Circuit breaker fault diagnosis method based on sound and vibration signal combination

Abstract

The invention discloses a circuit breaker fault diagnosis method based on sound and vibration signal joint analysis, and the method comprises the steps: obtaining a vibration signal and a sound signal of a circuit breaker under the conditions of normal switching-on, switching-on operation refusing, base loosening and connecting rod falling through a vibration sensor and a sound sensor; cEEMDAN decomposition is carried out on the vibration signals and the sound signals in the time domain, and then kurtosis and skewness are extracted through all modal components to serve as feature values. And then, fusing the characteristic values of the sound and vibration signals through a PSO-KCCA algorithm. And finally, taking the fused characteristic value as the input of the KNN, and carrying out fault diagnosis on the circuit breaker. According to the circuit breaker fault diagnosis method, blind source separation is carried out on sound signals of a circuit breaker through fast-kica, original sound signals are separated out, surrounding noise interference is removed, CEEMDAN decomposition is carried out on sound and vibration signals, kurtosis and skewness of the first five components are extracted, then fusion is carried out through a PSO-KCCA algorithm, circuit breaker sound-vibration fusion features are generated, and a circuit breaker fault diagnosis result is obtained. And a KNN classifier is trained. Sound and vibration signal features are fused in the most correlation direction in a high-dimensional space through a PSO-KCCA algorithm, and dimensionality reduction is performed, so that optimal feature fusion is obtained, feature values are more representative, the judgment accuracy can be improved, the fault-tolerant rate of circuit breaker fault diagnosis can be increased, and the fault diagnosis efficiency is improved. Moreover, the method can meet the requirements for the precision and speed of the fault diagnosis of the circuit breaker in operation on site, and achieves the quick and accurate recognition.

Description

Technical Field

[0001] The present invention relates to the technical field of electrical equipment fault detection, and in particular to a method for diagnosing circuit breaker faults. Background Art

[0002] High-voltage circuit breakers are important power equipment in power systems. They quickly cut off faults, protect power systems from damage, and ensure stable operation of power systems. If a circuit breaker fails, the loss caused is far greater than the value of the circuit breaker itself. Therefore, circuit breaker fault diagnosis is a hot topic in research.

[0003] At present, the fault diagnosis method of circuit breaker is mainly based on the research of single signals such as the current signal of the opening and closing coil, the mechanical vibration signal, and the sound signal. Among them, the vibration signal is the superposition of the impulse sub-wave of the operating mechanism. By adopting appropriate signal processing methods, the state-related information of some key components in the operating mechanism of the high-voltage circuit breaker can be mined from the vibration signal. The mechanical parts generate vibrations and send out sound signals, which also contain key information about the circuit breaker operation process. In order to extract relevant state information from the signal, it is necessary to select a suitable signal processing method. The signal processing method is mainly based on empirical wavelet transform, variational mode decomposition, empirical mode decomposition, Wigner distribution, etc. For the processed signal, the signal features are extracted to realize fault diagnosis. In order to use these features to better realize the fault diagnosis of circuit breaker, support vector machine and neural network based on machine learning are also used to greatly improve the accuracy of fault diagnosis.

[0004] At present, most of the online monitoring and fault identification technologies of mechanical characteristics of high-voltage circuit breakers are based on the monitoring results of a single characteristic quantity, so it is impossible to make a comprehensive evaluation of the various operating states of high-voltage circuit breakers. A single type of signal provides less state information and may cause large errors. Therefore, the diagnostic results obtained in many cases are unreliable, which in turn affects the reliable operation of the power system. Therefore, in order to improve the reliability of fault diagnosis, information fusion technology is introduced. Both vibration signals and sound signals contain important information about the circuit breaker operation process, and the two complement each other. Therefore, using information fusion technology to combine sound and vibration signals for circuit breaker fault diagnosis can improve the reliability of identification results. Summary of the invention

[0005] The purpose of the present invention is to provide a circuit breaker fault diagnosis method based on joint analysis of acoustic and vibration signals. First, the sound signal is blindly separated by the fast-kica algorithm to separate the circuit breaker sound signal. Secondly, the circuit breaker sound and vibration signals are decomposed by CEEMDAN respectively. Thirdly, the kurtosis and skewness of the first five components of the decomposed modal components are extracted as feature quantities. Then, the feature quantities of the extracted sound and vibration signals are fused by the PSO-KCCA algorithm, and the fused feature values ​​are used as the input of KNN to finally determine what state the vibration signal belongs to.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] A method for circuit breaker fault diagnosis based on joint analysis of acoustic and vibration signals is proposed. Vibration and sound sensors are used to obtain vibration and sound signals of the circuit breaker in the cases of normal closing, closing refusal, loose base and connecting rod falling off. The vibration and sound signals in the time domain are decomposed by CEEMDAN respectively, and then the kurtosis and skewness are extracted as eigenvalues ​​using each modal component. Then, the eigenvalues ​​of the sound and vibration signals are fused by the PSO-KCCA algorithm. Finally, the fused eigenvalues ​​are used as the input of KNN to diagnose the circuit breaker fault. This is a circuit breaker fault diagnosis method. The circuit breaker sound signal is blindly separated by fast-kica to separate the original sound signal and remove the surrounding noise interference. The sound and vibration signals are decomposed by CEEMDAN to extract the kurtosis and skewness of the first five components, and then fused by the PSO-KCCA algorithm to generate the circuit breaker sound-vibration fusion feature, which is used to train the KNN classifier. The PSO-KCCA algorithm is used to fuse the sound and vibration signal features in the most relevant direction in high-dimensional space and perform dimensionality reduction to obtain the optimal feature fusion, making the feature value more representative, which can improve the accuracy of judgment, increase the fault tolerance of circuit breaker fault diagnosis, and meet the accuracy and speed requirements of on-site circuit breaker fault diagnosis in operation, and quickly and accurately identify. The steps are as follows:

[0008] (1) Install the circuit breaker vibration sensor device and sound sensor, adjust the sensor parameters, collect the circuit breaker vibration signal and sound signal, and start the test and calculation analysis process.

[0009] (2) The collected circuit breaker sound signals are subjected to blind source separation using the fast-kica algorithm.

[0010] (3) The separated sound signal and vibration signal are decomposed by CEEMDAN, and the first five modal components are taken as the research objects. The kurtosis and skewness of the first five modal components are extracted as eigenvalues.

[0011] (4) The extracted kurtosis and skewness constitute the ten-dimensional features of the circuit breaker sound and vibration signals respectively. The PSO-KCCA algorithm is used to perform feature fusion and achieve the effect of dimensionality reduction to generate sound-vibration fusion features.

[0012] (5) The fused feature values ​​are used for pattern recognition using the KNN algorithm.

[0013] (6) The method is verified using the collected data samples to obtain the accuracy of fault diagnosis of the circuit breaker, which serves as a sign of the practicality of the method.

[0014] Step (1) Select AD50S piezoelectric acceleration sensor to collect vibration signals. The sensor is made of piezoelectric ceramics as piezoelectric materials and stainless steel as shell materials. The sensitivity is 50mV / g, the frequency response is 0.5-15000Hz, the installation resonance frequency is 27kHz, the range is 0-100g, the bias voltage is 8-12V, and the working environment temperature range is -40-120℃. The vibration sensor is installed on the outer shell of the circuit breaker box. The sound sensor required for collecting sound signals is WM-025N bullet-type waterproof sound sensor, and the main performance indicators are as follows: working voltage: 120dB SPL, frequency range: 85dB (1kHz at Max dB SPL) frequency response: 20Hz-20kHz, signal-to-noise ratio: 55dB (1 meter 40dB sound source) 35dB (10 meters 40dB sound source).

[0015] Step (2) performs blind source separation on the collected circuit breaker sound signal through the fast kernel independent component analysis (fast-KICA) algorithm, uses the Hilbert-Schmidt independence decision criterion as the objective function to measure the statistical independence of the variables, optimizes the objective function using the Newton method, minimizes the objective function, and obtains the separation matrix. The incomplete Cholesky decomposition method is used to improve the computing performance, eliminate redundant features, and reduce noise, and has strong nonlinear processing capabilities. The Gaussian radial basis function is used as the kernel function.

[0016] Step (3) performs CEEMDAN decomposition on the sound and vibration signals separated in step (2) respectively. CEEMDAN is an adaptive decomposition algorithm that effectively solves the problem of modal aliasing. White noise with standard normal distribution is added to the original signal, and the components obtained by multiple empirical mode decompositions are averaged as the final actual components. The problem of noise residual in the reconstructed signal can be effectively eliminated, and the problem of modal aliasing can be well handled.

[0017] Definition E j (·) is the jth modal component obtained by EMD decomposition, and the kth modal component obtained by CEEMDAN decomposition is recorded as Vi (t) represents the addition of a white noise sequence with a standard normal distribution in the i-th trial, ε j-1 Represents the signal-to-noise ratio coefficient of the jth modal component. The original signal s(t) plus the noise component constitutes s(t)+ε0V i (t) Carry out I tests. The first-order modal component is obtained by EMD decomposition And the residual signal r1(t) of the first stage is:

[0018]

[0019]

[0020] Then we can get the i-th residual signal r k (t) and the i+1th order modal component IMF k+1 (t) is:

[0021]

[0022]

[0023] Step (4) After step (3), the characteristic values ​​of the sound and vibration signals are extracted, and the dimensions are reduced by the proposed PSO-KCCA algorithm, and then fused in parallel. The KCCA algorithm mainly studies the correlation problem between two groups of random variables, and transforms the correlation study between the two groups of random variables into the correlation study between the few pairs of variables in the two groups that are unrelated to each other. The Gaussian radial basis function is used as the kernel function:

[0024]

[0025] The sample set is mapped to a high-dimensional space through a kernel function:

[0026]

[0027] According to the definition of kernel function, find the kernel matrix:

[0028]

[0029] According to the obtained kernel matrix, in order to avoid the influence of matrix singularity, the regularization parameter k is introduced, and the constraint condition can be written as:

[0030]

[0031] Construct a typical projection matrix W x and W y , and obtain the new fusion feature P through parallel fusion:

[0032] P=W xT X+W y T Y

[0033] In the process of feature fusion, not only the feature dimension is reduced, but also the fusion features with better discrimination are extracted by changing the distribution of features in space. Since the signals collected in the experiment are affected by the operating environment and equipment, the fixed parameters affect the recognition effect of the fusion features each time they are fused. The PSO-KCCA algorithm uses the PSO algorithm to optimize the KCCA regularization parameter K and the Gaussian radial basis function parameter σ, which greatly improves the efficiency of feature fusion.

[0034] Step (5) In many cases, when the data distribution is complex or the noise in the data is serious, there is a risk of error when making decisions based on the closest sample. The working environment of the circuit breaker is very complex and harsh, and there are noise and other influencing factors around it. An improved method is to introduce a voting mechanism to select the first several known samples closest to the sample and vote to determine the category of the new sample based on their category number. This is the k-nearest neighbor method (KNN). KNN can be expressed as: Suppose N known samples belong to c categories ω i , i = 1, ..., c, examine the first k neighbors in these samples, assuming k i ω i kind.

[0035] Then ω i The discriminant function is:

[0036] g i (x) = k i , i=1,…c

[0037] The decision rule is:

[0038] like Then x∈ω k

[0039] The distance measurement between samples can be calculated using Euclidean distance, Manhattan distance, Chebyshev distance, Minkowski distance, etc. Euclidean distance is:

[0040] δ(x i ,x j )=||x i -x j ||

[0041] The calculation formula of Minkowski distance is:

[0042]

[0043] Different p values ​​in the Minkowski distance formula represent different distance calculation methods: when p = 1, the formula is the Manhattan distance formula; when p = 2, formula (3) is the Euclidean formula. Similarly, in the distance calculation formula used in machine learning, the p value is generally a positive integer. When p is infinite, it is the Chebyshev distance. Depending on the actual problem, different distance metrics can be used, or the distance metric can be changed into a similarity metric.

[0044] Step (6) takes the test signal set and obtains the fault diagnosis accuracy through steps (1) to (5). BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 Circuit breaker fault diagnosis flow chart

[0046] Figure 2 CEEMDAN decomposition diagram of sound signal under normal conditions

[0047] Figure 3 CEEMDAN decomposition diagram of vibration signal under normal conditions

[0048] Figure 4 CEEMDAN breakdown diagram for closing refusal

[0049] Figure 5 CEEMDAN exploded view of the connecting rod falling off

[0050] Figure 6 Exploded view of CEEMDAN with loose base

[0051] Figure 7 Fault diagnosis recognition rate diagram of this method DETAILED DESCRIPTION

[0052] Taking the opening process of a ZN65-12 high-voltage circuit breaker as an example, the specific implementation of the present invention is as follows:

[0053] Step 1: Install the AD50S piezoelectric acceleration sensor on the circuit breaker housing and place the sound sensor 150 cm away from the left and right sides of the circuit breaker. Collect vibration and sound signals of the circuit breaker in normal, closing refusal, connecting rod detachment and base looseness. Collect 50 sets of signals for each case, for a total of 200 sets of data.

[0054] In the second step, the collected circuit breaker sound signal is blindly separated by the fast kernel independent component analysis (fast-KICA) algorithm. The Gaussian radial basis function is used as the kernel function. First, the observed data vector is whitened and preprocessed. Then, the Hessian matrix of the objective function is estimated by incomplete Cholesky. Finally, the separation matrix is ​​optimized and solved by the Newton-type method.

[0055] In the third step, the standard vibration and sound signals in the time domain are decomposed by CEEMDAN to calculate the power spectrum of each IMF component. The main peak frequency of the power spectrum of IMF1 to IMF5 with larger amplitudes is greater than 1kHz, which is the main component of the frequency structure of the original signal. The power spectrum frequency and amplitude drop sharply after IMF6, which is due to the significant attenuation of the vibration level after the signal propagates through the medium and the transmission of the operating mechanism. Therefore, the first five modal components are taken as the research objects to extract the kurtosis and skewness.

[0056] In the fourth step, the kurtosis and skewness features of each modal component of the extracted sound and vibration signals are fused through the PSO-KCCA algorithm. After fusion, the KNN algorithm is introduced for fault diagnosis. Twenty groups of samples (a total of 80 groups) of labeled data of normal, closing refusal, connecting rod detachment and base looseness are taken out from the 200 groups of data to train the PSO-KCCA algorithm and KNN classifier to determine the more appropriate regularization parameter K and kernel function σ value. The remaining 120 groups of data are used as test data sets to test the classification accuracy of the entire neural network.

[0057] In step 5, the distance metric of KNN uses Minkowski distance. The k value should be selected appropriately. Too large a k value may simplify the model and make it meaningless. Too small a k value may complicate the model and may be accompanied by overfitting. K is taken as an odd number, otherwise there may be an equal proportion situation, resulting in unsuccessful classification. Here, the cross-validation method is used. After the fourth step of training, the k value is finally determined to be 7. The KNN classifier can greatly improve the prediction efficiency while ensuring the prediction accuracy.

[0058] Step 6: The prediction accuracy of the fault diagnosis method is as follows: Figure 7 As shown in the figure, the accuracy rate can reach 99.7% in the case of closing refusal, and the accuracy rates of the other three types of normal, connecting rod falling off and base loosening are 98.6%, 98.9% and 97.8% respectively. It has practical significance.

Claims

1. A detection method for circuit breaker fault diagnosis based on joint analysis of acoustic and vibration signals, characterized in that: Includes steps: (1) Install the circuit breaker vibration sensor device and sound sensor, adjust the sensor parameters, collect the circuit breaker vibration signal and sound signal, and start the test and calculation analysis process. (2) The collected circuit breaker sound signals are subjected to blind source separation using the fast-kica algorithm. (3) The separated sound signal and vibration signal are decomposed by CEEMDAN, and the first five modal components are taken as the research objects. The kurtosis and skewness of the first five modal components are extracted as eigenvalues. (4) The extracted kurtosis and skewness constitute the ten-dimensional features of the circuit breaker sound and vibration signals respectively. The PSO-KCCA algorithm is used to perform feature fusion and achieve the effect of dimensionality reduction to generate sound-vibration fusion features. (5) The fused feature values ​​are used for pattern recognition using the KNN algorithm. (6) The method is verified using the collected data samples to obtain the accuracy of fault diagnosis of the circuit breaker, which serves as a sign of the practicality of the method.

2. The method for blind source separation of circuit breaker sound by fast-kica algorithm according to claim 1, characterized in that: The Hilbert-Schmidt independence criterion is used as the objective function to measure the statistical independence of variables. The objective function is optimized by the Newton method, and the objective function is minimized to obtain the separation matrix. The incomplete Cholesky decomposition method is used to improve the computing performance, eliminate redundant features, reduce noise, and has strong nonlinear processing capabilities. The Gaussian radial basis function is used as the kernel function.

3. The method for performing CEEMDAN decomposition on a sound signal and a vibration signal after blind source separation according to claim 1, characterized in that: The sound and vibration signals are decomposed by CEEMDAN respectively. CEEMDAN is an adaptive decomposition algorithm that effectively solves the problem of modal aliasing. White noise with standard normal distribution is added to the original signal, and the components obtained by multiple empirical mode decompositions are averaged as the final actual components. This can effectively eliminate the problem of noise residuals in the reconstructed signal and can handle the problem of modal aliasing well.

4. The method for fusing characteristic values ​​of sound and vibration signals according to claim 1, characterized in that: The ten-dimensional eigenvalues ​​of the sound and vibration signals are extracted, and the dimensionality is reduced by the proposed PSO-KCCA algorithm. The fusion is carried out in parallel, and the KCCA regularization parameter K and the Gaussian radial basis function parameter σ are optimized by the PSO algorithm, which greatly improves the efficiency of feature fusion. The KCCA algorithm mainly studies the correlation problem between two groups of random variables, and transforms the correlation study between the two groups of random variables into the correlation study between the few pairs of variables in the two groups that are not related to each other. The use of Gaussian radial basis function as the kernel function enables the eigenvalues ​​of the sound and vibration signals to be fused in the best way. The fused features are more targeted to the fault signals of the circuit breaker, which not only reduces the feature dimension, but also extracts the fusion features with better discrimination.

5. The method for classification using KNN according to claim 1, characterized in that: KNN's distance metric uses Minkowski distance. The k value should be selected appropriately. Too large a k value may simplify the model and make it meaningless. Too small a k value will complicate the model and may be accompanied by overfitting. K is taken as an odd number, otherwise an equal proportion may occur, resulting in unsuccessful classification. Here, the cross-validation method is used. After the fourth step of training, the k value is finally determined to be 7.

6. The method for fault diagnosis using fused features after feature fusion according to claim 1, characterized in that: The PSO-KCCA algorithm is used to fuse the kurtosis and skewness features of each modal component of the extracted sound and vibration signals. After fusion, the KNN algorithm is introduced for fault diagnosis. Twenty groups of samples (a total of 80 groups) of labeled data from the 200 groups of data, including normal, closing refusal, connecting rod detachment and base looseness, are taken out to train the PSO-KCCA algorithm and KNN classifier to determine the more appropriate regularization parameter K and kernel function σ value. The remaining 120 groups of data are used as test data sets to test the classification accuracy of the entire neural network.

Images