GIS circuit breaker opening and closing abnormal state identification method and system based on vibration signals

Abstract

The invention provides a GIS circuit breaker opening and closing abnormal state identification method and system based on vibration signals, and belongs to the technical field of high-voltage electrical equipment state monitoring and fault diagnosis. Comprising the following steps: collecting a vibration signal generated in a GIS tank body in a switching-on and switching-off process; the vibration sensor is arranged at a position close to the disconnecting link; fourier processing and modal analysis are carried out on the vibration signal, a discrete signal is converted into a one-dimensional time sequence vector with a finite length, time-frequency domain transformation and feature extraction processing are further carried out, and seven-dimensional features are obtained and spliced into a seven-dimensional feature vector; the seven-dimensional features comprise four frequency domain steady-state features of a unilateral power spectrum, a frequency spectrum variance, a frequency gravity center and a mean square frequency, and three time-frequency threshold transient-state features of wavelet packet sub-band energy, a sub-band energy ratio and wavelet packet time-frequency entropy; and a Bayesian optimized MLP + AdamW model is established, the seven-dimensional feature vector is input, the opening and closing faults of the GIS circuit breaker are identified, and model output results comprise opening abnormity, opening normal, closing abnormity and closing normal.

Description

Technical Field

[0001] This invention relates to the field of condition monitoring and fault diagnosis technology for high-voltage electrical equipment, and in particular to a method and system for identifying abnormal opening and closing states of GIS circuit breakers based on vibration signals. Background Technology

[0002] Gas-insulated switchgear (GIS) has been widely used in modern power systems due to its compact structure, high reliability, and low susceptibility to environmental influences. Circuit breakers, as the most important control and protection devices in GIS, directly affect the safe and stable operation of the entire power grid. During the opening and closing process, the operating mechanism and transmission components of a circuit breaker generate abundant mechanical vibration signals, which contain crucial information reflecting its mechanical state. Currently, monitoring methods for the opening and closing status of circuit breakers mainly include travel-time characteristic analysis and coil current analysis. However, these methods have some limitations: travel sensors are inconvenient to install and may affect equipment insulation; while coil current signals can reflect the condition of the electrical circuit, they are not sensitive enough to purely mechanical faults (such as loose screws or failed buffers).

[0003] Vibration signal analysis, as a non-invasive and information-rich monitoring method, is gradually gaining attention. However, existing technologies suffer from several problems: First, feature extraction is often limited, typically relying on single features in the time domain (e.g., peak value, RMS) or frequency domain (e.g., spectral peak value), making it difficult to comprehensively characterize the complex patterns of vibration signals, resulting in low accuracy and susceptibility to accidental interference. Second, it relies heavily on expert experience; feature selection and classification threshold settings depend heavily on expert experience, leading to poor generalization ability for different types of circuit breakers operating under different conditions. Third, it is insensitive to early faults; for minor mechanical anomalies, changes in features within a single domain are not significant, making early warning difficult with existing methods. Summary of the Invention

[0004] In view of this, the present invention provides a method and system for identifying abnormal opening and closing states of GIS circuit breakers based on vibration signals. It extracts multi-dimensional features of vibration signals and uses a deep learning model to achieve high-precision and robust identification of abnormal opening and closing states of GIS circuit breakers, thereby improving the sensitivity to early mechanical faults.

[0005] The technical solution adopted by the embodiments of the present invention to solve its technical problem is as follows:

[0006] A method for identifying abnormal opening and closing states of GIS circuit breakers based on vibration signals, comprising:

[0007] Step S1: Collect vibration signals generated during the opening and closing process inside the GIS tank; the vibration sensor is installed on the surface of the GIS circuit breaker, near the disconnect switch.

[0008] Step S2 involves performing Fourier processing and modal analysis on the vibration signal, transforming the discrete signal into a finite-length 1D time-series vector, and then performing time-frequency domain transformation and feature extraction on the 1D time-series vector to obtain 7D features, which are then concatenated into a 7D feature vector. The 7D features include four steady-state features in the frequency domain: one-sided power spectrum, spectral variance, frequency centroid, and mean square frequency; and three transient features in the time-frequency threshold: wavelet packet subband energy, subband energy ratio, and wavelet packet time-frequency entropy.

[0009] Step S3: Establish a Bayesian optimized MLP+AdamW model, input a 7-dimensional feature vector, identify the opening and closing faults of the GIS circuit breaker, and the model output results include opening abnormality, opening normality, closing abnormality, and closing normality; the MLP+AdamW model uses the MLP model for classification and uses AdamW as the optimizer.

[0010] Preferably, step S2 includes:

[0011] Step S21: Perform Fourier and modal analysis on the vibration signal x[n] to obtain a 1-dimensional time-series vector. : ;

[0012] In the formula, y[n] = x[n] ⋅ w[n], For window functions; ∈[0,N-1], where N is the length of the discrete-time sequence. For frequency index, For frequency domain signal processing;

[0013] according to Extracting steady-state features in the frequency domain:

[0014] Extracting the single-sided power spectrum P K : ;

[0015] In the formula, Indicates the amplitude of the frequency component. Represents a complex spectrum, P k ={ };

[0016] Extracting frequency centroid : ;

[0017] In the formula, For the first Point frequency, w k for Normalized weights;

[0018] Extracting spectral variance and mean square frequency : ; , ;

[0019] Step S22: Extract the time-frequency threshold steady-state feature signal by performing wavelet packet transform or empirical mode decomposition on the vibration signal x[n] to obtain the coefficient sequence of the wavelet packet nodes { }:

[0020] Extracting wavelet packet energy W1: ;

[0021] In the formula, For the first Layer, First The coefficient sequence of wavelet packet nodes, j∈[0,J-1], where J is the total decomposition level; For the first Node index within a layer ∈[1, ;]; Denotes the discrete energy of the subband, W1={ };

[0022] Extracted subband energy percentage W2: ;

[0023] In the formula, For the first The proportion of subband energy, ,∑ u,v E u,v W2 is the sum of the subband energy and the total subband energy; };

[0024] Extracting wavelet packet time-frequency entropy : ;

[0025] Step S23: Concatenate the frequency domain steady-state features and the time-frequency threshold transient features into a 7-dimensional feature vector f=[ , , , W1, W2, Furthermore, f is standardized as a whole.

[0026] Preferably, step S3 includes:

[0027] Initial data preparation stage: Form a matrix X∈R from N 7-dimensional eigenvectors f.N×7 It is divided into a training set and a validation set;

[0028] MLP model construction: The input layer takes a 7-dimensional matrix as input. Each hidden layer uses a linear layer, a one-dimensional batch normalization layer, an activation function, and a random deactivation layer, Dropout, in sequence for data processing. The output layer is a linear layer. The number of hidden units in the last layer is mapped to K categories and normalized to the category probabilities using Softmax during inference.

[0029] Optimizer selection: AdamW optimizer, and remove the bias parameter and normalization layer parameters from the weight decay regularization term; search for learning rate, weight decay coefficient, AdamW's first and second momentum coefficients betas, and batch size;

[0030] Bayesian optimization: The goal is to maximize the fold mean of cross-validation Macro-F1. The tree-structured Parzen estimation method first explores N1 times in a narrow space and then expands to a larger search range. Data preprocessing and model training are completed within each complete training and evaluation trial of each hyperparameter configuration. The pruner is enabled to terminate poorly performing trials early, and the training round with the best hyperparameter participation index is recorded.

[0031] Iteration: Repeat the selection, evaluation, and update process until the maximum number of iterations is reached, or stop early when the improvement value of the validation set metric does not exceed the preset threshold for M consecutive rounds;

[0032] Model output: For the four categories of normal closing, normal opening, abnormal closing, and abnormal opening, the category with the highest Softmax probability is directly selected.

[0033] A vibration signal-based GIS circuit breaker opening and closing status identification system, used to implement the aforementioned vibration signal-based GIS circuit breaker opening and closing abnormal status identification method, includes:

[0034] The acquisition module is used to collect vibration signals generated by the opening and closing of the gate inside the GIS tank;

[0035] The feature extraction and feature vector construction module is used to perform Fourier processing and modal analysis on vibration signals, transforming discrete signals into finite-length 1D time-series vectors, performing time-frequency domain transformation and feature extraction on the 1D time-series vectors to obtain 7D features, which are then concatenated into a 7D feature vector. The 7D features include four steady-state features in the frequency domain: one-sided power spectrum, spectral variance, frequency centroid, and mean square frequency; and three transient features in the time-frequency threshold: wavelet packet subband energy, subband energy ratio, and wavelet packet time-frequency entropy.

[0036] The model building module is used to build a Bayesian optimized MLP+AdamW model;

[0037] The fault identification module is used to identify the opening and closing faults of the GIS circuit breaker based on the 7-dimensional feature vector using the Bayesian optimized MLP+AdamW model. The fault types include abnormal opening, normal opening, abnormal closing, and normal closing.

[0038] As can be seen from the above technical solution, the method for identifying abnormal opening and closing states of GIS circuit breakers based on vibration signals provided in this embodiment of the invention collects vibration signals generated during the opening and closing process inside the GIS tank; performs Fourier processing and modal analysis on the vibration signals, transforming the discrete signals into finite-length 1D time-series vectors; performs time-frequency domain transformation and feature extraction on the 1D time-series vectors to obtain 7-dimensional features, which are then concatenated into a 7-dimensional feature vector; the 7-dimensional features include four steady-state features in the frequency domain: one-sided power spectrum, spectral variance, frequency centroid, and mean square frequency; and three transient features in the time-frequency threshold: wavelet packet subband energy, subband energy ratio, and wavelet packet time-frequency entropy; a Bayesian-optimized MLP+AdamW model is established, and the 7-dimensional feature vector is input to identify the opening and closing faults of the GIS circuit breaker. The model output results include opening abnormality, opening normality, closing abnormality, and closing normality. This invention, by extracting multi-dimensional features of vibration signals and utilizing a deep learning model, achieves high-precision and robust identification of abnormal opening and closing states of GIS circuit breakers, which can improve the sensitivity to early mechanical faults. Attached Figure Description

[0039] Figure 1 This is a flowchart of the method for identifying abnormal opening and closing states of GIS circuit breakers based on vibration signals according to the present invention.

[0040] Figure 2 This is a diagram illustrating the transformation of data from one dimension to seven dimensions in this invention.

[0041] Figure 3 This is a flowchart of the model training process of the present invention. Detailed Implementation

[0042] The technical solution and effects of the present invention will be further described in detail below with reference to the accompanying drawings.

[0043] refer to Figure 1 As shown, this invention provides a method for identifying abnormal opening and closing states of GIS circuit breakers based on vibration signals, the implementation steps of which include:

[0044] A method for identifying abnormal opening and closing states of GIS circuit breakers based on vibration signals, comprising:

[0045] Step S1: Collect vibration signals generated during the opening and closing process inside the GIS tank; the vibration sensor is installed on the surface of the GIS circuit breaker, near the disconnect switch.

[0046] Step S2 involves performing Fourier processing and modal analysis on the vibration signal, transforming the discrete signal into a finite-length 1D time-series vector, and then performing time-frequency domain transformation and feature extraction on the 1D time-series vector to obtain 7D features, which are then concatenated into a 7D feature vector. The 7D features include four steady-state features in the frequency domain: one-sided power spectrum, spectral variance, frequency centroid, and mean square frequency; and three transient features in the time-frequency threshold: wavelet packet subband energy, subband energy ratio, and wavelet packet time-frequency entropy.

[0047] Step S3: Establish a Bayesian optimized MLP+AdamW model, input a 7-dimensional feature vector, identify the opening and closing faults of the GIS circuit breaker, and the model output results include opening abnormality, opening normality, closing abnormality, and closing normality; the MLP+AdamW model uses the MLP model for classification and uses AdamW as the optimizer.

[0048] Step S2 is used to extract time-domain and time-frequency domain features for analyzing the fault state of GIS circuit breakers based on vibration signals. The specific implementation includes:

[0049] Step S21: The vibration signal x[n] is a discrete signal. Fourier and modal analysis of the vibration signal x[n] is performed to obtain a 1-dimensional time-series vector. : ;

[0050] In the formula, y[n] = x[n] ⋅ w[n], Window functions (such as the Hanning window) are used to reduce spectral leakage; ∈[0,N-1], where N is the length of the discrete-time sequence. For frequency index, For frequency domain signal processing;

[0051] according to Extracting steady-state features in the frequency domain:

[0052] Extracting the single-sided power spectrum P K : ;

[0053] In the formula, Indicates the amplitude (i.e., strength) of the frequency component. Represents a complex spectrum, P k ={ Argument represents phase, and argument and phase can be converted into each other.

[0054] The two-sided spectrum is merged into the positive frequency side, therefore Need to be multiplied by 2; DC Do not multiply by 2, divide by It is a common normalization, making Approaching the dimensions of "energy density" (consistent with the square of the signal amplitude), this step transforms the complex spectrum... Transforming it into a distribution of "energy / power as a function of frequency" facilitates subsequent feature calculations;

[0055] Extracting frequency centroid (Hz): ;

[0056] In the formula, For the first Point frequency, w k for Normalized weights; frequency centroid This indicates the frequency at which energy falls on average;

[0057] Extracting spectral variance and mean square frequency : ;

[0058] It is the second-order central moment of the frequency distribution relative to the centroid; For the single-sided power spectrum at the 1st The value of the point; broadband faults (loosening, friction, cavitation, severe bearing damage) will cause The larger the value, the "fatter" the spectrum. If the center of gravity remains unchanged but the variance increases, it indicates that the noise / sidebands on both sides have become stronger. ;

[0059] Mean square frequency As an integrated indicator of "spectral energy level × frequency weight", used for health trends, the mean square frequency is related to the variance / centroid as follows: .

[0060] Step S22: Extract the time-frequency threshold steady-state feature signal by performing wavelet packet transform or empirical mode decomposition on the vibration signal x[n] to obtain the coefficient sequence of the wavelet packet nodes { }:

[0061] Extract the wavelet packet band energy W1, decompose the signal into fine-grained frequency bands, and quantize the energy of each band. : ;

[0062] In the formula, For the first Layer, First The coefficient sequence of wavelet packet nodes, j∈[0,J-1], where J is the total decomposition level (starting from 0 or 1, a non-negative integer). The deeper the level, the narrower the bandwidth. For the first Node index within a layer ∈[1, ;]; Denotes the discrete energy of the subband, W1={ };

[0063] Extract the sub-band energy percentage W2, standardize the "energy of each band", and make it suitable for comparison between different loads / velocities: ;

[0064] In the formula, For the first The proportion of subband energy (dimensionless) ), ,∑ u,v E u,v W2 is the sum of the subband energy and the total subband energy; };

[0065] Extracting wavelet packet time-frequency entropy : ;

[0066] L WPT Entropy (dimensionless) based on subband energy distribution;

[0067] Step S23: Concatenate the frequency domain steady-state features and the time-frequency threshold transient features into a 7-dimensional feature vector f=[ , , , W1, W2, Furthermore, f is standardized as a whole.

[0068] Step S3 provides the establishment of an MLP (Multilayer Perceptron) model based on AdamW (Decoupled Weight Decay) optimization and Bayesian optimization. In the MLP+AdamW model, MLP (Multilayer Perceptron) + AdamW (Adam with decoupled weight decay) is a deep learning classification framework that learns nonlinear discriminant functions in an end-to-end manner. The key to MLP is to use multilayer fully connected layers and nonlinear activations (such as ReLU, Rectified LinearUnit) to represent the input features layer by layer, thereby learning a flexible nonlinear decision boundary. Structurally, tabular / engineering features are usually selected as "shallow and moderately wide" two or three-layer networks. If necessary, BatchNorm (Batch Normalization) is added to stabilize the distribution, and Dropout is added to suppress co-adaptation. Deeper networks are equipped with residuals and stronger regularization to prevent overfitting. MLP consists of multilayer fully connected layers and nonlinear activations. By minimizing the cross-entropy loss, it learns the decision hyperplane (or more accurately, the nonlinear decision boundary) that separates each category in the feature space. Unlike the traditional Adam, AdamW decouples weight decay from gradient updates: it directly applies a shrinkage term to the weights with each parameter update, so that the regularization strength is not affected by the adaptive learning rate scaling. This improves generalization ability and reduces the risk of overfitting while maintaining the same training stability. It is particularly suitable for industrial scenarios with limited sample size or high feature noise (such as circuit breaker status recognition).

[0069] The Bayesian optimization model uses the Macro-F1 score (a metric for overall performance in multi-class classification tasks) on the validation set as a black-box objective and employs Bayesian optimization to automatically search for hyperparameters. The search space covers: learning rate, weight decay, hidden layer width and number of layers, Dropout ratio, batch size, and learning rate scheduling parameters. Searches are performed separately for each of the seven features to avoid bias in the optimal solution due to differences in scale. The model building process is as follows: Figure 3 As shown:

[0070] Initial data preparation stage: The data input is 7-dimensional X∈R n×7 A matrix consisting of N samples, each sample being a 7-dimensional vector, with 90% of the data used as the training set and 10% as the test set;

[0071] Initialization: Using the 7-dimensional feature vector of the data as input, first randomly sample from the preset hyperparameter space (number / width of hidden layers, activation function, dropout, learning rate, weight decay, batch size, etc.). Group Configuration For each K-fold hierarchical cross-validation is performed: within each fold, only the training fold is used to fit the StandardScaler and any feature transformations, and the results are evaluated on the validation fold; AdamW is used for training (fixed random seed), and early stopping is enabled (based on validation set metrics / loss) to reduce variance and overfitting. The input layer is 7-dimensional, and the hidden layers are searched in layers 1–3. Each layer uses Linear → BatchNorm1d (optional one-dimensional batch normalization layer) → Activation → Dropout. The output layer is Linear(last,) (mapping the number of hidden units in the last layer to K classes) and uses Softmax (normalized to class probabilities) during inference. The optimizer is AdamW (which has the characteristic of decoupling weight decay), and the bias (bias parameter) and normalization layer parameters (scale and offset parameters of BatchNorm / LayerNorm) are removed from the weight decay (weight decay regularization term); the learning rate, weight_decay (weight decay coefficient), betas (first and second momentum coefficients of AdamW), and batch_size are searched.

[0072] Constructing a surrogate model: A Gaussian Process (GP) is used to model the black-box relationship between hyperparameters and Macro-F1 scores, explicitly considering evaluation noise (from random initialization and small samples). The space includes conditional dependencies (e.g., different layer width variables corresponding to 1 / 2 / 3 layers); for quantities with large spans (learning rate, weight decay), a logarithmic scale is used to define the search interval.

[0073] The acquisition function selects the next sampling point: on the current agent model, the acquisition function balances "exploring new regions" and "utilizing high-potential regions" to provide the next set of hyperparameters to be evaluated; mini-batch parallelism or asynchronous strategies are adopted to improve computing power utilization; at the same time, soft constraints on runtime and memory are applied to avoid unrealistic large model configurations.

[0074] Evaluation of the objective function: Train MLP+AdamW with selected hyperparameters and calculate the folded Macro-F1 during validation. The training pipeline is uniform and reproducible: only training folds are used to fit the StandardScaler; AdamW employs parameter grouping (without applying weight decay to bias and normalization layers); learning rate scheduling (OneCycleLR or Cosine), gradient clipping, and early stopping are enabled; class weights / minority class lightweight oversampling are added as needed to alleviate imbalance.

[0075] Update the surrogate model (Bayesian optimization): Add the "new hyperparameters + folded mean Macro-F1" as new observations to the sampling set D, and refit the surrogate model using a Gaussian process (or TPE) within the Bayesian optimization framework, updating the posterior to more accurately approximate the mapping of "hyperparameters → indicators"; then solve on the acquisition function (such as expectation improvement, upper confidence bound, or probability improvement), taking into account exploration – using trade-offs to propose the next set of candidate hyperparameters; use logarithmic scaling to model variables with large spans (such as learning rate, weight decay), and explicitly consider evaluation noise and conditional dependencies (such as the availability of layer width determined by the number of layers); to improve efficiency, maximize the folded mean of cross-validation Macro-F1, using the tree structure Parzen estimation method, first perform 50–80 explorations in a narrow space (first roughly search in a small range to find a roughly effective interval), and then expand to a larger search range. Within each trial (a complete training and evaluation experiment with hyperparameter configuration), data preprocessing and model training are completed. The pruner is enabled to terminate poorly performing trials early, and the training round with the best hyperparticipation index is recorded.

[0076] Iteration: Repeat "select point → evaluate → update" until the maximum number of iterations / time budget is reached, or stop early if there is no significant improvement after several rounds. Among them, the improvement of the validation set metric (such as loss or accuracy) relative to the historical best value exceeds a preset threshold (loss decrease of at least 0.001). If this minimum improvement is not reached for several consecutive rounds (such as 5 rounds), it is considered as no significant improvement and triggers early termination.

[0077] Output Results and Finalization: After the algorithm stops, output the optimal hyperparameters and corresponding validation metrics; retrain the final model completely on this optimal configuration using the "training set + validation set", and then report Macro-F1, Balanced Accuracy, Precision / Recall, confusion matrix, and probability calibration metrics on an independent test set; complete the threshold and hysteresis strategy settings according to engineering requirements; finally, export a deployable pipeline (normalizer + model + calibrator + threshold), and save the optimal hyperparameters, random seed, and environment information in a versioned manner. For the four categories of normal closing, normal opening, abnormal closing, and abnormal opening, directly select the category with the highest probability using Softmax (a function that transforms "any real number vector" into a "probability distribution"). The evaluation metric is Macro-F1, that is, calculate the F1 score for each of the four categories (normal closing, normal opening, abnormal closing, and abnormal opening), and then take the arithmetic mean of the four F1 scores to ensure that each category (especially the categories with potentially fewer samples such as abnormal closing and abnormal opening) is treated fairly.

[0078] The following is a specific embodiment to illustrate the method of the present invention:

[0079] First, an event window is extracted around the opening / closing trigger point, and basic preprocessing such as mean removal and bandpass filtering is performed. Then, the time-domain waveform is squared, divided into blocks of fixed length (e.g., every 1000 points), and the energy is calculated. The block energy sequence is then normalized to reduce continuous noise unrelated to the applied voltage and highlight transient energy related to the action. Based on the normalized signal, two parallel processing lines are used: the frequency domain line extracts four features—spectral centroid (overall energy bias), bandwidth / spreadability index (spectral dispersion, selecting either frequency variance or root mean square frequency), low / mid frequency band energy proportion (covering the inherent frequency of the mechanism and steady-state characteristics such as meshing / harmonics), and high-frequency impact band energy proportion (transient activities such as rebound / impact). All selected bands are determined at once on the training set through small grid search, and FFT is uniformly windowed (e.g., Hann), with fixed normalization and calculation order to ensure comparability across batches and devices.

[0080] Three complementary features are extracted from the time-frequency line: the energy proportion of the wavelet packet target sub-band (selecting the most sensitive impact sub-band based on spectral kurtosis / discrimination on the training set), the energy entropy of the wavelet packet (energy dispersion in each sub-band, normalized to 0–1), and the energy proportion of the first two IMFs of EMD (high-frequency transient intensity). Finally, F1–F4 and W1–W3 are concatenated into a 7-dimensional vector in a fixed order, standardized using training set statistics, and applied only once for validation / testing and online testing without further fitting. Simultaneously, the band selection boundary, WPT / EMD parameters, and version number are written into the configuration for easy traceability and reproduction. This 7-dimensional combination provides stable discrimination capability with low computational overhead through the complementarity of "frequency domain steady state + time-frequency transient," and is used for opening / closing fault classification and online early warning.

[0081] The signal acquired by the vibration sensor undergoes Fourier transform and time-frequency domain processing. The data, initially in the form f=[1.82,0.03,0.21,0.04,3.1,0.2,1.7,5.4,2.3,1.8,0.012,0.004,…], is transformed into a 7-dimensional feature vector f=[F1, F2, F3, F4, W1, W2, W3] containing 4 frequency domains and 3 time-frequency domains. For example… Figure 2 The diagram shows how 1D data can be transformed into a 7D data structure.

[0082] Comparison between traditional and optimized methods:

[0083] To verify the MLP model based on AdamW optimization and Bayesian optimization proposed in this invention, the invention uses real-time collected opening and closing discharge signals from GIS signal sources as input samples, constructs source domain data as input samples, and selects the recognition rates of the following models for comparison: traditional MLP + Adam, traditional MLP + Adam (Bayesian optimization), MLP + AdamW, and MLP + AdamW (Bayesian optimization).

[0084] As shown in the table, the baseline MLP (Multilayer Perceptron) + Adam model achieved recognition rates of 89.2% and 88.8% for "abnormal closing / abnormal opening," respectively. Introducing Bayesian optimization further improved these rates to 91.5% and 92.1% (+2.3 / +3.3 percentage points), indicating that hyperparameter search alone could provide stable gains. Replacing the optimizer with AdamW (Adam with decoupled weight decay) increased the rates to 94.8% and 95.4% even without Bayesian optimization (+5.6 / +6.6 percentage points relative to the baseline, and an additional +3.3 / +3.3 percentage points compared to "Adam + Bayesian optimization"). The Bayesian-optimized MLP + AdamW model achieved recognition rates of 97.6% and 96.9%, a significant improvement over traditional models.

[0085] Furthermore, this invention provides a GIS circuit breaker opening and closing status identification system based on vibration signals, used for implementation Figure 1 The vibration signal-based method for identifying abnormal opening and closing states of GIS circuit breakers includes:

[0086] The acquisition module is used to collect vibration signals generated by the opening and closing of the gate inside the GIS tank;

[0087] The feature extraction and feature vector construction module is used to perform Fourier processing and modal analysis on vibration signals, transforming discrete signals into finite-length 1D time-series vectors, performing time-frequency domain transformation and feature extraction on the 1D time-series vectors to obtain 7D features, which are then concatenated into a 7D feature vector. The 7D features include four steady-state features in the frequency domain: one-sided power spectrum, spectral variance, frequency centroid, and mean square frequency; and three transient features in the time-frequency threshold: wavelet packet subband energy, subband energy ratio, and wavelet packet time-frequency entropy.

[0088] The model building module is used to build a Bayesian optimized MLP+AdamW model;

[0089] The fault identification module is used to identify the opening and closing faults of the GIS circuit breaker based on the 7-dimensional feature vector using the Bayesian optimized MLP+AdamW model. The fault types include abnormal opening, normal opening, abnormal closing, and normal closing.

[0090] The present invention has the following advantages:

[0091] This study employs multidimensional features to analyze whether faults exist in the opening and closing of GIS circuit breakers. The 7-dimensional features integrate information from the time domain, frequency domain, and statistics, significantly improving the ability to represent subtle anomalies in opening and closing compared to single thresholds / single features. This allows normal and abnormal patterns to form separable differences in the high-dimensional feature space. The MLP can learn complex nonlinear decision boundaries, eliminating the need for experts to set multidimensional thresholds. It automatically finds the discriminant surface that distinguishes normal / abnormal opening and closing from the data. At the same time, AdamW is used for decoupling weight decay optimization, which effectively suppresses overfitting while ensuring convergence efficiency with adaptive learning rate. This allows the model to maintain good generalization performance under different operating conditions and noise levels. It can also effectively identify circuit breakers with only minor mechanical anomalies. Therefore, it can overcome the problem of insufficient sensitivity to mechanical faults, effectively improve the identification accuracy, reduce the possibility of accidental interference, and enhance the sensitivity to early mechanical faults.

[0092] The above-disclosed embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of the invention. Those skilled in the art will understand that implementing all or part of the above-described embodiments and making equivalent changes in accordance with the claims of the present invention are still within the scope of the invention.

Claims

1. A method for identifying abnormal opening and closing states of GIS circuit breakers based on vibration signals, characterized in that, include: Step S1: Collect vibration signals generated during the opening and closing process inside the GIS tank; The vibration sensor is installed on the surface of the GIS circuit breaker, near the disconnect switch. Step S2 involves performing Fourier processing and modal analysis on the vibration signal, transforming the discrete signal into a finite-length 1D time-series vector, and then performing time-frequency domain transformation and feature extraction on the 1D time-series vector to obtain 7D features, which are then concatenated into a 7D feature vector. The 7D features include four steady-state features in the frequency domain: one-sided power spectrum, spectral variance, frequency centroid, and mean square frequency; and three transient features in the time-frequency threshold: wavelet packet subband energy, subband energy ratio, and wavelet packet time-frequency entropy. Step S3: Establish a Bayesian optimized MLP+AdamW model, input a 7-dimensional feature vector, identify the opening and closing faults of the GIS circuit breaker, and the model output results include opening abnormality, opening normality, closing abnormality, and closing normality; the MLP+AdamW model uses the MLP model for classification and uses AdamW as the optimizer.

2. The method for identifying abnormal opening and closing states of GIS circuit breakers based on vibration signals as described in claim 1, characterized in that, Step S2 includes: Step 21: Perform Fourier and modal analysis on the vibration signal x[n] to obtain a 1-dimensional time-series vector. : ； In the formula, y[n] = x[n] ⋅ w[n], For window functions; ∈[0,N-1], where N is the length of the discrete-time sequence. For frequency index, For frequency domain signal processing; according to Extracting steady-state features in the frequency domain: Extracting the single-sided power spectrum P K : ； In the formula, Indicates the amplitude of the frequency component. Represents a complex spectrum, P k ={ }; Extracting frequency centroid : ； In the formula, For the first Point frequency, w k for Normalized weights; Extracting spectral variance and mean square frequency : ； ， ； Step 22, extract the time-frequency threshold steady-state feature signal, and perform wavelet packet transform or empirical mode decomposition on the vibration signal x[n] to obtain the coefficient sequence of the wavelet packet nodes { }: Extracting wavelet packet energy W1: ； In the formula, For the first Layer, First The coefficient sequence of wavelet packet nodes, j∈[0,J-1], where J is the total decomposition level; For the first Node index within a layer ∈[1, ;]; Denotes the discrete energy of the subband, W1={ }; Extracted subband energy percentage W2: ； In the formula, For the first The proportion of subband energy, ,∑ u,v E u,v W2 is the sum of the subband energy and the total subband energy; }; Extracting wavelet packet time-frequency entropy : ； Step 23: Concatenate the frequency domain steady-state features and the time-frequency threshold transient features into a 7-dimensional feature vector f=[ , , , W1, W2, Furthermore, f is standardized as a whole.

3. The method for identifying abnormal opening and closing states of GIS circuit breakers based on vibration signals as described in claim 2, characterized in that, Step S3 includes: Initial data preparation stage: Form a matrix X∈RN from N 7-dimensional eigenvectors f. n×7 It is divided into a training set and a validation set; MLP model construction: The input layer takes a 7-dimensional matrix X as input. Each hidden layer uses a linear layer, a one-dimensional batch normalization layer, an activation function, and a random deactivation layer, Dropout, in sequence for data processing. The output layer is a linear layer. The number of hidden units in the last layer is mapped to K categories and normalized to the category probabilities using Softmax during inference. Optimizer selection: AdamW optimizer, and remove the bias parameter and normalization layer parameters from the weight decay regularization term; search for learning rate, weight decay coefficient, AdamW's first and second momentum coefficients betas, and batch size; Bayesian optimization: The goal is to maximize the fold mean of cross-validation Macro-F1. The tree-structured Parzen estimation method first explores N1 times in a narrow space and then expands to a larger search range. Data preprocessing and model training are completed within each complete training and evaluation trial of each hyperparameter configuration. The pruner is enabled to terminate poorly performing trials early, and the training round with the best hyperparameter participation index is recorded. Iteration: Repeat the selection, evaluation, and update process until the maximum number of iterations is reached, or stop early when the improvement value of the validation set metric does not exceed the preset threshold for M consecutive rounds; Model output: For the four categories of normal closing, normal opening, abnormal closing, and abnormal opening, the category with the highest Softmax probability is directly selected.

4. A GIS circuit breaker opening and closing status identification system based on vibration signals, characterized in that, For implementing the method according to any one of claims 1-3, comprising: The acquisition module is used to collect vibration signals generated by the opening and closing of the gate inside the GIS tank; The feature extraction and feature vector construction module is used to perform Fourier processing and modal analysis on vibration signals, transforming discrete signals into finite-length 1D time-series vectors, performing time-frequency domain transformation and feature extraction on the 1D time-series vectors to obtain 7D features, which are then concatenated into a 7D feature vector. The 7D features include four steady-state features in the frequency domain: one-sided power spectrum, spectral variance, frequency centroid, and mean square frequency; and three transient features in the time-frequency threshold: wavelet packet subband energy, subband energy ratio, and wavelet packet time-frequency entropy. The model building module is used to build a Bayesian optimized MLP+AdamW model; The fault identification module is used to identify the opening and closing faults of the GIS circuit breaker based on the 7-dimensional feature vector using the Bayesian optimized MLP+AdamW model. The fault types include abnormal opening, normal opening, abnormal closing, and normal closing.

Images