A Method for Mechanical Fault Diagnosis of On-Load Tap Changers Based on DOA-BiNET

Abstract

A fault diagnosis method for on-load tap changers based on DOA-BiNET is disclosed, relating to the field of on-load tap changer fault diagnosis technology. This method addresses the problem of low accuracy in fault identification of on-load tap changers due to the difficulty in optimizing parameters of existing deep learning models. The method includes the following steps: S1, a nonlinear temporal residual enhancement mechanism is introduced into the bidirectional long short-term memory network BiLSTM to construct a bidirectional temporal network BiNET; S2, the DOA optimization algorithm is improved through Logistic-Tent chaotic mapping, adaptively optimizing the number of hidden layer neurons, learning rate, and L2 regularization coefficient of the BiNET network; S3, a fault diagnosis model based on DOA-BiNET is constructed, inputting the time-frequency domain features of the on-load tap changer vibration signal into the fault diagnosis model, and outputting the fault diagnosis result.

Description

Technical Field

[0001] This invention relates to the field of on-load tap changer fault monitoring technology, specifically a method for diagnosing mechanical faults in on-load tap changers based on DOA-BiNET. Background Technology

[0002] Currently, mechanical fault diagnosis of on-load tap changers (OLTCs) largely relies on artificial intelligence methods for pattern recognition of vibration signal characteristics. Among these, deep learning models have become the mainstream development direction due to their powerful feature representation capabilities and adaptability. Compared to traditional machine learning methods, deep learning exhibits higher stability and accuracy when processing complex, nonlinear, and multidimensional signals. Existing research has widely used optimized backpropagation (BP) neural networks and convolutional neural networks (CNNs) for diagnosis. However, BP neural networks have a simple structure and are easily affected by abnormal data, leading to overfitting. While CNNs possess local feature extraction capabilities, they struggle to capture the temporal dependencies of vibration signals. Summary of the Invention

[0003] The purpose of this invention is to provide a mechanical fault diagnosis method for on-load tap changers based on DOA-BiNET, which solves the problems of insufficient utilization of vibration signal features of existing on-load tap changers, difficulty in optimizing deep learning model parameters, and low fault identification accuracy, and achieves high-precision, interpretable, and stable identification of mechanical faults in OLTCs.

[0004] The technical solution adopted by the present invention to solve its technical problem is: a method for diagnosing mechanical faults of on-load tap changers based on DOA-BiNET, comprising the following steps.

[0005] S1. A nonlinear temporal residual enhancement mechanism is introduced into the bidirectional long short-term memory network BiLSTM to perform feature fusion and discrimination enhancement on the bidirectional hidden states, and to construct the bidirectional temporal network BiNET.

[0006] S2. Improve the DOA (Device Override) algorithm by using Logistic-Tent chaotic mapping. Adaptively optimize the number of hidden layer neurons, learning rate and L2 regularization coefficient of BiNET network using the improved DOA algorithm.

[0007] S3. Construct a mechanical fault diagnosis model for on-load tap changers based on DOA-BiNET, input the time-frequency domain characteristics of the vibration signal of the on-load tap changer into the fault diagnosis model, and output the fault diagnosis results.

[0008] Furthermore, in step S1, the multi-class cross-entropy is selected. As a loss function, it evaluates the difference between the prediction results of the BiLSTM network model and the actual results. (1); where C is the total number of categories; The first of the real labels Values ​​for each category; The first prediction for the BiLSTM network model The probability of each category.

[0009] Furthermore, in step S1, L2 regularization is introduced using formula (2) to reduce the complexity of the BiLSTM network model. (2); where, For regularization parameters; w is the square of the L2 norm of the weights. i Let be the i-th trainable weight parameter in the BiLSTM network; the formula for updating the weight parameters of the BiLSTM network is: (3); (4); where V f W represents the weight parameter matrix of the input layer in the updated BiLSTM network. f This is the updated weight parameter matrix of the hidden layer in the BiLSTM network; U f b is the weight parameter matrix of the gated structure in the updated BiLSTM network; f To update the bias parameters of each layer in the BiLSTM network; V b W represents the weight parameter matrix of the input layer in the BiLSTM network before updating. b To update the weight parameter matrix of the hidden layer in the BiLSTM network; U b b is the weight parameter matrix of the gating structure in the BiLSTM network before updating; b To update the bias parameters of each layer in the BiLSTM network before updating; This is the learning rate.

[0010] Furthermore, in step S1, the introduction of a nonlinear temporal residual enhancement mechanism specifically includes: S1.1, using formula (5) to perform differential operations on the integrated hidden states at adjacent time points, and constructing temporal residual features through a nonlinear activation function. (5); where, This represents the characteristics of the time-series residuals after nonlinear mapping; It is a linear rectification activation function; h is a trainable parameter vector used to adaptively adjust the enhancement strength of the difference features; t Let be the integrated hidden state of the BiLSTM network at time t; S1.2 represents the comprehensive hidden state of the BiLSTM network at time t-1; S1.2 uses formula (6) to concatenate and fuse the original bidirectional hidden features and residual features to form the enhanced temporal feature representation. (6); where, This refers to the enhanced hidden state after fusing steady-state features and transient residual features; This is a vector concatenation operation.

[0011] Further, in step S2, the Logistic-Tent mapping expression is: (7); where, This is the result of the (n+1)th iteration of the chaotic mapping; For control parameters, ; For chaotic mapping The result of the iteration ; For the modulo operation; when generating the initial population for the DOA optimization algorithm based on the Logistic-Tent mapping, it is assumed that... Number the prey , where n is the population size; For the problem variable dimension, Mapping the chaotic sequence generated by the Logistic-Tent mapping onto the function space yields the initial population of the hyenas. (8). In the formula, For the first The first individual hyena The value of the dimension; The lower bound of the search value; α∈[0,1] is the smoothing coefficient; The sequence is obtained based on the Logistic-Tent mapping; This represents the upper limit of the search values.

[0012] Further, in step S2, the improved DOA (Domain of Alternating Orbit) optimization algorithm is as follows: S2.1, Initialize algorithm parameters and population size, determining the population size N and the maximum number of iterations T; S2.2, Initialize the search space through Logistic-Tent mapping, generating N individuals with relatively uniform distribution; S2.3, Calculate the fitness value of each individual in the population, the optimal solution is the position of the male hyena, and the second-best solution is the position of the female hyena; S2.4, Calculate the prey escape energy E. (9); where, The initial state of the prey's energy is represented by E0 = 2ε-1 (10). Indicates the decreasing energy of the prey. (11); where ε represents any number between 0 and 1; It is a constant value equal to 1.5; Indicates the maximum number of iterations; γ is the nonlinear decay coefficient, γ>1; when |E|≥1, it adopts... (12) (13) and (14) Update the position of the prey in the next iteration; where f() is the fitness function; Y1(t) and Y2(t) are the coordinates of the prey after being updated by the influence of male and female hyenas; This indicates the location of the male hyena. This represents the current iteration count. Represents the position vector of the prey; This indicates the location of the female hyena. This indicates the distance between the male hyena and its prey during the exploration and reconnaissance phase; This indicates the distance between the female hyena and her prey during the exploration and reconnaissance phase. A random number vector based on the Lévy distribution. (15); η(t) is the adaptive step size scaling factor. η0 is the initial step size coefficient; Let Lévy's flight fitness function value be denoted as . (16) (17); where, and For the random variable used to generate the Lévy flight step size, and The values ​​of are all within the interval (0, 1); Γ is the default constant, with a value of 1.5; Γ represents the gamma function; σ is the scale parameter of the flight distribution; when |E| < 1, the default value is used. (18) (19) and (14) Update the prey's position for the next iteration; where, This indicates the distance between the male hyena and its prey during the encirclement and attack phase. This represents the distance between the female hyena and its prey during the encirclement and attack phase; S2.5, determine whether the algorithm meets the stopping condition. If the condition is met, exit the loop and return to the optimal position; if not, return to step S2.3 to continue execution.

[0013] Furthermore, in step S2, the fault diagnosis accuracy f is used as the fitness function in the optimization process. (20); where, The number of correctly classified samples; This represents the total number of samples.

[0014] Furthermore, in step S3, the on-load tap changer mechanical fault diagnosis model based on DOA-BiNET includes an input layer, a BiNET layer, a fully connected layer, a Softmax layer, and an output layer. The input layer is used to receive the feature vector extracted from the OLTC vibration signal. The BiNET layer is responsible for extracting deep features from the input OLTC feature data. The fully connected layer further performs feature fusion and dimensionality reduction. The Softmax layer acts as a classifier to predict the fault category of the feature vector output by the fully connected layer. The output layer finally gives the fault diagnosis result of the OLTC.

[0015] Furthermore, in step S3, the vibration signal is preprocessed using the improved adaptive noise complete set empirical mode decomposition ICEEMDAN, and kurtosis, fuzzy entropy, scatter entropy and energy entropy are extracted from each intrinsic mode function to construct the model input feature vector.

[0016] The beneficial effects of this invention are: by improving the DOA (Dog Optimization Algorithm) to optimize the hyperparameters of the BiNET model, this invention overcomes the hyperparameter optimization problem of traditional BiLSTM, avoids the instability caused by manual parameter tuning, accelerates model convergence and improves training efficiency, effectively enhances the fault mode recognition capability, and significantly improves the accuracy and robustness of OLTC mechanical fault diagnosis. It can be effectively used for OLTC mechanical condition monitoring and fault diagnosis. Attached Figure Description

[0017] Figure 1 Flowchart for DOA optimization process;

[0018] Figure 2 DOA-BiNET fault diagnosis flowchart;

[0019] Figure 3 This is a structural diagram of the Bidirectional Long Short-Term Memory (BiLSTM) network.

[0020] Figure 4 This is a structural diagram of the on-load tap changer fault diagnosis model based on DOA-BiNET of the present invention. Detailed Implementation

[0021] This invention uses the time-frequency domain characteristics of the vibration signal from an on-load tap changer (OLTC) as model input to construct a deep fault diagnosis method based on DOA-BiNET. For example... Figure 2As shown, this invention discloses a method for diagnosing mechanical faults in on-load tap changers based on DOA-BiNET, comprising the following steps: S1. To address the shortcomings of traditional BiLSTM in feature discrimination capability and parameter sensitivity, a nonlinear temporal residual enhancement mechanism is introduced into the bidirectional long short-term memory network BiLSTM to obtain the bidirectional temporal network BiNET. By performing feature fusion and discrimination enhancement on the bidirectional hidden states, key temporal features are more effectively highlighted and utilized in classification decisions. S2. To avoid the problems of network parameters relying on manual experience settings and model performance being easily affected by hyperparameters, a modified DOA algorithm based on Logistic-Tent chaotic mapping is proposed. The modified DOA algorithm is used to adaptively optimize the key hyperparameters of the BiNET network. The modified DOA algorithm improves the quality of the initial population through Logistic-Tent chaotic mapping and introduces a fitness-based dynamic weight mechanism for male and female hyenas to enhance search capability and avoid getting trapped in local optima, thereby achieving adaptive optimization of the number of hidden layer neurons, learning rate, and L2 regularization coefficient. S3. Construct a mechanical fault diagnosis model for on-load tap changers based on DOA-BiNET. Input the time-frequency domain characteristics of the on-load tap changer vibration signal into the fault diagnosis model and output the fault diagnosis results. Experimental results show that the model outperforms traditional methods in terms of convergence speed, diagnostic accuracy, and robustness, and can be effectively applied to OLTC mechanical condition monitoring and fault diagnosis under complex working conditions.

[0022] The following section introduces the Bidirectional Long Short-Term Memory (BiLSTM) network.

[0023] like Figure 3 As shown, the Bidirectional Long Short-Term Memory (BiLSTM) network is an extension of the Long Short-Term Memory (LSTM) network. By running LSTM networks in two opposite directions, BiLSTM simultaneously considers both past and future information of the sequence, thus more comprehensively grasping the dependencies within the sequence and improving the model's feature extraction capabilities when processing nonlinear data, thereby enhancing diagnostic accuracy. BiLSTM consists of two independent LSTM networks: a forward LSTM that processes the input in the sequence order, and a backward LSTM that processes the input in the reverse order. The computation process of BiLSTM is as follows: (1). In the formula, The hidden layer state of the feedforward LSTM network at time t; The hidden layer state of the feedback LSTM network at time t; h t The combined hidden state at time t is obtained by concatenating the hidden states of the forward LSTM network and the hidden states of the backward LSTM network. This is the final output; The hidden layer state of the feedforward LSTM network at time t-1; The hidden layer state of the feedback LSTM network at time t-1; g is the activation function of the output layer; This is the output layer weight matrix; For output layer bias; This is a vector concatenation operation; x t For input data.

[0024] This invention selects multi-class cross-entropy As a loss function, it is used to evaluate the difference between the model's predictions and the actual results. (2). In the formula, C is the total number of categories; The first of the real labels Values ​​for each category; The model predicts the first The probability of each category. To avoid overfitting during training, L2 regularization is introduced to reduce the complexity of the model, as shown in formula (3). (3). In the formula, This is a regularization parameter that controls the strength of the regularization term. w is the square of the L2 norm of the weights. i Let be the i-th trainable weight parameter in the network. The formula for updating the network's weight parameters is: (4); (5). In the formula, V f W represents the weight parameter matrix of the input layer in the updated BiLSTM network. f This is the updated weight parameter matrix of the hidden layer in the BiLSTM network; U f b is the weight parameter matrix of the gated structure in the updated BiLSTM network; f To update the bias parameters of each layer in the BiLSTM network; V b W represents the weight parameter matrix of the input layer in the BiLSTM network before updating. b To update the weight parameter matrix of the hidden layer in the BiLSTM network; U b b is the weight parameter matrix of the gating structure in the BiLSTM network before updating; b To update the bias parameters of each layer in the BiLSTM network before updating; The learning rate controls the step size for updating parameters.

[0025] The following describes the bidirectional temporal network BiNET of this invention.

[0026] Based on the BiLSTM network, and to address the problem that steady-state noise in vibration signals can easily interfere with the modeling of transient impact characteristics, an improved bidirectional long short-term memory network, namely the bidirectional temporal network BiNET, is proposed.

[0027] BiNET, while retaining the bidirectional temporal modeling capability of BiLSTM, introduces a nonlinear temporal residual enhancement mechanism to highlight the abrupt changes in vibration signals and suppress the influence of steady-state noise on feature representation. Specifically, it first performs differential operations on the combined hidden states at adjacent time points and constructs temporal residual features through a nonlinear activation function, as shown in Equation (6): (6). In the formula, The characteristics of the time-series residuals after nonlinear mapping; It is a linear rectification activation function; These are trainable parameter vectors used to adaptively adjust the enhancement strength of the difference features; This represents the synthesized hidden state of the BiLSTM at time t-1. Through this design, the residual component is activated only when the features change significantly between adjacent time steps, thus effectively suppressing the interference of steady-state noise. Subsequently, the original bidirectional hidden features and the residual features are concatenated and fused to form the enhanced temporal feature representation, as shown in formula (7): (7). In the formula, To enhance the hidden state by fusing steady-state and transient residual features, and to serve as input for subsequent classification layers. Through these improvements, without disrupting the original time-series modeling structure of BiLSTM, BiNET introduces the ability to explicitly model transient change features, enabling the network to simultaneously consider both the overall evolution trend and local abrupt changes of the dynamic signal, thereby improving the robustness and discriminative performance of fault diagnosis.

[0028] The following section introduces the DOA (Optimized Algorithm) of the Manedog algorithm.

[0029] The Dog Optimization Algorithm (DOA) searches for the optimal solution in the solution space by simulating the behavior of a dog. The implementation process of the Dog Optimization Algorithm (DOA) includes: (1) Initializing the search space; In the initial stage of the algorithm, the initial solution Y0 is uniformly distributed in the search space. (8). In the formula, The minimum value of the solution to be found; A random number in the range [0, 1]; The maximum value of the solution to be found is denoted as 'Prey'. The initialization process generates an initial prey matrix 'Prey', where the individuals with the highest and second-highest fitness form a pair of hyenas. (9). In the formula, For the first The first prey Dimensional parameter values, i=1,2…n, j=1,2…d; Population size; This represents the problem variable dimension. The prey's position corresponds to specific parameters in the problem solution. During optimization, a fitness function is used to evaluate the fitness value of each prey, filling the prey fitness matrix with the fitness values ​​of all prey. . (10). In the formula, The fitness function of the algorithm is denoted as . The male hyena is the individual with the highest fitness, followed by the female hyena. The hyena pair will occupy the corresponding prey's position. (2) Exploration and reconnaissance phase: The male hyena dominates the hunting process, and the female hyena follows its actions. The prey's updated coordinates are influenced by the male and female hyenas: (11); (12). In the formula, This indicates the location of the male hyena. This represents the current iteration count. Represents the position vector of the prey; This indicates the location of the female hyena. This indicates the distance between the male hyena and its prey during the exploration and reconnaissance phase; This indicates the distance between the female hyena and her prey during the exploration and reconnaissance phase. A vector of random numbers based on the Lévy distribution; Energy for the prey to escape. (13). In the formula, The decreasing energy for the prey; The initial energy state of the prey. E0 = 2ε-1 (14). (15). In the formula, ε is any number between 0 and 1; It is a constant value equal to 1.5; γ is the maximum number of iterations; γ is the nonlinear decay coefficient, which improves the transition speed of the algorithm from exploration to utilization, γ>1; (16). In the formula, η(t) is the adaptive step size scaling factor, η0 is the initial step size coefficient; LF(y) is the fitness function value of Lévy flight. (17). (18). In the formula, and For the random variable used to generate the Lévy flight step size, and The values ​​of are all within the interval (0, 1); Γ is a default constant with a value of 1.5; Γ represents the gamma function; σ is the scale parameter of the flight distribution. The prey position is updated by the mean of the results calculated by equations (11) and (12), and the calculation formula is as follows: (19). (3) Encirclement and Attack Phase; After the prey is continuously harassed by the hyenas, its escape energy gradually decreases. At this time, the hyenas will surround the prey detected in the previous phase. After completing the encirclement, they will launch a surprise attack and prey on the prey. The prey's location coordinates at this time are: (20); (21); where, This indicates the distance between the male hyena and its prey during the encirclement and attack phase. This indicates the distance between the female hyena and her prey during the encirclement and attack phase.

[0030] The prey's energy of escape Used to determine when to transition from the exploration and reconnaissance phase to the encirclement and attack phase. During the escape, the prey's energy will significantly decrease, from its initial energy... In each iteration, it randomly fluctuates between -1 and 1. When the value decreases from 0 to -1, it indicates that the prey's physical strength is declining; while when As the value increases from 0 to 1, it reflects an increase in prey strength. The escape energy E exhibits a dynamic decay trend during the iteration process. When |E|≥1, the hyena, in the exploration and reconnaissance phase, searches different areas to find prey tracks; when... At this point, the hyena enters the encirclement and attack phase, during which it will approach and attack its prey.

[0031] The following section introduces the improved Logistic-Tent mapping algorithm DOA, which is a dog optimization algorithm based on the present invention.

[0032] This invention improves the DOA (Dog Optimization Algorithm) from two aspects: initialization of the search space and the prey position update mechanism during the dog optimization process. It proposes a DOA improved by using Logistic-Tent mapping.

[0033] This invention introduces the Logistic-Tent mapping, which integrates the advantages of Logistic and Tent mappings, to generate highly diverse initial chaotic sequences, thus overcoming the problems of uneven population distribution and large differences caused by random initialization. The Logistic-Tent mapping expression is as follows: (22). In the formula, This is the result of the (n+1)th iteration of the chaotic mapping; For control parameters, ; This is the modulo operation; For chaotic mapping The result of the iteration When generating the initial population for the DOA (Dog Optimization Algorithm) based on Logistic-Tent mapping, it is assumed that... Number the prey ; For the problem variable dimension, Mapping the chaotic sequence generated by the Logistic-Tent mapping onto the function space yields the initial population of the hyenas. (23). In the formula, For the first The first individual hyena The value of the dimension; α represents the lower bound of the search value; α∈[0,1] is the smoothing coefficient to prevent the initial individuals from concentrating at the boundary or center. The sequence is obtained based on the Logistic-Tent mapping; This represents the upper limit of the search values.

[0034] As shown in equation (19), the prey position update is jointly decided by the male and female hyenas, and their decision weights are the same. In practice, the male hyena has a greater influence on the prey position update, so the average decision method has poor convergence. This invention assigns different weights to male and female hyenas based on their corresponding fitness values, and the calculation formula is as follows: (24), where f() is the fitness function. In the final calculation process, the prey position is updated by replacing formula (19) with formula (24).

[0035] Based on the above improvement strategies, the DOA optimization process of the improved bristle dog optimization algorithm proposed in this invention is as follows: Figure 1 As shown. The algorithm process of the improved hyena optimization algorithm DOA is as follows: S2.1, initialize the algorithm parameters and population size, and determine the population size N and the maximum number of iterations T. S2.2, initialize the search space through Logistic-Tent mapping to generate N individuals with relatively uniform distribution. S2.3, calculate the fitness value of each individual in the population. The optimal solution is the position of the male hyena, and the suboptimal solution is the position of the female hyena. S2.4, calculate the escape energy E of the prey based on equation (13). When |E|≥1, use equations (11), (12) and (24) to update the position of the prey in the next iteration; when |E|<1, use equations (20), (21) and (24) to update the position of the prey in the next iteration. S2.5, determine whether the algorithm meets the stopping condition. If the condition is met, exit the loop and return to the optimal position; if not, return to step S2.3 to continue execution.

[0036] The DOA-BiNET diagnostic model of this invention is described below.

[0037] like Figure 4 As shown, the OLTC fault diagnosis model based on BiNET in this invention includes an input layer, a BiNET layer, a fully connected layer, a Softmax layer, and an output layer.

[0038] The input layer is used to receive the feature vector extracted from the OLTC vibration signal. This invention uses an improved adaptive noise complete set empirical mode decomposition ICEEMDAN to preprocess the vibration signal and extracts kurtosis, fuzzy entropy, scatter entropy, and energy entropy from each intrinsic mode function to construct the model input feature vector.

[0039] The BiNET layer is the core of the diagnostic model, responsible for extracting deep features from the input OLTC feature data. The BiNET layer automatically learns and extracts key patterns in the data, enhancing feature representation capabilities to improve the accuracy of fault diagnosis.

[0040] The fully connected layer further performs feature fusion and dimensionality reduction, and enhances the separability of features through linear transformation, thereby improving the discriminative ability of the classifier.

[0041] The Softmax layer acts as a classifier, predicting the fault category from the feature vector output by the fully connected layer.

[0042] The output layer ultimately provides the fault diagnosis results for OLTC.

[0043] The hyperparameter configuration of the BiNET-based fault diagnosis model has a decisive impact on its construction and accuracy. Specifically, the number of hidden layer neurons in BiNET determines the model's complexity and learning ability; too large a number can lead to overfitting, while too small a number fails to capture effective features from complex data. The learning rate controls the step size of parameter updates; an excessively high learning rate can prevent convergence, while an excessively low rate results in slow convergence. L2 regularization limits the model's complexity by adding the sum of squared model parameters to the loss function. The L2 regularization coefficient determines the influence of the regularization term; an excessively high L2 regularization coefficient leads to insufficient model fitting ability, while an excessively low coefficient fails to effectively suppress overfitting. The residual enhancement coefficient... Used to adjust the model's capture strength of transient change features in time-series signals, reasonable The value can effectively amplify impact features while suppressing environmental noise interference, preventing the model from failing due to feature overload or noise amplification. Therefore, this invention optimizes the number of hidden layer neurons, learning rate, regularization parameter, and residual enhancement coefficient γ of BiNET based on DOA. The ultimate goal of this invention in optimizing BiNET hyperparameters is to improve the model's fault diagnosis capability; therefore, the fault diagnosis accuracy f is used as the fitness function in the optimization process. (25). In the formula, The number of correctly classified samples; This represents the total number of samples.

[0044] This invention optimizes the hyperparameters of the BiNET model using an improved DOA (Dog Optimization Algorithm), overcoming the challenges of hyperparameter optimization in traditional BiLSTM models. It avoids the instability caused by manual parameter tuning, accelerates model convergence, improves training efficiency, effectively enhances fault mode recognition capabilities, and significantly improves the accuracy and robustness of OLTC (Optical Continuous Traction Control) mechanical fault diagnosis. It features faster convergence speed, higher recognition accuracy, and stronger robustness, making it effective for OLTC mechanical condition monitoring and fault diagnosis.

Claims

1. A method for diagnosing mechanical faults in on-load tap changers based on DOA-BiNET, characterized in that, Includes the following steps: S1. Introduce a nonlinear temporal residual enhancement mechanism into the bidirectional long short-term memory network BiLSTM to perform feature fusion and discrimination enhancement on the bidirectional hidden states, and construct the bidirectional temporal network BiNET. S2. Improve the DOA (Device Override) algorithm by using Logistic-Tent chaotic mapping. Adaptively optimize the number of hidden layer neurons, learning rate and L2 regularization coefficient of BiNET network using the improved DOA algorithm. S3. Construct a mechanical fault diagnosis model for on-load tap changers based on DOA-BiNET, input the time-frequency domain characteristics of the vibration signal of the on-load tap changer into the fault diagnosis model, and output the fault diagnosis results. In step S1, the introduction of a nonlinear temporal residual enhancement mechanism specifically includes: S1.1, using formula (1) to perform differential operations on the integrated hidden states at adjacent time points, and constructing temporal residual features through a nonlinear activation function. (1); where, This represents the characteristics of the time-series residuals after nonlinear mapping; It is a linear rectification activation function; h is a trainable parameter vector. t Let be the integrated hidden state of the BiLSTM network at time t; S1.2 represents the comprehensive hidden state of the BiLSTM network at time t-1; S1.2 uses formula (2) to concatenate and fuse the original bidirectional hidden features and residual features to form the enhanced temporal feature representation. (2); where, This refers to the enhanced hidden state after fusing steady-state features and transient residual features; This is a vector concatenation operation; In step S2, the Logistic-Tent mapping expression is: (3); where, This is the result of the (n+1)th iteration of the chaotic mapping; For control parameters, ; For chaotic mapping The result of the iteration ; For the modulo operation; when generating the initial population for the DOA optimization algorithm based on the Logistic-Tent mapping, it is assumed that... Number the prey , where n is the population size; For the problem variable dimension, Mapping the chaotic sequence generated by the Logistic-Tent mapping onto the function space yields the initial population of the hyenas. (4); where, For the first The first individual hyena The value of the dimension; The lower bound of the search value; α∈[0,1] is the smoothing coefficient; The sequence is obtained based on the Logistic-Tent mapping; This is the upper limit of the search value; In step S2, the improved DOA (Domain of Alternating Orientation) algorithm for hyenas is as follows: S2.1, Initialize algorithm parameters and population size, determining the population size N and the maximum number of iterations T; S2.2, Initialize the search space using Logistic-Tent mapping, generating N individuals with a relatively uniform distribution; S2.3, Calculate the fitness value of each individual in the population, with the optimal solution being the position of the male hyena and the second-best solution being the position of the female hyena; S2.4, Calculate the escape energy E of the prey. (5); where, The initial state of the prey's energy is represented by E0 = 2ε-1 (6). Indicates the decreasing energy of the prey. (7); where ε represents any number between 0 and 1; It is a constant value equal to 1.5; Indicates the maximum number of iterations; γ is the nonlinear decay coefficient; when |E|≥1, the following is used: (8) (9) and (10) Update the position of the prey in the next iteration; where f() is the fitness function; Y1(t) and Y2(t) are the coordinates of the prey after being updated by the influence of male and female hyenas; This indicates the location of the male hyena. This represents the current iteration count. Represents the position vector of the prey; This indicates the location of the female hyena. This indicates the distance between the male hyena and its prey during the exploration and reconnaissance phase; This indicates the distance between the female hyena and her prey during the exploration and reconnaissance phase. A random number vector based on the Lévy distribution. (11); η(t) is the adaptive step size scaling factor, η0 is the initial step size coefficient; Let Lévy's flight fitness function value be denoted as . (12), (13); where, and For the random variable used to generate the Lévy flight step size, and The values ​​of are all within the interval (0, 1); Γ is the default constant, with a value of 1.5; Γ represents the gamma function; σ is the scale parameter of the flight distribution; when |E| < 1, the default value is used. (14) (15) and (10) Update the prey's position for the next iteration; where, This indicates the distance between the male hyena and its prey during the encirclement and attack phase. This represents the distance between the female hyena and its prey during the encirclement and attack phase; S2.5, determine if the algorithm meets the stopping condition. If the condition is met, exit the loop and return to the optimal position; if not, return to step S2.3 to continue execution; In step S3, the on-load tap changer mechanical fault diagnosis model based on DOA-BiNET includes an input layer, a BiNET layer, a fully connected layer, a Softmax layer, and an output layer. The input layer is used to receive the feature vector extracted from the OLTC vibration signal. The BiNET layer is responsible for extracting deep features from the input OLTC feature data. The fully connected layer further performs feature fusion and dimensionality reduction. The Softmax layer acts as a classifier to predict the fault category of the feature vector output by the fully connected layer. The output layer finally gives the fault diagnosis result of the OLTC.

2. The method for diagnosing mechanical faults in on-load tap changers based on DOA-BiNET according to claim 1, characterized in that, In step S1, select the multi-class cross-entropy. As a loss function, it evaluates the difference between the prediction results of the BiLSTM network model and the actual results. (16); where C is the total number of categories; The first of the real labels Values ​​for each category; The first prediction for the BiLSTM network model The probability of each category.

3. The method for diagnosing mechanical faults in on-load tap changers based on DOA-BiNET according to claim 2, characterized in that, In step S1, L2 regularization is introduced using formula (17) to reduce the complexity of the BiLSTM network model. (17); where, For regularization parameters; w is the square of the L2 norm of the weights. i Let be the i-th trainable weight parameter in the BiLSTM network; the formula for updating the weight parameters of the BiLSTM network is: (18); (19); where V f W represents the weight parameter matrix of the input layer in the updated BiLSTM network. f This is the updated weight parameter matrix of the hidden layer in the BiLSTM network; U f b is the weight parameter matrix of the gated structure in the updated BiLSTM network; f To update the bias parameters of each layer in the BiLSTM network; V b W represents the weight parameter matrix of the input layer in the BiLSTM network before updating. b To update the weight parameter matrix of the hidden layer in the BiLSTM network; U b b is the weight parameter matrix of the gating structure in the BiLSTM network before updating; b To update the bias parameters of each layer in the BiLSTM network before updating; This is the learning rate.

4. The method for diagnosing mechanical faults in on-load tap changers based on DOA-BiNET according to claim 1, characterized in that, In step S2, the fault diagnosis accuracy f is used as the fitness function in the optimization process. (20); where, The number of correctly classified samples; This represents the total number of samples.

5. The method for diagnosing mechanical faults in on-load tap changers based on DOA-BiNET according to claim 1, characterized in that, In step S3, the vibration signal is preprocessed using the improved adaptive noise complete set empirical mode decomposition ICEEMDAN, and kurtosis, fuzzy entropy, scatter entropy and energy entropy are extracted from each intrinsic mode function to construct the model input feature vector.

Images