CN103728551A - Analog circuit fault diagnosis method based on cascade connection integrated classifier - Google Patents

Abstract

The invention discloses a fault diagnosis method of an analog circuit and its realization. The content of the invention is divided into three parts: (1) extraction of analog circuit fault feature information; (2) fault classifier construction; (3) algorithm software implementation. The method includes the following steps: constructing the fault feature information database, adopting the principle of maximum information entropy, selecting the optimal mother wavelet, performing wavelet decomposition on the response nodes of the tested circuit, extracting the optimal features of the tested circuit, and using principal component analysis to analyze Dimensionality reduction of fault features; fault classification and intelligent diagnosis, according to the obtained fault feature information, using multi-classifier cascade model and classifier integration technology to construct a fault diagnostic device to identify existing faults and their causes; using C#.NET platform, combined with The Weka software implements the algorithm concretely. The fault diagnosis method and its realization method of the present invention have the advantages of higher fault diagnosis performance, wider diagnosis range, stronger algorithm robustness and stronger explainability.

Description

A Fault Diagnosis Method for Analog Circuits Based on Cascaded Integrated Classifiers

technical field

The invention relates to a fault diagnosis method of an analog circuit and its realization.

Background technique

Fault diagnosis of analog circuits began in the 1960s, and its theoretical research started from the solvability of network element parameters. However, due to its unique difficulties such as diversity of fault states, tolerance of element parameters, insufficient information and Due to the complexity of the structural model, etc., the development of research on fault diagnosis of analog circuits is relatively slow, and its testing and fault diagnosis has always been a problem that plagues the circuit testing industry. After the 1990s, with the development of artificial intelligence technology, fuzzy theory, wavelet technology and some machine learning methods have been applied in this field and achieved good results, but they all have one-sidedness and have no effect on solving actual analog circuit faults. There is still a certain gap between diagnosis and analysis. At the same time, the actual demand for fault diagnosis of analog circuits is constantly increasing. Therefore, it is of great significance to study an accurate and fast fault detection and fault location method for analog circuit boards, shorten the detection and maintenance time and reduce maintenance costs for the maintenance of analog circuit boards in electronic equipment.

Contents of the invention

The invention discloses a method for diagnosing faults of an analog circuit and its implementation, including three aspects: signal fault feature extraction, fault classification and identification, and algorithm software implementation. The method consists of the following steps: (1) Construct the fault feature information database, according to the signal characteristics of the circuit under test, adopt the maximum information entropy principle (MEP), select the optimal mother wavelet, and perform wavelet decomposition on the response nodes of the circuit under test, extract The optimal characteristics of the circuit under test, and then use principal component analysis (PCA) to reduce the dimensionality of each layer to obtain fault feature information. (2) Fault analysis and intelligent diagnosis, according to the obtained fault characteristic information parameters, using multi-classifier cascade model and ensemble (Ensemble) [homomorphism and heteromorphism] technology to construct an intelligent fault diagnostic device to identify possible faults and their reason. a) The fault diagnosis device uses a multi-classifier cascade model, firstly to solve the problem that it is difficult to distinguish between normal samples and early fault samples, that is, firstly, the characteristic samples of normal circuits and the samples of all fault circuits form two disjoint subsets, Using homomorphic integration technology to construct support vector machine classifier, forming a level G ₀ , which is used to distinguish normal and fault states; secondly, using heteromorphic integration technology to train base classifiers of different algorithms for fault samples, and then using weighted The voting algorithm merges the classifiers to form a level G ₁ , which is used to distinguish different fault states. This two-level structure forms a multi-classification cascading reasoning idea. b) The classifier construction of the cascaded model adopts the integration technology. First, for the level G ₀ , the homomorphic integration technology is adopted, that is, the Bagging algorithm of unilateral sampling is used to solve the problem of data imbalance to train an integrated support vector machine classifier. Then for level G ₁ , a synthetic classifier based on Bayesian, decision tree and support vector machine algorithms is trained by using heteromorphic integration technology to perform weighted voting output on the samples and increase the generalization accuracy of the fault diagnosis system. (3) The software implementation of the algorithm adopts the Microsoft C#.NET platform, and the weka.jar file of the Weka software project is converted into a weka.dll assembly that can be called by .NET through the IKVM.NET tool. Classes are rewritten, and after the specific implementation of the algorithm is completed, the software is written using a three-layer architecture model to achieve specific analysis and diagnosis of faults.

Description of drawings

Figure 1 Flow chart of fault feature extraction

Figure 2 Fault decision flow chart

Figure 3 software architecture diagram

Detailed ways

Knowledge-based analog circuit fault diagnosis technology is essentially a pattern recognition and classification problem. Therefore, how to extract effective features of faults is a key technology and an important part of analog circuit fault diagnosis. At the same time, the ultimate purpose of extracting features is to construct a classifier for test samples and realize correct classification and identification of different fault types. Ultimately, to achieve this goal and complete the real realization of fault diagnosis, it is necessary to implement software for the algorithm.

In order to achieve the above object, method of the present invention is achieved like this:

1. Optimal wavelet extraction of analog circuit fault feature information

The wavelet fault feature information extraction method as a signal processing is a current research hotspot. Wavelet analysis belongs to multi-resolution analysis, which is a fine time-frequency analysis method. Multi-layer decomposition of the signal is conducive to obtaining more local sampling signals. However, since different types of wavelets have different time-frequency characteristics, in order to more effectively extract the fault feature information of the circuit, the time-frequency characteristics of the wavelet should be matched with the time-frequency characteristics of the circuit response node. Therefore, the present invention uses An optimal mother wavelet selection method based on the principle of maximum information entropy is proposed to solve this problem. The specific steps are as follows, and the process is shown in Figure 1:

(1) Let any given node response signal be f(t), according to the definition of wavelet transform,

$W_f(a,b)=\frac{1}{\sqrt{a}}\&Integral;f\left(t\right){\mathrm{\&psi;}}^{*}\left(\frac{t-b}{a}\right)\mathrm{dt}=<f\left(t\right),{\mathrm{\&psi;}}_{a,b}\left(t\right)>,$ In the formula: ψ is the wavelet mother function, a is the scale parameter, b is the time center parameter, W _f (a, b) is the wavelet decomposition coefficient of the signal, and the wavelet transform is performed on the signal, but the transform formula can only be approximated by a computer Calculation, in order to achieve higher precision and simplicity of calculation, the trapezoidal method is generally used for approximate numerical integration operations, taking t=n·ΔT, b=k·ΔT, the calculation formula is

$W_f(a,k)=\frac{\mathrm{\&Delta;T}}{2\sqrt{a}}\underset{\mathrm{no}}{\mathrm{\&Sigma;}}[f\left(\mathrm{no}\right)\mathrm{\&psi;}\left(\frac{\mathrm{no}-k}{a}\right)+f(\mathrm{no}+1)\left(\frac{\mathrm{no}+1-k}{a}\right)],$ In the formula: ΔT is the sampling interval, assuming a=2 ⁱ (i=1, 2..., n), for each given a=2 ⁱ , the value of k is calculated sequentially at different times, and different wavelets are taken at the same time Generator function is calculated, and a set of wavelet transform coefficients when different wavelet mother functions are taken can be obtained.

根据信息熵基本理论，可定义电路信息熵

则可以根据电路信息熵的最大原则来选择最优的小波母函数。例如经过计算可得冲激响应信号适合于用Haar小波进行变换。(2) Use different wavelet transforms to calculate the information entropy of the circuit, and select the optimal mother wavelet to perform wavelet transform on the circuit node response. Let the normal signal of the response node be r(t), and the fault response signal f _i (t)(i=1,...,c), where i is the type of fault and c is the total number of faults, then the information entropy calculation method of the circuit is as follows : Carry out corresponding wavelet transform on the normal signal r(t) and the fault response signal f _i (t) respectively, and the decomposition level is n, then respectively take the nth layer low frequency approximation coefficient and the first,...,n layer high frequency approximation The coefficients form a vector, let the normal signal be R(k), and the faulty signal be F _i (k), where k is the number of signal sampling samples, and the normal signal R(k) is recorded as F ₀ (k), then all signals The vector formed after wavelet transformation can be expressed as F _i (k) (i=0, 1, ..., c), and then calculate the cosine similarity between each two vectors

According to the basic theory of information entropy, circuit information entropy can be defined

Then the optimal wavelet mother function can be selected according to the maximum principle of circuit information entropy. For example, the impulse response signal obtained through calculation is suitable for transformation by Haar wavelet.

其中l为第n层小波分解，

N为每层小波系数的个数。归一化后，将每层小波系数合并构成一个新的向量

将所有归一化的样本向量组合成矩阵X，建立相关矩阵，

m为样本数目，由R即可获得特征值λ_i和特征向量a_i(i＝1，2，...，n)，计算第i个主元对总方差的贡献率，按贡献率由大到小进行排列，依次选取k个主元使得积累贡献率之和大于90％。之后依据

计算出所需的各主元值，形成最终的特征向量样本。(3) Before performing PCA, in order to avoid the influence of the data dimension, the coefficients of each layer of wavelet transform of the response signal are normalized respectively, and the following formula is used to carry out:

where l is the wavelet decomposition of the nth layer,

N is the number of wavelet coefficients in each layer. After normalization, the wavelet coefficients of each layer are combined to form a new vector

Combine all normalized sample vectors into a matrix X to build a correlation matrix,

m is the number of samples, and the eigenvalue λ _i and eigenvector a _i (i=1, 2, ..., n) can be obtained from R, and the contribution rate of the i-th pivot to the total variance is calculated, according to the contribution rate by Arrange from large to small, and select k pivots in turn so that the sum of the cumulative contribution rate is greater than 90%. later based on

Calculate the required pivot values to form the final eigenvector samples.

2. Construction of fault classifier

The purpose of fault diagnosis is to classify and identify test samples. The usual method is to design different classification algorithms. For example, the commonly used classifiers include neural network classifiers, support vector machine classifiers, and Bayesian classifiers. Possibly good recognition performance, different classification schemes are often designed, but no matter which classifier is used for different problems, the effect is not always the best, so now the common practice is to improve various classifiers And technology based on classifier integration (Ensemble), and integrated learning can improve the accuracy of traditional classifiers by using the output of multiple base classifiers, and achieved good results. The present invention is based on the integration technology of classifiers, and adopts a cascade model to construct a reasoning fault diagnosis classifier. The samples of all faulty circuits respectively constitute two disjoint subsets, using homomorphic integration technology to construct a support vector machine classifier, forming a level G ₀ , which is used to distinguish normal and faulty states; secondly, for faulty samples, using abnormal state Based on the integrated technology of different algorithms, the base classifiers of different algorithms are trained, and then the classifiers are merged by using the weighted voting algorithm to form a level G ₁ , which is used to distinguish different fault states. Using such a two-level structure model, a diagnostic reasoning idea is formed. . The specific ideas and implementation steps are as follows, and the process is shown in Figure 2:

(1) Divide the feature samples of normal circuits and samples of all fault circuits into two disjoint subsets X _N and X _F , let X _N be the positive sample instance, X _F be the negative sample instance, however according to practical experience We know that the normal class has a high probability of appearing in the data set, while the probability of the fault class appearing is very small, which will cause the number of samples of the normal class to be significantly larger than the samples composed of other fault classes. This data set is called Unbalanced dataset. The main idea of the classification method to deal with this unbalanced data set is: (a) use a base classifier that has good adaptability to the unbalanced category, while the ensemble learning algorithm remains unchanged, (b) use a traditional classifier, by modifying the ensemble The learning algorithm is the resulting classifier that is able to adapt to the imbalanced class problem. The second line of thinking is currently dominant. As an important ensemble learning algorithm, Bagging is easy to implement and works well. The present invention adopts a method called homomorphic ensemble classification technology, and uses the unilateral sampling Bagging ensemble learning algorithm to train the sample sets X _N and X _F , and the base classifier uses the same classification learning algorithm-Support Vector Machine, which It overcomes the shortcomings of neural networks, and shows the characteristics of simple structure, global optimality and strong generalization ability in solving small sample, nonlinear and high-dimensional pattern recognition problems. The unilateral sampling Bagging integrated learning algorithm based on support vector machines is trained in this way. In each round, all positive sample instances are first extracted, and then randomly extracted from the negative sample instances with replacement as many positive samples. Instance samples, together with all positive examples constitute the training set T _i , and then use the base classification learning algorithm - support vector machine to train the base classifier from T _i , and finally fuse the base classifiers learned in each round to form The first level G ₀ of our fault diagnosis classifier can be used for the classification of normal and fault, and the output of normal is 0, while the output of fault is 1.

(2) When the output of the G ₀ classifier is 1, the sample is a fault sample, and it is necessary to further judge which kind of fault it is, so it is necessary to construct one or more fault classifiers. For problems with a large number of categories, in addition to some traditional classification algorithms, such as neural networks. At present, many improved algorithms have emerged, such as firstly performing rough classification based on clustering algorithms, and then constructing classifiers for each rough classification and a multi-classifier merging algorithm called heterogeneous integration technology, that is, using the same data sample , while the base classifiers are trained with different algorithms, and finally the obtained base classifiers are merged and so on. The present invention adopts the heteromorphic integrated classification technology to construct the second-level _G1 classifier, and the base classifier selects Bayesian algorithm, decision tree algorithm and support vector machine algorithm, and then evaluates various base classifiers, using The weighted voting method outputs the result, the classification curve of the integrated classifier constructed in this way will be obviously smooth, and it also has the characteristics of strong robustness.

3. A software implementation of the algorithm

Algorithms are the soul of solving problems, and the realization of algorithms enables the soul to have a body attached to it, and only then does it have real practical significance. It is a boring and difficult thing to implement the above-mentioned algorithm. Fortunately, Weka, as an open data mining work platform, has assembled a large number of machine learning algorithms that can undertake data mining tasks, including data pre-processing. Processing, classification, regression, clustering, etc., and more importantly, developers can improve the open source code, and even use Weka's architecture to develop more data mining algorithms. Therefore the present invention utilizes Weka project software based on C# platform to complete the realization of above-mentioned algorithm, concrete steps are as follows:

(1) The software implementation of the algorithm adopts the Microsoft C#.NET platform. To call the weka.jar file of the Weka software project under .NET, it is necessary to use the IKVM.NET tool to convert the weka.jar file into a weka that can be called by .NET .dll assembly, just execute ikvmc-target: library weka.jar, and import weka.dll into the .NET project reference.

(2) Export the one-dimensional wavelet transform function realized in C++ as a function that can be called in C#, adopt the following method:

using System.Runtime.InteropServices;

[DllImport("Wavelet1D.dll", CharSet=CharSet.Auto)]

public static extern int[]Wavelet1D(string filename, int level, string wname, refint[]length);

Then call the optimal mother wavelet in C# to perform wavelet transform function on the signal, get the wavelet transform coefficients and normalize them, and then call the PrincipalComponents class in weka to realize the principal component analysis of the sample. Note that before using the weka PrincipalComponents class, you need to use the using statement to guide the weka.filters.unsupervised.attribute, org.antlr.stringtemplate, org.antlr.stringtemplate.language namespace.

(3) Realize the rewriting of Bagging class, because what the weka.classifiers.meta.Bagging class in weka realizes is that the sampling mode that puts back ground is standard to operate training set, and what the present invention adopts is unilateral sampling Bagging Algorithm, so the method in its class needs to be rewritten and named as SSBagging class. Then directly call the LibSVM base classifier to train the first-level G ₀ classifier, named SSBaggingClassify class, and its basic code is as follows:

(4) After constructing the first-level G ₀ classifier, when its output is 1, it is necessary to further judge what kind of fault it is. According to the above ideas, it is necessary to construct a classifier based on Ensemble technology, which is used for classification in Weka key class

weka.classifiers.functions.LibSVM, weka.classifiers.tress.J48,

weka.classifiers.bayes.NaiveBayes and weka.classifiers.meta.Vote for ensemble techniques. Its basic code is as follows:

Finally, the software is implemented using a three-layer architecture model, and its structure diagram is shown in Figure 3, which will not be repeated here.

Claims (1)

1. The present invention relates to a fault diagnosis method of an analog circuit and its implementation. Its content mainly includes three aspects: signal fault feature extraction, fault classification and identification, and algorithm software implementation. It is characterized in that the method is carried out according to the following steps: (1) Construct the fault characteristic information database, according to the signal characteristics of the circuit under test, adopt the maximum principle of information entropy (MEP), select the optimal mother wavelet, and perform wavelet analysis on the response nodes of the circuit under test Decompose, extract the optimal features of the circuit under test, and then use Principal Component Analysis (PCA) to reduce the dimension of each layer to obtain fault feature information. (2) Fault analysis and intelligent diagnosis, according to the obtained fault characteristic information parameters, using multi-classifier cascade model and ensemble (Ensemble) [homomorphism and heteromorphism] technology to construct an intelligent fault diagnostic device to identify possible faults and their reason. a) The fault diagnosis device uses a multi-classifier cascade model, firstly to solve the problem that it is difficult to distinguish between normal samples and early fault samples, that is, firstly, the characteristic samples of normal circuits and the samples of all fault circuits form two disjoint subsets, Using homomorphic integration technology to construct support vector machine classifier, forming a level G ₀ , which is used to distinguish normal and fault states; secondly, using heteromorphic integration technology to train base classifiers of different algorithms for fault samples, and then using weighted The voting algorithm merges the classifiers to form a level G ₁ , which is used to distinguish different fault states. This two-level structure forms a multi-classification cascading reasoning idea. b) The classifier construction of the cascaded model adopts the integration technology. First, for the level G ₀ , the homomorphic integration technology is adopted, that is, the Bagging algorithm of unilateral sampling is used to solve the problem of data imbalance to train an integrated support vector machine classifier. Then for level G ₁ , a synthetic classifier based on Bayesian, decision tree and support vector machine algorithms is trained by using heteromorphic integration technology to perform weighted voting output on the samples and increase the generalization accuracy of the fault diagnosis system. (3) The software implementation of the algorithm adopts the Microsoft C#.NET platform, and the weka.jar file of the Weka software project is converted into a weka.dll assembly that can be called by .NET through the IKVM.NET tool. Classes are rewritten, and after the specific implementation of the algorithm is completed, the software is written using a three-layer architecture model to achieve specific analysis and diagnosis of faults. The fault diagnosis method of the present invention has the advantages of higher fault diagnosis performance, wider diagnosis range, stronger algorithm robustness and stronger explainability.

Images