CNN-SVM model transformer fault online diagnosis method based on LMD decomposition and frost ice optimization

Abstract

The invention discloses a CNN-SVM model transformer fault online diagnosis method based on LMD decomposition and frost ice optimization, and belongs to the technical field of fault diagnos.The method comprises the steps that a transformer vibration signal is decomposed into a plurality of PF components through local mean decomposition, effective PF components are extracted according to the fuzzy entropy value of each PF component for signal reconstruction, and the purpose of noise reduction is achieved; taking the reconstructed signal as input, training a DCGAN model, and continuously optimizing a generator network until a sample generated by the generator network is sufficiently similar to a real sample in statistical property; inputting the data set into the CNN-SVM optimized by the frost ice optimization algorithm for offline model training, and constructing an optimal transformer fault diagnosis model; and finally, completing online transformer fault diagnosis by using the optimal CNN-SVM model. According to the method, the problem of low diagnosis precision of the transformer fault vibration signal caused by small sample unbalanced data is effectively solved.

Description

Technical Field

[0001] The present invention relates to the technical field of transformer fault diagnosis, and specifically, to an online transformer fault diagnosis method based on an LMD decomposition and frost ice optimization CNN-SVM model. Background Art

[0002] Power transformers are one of the most important devices in the power system, and their operating conditions directly affect the safety and stability of the entire power grid. However, in actual operation, transformers face a complex and harsh working environment, and accidents caused by transformer faults occur from time to time. Therefore, in order to ensure the stability and reliability of the power system operation, it is very necessary to timely grasp the operating conditions of transformers and perform online fault diagnosis on transformers with potential safety hazards.

[0003] Transformer fault diagnosis technology based on vibration signals diagnoses faults by analyzing the vibration signals of the transformer tank. It has the advantages of online detection, no electrical connection with the test object, simple operation, and no change to the internal structure of the transformer. Online fault detection based on vibration signals can well make up for the deficiencies of offline detection.

[0004] When a transformer has internal mechanical faults, its vibration signals will generate non-linear and non-stationary signals. Currently, commonly used signal processing methods include wavelet decomposition method, empirical mode decomposition method (EMD), etc. Among them, the wavelet decomposition method requires artificial selection of a suitable wavelet basis and decomposition levels, and does not have self-adaptability; although EMD reduces the influence of human intervention compared with wavelet decomposition, there are problems such as end effects and mode mixing. Summary of the Invention

[0005] In order to solve the defects and deficiencies existing in the above-mentioned prior art, the present invention provides an online transformer fault diagnosis method based on an LMD decomposition and frost ice optimization CNN-SVM model, which can effectively solve the problem of low diagnostic accuracy caused by poor extraction of transformer fault vibration signal features.

[0006] The present invention is implemented by adopting the following technical solutions: An online transformer fault diagnosis method based on an LMD decomposition and frost ice optimization CNN-SVM model includes the following steps:

[0007] First step, preprocessing of transformer vibration signals: Perform LMD decomposition on the vibration signals to obtain multiple PF components, calculate the fuzzy entropy values of each PF component, and select the PF components with fuzzy entropy values higher than the threshold to reconstruct the signal;

[0008] Second step, sample expansion and its equalization: After performing convolutional pooling processing on the reconstructed signal, input it into the generator. After being judged by the discriminator and fed back to optimize the generator parameters, make the data generated by the generator realistic enough, and finally output to obtain the expanded data set;

[0009] Step 3, optimize the model parameters: Use the frost ice optimization algorithm to optimize the penalty factor c and the kernel function parameter h of CNN-SVM to obtain the optimal parameters;

[0010] Step 4, model training: Divide the balanced samples into a training set, a validation set, and a test set and input them into RIME-CNN-SVM for offline model training to construct an optimal transformer fault diagnosis model;

[0011] Step 5, input the real-time fault signal into the optimal model to complete the online diagnosis of transformer faults.

[0012] Preferably, the specific steps of the first step are as follows:

[0013] Step 1: Segment and decompose the collected vibration signal into k PF components;

[0014] Specifically, calculate the average value mi of adjacent local extreme points and the envelope estimation value ai, perform smoothing processing on the adjacent local extreme mi and the adjacent envelope estimation value ai to obtain the local mean function m11(t) and the envelope estimation function a11(t), then separate h11(t) from the original signal X(t), demodulate h11(t) to obtain s11(t), multiply a1n(t) by s1n(t) to obtain the first δ pF1 (t)

[0015] δ pF1 (t) = a 1 (t)s 1n (t)

[0016] Separate PF1 from the original signal X(t) to obtain the difference signal u1(t). At this time, the difference signal is used as the original signal and the above steps are repeated until the last difference signal becomes a monotonic function and stops. At this time, X(t) is decomposed into K PF components and a residual component uk(t):

[0017]

[0018] Step 2: Calculate the fuzzy entropy values of each PF component;

[0019] Specifically, use the PF components selected by the local mean decomposition algorithm as the input vector of the fuzzy entropy algorithm, denoted as u(i). Generate a group of m-dimensional vectors {Xmi} in the order of u(i), that is

[0020]

[0021] where u0(i) represents the average value of m consecutive u values.

[0022] Assume dm ij For two m-dimensional vectors X m i and X m j The maximum value of the difference in the distances between the pairwise corresponding elements in, that is:

[0023]

[0024] Given the parameters n and r, through the membership function μ(d m ij , n, r), the similarity D between Xmi and Xmj can be calculated, that is m ij , that is

[0025]

[0026] where n and r respectively represent the gradient and width of the boundary of the function μ(d m ij , n, r).

[0027] Define the function Q m (n, r), and regenerate a set of m + 1-dimensional vectors Q according to the above steps, that is m+1 , that is

[0028]

[0029] Obtain the negative natural logarithm of the deviation of the function Qm from Qm+1. When the length of the data set is finite, the fuzzy entropy can be equivalent to

[0030] FuzyEn(m,n,r) = lnQ m (n,r) - lnQ m+1 (n,r)

[0031] Step 3: Screen out the PF components whose entropy values meet the requirements for signal reconstruction to obtain the reconstructed signal x(tn). Specifically

[0032]

[0033] Preferably, the specific steps of the second step are as follows

[0034] Step 1: Input the real signal into the generator G, and at the same time, G randomly generates a vibration signal

[0035] Specifically, input the real data sample x(tn) into the generator G, and at the same time, the generator generates a random vibration signal sample x'.

[0036] The input of D is x(tn) and x′. D(x) represents the probability that x is from a real sample rather than from Pg. The purpose of training D is to maximize the accuracy of D in distinguishing between real and fake input samples, such that real samples are classified as 1 and fake samples are classified as 0, that is, the value of D(x) approaches 1 and the value of D(x′) approaches 0.

[0037] Step 2: The generator G outputs the vibration signal to the discriminator D. The discriminator calculates their similarity degree and then feeds back to train and generate vibration signal samples.

[0038] Specifically, keeping the parameters of G unchanged, then sending the generated sample x′ and the real sample x into D, and using the stochastic gradient ascent method to update the parameters of D through the following formula.

[0039]

[0040] Then train G, keeping the parameters of D unchanged, and using the stochastic gradient descent method to update the parameters of G through the following formula

[0041]

[0042] Until the two reach the Nash equilibrium P g = P data , at this time D(x) = D(x′) = 0.5, that is, the discriminator randomly discriminates the authenticity of the sample with a probability of 50%.

[0043] Preferably, the specific steps of the third step are as follows:

[0044] Step 1: Initialize the rime population.

[0045] Specifically, the types of individuals in the population include the penalty factor c, the kernel function parameter h, and the RIME frost ice particle population R positions as shown below:

[0046]

[0047] In the formula: the frost ice particle population can be represented by x ij where i represents the serial number of the frost ice crystal and j represents the serial number of the frost ice particle.

[0048] Step 2: Calculate the fitness of individuals in the population and select a part of the individuals with higher fitness as crystallization nuclei.

[0049] Specifically, the fitness values F(S i ) of all frost ice particles are shown in the following formula:

[0050]

[0051] f is the fitness of each frost ice particle. Regarding the motion properties of the frost ice particles, the positions of the frost ice particles are shown in the following formula:

[0052]

[0053] Where: R new ij is the position after particle update, and R best.j is the j-th particle of the optimal crystal in the frost ice population R, r 1 is a random number in (-1, 1), which controls the moving direction of the particle and will change with the iteration of the moving direction cosθ. β is the environmental factor representing the external influencing factor, and ψ is the particle adhesion force, which is a random number in (0, 1) and is used to control the distance between particles. U bij and U bij represent the upper and lower bounds of the particle motion space respectively.

[0054] Step 3: For each "crystal nucleus", a certain number of new individuals are generated through random mutation operations. According to the fitness values of the new individuals, a part of the new individuals are selected to replace the original individuals to form a new population, and the individuals in the population will gradually approach a better solution;

[0055] Step 4: Determine whether the termination condition is satisfied, such as reaching the maximum number of iterations or finding a sufficiently good solution. If the termination condition is satisfied, stop the search; otherwise, return to Step 2 to continue iterative optimization.

[0056] Specifically, in order to ensure convergence, it is necessary to set as shown in the following formula:

[0057]

[0058] Where: h is the current number of iterations, and H is the maximum number of iterations.

[0059]

[0060] Where: β is the step function, w is used to control the number of segments of the step function, E is the adhesion coefficient, which will increase with the number of iterations and affect the condensation probability of the particles; r2 is a random number in (0, 1), and it and E control whether the particles condense, that is, whether the particle positions are updated. The expression of E is:

[0061]

[0062] Step 5: Substitute the obtained optimal penalty factor c and optimal kernel function parameter h into CNN-SVM to obtain the optimized RIME-CNN-SVM;

[0063] Preferably, the specific steps of the fourth step are as follows:

[0064] Step 1: Add labels to the sample signals and divide them into a training set, a test set, and a validation set according to a ratio of 7:2:1. Use the training set to train the RIME-CNN-SVM model;

[0065] Step 2: Use the frost ice optimization algorithm to optimize the penalty factor c and the kernel function parameter h of the convolutional neural network support vector machine;

[0066] Step 3: Perform convolutional pooling on the training set samples;

[0067] Specifically, the output of the convolutional operation is obtained through the activation function:

[0068]

[0069] In the formula: f(·) is the convolutional activation function; M j is the input feature map set; x j l-1 is the (l-1)-th channel of the input feature map; k l ij is the weight matrix of the convolution; * is the convolution operation; b l j is the bias term.

[0070] After each convolutional layer, a downsampling layer is applied. The downsampling layer reduces the size of the input features and the number of parameters in the network, not only improving the accuracy but also reducing the dimension of the extracted features, and preventing overfitting to a certain extent. Its mathematical model is:

[0071]

[0072] In the formula: down(·) is the downsampling function; β l j is the multiplicative bias. The downsampling function uses average pooling sampling.

[0073] Step 4: The processed data is input into the SVM support vector machine for fault diagnosis

[0074] Specifically, the following mathematical model is adopted;

[0075]

[0076] ξ i ≥0, i = 1, 2, …, N

[0077] In the formula: ω is the optimal plane normal vector; C is the penalty factor; ξ i is the relaxation factor; m i is the category of the i-th sample; l i represents the feature vector of the i-th input sample; b is the classification threshold.

[0078] Here, the Gaussian radial basis function (RBF) kernel function is used to construct the decision function f(l);

[0079]

[0080] In the formula: sgn is the sign function; α i * is the weight coefficient; exp is the exponential function with base e; l is the feature vector of the sample to be measured; γ is the parameter of the decision function f(l); β* is the optimal threshold.

[0081] Step 4: Optimize the remaining parameters of the model using the validation set data to obtain the optimal diagnostic model;

[0082] Step 5: Test the RIME-CNN-SVM model using the test set data, obtain the test results, and save the optimal RIME-CNN-SVM model.

[0083] Preferably, the specific steps of the fifth step are as follows:

[0084] Step 1: Input the vibration signal of the transformer collected in real time into the optimal RIME-CNN-SVM model obtained in the offline training stage to realize the judgment of the fault type.

[0085] The beneficial effect of the present invention mainly overcomes the problem of low diagnostic accuracy in the prior art due to the imbalance of the transformer fault vibration signal sample quantity and poor feature extraction. Brief Description of the Drawings

[0086] Figure 1 is the flow block diagram of the method involved in the present invention;

[0087] Figure 2 is the fuzzy entropy value diagram of the PF component of some signals;

[0088] Figure 3 is the comparison diagram of some parameters between the vibration signal generated by the DCGAN and the real vibration signal;

[0089] Figure 4 is the diagnostic result of some vibration signals. Detailed Embodiment

[0090] In order to make the purpose, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the drawings and specific embodiments. The specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0091] Such as Figures 1-4As shown in the figure, an online fault diagnosis method for transformers based on the CNN-SVM model with LMD decomposition and frost ice optimization. In the first step, the collected vibration signal is segmented and decomposed into k PF components. The specific steps are as follows:

[0092] S101: Segment and decompose the collected vibration signal into k PF components;

[0093] Specifically, calculate the average value mi of adjacent local extreme points and the envelope estimation value ai, smooth the adjacent local extreme value mi and the adjacent envelope estimation value ai to obtain the local mean function m11(t) and the envelope estimation function a11(t), then separate h11(t) from the original signal X(t), demodulate h11(t) to obtain s11(t), multiply a1n(t) by s1n(t) to obtain the first δ pF1 (t)

[0094] δ pF1 (t) = a 1 (t)s 1n (t)

[0095] Separate PF1 from the original signal X(t) to obtain the difference signal u1(t). At this time, the difference signal is used as the original signal and the above steps are repeated until the last difference signal becomes a monotonic function and stops. At this time, X(t) is decomposed into K PF components and a residual component uk(t):

[0096]

[0097] S102; Calculate the fuzzy entropy values of each PF component;

[0098] Take the PF components selected by the local mean decomposition algorithm as the input vector of the fuzzy entropy algorithm, denoted as u(i). Generate a set of m-dimensional vectors {Xmi} in the order of u(i), that is

[0099]

[0100] where u0(i) represents the average value of m consecutive u values.

[0101] Assume d m ij is the maximum value of the difference between the corresponding elements of two m-dimensional vectors X m i and X m j That is:

[0102]

[0103] Given parameters n and r, through the fuzzy function μ(dm ij , n, r), the similarity D between Xmi and Xmj can be calculated m ij , that is

[0104]

[0105] where n and r respectively represent the gradient and width of the boundary of the function μ(d m ij , n, r).

[0106] Define the function Q m (n, r), and regenerate a set of m + 1-dimensional vectors Q according to the above steps m+1 , that is:

[0107]

[0108] Obtain the negative natural logarithm of the deviation of the function Qm from Qm+1. When the length of the data set is limited, the fuzzy entropy can be equivalent to

[0109] FuzyEn(m,n,r) = lnQ m (n,r) - lnQ m+1 (n,r)

[0110] After calculation, the fuzzy entropy values of the selected partial vibration signal PF components are almost all higher than 0.4 in the first 3 PF components, while the entropy values of the subsequent PF components are almost all less than 0.3, and the entropy value of the PF4 component hovers around 0.6. This indicates that the first 3 PF components contain a lot of noise signals, the subsequent PF components can be regarded as real components, and it is very appropriate to use 0.4 as the threshold of the fuzzy entropy.

[0111] S103; Screen out the PF components with entropy values less than 0.6 for signal reconstruction to obtain the reconstructed signal x(tn). Specifically,

[0112]

[0113] Second step, sample augmentation and its equalization: After performing convolutional pooling on the reconstructed signal, input it into the generator, and optimize the generator parameters through the judgment and feedback of the discriminator to make the data generated by the generator realistic enough, and finally output the augmented data set;

[0114] S201: Denote the distribution of the random noise z = {z(1), z(2),..., z(m)} as P z (z), and denote the distribution of the real data sample x(tn) = {x(t1), x(t2),..., x(tm)} as P data (x), and denote the distribution of the generated sample x' = G(z) as Pg (x), the purpose of training G is to make P data (x) and P g (x) as identical as possible.

[0115] The input of D is x(tn) and x′. D(x) represents the probability that x comes from a real sample rather than from Pg. The purpose of training D is to maximize the accuracy of D in discriminating whether the input sample is true or false, so that real samples are classified as 1 and fake samples are classified as 0, that is, the value of D(x) approaches 1 and the value of D(x′) approaches 0.

[0116] S202: The generator G outputs the vibration signal to the discriminator D, and the discriminator calculates its similarity degree and then feeds back to train the generation of vibration signal samples;

[0117] Specifically, keep the parameters of G unchanged, then send the generated sample x′ and the real sample x into D, and use the stochastic gradient ascent method to update the parameters of D through the following formula;

[0118]

[0119] Then train G, keep the parameters of D unchanged, and use the stochastic gradient descent method to update the parameters of G through the following formula

[0120]

[0121] Until the two reach the Nash equilibrium P g = P data , at this time D(x) = D(x′) = 0.5, that is, the discriminator randomly discriminates the true or false of the sample with a probability of 50%.

[0122] The third step, optimize the model parameters: use the rime ice optimization algorithm to optimize the penalty factor c and the kernel function parameter h of CNN-SVM to obtain the optimal parameters;

[0123] S301: Initialize the rime ice population. The types of individuals in the population include the penalty factor c and the kernel function parameter h. The position of the RIME rime ice particle population R is as follows:

[0124]

[0125] In the formula: the rime ice particle population can be represented by x ij represents, i represents the serial number of the rime ice crystal, and j represents the serial number of the rime ice particle.

[0126] S302: Calculate the fitness of the individuals in the population and select a part of the individuals with higher fitness as the crystallization nuclei

[0127] Specifically, the fitness values F(S i ) of all rime ice particles are as follows:

[0128]

[0129] Let \(f\) be the fitness of each frost ice particle. Regarding the motion properties of the frost ice particles, the position of the frost ice particles is shown as follows:

[0130]

[0131] In the formula: \(R\) new ij is the position after particle update, \(R\) best.j is the \(j\)-th particle of the optimal crystal in the frost ice population \(R\), \(r\) 1 is a random number in \((-1, 1)\), which controls the moving direction of the particle and will change with the iteration of the moving direction \(\cos\theta\). \(\beta\) is the environmental factor representing the external influencing factor, \(\psi\) is the particle adhesion force, which is a random number in \((0, 1)\) and is used to control the distance between particles, \(U\) bij and \(U\) bij represent the upper and lower bounds of the particle motion space respectively.

[0132] S303: For each "crystal nucleus", generate a certain number of new individuals through random mutation operations. According to the fitness values of the new individuals, select a part of the new individuals to replace the original individuals to form a new population, and the individuals in the population will gradually approach a better solution;

[0133] S304: Determine whether the termination condition is met, such as reaching the maximum number of iterations or finding a sufficiently good solution. If the termination condition is met, stop the search; otherwise, return to step 2 to continue iterative optimization.;

[0134] Specifically, in order to ensure convergence, it is necessary to set as shown in the following formula:

[0135]

[0136] In the formula: \(h\) is the current number of iterations, and \(H\) is the maximum number of iterations.

[0137]

[0138] In the formula: \(\beta\) is the step function, \(w\) is used to control the number of segments of the step function, \(E\) is the adhesion coefficient, which will increase with the number of iterations and affect the condensation probability of the particles; \(r2\) is a random number in \((0, 1)\), and it and \(E\) control whether the particles condense, that is, whether the particle position is updated. The expression of \(E\) is:

[0139]

[0140] S305: Substitute the obtained optimal penalty factor \(c\) and the optimal kernel function parameter \(h\) into CNN - SVM to obtain the optimized RIME - CNN - SVM;

[0141] Step 4, Model Training: Input the training set, validation set, and test set into RIME-CNN-SVM for offline model training to construct an optimal transformer fault diagnosis model. The specific steps are as follows:

[0142] S401: Add labels to the sample signals and divide them into a training set, a test set, and a validation set according to a ratio of 7:2:1, and use the training set to train the RIME-CNN-SVM model;

[0143] S402: Use the frost ice optimization algorithm to optimize the penalty factor c and the kernel function parameter h of the convolutional neural network support vector machine;

[0144] S403: Perform convolutional pooling on the training set samples;

[0145] Specifically, the result of the convolutional operation is output through the activation function:

[0146]

[0147] In the formula: f(·) is the convolutional activation function; M j is the input feature map set; x j l-1 is the (l-1)-th channel of the input feature map; k l ij is the weight matrix of the convolution; * is the convolution operation; b l j is the bias term.

[0148] After each convolutional layer, a downsampling layer is applied. The downsampling layer reduces the size of the input features and the number of parameters in the network, not only improving the accuracy but also reducing the dimension of the extracted features, and preventing overfitting to a certain extent. Its mathematical model is:

[0149]

[0150] In the formula: down(·) is the downsampling function; β l j is the multiplicative bias. The downsampling function uses average pooling sampling.

[0151] S404: Input the processed data into the support vector machine for fault diagnosis

[0152] Specifically, the following mathematical model is adopted;

[0153]

[0154] ξ i ≥0, i = 1, 2, …, N

[0155] where: ω is the optimal plane normal vector; C is the penalty factor; ξ i is the relaxation factor; m i is the category of the i-th sample; l i represents the feature vector of the i-th input sample; b is the classification threshold.

[0156] Here, the Gaussian radial basis function (RBF) kernel function is used to construct the decision function f(l);

[0157]

[0158] where: sgn is the sign function; α i * is the weight coefficient; exp is the exponential function with base e; l is the feature vector of the sample to be measured; γ is the parameter of the decision function f(l); β* is the optimal threshold.

[0159] Step 5, input the real-time fault signal into the optimal model to complete the on-line diagnosis of transformer faults. The specific steps are as follows:

[0160] S501: Input the vibration signal of the transformer collected in real time into the optimal RIME-CNN-SVM model obtained in the off-line training stage to realize the judgment of the fault type.

[0161] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; thus, these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope defined by the claims of the present invention.

Claims

1. A transformer fault online diagnosis method based on CNN-SVM model with LMD decomposition and frost optimization, characterized in that: The following steps are involved: The first step is transformer vibration signal preprocessing: the vibration signal is decomposed by LMD to obtain multiple PF components, and the fuzzy entropy value of each PF component is calculated, and the PF component with a fuzzy entropy value higher than the threshold is selected to reconstruct the signal; The second step is sample expansion and equalization: the reconstructed signal is processed by convolution and pooling and then input into the generator. The discriminator judges and feeds back to optimize the generator parameters so that the data generated by the generator is realistic enough. Finally, the expanded data set is output. The third step is to optimize the model parameters: use the frost optimization algorithm to optimize the penalty factor c of CNN-SVM and the kernel function parameter h to obtain the optimal parameters; Step 4: Model training: Divide the balanced samples into training set, validation set and test set and input them into RIME-CNN-SVM for offline model training to build the optimal transformer fault diagnosis model; Step 5: Online fault diagnosis: Input the optimal model with real-time fault signals to complete online diagnosis of the transformer fault.

2. The transformer fault online diagnosis method based on CNN-SVM model of LMD decomposition and frost optimization according to claim 1 is characterized in that: The specific steps of the first step are as follows: Step 1: Decompose the collected vibration signal into k PF components; Specifically, the average value mi and the envelope estimation value ai of the adjacent local extreme value points are calculated, and the adjacent local extreme value mi and the adjacent envelope estimation value ai are smoothed to obtain the local mean function m11(t) and the envelope estimation function a11(t). Then h11(t) is separated from the original signal X(t), h11(t) is demodulated to obtain s11(t), and a1n(t) is multiplied by s1n(t) to obtain the first δ pF1 (t) δ pF1 (t)=a1(t)s 1n (t) Separate PF1 from the original signal X(t) to obtain the difference signal u1(t); the difference signal is used as the original signal again, and the above steps are repeated until the last difference signal becomes a monotonic function; at this time, X(t) is decomposed into K PF components and a residual component uk(t): Step 2: Calculate the fuzzy entropy value of each PF component and select the appropriate threshold; Specifically, the PF component selected by the local mean decomposition algorithm is used as the input vector of the fuzzy entropy algorithm, denoted as u(i); a set of m-dimensional vectors {Xmi} are generated in sequence according to the arrangement order of u(i), namely Where u0(i) represents the average value of m consecutive u values; Assumption d m ij are two m-dimensional vectors X m i and X m j The maximum value of the distance difference between the two corresponding elements in , that is: Given parameters n and r, through the fuzzy function μ(d m ij , n, r) can calculate the similarity D between Xmi and Xmj m ij ,Right now Where n and r represent the function μ(d m ij , n, r) the gradient and width of the boundary; Define the function Q m (n, r), and regenerate a set of m+1-dimensional vectors Q according to the above steps m+1 ,Right now: The negative natural logarithm of the deviation of function Qm is obtained from Qm+1; when the length of the data set is finite, the fuzzy entropy can be equivalent to FuzyEn(m,n,r)=lnQ m (n,r)-lnQ m+1 (n,r) Step 3: Filter out the PF components whose entropy values ​​meet the requirements for signal reconstruction to obtain the reconstructed signal x(tn); Specifically, the PF component whose fuzzy entropy value meets the requirements is selected to reconstruct the signal according to the following formula; 3. The online transformer fault diagnosis method based on CNN-SVM model of LMD decomposition and frost optimization according to claim 1 is characterized in that: The specific steps of the second step are as follows: Step 1: Input the real signal into the generator G, and G randomly generates a vibration signal; Specifically, the real data sample x(tn) is input into the generator G, and the generator generates a random vibration signal sample x′; The input of D is x(tn) and x′. D(x) represents the probability that x comes from the real sample instead of Pg. The purpose of training D is to maximize the accuracy of D in distinguishing the true and false input samples, so that the real samples are classified as 1 and the false samples are classified as 0, that is, the value of D(x) approaches 1 and the value of D(x′) approaches 0. Step 2: The generator G outputs the vibration signal to the discriminator D, which calculates its authenticity and then feeds back to train and optimize the generator until the samples generated by the generator are realistic enough; Specifically, keep the parameters of G unchanged, then send the generated sample x′ and the real sample x into D, and use the stochastic gradient ascent method to update the parameters of D through the following formula; Then train G, keep the parameters of D unchanged, and use the stochastic gradient descent method to update the parameters of G through the following formula Until the two reach Nash equilibrium P g =P data , at this time, D(x)=D(x′)=0.5, that is, the discriminator randomly judges the authenticity of the sample with a probability of 50%.

4. The online transformer fault diagnosis method based on CNN-SVM model of LMD decomposition and frost optimization according to claim 1 is characterized in that: The specific steps of the third step are as follows: Step 1: Initialize the rime population. Specifically, the individual types in the population include the penalty factor c, the kernel function parameter h, and the position of the RIME frost particle population R as shown below: Where: The frost particle population can be obtained through x ij Indicates, i represents the serial number of the frost ice crystal, and j represents the serial number of the frost ice particle; Step 2: Calculate the fitness of individuals in the population and select a portion of individuals with higher fitness as crystallization nuclei; Specifically, the fitness value F(S i ) is shown in the following formula: f is the fitness of each frost particle. According to the motion properties of frost particles, the position of frost particles is as follows: Where: R new ij is the updated position of the particle, R best.j is the jth particle of the optimal crystal in the frost population R, r1 is a random number (-1, 1) that controls the moving direction of the particle and changes with the iteration of the particle moving direction cosθ, β is the environmental factor representing the external influencing factors, ψ is the particle adhesion force, which is a random number (0, 1) used to control the distance between particles, and U bij , U bij They represent the upper and lower bounds of the particle motion space respectively; Step 3: For each "crystallization core", a certain number of new individuals are generated through random mutation operations. According to the fitness values ​​of the new individuals, some new individuals are selected to replace the original individuals to form a new population. The individuals in the population will gradually approach a better solution; Step 4: Determine whether the termination condition is met, such as reaching the maximum number of iterations or finding a good enough solution; if the termination condition is met, stop searching; otherwise, return to step 2 to continue iterative optimization; Specifically, to ensure convergence, the following formula needs to be set: Where: h is the current iteration number, H is the maximum iteration number; Where: β is a step function, w is used to control the number of segments of the step function, E is the adhesion coefficient, which increases with the number of iterations and affects the probability of particle condensation; r2 is a random number (0, 1), which controls whether the particle condenses with E, that is, whether the particle position is updated; the expression of E is: Step 5: Substitute the obtained optimal penalty factor c and optimal kernel function parameter h into CNN-SVM to obtain the optimized RIME-CNN-SVM.

5. The transformer fault online diagnosis method based on the CNN-SVM model of LMD decomposition and frost optimization according to claim 1 is characterized in that: The specific steps of the fourth step are as follows: Step 1: Label the sample signals and divide them into training set, test set and validation set in a ratio of 7:2:

1. Use the training set to train the RIME-CNN-SVM model. Step 2: Use the frost optimization algorithm to optimize the penalty factor c and kernel function parameter h of the convolutional neural network support vector machine; Step 3: Perform convolution pooling on the training set samples; Specifically, the result of the convolution operation is output through the activation function: Where: f(·) is the convolution activation function; M j is the input feature atlas; x j l-1 is the l-1th channel of the input feature map; k l ij is the weight matrix of the convolution; * is the convolution operation; b l j is the bias term; A downsampling layer is applied after each convolutional layer. The downsampling layer reduces the size of the input features and the number of parameters in the network, which not only improves the accuracy but also reduces the dimension of the extracted features, thus preventing overfitting to a certain extent. Its mathematical model is: Where: down(·) is the downsampling function; β l j is a multiplication bias; the downsampling function uses average pooling sampling; Step 4: The processed data is input into the SVM support vector machine for fault diagnosis Specifically, the following mathematical model is adopted; Where: ω is the optimal plane normal vector; C is the penalty factor; ξ i is the relaxation factor; m i is the category of the i-th sample; l i represents the feature vector of the i-th input sample; b is the classification threshold; The decision function f(l) is constructed using the Gaussian radial basis function (RBF) kernel function; Where: sgn is the sign function; α i * is the weight coefficient; exp is the exponential function with e as the base; l is the feature vector of the sample to be tested; γ is the parameter of the decision function f(l); β* is the optimal threshold; Step 4: Use the validation set data to optimize the remaining parameters of the model to obtain the optimal diagnostic model; Step 5: Use the test set data to test the RIME-CNN-SVM model, obtain the test results and save the RIME-CNN-SVM optimal model.

6. The online transformer fault diagnosis method based on CNN-SVM model of LMD decomposition and frost optimization according to claim 1 is characterized in that: The specific steps of the fifth step are as follows: Step 1: Input the transformer vibration signal collected in real time into the optimal RIME-CNN-SVM model obtained in the offline training phase to realize fault type judgment.

Images