On-load tap-changer mechanical fault diagnosis method based on SABO optimization hybrid network

Abstract

The invention discloses an on-load tap-changer mechanical fault diagnosis method based on an SABO optimized hybrid network, relates to the technical field of on-load tap-changer fault diagnosis, and is used for solving the problems that a traditional fault diagnosis method generally depends on expert experience and preset rules, and the applicability and accuracy of the traditional fault diagnosis method are limited when the traditional fault diagnosis method faces complex multi-mode signals and nonlinear fault modes. Comprising the steps of signal acquisition, signal division, search optimization on a bandwidth threshold value and a B spline order of time-varying filtering empirical mode decomposition by using an SABO algorithm, search agent updating, signal decomposition, TCN-GUR model training and fault diagnosis output. According to the method, the accuracy is superior to that of a traditional method, higher robustness is shown, and particularly higher accuracy and stability are achieved in multi-fault-mode recognition.

Description

Technical Field

[0001] The present invention relates to the technical field of on-load tap-changer fault diagnosis, and specifically to a mechanical fault diagnosis method for on-load tap-changers based on an SABO-optimized hybrid network. Background Technique

[0002] In recent years, vibration signal analysis technology has been widely studied and applied in the field of mechanical equipment fault diagnosis. Since during the operation of an on-load tap-changer, mechanical movements and electrical switching operations of its internal components will inevitably generate vibrations, and these vibration signals contain rich information closely related to the equipment operating state. The existing on-load tap-changer fault diagnosis methods based on vibration signals have significant advantages of being non-intrusive, capable of online monitoring, and sensitive to early faults, providing a new and effective technical approach for on-load tap-changer state monitoring and fault diagnosis.

[0003] Empirical Mode Decomposition (EMD) and its derivative methods show certain potential in processing non-stationary and non-linear vibration signals. For example, Ensemble Empirical Mode Decomposition (EEMD) can overcome the mode mixing problem to a certain extent, but there are still deficiencies in residual noise.

[0004] Variational Mode Decomposition (VMD), as a new signal decomposition method, shows good decomposition accuracy and anti-noise performance when processing complex vibration signals. By constructing and solving a variational problem, the original signal is decomposed into modal components with specific bandwidths, which can effectively extract signal feature information.

[0005] However, for signals with strong background noise, multi-modal coupling, and complex time-varying characteristics, these decomposition methods cannot fully exploit their inherent subtle fault features and have limitations in extracting some weak fault features and distinguishing complex fault patterns. Traditional fault diagnosis methods usually rely on expert experience and preset rules, but their applicability and accuracy are often limited when facing complex multi-modal signals and non-linear fault patterns. Summary of the Invention

[0006] The purpose of the present invention is to provide a mechanical fault diagnosis method for on-load tap-changers based on an SABO-optimized hybrid network, which is used to solve the problem that traditional fault diagnosis methods usually rely on expert experience and preset rules, and their applicability and accuracy are limited when facing complex multi-modal signals and non-linear fault patterns.

[0007] The technical solution adopted by the present invention to solve its technical problems is: A mechanical fault diagnosis method for on-load tap-changers based on an SABO-optimized hybrid network, including the following steps:

[0008] S1: Simulate typical mechanical faults of on-load tap changers, and collect the generated vibration signals and drive motor current signals.

[0009] S2: Divide the vibration signal and drive motor current signal samples obtained in step S1 into a training set and a validation set to prepare for subsequent model training and testing.

[0010] S3: Use the SABO algorithm to search and optimize the bandwidth threshold and B-spline order of the time-varying filtering empirical mode decomposition (TVFEMD), backhaul the parameter combinations, decompose the signal using TVFEMD, and evaluate the quality of each set of parameters with the minimum information entropy as the fitness function.

[0011] S4: Update the search agent, continuously adjust the bandwidth threshold and B-spline order until the stopping criterion is met.

[0012] S5: Adopt the optimal parameter combination obtained by optimization, apply TVFEMD to finally decompose the vibration signal, and generate IMF components.

[0013] S6: Input the decomposed IMF components and the drive motor current signal into the TCN-GUR classification model for training and classification.

[0014] S7: Output the fault diagnosis result.

[0015] Furthermore, the specific steps for the SABO algorithm to search and optimize the bandwidth threshold and B-spline order include:

[0016] S3.1: First, define the operator as: (1). A and B are different individuals, sign is the sign function; F(A) is the fitness value of A; F(B) is the fitness value of B; is a vector with the same dimension as the search space, a random number generated within [1, 2]; represents the Hadamard product of two vectors.

[0017] S3.2: Randomly initialize the population in the optimization space: (2). Where: is an individual; is the upper boundary of optimization; is the lower boundary of optimization, a random number between [0, 1]; r is a random number vector uniformly distributed within the interval [0, 1].

[0018] S3.3. Calculate each search agent New location : (3). Where, is a vector with the same dimension as the search space, consisting of random numbers between [0,1]; is the total number of individuals; X i is the current position vector of the i-th search agent; X j is the current position vector of the j-th search agent; v is the random perturbation vector; is the latest calculated position of the i-th individual. If the new position is better, it replaces the previous position, otherwise it keeps the original position. (4). Search Agent The objective function value of Search Agent The objective function value of .

[0019] Furthermore, TVFEMD includes two stages: S3.4 local cutoff frequency rearrangement and S3.5 screening based on time-varying filtering; the steps of local cutoff frequency rearrangement are:

[0020] S3.4.1. For real-valued signals Perform Hilbert transform to obtain analytical signal , (5). Where A(t) is the amplitude; a1(t) is the real component of the analytical signal; a2(t) is the imaginary component of the analytical signal; j is the imaginary unit; for Hilbert transform of ; φ(t), φ1(t), φ2(t) are phases, and (6).

[0021] S3.4.2, when When , A(t) has a local maximum; when When , A(t) has a local minimum. For all the maximum points of A(t) Interpolate to get the curve , for all minimum points of A(t) Interpolate to get the curve ,get (7).

[0022] S3.4.3, for A 2 All maximum points of (t) Interpolation is performed to obtain , for A 2 All minimum points of (t) Interpolation is performed to obtain , and order (8). Among them, is 's derivative, and φ2 ' (t) is 's derivative. The local cut-off frequency is: (9).

[0023] S3.4.4. Define the time series of the maximum points of the signal as , (10). (u i : u i+1 ) is the time interval between two consecutive local extreme points. Take ρ = 0.25. If the above equation holds, then u i is called a discontinuity point. Let , represents the sequence of discontinuity points. If then is called 's rising edge, and is regarded as the lowest value; if then is called 's falling edge, and is regarded as the lowest value. The remaining part is regarded as the peak value.

[0024] S3.4.5. Interpolate between all peak values to obtain the rearranged local cut-off frequency .

[0025] Further, the steps of the screening stage based on time-varying filtering include:

[0026] S3.5.1. Reconstruct the signal h(t) according to the rearranged cut-off frequency: (11). Take the extreme points of as nodes, divide into n segments, with each segment having a step size of m. Among them, n is called the order of the B-spline function. Perform B-spline interpolation approximation between the extreme points, and the approximation result is denoted as , which is the local mean curve.

[0027] S3.5.2. Judge whether the residual signal satisfies the stopping criterion , where is the given bandwidth threshold. If it is satisfied, then is the IMF; otherwise, let , and repeat all the above steps until the stopping criterion is satisfied. Among them, (12); Loughlin instantaneous bandwidth B loughlinThe calculation formula of (t) is: (13); the weighted average instantaneous frequency φ avg The calculation formula of (t) is: (14). a1 ′2 (t) and a2 ′2 (t) are the square terms of the signal change rate.

[0028] Furthermore, the TCN-GRU model uses TCN to efficiently extract the temporal features of multimodal signals to mine the local and global patterns in the time dimension. Subsequently, the extracted temporal features are input into the gated recurrent unit GRU network to further enhance the modeling ability of dynamic features, and finally achieve the accurate classification and recognition of different fault types.

[0029] The beneficial effects of the present invention are as follows: The present invention simulates three typical mechanical faults through experiments: the gear jamming of the transmission shaft, the loosening of the transmission shaft screws, and the loosening of the arc plate, and reveals the vibration feature differences under different fault conditions through vibration signal analysis. The SABO algorithm is used to optimize the bandwidth threshold and B-spline order of TVFEMD, and then effectively extract the intrinsic mode components of the signal. By comparing different classification models, the SABO-TVFEMD-TCN-GRU model shows significant advantages in the fault diagnosis task. Compared with other common classification models, the present invention not only surpasses the traditional methods in terms of accuracy but also demonstrates stronger robustness, especially with higher accuracy and stability in the recognition of multi-fault modes. In summary, the SABO-TVFEMD-TCN-GRU model provides an innovative and efficient solution for the mechanical fault diagnosis of on-load tap-changers, which can not only improve the accuracy of fault diagnosis but also adapt to the signal changes in complex environments, and has important application value for the online monitoring and fault warning of smart grid equipment. Description of the Drawings

[0030] Figure 1 is the flowchart of the fault diagnosis method of the present invention;

[0031] Figure 2 is the convergence curve of the SABO algorithm, particle swarm optimization algorithm, and grey wolf optimization algorithm on the F1 test function;

[0032] Figure 3 is the convergence curve of the SABO algorithm, particle swarm optimization algorithm, and grey wolf optimization algorithm on the F2 test function;

[0033] Figure 4 is the convergence curve of the SABO algorithm, particle swarm optimization algorithm, and grey wolf optimization algorithm on the F9 test function;

[0034] Figure 5Convergence curves of the SABO algorithm, particle swarm optimization algorithm, and grey wolf optimization algorithm on the F10 test function;

[0035] Figure 6 Structural diagram of the TCN model;

[0036] Figure 7 Structural diagram of the residual connection block;

[0037] Figure 8 Structural diagram of the dilated causal convolution;

[0038] Figure 9 Structural diagram of the GRU cell;

[0039] Figure 10 Combined model diagram of TCN-GRU;

[0040] Figure 11 Time-domain waveform diagram of the vibration signal decomposed by SABO-TVFEMD under normal operating conditions;

[0041] Figure 12 Spectrum diagram of the vibration signal decomposed by SABO-TVFEMD under normal operating conditions;

[0042] Figure 13 Time-domain waveforms and spectrum diagrams of IMFs 1–IMF5 of the drive shaft gear jamming fault condition signal;

[0043] Figure 14 Time-domain waveforms and spectrum diagrams of IMFs 1–IMF5 of the drive shaft screw loosening fault condition signal;

[0044] Figure 15 Time-domain waveforms and spectrum diagrams of IMFs 1–IMF5 of the contact and arc plate loosening fault condition signal;

[0045] Figure 16 Loss curve diagram of the TCN-GRU model on the training set and test set;

[0046] Figure 17 Accuracy curve diagram of the TCN-GRU model on the training set and test set;

[0047] Figure 18 Classification accuracy diagram of each model in 10 experiments under different decomposition methods;

[0048] Figure 19 Confusion matrix diagram of the classification results of a certain experiment of the model;

[0049] Figure 20 Classification accuracy diagram of four classification models in 10 experiments;

[0050] Figure 21It is the confusion matrix diagram of the classification results of a certain experiment of each model. Specific implementation manner

[0051] Aiming at the key challenges in the fault diagnosis of on-load tap-changer (OLTC) vibration signals, a combined optimization method based on SABO-TVFEMD and TCN-GRU is proposed by integrating vibration signal processing technology and machine learning algorithms. The temporal convolutional neural network (TCN) has gradually become a popular method for time series analysis due to its ability to effectively capture the long-range dependencies of sequential signals and its parallel computing advantages. As a simplified recurrent neural network, the gated recurrent unit (GRU) has low computational complexity and good performance, and performs excellently in processing dynamic time series data. By introducing the SABO algorithm to adaptively optimize the parameter settings of TVFEMD, the advantages of retaining time-varying features and reducing mode mixing are utilized to accurately decompose the vibration signals. At the same time, after extracting features by combining TCN and feeding them into GRU for fault classification and recognition, the diagnostic performance is further improved. This method fully considers the non-stationarity and complexity of vibration signals, makes up for the deficiencies of traditional diagnostic methods in dealing with complex fault modes, and provides a more accurate, reliable and intelligent fault diagnosis solution for the reliable operation of on-load tap-changers in power systems.

[0052] The mechanical fault diagnosis method of on-load tap-changer based on SABO-optimized hybrid network of the present invention includes the following steps:

[0053] S1: Simulate typical mechanical faults of on-load tap-changers through experiments, and collect the vibration signals and drive motor current signals generated during the experiments.

[0054] S2: Divide the vibration signal and drive motor current signal samples into a training set and a validation set to prepare for subsequent model training and testing.

[0055] S3: Use the SABO algorithm to search and optimize the key parameters bandwidth threshold and B-spline order of the time-varying filtering empirical mode decomposition (TVFEMD), bring back the parameter combinations and use TVFEMD to decompose the signals, and use the minimum information entropy as the fitness function to evaluate the advantages and disadvantages of each group of parameters.

[0056] S4: Update the search agent, and continuously adjust the bandwidth threshold and B-spline order until the stopping criterion is met.

[0057] S5: Adopt the best parameter combination obtained by optimization, apply TVFEMD to finally decompose the vibration signal, and generate IMF components.

[0058] S6: Input the decomposed IMF components and the drive motor current signal into the TCN-GUR classification model for training and classification.

[0059] S7: Output the fault diagnosis result.

[0060] The working principles of each step of the present invention are described below:

[0061] I. Signal decomposition principle based on SABO-TVFEMD.

[0062] 1. Time-varying filtering empirical mode decomposition TVFEMD.

[0063] Essentially, TVFEMD completes the iterative mean removal operation in the EMD decomposition process by constructing a low-pass filter with a cut-off frequency varying with time, and uses a local narrow-band signal instead of the intrinsic mode function as the iteration stop condition to solve the intermittency and mode mixing problems that occur when processing non-linear and non-stationary signals. TVFEMD is mainly divided into two stages: local cut-off frequency rearrangement and screening based on time-varying filtering.

[0064] 1.1 Local cut-off frequency rearrangement, used to eliminate mode mixing. The steps of this stage are:

[0065] S1: For the real-valued signal Obtain the analytical signal through Hilbert transform, and represent it as a two-component signal in a complex form: (1). Where, A(t) is the amplitude; a1(t) is the real part component of the analytical signal, a2(t) is the imaginary part component of the analytical signal; j is the imaginary unit; is the Hilbert transform of. φ(t), φ1(t), φ2(t) are phases, and (2).

[0066] S2: When , A(t) has a local maximum; when , A(t) has a local minimum. Interpolate all the maximum points of A(t) to obtain the curve , interpolate all the minimum points of A(t) to obtain the curve , and obtain (3).

[0067] S3: Interpolate all the maximum points 2 of A (t) to obtain , interpolate all the minimum points 2 of A Interpolate to obtain , and let (4). Among them, is 's derivative, and φ2 ' (t) is 's derivative. Define the local cut-off frequency as: (5).

[0068] S4. Define the time series of the maximum points of the signal as , (6). (u i : u i+1 ) is the time interval between two consecutive local extreme points. In the present invention, ρ = 0.25 is taken. If the above formula holds, then u i is called a discontinuity point. Let , represents the sequence of discontinuity points. If then it is called is 's rising edge, and is regarded as the lowest value; if then it is called is 's falling edge, and is regarded as the lowest value. The remaining part is regarded as the peak value.

[0069] S5. Interpolate between all peak values to obtain the rearranged local cut-off frequency .

[0070] 1.2 Screening stage based on time-varying filtering. The steps of this stage include:

[0071] S1. Reconstruct the signal h(t) according to the rearranged cut-off frequency: (7). Take the extreme points of as nodes, and divide into n segments, with each segment having a step size of m. Among them, n is called the order of the B-spline function. Perform B-spline interpolation approximation between the extreme points, and the approximation result is denoted as , which is the local mean curve.

[0072] S2. Judge whether the residual signal satisfies the stopping criterion , where is the given bandwidth threshold. If it is satisfied, then is the IMF; otherwise, let , and repeat all the above steps until the stopping criterion is satisfied. Among them, (8); Loughlin instantaneous bandwidth Bloughlin The calculation formula of (t) is: (9); weighted average instantaneous frequency φ avg The calculation formula of (t) is: (10). a1 ′2 (t) and a2 ′2 (t) is the square term of the signal change rate.

[0073] 2. Optimization algorithm based on subtraction average

[0074] During the execution of TVFEMD, it is necessary to manually specify two parameters, the bandwidth threshold and the B-spline order There is a large degree of blindness and uncertainty. To obtain the optimal parameters, the present invention selects the optimization algorithm SABO based on subtraction average to optimize the two parameters. SABO is a meta-heuristic optimization algorithm. Its basic idea is to update the position of each individual by subtracting the average of all individuals in the search space, so as to continuously iterate to find the optimal solution. The specific steps are as follows:

[0075] S1. First, define the operator as: (11). A and B are different individuals, and sign is the sign function. F(A) is the fitness value of A; F(B) is the fitness value of B; is a vector with the same dimension as the search space, and is a random number generated within [1,2]; represents the Hadamard product of two vectors. Randomly initialize the population in the optimization space: (12). In the formula: is an individual; is the upper boundary of optimization; is the lower boundary of optimization, and is a random number between [0,1]; r is a random number vector uniformly distributed in the interval [0,1]. In SABO, the displacement of any search agent in the search space is calculated by the arithmetic mean of each search agent 's The new position of each search agent is calculated as follows: In the formula, is a vector with the same dimension as the search space, composed of random numbers between [0,1]; is the total number of individuals; X i is the current position vector of the i-th search agent; X j is the current position vector of the j-th search agent; v is a random perturbation vector; is the latest calculated position of the i-th individual. If the new position is better, it replaces the previous position, otherwise it keeps the original position. That is, (14). Search Agent The objective function value of Search Agent The objective function value of .

[0076] To verify the SABO algorithm's optimization performance, we selected the CEC2005 optimization benchmark function set for testing. Unimodal functions F1 and F2 were selected to test the algorithm's optimization accuracy and speed, while multimodal functions F9 and F10 were used to test the algorithm's global optimization capability and whether it could achieve an optimal solution. The expressions and search spaces for the test functions are shown in Table 1.

[0077]

[0078] The test functions in Table 1 are used to test and compare the SABO algorithm, the particle swarm optimization algorithm PSO and the grey wolf optimization algorithm GWO. The experiment is set to 1000 iterations and a population size of 100. Figures 2 - 5 The three-dimensional graphs of the four test functions and the convergence curves of each algorithm are displayed. The results in the graphs show that the SABO algorithm is the first to reach the optimal value on each test function. This not only demonstrates its significant advantages in global optimization and convergence speed, but also reflects its strong ability to escape from local optimal solutions.

[0079] 2. Signal fusion and fault diagnosis method based on TCN-GRU model

[0080] 1. Time Domain Convolutional Neural Network TCN

[0081] To fully capture the intrinsic mode function (IMF) features of the decomposed vibration signal and the deep-level information in the drive motor current signal, a deep learning method, the TCN-GRU model, was introduced that combines time series modeling and feature extraction capabilities. This model first uses TCN to efficiently extract time series features from multimodal signals to explore their local and global patterns in the time dimension. The extracted time series features are then input into a gated recurrent unit (GRU) network to further enhance the modeling capabilities of dynamic features and ultimately achieve accurate classification and identification of different fault types. This method not only improves diagnostic accuracy but also enhances the model's robust extraction of fault features under complex operating conditions, effectively promoting the improvement of OLTC intelligent fault diagnosis performance.

[0082] TCN is a deep learning architecture designed specifically for sequence data modeling. Due to its ability to efficiently capture the long-range dependencies of time series signals and its parallel computing advantages, it has gradually become a popular method in the field of time series analysis. Through dilated convolution, long-term dependency modeling, and residual connection techniques, TCN can better capture the characteristics of time series data, has strong capabilities in modeling long-term dependencies, high efficiency in processing variable-length sequences, and translational invariance advantages, and avoids the problems of gradient vanishing and gradient explosion. The specific structure of the TCN model is as shown in Figures 6 - 8 shown.

[0083] The network structure of TCN includes residual connection blocks and dilated causal convolutions. (1) Residual connection blocks: To alleviate the problems of gradient vanishing and gradient explosion in deep networks, TCN adopts residual connections, which improve the stability and training efficiency of the network while maintaining the model complexity. As shown in Figure 6 and Figure 7 shown, the TCN module is mainly composed of multiple residual connection blocks, and each residual connection block includes a dilated causal convolution layer, a normalization layer, an activation function layer, and a regularization layer. (2) Dilated causal convolution is the core operation of TCN. By combining the time dependence of causal convolution with the long-range modeling ability of dilated convolution, TCN can efficiently handle the long-range dependencies in time series data. Its structure is as shown in Figure 8 shown. Different from the recurrent neural networks LSTM and GRU that process sequence data step by step, dilated causal convolution can achieve parallel computing through convolution operations, thereby improving the computing efficiency. It not only ensures the causality in time series modeling but also expands the receptive field, captures richer time series features with a shallower network, and supports efficient parallel computing.

[0084] 2. Gated Recurrent Unit GRU

[0085] GRU is a variant of the recurrent neural network RNN and can more effectively model sequence data. Compared with the long short-term memory network, GRU integrates the input gate and forget gate in LSTM into an update gate, which plays a crucial role in controlling the degree of information retention. And GRU abandons the output gate in LSTM and instead introduces a reset gate, whose function is to flexibly adjust the way of fusing the current state and previous information. Figure 9 shows the structure of GRU, and its mathematical description is as shown in Equation (15). Among them, is the input at time ; is the output or state at time ; is the output or state at time ; are the weights; is the activation function sigmoid; is the activation function; W T （x，z） is the input weight matrix of the update gate; W T （h，z） is the hidden state weight matrix of the update gate; W T （x，r） is the input weight matrix of the reset gate; W T （h，r） is the hidden state weight matrix of the reset gate; W T （x，g） is the input weight matrix of the candidate state; W T （h，g） is the hidden state weight matrix of the candidate state; g t is the candidate hidden state.

[0086] Figure 10 Figure shows the TCN-GRU combined model structure. The TCN-GRU method combines the structural advantages of TCN and GRU, and has strong temporal feature modeling and expression capabilities. TCN adopts a one-dimensional causal convolution and residual connection structure, which can effectively capture long-term dependence information, avoid the problem of gradient disappearance, and at the same time has the advantage of parallel computing, with high training efficiency. Compared with the traditional RNN structure, TCN is more stable in the overall modeling of the input sequence. As a lightweight recurrent neural network, GRU dynamically controls the update and forgetting of information through a gating mechanism, further enhancing the model's sensitivity and memory ability to key temporal features. The combination of the two not only improves the accuracy and timeliness of feature extraction, but also enhances the robustness and generalization ability of the model under complex and variable working conditions, and is suitable for fault diagnosis tasks driven by multi-modal signals.

[0087] III. Fault Diagnosis Model Based on SABO-TVFEMD-TCN-GRU

[0088] 3.1 Overall Architecture of the Model

[0089] Based on the advantages of the SABO algorithm, the present invention optimizes TVFEMD with the subtraction average-based optimization algorithm SABO to search for the optimal bandwidth threshold and B-spline order . Select the minimum information entropy as the fitness function, as shown in the following formula (16). In the formula, represents the th component's normalized energy.

[0090] Use SABO-TVFEMD to decompose the signal into IMF components with different time-frequency characteristics, and input these components into the TCN-GRU model for fault classification and diagnosis.

[0091] 3.2 Decomposition Results of Vibration Signals Based on SABO-TVFEMD

[0092] According to the aforementioned method principles and steps, the search range of SABO and the optimal TVFEMD parameter results for signals under each working condition are set as shown in Table 2.

[0093]

[0094] Figure 11 The results of decomposing the vibration signal under normal operating conditions by the SABO-TVFEMD method are shown, clearly separating multiple intrinsic mode functions IMF. Figure 12 Furthermore, the spectrograms of each IMF component are given, intuitively presenting their frequency distribution characteristics. For the convenience of comparative analysis, Table 3 further gives Figure 12 the energy proportion of each IMF in it and its corresponding peak frequency. Through analysis, it can be seen that IMF1 to IMF5 together account for 98.88% of the total signal energy, and no obvious mode mixing phenomenon is found between each mode, fully verifying the superiority of the SABO-TVFEMD method in signal separation and feature retention. This result shows that the main effective vibration information is concentrated in IMF1 to IMF5, while the remaining high-order IMFs are mainly composed of noise components and have low analysis value.

[0095]

[0096] Figures 13 - 15 The time-domain waveforms and spectrograms of IMF1 to IMF5 obtained by decomposing the vibration signal by SABO-TVFEMD are shown respectively under three fault conditions of drive shaft gear jamming, drive shaft screw loosening, and contact and arc plate loosening. The results show that IMF1 to IMF5 show obvious fluctuation characteristics in the time domain, and their spectrograms also clearly depict the main frequency ranges where each component is located, and no frequency mixing phenomenon occurs, further verifying the high quality of the decomposition results. By comparing the decomposition results under different fault conditions, significant differences can be observed in both time-domain and frequency-domain characteristics.

[0097] 3.3 Fault Diagnosis Based on TCN-GRU Model

[0098] In the above-mentioned on-load tap-changer typical mechanical fault simulation experiment, 200 sets of vibration data were collected for each operating condition. The data was divided into a training set and a test set according to a ratio of 8:2. Specifically, the training set contains 160 sets of signals for each condition, totaling 640 sets; the test set contains 40 sets of signals for each condition, totaling 160 sets. All vibration signal samples were preprocessed by the SABO-TVFEMD method, and IMF1 to IMF5 obtained by decomposition were extracted. The IMF1-IMF5 obtained by decomposing the signals of each condition and the driving motor current signal were used as the input features of the TCN-GRU model. To improve the model performance, the present invention optimized some key parameters in the TCN-GRU model using the grid search algorithm. The final parameter settings are shown in Table 4. Among them, / indicates that the parameter was not optimized and a fixed value was used.

[0099]

[0100] During the model training process, the loss value Loss and the accuracy Accuracy are the key indicators for evaluating the model performance. Figures 16 - 17 Shows the change curves of the loss and accuracy of the TCN-GRU model on the training set and the test set. Figure 16 Shows the change process of the training loss and the test loss. In the initial stage of training, the model's fitting ability to the data was weak, and both the training loss and the test loss were at a relatively high level. The training loss was 1.77, and the test loss was 1.75. As the training progressed, the two loss curves decreased significantly, indicating that the model's fitting effect on the data was continuously improved. In the later stage of training, the training loss dropped to 0.05, and the test loss dropped to 0.11, and the curves tended to be stable, indicating that the model achieved a good fitting effect on both the training set and the test set. Figure 17 Shows the change of the training accuracy and the test accuracy. In the initial stage of training, the training accuracy was 28.3%, and the test accuracy was 24.9%. As the number of training rounds increased, the training accuracy increased rapidly and reached 99.3% in the later stage of training, indicating that the model's recognition ability for the training data was significantly enhanced. At the same time, the test accuracy also showed a steady upward trend and finally stabilized at 96.3%, fully demonstrating the excellent generalization ability of the model on the test set.

[0101] 3.4. Ablation Experiment and Comparative Experiment

[0102] To verify the effectiveness of the SABO-TVFEMD decomposition method adopted in the present invention, models with different decomposition methods were trained using the same training set, and their performances were evaluated on the test set. To exclude the contingency of the experimental results, each model was independently trained 10 times without using pre-trained weights. Table 5 shows the average classification accuracy of each model in 10 experiments. Figure 18Shows the classification accuracy of each model in 10 experiments. Figure 19 Shows the confusion matrix of the classification results of each model in a certain experiment in Table 5. Fault 1 is the jamming of the drive shaft gear, Fault 2 is the loosening of the drive shaft screw, and Fault 3 is the loosening of the contact and the arc plate, which are three fault conditions.

[0103]

[0104] Table 5 shows the average classification accuracy after combining four different decomposition methods with the TCN-GRU model. The average accuracy of the SABO-TVFEMD-TCN-GRU model is 96.38%, which performs the best among all models. The average accuracy of the TVFEMD-TCN-GRU model is 91.62%. The average accuracy of the VMD-TCN-GRU model is 88.12%, while the average accuracy of the EMD-TCN-GRU model is 84.62%. Obviously, the performance of the SABO-TVFEMD-TCN-GRU model is better than that of other models. This indicates that compared with the TVFEMD, VMD, and EMD decomposition methods, the SABO-TVFEMD decomposition method significantly improves the classification accuracy of fault diagnosis.

[0105] Subsequently, to verify the improvement effect of the TCN-GRU model in improving the accuracy of fault diagnosis, using the same training set, based on the SABO-TVFEMD decomposition method, models composed of TCN-GRU, TCN, GRU, and LSTM were trained respectively. Table 6 shows the average classification accuracy of each model on the test set. Figure 20 Shows the classification accuracy of each model in 10 experiments in Table 6, Figure 21 Shows the confusion matrix of one of the experiments.

[0106]

[0107] Table 6 shows the average classification accuracy of each model on the test set, clearly comparing their performance in the fault diagnosis task. Among the four models evaluated, the SABO-TVFEMD-TCN-GRU model achieved the highest average accuracy of 96.38%, indicating that combining the TCN-GRU architecture with the SABO-TVFEMD decomposition method can significantly improve the diagnostic performance. The SABO-TVFEMD-TCN model outperformed the SABO-TVFEMD-GRU and SABO-TVFEMD-LSTM models, reaching an average accuracy of 91.88%, but still lower than the performance of the SABO-TVFEMD-TCN-GRU model. The results show that although the LSTM, GRU, and TCN architectures each have certain advantages, they may have limitations in fully capturing the complex features of fault data. In contrast, the combined TCN-GRU structure shows stronger capabilities in extracting temporal dependencies and processing the non-linear features of signals, thus improving the accuracy of fault classification.

[0108] The present invention simulates three typical mechanical faults through experiments: stuck drive shaft gears, loose drive shaft screws, and loose arc plates, and reveals the vibration characteristic differences under different fault conditions through vibration signal analysis. The SABO algorithm is used to optimize the bandwidth threshold and B-spline order of TVFEMD, and then effectively extract the intrinsic mode components of the signal, providing rich feature information for subsequent fault classification. By comparing different classification models, the SABO-TVFEMD-TCN-GRU model shows significant advantages in the fault diagnosis task. Compared with other common classification models, the present invention not only surpasses the traditional method in terms of accuracy but also demonstrates stronger robustness, especially with higher accuracy and stability in multi-fault mode recognition. In summary, the SABO-TVFEMD-TCN-GRU model provides an innovative and efficient solution for the mechanical fault diagnosis of on-load tap changers, which can not only improve the accuracy of fault diagnosis but also adapt to signal changes in complex environments, and has important application value for the online monitoring and fault warning of smart grid equipment.

Claims

1. A method for diagnosing mechanical faults of on-load tap-changers based on optimizing hybrid networks with SABO, characterized in that The following steps are involved: S1: Simulates a typical mechanical fault of the on-load tap-changer and collects the generated vibration signal and the drive motor current signal; S2: Divide the vibration signal and drive motor current signal samples obtained in step S1 into a training set and a validation set; S3: Bandwidth threshold of time-varying filtered empirical mode decomposition TVFEMD using SABO algorithm and B-spline order Perform search optimization, retrieve parameter combinations and use TVFEMD to decompose the signal, and use the minimum information entropy as the fitness function to evaluate the quality of each set of parameters; S4: Update the search agent and continuously adjust the bandwidth threshold and the B-spline order until the stopping criterion is met; S5: Using the best parameter combination obtained by optimization, TVFEMD is applied to perform the final decomposition of the vibration signal to generate IMF components; S6: Input the decomposed IMF components and the driving motor current signal into the TCN-GUR classification model for training and classification; S7: Output fault diagnosis results.

2. The on-load tap-changer mechanical fault diagnosis method based on the SABO-optimized hybrid network according to claim 1, wherein, The SABO algorithm searches and optimizes the bandwidth threshold and the B-spline order The specific steps are as follows: S3.

1. First, define the operator as: (1); A and B are different individuals, sign is the sign function; F(A) is the fitness value of A; F(B) is the fitness value of B; is a vector with the same dimension as the search space and is a random number generated within [1, 2]; represents the Hadamard product of two vectors; S3.

2. Randomly initialize the population in the optimization space: (2); where: is an individual; is the upper boundary of optimization; is the lower boundary of optimization, and is a random number between [0, 1]; r is a random number vector uniformly distributed in the interval [0, 1]; S3.

3. Calculate the new position of each search agent as follows : (3); where is a vector with the same dimension as the search space, composed of random numbers between [0, 1]; is the total number of individuals; X i is the current position vector of the i-th search agent; X j is the current position vector of the j-th search agent; v is a random perturbation vector; is the newly calculated position of the i-th individual. If the new position is better, replace the previous position, otherwise keep the original position; that is, (4); is the objective function value of the search agent , is the objective function value of the search agent .

3. The on-load tap-changer mechanical fault diagnosis method based on the SABO-optimized hybrid network according to claim 1, wherein TVFEMD includes two stages: S3.4 local cutoff frequency rearrangement and S3.5 screening based on time-varying filtering. The steps of local cutoff frequency rearrangement are as follows: S3.4.

1. Perform Hilbert transform on the real-valued signal to obtain the analytic signal , (5); where, A(t) is the amplitude; a1(t) is the real component of the analytic signal, a2(t) is the imaginary component of the analytic signal; j is the imaginary unit; is the Hilbert transform of; φ(t), φ1(t), φ2(t) are phases, and (6); S3.4.

2. When occurs, A(t) has a local maximum; when occurs, A(t) has a local minimum; interpolate all the maximum points of A(t) to obtain the curve , interpolate all the minimum points of A(t) to obtain the curve , and obtain (7); S3.4.

3. Interpolate all the maximum points of A 2 at (t) to obtain , and interpolate all the minimum points of A at (t) to obtain 2 , and let be obtained by interpolation; where , (8); where is the derivative of , and φ2 ' at (t) is the derivative of ; the local cut-off frequency is: (9); S3.4.

4. Define signals The time series of the maximum points of is (10); (u i : u i+1 )is the time interval between two consecutive local extreme points; Take ρ = 0.

25. If the above formula holds, then u i is called a discontinuous point. Let , denote the discontinuous point sequence; If then it is called as 's rising edge, is regarded as the lowest value; If then it is called as 's falling edge, is regarded as the lowest value; The remaining part is regarded as the peak value; S3.4.

5. Interpolate between all the peaks to obtain the rearranged local cut-off frequency .

4. The on-load tap-changer mechanical fault diagnosis method based on the SABO-optimized hybrid network according to claim 3, characterized in that The steps of the screening phase based on time-varying filtering include: S3.5.

1. Reconstruct the signal h(t) according to the rearranged cut-off frequency: (11); Take 's extreme points as nodes, divide into n segments, with each segment having a step size of m; where n is called the order of the B-spline function; perform B-spline interpolation approximation between the extreme points, and the approximation result is denoted as , which is the local mean curve; S3.5.2, Determine the residual signal Whether it satisfies the stopping criterion , where is the given bandwidth threshold; if it is satisfied, then is the IMF; otherwise, let , repeat all the above steps until the stopping criterion is satisfied; where (12); The Loughlin instantaneous bandwidth B loughlin (t) is calculated by the formula: (13); The weighted average instantaneous frequency φ avg (t) is calculated by the formula: (14); a1 ′2 (t) and a2 ′2 (t) are the squared terms of the signal change rate.

5. The on-load tap-changer mechanical fault diagnosis method based on the SABO-optimized hybrid network according to claim 4, wherein The TCN-GRU model uses TCN to efficiently extract temporal features from multimodal signals to explore their local and global patterns in the time dimension. The extracted temporal features are then input into the gated recurrent unit (GRU) network to further enhance the modeling capabilities of dynamic features, ultimately achieving accurate classification and identification of different fault types.

Images