CN116559650A - A fault identification method for automatic transfer switches based on multi-dimensional entropy distance - Google Patents

Abstract

The invention discloses a multi-dimensional sample entropy-based automatic transfer switch fault identification method, which belongs to the technical field of fault diagnosis. This method first collects the original fault current signal of the electromagnetic mechanism of the automatic transfer switch, and then decomposes the current signal on the set empirical mode, extracts the wavelet energy entropy, frequency spectrum entropy and sample entropy of each empirical mode component, and performs three kinds of information entropy. Standardize and calculate its multidimensional entropy distance as the feature vector, and then select the principal component analysis method to reduce the dimensionality of all feature vectors to obtain the final feature matrix. Divide the fault features into multiple training samples and test samples, then use the training samples to train and classify the fault identification model based on the support vector machine optimized by the grid search-cross-validation algorithm, and finally identify the fault type of the samples according to the classification results . The model proposed by the invention has high innovation in dealing with multi-feature fusion faults, and has high accuracy in the fault identification process.

Description

A fault identification method for automatic transfer switches based on multi-dimensional entropy distance

technical field

The invention belongs to the technical field of fault identification, and in particular relates to a multi-dimensional entropy distance-based automatic transfer switch fault identification method.

Background technique

The automatic transfer switch is composed of one (or several) transfer switch appliances and other necessary appliances, which are used to monitor the power circuit and automatically switch the load circuit from one power supply to another when the circuit fails. It is a guarantee system Necessary appliances for continuity of power supply. However, with the increase in the number of actions and environmental stresses such as vibration and temperature, the health of the automatic transfer switch will degrade, and in severe cases, failures such as mechanical structure jams will occur, making it impossible to switch the circuit normally. Therefore, detecting the health status of automatic transfer switches is the focus of research in recent years.

In fault diagnosis, the selection of characteristic signals is the premise to ensure the accuracy of diagnosis. When the automatic transfer switch switches the circuit, the coil current will change accordingly with the action of the mechanism. The current signal contains rich mechanical structure information. , which can be used as the basis for subsequent fault identification. Acquisition of time-domain signals for automatic transfer switch fault current is simple, but it is susceptible to interference from other signals and has poor anti-interference ability. In addition, the time-domain characteristics caused by the nonlinear and non-stationary characteristics of current signals are difficult to carry out subsequent analysis. Therefore, in frequency The method of feature extraction in domain aspect came into being. In order to better deal with a large amount of detection information, the theory of information entropy is introduced. Traditional wavelet analysis can make signal analysis more refined and high-resolution for high-frequency signal processing, but the information to be processed is also greatly increased. Wavelet energy entropy can effectively reduce the complexity of the system, and power spectrum entropy can be better It describes the change of spectrum energy when the current signal fails, and the sample entropy has a significant advantage in dealing with the statistical characteristics of complex sequence samples. However, feature extraction only from single information entropy is contingency, and the scope of application is limited and cannot deal with complex multi-fault signals, while the fusion of multi-feature signals often causes noise and redundancy.

Contents of the invention

Aiming at the deficiency of the current single information entropy method, a fault identification method of automatic transfer switch based on multi-dimensional entropy distance is provided.

This method first collects the original fault current signal of the electromagnetic mechanism of the automatic transfer switch, and then decomposes the current signal on the set empirical mode to extract the wavelet energy entropy, frequency spectrum entropy and sample entropy of each empirical mode component, so as to solve the problem as comprehensively as possible. The contingency problem of single-signal feature quantity. In order to perform better data fusion on multi-dimensional signals and eliminate the influence of different magnitudes of the three types of information entropy, the three types of information entropy are standardized and their multi-dimensional entropy distances are calculated as feature vectors. Secondly, the principal component analysis method is selected for all The eigenvectors are dimensionally reduced to obtain the final eigenmatrix. Divide the fault features into multiple training samples and test samples, and then use the training samples to train the fault recognition model based on the grid search-cross-validation algorithm optimized support vector machine, and use the trained fault time-awareness model to test the test samples Carry out classification, and finally identify the fault type and fault degree of the sample according to the classification result. The results show that the model proposed by the present invention has higher accuracy in detecting multi-feature fusion faults, and has stronger robustness in the process of fault identification.

A kind of automatic transfer switch fault identification method based on multidimensional entropy distance provided by the present invention comprises the following steps:

(1) Collect the original fault current signal of the object to be diagnosed;

(2) Using ensemble empirical mode decomposition to process the original fault current signal;

(3) Extract the multidimensional entropy value of each modal component after decomposing and calculate its multidimensional entropy distance;

(4) Divide the fault feature samples into training samples and test samples, and use principal component analysis to reduce the dimensionality of the feature matrix;

(5) Using multiple training samples to train the support vector machine fault classification model optimized by the grid search-cross-validation algorithm;

(6) Classify the model of the test sample by using the fault identification model that has been trained;

(7) Identify the fault type and degree of the measured object according to the classification result;

Further, the process of performing ensemble empirical mode decomposition of the original fault information extracted in step (2) includes:

(2-1) Add white noise signals of opposite signs to the original signal in pairs to form two new signals;

(2-2) Carry out empirical mode decomposition to target signal;

(2-3) cyclic steps (2-1) and (2-2);

(2-4) Perform the overall average operation on the above decomposition results to obtain the decomposition results:

in, is the decomposed fault current signal; f _i (t(i=1, 2, ..., n) is the i-th modal component; r _n (t) is the residual component;

The decomposed modal component f _i (t) needs to meet the following conditions:

The number of extreme points and zero crossings in the entire time series differs by at most one;

The mean value of the envelope obtained by the local maximum and local minimum at any time is zero;

Further, the process of extracting multidimensional entropy value as feature value in each modal component in step (3) includes:

(3-1) Using the concepts of energy weight and energy entropy, calculate the ratio of each modal component f(t) to the energy of the original current signal I(t) and use it as

is the energy weight λ _k of the corresponding modal component, as follows:

According to the energy weight of each component, the energy weight entropy J is defined as follows:

Among them, h is the number of data points per cycle, and k is the number of modes;

(3-2) According to the principle of sample entropy, calculate the probability of producing a new mode in each modal signal;

Divide the N sampling points in one period of the fault current signal into a group of vector sequences I _m (1), _Im (2), ..., I _m (N-M+1) with a number of digits m, and define the vector sequence The distance is:

d=max _{k=0, 1, ..., m-1} (|i(n+k)-i(t+k)|)

Where d is the distance between the vector sequence I _m (n) and I _m (t);

For a given _Im (n), count the number of t whose distance between I _m (n) and _Im (t) is less than or equal to r, denoted as B ^m (r), let k=m+1, this The number at time is recorded as B ^m+1 (r);

Among them, B ^m (r) is the probability that two sequences match m points under the similarity tolerance r, and B ^m+1 (r) is the probability that two sequences match m+1 points. Sample entropy is defined as:

(3-3) According to the principle of energy conservation, first perform Fourier transform on the original time-domain current signal, and calculate the power spectrum value S _i of the frequency spectrum sequence y _l at the corresponding frequency, and express it according to the corresponding logarithmic power The corresponding power spectrum entropy value is as follows:

Among them, q _i is the percentage of the i-th power spectrum in the entire power spectrum:

Further, the process of constructing feature matrix after extracting multidimensional entropy value to fault current signal in step (3) includes:

Considering each of the above information entropy as a dimension, these three types of information entropy are regarded as three dimensions, thus forming a three-dimensional space;

The wavelet energy entropy, power spectrum entropy, and sample entropy are respectively used as the x, y, and z axes of the three-dimensional space;

In order to eliminate the influence of different magnitudes of the three kinds of information entropy, the Z-score standardization process is performed on the three kinds of information entropy;

Calculate the mean and standard deviation of the three kinds of information entropy of signal samples in advance, and then take the first ten normal cycles as a group of samples, and get the three kinds of entropy values (H _n , H _p , H _s );

Then treat each cycle as a sample, and after the Z-score standardization process, the entropy point (H _nx , H _px , H _sx ) corresponding to each cycle can be obtained; define the three-dimensional space distance _SH as "information Entropy distance" has:

Select samples of each variable in different states to form a feature matrix X;

Further, the process of dividing fault feature samples into training samples and test samples in step (4) includes:

Select the sample data of different entropy values in the non-fault state to form the training sample matrix X _train :

Among them, X _train is a matrix composed of 4 sample entropy values in ten periodic samples;

Increase the cycle number of fault samples and use it as the detection sample matrix Z _test ;

Further, the process of carrying out principal component analysis to the training sample and the test sample in step (4) includes:

(1) carry out standardized processing to X _train , X _test obtains X _train and X _stest ;

(2) Calculate the covariance matrix of the X _strain , and obtain u eigenvalues and corresponding eigenvectors p ₁ , p ₂ ,...p _u ;

(3) Establish the model of principal component analysis for:

Among them, t _i represents the i-th pivot;

(4) Utilize the eigenvalues and corresponding eigenvectors obtained by solving the Xstest matrix to calculate the upper control limit T _ucl of the T ² statistic and the upper limit Q _ucl of the SPE statistic;

(5) Collect face-changing matrices of different periods, repeat steps (1) to (4), and obtain the T ² statistic and SPE statistic of each sample and the corresponding upper limit T _ucl and Q _ucl , as the fault characteristics after dimensionality reduction value, the overall process is shown in Figure 2;

Further, in the step (5), a plurality of training samples are used to train the support vector machine fault classification model optimized by the grid search-cross-validation algorithm, and the training process includes:

(1) Determine the parameters of the grid search-cross-validation algorithm, set the search range of C and g, the search step size of the grid search method and the fold number of cross-validation;

(2) According to the search step size, search for the values of parameters C and g within the search range, and use the exhaustive method to list all possible parameter combinations.

(3) According to the cross-validation principle, the performance degradation data obtained by the accelerated degradation test are divided into N groups on average, and one group is taken out as a test set, and the remaining N-1 groups are used as a training set, so that each group of data serves as a test set. A total of N combinations are obtained. When training and predicting the SVM model, the average value of N prediction results is taken as the final result.

(4) Input the test sample into the support vector machine model optimized by the grid search-cross-validation algorithm, and output the fault type and degree of the object.

The invention extracts multi-dimensional signal entropy from the current signal of faulty items, and realizes the fusion of three kinds of information entropy by using the method of multi-dimensional entropy distance. Multi-dimensional entropy distance is very effective in measuring nonlinear and random mutation time series, and can accurately extract deep eigenvalues in signals, and is innovative in signal processing and fault feature extraction. In addition, the combination of feature extraction based on multidimensional entropy distance and support vector machine optimized by grid search-cross-validation algorithm can make the fault identification model maintain a high accuracy rate. The present invention has following salient features:

(1) The multidimensional entropy distance extraction feature vector proposed by the present invention eliminates the influence of the contingency of a single feature quantity, and can effectively avoid the misjudgment phenomenon that is easy to occur during traditional feature extraction. Dimensionality reduction processing can reduce noise and redundancy caused by multi-dimensional feature fusion.

(2) the support vector machine optimized by the grid search-cross-validation algorithm proposed by the present invention is a kind of intelligent algorithm, which can effectively solve the problems of time-consuming manual operation and poor precision when support vector machine parameters are set, and Effectively improve the fault identification rate.

(3) The present invention combines the feature vector extraction algorithm based on the multidimensional entropy distance with the support vector machine optimized by the grid search-cross-validation algorithm, and systematically proposes a new fault diagnosis method.

Description of drawings

Fig. 1 is the flow chart of the automatic transfer switch fault identification method based on multidimensional entropy distance of the present invention;

Fig. 2 is the flow chart that the present invention extracts fault current signal multidimensional sample entropy;

Figure 3(a) is the extracted wavelet energy entropy value under the short-circuit fault of the automatic transfer switch coil of a company;

Figure 3(b) is the extracted wavelet energy entropy value under the jamming fault of the automatic transfer switch of a company;

Figure 3(c) shows the wavelet energy entropy value extracted under the jamming fault of the mechanical structure of an automatic transfer switch of a certain company;

Figure 3(d) shows the wavelet energy entropy value extracted under the fault of insufficient armature stroke of an automatic transfer switch of a company;

Figure 4(a) is the power spectrum entropy value extracted under the short-circuit fault of the automatic transfer switch coil of a company;

Figure 4(b) shows the power spectrum entropy value extracted from a company's automatic transfer switch core jamming fault;

Figure 4(c) shows the power spectrum entropy value extracted under the mechanical structure jamming fault of an automatic transfer switch of a certain company;

Figure 4(d) is the power spectrum entropy value extracted under the fault of insufficient armature stroke of an automatic transfer switch of a company;

Figure 5(a) is the sample entropy value extracted under the short-circuit fault of the automatic transfer switch coil of a certain company;

Figure 5(b) is the sample entropy value extracted from a company's automatic transfer switch core jamming fault;

Figure 5(c) is the sample entropy value extracted under the jamming fault of the automatic transfer switch mechanical structure of a certain company;

Figure 5(d) is the sample entropy value extracted under the fault of insufficient armature stroke of an automatic transfer switch of a certain company;

Fig. 6 is the fault recognition rate of the fault signal training set based on the support vector machine of multidimensional sample entropy;

Fig. 7 is the fault recognition rate of the fault signal test set of the support vector machine after optimization based on multidimensional sample entropy;

Detailed ways

The inventive method comprises the steps:

(1) Collect the original fault current signal of the object to be diagnosed;

(2) Using ensemble empirical mode decomposition to process the original fault current signal;

(3) Extract the multi-dimensional entropy value of each modal component after decomposition and calculate its three-dimensional entropy distance

(4) Divide the fault feature samples into training samples and test samples, and use principal component analysis to reduce the dimensionality of the feature matrix;

(5) Using multiple training samples to train the support vector machine fault classification model optimized by the grid search-cross-validation algorithm;

(6) Classify the test samples by using the trained fault recognition model;

(7) Identify the fault type and degree of the measured object according to the classification result;

When dealing with non-stationary signals in engineering, wavelet packet analysis is generally selected to process the collected signals, but there is a problem of energy leakage when the fault signal is decomposed by wavelet packet transform, and its decomposition is not adaptive. Directly extracting feature vectors from components will greatly increase the information to be processed and increase the complexity of the system. In this regard, the theory of wavelet energy entropy is introduced, but the single feature quantity is easily affected by chance, which makes the traditional feature extraction process prone to misjudgment.

In view of the above problems, in order to overcome the problem of low accuracy of single feature quantity fault identification, this example innovatively proposes to use multi-dimensional entropy value as the feature vector of the original fault current signal, which can comprehensively combine the advantages of different sample entropy and avoid the process of feature extraction. of misjudgment. In order to illustrate the superiority of multi-dimensional entropy, this example uses the automatic transfer switch as the fault object to illustrate the effectiveness of the method, and analyzes the fault current signals of the automatic transfer switch under different fault conditions. The experiment uses the test data of a company’s product quality inspection department. The full-life test bench used in the test includes power supply device, support device, control device, data acquisition device and monitoring platform. The model of the test appliance is ATSE-100/63A/4P, artificially simulated There are four kinds of faults: coil short circuit, iron core jam, mechanical structure jam and lack of armature stroke. The sampling frequency is 20KHZ. Each state takes 40 sets of data, and the length of each set of data is 1200 data points. After the data is collected, the aggregate empirical mode decomposition is performed, and the wavelet energy entropy, power spectrum entropy and sample entropy analysis are performed on each modal component after decomposition, as shown in Figure 3, Figure 4 and Figure 5.

Analysis of Figure 3 shows that there are significant differences in the wavelet energy entropy values of different fault types. The wavelet energy entropy in the normal cycle basically presents a horizontal line, and the overall fluctuation is small, while the entropy value of the fault cycle fluctuates greatly and tends to increase. And when a core fault occurs, the growth rate is significantly greater than other faults, so the wavelet energy entropy has a significant advantage in distinguishing core faults.

Analysis of Figure 4 shows that, for the power spectrum entropy, the waveform is still approximately linear in the normal cycle, but in the fault cycle, different types of fault waveforms all fluctuate, and when a coil short circuit fault occurs, the power spectrum entropy value is at the end It is in a stable rising stage, while other types of faults have downward fluctuations, so the power spectrum entropy has obvious advantages in distinguishing coil short-circuit faults.

Analyzing Figure 5, it can be seen that for the sample entropy, there is a significant gap between the sample entropy of different fault types. When the fault occurs, the change of sample entropy is not obvious compared with other faults, and the sample entropy value only increases rapidly at the end of the cycle. Therefore, the sample entropy has obvious advantages in distinguishing mechanical structure jamming and armature stroke deficiency.

To sum up, based on the identification ability of different sample entropy for each fault type, multi-dimensional entropy can effectively extract the feature quantity of each fault sample. In order to integrate the advantages of each information entropy and reduce the amount of calculation, the multi-dimensional entropy value is reduced through principal component analysis, so as to obtain the T ² and SPE statistics of each sample and their upper control limits T _ucl and Q _ucl , as shown in Table 1 and Table 1. 2.

Table 1 The test data T ² statistic and the control upper limit T _ucl of the product quality inspection department of a certain company

Table 2 SPE statistics and control upper limit Q _ucl of experimental test data of a company's product quality inspection department

In the above experimental data, there are 4 types of faults, each type of fault data has 60 sets, a total of 240 sets of data, now randomly select 30 sets of samples from 60 sets of samples of each type of fault as training samples, and the remaining 30 sets Samples are used as test samples, with a total of 120 sets of training samples and 120 sets of test samples; first, the test set is input into the support vector machine model to be optimized, and the optimal network structure parameters are obtained through learning, and then the parameters are reassigned , input the test set data into the optimized support vector machine model, and the fault recognition rates obtained from the training set and the test set are shown in Figure 6 and Figure 7.

It can be seen from Figure 6 that the classification of the first two types of faults by the unoptimized support vector machine model is not accurate, and there is a case of misclassification. The classification of the optimized support vector machine model is shown in Figure 7. The optimized model has a classification accuracy of 100% for the first two faults, and the model can still maintain a high accuracy for the latter two faults. rate, there is only one series of mechanical structure jammed faults that are misclassified as faults with insufficient armature travel, which fully demonstrates that the multidimensional information entropy of fault signals can effectively reflect the characteristics of different faults, and the optimized support vector machine model The accuracy of fault identification has been significantly improved, and it can accurately identify different faults.

Claims (4)

1. An automatic transfer switch fault identification method based on multidimensional entropy distance is characterized by comprising the following steps:

(1) Collecting an original fault current signal of an object to be diagnosed;

(2) Processing the original fault current signal by adopting ensemble empirical mode decomposition;

(3) Extracting the multi-dimensional entropy value of each decomposed modal component and calculating the multi-dimensional entropy distance;

(4) Dividing the fault characteristic sample into a training sample and a test sample, and performing dimension reduction treatment on the characteristic matrix by principal component analysis;

(5) Training a support vector machine fault classification model optimized by a grid search-cross verification algorithm by adopting a plurality of training samples;

(6) Classifying the model of the test sample by using the trained fault recognition model;

(7) Identifying the fault type and degree of the detected object according to the classification result;

the process of performing the aggregate empirical mode decomposition of the original fault information extracted in the step (2) includes:

(2-1) adding white noise signals of opposite signs to the original signal in pairs to form two new signals;

(2-2) performing empirical mode decomposition on the target signal;

(2-3) cycling steps (2-1) and (2-2);

(2-4) carrying out ensemble average operation on the decomposition result to obtain a decomposition result:

wherein ,a decomposed fault current signal; f (f) _i (t) (i=1, 2, …, n) is the i-th modal component; r is (r) _n (t) is a residual component;

the decomposed modal component f _i (t) the following conditions are required to be satisfied:

the number of extreme points and zero crossing points in the whole time sequence is different by one at most;

the average value of the envelope curve obtained by the local maximum value and the local minimum value at any moment is zero;

further, the process of extracting the multi-dimensional entropy value as the characteristic value in each modal component in the step (3) includes:

the process of extracting the multi-dimensional entropy value as the characteristic value in each modal component in the step (3) comprises the following steps:

(3-1) applying the concepts of energy weight and energy entropy, and calculating the energy weight entropy value under each modal component;

(3-2) calculating the probability of generating a new mode of each mode signal according to the principle of sample entropy, thereby obtaining the sample entropy value of each mode component;

(3-3) according to the principle of conservation of energy, carrying out Fourier transform on the time domain signals, and drawing a logarithmic power spectrum to reflect the power spectrum entropy value of each component;

the process of separating the fault signature sample into the training sample and the test sample in step (4) comprises:

selecting sample data with different entropy values under non-fault state to form training sample matrix X _train ：

wherein ,X_train A matrix of 4 sample entropy values in ten periodic samples;

increase the cycle number of the fault sample and take it as a detection sample matrix X _test ；

In the step (5), the training process of the support vector machine fault classification model optimized by the grid search-cross verification algorithm by adopting a plurality of training samples comprises the following steps:

(5-1) inputting the multidimensional entropy into a support vector machine model after parameter optimization, and searching possible values of parameters to be optimized of the support vector machine through an exhaustion method;

(5-2) outputting the different fault states as support vector machines, and selecting optimal parameters by using cross validation.

2. The method for identifying faults of automatic transfer switches based on multidimensional entropy distance according to claim 2, wherein the characteristic quantity extraction is characterized by the steps (3-1) and (3-3), and the specific steps are as follows:

(1) The concept of energy weight and energy entropy is applied, and the ratio of each modal component f (t) to the energy of the original current signal I (t) is calculated and used as the energy weight lambda of the corresponding modal component _k The method is characterized by comprising the following steps:

according to the energy weight of each component, an energy weight entropy J is defined as follows:

wherein h is the number of data points per period, and k is the number of modes;

(2) According to the principle of sample entropy, calculating the probability of generating a new mode in each mode signal, B ^m (r) probability of two sequences matching m points with similar margin r, and B ^m+1 (r) probability of matching m+1 points for two sequences, sample entropy is defined as:

(3) Firstly, carrying out Fourier transform on an original time domain current signal, and calculating a frequency spectrum sequence y _i Power spectrum value S at corresponding frequency _i Solving the percentage q of the ith power spectrum in the whole power spectrum _i And the corresponding power spectrum entropy value is represented according to the corresponding logarithmic power, and the specific steps are as follows:

3. the method for identifying the fault of the automatic transfer switch based on the multidimensional entropy distance according to claim 2, wherein the step (4) optimizes parameters of a support vector machine by using a main analysis algorithm, and comprises the following specific steps:

(1) For X _train ，X _test Standardized treatment is carried out to obtain X _strain and X_stest ；

(2) Find X _strain Obtaining u eigenvalues and corresponding eigenvectors p by covariance matrix of (2) ₁ ，p ₂ ，…p _u ；

(3) Modeling principal component analysisThe method comprises the following steps:

wherein ,t_i Representing the ith principal element;

(4) Calculating T by using characteristic values obtained by Xstest matrix solution and corresponding characteristic vectors thereof ² Upper control limit of statistics Xstest and upper limit of SPE statistics Q _ucl ；

(5) Collecting face matrixes of different periods, and repeating the steps (1) to (4) to obtain T of each sample ² Statistics and SPE statistics and corresponding upper limit T _ucl and Q_ucl As a fault characteristic value after dimension reduction.

4. The method for identifying the fault of the automatic transfer switch based on the multidimensional entropy distance according to claim 2, wherein the step (5) optimizes parameters of the support vector machine by using a grid search-cross validation algorithm, and comprises the following specific steps:

1) Determining parameters of a grid search-cross verification algorithm, and setting search ranges of C and g, search step length of a grid search method and the number of folds of cross verification;

2) Searching the values of the parameters C and g in a searching range according to the searching step length, and listing all possible parameter combinations by using an exhaustion method;

3) According to the cross verification principle, performance degradation data obtained by an accelerated degradation test are equally divided into N groups, one group is taken out to serve as a test set, the other N-1 groups are taken as training sets, and each group of data serves as a test set to obtain N combinations; and when training and predicting the SVM model, taking the average value of N times of prediction results as a final result.