CN-121682453-B - DOA-BiNET-based mechanical fault diagnosis method for on-load tap-changer

CN121682453BCN 121682453 BCN121682453 BCN 121682453BCN-121682453-B

Abstract

A DOA-BiNET-based on-load tap-changer mechanical fault diagnosis method relates to the technical field of on-load tap-changer fault diagnosis and is used for solving the problems that parameters of an existing deep learning model are difficult to optimize and the on-load tap-changer fault identification precision is low. The method comprises the following steps of S1 introducing a nonlinear time sequence residual enhancement mechanism into a bidirectional long-short-term memory network BiLSTM to construct a bidirectional time sequence network BiNET, S2 improving a bristle dog optimization algorithm DOA through Logistic-Tent chaotic mapping, carrying out self-adaptive optimization on the number of hidden layer neurons, learning rate and L2 regularization coefficient of a BiNET network, S3 constructing a fault diagnosis model based on DOA-BiNET, inputting time-frequency domain characteristics of a vibration signal of an on-load tap-changer into the fault diagnosis model, and outputting a fault diagnosis result.

Inventors

ZHAO TONG
CHEN ZHIXIN
WANG XIAOLONG
ZHANG YUANTAO
SUN YING
LIU YADI
QI RUNZE
DUAN TIANYU

Assignees

山东大学

Dates

Publication Date: 20260505
Application Date: 20260210

Claims (5)

1. The on-load tap-changer mechanical fault diagnosis method based on DOA-BiNET is characterized by comprising the following steps of: S1, introducing a nonlinear time sequence residual error enhancement mechanism into a two-way long-short-term memory network BiLSTM, performing feature fusion and discrimination enhancement on a two-way hidden state, and constructing a two-way time sequence network BiNET; S2, improving a bristle dog optimization algorithm DOA through Logistic-Tent chaotic mapping, and performing self-adaptive optimization on the number of hidden layer neurons, the learning rate and the L2 regularization coefficient of the BiNET network by utilizing the improved bristle dog optimization algorithm DOA; S3, constructing a mechanical fault diagnosis model of the on-load tap-changer based on DOA-BiNET, inputting the time-frequency domain characteristics of the vibration signal of the on-load tap-changer into the fault diagnosis model, and outputting a fault diagnosis result; in the step S1, the nonlinear time sequence residual enhancement mechanism is introduced specifically comprising S1.1, performing differential operation on the comprehensive hidden states at adjacent moments by using a formula (1), constructing time sequence residual characteristics by a nonlinear activation function, (1) In the formula (I), in the formula (II), Representing the time sequence residual error characteristics after nonlinear mapping; is a linear rectification activation function; h t is the comprehensive hiding state of the BiLSTM network at the moment t; S1.2, splicing and fusing the original bidirectional hidden characteristic and the residual characteristic by using a formula (2) to form an enhanced time sequence characteristic representation, (2) In the formula (I), in the formula (II), The method is an enhanced hidden state after the steady state characteristic and the transient residual error characteristic are fused; Vector splicing operation is carried out; In step S2, the Logistic-Tent mapping expression is (3) In the formula (I), in the formula (II), The result of the n+1st iteration of the chaotic mapping; In order to control the parameters of the device, ; Mapping for chaos As a result of the number of iterations, ; When generating an initialization population of the bristle dog optimization algorithm DOA based on Logistic-Tent mapping, the following is assumed The number of the hunting object is given to the hunting object, N is population size; As a dimension of the problem variable, Mapping the chaotic sequence generated by the Logistic-Tent mapping into a function space to obtain initial population of the mane as (4) In the formula (I), in the formula (II), Is the first Individual mane dog's first Values of dimensions; alpha epsilon [0,1] is a smoothing coefficient; Sequences obtained based on Logistic-Tent mapping; Is the upper limit of the search value; In the step S2, the algorithm process of the improved bristle dog optimization algorithm DOA comprises the steps of S2.1, initializing algorithm parameters and population scale, determining population scale N and maximum iteration times T, S2.2, initializing a search space through Logistic-Tent mapping to generate N individuals with uniform distribution, S2.3, calculating fitness values of the individuals of the population, wherein the optimal solution is the position of a male bristle dog, the suboptimal solution is the position of a female bristle dog, S2.4, calculating escape energy E of a game, (5) In the formula (I), in the formula (II), Representing an initial state of prey energy, E 0 =2ε -1 (6); Representing the decreasing energy of the prey, (7) Wherein ε represents any number between 0 and 1; A constant value equal to 1.5; Represents the maximum iteration number, gamma is a nonlinear attenuation coefficient, and when |E| is more than or equal to 1, the method adopts (8)、 (9) And (10) Updating the position of the next iteration of the game, wherein f () is an fitness function, and Y 1 (t) and Y 2 (t) are coordinates of the game after being updated under the influence of male and female dogs; is the position of male mane; the iteration times are the current stage; a position vector representing a prey; is the location of female mane; representing the distance between the male bristle dog and the prey during the scout stage of exploration; representing the distance between the female bristle dog and the prey during the scout stage of exploration; For random number vectors based on the distribution of the rice, (11) Η (t) is an adaptive step size scaling factor, Η 0 is the initial step size coefficient; For the fitness function value of the lewy flight, (12), (13) In the formula (I), in the formula (II), And To generate the random variable for the lewy flight step, And The value ranges of the two are all (0, 1) intervals; takes a default constant of 1.5, Γ represents a gamma function, σ represents a scale parameter of the flight profile, and |E| <1 is used (14)、 (15) And (10) Updating the next iteration position of the hunting object, wherein, Representing the distance between the male bristle dog and the prey surrounding the attack phase; S2.5, judging whether the algorithm meets a stopping condition, if so, exiting the loop and returning to an optimal position, and if not, returning to the step S2.3 for continuous execution; In the step S3, the on-load tap-changer mechanical fault diagnosis model based on DOA-BiNET comprises an input layer, a BiNET layer, a full-connection layer, a Softmax layer and an output layer, wherein the input layer is used for receiving a feature vector extracted by an OLTC vibration signal, the BiNET layer is responsible for extracting deep features from input OLTC feature data, the full-connection layer further performs feature fusion and dimension reduction, the Softmax layer is used as a classifier for performing fault category prediction on the feature vector output by the full-connection layer, and the output layer finally gives out an OLTC fault diagnosis result.
2. The method for diagnosing mechanical failure of an on-load tap changer based on DOA-BiNET as claimed in claim 1, wherein in step S1, multi-class cross entropy is selected The difference of the predicted result of the network model from the actual result is evaluated BiLSTM as a loss function, (16) Wherein, C is the total number of categories; The first to be the true label Values of the individual categories; predicted first for BiLSTM network model Probability of individual categories.
3. The method for diagnosing mechanical failure of an on-load tap changer based on DOA-BiNET as claimed in claim 2, wherein in step S1, L2 regularization is introduced by using formula (17) to reduce the complexity of BiLSTM network model, (17) In the formula (I), in the formula (II), Is a regularization parameter; The weight is the square of L2 norm of the weight, w i is the i-th trainable weight parameter in BiLSTM network, and the weight parameter updating formula of BiLSTM network is as follows: (18); (19) Wherein V f is the weight parameter matrix of the input layer in the BiLSTM network after updating, W f is the weight parameter matrix of the hidden layer in the BiLSTM network after updating, U f is the weight parameter matrix of the gating structure in the BiLSTM network after updating, b f is the bias parameters of each layer in the BiLSTM network after updating, V b is the weight parameter matrix of the input layer in the BiLSTM network before updating, W b is the weight parameter matrix of the hidden layer in the BiLSTM network before updating, U b is the weight parameter matrix of the gating structure in the BiLSTM network before updating, and b b is the bias parameters of each layer in the BiLSTM network before updating; Is the learning rate.
4. The mechanical fault diagnosis method for on-load tap-changer based on DOA-BiNET as claimed in claim 1, wherein in step S2, the fault diagnosis accuracy f is adopted as the fitness function in the optimizing process, (20) In the formula (I), in the formula (II), To classify the correct number of samples; Is the total number of samples.
5. The method for diagnosing mechanical failure of an on-load tap changer based on DOA-BiNET as claimed in claim 1, wherein in step S3, the vibration signal is preprocessed by using improved adaptive noise complete set empirical mode decomposition ICEEMDAN, and kurtosis, fuzzy entropy, dispersion entropy and energy entropy are extracted from each intrinsic mode function to construct model input feature vectors.

Description

DOA-BiNET-based mechanical fault diagnosis method for on-load tap-changer Technical Field The invention relates to the technical field of on-load tap-changer fault monitoring, in particular to a DOA-BiNET-based on-load tap-changer mechanical fault diagnosis method. Background At present, the mechanical fault diagnosis of the on-load tap-changer OLTC is mostly dependent on an artificial intelligence method to perform pattern recognition on vibration signal characteristics, wherein a deep learning model has become a mainstream development direction due to strong characteristic expression capability and adaptability. Compared with the traditional machine learning method, the deep learning has higher stability and accuracy when processing complex, nonlinear and multidimensional signals. The optimized BP neural network and the convolutional neural network CNN are widely used for diagnosis in the prior art, but the BP neural network has a simple structure and is easily influenced by abnormal data to generate overfitting, and the CNN has local feature extraction capability but is difficult to capture the time dependence of vibration signals. Disclosure of Invention The invention aims to provide a DOA-BiNET-based on-load tap-changer mechanical fault diagnosis method, which is used for solving the problems that the existing on-load tap-changer vibration signal characteristics are not sufficiently utilized, deep learning model parameters are difficult to optimize and fault identification accuracy is not high, and realizing high-accuracy, interpretable and stable identification of an OLTC mechanical fault. The technical scheme adopted by the invention for solving the technical problems is that the mechanical fault diagnosis method of the on-load tap-changer based on DOA-BiNET comprises the following steps. S1, introducing a nonlinear time sequence residual error enhancement mechanism into a two-way long-short-term memory network BiLSTM, and carrying out feature fusion and discrimination enhancement on the two-way hidden state to construct a two-way time sequence network BiNET. S2, improving a bristle dog optimization algorithm DOA through Logistic-Tent chaotic mapping, and performing self-adaptive optimization on the number of hidden layer neurons, the learning rate and the L2 regularization coefficient of the BiNET network by utilizing the improved bristle dog optimization algorithm DOA. S3, constructing a mechanical fault diagnosis model of the on-load tap-changer based on DOA-BiNET, inputting the time-frequency domain characteristics of the vibration signal of the on-load tap-changer into the fault diagnosis model, and outputting a fault diagnosis result. Further, in step S1, multi-class cross entropy is selectedThe difference of the predicted result of the network model from the actual result is evaluated BiLSTM as a loss function,(1) Wherein, C is the total number of categories; The first to be the true label Values of the individual categories; predicted first for BiLSTM network model Probability of individual categories. Further, in step S1, introducing L2 regularization using equation (2) reduces BiLSTM the complexity of the network model,(2) In the formula (I), in the formula (II),Is a regularization parameter; The weight is the square of L2 norm of the weight, w i is the i-th trainable weight parameter in BiLSTM network, and the weight parameter updating formula of BiLSTM network is as follows: (3);(4) Wherein V f is the weight parameter matrix of the input layer in the BiLSTM network after updating, W f is the weight parameter matrix of the hidden layer in the BiLSTM network after updating, U f is the weight parameter matrix of the gating structure in the BiLSTM network after updating, b f is the bias parameters of each layer in the BiLSTM network after updating, V b is the weight parameter matrix of the input layer in the BiLSTM network before updating, W b is the weight parameter matrix of the hidden layer in the BiLSTM network before updating, U b is the weight parameter matrix of the gating structure in the BiLSTM network before updating, and b b is the bias parameters of each layer in the BiLSTM network before updating; Is the learning rate. In step S1, the nonlinear time sequence residual enhancement mechanism is introduced specifically including S1.1, performing differential operation on the comprehensive hidden states of adjacent moments by using a formula (5), constructing time sequence residual characteristics by nonlinear activation functions,(5) In the formula (I), in the formula (II),Representing the time sequence residual error characteristics after nonlinear mapping; is a linear rectification activation function; H t is the comprehensive hiding state of the BiLSTM network at the moment t; S1.2, splicing and fusing the original bidirectional hidden characteristic and the residual characteristic by using a formula (6) to form an enhanced time sequence characteristic representation, (6) In th