Wavelet transform-multilayer perceptron based fault diagnosis method for transmission lines

Abstract

The application discloses a kind of wavelet transform-multilayer perception based on transmission line fault diagnosis method, steps include:1) the original three-phase current signal of power transmission system is collected, and the fault type is numbered;Again, the original three-phase current signal is handled with noise, and sample is divided into training set, test set;2) different frequency band wavelet coefficient is obtained by wavelet decomposition, threshold denoising is carried out by determining wavelet denoising threshold and denoising mode, and the best wavelet basis function of denoising effect is selected to carry out wavelet decomposition;3) wavelet coefficient energy value is extracted, and the feature matrix of fault energy is constructed, while the feature matrix is preprocessed in different ways, and the data processing mode with the highest diagnostic accuracy is selected;4) adjust network full connection layer structure and neuron number, join Dropout network optimization, and obtain MLP fault diagnosis model.The method of the application reduces the influence of noise, and significantly improves training speed and diagnostic accuracy.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of fault diagnosis of power transmission lines, and relates to a power transmission line fault diagnosis method based on wavelet transform and a multilayer perceptron. BACKGROUND

[0002] With the continuous development of the power system, it is of great significance to realize the rapid fault diagnosis and early warning of the power transmission line and to ensure the reliable operation of the power system. However, the existing fault diagnosis technology has the problems of complex network model, slow diagnosis speed and low diagnosis accuracy. In view of this problem, if the early fault features in the real-time monitoring data can be found and real-time diagnosis and early warning can be performed before the permanent fault occurs, an adjustable maintenance plan can be provided for the operation and maintenance personnel, and the safe and stable operation of the power system can be ensured.

[0003] Recently, deep learning is more and more widely used in power system fault diagnosis. The classic multilayer perceptron (MLP) in deep learning can realize online fault diagnosis. Through multiple full connection nonlinear operations, high-dimensional features in fault data can be extracted to realize accurate diagnosis of complex faults. However, the original data is used as the input of the diagnosis network, and there is much redundant information, which increases the model training time. In addition, the noise contained in the real-time monitoring data will also affect the accuracy of the diagnosis model. SUMMARY

[0004] The purpose of the application is to provide a power transmission line fault diagnosis method based on wavelet transform and a multilayer perceptron, which solves the problems of too much redundant information, increased model training time, too much noise in monitoring data and low accuracy of power transmission line fault diagnosis in the prior art.

[0005] The technical solution adopted by the application is a power transmission line fault diagnosis method based on wavelet transform and a multilayer perceptron, which is implemented according to the following steps:

[0006] Step 1, collect the original three-phase current signal of the power transmission system, number the fault types to determine the fault type number corresponding to the original three-phase current signal, and then add noise to the original three-phase current signal to obtain a fault sample containing noise, and divide the sample into a training set and a test set;

[0007] Step 2, obtain wavelet coefficients of different frequency bands through wavelet decomposition, determine a wavelet denoising threshold and a denoising mode for threshold denoising, verify the denoising effect of different wavelet basis functions, and select the best wavelet basis function for wavelet decomposition according to the denoising effect;

[0008] Step 3, the wavelet coefficient energy value is extracted, a fault energy feature matrix is constructed, and different ways of pretreatment are carried out on the feature matrix, and the data processing mode with the highest diagnostic accuracy is selected;

[0009] Step 4, the MLP fault diagnosis model is established, the network full connection layer structure and the number of neurons are adjusted, and the Dropout network optimization is added, and the fault diagnosis model with the highest fault diagnosis accuracy is selected.

[0010] The beneficial effects of the application are as follows:

[0011] 1) The advantages of two methods are combined, noise suppression is carried out through wavelet change, the influence of noise on the diagnostic model is reduced, and the energy feature matrix extracted can be used as the input of the diagnostic network and can be further analyzed.

[0012] 2) The multi-layer perception machine model is established for fault diagnosis, the feature energy matrix is used as the network input, and the standard normal distribution is selected for processing, which can not only reduce the model calculation load, but also significantly improve the model training speed and diagnostic accuracy, and provides a reference for power transmission line fault diagnosis.

[0013] 3) The fault diagnosis accuracy of the power transmission line proposed in the application is better than that of the traditional machine learning method, and more efficient fault diagnosis can be realized. DETAILED DESCRIPTION

[0014] Figure 1 is an embodiment architecture diagram of the multi-layer perception machine used in the application;

[0015] Figure 2a is a standard full connection layer network of the Dropout network optimization; Figure 2b is a full connection layer network using the Dropout technology;

[0016] Figure 3 is a whole process diagram of the method of the application;

[0017] Figure 4a is a wavelet coefficient energy value of the multi-cycle early fault A-phase current; Figure 4b is a wavelet coefficient energy value of the half-cycle early fault A-phase current; Figure 4c is a wavelet coefficient energy value of the constant impedance grounding fault A-phase current;

[0018] Figure 5a is a schematic diagram of the multi-cycle early fault of the electrical diagram model; Figure 5b is a schematic diagram of the half-cycle early fault; Figure 5c is a constant impedance grounding schematic diagram;

[0019] Figure 6is a fault original energy value schematic diagram of the embodiment of the application;

[0020] Figure 7 is a wavelet energy pretreatment schematic diagram of the embodiment of the application;

[0021] Figure 8 is a training accuracy schematic diagram of the embodiment of the application;

[0022] Figure 9a is an original current signal recognition result; Figure 9b is a standardized energy feature recognition result of the embodiment of the application;

[0023] Figure 10a is a diagnosis result of the method of the application, Figure 10b is a comparison diagram of the diagnosis result of the method of the application and a decision tree; Figure 10c is a comparison diagram of the diagnosis result of the method of the application and K nearest neighbor. DETAILED DESCRIPTION

[0024] The application will be described in detail below in combination with the drawings and specific embodiments.

[0025] As a classical deep learning network, the multi-layer perception has fast response speed and low calculation example demand, and through model structure and parameter adjustment, the most suitable network model is selected, so that the response speed and accuracy of line fault diagnosis are obviously improved. In order to reduce the influence of noise on the fault diagnosis accuracy of the multi-layer perception, the original data is denoised through wavelet denoising, and at the same time, in order to further improve the fault diagnosis accuracy, the fault features are better represented through data pretreatment, and the performance of the multi-layer perception fault diagnosis model is further improved. According to this, the application proposes an algorithm combining wavelet transformation and multi-layer perception for power transmission line fault recognition, constructs a fault feature matrix through wavelet transformation as the input of the multi-layer perception diagnosis model, selects the diagnosis accuracy optimal model through adjustment of the model structure and parameters, and verifies the superiority of the multi-layer perception in power transmission line fault diagnosis through comparison with traditional machine learning.

[0026] The power transmission line fault diagnosis method of the application is based on the above wavelet transformation-multi-layer perception principle and implemented according to the following steps:

[0027] Step 1, original three-phase current signals of a power transmission system are collected, the fault types are numbered to determine the fault type numbers corresponding to the original three-phase current signals, the original three-phase current signals are then subjected to noise adding treatment, the fault samples containing noise are obtained, and the samples are divided into a training set and a test set, and the specific process is as follows:

[0028] A Gaussian white noise signal e(t) is superimposed in the original three-phase current signal x(t) to obtain a current signal f(t) containing noise, and the function formula is as formula (1):

[0029] f(t) = x(t) + e(t) (1)

[0030] In formula (1), t represents time, and the ratio of the training set to the test set in the embodiment is 4:1.

[0031] Step 2, obtain wavelet coefficients of different frequency bands through wavelet decomposition, determine a wavelet denoising threshold and a denoising mode for threshold denoising, verify the denoising effect of different wavelet basis functions, and select the wavelet basis function with the best denoising effect for wavelet decomposition, and the specific process is as follows:

[0032] First, wavelet decomposition is used to obtain wavelet coefficients W j,k (t) of the kth frequency band on the jth layer, and the functional expression of the wavelet coefficients W j,k (t) is as formula (2):

[0033]

[0034] In formula (2), a0 j is a scaling factor, ka0 j b0 is a translation factor, a0>1, b0>0; f(t) is a current signal containing noise; and Ψ is a wavelet basis function.

[0035] Second, an adaptive threshold (Rigrsure) is used to determine the threshold value λ i size, that is, the threshold value size is determined by calculating the local mean μ i and the local variance The functional expression is as follows:

[0036]

[0037]

[0038] Where x(i) represents the value of the ith current signal, and N represents the number of current value samples; when the threshold value λ i size is determined according to the local variance , generally k times of σ, and k generally takes a value of 2.5, then the functional expression of the adaptive threshold value λ i is:

[0039] λ i = kσ i (5)

[0040] After the threshold value λ i is determined, the hard threshold denoising mode and the soft threshold denoising mode commonly used in wavelet denoising are used, and the functional expressions of the two modes are formula (6) and (7) respectively:

[0041]

[0042]

[0043] wherein, W i is the current wavelet coefficient vector; W δ is the wavelet coefficient after threshold quantization processing;

[0044] Finally, the evaluation index of denoising is constructed, including the signal to noise ratio (SNR) and the root mean square error (RMS), and the two function expressions are as follows:

[0045]

[0046]

[0047] wherein, x(i) is the original signal; x δ (i) is the signal after denoising; is the power of the original signal; is the power of the noise; N is the number of current samples;

[0048] Step 3, the wavelet coefficient energy value is extracted, the feature matrix of fault energy is constructed, and different ways of preprocessing are performed on the feature matrix, and the data processing method with the highest diagnosis accuracy is selected, and the specific process is:

[0049] The energy E(j, k) is obtained through the wavelet coefficient W j,k (t), and the function expression of the energy E(j, k) is as follows:

[0050]

[0051] The standard normal distribution processing can reduce the correlation between features, and converting data of different orders of magnitude to the same order of magnitude is conducive to comparison, and the function expression of the standard normal distribution processing (i.e. standardization processing) is as follows:

[0052]

[0053]

[0054] wherein, x′ i is the data after standardization processing; μ is the average value of the original data; σ is the standard deviation of the original data;

[0055] Step 4, the MLP fault diagnosis model is established, the network full connection layer structure and the number of neurons are adjusted, and the Dropout network optimization is added, and the one with the highest fault diagnosis accuracy is taken as the MLP fault diagnosis model,

[0056] A fault diagnosis model of a multi-layer perceptron is established, all inputs are fully connected by multiple neurons to extract high-dimensional fault features, and the neuron output z w,b after an activation function is calculated as formula (13):

[0057] z w,b (x)=ρ(w i x i +b) (13)

[0058] In formula (13), z w,b represents the neuron output after the activation function, and p is the activation function operation;

[0059] The output a n of the nth layer of the multi-layer perceptron needs to satisfy formula (14):

[0060] a n =ρ(z n )=ρ(W n a n-1 +b n ) (14)

[0061] In formula (14), a n is the output after the activation function p operation of z n , z n is the calculation result of the nth layer of the multi-layer perceptron, W n is the weight of the nth layer, and b n is the bias coefficient of the nth layer;

[0062] The output a n of each layer is calculated by formula (14), and the matrix expression is formula (15):

[0063]

[0064] a n is the output of the nth layer; a n 1, a n 2, …, a n k 1, a k,n 2, …, a n n are respectively the first neuron calculation value, the second neuron calculation value, …, the kth neuron calculation value of the nth layer output; w n k are respectively the first neuron bias coefficient, the second neuron bias coefficient, …, the kth neuron bias coefficient of the nth layer;

[0065] ​In order to avoid overfitting in network training process, therefore, the Dropout technology is added, and the Mean Squared Error (MSE) loss function is used to evaluate the model training, and the expression of the MSE loss function is as formula (16):

[0066]

[0067] In formula (16), w is a weight matrix, x is a network input, y is a network output, that is, the output of the network.

[0068] Experimental verification:

[0069] The method for power transmission line fault diagnosis based on wavelet-multilayer perceptron of the application, the five-layer multilayer perceptron fault diagnosis model structure constructed is shown in Figure 1 , the Dropout network optimization used is shown in Figure 2a 、 Figure 2b , the diagnosis process is shown in Figure 3 , and the specific implementation steps are as follows:

[0070] Step 1, collect the original three-phase current signal of the power transmission system, number the fault types to determine the fault type number corresponding to the original three-phase current signal. The original three-phase current signal is subjected to noise processing to obtain a fault sample containing noise, and the sample is divided into a training set and a test set according to a ratio of 4:1;

[0071] The original three-phase current signal x(t) is superimposed with Gaussian white noise e(t) to obtain a current signal f(t) containing noise:

[0072] f(t) = x(t) + e(t) (1)

[0073] Step 2, first, wavelet decomposition is used to obtain the wavelet coefficient W j,k (t) on the kth frequency band on the jth layer, and the function formula is as formula (2):

[0074]

[0075] The threshold value λ i size is determined by calculating the local mean μ and the local variance i , and the function formula is:

[0076]

[0077]

[0078] The threshold value λ i obtained is as formula (5):

[0079] λ i =kσi (5)

[0080] k=2.5, after determining the threshold, select hard threshold and soft threshold denoising method, function formula as formula (6) and (7):

[0081]

[0082]

[0083] Where, W i is the current wavelet coefficient vector; W δ is the wavelet coefficient obtained by threshold quantization processing; λ i is the determined threshold value;

[0084] Finally, the evaluation index of signal-to-noise ratio and root mean square error is constructed to verify the denoising effect, function formula as formula (8) and formula (9):

[0085]

[0086]

[0087] Where, is the power of the original signal; is the power of the noise; x(i) is the original signal; x δ (i) is the denoised signal; N is the number of current value samples;

[0088] Select sym6 wavelet basis function for denoising, and the evaluation index is shown in Table 1.

[0089] Table 1, select sym6 wavelet basis function for denoising evaluation index

[0090]

[0091] Through the analysis of the evaluation index of soft threshold and hard threshold denoising, it is known that when the sym6 wavelet basis function is selected, the value of SNR is 45.3277, which is the maximum value, and the value of RMSE is the minimum value at this time, confirming that the sym6 wavelet basis function has the best denoising effect, so the sym6 wavelet basis function is selected for signal decomposition.

[0092] Step 3, select wavelet basis function sym6 for signal 6-layer decomposition, and extract each wavelet coefficient energy E(j, k), function formula as formula (10):

[0093]

[0094] Where, W j,k (t) is the wavelet coefficient of the kth frequency band on the jth layer;

[0095] The original energy values of 7 frequency bands of fault 7 are obtained according to formula (10), as shown in Table 4. Figure 4a 、 Figure 4b 、 Figure 4c .

[0096] The wavelet coefficient energy features are preprocessed by using the standard normal distribution to reduce the correlation between the features and convert data of different orders of magnitude into the same order of magnitude. The standardization calculation is as shown in formula (11) and formula (12):

[0097]

[0098]

[0099] x'i is the data after standardization processing; μ is the average value of the original data; σ is the standard deviation of the original data;

[0100] The fault energy values after standardization processing are shown in Table 5. Figure 5a 、 Figure 5b 、 Figure 5c .

[0101] Step 4, the MLP fault diagnosis model is established, all inputs are fully connected by multiple neurons to extract high-dimensional fault features, and the neuron output z w,b as shown in formula (13):

[0102] z w,b (x)=ρ(w i x i +b) (13)

[0103] The input of the n-th layer satisfies formula (14):

[0104] a n =ρ(z n )=ρ(W n a n-1 +b n ) (14)

[0105] The output a n of the last layer is calculated by formula (14), as shown in formula (15):

[0106]

[0107] The error function used in the network training process is the mean squared error (MSE), and the MSE calculation is as shown in formula (16):

[0108]

[0109] Wherein, the weight matrix is W, N is the sample number, x is the network input; y is the network output;

[0110] The present application carries out 8 examples of tests in the fault simulation process, and the fault simulation parameter settings of example 1, example 2, example 3, example 4, example 5, example 6, example 7 and example 8 are shown in table 2.

[0111] Table 2, fault simulation parameter settings of 8 examples of the present application

[0112]

[0113] From the data of the above 8 examples, it can be seen that the power transmission line fault diagnosis method based on wavelet-multilayer perception machine of the present application has more reliable and stable effect on line fault monitoring. Figure 6 is an equivalent circuit diagram of the power transmission line, and the electrical diagram is built in PSCAD according to the line structure, the line is a 5.0km long 10kV power transmission line, a 10kV type YJLY-10kV cross-linked polyethylene power transmission line is selected for verification, the parameters are shown in table 3, and the power transmission line fault PSCAD model is shown in Figure 7 .

[0114] Table 3, power transmission line parameters

[0115]

[0116]

[0117] In order to increase the diversity of fault characteristics, the diversity of faults is increased by modifying the time constant and arc loss energy, the initial angle of type fault occurrence is a random angle in a period, the fault distance is a random position of 0 to 5km, the sample data are all sampled at 10kHz, 600 samples of each type of fault are generated, and then the samples are divided into a training set and a verification set according to a ratio of 4:1. The results show that the recognition accuracy of the power transmission line fault diagnosis of the method of the present application is obviously improved compared with traditional machine learning random forest (RF) and K nearest neighbor (KNN) methods.

[0118] Example 9 is a hardware test example of the present application. The wavelet-multilayer perception machine of the MLP fault diagnosis model of the method of the present application adopts 3 layers of hidden layers, the neuron numbers of the 3 layers of hidden layers are 100-50-100 respectively, and the Dropout technology is added in the training process to avoid overfitting phenomenon. In addition, different inputs are also used as the inputs of the multilayer perception machine, the Dropout is set to 0.4, the learning rate is 0.05, the iteration is 600, and the model training accuracy result is shown in Figure 8 . From Figure 8It can be seen that, taking the original current as the input, the training accuracy fluctuates greatly in the initial training stage due to the existence of redundant information; while taking the standardized energy as the input, the training accuracy rises smoothly, and after 250 iterations, the accuracy basically reaches the highest, which is 98.72%; and the accuracy of the original current as the input still fluctuates greatly after 600 iterations. The test set results are shown in Figure 9a 、 Figure 9b For 600 samples, 480 are divided into a training set and 120 are divided into a test set, and the fault recognition accuracy is 98.39%.

[0119] Finally, in order to verify the advantages of the multi-layer perceptron used in the method, the traditional machine learning RF and K nearest neighbor method are compared, and the fault diagnosis results are displayed through the t-SNE visualization technology, further verifying the feasibility of the method for fault diagnosis, and the diagnosis results are shown in Figure 10a 、 Figure 10b 、 Figure 10c , wherein XY represents the spatial position. Analyzing the fault diagnosis scatter plot of each model, Figure 10a The MLP fault diagnosis model of the application has a model test accuracy of 98.39%, and from the scatter plot it can be seen that after multi-layer full connection layer calculation, the feature vector is divided into 8 obvious categories, and the distance between the same faults is small, and the distance between different features is large, indicating that the MLP fault diagnosis model of the application has less confusion between different faults in the fault recognition process, and the generalization of the MLP fault diagnosis model of the application is stronger. Figure 10b The RF classification result is shown in Table 4, and the model test set accuracy is 93.75%. Since RF is composed of multiple DTs, the model recognition accuracy is relatively improved by 4.51%, and from the scatter plot it can be seen that the overall classification result of the model is good, and only the distance between sample 4 and sample 5 is close, and there is a small amount of confusion. Figure 10c The KNN classification result is shown in Table 5, and the model test set accuracy is 89.24%. The effect is good in the classification of sample 1, sample 6, sample 7 and sample 8, sample 2, sample 3 and sample 4 are divided into different sets, and sample 4 and sample 5 have some overlap. Through the above comparison, each traditional machine learning has a small amount of confusion between sample 4 and sample 5, and the MLP fault diagnosis model of the application has strong generalization in fault diagnosis.

Claims

1. A power transmission line fault diagnosis method based on wavelet transform-multilayer perceptron, characterized by, The following steps are implemented: Step 1, collect the original three-phase current signal, number the fault type; add noise to the original three-phase current signal, obtain the fault sample containing noise, divide it into training set and test set, the specific process is: Collect the original three-phase current signal of the power transmission system, number the fault type; then add noise to the original three-phase current signal, obtain the fault sample containing noise, and divide the sample into training set and test set; In the original three-phase current signal a Gaussian white noise signal is superimposed , resulting in a noisy current signal f ( t ) with the functional form as in equation (1): （1） In formula (1), t representing time; Step 2, select the best wavelet basis function for wavelet decomposition, the specific process is: Through wavelet decomposition, the wavelet coefficients of different frequency bands are obtained, the wavelet denoising threshold and denoising method are determined for threshold denoising, the best wavelet basis function for denoising is selected for wavelet decomposition, First, wavelet decomposition is used to obtain the th j The first layer k Wavelet coefficients W in each frequency band j,k ( t Wavelet coefficients W j,k ( t The function expression is shown in equation (2): （2） In formula (2), a 0 j is a scaling factor, ka 0 j b 0 is a translation factor, a 0>1, b 0>0; f ( t ) is a noisy current signal; Ψ is a wavelet basis function; Secondly, the adaptive threshold method is used to determine the threshold λ i Size, i.e. by calculating local mean μ i And local variance σ i 2 Determine the threshold size, the function is as follows: （3） （4） wherein, represents the value of the i current signal, N represents the number of current value samples; the threshold value σ i 2 is determined as a function of the local variance λ i size, then the adaptive threshold λ i function is: （5） Threshold value λ i After the determination, the hard threshold denoising method and the soft threshold denoising method commonly used in wavelet denoising are adopted, and the function expressions of the two methods are formula (6) and formula (7) respectively: （6） （7） wherein, W i is the current wavelet coefficient vector; W δ is the wavelet coefficient after threshold quantization. Finally, the denoising evaluation index is constructed, including signal-to-noise ratio SNR and root mean square error, the two function expressions are as follows: （8） （9） wherein, x i is the original signal; x δ i is the denoised signal; is the power of the original signal; is the power of the noise; N is the number of current value samples;​​ Step 3, extract the wavelet coefficient energy value, construct the feature matrix of fault energy, and preprocess the feature matrix in different ways to select the data processing method with the highest diagnosis accuracy; Step 4, adjust the network full connection layer structure and the number of neurons, and add Dropout network optimization, and take the one with the highest fault diagnosis accuracy as the MLP fault diagnosis model.

2. The wavelet transform-muti-layer perceptron based transmission line fault diagnosis method according to claim 1, characterized in that, In step 3, the specific process is: The energy E is obtained by the wavelet coefficients W j,k ( t ) of the image I E(j,k), The energy E is obtained by the wavelet coefficients W E(j,k) The functional expression of the energy E is as follows: （10） The standard normal distribution processing can reduce the correlation between features, and converting data of different orders of magnitude to the same order of magnitude is conducive to comparison. The function expression of standard normal distribution processing is as follows: （11） （12） wherein, x' i is the standardized data; μ is the average of the raw data; σ is the standard deviation of the raw data. 3.The wavelet transform-muti-layer perception machine based power transmission line fault diagnosis method according to claim 1, wherein, In step 4, the specific process is: The fault diagnosis model of the multilayer perceptron is established, all inputs are fully connected by multiple neurons, high-dimensional fault features are extracted, and neuron outputs are obtained through an activation function z w,b , the function is as formula (13): （13） In formula (13), z w,b denotes the neuron output through an activation function, ρ is an activation function operation; then the output of the multi-layer perceptron is n the output of the layer a n needs to satisfy equation (14): （14） In equation (14), a n for z n After activation function ρ The output after the calculation z n This represents the computation result of the nth layer of the multilayer perceptron. W n For the first n Layer weights b n For the first n Layer bias coefficient; The output of each layer is calculated by equation (14) a n The matrix expression is given by equation (15): （15） a n For the first n Layer output; a n 1, a n 2, …, a n k The first n The calculated values ​​of the first neuron, the second neuron, ..., the output of the layer. k Each neuron calculates a value; w k,n The weight values ​​between each layer, b n 1, b n 2, …, b n k The first n The bias coefficients of the first neuron, the second neuron, ..., the first neuron of the layer k Bias coefficients of individual neurons; In order to avoid overfitting during network training, Dropout technology is added, and mean square error loss function is used to evaluate model training, the expression of MSE loss function is as formula (16): （16） In formula (16), w is a weight matrix; x is a network input; y is a network output.

Images