Transformer early fault diagnosis method based on ISSA-MKELM

Abstract

The present invention discloses a transformer early fault diagnosis method based on ISSA-MKELM, which includes the following steps: collecting data of transformer operation in different states to obtain corresponding envelope spectrum signals, normalizing the corresponding envelope spectrum signals, and randomly dividing the processed data into training data and test data; extracting early fault characteristics of different categories from the input data, the number of hidden layer nodes of the network, and the learning rate of each layer of DA, and optimizing them using the ISSA algorithm; constructing a hybrid kernel extreme learning machine (MKELM), introducing Bernoulli chaotic mapping to optimize the initial sparrow population, using an improved sparrow search algorithm (ISSA) to optimize the parameters of the hybrid kernel extreme learning machine (MKELM), establishing an ISSA-MKELM fault diagnosis model for testing, and calculating the diagnostic accuracy of the model for test samples. Compared with the existing technology, the present invention has the advantages of good diagnostic accuracy and robustness.

Description

Technical Field

[0001] The present invention relates to the technical field of transformer fault diagnosis, and in particular to an ISSA-MKELM-based transformer early fault diagnosis method. Background Art

[0002] Transformers, especially large ones, are critical components of modern power grids. Their operational reliability is directly linked to the safety and stability of grid operations. Mechanical failures within transformers are latent. When mechanical performance degrades to a certain level, they can lead to a series of chain reactions, including electrical and insulation failures, and can even develop into catastrophic failures. Among mechanical failures, transformer windings and cores are the most prone to failure. Therefore, diagnosing latent faults in transformer windings and cores is crucial.

[0003] Currently, there are various methods for diagnosing transformer faults, such as chromatography, insulation performance testing, partial discharge testing, DC leakage current measurement, and low-voltage short-circuit impedance measurement. However, these methods are ineffective for diagnosing localized faults. With the continuous development of transformer diagnostic technology, dissolved gas analysis (DGA) has received extensive attention and research. However, DGA can only determine whether a transformer fault exists, but cannot determine the severity or location of the fault. Furthermore, it has little effect on diagnosing the most destructive mechanical faults.

[0004] At the same time, with the rapid development of modern sensing and signal processing technologies, various condition monitoring principles, sensing technologies, and signal processing techniques have been applied to power equipment fault diagnosis. Common methods include the short-circuit impedance method, capacitance change method, low-voltage pulse method, frequency response method, ultrasonic method, and vibration analysis method. Vibration analysis is a non-destructive online monitoring method. The sensor is placed on the outer wall of the transformer oil tank, with no direct electrical connection to the transformer, which does not affect the normal operation of the transformer. This is the advantage of vibration analysis and is currently the mainstream analysis method. However, the collected transformer vibration signal inevitably contains a certain amount of noise, and in the early stages of transformer failure, the fault characteristics are not obvious, making feature extraction difficult. Common feature extraction methods such as principal component analysis, wavelet analysis, and blind source separation are shallow feature extraction methods that are difficult to achieve for early transformer fault signals. Therefore, how to provide a transformer early fault diagnosis method based on ISSA-MKELM is an urgent problem that those skilled in the art need to solve. Summary of the Invention

[0005] An object of the present invention is to propose a transformer early fault diagnosis method based on ISSA-MKELM. Compared with the prior art, the present invention has the advantages of good diagnostic accuracy and robustness.

[0006] A transformer early fault diagnosis method based on ISSA-MKELM according to an embodiment of the present invention includes the following method steps:

[0007] Data on transformer operation under different conditions are collected. Early vibration signals of different transformer faults are preprocessed to obtain corresponding envelope spectrum signals. The corresponding envelope spectrum signals are normalized and randomly divided into training data and test data.

[0008] Extract the early fault characteristics of different categories from the input data, the number of hidden layer nodes of the network and the learning rate of each layer of DA, and optimize them through the ISSA algorithm;

[0009] A hybrid kernel extreme learning machine (MKELM) was constructed, and the Bernoulli chaotic map was introduced to optimize the initial sparrow population. The Cauchy-Gaussian mutation strategy and nonlinear dynamic inertia weight were used, and the improved sparrow search algorithm (ISSA) was adopted to optimize the parameters of the hybrid kernel extreme learning machine (MKELM). An ISSA-MKELM fault diagnosis model was established.

[0010] The early fault diagnosis model ISSA-MKELM is tested using test data, and the diagnostic accuracy of the model for the test samples is calculated.

[0011] Optionally, the transformer is divided into nine states of operating data, the training data includes SSDAE pre-training and global fine-tuning SSDAE, and the training data is used to train a deep feature extraction neural network SSDAE.

[0012] Optionally, the preprocessing of vibration signals of different transformer faults to obtain corresponding envelope spectrum signals further includes:

[0013] The original vibration signal x(t) is decomposed using CEEMD to obtain several IMFs and residual terms;

[0014] Select the IMF with higher correlation and calculate the Hilbert spectrum;

[0015] The corresponding envelope value is obtained based on the Hilbert spectrum;

[0016] The obtained envelope value is subjected to Fourier transform to obtain the corresponding envelope spectrum.

[0017] Optionally, the hybrid kernel extreme learning machine MKELM is constructed by a Gaussian kernel function and a polynomial kernel function.

[0018] Optionally, the unsupervised stacked sparse denoising autoencoder SSDAE pre-training further includes:

[0019] The mean square error function is selected as the reconstruction error function of SSDAE, and a batch processing strategy is used to process large-scale training sample sets;

[0020] After obtaining the error function, the batch stochastic gradient descent method is used to find the minimum value of the error function and the partial derivatives of each parameter with respect to the error function are obtained;

[0021] The network activation function selects the sigmoid function, and its first-order derivative and second-order derivative function are obtained;

[0022] By expressing SSDAE through weights and bias terms, we can get the effect of each parameter on z ik The partial derivative of

[0023] Use the chain rule to find the partial derivatives of each parameter with respect to the loss function;

[0024] The parameter set is updated according to the chain rule and the gradient descent method, and the parameter set is updated according to the update rule until the error convergence criterion is met. After the training is completed, the reconstruction layer is removed, and the output result of the last hidden layer is the result of feature extraction.

[0025] Optionally, the supervised global fine-tuning SSDAE further includes:

[0026] After the SSDAE pre-training is completed, the SOFTMAX classifier is added to the last hidden layer of SSDAE to form a complete network;

[0027] Input labeled data and use the backpropagation algorithm to perform global training on the entire network from top to bottom. The partial derivative of the classifier loss function will propagate downward layer by layer, making slight adjustments to the network weights originally initialized by SSDAE.

[0028] The partial derivatives of the classifier weight parameters with respect to the loss function are used to fine-tune the parameters according to the gradient descent rule:

[0029] According to the back-propagation algorithm, the relevant weights of all layers of the network are adjusted from top to bottom until the convergence criteria are met.

[0030] Optionally, the optimizing the hybrid kernel extreme learning machine classification model using the ISSA algorithm further includes:

[0031] Use hybrid kernel function to realize the mapping from input space to high-dimensional hidden layer feature space;

[0032] In the population initialization stage, Bernoulli chaotic mapping is introduced to generate chaotic sequences;

[0033] After obtaining the chaotic sequence that traverses the entire two-dimensional space, the chaotic particles are mapped to the search range required for environmental modeling;

[0034] Dynamic inertia weights are introduced to optimize the position update method of sparrow population discoverers, reconcile the global optimization and local mining capabilities of the equilibrium algorithm, and in the early stage of algorithm evolution, the main focus is on exploring the search space to reach a better area and find the optimal solution.

[0035] Optionally, the mathematical expression of the Bernoulli map is:

[0036]

[0037] Where: Z k is the current value of the kth generation chaotic sequence, and λ is the control parameter;

[0038] Optionally, after obtaining the chaotic sequence that traverses the entire two-dimensional space, the chaotic particles are mapped to the search range required for environmental modeling, that is:

[0039]

[0040] Where, X min and X max The minimum and maximum values ​​of the horizontal axis are set to 10 and 200 respectively; min and Y max Set the minimum and maximum values ​​of the vertical axis to -10 and 10 respectively; is the position of the kth generation sparrow.

[0041] Optionally, the finder position with the dynamic inertia weight w added is updated as follows:

[0042]

[0043]

[0044] Where t represents the current iteration number, j = 1, 2, ....., d, where d represents the dimension of the solution, iter max is the maximum number of iterations. i,j represents the position of the i-th sparrow in the j-th dimension, α∈(0,1] is a random number, R2(R2∈[0,1]) and ST(ST∈[0.5,1]) represent the warning value and safety value respectively, Q is a random number that obeys the normal distribution, L is a 1×d matrix of all ones; w start is the initial inertia weight; w end is the inertia weight at the maximum number of iterations.

[0045] The beneficial effects of the present invention are:

[0046] (1) The present invention adopts a deep learning technology-based SSDAE-Softmaxs deep feature extraction network to perform feature extraction, which solves the problem that traditional feature extraction methods cannot obtain deep-level features of vibration data;

[0047] (2) The present invention addresses the defects of the kernel extreme learning machine classifier KELM and combines the Gaussian kernel function and the polynomial kernel function to construct a hybrid kernel extreme learning machine MKELM with stronger generalization ability;

[0048] (3) In order to solve the problem of random initialization of the sparrow search algorithm and its easy falling into local extreme values, the present invention introduces Bernoulli chaos mapping to optimize the initial sparrow population, and uses Cauchy-Gauss mutation strategy and nonlinear dynamic inertia weight to improve the problem of the sparrow search algorithm easily falling into local optimal value. Then, the improved sparrow search algorithm ISSA is used to optimize the parameters of the hybrid kernel extreme learning machine MKELM and establish a transformer early fault diagnosis model based on ISSA-MKELM. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In the attached figure:

[0050] Figure 1 This is a flowchart of a transformer early fault diagnosis method based on ISSA-MKELM proposed in the present invention;

[0051] Figure 2 This is a structural diagram of the automatic encoder in the transformer early fault diagnosis method based on ISSA-MKELM proposed in the present invention;

[0052] Figure 3 This is a structural diagram of SOFTMAX in the transformer early fault diagnosis method based on ISSA-MKELM proposed in the present invention;

[0053] Figure 4 This is a feature extraction flow chart based on SSDAE in a transformer early fault diagnosis method based on ISSA-MKELM proposed in the present invention;

[0054] Figure 5 A comparison chart of the ISSA optimization algorithm in the ISSA-MKELM-based transformer early fault diagnosis method proposed in the present invention;

[0055] Figure 6 This is a two-dimensional distribution diagram of the original data feature distribution in the transformer early fault diagnosis method based on ISSA-MKELM proposed in the present invention;

[0056] Figure 7This is a two-dimensional feature distribution diagram after t-SNE processing in the transformer early fault diagnosis method based on ISSA-MKELM proposed in the present invention;

[0057] Figure 8 This is the Bernoulli chaotic sequence distribution in the transformer early fault diagnosis method based on ISSA-MKELM proposed by the present invention;

[0058] Figure 9 The probability density function curves of Cauchy distribution and Gaussian distribution in the transformer early fault diagnosis method based on ISSA-MKELM proposed by the present invention;

[0059] Figure 10 This is an extreme learning machine network model in the transformer early fault diagnosis method based on ISSA-MKELM proposed by the present invention;

[0060] Figure 11 This is a diagram of the ISSA-MKELM diagnosis results in the ISSA-MKELM-based transformer early fault diagnosis method proposed in the present invention;

[0061] Figure 12 This is a diagram of the SSA-MKELM diagnosis results in the transformer early fault diagnosis method based on ISSA-MKELM proposed in the present invention;

[0062] Figure 13 This is a diagram of the SSA-KELM diagnosis results in the transformer early fault diagnosis method based on ISSA-MKELM proposed in the present invention;

[0063] Figure 14 This is the SSA-SVM diagnosis result diagram in the transformer early fault diagnosis method based on ISSA-MKELM proposed in this invention. DETAILED DESCRIPTION

[0064] refer to Figure 1 A transformer early fault diagnosis method based on ISSA-MKELM includes the following steps:

[0065] Data from transformers operating under different conditions were collected. Early vibration signals of different transformer faults were preprocessed to obtain corresponding envelope spectrum signals. The corresponding envelope spectrum signals were normalized and randomly divided into training data and test data. The transformer operating data was divided into nine conditions. The training data included SSDAE pre-training and global fine-tuning SSDAE. The training data was used to train the deep feature extraction neural network SSDAE.

[0066] In this embodiment, the vibration signals of different transformer faults are preprocessed to obtain corresponding envelope spectrum signals, which also includes:

[0067] The original vibration signal x(t) is decomposed using CEEMD to obtain several IMFs and residual terms;

[0068]

[0069] Among them, c i (t) is IMF, r n is the residual item.

[0070] Select an IMF with higher correlation. The correlation between the IMF and the original signal can be calculated using the following formula:

[0071]

[0072] Among them, s i It represents the correlation between the ith IMF and the original signal. When it is greater than 0.5, it can be considered to be highly correlated.

[0073] The Hilbert spectrum is calculated from the relevant IMF obtained in the previous step:

[0074]

[0075] Based on the Hilbert spectrum, the corresponding envelope value can be obtained by the following formula:

[0076]

[0077] The obtained envelope value is subjected to Fourier transform to obtain the corresponding envelope spectrum.

[0078] Extract the early fault characteristics of different categories from the input data, the number of hidden layer nodes of the network and the learning rate of each layer of DA, and optimize them through the ISSA algorithm;

[0079] In this embodiment, the unsupervised stacked sparse denoising autoencoder (SSDAE) pre-training further includes: selecting a mean square error function as the reconstruction error function of the SSDAE, and processing a large-scale training sample set using a batch processing strategy. The error function of the batch sample set can be described as:

[0080]

[0081] In the formula, n represents the dimension of the input data, m represents the total number of samples processed in batches, and x ij represents the jth component in the i-th sample.

[0082] After obtaining the error function, the batch stochastic gradient descent method is used to find the minimum value of the error function and the partial derivatives of each parameter with respect to the error function are obtained;

[0083] The network activation function selects the sigmoid function, and its first-order derivative and second-order derivative function are obtained;

[0084] By expressing SSDAE through weights and bias terms, we can get the parameter pairs zi k Partial derivatives of :

[0085]

[0086] The expression of SSDAE can be expressed by weight and bias terms:

[0087]

[0088] Where u ip represents the weighted sum of all input components of the i-th training sample to the p-th unit of the hidden layer, v ik Represents the weighted sum of all input components of the i-th training sample to the k-th unit of the reconstruction layer.

[0089] Obtain the parameters for z ik Partial derivatives of :

[0090]

[0091]

[0092]

[0093] Where w rs is the weight from the rth unit in the input layer to the sth unit in the hidden layer, b hr is the bias term of r units in the hidden layer, b zr is the bias term of r unit in the reconstruction layer.

[0094] The partial derivatives of each parameter with respect to the loss function can be obtained by the chain rule:

[0095]

[0096] According to the chain rule and gradient descent, the update rule of the parameter set can be described as follows:

[0097]

[0098] The parameter set is updated according to the update rule described in the above formula until the error convergence criterion is met; after the training is completed, the reconstruction layer is removed, and the output result of the last hidden layer is the result of feature extraction.

[0099] In this implementation, the supervised global fine-tuning SSDAE also includes:

[0100] After SSDAE pre-training is complete, the SOFTMAX classifier is added to the last hidden layer of the SSDAE to complete the network. Labeled data is input and the entire network is trained from top to bottom using the backpropagation algorithm. The partial derivatives of the classifier loss function are propagated downward layer by layer, making minor adjustments to the originally initialized SSDAE network weights. This global fine-tuning further optimizes the feature extraction capabilities of the SSDAE network. The loss function of the SOFTMAX classifier is as follows:

[0101]

[0102] Where w and b represent the weight and bias of the classifier respectively, Q j represents the feature vector of the jth training sample, N represents the total number of failure modes, Y j is the class label of the training sample, 1{·} is the indicator function, which is 1 when the sample belongs to the i-th type of fault and 0 otherwise.

[0103] The partial derivative of the classifier weight parameter with respect to the loss function can be obtained:

[0104]

[0105] According to the gradient descent rule, fine-tune the parameters:

[0106]

[0107] According to the back-propagation algorithm, the relevant weights of all layers of the network are adjusted from top to bottom until the convergence criteria are met.

[0108] Aiming at the defects of the kernel extreme learning machine classifier KELM, a hybrid kernel extreme learning machine MKELM with stronger generalization ability is constructed by combining Gaussian kernel function and polynomial kernel function. Secondly, to address the problems of random initialization of population and easy falling into local extreme values ​​of the sparrow search algorithm, Bernoulli chaos mapping is introduced to optimize the initial sparrow population. The problem of the sparrow search algorithm easily falling into local optimal value is improved by using Cauchy-Gauss mutation strategy and nonlinear dynamic inertia weight. Then, the improved sparrow search algorithm ISSA is used to optimize the parameters of the hybrid kernel extreme learning machine MKELM and an ISSA-MKELM fault diagnosis model is established.

[0109] In this embodiment, the ISSA algorithm is used to optimize the hybrid kernel extreme learning machine classification model, which also includes:

[0110] The classification performance of the kernel extreme learning machine depends largely on the kernel function. Generally speaking, a single global kernel function or a local kernel function has its limitations and cannot strike a good balance between fitting and generalization. Since the Gaussian kernel function is a typical global kernel function and the polynomial kernel function is a typical local kernel function, it is possible to consider mixing these two kernel functions and replacing the original single kernel function with a hybrid kernel function. This can give the model better learning and generalization capabilities and more accurately achieve the mapping from the input space to the high-dimensional hidden layer feature space. The specific expression of the hybrid kernel function is:

[0111]

[0112] Where: δ, n, d and λ are the parameters of the mixing kernel function;

[0113] In the population initialization stage, Bernoulli chaotic mapping is introduced to generate chaotic sequences;

[0114] After obtaining the chaotic sequence that traverses the entire two-dimensional space, the chaotic particles are mapped to the search range required for environmental modeling;

[0115] Dynamic inertia weights are introduced to optimize the position update method of sparrow population discoverers, reconcile the global optimization and local mining capabilities of the equilibrium algorithm, and in the early stage of algorithm evolution, the main focus is on exploring the search space to reach a better area and find the optimal solution.

[0116] In this embodiment, the traditional sparrow algorithm uses random numbers to initialize the population positions during the initialization phase of the sparrow population. This results in a high and low convergence rate due to the uneven positions of the sparrows in the population and the inability to traverse all positions in the environment. Therefore, the present invention proposes to introduce the Bernoulli chaotic map during the population initialization phase to generate a chaotic sequence. This chaotic sequence has the characteristics of nonlinearity, ergodicity, and randomness, and can achieve better optimization results than random numbers. The mathematical expression of the Bernoulli map can be expressed as:

[0117]

[0118] Where: Z k is the current value of the kth generation chaotic sequence, and λ is the control parameter;

[0119] In this embodiment, after obtaining the chaotic sequence that traverses the entire two-dimensional space, the chaotic particles are mapped to the search range required for environmental modeling, that is:

[0120]

[0121] Where, X min and X maxThe minimum and maximum values ​​of the horizontal axis are set to 10 and 200 respectively; min and Y max Set the minimum and maximum values ​​of the vertical axis to -10 and 10 respectively; is the position of the kth generation sparrow.

[0122] In this embodiment, the finder position is updated with the dynamic inertia weight w as:

[0123]

[0124]

[0125] Where t represents the current iteration number, j = 1, 2, ....., d, where d represents the dimension of the solution, iter max is the maximum number of iterations. i,j represents the position of the i-th sparrow in the j-th dimension, α∈(0,1] is a random number, R2(R2∈[0,1]) and ST(ST∈[0.5,1]) represent the warning value and safety value respectively, Q is a random number that obeys the normal distribution, L is a 1×d matrix of all ones; w start is the initial inertia weight, which is 0.9 in this embodiment; w end is the inertia weight at the maximum number of iterations, which is 0.4 in this embodiment.

[0126] The early fault diagnosis model ISSA-MKELM is tested using test data, and the diagnostic accuracy of the model for the test samples is calculated. The above method can complete the early fault diagnosis of transformers.

[0127] refer to Figure 2 In this embodiment, the autoencoder is an unsupervised neural network model, including an input layer, a hidden layer, and an output layer. The input layer and the hidden layer constitute the encoder, and the hidden layer and the output layer constitute the decoder.

[0128] refer to Figure 3 ,The SSDAE network after adding the Softmax classifier is still a multi-layer neural network in essence, so the error back propagation algorithm can be used for training.

[0129] refer to Figure 4SSDAE-Softmax is established by stacking multiple SDAEs and Softmax classifiers with a global fine-tuning strategy using a hierarchical pre-training process. With the assistance of the supervised training method of the Softmax classifier, SSDAE's ability to extract deep features of data is greatly enhanced. The method of extracting data features based on SSDAE mainly includes two steps: SSDAE pre-training and global fine-tuning. The pre-training process of SSDAE is an unsupervised learning process, which is to obtain SSDAE and initialize its model parameters; global fine-tuning is a supervised learning process for the global framework composed of SSDAE and Softmax, so as to obtain the optimal model parameters. After the global fine-tuning is completed, the Softmax classifier is removed, and the remaining SSDAE can be used for feature extraction, and its final output is the deep feature expression of the data.

[0130] In order to verify the effect of ISSA, this paper uses SSA, PSO and GWO three optimization algorithms for comparison. Figure 5 , which shows the change in fitness function values ​​during the iteration process of the above algorithms. As can be seen from the data in the figure, ISSA only needs about 10 iterations to converge and can obtain the minimum fitness function value, which shows the superiority of this algorithm.

[0131] In this embodiment, the SSDAE-Softmax network structure is built using the deep learning library in MATLAB. The best feature extraction effect cannot be achieved by setting the network parameters using experience. The present invention uses the ISSA algorithm to optimize the number of hidden layer units in the network and the learning rate of each layer DA and other hyperparameters to obtain the optimal feature extraction network model. The optimal parameter set is shown in Table 1:

[0132] Table 1 Optimal parameter set

[0133] Hidden layer structure Sparsity parameter DA1 learning rate DA2 learning rate 15370-8595 0.0527 0.1856 0.1747

[0134] refer to Figure 6-Figure 7 The number of hidden layer nodes, sparsity parameters, and learning rates generated by ISSA are used to construct an optimized SSDAE. This optimized network is then used for feature extraction. While feature extraction reduces dimensionality, the data remains high-dimensional, making it difficult to visualize the effects of feature extraction. To visualize the scatter distribution of features extracted from the nine fault instances across different fault categories, this implementation uses the t-SNE tool to visualize the raw data and extracted features, displaying their 2D distribution. The features extracted from the last hidden layer serve as the output features.

[0135] refer to Figure 8, is the Bernoulli chaotic sequence distribution diagram. The Bernoulli map is iterated 200 times to obtain the distribution of values ​​within the interval [0,1]. This invention proposes introducing the Bernoulli chaotic map during population initialization to generate a chaotic sequence. This chaotic map exhibits nonlinearity, ergodicity, and randomness. Using it to replace random numbers for population initialization affects the overall algorithm process, resulting in better optimization results than random numbers.

[0136] refer to Figure 9 , are the probability density function curves of the Cauchy distribution and the Gaussian distribution. As can be seen from the figure, the Gaussian distribution is more local on the horizontal axis, which is conducive to the algorithm efficiently finding the minimum point in the current search area. The Cauchy distribution is slightly smaller than the Gaussian distribution in the vertical direction, and has a larger distribution range in the horizontal direction. It is easier to generate a random number far away from the origin, which is conducive to helping the algorithm jump out of the local optimum and is more suitable for global search. Therefore, the present invention uses Cauchy perturbation in the early stage of algorithm optimization to improve the global search capability, and uses Gaussian perturbation in the later stage of optimization to perform a more detailed small-scale search to accelerate the convergence speed.

[0137] refer to Figure 10 , a new single-hidden-layer feedforward neural network learning algorithm, has been widely applied in many different fields. To address the shortcomings of ELM (Electronic Modeling Machine), which randomly sets the initial weights and biases of the hidden layer, leading to unstable classification accuracy, the kernel function concept from the support vector machine (SVM) was introduced into the ELM to propose the kernel extreme learning machine (KELM). However, single global kernel functions or local kernel functions have their limitations and cannot effectively balance the contradiction between fitting and generalization. Therefore, a hybrid kernel function is used instead of the original single kernel function to effectively improve model performance.

[0138] Meanwhile, Table 2 shows various types of fault information used for diagnosis in the present invention:

[0139] Table 2 Fault types

[0140]

[0141]

[0142] refer to Figure 11-14To verify the effectiveness of the improved algorithm and optimized model in this paper, the diagnostic accuracy of four models, (a) ISSA-MKELM, (b) SSA-MKELM, (c) SSA-KELM, and (d) SSA-SVM, was tested. The ISSA-MKELM method, using the method described in this paper, achieved 100% classification accuracy for fault identification, demonstrating excellent results for transformer fault diagnosis and classification. Comparison shows that the kernel extreme learning machine achieves higher classification accuracy than the support vector machine network, and the hybrid kernel function and improved SSA algorithm significantly improve the classification accuracy of the ELM network model.

[0143] This invention utilizes a deep feature extraction network based on SSDAE-Softmaxs, built using deep learning technology, to extract features, resolving the inability of traditional feature extraction methods to capture deep-level features of vibration data. Addressing the shortcomings of the kernel extreme learning machine (KELM) classifier, a hybrid kernel extreme learning machine (MKELM) with enhanced generalization capabilities is constructed by combining Gaussian and polynomial kernel functions. Furthermore, to address the issues of the sparrow search algorithm with random population initialization and its tendency to fall into local extrema, a Bernoulli chaos map is introduced to optimize the initial sparrow population. Furthermore, a Cauchy-Gaussian mutation strategy and nonlinear dynamic inertia weights are used to improve the sparrow search algorithm's tendency to fall into local optima. The improved sparrow search algorithm (ISSA) is then used to optimize the parameters of the hybrid kernel extreme learning machine (MKELM) and establish an ISSA-MKELM-based transformer early fault diagnosis model. Compared to existing technologies, this invention offers the advantages of superior diagnostic accuracy and robustness.

Claims

1. A transformer early fault diagnosis method based on ISSA-MKELM, characterized in that: The method comprises the following steps: Data on transformer operation under different conditions are collected. Early vibration signals of different transformer faults are preprocessed to obtain corresponding envelope spectrum signals. The corresponding envelope spectrum signals are normalized and randomly divided into training data and test data. Extract the early fault characteristics of different categories from the input data, the number of hidden layer nodes of the network and the learning rate of each layer of DA, and optimize them through the ISSA algorithm; A hybrid kernel extreme learning machine (MKELM) was constructed, and the Bernoulli chaotic map was introduced to optimize the initial sparrow population. The Cauchy-Gaussian mutation strategy and nonlinear dynamic inertia weight were used, and the improved sparrow search algorithm (ISSA) was adopted to optimize the parameters of the hybrid kernel extreme learning machine (MKELM). An ISSA-MKELM fault diagnosis model was established. The early fault diagnosis model ISSA-MKELM is tested using test data, and the diagnostic accuracy of the model for the test samples is calculated.

2. The transformer early fault diagnosis method based on ISSA-MKELM according to claim 1 is characterized in that: The transformer is divided into nine states of operating data, the training data includes SSDAE pre-training and global fine-tuning SSDAE, and the training data is used to train a deep feature extraction neural network SSDAE.

3. The transformer early fault diagnosis method based on ISSA-MKELM according to claim 1 is characterized in that: Preprocess the vibration signals of different transformer faults and obtain the corresponding envelope spectrum signals, including: The original vibration signal x(t) is decomposed using CEEMD to obtain several IMFs and residual terms; Select the IMF with higher correlation and calculate the Hilbert spectrum; The corresponding envelope value is obtained based on the Hilbert spectrum; The obtained envelope value is subjected to Fourier transform to obtain the corresponding envelope spectrum.

4. The transformer early fault diagnosis method based on ISSA-MKELM according to claim 1 is characterized in that: The hybrid kernel extreme learning machine MKELM is constructed by a Gaussian kernel function and a polynomial kernel function.

5. The transformer early fault diagnosis method based on ISSA-MKELM according to claim 2 is characterized in that: Unsupervised Stacked Sparse Denoising Autoencoder (SSDAE) pre-training, also includes: The mean square error function is selected as the reconstruction error function of SSDAE, and a batch processing strategy is used to process large-scale training sample sets; After obtaining the error function, the batch stochastic gradient descent method is used to find the minimum value of the error function and the partial derivatives of each parameter with respect to the error function are obtained; The network activation function selects the sigmoid function, and its first-order derivative and second-order derivative function are obtained; By expressing SSDAE through weights and bias terms, we can get the parameter pairs zi k The partial derivative of Use the chain rule to find the partial derivatives of each parameter with respect to the loss function; The parameter set is updated according to the chain rule and the gradient descent method, and the parameter set is updated according to the update rule until the error convergence criterion is met. After the training is completed, the reconstruction layer is removed, and the output result of the last hidden layer is the result of feature extraction.

6. The transformer early fault diagnosis method based on ISSA-MKELM according to claim 5, characterized in that: Supervised global fine-tuning SSDAE, also includes: After the SSDAE pre-training is completed, the SOFTMAX classifier is added to the last hidden layer of SSDAE to form a complete network; Input labeled data and use the backpropagation algorithm to perform global training on the entire network from top to bottom. The partial derivative of the classifier loss function will propagate downward layer by layer, making slight adjustments to the network weights originally initialized by SSDAE. The partial derivative of the classifier weight parameter with respect to the loss function is used to fine-tune the parameters according to the gradient descent rule: According to the back-propagation algorithm, the relevant weights of all layers of the network are adjusted from top to bottom until the convergence criteria are met.

7. The transformer early fault diagnosis method based on ISSA-MKELM according to claim 1 is characterized in that: The ISSA algorithm is used to optimize the hybrid kernel extreme learning machine classification model, which also includes: Use hybrid kernel function to realize the mapping from input space to high-dimensional hidden layer feature space; In the population initialization stage, Bernoulli chaotic mapping is introduced to generate chaotic sequences; After obtaining the chaotic sequence that traverses the entire two-dimensional space, the chaotic particles are mapped to the search range required for environmental modeling; Dynamic inertia weights are introduced to optimize the position update method of sparrow population discoverers, reconcile the global optimization and local mining capabilities of the equilibrium algorithm, and in the early stage of algorithm evolution, the main focus is on exploring the search space to reach a better area and find the optimal solution.

8. The transformer early fault diagnosis method based on ISSA-MKELM according to claim 7, characterized in that: The mathematical expression of Bernoulli map is: Where: Z k is the current value of the kth generation chaotic sequence, and λ is the control parameter.

9. The transformer early fault diagnosis method based on ISSA-MKELM according to claim 7, characterized in that: After obtaining the chaotic sequence that traverses the entire two-dimensional space, the chaotic particles are mapped to the search range required for environmental modeling, that is: Where, X min and X max The minimum and maximum values ​​of the horizontal axis are set to 10 and 200 respectively; min and Y max Set the minimum and maximum values ​​of the vertical axis to -10 and 10 respectively; is the position of the kth generation sparrow.

10. The transformer early fault diagnosis method based on ISSA-MKELM according to claim 7, characterized in that: The position of the discoverer with the dynamic inertia weight w is updated as follows: Where t represents the current iteration number, j = 1, 2, ....., d, where d represents the dimension of the solution, iter max is the maximum number of iterations, X i,j represents the position of the i-th sparrow in the j-th dimension, α∈(0,1] is a random number, R2(R2∈[0,1]) and ST(ST∈[0.5,1]) represent the warning value and safety value respectively, Q is a random number that obeys the normal distribution, L is a 1×d matrix of all ones; w start is the initial inertia weight; w end is the inertia weight at the maximum number of iterations.

Images