Power transformer winding mechanical fault diagnosis method based on CEEMDAN-BWO-SVM

Abstract

The invention provides a power transformer winding mechanical fault diagnosis method based on CEEMDAN-BWO-SVM. The method comprises the following steps: S1, obtaining a vibration acceleration signal of the outer side wall of a sample power transformer box body; s2, decomposing the vibration acceleration signal by adopting a CEEMDAN method to form N IMF components; s3, calculating a Pearson's correlation coefficient of each IMF component and the initial vibration acceleration signal, and screening characteristic components based on the Pearson's correlation coefficients; s4, calculating the exponential entropy of the feature component and forming a feature vector; s5, constructing a BWO-SVM prediction model, and inputting the feature vector into the BWO-SVM prediction model for training; and S6, obtaining a vibration acceleration signal of the outer side wall of the box body of the power transformer to be detected, processing the vibration acceleration signal of the power transformer to be detected through the steps S2 to S4 to form a feature vector to be detected, and inputting the feature vector to be detected into the trained BWO-SVM prediction model to obtain a mechanical fault diagnosis result.

Description

Technical Field

[0001] The invention relates to a method for diagnosing faults of electric equipment, and in particular to a method for diagnosing mechanical faults of electric transformer windings based on CEEMDAN-BWO-SVM. Background Art

[0002] In recent years, the power industry has developed rapidly, compact ultra-high voltage and ultra-high voltage transmission networks have been steadily built, and the voltage level and overall capacity of the power grid have been continuously improved, which has put forward more stringent technical requirements for the safety and reliability of the power system. As the core equipment for power conversion, the importance of power transformers has become increasingly prominent.

[0003] Winding mechanical failure is the main type of transformer failure. It is difficult to detect through conventional insulation tests and is a latent failure. As the transformer operation time increases, the short-circuit resistance of the winding insulation gradually decreases. After the short-circuit current impact, the winding is subjected to huge non-uniform electromotive force under the action of the internal magnetic field of the transformer, which damages the winding insulation structure, causing the winding to loosen and deform, directly affecting the performance and life of the transformer. Therefore, studying the winding mechanical fault diagnosis method is particularly critical to ensure the stable operation of the power system.

[0004] At present, the transformer winding mechanical fault diagnosis methods are mainly divided into electrical signal method and vibration analysis method. The electrical signal method is to identify the fault by measuring and comparing the electrical signals of the transformer winding mechanical fault and normal operation, including frequency response method, short-circuit impedance method and low-voltage pulse method. This method can directly reflect the operating status and performance parameters of the transformer and is applicable to a variety of power equipment. However, this method belongs to the offline fault diagnosis method, which will affect the continuous power supply of the power system, cannot perform real-time fault detection, and it is difficult to find potential fault hazards; and when the transformer winding is slightly loose or slightly deformed, the frequency response function, short-circuit impedance and output response signal measured based on the electrical signal change little, resulting in reduced fault diagnosis accuracy.

[0005] Therefore, in order to solve the above technical problems, it is urgent to propose a new technical means. Summary of the invention

[0006] In view of this, the purpose of the present invention is to provide a power transformer winding mechanical fault diagnosis method based on CEEMDAN-BWO-SVM. Based on the vibration acceleration signal of the power transformer, an adaptive noise complementary total empirical mode decomposition method is used to determine the IMF component and form a feature vector, and then a recognition model based on the combination of BWO algorithm and support vector machine classifier is constructed and the feature vector is used to train the recognition model and finally identify it, thereby comprehensively considering the complexity and uncertainty of the vibration acceleration signal, and without a large number of samples, effectively ensuring the accuracy of the final fault identification result, and providing guarantee for the stable operation of the power system.

[0007] The present invention provides a method for diagnosing mechanical faults of power transformer windings based on CEEMDAN-BWO-SVM, comprising the following steps:

[0008] S1. Obtain the vibration acceleration signal of the outer wall of the sample power transformer box;

[0009] S2. Use CEEMDAN method to decompose the vibration acceleration signal into N IMF components;

[0010] S3. Calculate the Pearson correlation coefficient between each IMF component and the initial vibration acceleration signal, and select the characteristic component based on the Pearson correlation coefficient;

[0011] S4. Calculate the exponential entropy of the characteristic component and form a characteristic vector;

[0012] S5. construct a BWO-SVM prediction model, and input the feature vector into the BWO-SVM prediction model for training;

[0013] S6. Obtain the vibration acceleration signal of the outer wall of the power transformer box to be tested, and process the vibration acceleration signal of the power transformer to be tested through steps S2 to S4 to form a feature vector to be tested, and input the feature vector to be tested into the trained BWO-SVM prediction model to obtain the mechanical fault diagnosis result.

[0014] Furthermore, the CEEMDAN method is used to decompose the vibration acceleration signal into N IMF components, including:

[0015] S21. Determine M Gaussian white noises with the same standard deviation n i (t), where i = 1, 2, ..., M;

[0016] S22. Add Gaussian white noise to the initial signal X(t), and then perform EMD decomposition on the initial signal X(t) to which the Gaussian white noise is added to obtain component IMF1;

[0017] S23. Calculate the residual component r1(t): r1(t)=X(t)-IMF1;

[0018] S24. Perform EMD decomposition on the original Gaussian white noise and convert the decomposed first-order component E1(n i (t)) is superimposed with the residual component r1(t) to obtain M synthetic signals, and the EMD decomposition is performed to obtain IMF2;

[0019] S25. Repeat steps S22 to S24 until the set number N is reached to obtain the component IMF N and the remaining component r N (t).

[0020] Further, step S3 specifically includes:

[0021] Calculate the Pearson correlation coefficient ρ n :

[0022]

[0023] Where: n = 1, 2, ..., N, cov (IMF n ,X) is the covariance between IMFn and the initial signal X(t), σ IMFn , σ X are the standard deviation of IMFn and the initial signal X(t), μ IMFn , μ X are the means of IMFn and the initial signal X(t) respectively;

[0024] Set the feature extraction threshold η; set ρ n The IMF components with >η are selected to form the characteristic component S k , k = 1, 2,…, K, K is the total number of characteristic components.

[0025] Further, step S4 specifically includes:

[0026]

[0027] Where: h k is the characteristic component S k The total number of sample points, P kj is the characteristic component S k The energy value of the jth sample point in S k Ratio of energy values;

[0028] Construct feature vector S: S = [H1, H2, ..., H K ].

[0029] Furthermore, the initial parameters of the BWO-SVM prediction model include the beluga population size n, the maximum number of iterations T maxand the number of optimization parameters d, as well as the initialization parameter penalty factor c and kernel function parameter g of the SVM classifier, where: c∈[0,100], g∈[-100,100].

[0030] The beneficial effects of the present invention are as follows: through the present invention, based on the vibration acceleration signal of the power transformer, an adaptive noise complementary overall empirical mode decomposition method is used to determine the IMF component and form a feature vector, and then a recognition model based on the combination of the BWO algorithm and the support vector machine classifier is constructed and the feature vector is used to train the recognition model and finally identify it, thereby comprehensively considering the complexity and uncertainty of the vibration acceleration signal, and without the need for a large number of samples, effectively ensuring the accuracy of the final fault identification result, and providing guarantee for the stable operation of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:

[0032] Figure 1 It is a flow chart of the present invention.

[0033] Figure 2 This is a schematic diagram of the distribution of acceleration sensors of the present invention.

[0034] Figure 3 This is the decomposition result diagram of the vibration signal of the power transformer winding deformation fault.

[0035] Figure 4 It is a bar chart of the Pearson correlation coefficient between each IMF component and the original signal. DETAILED DESCRIPTION

[0036] The present invention is further described in detail below:

[0037] The present invention provides a method for diagnosing mechanical faults of power transformer windings based on CEEMDAN-BWO-SVM, comprising the following steps:

[0038] S1. Obtain the vibration acceleration signal of the outer wall of the sample power transformer box; Figure 2 As shown in the figure, a piezoelectric acceleration sensor is selected and a vibration signal acquisition platform is built. The measuring points are evenly arranged on the surface of the transformer box. The excitation voltage is applied to the high-voltage side of the transformer. The low-voltage winding is short-circuited with a copper wire to measure the vibration signals of the power transformer in normal operation and in the winding deformation and loose fault states.

[0039] S2. The vibration acceleration signal is decomposed into N IMF components by using the CEEMDAN method; wherein the CEEMDAN method is the abbreviation of the adaptive noise complementary total empirical mode decomposition method;

[0040] S3. Calculate the Pearson correlation coefficient between each IMF component and the initial vibration acceleration signal, and select the characteristic component based on the Pearson correlation coefficient;

[0041] S4. Calculate the exponential entropy of the characteristic component and form a characteristic vector;

[0042] S5. construct a BWO-SVM prediction model, and input the feature vector into the BWO-SVM prediction model for training;

[0043] S6. Obtain the vibration acceleration signal of the outer wall of the power transformer box to be tested, and process the vibration acceleration signal of the power transformer to be tested through steps S2-S4 to form a feature vector to be tested, and input the feature vector to be tested into the trained BWO-SVM prediction model to obtain the mechanical fault diagnosis result. Through the above method, based on the vibration acceleration signal of the power transformer, the adaptive noise complementary overall empirical mode decomposition method is used to determine the IMF component and form a feature vector, and then a recognition model based on the combination of the BWO algorithm and the support vector machine classifier is constructed and the feature vector is used to train the recognition model and finally identify, so as to comprehensively consider the complexity and uncertainty of the vibration acceleration signal, and without a large number of samples, effectively guarantee the accuracy of the final fault identification result, and provide guarantee for the stable operation of the power system.

[0044] In this embodiment, the CEEMDAN method is used to decompose the vibration acceleration signal to form N IMF components, which specifically includes:

[0045] S21. Determine M Gaussian white noises with the same standard deviation n i (t), where i = 1, 2, ..., M;

[0046] S22. Add Gaussian white noise to the initial signal X(t), and then perform EMD decomposition on the initial signal X(t) to which the Gaussian white noise is added to obtain component IMF1;

[0047]

[0048] S23. Calculate the residual component r1(t): r1(t)=X(t)-IMF1;

[0049] S24. Perform EMD decomposition on the original Gaussian white noise and convert the decomposed first-order component E1(n i (t)) is superimposed with the residual component r1(t) to obtain M synthetic signals, and the EMD decomposition is performed to obtain IMF2;

[0050]

[0051] S25. Repeat steps S22 to S24 until the set number N is reached to obtain the component IMF N and the remaining component r N (t).

[0052]

[0053] In this embodiment, step S3 specifically includes:

[0054] Calculate the Pearson correlation coefficient ρ n :

[0055]

[0056] Where: n = 1, 2, ..., N, cov (IMF n ,X) is the covariance between IMFn and the initial signal X(t), σ IMFn , σ X are the standard deviation of IMFn and the initial signal X(t), μ IMFn , μ X are the means of IMFn and the initial signal X(t) respectively;

[0057] Set the feature extraction threshold η; set ρ n The IMF components with >η are selected to form the characteristic component S k , k = 1, 2,…, K, K is the total number of characteristic components.

[0058] In this embodiment, step S4 specifically includes:

[0059]

[0060] Where: h k is the characteristic component S k The total number of sample points, P kj is the characteristic component S k The energy value of the jth sample point in S k Ratio of energy values;

[0061] Construct feature vector S: S = [H1, H2, ..., H K ]. Before constructing the feature vector S, the exponential entropy needs to be normalized:

[0062]

[0063] In this embodiment, the initial parameters of the BWO-SVM prediction model include the beluga population size n, the maximum number of iterations T maxand the number of optimization parameters d, as well as the initialization parameter penalty factor c and kernel function parameter g of the SVM classifier, where: c∈[0,100], g∈[-100,100].

[0064] The BWO-SVM prediction model is a model that combines the White Whale optimization algorithm BWO with the support vector machine classifier SVM; the existing model can be used. The following is a brief description of the BWO algorithm:

[0065] Take the position of each beluga whale as a solution vector P i , the vector dimension d is the number of optimization parameters, then the position matrix P of the beluga population is as shown in formula (1), combined with the fitness function F P =f(p i,1 ,p i,2 ,…,p i,d ) can get the fitness matrix F of the beluga population P , where i = 1, 2,…, n.

[0066]

[0067] By balancing factor B f Realize the transformation between the exploration phase and the development phase, as shown in formula (2).

[0068]

[0069] Where T is the number of iterations of the current algorithm; B0 is a random number between 0 and 1 generated at each algorithm iteration; at the initial stage of the algorithm iteration, B f >0.5, the probability of the search agent entering the exploration phase is high. As the number of iterations increases, B f Gradually decreases, when B f When <0.5, the probability of the development stage increases significantly.

[0070] In the exploration phase, the paired swimming behavior of beluga whales is used to establish a mathematical model, and the position of the beluga whale is updated according to the parity of the position parameter in the jth dimension, as shown in formula (3).

[0071]

[0072] in, is the position parameter of the ith beluga whale in the jth dimension during the T+1th algorithm iteration; r is a random integer between 0 and n, used to extract the position parameter of the randomly selected beluga whale; x j is a random integer between 0 and d; r1 and r2 are random numbers between 0 and 1, sin(2πr2) and cos(2πr2) indicate that the two mirror-image beluga fins are facing the water surface; k is a positive integer.

[0073] In the development phase, the Levy flight strategy and the optimal beluga position are combined to simulate the cooperative predation behavior of beluga whales, as shown in formula (4).

[0074]

[0075] Among them, r3 and r4 are random numbers between 0 and 1; is the optimal position of the beluga individual; C1 is the random jump intensity of Levy flight, as shown in formula (5); L F is the Levy flight function, as shown in equations (6) and (7).

[0076]

[0077] Among them, u and v are random numbers that obey the normal distribution, and β is a constant of 1.5.

[0078] The BWO algorithm introduces the whale fall process to simulate the randomness of the optimization process and sets the white whale fall step length P step And calculate the whale fall probability W f , as shown in equations (8) and (9), when B f <W f When , the population position matrix is ​​updated based on formula (10).

[0079]

[0080] P i T+1 =r5P i T -r6P i T +r7P step (10)

[0081] Among them, r5 and r6 are random numbers between 0 and 1; u b and l b are the upper and lower limits of the optimization parameters; C2 is the step factor related to the fall of the white whale, as shown in formula (11).

[0082] C2=2W f n(11)

[0083] The model evaluation index is the same as the fitness function of the BWO model, which is the fault diagnosis accuracy δ, as shown in the following formula.

[0084]

[0085] Among them, q is the number of samples in which the model accurately identifies faults, and n is the total number of samples.

[0086] Taking the deformation fault state of power transformer winding as an example, the vibration signal decomposition results are as follows: Figure 3 As shown, the Pearson correlation coefficient between each IMF component and the original signal is as follows Figure 4 As shown in the figure, the feature extraction thresholds under this condition are calculated to be 0.3049. Figure 4 From the correlation coefficient value and feature extraction threshold in , it can be seen that the Pearson correlation coefficients of modes IMF4, IMF5, IMF6 and the original signal are greater than the feature extraction threshold under this working condition. Therefore, modes IMF4, IMF5, and IMF6 are selected as feature components under the winding deformation fault state. In order to ensure that the feature vector lengths of each working condition are the same, the first three modes that meet the fault extraction conditions are selected as feature components. The unnormalized feature vector values ​​are shown in Table 1.

[0087] Table 1 Eigenvector values ​​under three working conditions

[0088]

[0089] (3) BWO-SVM model fault feature vector classification

[0090] The BWO model is used to optimize the parameters of the SVM classifier, the BWO-SVM model parameters are initialized, the beluga population size n=20, the maximum number of iterations Tmax=100, the number of optimization parameters d=2, which is the penalty factor c and kernel function parameter g of the traditional SVM classifier, the parameter optimization range is set to c∈[0,100], g∈[-100,100], and the position matrix P of the beluga population is randomly generated in the search space.

[0091] The fitness function of the BWO model is the fault diagnosis accuracy δ, as shown in the following formula.

[0092]

[0093] Among them, q is the number of samples in which the model accurately identifies faults, and n is the total number of samples. The position matrix of the white whale population is updated and the fitness function value is calculated through the exploration, development, and whale fall stages. The maximum number of iterations is combined to determine whether the iteration termination condition is reached, and the optimal parameter value is substituted into the SVM model.

[0094] Combined with the feature extraction threshold, 100 groups of feature vectors are obtained under the normal operation of the power transformer and the winding deformation and loose fault states. The category labels are 1, 2 and 3 respectively. 80 groups of feature vectors under each of the three working conditions are taken as training sets and input into the SVM model. The remaining 20 groups of feature vectors are used to test the performance of the trained BWO-SVM classifier.

[0095] In order to verify the superiority of the power transformer winding mechanical fault diagnosis method based on CEEMDAN-BWO-SVM proposed in this invention, a comparison is made from two aspects: the feature vector extraction method and the SVM parameter optimization algorithm. For the feature vector extraction method, the exponential entropy (EE) is converted into singular spectrum entropy (SSE), approximate entropy (APE) and sample entropy (SEN). For the SVM parameter optimization algorithm, the BWO model is converted into a particle swarm optimization algorithm (PSO) and a whale optimization algorithm (WOA), and a classifier is constructed in combination with the SVM model for comparative experiments. The initialization parameter setting method of different optimization algorithms is the same as that of the above-mentioned BWO model. The number of samples under the three working conditions is 100, and the ratio of training set to test set is 4:1. The influence of different feature vector extraction methods and optimization algorithms on the fault diagnosis accuracy is shown in Table 2.

[0096] Table 2 Comparison of classification effects of different models

[0097]

[0098] It can be seen from the results that, compared with other feature vector extraction methods, the feature vectors generated based on CEEMDAN and exponential entropy (EE) can effectively evaluate the signal characteristics under different working conditions. The introduction of the BWO algorithm can enhance the global search capability and convergence speed, thereby achieving accurate optimization of SVM model parameters and further improving the accuracy of fault diagnosis. Finally, the fault diagnosis accuracy of the model of the present invention is 98.98%, indicating that the model can effectively perform mechanical fault diagnosis of power transformer windings based on CEEMDAN-BWO-SVM.

[0099] In summary, traditional parameters such as information entropy have problems such as zero value and non-convergence. Exponential entropy (EE) extracts features based on exponential functions, thereby effectively evaluating signal characteristics under different working conditions. The use of the BWO optimization algorithm can fully avoid the problems of parameters falling into local optimal solutions and high computational complexity. The position matrix of the white whale population is updated through the exploration, development and whale fall stages to achieve efficient parameter optimization of the SVM model, so that the trained SVM model can be used to classify and identify mechanical faults of power transformer windings, which is convenient for maintaining the safe operation of power transformers.

[0100] The method of the present invention is based on the CEEMD method, exponential entropy (EE) and BWO-SVM model. It realizes winding mechanical fault identification by extracting the fault characteristics of the vibration signal on the surface of the transformer box wall and classifying them, thereby reducing the operating risk of the power transformer and improving the reliability and safety of the power system.

[0101] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.

Claims

1. A method for diagnosing mechanical faults of power transformer windings based on CEEMDAN-BWO-SVM, characterized by: The following steps are involved: S1. Obtain the vibration acceleration signal of the outer wall of the sample power transformer box; S2. Use CEEMDAN method to decompose the vibration acceleration signal into N IMF components; S3. Calculate the Pearson correlation coefficient between each IMF component and the initial vibration acceleration signal, and select the characteristic component based on the Pearson correlation coefficient; S4. Calculate the exponential entropy of the characteristic component and form a characteristic vector; S5. construct a BWO-SVM prediction model, and input the feature vector into the BWO-SVM prediction model for training; S6. Obtain the vibration acceleration signal of the outer wall of the power transformer box to be tested, and process the vibration acceleration signal of the power transformer to be tested through steps S2 to S4 to form a feature vector to be tested, and input the feature vector to be tested into the trained BWO-SVM prediction model to obtain the mechanical fault diagnosis result.

2. The power transformer winding mechanical fault diagnosis method based on CEEMDAN-BWO-SVM according to claim 1 is characterized by: The CEEMDAN method is used to decompose the vibration acceleration signal into N IMF components, including: S21. Determine M Gaussian white noises with the same standard deviation n i (t), where i = 1, 2, ..., M; S22. Add Gaussian white noise to the initial signal X(t), and then perform EMD decomposition on the initial signal X(t) to which the Gaussian white noise is added to obtain component IMF1; S23. Calculate the residual component r1(t): r1(t)=X(t)-IMF1; S24. Perform EMD decomposition on the original Gaussian white noise and convert the decomposed first-order component E1(n i (t)) is superimposed with the residual component r1(t) to obtain M synthetic signals, and the EMD decomposition is performed to obtain IMF2; S25. Repeat steps S22 to S24 until the set number N is reached to obtain the component IMF N and the remaining component r N (t).

3. The power transformer winding mechanical fault diagnosis method based on CEEMDAN-BWO-SVM according to claim 2 is characterized by: Step S3 specifically includes: Calculate the Pearson correlation coefficient ρ n : Where: n = 1, 2, ..., N, cov (IMF n ,X) is the covariance between IMFn and the initial signal X(t), σ X are the standard deviations of IMFn and the initial signal X(t), μ X are the means of IMFn and the initial signal X(t) respectively; Set the feature extraction threshold η; set ρ n The IMF components with >η are selected to form the characteristic component S k , k = 1, 2,…, K, K is the total number of characteristic components.

4. The power transformer winding mechanical fault diagnosis method based on CEEMDAN-BWO-SVM according to claim 3 is characterized by: Step S4 specifically includes: Where: h k is the characteristic component S k The total number of sample points, P kj is the characteristic component S k The energy value of the jth sample point in S k Ratio of energy values; Construct feature vector S: S = [H1, H2, ..., H K ].

5. The power transformer winding mechanical fault diagnosis method based on CEEMDAN-BWO-SVM according to claim 1 is characterized by: The initial parameters of the BWO-SVM prediction model include the beluga population size n, the maximum number of iterations T max and the number of optimization parameters d, as well as the initialization parameter penalty factor c and kernel function parameter g of the SVM classifier, where: c∈[0,100], g∈[-100,100].

Images