Circuit breaker fault diagnosis method based on global-local feature extraction

Abstract

The invention discloses a circuit breaker fault diagnosis method based on global-local feature extraction. Comprising the following steps: 1, collecting vibration signals of a circuit breaker in a normal working condition and a fault state as original data; 2, extracting original time domains and frequency domains of the vibration signals of the circuit breaker operating mechanism in different states and time-frequency characteristics of a generalized S transformation matrix, and regarding the original time domains and frequency domains as global characteristics; 3, extracting local features of the vibration signals by using a texture feature extraction mode of matching a local binary pattern with a gray-level co-occurrence matrix; 4, further screening the features by using maximum correlation and minimum redundancy to obtain a feature set; and 5, in combination with a kernel extreme learning machine classifier and GL-mRMR feature extraction, obtaining a GL-mRMR-KELM model applied to circuit breaker vibration data classification. The method is used for fault diagnosis of a high-voltage circuit breaker, and solves the technical problems of low universality and poor accuracy under small samples in a current operating mechanism state diagnosis algorithm.

Description

Technical Field

[0001] The invention relates to a circuit breaker fault diagnosis method based on global-local feature extraction, belonging to the technical field of power equipment fault diagnosis. Background Art

[0002] As a key device for breaking circuits in power grids, high-voltage circuit breakers may fail during long-term operation, most of which are mechanical failures of the operating mechanism. Therefore, early diagnosis and prediction of mechanical failures of circuit breaker operating mechanisms are of great significance for the safe and stable operation of power systems.

[0003] Domestic and foreign research on mechanical fault state diagnosis of circuit breaker operating mechanisms focuses on vibration signal feature extraction methods and fault classification models. During the operation of high-voltage circuit breakers, the components interact and influence each other, so the vibration signal contains a lot of redundant information. When extracting features, there are problems of complex features and high dimensions, which greatly increases the requirements for feature extraction algorithms. If the number of extracted features is insufficient, it will lead to insufficient expression; and if the feature dimension is too high, it will lead to repeated expression. Both situations will affect the diagnosis effect.

[0004] The main methods for extracting features of existing circuit breaker vibration signals include short-time energy analysis, wavelet packet analysis, phase space reconstruction, empirical mode decomposition (EMD), variational mode decomposition (VMD), etc. Among them, using wavelet or modal analysis to decompose the signal and then extract features is the main feature extraction method. However, due to the differences in the mechanical structures of different equipment, when these methods are extended to other types of equipment, their feature extraction parameters are often no longer applicable and need to be re-selected through experiments.

[0005] At present, the models used as classifiers in the state diagnosis of circuit breaker operating mechanisms mainly include support vector machines (SVM), clustering, random forests, neural networks, etc. As the most widely used model, the support vector machine (SVM) has the disadvantage that it is difficult to train in the case of small samples and is prone to fall into local optimality. In addition, it is difficult to obtain the fault state data of high-voltage circuit breaker operating mechanisms, and it is difficult to collect a large number of samples. The application of typical neural networks such as convolutional neural networks (CNN) and residual networks (ResNet) is more restricted by the problem of effective data volume. The high training cost and data cost also make it difficult to migrate and apply them between different types of equipment. The applicability of them in the state diagnosis of circuit breaker operating mechanisms needs further research and improvement. Summary of the invention

[0006] The purpose of the present invention is to provide a circuit breaker fault diagnosis method based on global-local feature extraction, which is used in high-voltage circuit breaker fault diagnosis to solve the technical problems of low versatility and poor accuracy under small samples in the current operating mechanism state diagnosis algorithm. The purpose of the present invention is to provide a circuit breaker global-local feature fault extraction scheme for processing one-dimensional vibration data of circuit breakers, which is used in high-voltage circuit breaker fault diagnosis. The method can comprehensively extract signal features and thus complete the diagnosis. The feature vector extracted by the method has better intra-class compactness and inter-class separability. First, the global features are extracted from the one-dimensional vibration signal, and the one-dimensional signal is further subjected to local binary processing. The gray-level co-occurrence matrix is ​​obtained for the processed signal, and finally the texture features of the gray-level co-occurrence matrix are extracted as local features. In the feature extraction stage, the global features and local features are processed separately to reduce the demand for obtaining detailed features through modal decomposition, improve the versatility of the feature extraction method, and use the maximum relevance minimum redundancy (mRMR, minimum Redundancy-Maximum Relevance) method to complete feature screening. The present invention refers to this method as the global-local feature extraction GL (Global-Local)-mRMR method.

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

[0008] A circuit breaker fault diagnosis method based on global-local feature extraction comprises the following steps:

[0009] 1. Collect vibration signals of circuit breakers under normal working conditions and fault conditions as raw data:

[0010] 1.1 Collect normal status data;

[0011] 1.2 Collect vibration data under various fault conditions, including loose closing electromagnet, abnormal opening spring output, and loose oil buffer fastening bolts;

[0012] 1.3 Construct the circuit breaker mechanical state fault data set through the collected fault data and normal data;

[0013] Second, the original time domain, frequency domain and time-frequency features of the generalized S transform matrix of the vibration signal of the circuit breaker operating mechanism under different states are extracted and regarded as global features. The specific implementation steps are as follows:

[0014] 2.1 First, extract the global features of the collected data set in the original state. Such features belong to the global features in the time domain. The specific types of time domain features collected include mean, standard deviation (Std), root mean square (RMS), skewness, kurtosis, peak-to-peak value, impulse factor, and shape factor;

[0015] 2.2 Perform Fourier transform on the original one-dimensional vibration data to obtain the signal frequency domain data set, and extract the global frequency domain features of the transformed frequency domain data. The specific extracted features are: center frequency (CF), mean square frequency (MSF), root mean square frequency (RMSF), frequency variance (VF), frequency standard deviation, frequency amplitude median, frequency maximum value, and frequency median value;

[0016] 2.3 Perform generalized S transform on the original one-dimensional vibration data to obtain the signal time-frequency domain feature matrix data set, and extract the sum of column standard deviation, row standard deviation, maximum value, sum of column variance, sum of main diagonal variance, sum of column kurtosis, and sum of row skewness from the time-frequency domain feature matrix, a total of 7 eigenvalues ​​as generalized time-frequency domain features;

[0017] 3. Use the local binary pattern (1D-LBP) and gray-level co-occurrence matrix (GLCM) texture feature extraction method to extract the local features of the vibration signal;

[0018] 3.1 First, the signal is processed using the local binary pattern 1D-LBP to obtain the local changes of the signal. This step is called local binary coding;

[0019] 3.2 Obtain the gray-level co-occurrence matrix of the signal after 1D-LBP encoding through the GLCM algorithm, and extract the angular second-order moment, entropy, correlation, contrast and homogeneity of the gray-level co-occurrence matrix as the texture features of the signal;

[0020] Fourth, the features are further screened using maximum correlation and minimum redundancy to obtain a feature set;

[0021] 5. Combining the kernel extreme learning machine classifier with GL-mRMR feature extraction, the GL-mRMR-KELM model applied to circuit breaker vibration data classification is obtained.

[0022] The purpose of the present invention can also be further achieved by the following technical measures:

[0023] In the aforementioned circuit breaker fault diagnosis method based on global-local feature extraction, in step 2:

[0024] The mean, standard deviation Std, root mean square RMS, skewness, kurtosis, peak-to-peak value, impulse factor, and shape factor of the time domain feature types are as follows;

[0025] The mean μ is the average value of the data set, calculated as:

[0026]

[0027] n represents the number of data points, x iThe ith data point.

[0028] The standard deviation Std is the degree of dispersion between the data points and the mean, which can measure the volatility of the data. The calculation formula is:

[0029]

[0030] The root mean square RMS is a statistic that measures the size of a data set. The calculation formula is:

[0031]

[0032] Skewness can measure the symmetry of data distribution and reflect whether the data is tilted to one side. The calculation formula is:

[0033]

[0034] Kurtosis measures the peak degree of data distribution and reflects the concentration of extreme values ​​of data. The calculation formula is:

[0035]

[0036] The peak-to-peak value PP is the difference between the maximum and minimum values ​​in the data set, reflecting the maximum fluctuation amplitude of the signal. The calculation formula is:

[0037] PP=x max -x min (8)

[0038] x max Indicates the maximum value of the data set, x min Represents the minimum value in the data set.

[0039] Crest Factor is the ratio of peak value to root mean square, usually used to describe the sharpness of the signal. The calculation formula is:

[0040]

[0041] Shape Factor is the ratio of the root mean square to the absolute average value (or average amplitude), which is used to characterize the complexity of the signal shape. The calculation formula is:

[0042]

[0043] The global frequency domain features center frequency (CF), mean square frequency (MSF), root mean square frequency (RMSF), frequency variance (VF), frequency standard deviation, frequency amplitude median, frequency maximum, and frequency median are as follows:

[0044] The center frequency CF is the weighted average frequency of the signal spectrum, calculated as:

[0045]

[0046] Where f i is the ith frequency component, n is the number of frequency components, |X(f i )| 2 Represents the power spectral density of the signal at the i-th frequency;

[0047] The mean square frequency (MSF) is the weighted average of the square of the frequency and is used to characterize the frequency distribution of signal energy. The calculation formula is:

[0048]

[0049] The RMS frequency is the square root of the mean square frequency and reflects the standard deviation of the signal frequency distribution. Calculation formula:

[0050]

[0051] Frequency variance VF indicates the breadth of signal energy distribution in the frequency domain and reflects the discreteness of the spectrum. The calculation formula is:

[0052]

[0053] The frequency standard deviation is the square root of the frequency variance and is used to measure the degree of dispersion of the frequency distribution. The calculation formula is:

[0054]

[0055] The frequency amplitude median is the median of the frequency component amplitude, which is used to describe the central trend of the frequency component amplitude. The calculation formula is:

[0056]

[0057] The maximum frequency is the frequency at which the maximum amplitude of the frequency domain signal is located, indicating the highest frequency component of the signal. The calculation formula is:

[0058] f max = arg max(f i ,|X(f i )|) (17)

[0059] |X(f i )| is the frequency f i The corresponding frequency component amplitude;

[0060] The median frequency is the frequency that divides the power spectrum of the frequency domain signal into two parts with equal energy. It represents the middle frequency position of the signal energy, and the calculation formula is:

[0061]

[0062] f m is the median frequency;

[0063] The sum of the column standard deviation, the sum of the row standard deviation, the maximum value, the sum of the column variance, the sum of the main diagonal variance, the sum of the column kurtosis, and the sum of the row skewness of the time-frequency domain feature matrix are as follows:

[0064] The sum of the column standard deviations H1 is to calculate the standard deviations of all columns and sum them up. The calculation formula is:

[0065]

[0066] n is the number of matrix rows, m is the number of matrix columns, x ij represents the element in the i-th row and j-th column of the matrix, μ j represents the mean of the jth column;

[0067] The sum of the row standard deviations H2 is calculated by summing up the standard deviations of all rows. The calculation formula is:

[0068]

[0069] μ i represents the mean of the i-th row;

[0070] The maximum value H3 is the maximum value of all elements in the matrix, calculated by:

[0071] H3=max(x ij ) (twenty one)

[0072] The sum of column variances H4 is the sum of the variances of each column, calculated using the formula:

[0073]

[0074] The sum of the main diagonal variances H5 is the sum of the variances of the main diagonal elements, and the calculation formula is:

[0075]

[0076] The sum of the column kurtosis H6 is the steepness of the data distribution, calculated as:

[0077]

[0078] The sum of the row skewness H7 is the degree of skewness of the data distribution, calculated by:

[0079]

[0080] In the aforementioned circuit breaker fault diagnosis method based on global-local feature extraction, in step three:

[0081] The formula for LBP to process one-dimensional data is as follows:

[0082]

[0083] x[i] represents the sampling point; P represents the neighborhood size; r represents the index from 0 to P / 2-1, which is used to indicate the relative position;

[0084] The length of the original signal is set to N, and the P+1 points centered at a certain sampling point x[i] are thresholded. The range of the original signal after processing is [P / 2, NP / 2]. The difference between the sample points is converted into a P-bit binary code through the Sign function S[x]. At the same time, the binomial weight is used to convert the binary code into a range of [0,2 P ]; where S[x] is set as follows:

[0085]

[0086] GLCM creates a grayscale image and extracts features from it based on the distance (D) between pixels, the pixel angle (θ) and the grayscale level (0-64). GLCM rescales the one-dimensional grayscale data according to the determined number of grayscale levels, and then assigns the sum of the number of neighboring pixels at the specified angle and distance in the scaled data with the specified grayscale to the GLCM matrix.

[0087] Rescale the image grayscale from 256 to 0-64 to reduce the amount of calculation. The calculation formula is:

[0088]

[0089] Select the distance D and angle θ of adjacent pixels to define the positional relationship of pixel pairs. Create a G×G two-dimensional matrix P, where G is the number of gray levels after rescaling. Each element P(i,j) in the matrix is ​​used to store the number of times a pixel with gray level i appears with a pixel with gray level j at a given distance and direction. For each pixel in the image, find the neighboring pixels of the point based on the selected distance D and angle θ, and count their gray level relationships. Normalize the GLCM matrix so that the value of each element represents the probability of co-occurrence of the gray level pair, as shown in the following formula:

[0090]

[0091] Among the gray-level co-occurrence matrix features, the calculation formulas for angular second moment, entropy, correlation, contrast, and homogeneity are as follows:

[0092]

[0093] Where p(i,j) represents the value of the gray-level co-occurrence matrix, μ represents the corresponding value of the gray level, and σ represents the standard deviation of the gray level of the GLCM matrix.

[0094] In the aforementioned circuit breaker fault diagnosis method based on global-local feature extraction, the step 4 includes the following steps:

[0095] 4.1 Correlation calculation

[0096] The mRMR method uses mutual information to measure correlation. The calculation formula of mutual information is as follows:

[0097]

[0098] Where x and y represent two variables, p(x, y) represents the joint probability density function, p(x) and p(y) are the probability densities of x and y respectively; I(x, y) represents the mutual information between x and y;

[0099] Set the circuit breaker fault type to c and the original feature vector to x i , then calculate x i The mutual information I(x i ,c), and then get the maximum correlation between features and categories and the minimum redundancy between features. The calculation formula is as follows:

[0100]

[0101] In formula (32), S is the circuit breaker fault feature set, |S| is the number of features, and the maximum correlation can filter out the features in the feature set that are highly correlated with category c. When there is dependency between features, or effective features are repeated due to high feature dimensions, the features become redundant, and the maximum correlation criterion will fail. It is necessary to filter the features for minimum redundancy.

[0102] 4.2 Feature Screening

[0103] The result of mRMR is the difference between D and R in formula (32), that is:

[0104] maxΦ(D,R),Φ=DR (33)

[0105] Finally, mRMR is applied to the feature selection task, and the first m features are selected according to the descending sorting of Φ.

[0106] In the aforementioned circuit breaker fault diagnosis method based on global-local feature extraction, the specific method of step 5 is as follows:

[0107] After feature extraction, sample features are classified by ELM:

[0108] Assume that the connection weights between the ELM hidden layer and the input layer and the output layer are ω and β respectively, set the bias of the hidden layer to b, and the activation function to f(x). The ELM model is expressed as:

[0109]

[0110] In formula (9), y j is the network output, H is the hidden layer output, the network output is T, and the matrix form of ELM is expressed as:

[0111] Hβ=T T (35)

[0112] The output weight β can be obtained through the least squares solution, and its solution is expressed as:

[0113] β=H T (HH T ) -1 T (36);

[0114] In the extreme learning machine with kernel function, the kernel function is used as an activation function to realize the nonlinear mapping from input to output. Therefore, the kernel function is the key to the network's nonlinear classification ability. The choice of kernel function needs to satisfy Mercer's theorem. i ,y i ) is expressed as a kernel function, which is a mixed kernel function combining the Gaussian kernel function and the sigmoid function. Its formula is as follows:

[0115]

[0116] A regularization penalty term is introduced into the kernel function to improve the nonlinear level of the kernel function, so that it has better mapping ability for non-periodic and drastically changing data, thereby improving the versatility of the classification model.

[0117] Compared with the prior art, the present invention has the following beneficial effects:

[0118] The present invention studies the fault state diagnosis of the circuit breaker operating mechanism based on vibration signals. Aiming at the problems of low versatility and poor accuracy under small samples in the current operating mechanism state diagnosis algorithm, the GL-mRMR-KELM fault diagnosis model of high-voltage circuit breakers is constructed by using a feature extraction method combining global features with local texture features based on gray-level co-occurrence matrix. The following technical effects are achieved:

[0119] 1) For the early failure of the circuit breaker mechanical mechanism, the proposed GL-mRMR-KELM fault diagnosis model can effectively distinguish different working conditions through the combined feature extraction of global features and local features;

[0120] 2) Using mRMR for feature selection has a significant effect on filtering redundant features. Compared with the original feature space, its feature distribution is significantly improved.

[0121] The present invention is used in high-voltage circuit breaker fault diagnosis, and can comprehensively extract signal features and thereby complete the diagnosis. The feature vector extracted by the method has better intra-class compactness and inter-class separability. The present invention first extracts global features from a one-dimensional vibration signal, further performs local binary processing on the one-dimensional signal, obtains a grayscale co-occurrence matrix from the processed signal, and finally extracts texture features of the grayscale co-occurrence matrix as local features. In the feature extraction stage, global features and local features are processed separately to reduce the need for obtaining detailed features through modal decomposition, thereby improving the versatility of the feature extraction method. BRIEF DESCRIPTION OF THE DRAWINGS

[0122] Figure 1 It is the vibration data when the circuit breaker is normally opened and closed; Figure 1 (a) is the closing state, Figure 1 (b) is the open state;

[0123] Figure 2 It is the spectrum diagram when the closing signal is extracted by generalized S variation; Figure 2 (a) is the normal closing state, Figure 2 (b) is the state where the closing electromagnet is loose. Figure 2 (c) is the abnormal state of the opening spring output. Figure 2 (d) The oil buffer fastening bolts are loose;

[0124] Figure 3 It is a display diagram of the gray-level co-occurrence matrix of different types of data; Figure 3 (a) is normal closing, Figure 3 (b) The circuit breaker is closed when the closing electromagnet is loose. Figure 3 (c) is normal opening. Figure 3 (d) The trip spring output is abnormal;

[0125] Figure 4 It is a flow chart of the GL-mRMR-KELM method of the present invention. DETAILED DESCRIPTION

[0126] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0127] The state diagnosis of the circuit breaker operating mechanism requires comprehensive analysis of the features of the entire process. The present invention proposes a multi-level global-local feature extraction method (GL-mRMR) for the vibration signal of the circuit breaker operating mechanism under abnormal state. The features are divided into global features and local features for independent extraction, and then the features are screened using the maximum relevance minimum redundancy method to finally obtain a feature set for classification.

[0128] 1. Collect vibration signals of circuit breakers under normal working conditions and fault conditions as raw data:

[0129] 1.1 Collect normal status data;

[0130] When the circuit breaker operating mechanism is in action, the vibration signal will change dramatically in a short time, such as Figure 1 shown.

[0131] During the startup phase, the frequency and amplitude of the vibration signal suddenly change, and a high-energy signal appears in a short period of time; during the acceleration phase, a high-frequency signal appears again, reaching the second peak; during the deceleration phase, the frequency and amplitude decrease. If abnormal vibration occurs at this time, it may be caused by damage, jamming or looseness of the mechanism parts;

[0132] 1.2 Collect vibration data under various fault conditions, including loose closing electromagnet, abnormal opening spring output, and loose oil buffer fastening bolts;

[0133] 1.3 Construct the circuit breaker mechanical state fault data set through the collected fault data and normal data.

[0134] Second, the original time domain, frequency domain and time-frequency features of the generalized S transform matrix of the vibration signal of the circuit breaker operating mechanism under different states are extracted and regarded as global features. The specific implementation steps are as follows:

[0135] By using hyperbolic functions to improve the window function and realize the generalized S transform, a higher frequency resolution can be provided for high-frequency components, and a higher time resolution can be provided for low-frequency components. This has a better processing effect on non-periodic and unstable signals such as the vibration signal of the circuit breaker operating mechanism. The calculation formula is:

[0136]

[0137] Where V() is the window function, V(τ-t,{γ f ,γ b λ 2 The main function of}) is to generate an adaptive window so that it can adjust the time-frequency resolution according to the frequency characteristics of the signal (controlled by γf and γb), and the window shape is also adjusted according to λ and the hyperbolic function.

[0138] In formula (1), S represents the result obtained after S transformation, h(t) is the vibration signal to be processed, f is the signal frequency, τ is the signal offset, and γ f and γ b are the forward and backward taper parameters, λ is the positive curvature parameter, and ξ is the translation factor used to make the signal peak appear at τ = t. The calculation formula for ξ is:

[0139]

[0140] γ f , γ b And the parameter selection of λ is determined based on the empirical parameter method (the parameter selection based on the empirical parameter method refers to selecting the appropriate parameter value through experience or observation in the setting of the model or algorithm). In the embodiment, the parameter is determined to be γ by the empirical parameter method. f =0.5,γ b =1.32,λ=0.002, the feature extraction effect is as follows Figure 2 shown.

[0141] In the selection of global features, the collected circuit breaker vibration data is converted into a multi-domain data set, and feature extraction is performed on the time domain, frequency domain, and time-frequency domain data. In the time domain, there are: mean, standard deviation (Std), root mean square (RMS), skewness, kurtosis, peak-to-peak value, pulse factor, and shape factor; frequency domain features include: center frequency (CF), mean square frequency (MSF), root mean square frequency (RMSF), frequency variance (VF), frequency standard deviation, median frequency amplitude, maximum frequency, and median frequency. The extraction of time-frequency domain features is achieved with the help of generalized S transform (Stockwel transform). The S transform combines short-time Fourier transform and wavelet transform, overcomes the problem of fixed window of short-time Fourier transform, and has the multi-resolution analysis capability of wavelet transform.

[0142] 2.1 First, extract the global features of the collected data set in the original state. Such features belong to the global features in the time domain. The specific types of time domain features collected include mean, standard deviation (Std), root mean square (RMS), skewness, kurtosis, peak-to-peak value, impulse factor, and shape factor;

[0143] The mean μ is the average value of the data set, calculated as:

[0144]

[0145] n represents the number of data points, x i The ith data point.

[0146] The standard deviation Std is the degree of dispersion between the data points and the mean, which can measure the volatility of the data. The calculation formula is:

[0147]

[0148] The root mean square RMS is a statistic that measures the size of a data set. The calculation formula is:

[0149]

[0150] Skewness can measure the symmetry of data distribution and reflect whether the data is tilted to one side. The calculation formula is:

[0151]

[0152] Kurtosis measures the peak degree of data distribution and reflects the concentration of extreme values ​​of data. The calculation formula is:

[0153]

[0154] The peak-to-peak value PP is the difference between the maximum and minimum values ​​in the data set, reflecting the maximum fluctuation amplitude of the signal. The calculation formula is:

[0155] PP=x max -x min (8)

[0156] x max Indicates the maximum value of the data set, x min Represents the minimum value in the data set.

[0157] Crest Factor is the ratio of peak value to root mean square, usually used to describe the sharpness of the signal. The calculation formula is:

[0158]

[0159] Shape Factor is the ratio of the root mean square to the absolute average value (or average amplitude), which is used to characterize the complexity of the signal shape. The calculation formula is:

[0160]

[0161] 2.2 Perform Fourier transform on the original one-dimensional vibration data to obtain the signal frequency domain data set, and extract the global frequency domain features of the transformed frequency domain data. The specific extracted features are: center frequency (CF), mean square frequency (MSF), root mean square frequency (RMSF), frequency variance (VF), frequency standard deviation, frequency amplitude median, frequency maximum value, and frequency median value;

[0162] The center frequency CF is the weighted average frequency of the signal spectrum, calculated as:

[0163]

[0164] Where f i is the ith frequency component, n is the number of frequency components, |X(f i )| 2 Represents the power spectral density (energy of frequency components) of the signal at the i-th frequency.

[0165] The mean square frequency (MSF) is the weighted average of the square of the frequency and is used to characterize the frequency distribution of signal energy. The calculation formula is:

[0166]

[0167] The RMS frequency is the square root of the mean square frequency and reflects the standard deviation of the signal frequency distribution. Calculation formula:

[0168]

[0169] Frequency variance VF indicates the breadth of signal energy distribution in the frequency domain and reflects the discreteness of the spectrum. The calculation formula is:

[0170]

[0171] The frequency standard deviation is the square root of the frequency variance and is used to measure the degree of dispersion of the frequency distribution. The calculation formula is:

[0172]

[0173] The frequency amplitude median is the median of the frequency component amplitude, which is used to describe the central trend of the frequency component amplitude. The calculation formula is:

[0174]

[0175] The maximum frequency is the frequency at which the maximum amplitude of the frequency domain signal is located, indicating the highest frequency component of the signal. The calculation formula is:

[0176] f max = arg max(f i ,|X(f i )|) (17)

[0177] |X(f i )| is the frequency f i The corresponding frequency component amplitude.

[0178] The median frequency is the frequency that divides the power spectrum of the frequency domain signal into two parts with equal energy. It represents the middle frequency position of the signal energy, and the calculation formula is:

[0179]

[0180] f m is the median frequency.

[0181] 2.3 Perform a generalized S transform on the original one-dimensional vibration data to obtain the signal time-frequency domain feature matrix data set, and extract the sum of column standard deviations, row standard deviations, maximum value, sum of column variances, sum of main diagonal variances, sum of column kurtosis, and sum of row skewness from the time-frequency domain feature matrix, a total of 7 eigenvalues ​​as generalized time-frequency domain features.

[0182] The sum of the column standard deviations H1 is to calculate the standard deviations of all columns and sum them up. The calculation formula is:

[0183]

[0184] n is the number of matrix rows, m is the number of matrix columns, x ij represents the element in the i-th row and j-th column of the matrix, μ j represents the mean of the j-th column.

[0185] The sum of the row standard deviations H2 is calculated by summing up the standard deviations of all rows. The calculation formula is:

[0186]

[0187] μ i represents the mean of the i-th row.

[0188] The maximum value H3 is the maximum value of all elements in the matrix, calculated by:

[0189] H3=max(x ij ) (twenty one)

[0190] The sum of column variances H4 is the sum of the variances of each column, calculated using the formula:

[0191]

[0192] The sum of the main diagonal variances H5 is the sum of the variances of the main diagonal elements, and the calculation formula is:

[0193]

[0194] The sum of the column kurtosis H6 is the steepness of the data distribution, calculated as:

[0195]

[0196] The sum of the row skewness H7 is the degree of skewness of the data distribution, calculated by:

[0197]

[0198] 3. Use one-dimensional local binary pattern (1D-LBP) and gray-level co-occurrence matrix (GLCM) texture feature extraction method to extract local features of vibration signals separately;

[0199] The acquisition of local features is carried out in three steps. First, one-dimensional local binary patterns (1D-LBP) are used to obtain a one-dimensional signal reflecting local change characteristics; second, the gray-level co-occurrence matrix (GLCM) of the signal is extracted; finally, the matrix features are extracted as local features.

[0200] 3.1 First, the signal is processed using 1D-LBP to obtain the local changes of the signal. This step is called local binary coding;

[0201] The formula for LBP to process one-dimensional data is as follows:

[0202]

[0203] x[i] represents the sampling point; P represents the neighborhood size; r represents the index from 0 to P / 2-1, which is used to indicate the relative position;

[0204] The length of the original signal is set to N, and the P+1 points centered at a certain sampling point x[i] are thresholded. The range of the original signal after processing is [P / 2, NP / 2]. The difference between the sample points is converted into a P-bit binary code through the Sign function S[x]. At the same time, the binomial weight is used to convert the binary code into a range of [0,2 P ]; where S[x] is set as follows:

[0205]

[0206] 3.2 The gray level co-occurrence matrix of the signal after 1D-LBP encoding is obtained through the GLCM algorithm, and the angular second moment (energy), entropy, correlation, contrast and homogeneity of the gray level co-occurrence matrix are extracted (see formula (5)) as the texture features of the signal.

[0207] The vibration of the circuit breaker is highly instantaneous, and a small fault may cause abnormal fluctuations in the local signal. The LBP method can be used to obtain a description of the local data differences of the sample, and then the GLCM is used to obtain the gray-level co-occurrence matrix from the LBP signal to achieve the effect of amplifying the local data features.

[0208] GLCM creates a grayscale image and extracts features from it based on the distance (D) between pixels, the pixel angle (θ) and the grayscale level (0-64). GLCM rescales the one-dimensional grayscale data according to the determined number of grayscale levels, and then assigns the sum of the number of neighboring pixels at the specified angle and distance in the scaled data with the specified grayscale to the GLCM matrix. Here, GLCM processes one-dimensional grayscale data, so there is no angle information.

[0209] Rescale the image grayscale from 256 to 0-64 to reduce the amount of calculation. The calculation formula is:

[0210]

[0211] Select the distance D and angle θ of adjacent pixels to define the positional relationship of pixel pairs. Create a G×G two-dimensional matrix P, where G is the number of gray levels after rescaling. Each element P(i,j) in the matrix is ​​used to store the number of times a pixel with gray level i appears with a pixel with gray level j at a given distance and direction. For each pixel in the image, find the neighboring pixels of the point based on the selected distance D and angle θ, and count their gray level relationships. Normalize the GLCM matrix so that the value of each element represents the probability of co-occurrence of the gray level pair, as shown in the following formula:

[0212]

[0213] Figure 3 The image display of the gray level co-occurrence matrix calculated from two sets of data with different fault status types is shown. The LBP parameter P is selected as 6 and the GLCM gray level is 64.

[0214] In the study of gray-level co-occurrence matrix features, the angular second-order moment, namely energy, entropy, correlation, contrast and homogeneity, is the key point of attention. The calculation formula is as follows:

[0215]

[0216] Where p(i,j) represents the value of the gray-level co-occurrence matrix, μ represents the corresponding value of the gray level, and σ represents the standard deviation of the gray level of the GLCM matrix.

[0217] Fourth, the features are further screened using maximum correlation and minimum redundancy (mRMR) to obtain a feature set;

[0218] 4.1 Correlation calculation

[0219] The mRMR method uses mutual information to measure correlation. The calculation formula of mutual information is as follows:

[0220]

[0221] Where x and y represent two variables, p(x, y) represents the joint probability density function, p(x) and p(y) are the probability densities of x and y respectively; I(x, y) represents the mutual information between x and y.

[0222] Set the circuit breaker fault type to c and the original feature vector to x i , then calculate x i The mutual information I(x i ,c), and then get the maximum correlation between features and categories and the minimum redundancy between features. The calculation formula is as follows:

[0223]

[0224] In formula (32), S is the circuit breaker fault feature set, |S| is the number of features, and the maximum correlation can filter out the features in the feature set that are highly correlated with category c. When there is dependency between features, or effective features are repeated due to high feature dimensions, the features become redundant, the maximum correlation criterion will fail, and the features need to be screened for minimum redundancy.

[0225] 4.2 Feature Screening

[0226] The result of mRMR is the difference between D and R in formula (32), that is:

[0227] maxΦ(D,R),Φ=DR (33)

[0228] Finally, mRMR is applied to the feature selection task, and the first m features are selected according to the order of Φ from large to small. In the embodiment of the present invention, the number of retained features is set to 32. Experiments show that the diagnostic effect is best when this number of features is retained.

[0229] 5. Combining the kernel extreme learning machine (KELM) (kernel extreme learning machine is an improved algorithm based on extreme learning machine, which introduces kernel function to improve the nonlinear fitting ability of the model) classifier with GL-mRMR feature extraction, a GL-mRMR-KELM model for circuit breaker vibration data classification is provided.

[0230] All hidden layer node weights and biases in the extreme learning machine (ELM) network are randomly generated, and the output weights are obtained by batch least squares analysis. In this embodiment, after feature extraction, sample features are classified by ELM.

[0231] Assume that the connection weights between the ELM hidden layer and the input layer and the output layer are ω and β respectively, set the bias of the hidden layer to b, and the activation function to f(x). The ELM model is expressed as:

[0232]

[0233] In formula (9), y j is the network output, H is the hidden layer output, the network output is T, and the matrix form of ELM is expressed as:

[0234] Hβ=TT (35)

[0235] The output weight β can be obtained through the least squares solution, and its solution is expressed as:

[0236] β=H T (HH T ) -1 T (36)

[0237] The above content is the extreme learning machine. The kernel function is introduced below. The two constitute the kernel extreme learning machine.

[0238] In the extreme learning machine with kernel function, namely kernel extreme learning machine (KELM), kernel function is used as activation function to realize nonlinear mapping from input to output, so kernel function is the key to the network's nonlinear classification ability. The choice of kernel function needs to satisfy Mercer's theorem. i ,y i ) is expressed as a kernel function. In the present invention, a mixed kernel function combining a Gaussian kernel function and a sigmoid function is used, and its formula is as follows:

[0239]

[0240] A regularization penalty term is introduced into the kernel function to improve the nonlinear level of the kernel function, so that it has better mapping ability for non-periodic and drastically changing data, thereby improving the versatility of the classification model.

[0241] In general, the present invention provides a new fault feature extraction method called GL-mRMR, and combines the KELM classifier GL-mRMR to propose a complete circuit breaker fault diagnosis method based on global local feature extraction. The complete model is as follows Figure 4 shown.

[0242] In addition to the above embodiments, the present invention may also have other implementation modes. Any technical solutions formed by equivalent replacement or equivalent transformation shall fall within the protection scope required by the present invention.

Claims

1. A circuit breaker fault diagnosis method based on global-local feature extraction, characterized in that: The following steps are involved:

1. Collect vibration signals of circuit breakers under normal working conditions and fault conditions as raw data: 1.1 Collect normal status data; 1.2 Collect vibration data under various fault conditions, including loose closing electromagnet, abnormal opening spring output, and loose oil buffer fastening bolts; 1.3 Construct the circuit breaker mechanical state fault data set through the collected fault data and normal data; Second, the original time domain, frequency domain and time-frequency features of the generalized S transform matrix of the vibration signal of the circuit breaker operating mechanism under different states are extracted and regarded as global features. The specific implementation steps are as follows: 2.1 First, extract the global features of the collected data set in the original state. Such features belong to the global features in the time domain. The specific types of time domain features collected include mean, standard deviation, root mean square, skewness, kurtosis, peak-to-peak value, impulse factor, and shape factor; 2.2 Perform Fourier transform on the original one-dimensional vibration data to obtain the signal frequency domain data set, and extract the global frequency domain features of the transformed frequency domain data. The specific extracted features are: center frequency, mean square frequency, root mean square frequency, frequency variance, frequency standard deviation, frequency amplitude median, frequency maximum value, and frequency median value; 2.3 Perform generalized S transform on the original one-dimensional vibration data to obtain the signal time-frequency domain feature matrix data set, and extract the sum of column standard deviation, row standard deviation, maximum value, sum of column variance, sum of main diagonal variance, sum of column kurtosis, and sum of row skewness from the time-frequency domain feature matrix, a total of 7 eigenvalues ​​as generalized time-frequency domain features; 3. Extract local features of vibration signals using texture feature extraction method of local binary pattern combined with gray-level co-occurrence matrix; 3.1 First, the signal is processed using the local binary pattern 1D-LBP to obtain the local changes of the signal. This step is called local binary coding; 3.2 Obtain the gray-level co-occurrence matrix of the signal after 1D-LBP encoding through the GLCM algorithm, and extract the angular second-order moment, entropy, correlation, contrast and homogeneity of the gray-level co-occurrence matrix as the texture features of the signal; Fourth, the features are further screened using maximum correlation and minimum redundancy to obtain a feature set; 5. Combining the kernel extreme learning machine classifier with GL-mRMR feature extraction, the GL-mRMR-KELM model applied to circuit breaker vibration data classification is obtained.

2. The circuit breaker fault diagnosis method based on global-local feature extraction according to claim 1, characterized in that: In the step 2: The mean, standard deviation, root mean square, skewness, kurtosis, peak-to-peak value, impulse factor, and shape factor of the time domain feature types are as follows; The mean μ is the average value of the data set, calculated as: n represents the number of data points, x i The i-th data point; The standard deviation Std is the degree of dispersion between the data points and the mean, which can measure the volatility of the data. The calculation formula is: The root mean square RMS is a statistic that measures the size of a data set. The calculation formula is: Skewness can measure the symmetry of data distribution and reflect whether the data is tilted to one side. The calculation formula is: Kurtosis measures the peak degree of data distribution and reflects the concentration of extreme values ​​of data. The calculation formula is: The peak-to-peak value PP is the difference between the maximum and minimum values ​​in the data set, reflecting the maximum fluctuation amplitude of the signal. The calculation formula is: P-P=x max -x min (8) x max Indicates the maximum value of the data set, x min Indicates the minimum value in the data set; Crest Factor is the ratio of peak value to root mean square, usually used to describe the sharpness of the signal. The calculation formula is: Shape Factor is the ratio of the root mean square to the absolute average value, which is used to characterize the complexity of the signal shape. The calculation formula is: The global frequency domain features center frequency CF, mean square frequency MSF, root mean square frequency RMSF, frequency variance VF, frequency standard deviation, frequency amplitude median, frequency maximum value, and frequency median value are as follows: The center frequency CF is the weighted average frequency of the signal spectrum, calculated as: Where f i is the ith frequency component, n is the number of frequency components, |X(f i ) 2 Represents the power spectral density of the signal at the i-th frequency; The mean square frequency (MSF) is the weighted average of the square of the frequency and is used to characterize the frequency distribution of signal energy. The calculation formula is: The RMS frequency is the square root of the mean square frequency and reflects the standard deviation of the signal frequency distribution. Calculation formula: Frequency variance VF indicates the breadth of signal energy distribution in the frequency domain and reflects the discreteness of the spectrum. The calculation formula is: The frequency standard deviation is the square root of the frequency variance and is used to measure the degree of dispersion of the frequency distribution. The calculation formula is: The frequency amplitude median is the median of the frequency component amplitude, which is used to describe the central trend of the frequency component amplitude. The calculation formula is: The maximum frequency is the frequency at which the maximum amplitude of the frequency domain signal is located, indicating the highest frequency component of the signal. The calculation formula is: f max =arg max(f i ,|X(f i )) (17) |X(f i )| is the frequency f i The corresponding frequency component amplitude; The median frequency is the frequency that divides the power spectrum of the frequency domain signal into two parts with equal energy. It represents the middle frequency position of the signal energy. The calculation formula is: f m is the median frequency; The sum of the column standard deviation, the sum of the row standard deviation, the maximum value, the sum of the column variance, the sum of the main diagonal variance, the sum of the column kurtosis, and the sum of the row skewness of the time-frequency domain feature matrix are as follows: The sum of the column standard deviations H1 is to calculate the standard deviations of all columns and sum them up. The calculation formula is: n is the number of matrix rows, m is the number of matrix columns, x ij represents the element in the i-th row and j-th column of the matrix, μ j represents the mean of the jth column; The sum of the row standard deviations H2 is calculated by summing up the standard deviations of all rows. The calculation formula is: μ i represents the mean of the i-th row; The maximum value H3 is the maximum value of all elements in the matrix, calculated by: H3=max(x ij ) (21) The sum of column variances H4 is the sum of the variances of each column, calculated using the formula: The sum of the main diagonal variances H5 is the sum of the variances of the main diagonal elements, and the calculation formula is: The sum of the column kurtosis H6 is the steepness of the data distribution, calculated as: The sum of the row skewness H7 is the degree of skewness of the data distribution, calculated by:

3. The circuit breaker fault diagnosis method based on global-local feature extraction according to claim 1, characterized in that: In the step three: The formula for LBP to process one-dimensional data is as follows: x[i] represents the sampling point; P represents the neighborhood size; r represents the index from 0 to P / 2-1, which is used to indicate the relative position; The length of the original signal is set to N, and the P+1 points centered at a certain sampling point x[i] are thresholded. The range of the original signal after processing is [P / 2, NP / 2]. The difference between the sample points is converted into a P-bit binary code through the Sign function S[x]. At the same time, the binomial weight is used to convert the binary code into a range of [0,2 P ]; where S[x] is set as follows: GLCM creates a grayscale image and extracts features from it based on the distance D between pixels, the pixel angle θ, and the grayscale level. GLCM rescales the one-dimensional grayscale data according to the determined number of grayscale levels, and then assigns the sum of the number of adjacent pixels at the specified angle and distance in the scaled data with the specified grayscale to the GLCM matrix. Rescale the image grayscale from 256 to 0-64 to reduce the amount of calculation. The calculation formula is: Select the distance D and angle θ of adjacent pixels to define the positional relationship of pixel pairs; Create a G×G two-dimensional matrix P, where G is the number of gray levels after rescaling; each element P(i,j) in the matrix is ​​used to store the number of times a pixel with gray level i appears with a pixel with gray level j at a given distance and direction; for each pixel in the image, find the neighboring pixels of the point according to the selected distance D and angle θ, and count their gray level relationships; normalize the GLCM matrix so that the value of each element represents the probability of co-occurrence of the gray level pair, as shown in the following formula: Among the gray-level co-occurrence matrix features, the calculation formulas for angular second moment, entropy, correlation, contrast, and homogeneity are as follows: Where p(i,j) represents the value of the gray-level co-occurrence matrix, μ represents the corresponding value of the gray level, and σ represents the standard deviation of the gray level of the GLCM matrix.

4. The circuit breaker fault diagnosis method based on global-local feature extraction according to claim 1, characterized in that: The step 4 includes the following steps: 4.1 Correlation calculation The mRMR method uses mutual information to measure correlation. The calculation formula of mutual information is as follows: Where x and y represent two variables, p(x, y) represents the joint probability density function, p(x) and p(y) are the probability densities of x and y respectively; I(x, y) represents the mutual information between x and y; Set the circuit breaker fault type to c and the original feature vector to x i , then calculate x i The mutual information I(x i ,c), and then get the maximum correlation between features and categories and the minimum redundancy between features. The calculation formula is as follows: In formula (32), S is the circuit breaker fault feature set, |S| is the number of features, and the maximum correlation can filter out the features in the feature set that are highly correlated with category c; when there is dependency between features, or effective features are repeated due to high feature dimension, the features are redundant, and the maximum correlation criterion will fail, and the features need to be screened for minimum redundancy; 4.2 Feature Screening The result of mRMR is the difference between D and R in formula (32), that is: maxΦ(D,R),Φ=DR (33) Finally, mRMR is applied to the feature selection task, and the first m features are selected according to the descending sorting of Φ.

5. The circuit breaker fault diagnosis method based on global-local feature extraction according to claim 1, characterized in that: The specific method of step 5 is as follows: After feature extraction, sample features are classified by ELM: Assume that the connection weights between the ELM hidden layer and the input layer and the output layer are ω and β respectively, set the bias of the hidden layer to b, and the activation function to f(x). The ELM model is expressed as: In formula (9), y j is the network output, H is the hidden layer output, the network output is T, and the matrix form of ELM is expressed as: Hβ=T T (35) The output weight β can be obtained through the least squares solution, and its solution is expressed as: β=H T (HH T ) -1 T (36); In the extreme learning machine with kernel function, the kernel function is used as an activation function to realize the nonlinear mapping from input to output. Therefore, the kernel function is the key to the network's nonlinear classification ability. The choice of kernel function needs to satisfy Mercer's theorem. i ,y i ) is expressed as a kernel function, which is a mixed kernel function combining the Gaussian kernel function and the sigmoid function. Its formula is as follows: A regularization penalty term is introduced into the kernel function to improve the nonlinear level of the kernel function, so that it has better mapping ability for non-periodic and drastically changing data, thereby improving the versatility of the classification model.

Images