteriorate. If not handled in time, it will evolve into a hard failure.

[0005] The main task of soft fault diagnosis of analog circuits is to determine the physical location and parameters of faulty components when the topology of the network, the input excitation signal and the response under the fault are known. There are four main steps for analog circuit fault testing and diagnosis, namely fault detection, fault location, fault identification, and fault isolation. Fault detection mainly uses technical means to determine whether a circuit fault has occurred, that is, it can distinguish between a fault-free circuit condition and a faulty circuit condition; the method of confirming the location of the fault in the circuit is fault location; the type of fault is identified by fault identification; and finally, technical means are used to isolate the fault. After decades of development, although many important results have been achieved in the testing and fault diagnosis of analog circuits, the research on circuit diagnostic methods, especially the research on soft fault diagnosis of analog circuits, is still relatively backward. The main difficulties are:

[0006] (1) Lack of good fault models. Since the input stimulus of the circuit under test and the output response of the port are continuous quantities, and the parameters of each component are also continuous quantities, a component may have infinite faults, resulting in the failure mode of the entire circuit being uncountable, and it is impossible to construct a "dictionary" to check all faults.

[0007] (2) Since component parameters often have tolerances, i.e. parameter values ​​are allowed to vary around nominal values, the tolerances of many components may be equivalent to a “major fault” of one component, making it impossible to accurately determine the physical location of the actual fault, making many diagnostic methods lose accuracy and stability. The existence of component parameter tolerances is the biggest challenge facing analog circuit fault diagnosis.

[0008] (3) Nonlinear problems are common in analog circuits. As the circuit scale increases linearly, the computation increases exponentially. The presence of a large number of feedback loops also increases the complexity of computation and testing.

[0009] (4) Difficulty and inaccuracy in measurement. The measurable nodes of analog circuits are limited. Since modern electronic circuits are usually multi-layered or packaged, the accessible nodes are limited, making the collected fault information incomplete. In addition, it is impossible to collect branch currents at accessible nodes because the branch needs to be disconnected first and then the measuring instrument is connected in series, which is impossible in actual measurement. In short, the output signal of analog circuits generally uses the voltage at limited measurable nodes, which results in insufficient information that can be used for fault diagnosis, resulting in uncertainty and ambiguity in fault location.

[0010] In summary, the soft fault diagnosis method of analog circuits is closely related to our lives. Research on analog circuit fault diagnosis not only helps to reduce equipment failure rate, improve maintenance rate and reduce economic losses, but also contributes to the development of electronic R&D technology and manufacturing industry. On a deeper level, the research on fault diagnosis methods of electronic systems and equipment is not only related to the sustained and healthy development of the national economy, but also one of the potential determinants of national security and social progress. Therefore, the research on intelligent fault diagnosis methods has become a research topic with important theoretical significance and practical use value. Summary of the invention

[0011] The purpose of the present invention is to solve the problem that the existing analog circuit signal has nonlinearity and the fault characteristics are seriously confused, and to propose an analog circuit fault signal processing method based on empirical wavelet transform decomposition.

[0012] The above purpose is achieved through the following technical solutions:

[0013] A method for processing analog circuit fault signals based on empirical wavelet transform decomposition is implemented by the following steps:

[0014] Step 1: Decompose the circuit fault signal by adaptive wavelet transform, then calculate the initial demarcation point obtained by scale space method to obtain the mutual information value of the component, so as to re-divide the spectrum to obtain new decomposed signal components;

[0015] Step 2: Extract feature information from the signal reconstructed by selecting the optimal component using the kurtosis index;

[0016] Step 3: Use K-means clustering analysis to classify the characteristic parameters, so as to realize the diagnosis of analog circuit soft faults.

[0017] Furthermore, the step of performing adaptive wavelet transform decomposition on the circuit fault signal in step 1 is specifically:

[0018] Normalize the Fourier axis of the original signal R(t) to [0, 2π], and limit the discussion range to [0, π] to follow Shannon's theorem; divide the Fourier support interval [0, π] into N continuous segments, each segment is represented by Λ n =[ω n-1 ,ω n ], where ω n-1 and ω n Respectively represent the boundaries of each segment, where n = 1, 2.....N, ω0 = 0, ω N =π, then we have

[0019] ω n Centered at 2τ n The gray area represents a transition zone, τ n =γω n τ, 0<γ<1; In order to ensure that there is no overlap between any two transition zones, γ is limited to less than within the scope, that is:

[0020]

[0021] Constructing wavelet function ψ based on the idea of ​​Littlewood-Paley and Meyer wavelets n (t) and the scaling function φ n (t), ψ n (t) and φ n The Fourier transform of (t) is expressed as ψ n (ω), φ n (ω), and satisfies:

[0022]

[0023] In the formula, β(x) selects the polynomial proposed by Daubechies:

[0024] β(x)=x 4 (35-84x+70x 2 -20x 3 )

[0025] The detail coefficient and approximation coefficient of EWT are defined as:

[0026]

[0027]

[0028] Among them, ψ n (t) represents the empirical wavelet function, φ1(t) represents the scaling function, and In turn, ψ n Fourier transform of φ(t) and φ1(t); is n The conjugate complex numbers of (t) and φ1(t), F(), F -1 () represents Fourier transform and inverse Fourier transform;

[0029] According to the above formula, the reconstructed signal of the original signal is expressed as:

[0030]

[0031] Where: * represents convolution operation; and Respectively and The Fourier transform of .

[0032] Furthermore, the step of calculating the initial demarcation point obtained by the scale space method in step 1 to obtain the mutual information value of the component, thereby re-dividing the spectrum to obtain new decomposed signal components, specifically, is:

[0033] First, design and improve the IEWT method:

[0034] In order to overcome the problem of too many spectrum divisions in wavelet transform, the present invention re-divides and merges the spectrum according to the mutual information value to reduce the number of frequency bands; the relative size of the mutual information value more accurately and directly expresses the correlation between two random variables than the correlation coefficient;

[0035] MI(X, Y) = H(Y) - H(Y|X)

[0036] Among them, H(Y) is the entropy of Y, and H(Y|X) is the conditional entropy of Y when X is known;

[0037] Second, design the IEWT signal decomposition steps:

[0038] First, perform FFT on the original signal to obtain the spectrum of the vibration signal; then, determine the initial frequency band boundary points according to the scale space method to obtain the initial spectrum boundary points; then, calculate the mutual information values ​​of the components obtained by spectrum division according to the initial dividing points, and then redefine the dividing points according to the relative sizes of the mutual information values; then, redivide the spectrum according to the newly determined dividing points to obtain new decomposed signal components; then, select the optimal component according to the kurtosis index to reconstruct the signal.

[0039] Furthermore, the step of extracting feature information from the signal reconstructed by selecting the optimal component using the kurtosis index in step 2 is specifically:

[0040] Use Multisim simulation software to perform transient analysis on the circuit, sample the output response, and obtain the original fault information; perform 50 Monte Carlo analyses on each type of fault, and randomly divide the samples under each fault mode into two parts, one part of the data is used for training, and the other part of the data is used for testing;

[0041] Perform Fourier transform on the fault signal to transform the time domain problem into the frequency domain, determine the initial frequency band boundary point according to the scale space method, and complete the adaptive frequency band division of the frequency domain signal;

[0042] The frequency band boundary points are reselected by analyzing the mutual information MI(X,Y) between each frequency band in the initial spectrum. The calculation formula of MI(X,Y) is as follows:

[0043] MI(X, Y) = H(Y) - H(Y|X)

[0044] Where H(Y) is the entropy of Y, and H(Y|X) is the conditional entropy of Y when X is known;

[0045] The original frequency bands are re-divided according to the average value of the mutual information to obtain new spectrum demarcation points, and the spectrum is re-divided; the original 15 frequency bands are merged into new 5 frequency bands;

[0046] According to the new spectrum demarcation point, the wavelet basis function is constructed to perform EWT decomposition on the signal to obtain the processed signal components and the reconstructed signal;

[0047] A one-dimensional vector consisting of wavelet packet entropy, root mean square, variance, frequency kurtosis, frequency skewness and energy is used to characterize the fault signal characteristics.

[0048] Furthermore, the step of classifying the characteristic parameters by using K-means cluster analysis in step 3 to realize the diagnosis of analog circuit soft faults is specifically:

[0049] Based on the convolutional neural network, an improved capsule network is introduced to diagnose analog circuit soft faults, and the network structure is designed, including convolution layer, pooling layer, squeeze excitation module, primary capsule layer and advanced capsule layer;

[0050] First, the extracted feature information is passed through the convolution layer, and a wide convolution kernel is used to increase the receptive field of the model and retain more feature information; then, the feature dimension is reduced through the maximum pooling layer; two convolution layers are used in the middle, and small convolution kernels are used to fully extract the underlying features; then, a squeeze excitation module is added to recalibrate the channel dimension features and pass the weighted feature information to the primary capsule layer; then, a batch normalization layer is used after each convolution layer in the feature extraction part; all convolution layers in the model use L2 regularization; the primary capsule layer first performs a convolution operation, and constructs the convolution result into a vector form as the input of the advanced capsule layer; the advanced capsule layer calculates the correlation between capsule layers through the self-attention routing algorithm to realize the classification of the extracted features, and obtains x output vectors, where x represents the number of single signal categories, and the value is 9; the output layer calculates the modulus length of the output vectors respectively, that is, the prediction result of the sample is obtained.

[0051] The beneficial effects of the present invention are:

[0052] 1) The fault feature information in the original output signal of the circuit is not obvious and is easily affected by environmental noise, making feature extraction difficult.

[0053] Proposed solution: For the original signal output by the faulty circuit, the present invention intends to use empirical wavelet transform for processing. The original signal is first decomposed into components of different scales using empirical wavelet transform, and the influence of noise on the signal is suppressed by retaining the main information and removing redundant information, thereby improving the quality and accuracy of the signal. EWT can analyze the local characteristics of the signal from multiple scales, and can well discover subtle fault characteristic information, thereby improving the accuracy of fault diagnosis.

[0054] In the signal processing process, the empirical wavelet transform is combined with the improved scale space method. First, the circuit fault signal is normalized, and the frequency band is initially divided by the scale space method. Then, the spectrum is re-divided according to the mutual information value, which effectively improves the problem of excessive spectrum division of the wavelet transform. The optimal signal component is selected according to the kurtosis index to reconstruct the signal; finally, the eigenvalues ​​in the reconstructed signal are calculated by signal processing to obtain the eigenvector, which is input into the neural network as a data set, providing a basis for subsequent fault diagnosis.

[0055] In feature extraction, a one-dimensional vector consisting of root mean square, energy, variance, frequency kurtosis, frequency skewness and wavelet packet entropy is used to characterize the fault signal characteristics. Wavelet packet entropy, root mean square, variance, frequency kurtosis, frequency skewness and energy are six commonly used features, and they have fewer dimensions, which can reduce the computational complexity and improve the efficiency of the algorithm. Wavelet packet entropy can reflect the complexity of the signal, root mean square and variance can reflect the amplitude change of the signal, frequency kurtosis and frequency skewness can reflect the frequency distribution characteristics of the signal, and energy can reflect the strength of the signal. These features combined can comprehensively reflect the characteristics of the signal and can identify different types of faults.

[0056] 2) When performing empirical wavelet decomposition on the signal, the steps of selecting EWT parameters and extracting local extreme points require prior knowledge and are easily affected by human factors. The present invention uses a scale space method to adaptively segment the spectrum, and realizes adaptive detection of the boundary of the segmented Fourier spectrum without prior knowledge.

[0057] 3) The core of IEWT decomposition is to reasonably divide the Fourier spectrum. It is necessary to first determine the dividing points of the frequency band. Too few dividing points will lead to component aliasing and data distortion; too many IMF decompositions will increase the amount of calculation and increase the processing time.

[0058] Proposed solution: Use mutual information value to re-divide the frequency band. Merge adjacent frequency bands whose mutual information value is greater than the average value into the same frequency band, and merge adjacent frequency bands whose mutual information value is less than the average value into one frequency band, which can effectively improve the problem of too many spectrum divisions of wavelet transform.

[0059] 4) Due to the simulation circuit and continuous testing, a large amount of fault data is obtained. When machine learning and artificial neural networks are used to train and classify the fault set, the training time is long and the efficiency is low. How to solve this problem?

[0060] Proposed solution: In order to solve the problem of long training time and low efficiency of artificial neural networks, the present invention adopts convolutional neural networks for classification. Traditional neural networks need to perform matrix multiplication on the neural layer to activate the neural layer. Since the kernel of the convolutional neural network is smaller than the input size, sparse connection can be achieved, thereby reducing the number of weight parameters required in the calculation process, reducing the calculation burden, and improving classification efficiency. The diagnostic method adopts a deep learning method. After the above steps, samples are obtained, input and trained into the convolutional neural network, the fault type is classified, the fault diagnosis accuracy is monitored in real time through simulation circuits, the neural network is continuously trained, the initial weight parameters are adjusted, and the neural network is optimized. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1It is a Fourier spectrum segmentation diagram involved in the present invention;

[0062] Figure 2 It is a signal decomposition flow chart of the IEWT algorithm involved in the present invention;

[0063] Figure 3 The present invention relates to an IEWT analog circuit fault signal diagnosis flow chart;

[0064] Figure 4 The present invention relates to a Sallen-key bandpass filter simulation circuit diagram;

[0065] Figure 5 It is a fault signal data diagram involved in the present invention;

[0066] Figure 6 It is a time domain diagram of a fault signal involved in the present invention;

[0067] Figure 7 The present invention relates to an adaptive frequency band division spectrum diagram;

[0068] Figure 8 is the component mutual information value involved in the present invention;

[0069] Fig. 9 The present invention relates to a processed frequency band division spectrum diagram;

[0070] Fig.10 It is the IEWT decomposition signal component involved in the present invention;

[0071] Fig.11 It is a time domain diagram of a reconstructed signal involved in the present invention;

[0072] Fig.12 The present invention relates to principal component analysis feature extraction;

[0073] Fig.13 The present invention relates to the K-means cluster analysis method feature grouping;

[0074] Fig.14 The present invention relates to a capsule network structure;

[0075] Fig.15 This is the accuracy curve involved in the present invention. DETAILED DESCRIPTION

[0076] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. Specific implementation method one:

[0078] The present embodiment provides a method for processing analog circuit fault signals based on empirical wavelet transform decomposition, and the method is implemented by the following steps:

[0079] Step 1: Decompose the circuit fault signal by adaptive wavelet transform, then calculate the initial demarcation point obtained by scale space method to obtain the mutual information value of the component, so as to re-divide the spectrum to obtain new decomposed signal components;

[0080] Step 2: Extract feature information from the signal reconstructed by selecting the optimal component using the kurtosis index;

[0081] Step 3: Use K-means clustering analysis to classify the characteristic parameters, so as to realize the diagnosis of analog circuit soft faults. The simulation experiment of bandpass filter circuit shows that the IEWT method can effectively process circuit fault signals for circuit fault diagnosis. Specific implementation method 2:

[0083] Different from the first specific implementation mode, in the present implementation mode, a method for processing analog circuit fault signals based on empirical wavelet transform decomposition is described in step 1, wherein the step of performing adaptive wavelet transform decomposition on the circuit fault signal is specifically:

[0084] The essence of the empirical wavelet transform method is to divide the Fourier spectrum of the signal into continuous intervals, then construct a wavelet filter bank for filtering in each interval, and finally obtain a set of amplitude modulation and frequency modulation components through signal reconstruction. This method can use a wavelet filter bank with compact support characteristics to identify the location of the characteristic information in the Fourier spectrum of the signal and adaptively extract different frequency components of the signal.

[0085] Normalize the Fourier axis of the original signal R(t) to [0, 2π], and limit the discussion range to [0, π] to follow Shannon's theorem; divide the Fourier support interval [0, π] into N continuous segments, each segment is represented by Λ n =[ω n-1 ,ω n ], where ω n-1 and ω nRespectively represent the boundaries of each segment, where n = 1, 2.....N, ω0 = 0, ω N =π, then we have

[0086] like Figure 1 As shown in the Fourier spectrum segmentation diagram, ω n Centered at 2τ n A gray area is represented as a transition zone, τ n =γω n τ, 0<γ<1; In order to ensure that there is no overlap between any two transition zones, γ is limited to less than within the scope, that is:

[0087]

[0088] The horizontal axis in the Fourier spectrum segmentation diagram represents the frequency of the signal R(t) normalized to [0, 2π] after Fourier transformation, and the vertical axis represents the amplitude;

[0089] The empirical wavelet is equivalent to a series of bandwidth Λ n The bandpass filter bank is constructed by the filter of . The wavelet function ψ is constructed based on the ideas of Littlewood-Paley and Meyer wavelets. n (t) and the scaling function φ n (t), ψ n (t) and φ n The Fourier transform of (t) is expressed as ψ n (ω), φ n (ω), and satisfies:

[0090]

[0091] In the formula, β(x) selects the polynomial proposed by Daubechies:

[0092] β(x)=x 4 (35-84x+70x 2 -20x 3 )

[0093] The detail coefficient and approximation coefficient of EWT are defined as:

[0094]

[0095] Among them, ψ n (t) represents the empirical wavelet function, φ1(t) represents the scaling function, and In turn, ψ n Fourier transform of φ(t) and φ1(t); is n The conjugate complex numbers of (t) and φ1(t), F(), F -1 () represents Fourier transform and inverse Fourier transform;

[0096] According to the above formula, the reconstructed signal of the original signal is expressed as:

[0097]

[0098] Where: * represents convolution operation; and Respectively and After signal preprocessing, the wavelet function ψ is constructed n (t) and the scaling function φ n (t), highlighting the transient and local frequency changes in the signal, which helps to extract the characteristics of the fault signal later. Specific implementation method three:

[0100] Different from the second specific implementation mode, in the present implementation mode, a method for processing analog circuit fault signals based on empirical wavelet transform decomposition is described in step 1, wherein the initial demarcation point obtained by the scale space method is calculated to obtain the mutual information value of the component, thereby re-dividing the spectrum to obtain a new decomposed signal component, specifically:

[0101] The scale space method is a local maximum method proposed by Gills in Empirical wavelet transform that can be used to obtain the boundary of the segmented Fourier spectrum. The scale space method is mainly based on the scale space representation theory and the binary clustering algorithm to achieve the segmentation of the Fourier spectrum by finding "meaningful patterns" in the histogram. Compared with the frequency band division of traditional EWT, the parameter-free scale space method can detect the boundary of the segmented Fourier spectrum without prior knowledge, avoiding the influence of human factors. For the circuit fault diagnosis signal f(x) of length N, the steps of the scale space method are as follows:

[0102] Define the Gaussian kernel function The continuous scale space representation of f is as follows:

[0103] L(x, t) = g(x; t) * f(t)

[0104] Where: t is the scale parameter, which indicates the number of convolutions with the Gaussian kernel;

[0105] The circuit fault signal is a signal of finite length, so when t is greater than t max t0 is the initial value of t, t max represents the maximum value of t, In practical applications, from the initial scale t0 to the final scale t max The number of steps N step is limited, at this time there is That is N step =2x max ; So the scale space is expressed in discrete space as:

[0106]

[0107] Define the threshold T to standardize the spectrum segmentation: the number of minimum values ​​in L(x, t) is a decreasing function of the scale parameter t, and no new minimum value will appear as t increases. In order to make the spectrum segmentation meaningful, the threshold T is determined by the Otsu method;

[0108] Although the non-parametric scale space method can detect the boundary of the segmented Fourier spectrum without prior knowledge, it is easily affected by noise, resulting in the inability to accurately obtain the boundary of the segmented Fourier spectrum, thereby increasing the workload of EWT signal processing.

[0109] First, design and improve the IEWT method:

[0110] In order to overcome the problem of too many spectrum divisions in wavelet transform, the present invention re-divides and merges the spectrum according to the mutual information (MI) value to reduce the number of frequency bands; the relative size of the mutual information value more accurately and directly expresses the correlation between two random variables than the correlation coefficient;

[0111] MI(X, Y) = H(Y) - H(Y|X)

[0112] Among them, H(Y) is the entropy of Y, and H(Y|X) is the conditional entropy of Y when X is known:

[0113] Second, design the IEWT signal decomposition steps, such as Figure 2 As shown:

[0114] First, perform FFT on the original signal to obtain the spectrum of the vibration signal; then, determine the initial frequency band boundary points according to the scale space method to obtain the initial spectrum boundary points; then, calculate the mutual information values ​​of each component obtained by spectrum division according to the initial demarcation points, and then redetermine the demarcation points according to the relative size of the mutual information values; then, redivide the spectrum according to the newly determined demarcation points to obtain new decomposed signal components; then, select the optimal component according to the kurtosis index to reconstruct the signal. Dividing the frequency band according to the signal mutual information value effectively overcomes the need to determine the number of modal components in traditional EWT based on personal experience, so that the processed input signal can contain useful information to the greatest extent. Specific implementation method four:

[0116] Different from the third specific implementation mode, in the present implementation mode, a method for processing analog circuit fault signals based on empirical wavelet transform decomposition, the step of extracting feature information from the signal reconstructed by selecting the optimal component using the kurtosis index in step 2 is specifically:

[0117] In order to verify the effectiveness of the experimental method, the present invention uses Figure 4 The bandpass filter circuit shown in the figure is used as a diagnostic example. A circuit diagram model is established in NI Multisim 14.0. Figure 4 shown.

[0118] Data processing was performed in MATLAB R2021b. The parameters of each component in the circuit are as follows Figure 4 As shown, the tolerance of the resistor is 5%, and the tolerance of the capacitor is 10%. Parameters of components within this range are considered to be in a normal state, and those exceeding 30% are considered to be in a fault state.

[0119] The test excitation signal in the experiment is f(t)=sin(800πt)+sin(sin(1000πt))+sin(sin(2800πt))+noise. Through sensitivity analysis, C1, C2, R2, and R3 are selected for test analysis. The positive deviation, negative deviation, and normal operation of the parameters form 9 state modes. The specific fault mode settings are shown in Table 1:

[0120] Table 1 Circuit failure modes of bandpass filter

[0121]

[0122] Multisim simulation software was used to perform transient analysis on the circuit, sample the output response, and obtain the original fault information. Monte Carlo analysis was performed 50 times for each type of fault in Table 1. The samples under each fault mode were randomly divided into two parts, of which 30 groups of data were used for training and the other 20 groups of data were used for testing.

[0123] The signal processing logic and process of various faults are consistent. The following is an explanation of the overall process by processing the output signal of the simulation circuit for fault 2 state.

[0124] The fault signal is transformed into the frequency domain by Fourier transform. The initial frequency band boundary points are determined according to the scale space method to complete the adaptive frequency band division of the frequency domain signal. The initial frequency band division obtained after data processing is as follows Figure 7 shown.

[0125] At this time, the spectrum is adaptively divided into 15 frequency bands. The increase in frequency bands may lead to information redundancy, especially when multiple frequency bands capture similar information. This not only increases the complexity of processing, but also may mask more important fault features in the signal. Moreover, the more frequency bands are divided, the greater the computational complexity of the processing process, which will also cause the program to process the signal more slowly.

[0126] Therefore, the mutual information MI(X,Y) between each band in the initial spectrum is analyzed to reselect the band boundary points. The mutual information value of the band components is as follows Figure 8 The calculation formula of MI(X,Y) is as follows:

[0127] MI(X, Y) = H(Y) - H(Y|X)

[0128] Where H(Y) is the entropy of Y, and H(Y|X) is the conditional entropy of Y when X is known;

[0129] The original frequency bands are re-divided according to the average value of the mutual information to obtain new spectrum demarcation points, and the spectrum is re-divided; the original 15 frequency bands are merged into new 5 frequency bands; as follows Fig. 9 As shown, the complexity of subsequent data calculation is greatly reduced while retaining the fault characteristics.

[0130] According to the new spectrum demarcation point, the wavelet basis function is constructed to perform EWT decomposition on the signal to obtain the processed signal components and reconstructed signals as follows Fig.10 and Fig.11 shown.

[0131] In order to obtain the fault feature information contained in the fault signal efficiently and quickly, a one-dimensional vector consisting of wavelet packet entropy, root mean square, variance, frequency kurtosis, frequency skewness and energy is used to characterize the fault signal characteristics. Wavelet packet entropy, root mean square, variance, frequency kurtosis, frequency skewness and energy are six common features, and they have fewer dimensions, which can reduce the computational complexity and improve the efficiency of the algorithm.

[0132] Through the simulation calculation process of Matlab software and the plotting of the spectrum, the following is a test of 50 groups of samples for each state, including fault 2, fault 3, fault 4, fault 5, fault 6, fault 7, fault 8, and fault 9. A total of 8*50 feature vectors constitute the feature set samples. Table 2 shows three groups of samples randomly selected from the sample set.

[0133] Table 2 Feature dataset

[0134]

[0135] In order to verify the effectiveness of the IEWT method for signal processing and feature extraction, the principal component analysis method is used to reduce the six-dimensional features to three-dimensional features for visualization of the fault signal feature set composed of the simulation experiment structure as follows: Fig.12 shown.

[0136] In order to further better reflect the effectiveness of the IEWT method, the present invention classifies the extracted feature vectors using the K-means clustering analysis method. 40 samples in each state are used for K-means clustering training, and the remaining 10 samples are used for testing. The test results are shown in Figure 2. Fig.13 As shown, various fault types can be clearly distinguished from the figure. Specific implementation method five:

[0138] Different from the fourth embodiment, in the present embodiment, a method for processing analog circuit fault signals based on empirical wavelet transform decomposition is described in step 3, wherein the characteristic parameters are classified by using K-means clustering analysis to achieve the step of diagnosing analog circuit soft faults, specifically:

[0139] With the rise of neural networks, the theoretical models of analog circuit soft fault diagnosis are mostly based on the combination of feature extraction and neural networks. In order to optimize the fault classification method, we understand and learn the application of deep learning in fault diagnosis. We learn the basic structure and principle of convolutional neural networks and introduce the improved capsule network to diagnose analog circuit soft faults based on it;

[0140] As an outstanding machine learning algorithm for deep learning, convolutional neural network has gradually made breakthrough achievements in the field of fault diagnosis. However, it also has its own shortcomings:

[0141] (1) A large amount of data is required for training: CNNs usually require a large amount of labeled data for training, especially when dealing with complex tasks, more data is needed to obtain good generalization performance. This may be a challenge for some fields because it is costly to obtain large-scale labeled data.

[0142] (2) When CNN processes circuit fault features, especially complex circuits, it lacks spatial stratification and spatial reasoning capabilities, has difficulty in handling ambiguity and loss of detail information, and other problems.

[0143] Therefore, capsule networks have attracted the research attention of many scholars. The capsule network architecture is simple, with only three layers, namely convolutional layer, primary capsule layer, and advanced capsule layer.

[0144] The traditional capsule network has a simple architecture, so the quality of circuit fault feature classification is not very high. Based on this, the present invention proposes an improved capsule network model for simulating the diagnosis of circuit soft faults. Its core includes convolution layer, pooling layer, squeeze excitation module, primary capsule layer and advanced capsule layer; the network structure is as follows Fig.14 shown.

[0145] First, the extracted feature information is passed through the convolution layer, and a wide convolution kernel is used to increase the receptive field of the model and retain more feature information; then, the feature dimension is reduced through the maximum pooling layer to improve the operation efficiency; two convolution layers are used in the middle, and the underlying features are fully extracted with a small convolution kernel; then, a squeeze excitation module is added to fully consider the importance of different channel features, re-correct the channel dimension features, and pass the weighted feature information to the primary capsule layer; then, in order to further optimize the model, a batch normalization layer is used after each convolution layer in the feature extraction part to prevent gradient disappearance and enhance generalization ability; all convolution layers in the model use L2 regularization to prevent overfitting; the primary capsule layer first performs a convolution operation, and constructs the convolution result into a vector form as the input of the advanced capsule layer; the advanced capsule layer calculates the correlation between the capsule layers through the self-attention routing algorithm to realize the classification of the extracted features, and obtains x output vectors, where x represents the number of single signal categories, and the value is 9 in the present invention; the output layer calculates the modulus length of the output vectors respectively, that is, obtains the prediction result of the sample.

[0146] Table 3 Grayscale of fault samples

[0147]

[0148] This experiment uses Python programming language and Tensorflow deep learning framework, and the GPU is equipped with NVIDIA Tesla V100. The specific parameter settings of the model are shown in Table 4. The experiment uses Adam optimizer, sets Batchsize to 40, and adjusts the learning rate using exponential decay, with an initial learning rate of 0.001 and a decay rate of 0.9.

[0149] Table 4 Model parameters of capsule network

[0150]

[0151] The grayscale feature vector is converted into a grayscale image and input into the improved capsule network model for training. The loss function of the model is obtained. Fig.15 .pass Fig.15It can be seen that in the early stage of training, the loss rate fluctuated too much and there was an overfitting phenomenon, but as the number of iterations increased, when the number of iterations reached 120, the loss function slowly began to converge until it finally stabilized. The training results show that the parameter settings for model training are basically correct.

[0152] After iterating 200 sets of data to train the network model, the success rate of diagnosing fault categories tends to be stable, as shown in Table 5.

[0153] Table 5 Fault classification accuracy

[0154]

[0155] The fault data set is used as input for deep learning neural network training, and the weight parameters are continuously adjusted. The trained neural network can accurately distinguish various types of faults.

[0156] The embodiments of the present invention disclose preferred embodiments, but are not limited thereto. A person skilled in the art can easily understand the spirit of the present invention based on the above embodiments and make different extensions and changes. However, as long as they do not deviate from the spirit of the present invention, they are all within the protection scope of the present invention.

Claims

1. A method for processing analog circuit fault signals based on empirical wavelet transform decomposition, characterized in that: The method is implemented by the following steps: Step 1: Decompose the circuit fault signal by adaptive wavelet transform, then calculate the initial demarcation point obtained by scale space method to obtain the mutual information value of the component, so as to re-divide the spectrum to obtain new decomposed signal components; Step 2: Extract feature information from the signal reconstructed by selecting the optimal component using the kurtosis index; Step 3: Use K-means clustering analysis to classify the characteristic parameters, so as to realize the diagnosis of analog circuit soft faults.

2. The method for processing analog circuit fault signals based on empirical wavelet transform decomposition according to claim 1, characterized in that: The step of performing adaptive wavelet transform decomposition on the circuit fault signal in step 1 is specifically: Normalize the Fourier axis of the original signal R(t) to [0, 2π], and limit the discussion range to [0, π] to follow Shannon's theorem; divide the Fourier support interval [0, π] into N continuous segments, each segment is represented by Λ n =[ω n-1 ,ω n ], where ω n-1 and ω n Respectively represent the boundaries of each segment, where n = 1, 2.....N, ω0 = 0, ω N =π, then we have ω n Centered at 2τ n The gray area represents a transition zone, τ n =γω n τ, 0<γ<1; In order to ensure that there is no overlap between any two transition zones, γ is limited to less than within the scope, that is: Constructing wavelet function ψ based on the idea of ​​Littlewood-Paley and Meyer wavelets n (t) and the scaling function φ n (t), ψ n (t) and φ n The Fourier transform of (t) is expressed as ψ n (ω), φ n (ω), and satisfies: In the formula, β(x) selects the polynomial proposed by Daubechies: β(x)=x 4 (35-84x+70x 2 -20x 3 ) The detail coefficient and approximation coefficient of EWT are defined as: Among them, ψ n (t) represents the empirical wavelet function, φ1(t) represents the scaling function, and In turn, ψ n Fourier transform of φ(t) and φ1(t); is n The conjugate complex numbers of (t) and φ1(t), F(), F -1 () represents Fourier transform and inverse Fourier transform; According to the above formula, the reconstructed signal of the original signal is expressed as: Where: * represents convolution operation; and Respectively and ) is the Fourier transform of .

3. The method for processing analog circuit fault signals based on empirical wavelet transform decomposition according to claim 2, characterized in that: The step of calculating the initial demarcation point obtained by the scale space method in step 1 to obtain the mutual information value of the component, thereby re-dividing the spectrum to obtain new decomposed signal components, specifically, is: First, design and improve the IEWT method: In order to overcome the problem of too many spectrum divisions in wavelet transform, the present invention re-divides and merges the spectrum according to the mutual information value to reduce the number of frequency bands; the relative size of the mutual information value more accurately and directly expresses the correlation between two random variables than the correlation coefficient; MI(X, Y) = H(Y) - H(Y|X) Among them, H(Y) is the entropy of Y, and H(Y|X) is the conditional entropy of Y when X is known; Second, design the IEWT signal decomposition steps: First, perform FFT on the original signal to obtain the spectrum of the vibration signal; then, determine the initial frequency band boundary points according to the scale space method to obtain the initial spectrum boundary points; then, calculate the mutual information values ​​of the components obtained by spectrum division according to the initial dividing points, and then redefine the dividing points according to the relative sizes of the mutual information values; then, redivide the spectrum according to the newly determined dividing points to obtain new decomposed signal components; then, select the optimal component according to the kurtosis index to reconstruct the signal.

4. The method for processing analog circuit fault signals based on empirical wavelet transform decomposition according to claim 3 is characterized in that: The step of extracting feature information from the signal reconstructed by selecting the optimal component using the kurtosis index in step 2 is specifically: Use Multisim simulation software to perform transient analysis on the circuit, sample the output response, and obtain the original fault information; perform 50 Monte Carlo analyses on each type of fault, and randomly divide the samples under each fault mode into two parts, one part of the data is used for training, and the other part of the data is used for testing; Perform Fourier transform on the fault signal to transform the time domain problem into the frequency domain, determine the initial frequency band boundary point according to the scale space method, and complete the adaptive frequency band division of the frequency domain signal; The frequency band boundary points are reselected by analyzing the mutual information MI(X,Y) between each frequency band in the initial spectrum. The calculation formula of MI(X,Y) is as follows: MI(X,Y)=H(Y)-H(Y|X) Where H(Y) is the entropy of Y, and H(Y|X) is the conditional entropy of Y when X is known; The original frequency bands are re-divided according to the average value of the mutual information to obtain new spectrum demarcation points, and the spectrum is re-divided; the original 15 frequency bands are merged into new 5 frequency bands; According to the new spectrum demarcation point, the wavelet basis function is constructed to perform EWT decomposition on the signal to obtain the processed signal components and the reconstructed signal; A one-dimensional vector consisting of wavelet packet entropy, root mean square, variance, frequency kurtosis, frequency skewness and energy is used to characterize the fault signal characteristics.

5. The method for processing analog circuit fault signals based on empirical wavelet transform decomposition according to claim 4 is characterized in that: The step of classifying the characteristic parameters by using K-means cluster analysis to realize the diagnosis of analog circuit soft faults is as follows: Based on the convolutional neural network, an improved capsule network is introduced to diagnose analog circuit soft faults, and the network structure is designed, including convolution layer, pooling layer, squeeze excitation module, primary capsule layer and advanced capsule layer; First, the extracted feature information is passed through the convolution layer, and a wide convolution kernel is used to increase the receptive field of the model and retain more feature information; then, the feature dimension is reduced through the maximum pooling layer; two convolution layers are used in the middle, and small convolution kernels are used to fully extract the underlying features; then, a squeeze excitation module is added to recalibrate the channel dimension features and pass the weighted feature information to the primary capsule layer; then, a batch normalization layer is used after each convolution layer in the feature extraction part; all convolution layers in the model use L2 regularization; the primary capsule layer first performs a convolution operation, and constructs the convolution result into a vector form as the input of the advanced capsule layer; the advanced capsule layer calculates the correlation between capsule layers through the self-attention routing algorithm to realize the classification of the extracted features, and obtains x output vectors, where x represents the number of single signal categories, and the value is 9; the output layer calculates the modulus length of the output vectors respectively, that is, the prediction result of the sample is obtained.

Images