A high-voltage circuit breaker fault diagnosis method

Abstract

The application relates to the technical field of power system equipment monitoring, and discloses a high-voltage circuit breaker fault diagnosis method, which comprises the following steps: step one, collecting a vertical direction vibration signal in the operation process of a high-voltage circuit breaker, and dividing the vibration signal into a training set and a test set according to a preset ratio after digital processing; step two, dynamically optimizing and improving key parameters of an adaptive noise complete ensemble empirical mode decomposition algorithm by using a Newton-Raphson optimizer to obtain an optimized signal decomposition algorithm; and step three, recursively decomposing the vibration signals in the training set and the test set by using the optimized signal decomposition algorithm to generate a plurality of intrinsic mode function components. The key parameters of the adaptive noise complete ensemble empirical mode decomposition algorithm are dynamically optimized and improved by using the Newton-Raphson optimizer, the modal aliasing phenomenon in the vibration signal of the high-voltage circuit breaker is effectively inhibited, and the signal decomposition precision and feature clarity are improved.

Description

Technical Field

[0001] This invention relates to the field of power system equipment monitoring technology, specifically a method for diagnosing faults in high-voltage circuit breakers. Background Technology

[0002] As the core equipment of the power transmission and distribution system, high-voltage circuit breakers effectively protect electrical equipment from short circuits and overcurrents by isolating fault areas through opening and closing circuits, which is crucial for ensuring the stable operation of the power system. However, under complex operating environments and long-term mechanical stress, they are prone to mechanical failures due to mechanical wear, contact aging, spring fatigue, and other problems, which can endanger the reliability and safety of the power grid, causing large-scale power outages and significant economic losses. Therefore, it is of great significance to quickly and accurately identify the fault types of high-voltage circuit breakers and deal with them in a timely manner. Currently, high-voltage circuit breaker fault diagnosis technologies are mainly divided into three categories: model-driven methods, hybrid methods, and data-driven methods. Model-driven methods analyze faults by constructing physical or mathematical models of signals, but their accuracy is limited due to the nonlinear and non-stationary characteristics of high-voltage circuit breaker vibration signals and their sensitivity to external factors, making them susceptible to noise interference under complex operating conditions. Hybrid methods combine physical models with machine learning techniques, using simulation data to train models to improve diagnostic robustness and adaptability, but they have shortcomings when applied to complex systems and are costly to develop. Data-driven methods automatically extract signal features using algorithms such as support vector machines, random forests, and convolutional neural networks, significantly improving diagnostic accuracy and efficiency, and have become the main research method in this field. However, their performance degrades significantly and their generalization ability is limited in scenarios with scarce fault samples or high noise levels. Therefore, optimizing signal decomposition techniques and feature extraction methods is key to improving the performance of data-driven methods.To extract fault features under complex background noise interference, the vibration signal of a high-voltage circuit breaker needs to be decomposed and processed. Traditional signal processing techniques cannot balance frequency and time resolution, limiting their application to non-stationary signals. Wavelet transform improves the analysis of non-stationary signals through multi-scale decomposition, but its high-frequency resolution remains limited. Wavelet packet transform improves the accuracy of feature extraction from non-stationary data by utilizing full-band decomposition, but its ability to process complex nonlinear signals is limited and it is susceptible to noise. Empirical mode decomposition (EMD) enhances its adaptability to the nonlinear and non-stationary characteristics of vibration signals, but it is prone to mode aliasing, reducing feature extraction accuracy. Integrated EMD improves the stability and accuracy of signal decomposition by adding Gaussian white noise and enhances diagnostic accuracy by combining it with a BP neural network for fault classification, but it suffers from mode aliasing, reconstruction errors, and high computational costs. A fully integrated EMD and adaptive noise method improves decomposition stability and eliminates reconstruction errors by introducing adaptive noise, alleviating problems such as low computational efficiency, residual noise, and mode aliasing to some extent, but it still has limitations in processing... In complex signals and strong noise backgrounds, the problems of mode aliasing and noise residue remain unresolved. Fault diagnosis techniques based on Hilbert marginal spectral energy entropy and variational mode decomposition (VMD) have eliminated mode aliasing through accurate decomposition and feature extraction. However, the performance of classical VMD is highly dependent on the selection of the quadratic penalty factor and the number of mode decompositions, requiring parameter adjustment through optimization techniques. A diagnostic method based on improved fully integrated empirical mode decomposition with adaptive noise has improved denoising performance through multi-resolution singular value decomposition, solving the problems of mode aliasing and residual noise. It also combines variational mode decomposition and power spectral entropy quantization to extract vibration signal features, improving diagnostic robustness and accuracy. However, under strong noise conditions, it is highly sensitive to parameter settings, which may lead to a significant decrease in decomposition performance. The whale optimization algorithm has been used to optimize the noise realization number and noise standard deviation parameters of the improved fully integrated empirical mode decomposition with adaptive noise, improving the decomposition stability and denoising performance of ultrasonic signals under strong noise interference. However, the whale optimization algorithm is prone to slow convergence, local optima, and insufficient accuracy during the optimization process.

[0003] In the process of feature extraction and fault classification, there are still problems such as small sample size, complex and diverse features, and the need for high classification accuracy. Existing fault diagnosis methods based on convolutional neural networks, improved support vector machines, deep belief networks, convolutional attention networks, gray wolf optimized multi-granularity cascaded forests, variational mode decomposition, and whale optimization algorithms have achieved certain results in their respective application scenarios. However, the overall diagnostic performance under conditions of small sample size, high noise, and ambiguous signal features still needs to be improved. Although the hybrid diagnostic model combining convolutional neural networks and support vector machines can improve diagnostic robustness and accuracy and make up for the shortcomings of single technologies, the diagnostic accuracy will decrease when identifying ambiguous fault signals. It is urgent to improve the diagnostic accuracy and noise resistance of the model through optimization. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a fault diagnosis method for high-voltage circuit breakers, solving the problems mentioned in the background section.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a high-voltage circuit breaker fault diagnosis method, characterized by comprising the following steps: Step 1: Collect vertical vibration signals during the operation of the high-voltage circuit breaker. After digital processing of the vibration signals, divide them into training set and test set according to a preset ratio. Step 2: The key parameters of the improved adaptive noise complete set empirical mode decomposition algorithm are dynamically optimized using the Newton-Raphson optimizer to obtain the optimized signal decomposition algorithm. Step 3: Use the optimized signal decomposition algorithm to recursively decompose the vibration signals in the training set and the test set to generate multiple intrinsic mode function components. Step 4: Based on the variance contribution rate, select a specified number of intrinsic mode function components, extract time-domain features from each selected intrinsic mode function component, and construct a feature matrix; Step 5: The hyperparameters of the convolutional neural network-support vector machine model are optimized using the crowned porcupine optimizer to obtain the optimized classification model; Step 6: Input the feature matrix corresponding to the training set into the optimized classification model for training to obtain the trained fault diagnosis model; Step 7: Input the feature matrix corresponding to the test set into the trained fault diagnosis model and output the fault diagnosis results of the high-voltage circuit breaker.

[0006] Preferably, in step one, a measurement range of 0-5000 m / s is used. 2Vibration signals are collected by an accelerometer with a sampling rate of 10kHz. The signals are digitized by a data acquisition card. The collected vibration signals cover four working conditions: normal state, electromagnet blockage, spindle blockage, and half-shaft blockage. The vibration signal for each working condition is a 1×1000 time-domain sequence. The training set and test set are divided by stratified sampling to ensure that the sample ratio of each category is consistent.

[0007] Preferably, in step two, the key parameters of the improved adaptive noise complete set empirical mode decomposition algorithm include the noise standard deviation and the maximum number of modes. The Newton-Raphson optimizer optimizes the parameters by minimizing the envelope entropy of the first K intrinsic mode function components. During the optimization process, the fast convergence of gradient optimization and the perturbation strategy are combined to enhance the stability of parameter optimization and global optimization capability.

[0008] Preferably, in step three, the optimized signal decomposition algorithm extracts intrinsic mode function components by adaptively adding white noise and recursively decomposing the residuals. During the decomposition process, the noise amplitude is dynamically adjusted according to the statistical characteristics of the residual signal until the residuals cannot be decomposed. The final generated intrinsic mode function components cover high-frequency components to low-frequency components and maintain the physical characteristics and hierarchy of the signal.

[0009] Preferably, in step four, the specified number of intrinsic mode function components is 5, and the extracted time-domain features include signal entropy, high-frequency entropy, low-frequency entropy, root mean square, kurtosis, and skewness. The dimension of the feature matrix is ​​determined according to the number of selected intrinsic mode function components and the number of extracted time-domain features.

[0010] Preferably, in step five, the hyperparameters of the convolutional neural network-support vector machine model include the initial learning rate, mini-batch size, and learning rate reduction factor. The porcupine optimizer optimizes the hyperparameters by simulating the porcupine's defense mechanism. During the optimization process, the mean squared error of the classification model is used as the fitness function to dynamically adjust the population position and balance global exploration with local development.

[0011] Preferably, in step five, the convolutional neural network includes two convolutional layers, each followed by a batch normalization layer and a ReLU activation function, and the second convolutional layer is followed by a Dropout layer. The support vector machine uses a Gaussian kernel function to construct a nonlinear classification boundary.

[0012] Preferably, in step six, the classification model is trained in an end-to-end manner, the parameters are optimized by the backpropagation algorithm, the prediction error is quantified by the cross-entropy loss function, and the parameters are updated by the stochastic gradient descent momentum optimizer.

[0013] Preferably, in step seven, the fault diagnosis results include the operating status category of the high-voltage circuit breaker and the corresponding confidence level. The operating status category includes normal state, tripping closing electromagnet blockage, main shaft blockage, and half shaft blockage. The performance of the fault diagnosis model is evaluated by accuracy, precision, recall rate, and F1 score.

[0014] Preferably, the method also includes a real-time verification step for the fault diagnosis model. By analyzing the computational complexity of the training and inference phases, it ensures that the total time for diagnosing a single set of vibration signals during the inference phase does not exceed 2 seconds, thus meeting the real-time requirements for online fault diagnosis of high-voltage circuit breakers.

[0015] This invention provides a method for fault diagnosis of high-voltage circuit breakers. It has the following beneficial effects: 1. This invention improves the key parameters of the adaptive noise complete set empirical mode decomposition algorithm by dynamically optimizing the Newton-Raphson optimizer, effectively suppressing mode aliasing in the vibration signal of high-voltage circuit breaker, improving the signal decomposition accuracy and feature clarity, providing high-quality data support for subsequent fault feature extraction, and realizing efficient decomposition of complex nonlinear and non-stationary vibration signals.

[0016] 2. This invention utilizes the Crowned Porcupine Optimizer to precisely adjust the hyperparameters of the convolutional neural network-support vector machine model, enhancing the model's ability to extract and discriminate deep fault features, while also improving the model's robustness in noisy environments, thus achieving accurate identification and classification of various fault types in high-voltage circuit breakers.

[0017] 3. This invention constructs an integrated diagnostic framework that combines signal decomposition, feature screening, deep feature extraction, and optimized classification. Through the synergistic effect of front-end signal decomposition optimization and back-end classification model optimization, it balances the accuracy and real-time performance of fault diagnosis, meets the stringent requirements of online fault diagnosis of high-voltage circuit breakers, and achieves a dual improvement in diagnostic performance and engineering practicality. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the fault classification framework of CNN-SVM optimized by CPO according to the present invention.

[0019] Figure 2 This is a schematic diagram of the diagnostic process of the NRBO-ICEEMDAN-CPO-CNN-SVM framework of the present invention.

[0020] Figure 3 This is a time-domain plot of vibration signals under four mechanical states acquired in real time based on the YD-39 accelerometer.

[0021] Figure 4 The image shows the time-domain decomposition results of the NRBO-ICEEMDAN vibration signal.

[0022] Figure 5 The graph shows the frequency domain decomposition results of the EMD vibration signal.

[0023] Figure 6 The image shows the frequency domain decomposition results of the EEMD vibration signal.

[0024] Figure 7 The image shows the frequency domain decomposition results of the ICEEMDAN vibration signal.

[0025] Figure 8 The result of frequency domain decomposition of VMD vibration signal is shown in the figure.

[0026] Figure 9 The image shows the frequency domain decomposition results of the NRBO-ICEEMDAN vibration signal.

[0027] Figure 10 The graph shows the mean values ​​of the IMF mode aliasing quantization results for spectral overlap and frequency dispersion under four operating conditions.

[0028] Figure 11 The figure shows a comparison of reconstruction errors for different decomposition methods.

[0029] Figure 12 Example diagram of the IMF screening process and indicator optimization modality.

[0030] Figure 13 This is a performance comparison chart of different screening criteria under varying IMF numbers.

[0031] Figure 14 A comparison chart showing the classification accuracy of different decomposition algorithms combined with CNN.

[0032] Figure 15 A comparison chart showing the classification performance (precision, recall, F1 score) of different decomposition algorithms combined with CNN.

[0033] Figure 16 The image shows the confusion matrix of the NRBO-ICEEMDAN-CPO-CNN-SVM classifier on the test set.

[0034] Figure 17 This is a comparison chart of the convergence of the training process of five algorithms.

[0035] Figure 18 This is a comparison chart of the classification accuracy of five algorithms on the training and test sets.

[0036] Figure 19 The graph shows a comparison of the classification performance (F1 score) of the five algorithms under different working conditions.

[0037] Figure 20 This is a performance comparison chart of five algorithms in terms of precision, recall, and F1 score.

[0038] Figure 21 The graph shows the convergence characteristics of ICEEMDAN fitness based on different optimization algorithms.

[0039] Figure 22 The graph shows the convergence characteristics of the fitness of CNN-SVM based on different optimization algorithms. Detailed Implementation

[0040] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] Please see the appendix Figure 1 - Appendix Figure 22 This invention provides a method for diagnosing faults in high-voltage circuit breakers, including...

[0042] The remaining structure of this paper is as follows: Section 2 establishes the proposed NRBO-ICEEMDAN-CPO- A CNN-SVM framework for vibration signal fault diagnosis. Section 3 verifies the rationality and effectiveness of the proposed method and conducts an overall performance evaluation. Section 4 summarizes this paper.

[0043] For non-stationary and nonlinear vibration signals, the ICEEMDAN algorithm, through adaptive addition of white noise and recursive decomposition of residuals, can extract high-quality intrinsic mode functions (IMFs), suppress noise interference, and reduce mode aliasing. The specific steps of the ICEEMDAN algorithm are as follows: generate NR A set of independent white noise sequences were obtained, and their IMFs were derived through EMD decomposition. The original signal was then added to the controlled noise IMFs to construct a noisy signal. (1) In the formula, The original signal, Nstd The standard deviation of noise. For the first i The first IMF of white noise, For the first i A noisy signal.

[0044] Perform EMD decomposition on each noisy signal, extract the first IMF and average it, then calculate the residual: (2) In the formula, NR To achieve the quantity of noise, For the first i The first IMF with a noisy signal, This is the first and final IMF.

[0045] (3) In the formula, This represents the residual after the first round of decomposition.

[0046] In each level of decomposition, based on the residual signal The statistical characteristics dynamically adjust the noise amplitude: (4) In the formula, To adjust the factor, is the standard deviation of the residual signal.

[0047] For residuals Perform recursive decomposition, introducing independent white noise and averaging it, until the residuals are no longer decomposable: (5) In the formula, For the first k -1 round residual, For the first i The first white noise k indivual IMF , For the first k indivual IMF , To update the residuals.

[0048] In summary, the original signal can be reconstructed as: (6) In the formula, P represents the total number of IMFs obtained from the adaptive decomposition. This is the final residual.

[0049] The NRBO algorithm mainly consists of two parts: the Newton-Raphson Search Rule (NRSR) and the Trap Avoidance Operator (TAO). The NRSR employs a Newton-Raphson optimization strategy to accelerate the global search and improve convergence efficiency; the TAO enhances the algorithm's global search capability by dynamically adjusting population diversity, thus preventing the algorithm from getting trapped in local optima. This dual optimization mechanism ensures the efficiency and robustness of the NRBO algorithm in the parameter optimization process of complex model optimization problems.

[0050] NRSR is based on the gradient-driven search idea of ​​the Newton-Raphson method, which quickly updates the position of the solution using the second derivative. This rule originates from the following formula of Newton's method: (7) In the formula, Let n be the position vector of the nth iteration; and The first / second derivative of the objective function.

[0051] For non-differentiable optimization objective functions The first and second derivatives are approximated using numerical methods, as shown below: (8) In the formula, h is the difference step size.

[0052] Combining these approximate expressions for derivatives, the NRSR update formula can be expressed as: (9) To adapt to the characteristics of group search, the NRBO algorithm first optimizes NRSR into the following preliminary expression: (10) In the formula, and These represent the positions of the best and worst solutions in the current population, respectively. It is a normally distributed random vector used for directional perturbation to enhance search diversity; The range of population diversity can be automatically updated using the following formula: (11) In the formula, rand(·) is a uniform random vector in the range [0,1], used to scale the step size proportionally; For the decision vector dimension.

[0053] Then, the initial NRSR location needs to be revised and re-evaluated, based on which two candidate locations are constructed as follows: (12) In the formula, and There are two candidate positions; ,in , , To randomly select other individual indexes, The step size vector, derived from the idea of ​​differential evolution, aims to accelerate convergence and maintain the balance of population diversity.

[0054] To further improve convergence accuracy, and Second-order correction is performed on the midpoint between them: (13) In the formula, . This leads to the second-order refined NRSR expression: (14) In the formula, It is a normally distributed random vector.

[0055] Finally, by substituting the second-order step size mentioned above, we obtain the updated position of the individual: (15) NRSR uses adaptive parameters The dynamic equilibrium exploration and development phase effectively improves the population's search ability and accelerates the convergence to the global optimum. Adaptive parameters... The specific expression is: (16) In the formula, Indicates the current iteration number; This indicates the maximum number of iterations.

[0056] To enhance population diversity and avoid local optima, TAO is triggered by a probability of a trap avoidance factor (DF). This involves introducing random perturbation parameters and dynamically adjusting individual distribution by combining the global optimum with the population mean. The execution process of TAO can be divided into the following three steps: (a) Trigger determination: First, based on decision factors DF Controlling the trigger probability: (17) (b) Disturbance generation: Secondly, based on the current iteration number and the population state, two types of random factors are generated: (18) In the formula, and The disturbance coefficient determines the direction and intensity of the disturbance.

[0057] (19) In the formula, and For perturbation weights, by and Joint decision; These are binary control parameters.

[0058] And calculate the population mean. Used to guide overall movement: (20) (c) Location update: Finally, according to The size is updated using two different methods: (twenty one) In the formula, Let n be the position vector of the nth iteration; Taken from formula (16).

[0059] Update via TAO Parameters can enhance population diversity during the development phase and avoid the trap of local optima.

[0060] To avoid modal aliasing or inaccurate decomposition caused by improper parameter selection, this paper combines NRBO with ICEEMDAN. By leveraging NRBO's fast global search capability and TAO's dynamic optimization of Nstd and K in the ICEEMDAN algorithm, and through reasonable selection of hyperparameters, the decomposition accuracy, modal separation capability, and stability of ICEEMDAN can be improved.

[0061] In this study, Nstd and K These are respectively set as the first and second decision dimensions of the NRBO algorithm, as follows: The Nstd parameter controls the amplitude of noise injection and is typically selected empirically, ranging from 0.1 to 0.9. It is a core hyperparameter affecting the decomposition accuracy, stability, and modal aliasing degree of the ICEEMDAN algorithm. NRBO is dynamically adjusted through feedback from the fitness function. Nstd This avoids modal aliasing caused by excessively large or small noise amplitude, or inadequate decomposition due to improper noise amplitude selection.

[0062] First, perform population initialization: (twenty two) In the formula, For the first i The first dimension of each particle ( Nstd ); and These are the upper and lower bounds of the noise standard deviation, respectively.

[0063] Secondly, iterative updates are performed by first applying NRSR formulas (10)–(14) and TAO formulas (16–21) to calculate the increment: (twenty three) In the formula, and These are the components of the NRSR and TAO vectors in the noise dimension, respectively.

[0064] At the same time, based on probability DF Determine if a TAO strong perturbation has been triggered: (twenty four) Finally, after the update, it is truncated to [ ]. In the optimization framework, K The size of the modal analysis window is defined to specify the number of IMFs evaluated for the fitness function, and indirectly affects the number of decomposition layers in ICEEMDAN. NRBO dynamically adjusts this window. K This involves finding a balance between signal representation depth and computational efficiency. For example, when K=x, only the information entropy characteristics of the first x IMFs are evaluated, thereby optimizing frequency coverage and avoiding modal redundancy.

[0065] Similarly, population initialization should be performed first: (25) In the formula, For the first i The second dimension of each particle ( K ); This indicates rounding to the nearest integer. for K The upper boundary.

[0066] Iteration, the increment is calculated according to equations (10)–(14) and (16–21): (26) In the formula, and These are the components of the NRSR and TAO vectors in the modal dimension, respectively.

[0067] Then: (27) And guarantee . This paper introduces the envelope entropy of the IMF as a fitness function index; the lower the envelope entropy value, the purer the mode. The algorithm balances decomposition accuracy and mode aliasing suppression by minimizing the minimum value among the first K IMF envelope entropies and the optimal Nstd combination.

[0068] The envelope entropy of the IMF is calculated as follows: First, for each IMF component Performing the Hilbert transform yields the analytic signal: (28) Then, the modulus of the analytic signal is used as the envelope: (29) Then, normalize the envelope to obtain the probability distribution: (30) In the formula, Indicates the first j At a certain point in time, N Let be the signal length. Based on this, we obtain the _th _. i Envelope entropy of an IMF: (31) Finally, the fitness function, which is the minimum value of the envelope entropy of all IMFs, is obtained: (32) Optimize through NRBO iteration It can achieve Nstd and K Dynamic parameter adjustment optimizes the quality of ICEEMDAN decomposition results.

[0069] The number of noisy implementations (NR) for ICEEMDAN is set to 50, and the maximum number of screening iterations (MaxIter) is set to 300. The dynamic optimization process of NRBO-ICEEMDAN is as follows: First, parameters are randomly generated. Nstd and K The initial population and fitness were calculated. H Subsequently, NRSR is used to adjust particle positions, and TAO is combined to avoid local optima, updating the optimal parameter combination in real time. Nstd* , K*） After optimization, the optimal parameters are used for ICEEMDAN signal decomposition, and the entropy of the IMFs is used as the decomposition result. H Evaluation of decomposition effect The CPO optimization process draws on four defense mechanisms of porcupines in adapting to the environment: global exploration, directional search, local exploitation, and fine-tuning

[52] . First, CPO randomly initializes the population position to generate an initial parameter combination. The position of each porcupine individual corresponds to a set of hyperparameters, such as the initial learning rate, mini-batch size, and learning rate decrease factor. The initial position is generated by the following formula: (33) In the formula, Indicates the first i The first individual j Parameter values, and The first j Upper and lower bounds of each parameter, N For population size, dim The dimension of the parameter.

[0070] In each iteration, the CPO evaluates individuals in the population based on the fitness function and adjusts their positions through a dynamic update mechanism. The fitness function is defined as the mean squared error (RMSE) of the model, used to measure the quality of the hyperparameter combination, and its expression is: (34) In the formula, and These are the predicted value and the actual value, respectively. n The sample size is represented by the number of samples. Through feedback from fitness values, the CPO dynamically adjusts the positions of individuals, gradually converging the population to the global optimum.

[0071] During the search phase, the CPO updates the individual's location using the following formula to expand the search scope: (35) In the formula, This represents the current globally optimal solution. For random individuals in the population, randn and rand introduce random perturbations to increase search diversity.

[0072] During the development phase, the CPO optimizes its current position by adjusting its direction and changing the search radius, further balancing global exploration and local development. Its update formula is: (36) In the formula, The random weights representing directional search, A random individual in the population. and For two randomly selected different individuals r It is a random number in the interval [0,1].

[0073] During the local development phase, the CPO introduces the global optimal solution through the dynamic adjustment mechanism of formula (35). The random perturbation and randn() method maintain global search capability and diversity during the optimization process. The introduction of random weights and perturbations prevents individuals from getting trapped in local optima and expands the search scope in the early stages while gradually focusing on local development in the later stages.

[0074] To ensure the effectiveness of parameter updates, CPO uses the upper and lower bounds defined by formula (33) and processes parameter values ​​that exceed the range using a direct truncation method: (37) This mechanism ensures that parameter values ​​are always within a valid range, thus improving the stability of the optimization process.

[0075] To improve convergence performance in later stages, CPO employs a dynamic population size adjustment mechanism, the calculation formula of which is: (38) The dynamic reduction of population size allows the algorithm to concentrate resources on individuals with higher fitness in later stages, accelerating the local development process. This is combined with dynamic weight switching based on defense mechanisms. The CPO gradually shifts from global exploration to local development, eventually approaching the global optimal solution.

[0076] The performance of fault diagnosis models is highly dependent on structural design and hyperparameter configuration, especially under complex operating conditions, where the sensitivity to parameter selection is significant. Therefore, this paper proposes a model combining CPO to optimize CNN-SVM hyperparameters (CPO-CNN-SVM). This method leverages the strong global search capability and adaptability of the CPO algorithm in high-dimensional non-convex spaces, using the RMSE of the CNN-SVM model as the fitness function to dynamically optimize the key hyperparameters of the CNN (initial learning rate, mini-batch size, and learning rate descent factor). Subsequently, the deep features extracted by the optimized CNN are input into the SVM classifier to achieve the final fault identification.

[0077] First, CPO initializes the population and targets the RMSE error of CNN-SVM on the training set (Equation 34), aiming for the smallest possible value. It then optimizes according to (Equations 35-38) to finally output the optimal hyperparameter combination. CPO's optimization process is particularly suitable for non-convex hyperparameter spaces. Its bio-inspired global exploration and directional search strategy enables CPO to effectively search in high-dimensional non-convex spaces, avoiding getting trapped in local optima and ensuring a robust hyperparameter combination for the CNN-SVM model.

[0078] Secondly, the optimized CNN is used to extract high-dimensional features from the input data. Its core structure includes two convolutional layers (Conv1 and Conv2), each followed by batch normalization and ReLU activation functions to normalize the feature distribution and enhance the non-linear expressive power of the model.

[0079] ConvolutionalLayer extracts features from the input data through convolutional operations. Its formula is as follows: (39) In the formula, Input is the input data; Kernel is the convolution kernel; Bias is the bias term; and k is the size of the convolution kernel.

[0080] The BatchNormalizationLayer is used to accelerate training and reduce internal covariate bias. Its formula is as follows: (40) In the formula, It is the input data; and These are the mean and variance of the current batch of data; It is a small constant used for numerical stability; and These are learnable scaling and offset parameters.

[0081] ReLULayer is used to introduce nonlinearity. Its formula is as follows: (41) After the second convolutional layer, a Dropout layer is added to reduce the risk of overfitting. The formula is as follows: (42) In the formula, mask is a randomly generated binary matrix used to determine which neurons are discarded.

[0082] During training, the CNN employs an end-to-end architecture for optimization, including convolutional layers, fully connected layers, and softmax layers. This end-to-end training mechanism allows the convolutional layers to receive the complete gradient signal from the classification loss, thereby adaptively learning deep features of the input data. Specifically, during end-to-end training, the network optimizes its parameters using the backpropagation algorithm. For multi-class classification tasks, the cross-entropy loss function is used to quantify the prediction error, with the following form: (43) In the formula, Indicates the number of samples. Indicates the number of categories. For real labels, The network predicts the output. Through backpropagation, the gradient is passed from the output layer to each part of the network.

[0083] For a fully connected layer, the weight gradient is calculated as follows: (44) In the formula, Indicates the weights of the fully connected layer. For the output of the fully connected layer, This is the activation value of the previous layer.

[0084] For convolutional layers, considering the characteristics of convolution operations, the weight gradients are calculated in the following way: (45) Furthermore, the batch normalization layer also involves gradient propagation during training, and its learnable parameters... and The gradients are respectively: (46) (47) In the formula, For standardized input, This is for batch normalization layer output.

[0085] Based on these gradients, parameter updates are implemented using the stochastic gradient descent momentum method (SGDM) optimizer. The update process is as follows: (48) (49) In the formula, For momentum term, The momentum coefficient, For learning rate, This represents network parameters.

[0086] To ensure the effective propagation of gradients during backpropagation and avoid gradient vanishing or exploding problems, hyperparameters are dynamically adjusted through the CPO optimization strategy, thereby improving feature extraction quality and model convergence performance.

[0087] In the feature extraction stage, the ReLU2 layer output is chosen as the input feature for the SVM because it has been fully optimized through backpropagation. Through end-to-end training, the ReLU2 layer receives complete gradient information from the classification loss, resulting in a more discriminative feature representation. Compared to the fully connected layer output, the ReLU2 layer features retain more spatial structure information while avoiding information loss due to over-abstraction. Finally, to fully utilize these deep features for classification, this paper selects the output of the second convolutional layer (ReLU2 layer) in the trained CNN as the high-dimensional feature input to the SVM classifier, rather than relying on the fully connected layer output, to avoid introducing redundant information. Based on the principle of knowledge transfer, the abstract features extracted by the CNN are directly mapped to the decision space of the SVM, thereby improving classification performance.

[0088] Given the nonlinear, non-stationary, and noise-interference characteristics of HVCB vibration signals, this paper selects the Gaussian kernel (RBF kernel) as the SVM kernel function. The Gaussian kernel constructs nonlinear boundaries through high-dimensional mapping, effectively capturing complex signal modes. Its locality significantly reduces noise influence and improves robustness, making it suitable for fault diagnosis of vibration signals. Through CPO-optimized hyperparameters, the CNN-SVM model can better adapt to the non-convex decision boundaries constructed by the SVM Gaussian kernel, improving the model's diagnostic performance and generalization ability in complex signals.

[0089] SVM constructs a nonlinear classification boundary using a Gaussian kernel function, which is defined as follows: (50) In the formula, and For the input sample, These are kernel parameters that control the sensitive range of the kernel function.

[0090] SVM training is achieved by solving the following optimization problem: (51) Constraints: (52) In the formula, and b Define the orientation and position of the hyperplane respectively. These are slack variables, allowing for a certain classification error; C The regularization parameter adjusts the tradeoff between boundary maximization and error tolerance. Combined with a Gaussian kernel, SVM optimizes nonlinear boundaries in high-dimensional space.

[0091] To accurately identify the fault characteristics of HVCB vibration signals, this study synergistically optimizes signal decomposition and classification techniques. Therefore, this paper combines the NRBO-ICEEMDAN signal decomposition method with the CPO-CNN-SVM model to construct the NRBO-ICEEMDAN-CPO-CNN-SVM framework (see details). Figure 2 To achieve efficient collaboration from signal decomposition to fault classification.

[0092] Signal decomposition and feature extraction form the foundation of this framework. This paper employs NRBO-ICEEMDAN to recursively decompose vibration signals, generating multiple Integrated Mode Factors (IMFs) containing signal features from different frequency bands. To remove redundancy and improve classification efficiency, IMF components are selected based on indicators such as sample entropy, variance contribution rate, and correlation coefficient, retaining representative key modes. Then, six statistical features—signal entropy, high-frequency entropy, low-frequency entropy, root mean square (RMS), kurtosis, and skewness—are extracted to construct the input feature vector for the classifier.

[0093] The constructed feature vectors are input into the CPO-CNN-SVM classifier. CNN deeply extracts high-dimensional features to enhance discriminative ability; SVM uses Gaussian kernels to construct nonlinear classification boundaries to accurately identify fault modes. By optimizing the initial learning rate, mini-batch size, and learning rate reduction factor in the CNN-SVM model through CPO, the parameters are reasonably adjusted to achieve a balance between classification performance, convergence speed, and model robustness.

[0094] This experiment collected data under four operating conditions (each a 1×1000 time-domain sequence): normal state, tripped electromagnet blockage, spindle blockage, and half-shaft blockage. Each condition was tested 100 times, resulting in a total data size of 400×1000 (typical signals are shown in...). Figure 3 Stratified sampling was used to divide the dataset into a training set (350 groups) and a test set (50 groups), ensuring that the proportion of each category was close to 25% (see Table 3). Specifically, the allocation was: 87 training groups and 13 test groups for each of the two work conditions, and 88 training groups and 12 test groups for each of the other two categories. To ensure unbiasedness, the number of training samples (87 or 88 groups) was randomly assigned using random sampling. Experimental results showed that the difference in allocation had a minimal impact on model performance (<0.1%), with no statistical significance. Figure 3 As shown in (a)-3(d), the time-domain differences of vibration signals are not obvious, and the analysis of weak signal characteristics is crucial for state identification.

[0095] To verify the performance of NRBO-ICEEMDAN in signal decomposition, this paper first conducts a time-domain decomposition experiment on HVCB vibration signal based on NRBO-ICEEMDAN. Figure 4 The time-domain decomposition results of the vibration signal are presented. It is evident that the 10 IMF components are clearly visible in the time domain, demonstrating significant modal separation. From high-frequency components (IMF1-IMF4) to low-frequency components (IMF7-IMF10), the hierarchical structure of the decomposition and the physical characteristics of the signal are well preserved, verifying the superior performance of NRBO-ICEEMDAN in time-domain signal decomposition.

[0096] To further verify the decomposition accuracy of the NRBO-ICEEMDAN method, this paper quantitatively evaluates it from the perspective of reconstruction error. By comparing the error magnitudes of five methods in signal reconstruction, the accuracy and reliability of the NRBO-ICEEMDAN decomposition method are measured. The error calculation results intuitively reflect the performance of different decomposition methods in signal reconstruction capability, providing a quantitative basis for the evaluation of methods. The comparison results of reconstruction errors of different decomposition methods are shown below. Figure 11 As shown. Figure 11 As shown, they differ significantly in error magnitude and distribution characteristics. EMD exhibits smaller reconstruction error fluctuations (10). -15 While stable (on the order of magnitude), it has a weak ability to adapt to complex signals; EEMD's error increases significantly (on the order of 100) due to the introduction of random noise, and fluctuates frequently; VMD's error is also on the order of 100, with relatively small overall fluctuations, but there are spikes at local time points (such as around 460ms). ICEEMDAN significantly reduces the error (on the order of 10) based on the improvement of EEMD. -15 With a relatively small fluctuation (on the order of magnitude), it balances accuracy and noise resistance. NRBO-ICEEMDAN further reduces the error to 10^10 through non-recursive optimization. -15 The magnitude is small, the fluctuation is minimal, and the performance is optimal. This has been verified through multiple experiments. Figure 11 The reconstruction error range for a representative operating condition is shown to be -1×10. -15 Up to 1×10-15 The errors for the other three operating conditions are also within this range, with a standard deviation of 10. -16 The order of magnitude is below 1. Therefore, this result demonstrates that NRBO-ICEEMDAN maintains extremely high reconstruction accuracy and consistency under different operating conditions, validating its adaptability and stability in varied and complex working conditions.

[0097] Figure 16 The figure shows the confusion matrix results of NRBO-ICEEMDAN-CPO-CNN-SVM on the test set. As shown, this algorithm achieved a 100% detection rate for the Normal and Coil_Jam categories. For the Spindle_Jam category, the detection rate was 100%, but the precision was 92.9% (13 / 14), meaning that 7.1% (1 / 14) of all samples predicted as 'Spindle_Jam' were actually HalfShaft_Jam. For the HalfShaft_Jam category, the detection rate was 92.3% (12 / 13), and the precision was 100%, indicating that one HalfShaft_Jam sample was misclassified as Spindle_Jam. Despite this slight confusion, the model achieved an overall classification accuracy of 98%, demonstrating excellent performance and effectiveness and robustness in handling these failure modes.

[0098] Furthermore, to verify the effectiveness of the proposed model across various modules, the performance of five different algorithms was compared and analyzed from multiple dimensions, including training convergence, classification accuracy, recognition performance of each fault category, and core evaluation metrics: CNN, CNN-SVM, CPO-CNN-SVM, ICEEMDAN-CNN-SVM, and NRBO-ICEEMDAN-CPO-CNN-SVM (where ICE represents ICEEMDAN). A comparison of the convergence of the training processes of the five algorithms can be found in [link to documentation]. Figure 17 .

[0099] Figure 22The study demonstrates the fitness changes of CPO, GA, and WOA during the optimization of CNN-SVM model hyperparameters (using the CNN-SVM root mean square error (RMSE) as the evaluation criterion, with lower values ​​indicating better performance). The results show that CPO performs best, with an initial fitness value of 0.0914, which rapidly decreases to 0.0834 in the third iteration and stabilizes after converging to a minimum value of 0.0416 in the fourth iteration, exhibiting both fast convergence and high accuracy. GA's initial fitness value is 0.0922, showing limited improvement in the first four iterations, only decreasing to 0.0589 and stabilizing in the fifth iteration, indicating insufficient exploration ability and slow convergence speed, but with moderate final accuracy. WOA's initial fitness value is the highest at 0.0934, only decreasing to 0.0789 and stabilizing in the fourth iteration, lagging behind CPO in both convergence speed and final accuracy, making it prone to suboptimal solutions. Overall, CPO outperforms GA (5 iterations, 0.0589) and WOA (4 iterations, 0.0789) in both the number of iterations (4) and the final accuracy (0.0416), providing a more efficient and reliable solution for hyperparameter tuning of CNN-SVM models.

[0100] Furthermore, the actual running times of the three algorithms, CPO, WOA, and GA, were compared to evaluate their computational efficiency. CPO completed parameter optimization in a total time of 1500.75 seconds, significantly lower than WOA's 2050.65 seconds and GA's 6700.37 seconds. CPO saved approximately 549.90 seconds compared to WOA (a reduction of approximately 26.82% in computation time) and approximately 5199.62 seconds compared to GA (a reduction of approximately 77.60% in computation time). Combined with the convergence performance comparison analysis, CPO achieved a multi-dimensional balance and leadership in optimizing CNN-SVM model hyperparameters, considering optimization accuracy, convergence iterations, and computation time.

[0101] In summary, NRBO achieved equivalent or better solutions in the ICEEMDAN parameter optimization experiments with faster convergence speed and lower computational cost, while CPO also demonstrated excellent efficiency and accuracy in CNN-SVM model hyperparameter tuning. These results validate the rationality and necessity of using NRBO and CPO in the signal decomposition and model tuning stages, respectively.

[0102] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A fault diagnosis method for high-voltage circuit breakers, characterized in that, Includes the following steps: Step 1: Collect vertical vibration signals during the operation of the high-voltage circuit breaker. After digital processing of the vibration signals, divide them into training set and test set according to a preset ratio. Step 2: The key parameters of the improved adaptive noise complete set empirical mode decomposition algorithm are dynamically optimized using the Newton-Raphson optimizer to obtain the optimized signal decomposition algorithm. Step 3: Use the optimized signal decomposition algorithm to recursively decompose the vibration signals in the training set and the test set to generate multiple intrinsic mode function components. Step 4: Based on the variance contribution rate, select a specified number of intrinsic mode function components, extract time-domain features from each selected intrinsic mode function component, and construct a feature matrix; Step 5: The hyperparameters of the convolutional neural network-support vector machine model are optimized using the crowned porcupine optimizer to obtain the optimized classification model; Step 6: Input the feature matrix corresponding to the training set into the optimized classification model for training to obtain the trained fault diagnosis model; Step 7: Input the feature matrix corresponding to the test set into the trained fault diagnosis model and output the fault diagnosis results of the high-voltage circuit breaker.

2. The high-voltage circuit breaker fault diagnosis method according to claim 1, characterized in that, In step one, a measurement range of 0-5000 m / s is used. 2 Vibration signals are collected by an accelerometer with a sampling rate of 10kHz. The signals are digitized by a data acquisition card. The collected vibration signals cover four working conditions: normal state, electromagnet blockage, spindle blockage, and half-shaft blockage. The vibration signal for each working condition is a 1×1000 time-domain sequence. The training set and test set are divided by stratified sampling to ensure that the sample ratio of each category is consistent.

3. The high-voltage circuit breaker fault diagnosis method according to claim 1, characterized in that, In step two, the key parameters of the improved adaptive noise complete set empirical mode decomposition algorithm include the noise standard deviation and the maximum number of modes. The Newton-Raphson optimizer optimizes the parameters by minimizing the envelope entropy of the first K intrinsic mode function components. During the optimization process, the fast convergence of gradient optimization and the perturbation strategy are combined to enhance the stability of parameter optimization and global optimization capability.

4. The high-voltage circuit breaker fault diagnosis method according to claim 1, characterized in that, In step three, the optimized signal decomposition algorithm extracts intrinsic mode function components by adaptively adding white noise and recursively decomposing the residuals. During the decomposition process, the noise amplitude is dynamically adjusted according to the statistical characteristics of the residual signal until the residuals cannot be decomposed. The final generated intrinsic mode function components cover high-frequency components to low-frequency components and maintain the physical characteristics and hierarchy of the signal.

5. The high-voltage circuit breaker fault diagnosis method according to claim 1, characterized in that, In step four, the specified number of intrinsic mode function components is five, and the extracted time-domain features include signal entropy, high-frequency entropy, low-frequency entropy, root mean square, kurtosis, and skewness. The dimension of the feature matrix is ​​determined based on the number of selected intrinsic mode function components and the number of extracted time-domain features.

6. The high-voltage circuit breaker fault diagnosis method according to claim 1, characterized in that, In step five, the hyperparameters of the convolutional neural network-support vector machine model include the initial learning rate, mini-batch size, and learning rate reduction factor. The porcupine optimizer optimizes the hyperparameters by simulating the porcupine's defense mechanism. During the optimization process, the mean squared error of the classification model is used as the fitness function to dynamically adjust the population position and balance global exploration with local development.

7. The high-voltage circuit breaker fault diagnosis method according to claim 1, characterized in that, In step five, the convolutional neural network contains two convolutional layers. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function. The second convolutional layer is followed by a Dropout layer. The support vector machine uses a Gaussian kernel function to construct a non-linear classification boundary.

8. The high-voltage circuit breaker fault diagnosis method according to claim 1, characterized in that, In step six, the classification model is trained in an end-to-end manner. The parameters are optimized by backpropagation algorithm, the prediction error is quantified by cross-entropy loss function, and the parameters are updated by stochastic gradient descent momentum optimizer.

9. The high-voltage circuit breaker fault diagnosis method according to claim 1, characterized in that, In step seven, the fault diagnosis results include the operating status category of the high-voltage circuit breaker and the corresponding confidence level. The operating status categories include normal state, tripping and closing electromagnet blockage, main shaft blockage, and half shaft blockage. The performance of the fault diagnosis model is evaluated by accuracy, precision, recall rate, and F1 score.

10. The high-voltage circuit breaker fault diagnosis method according to claim 1, characterized in that, The method also includes a real-time verification step for the fault diagnosis model. By analyzing the computational complexity of the training and inference phases, it ensures that the total time for diagnosing a single set of vibration signals during the inference phase does not exceed 2 seconds, thus meeting the real-time requirements for online fault diagnosis of high-voltage circuit breakers.

Images