A method and system for evaluating the surface condition of vacuum interrupter contacts

Abstract

The present invention discloses a method and system for evaluating the surface condition of contacts of a vacuum interrupter, which belongs to the technical field of vacuum circuit breakers. The method uses adaptive window length to improve short-time Fourier transform for noise reduction, combines dual-tree complex wavelet transform to enhance signals, and dynamically normalizes data through improved K-means clustering; a lightweight CRNN model is constructed, the convolution module introduces residual connection and SE module, and the recurrent module uses a bidirectional quasi-recurrent neural network to capture time series features; the hybrid loss function is combined with the AdamW optimizer to improve the convergence stability through L2 regularization; and the degree of contact ablation is dynamically evaluated based on a fuzzy logic algorithm. The method achieves high-precision classification of DC pre-breakdown waveforms through time-frequency feature fusion and lightweight network design, solves the problems of traditional model degradation, overfitting and slow convergence, and provides an intelligent solution for the insulation performance evaluation of vacuum interrupters.

Description

Technical Field

[0001] The present invention belongs to the technical field of vacuum circuit breakers and relates to a method and system for evaluating the surface condition of contacts of a vacuum interrupter. Background Art

[0002] Power systems often require capacitor bank switching for reactive power compensation. Vacuum circuit breakers (VCBs) or SF6 circuit breakers are the primary devices used in these applications. Vacuum circuit breakers are widely used in capacitor bank applications due to their high reliability, low maintenance requirements, frequent operation, and environmental friendliness. However, compared to SF6 circuit breakers, VCBs face a higher risk of re-breakdown in capacitor bank applications. Re-breakdown can trigger overvoltages, damaging not only the equipment but also the safe operation of the entire power system. Pre-breakdown occurs during the closing of capacitor banks, often accompanied by varying degrees of inrush current. This inrush current can cause erosion on the contact surfaces of the vacuum interrupter, a key factor contributing to re-breakdown. To minimize contact damage from inrush current, modern technologies employ phase-controlled closing to control the vacuum interrupter. Research has found that the surface condition of the vacuum interrupter contacts significantly affects the performance of phase-controlled closing. Therefore, accurate assessment of the contact surface condition of vacuum interrupters is of great practical significance for ensuring the safe and stable operation of power systems.

[0003] Previous methods for measuring contact surface erosion levels involved invasive methods, which destroyed the interrupter and were only suitable for macroscopic investigations. Non-invasive methods, on the other hand, are time-consuming, labor-intensive, and complex to operate, potentially affecting the mechanical properties of the circuit breaker. Pre-breakdown in a vacuum interrupter typically produces different types of pre-breakdown waveforms, which can, to a certain extent, reflect the contact surface condition. Typically, based on the voltage across the interrupter, field emission current, and contact travel waveform during the closing process, pre-breakdown types can be categorized into three types: field emission pre-breakdown (FEPS), particle pre-breakdown (PPS), and field emission-particle pre-breakdown (FE-PPS). Traditional identification and detection methods primarily rely on manual feature extraction and threshold determination. These methods require setting threshold ranges for waveform characteristic parameters for different breakdown types, calculating the characteristic parameters of the waveform to be classified, and comparing these parameters with the thresholds to determine the waveform's category. However, the characteristic parameters rely on empirical settings, lack universality, and are difficult to process complex waveforms. The classification accuracy is limited, and it cannot adapt to changes in device status and lacks real-time performance.

[0004] With the development of artificial intelligence, image detection and recognition technology is becoming increasingly intelligent, and waveform classification using the pattern recognition capabilities of machine learning is becoming increasingly common. Decision tree algorithms construct decision tree models based on waveform features. By comparing eigenvalues ​​and making decisions step by step according to a tree-like structure, waveforms are classified into different branches. However, they are prone to overfitting and strong feature dependence. The random forest algorithm, consisting of multiple decision trees, is trained on large amounts of waveform data and combines the classification results of multiple decision trees. However, its computational complexity is high. While it can effectively improve classification accuracy and stability, the explanation of the classification reasons is unclear. Support vector machines, on the other hand, separate waveforms of different breakdown types in feature space by finding the optimal classification hyperplane. While suitable for small sample and nonlinear waveform classification problems, they are sensitive to the choice of kernel parameters and are not suitable for large-scale data. Summary of the Invention

[0005] In response to the problems existing in the prior art, the present invention provides a method and system for evaluating the surface condition of vacuum interrupter contacts, thereby solving the technical problems that the machine learning algorithm in the prior art cannot take into account both high-precision classification and dynamic adaptation requirements in the process of evaluating contact ablation, and the traditional model is prone to degradation, overfitting, and slow convergence.

[0006] The present invention is achieved through the following technical solutions:

[0007] A method for evaluating the surface condition of a vacuum interrupter contact comprises the following steps:

[0008] S1: Acquire the field emission current signal, arc chamber voltage signal and contact stroke curve signal during the pre-breakdown process of the vacuum interrupter;

[0009] S2: Denoising the acquired field emission current signal, arc extinguishing chamber voltage signal, and contact travel curve signal to obtain a denoised signal; decomposing and reconstructing the denoised signal by dual-tree complex wavelet transform to obtain an enhanced multi-channel time-frequency signal; and dynamically normalizing the enhanced multi-channel time-frequency signal by using an improved K-means clustering method to obtain a preprocessed signal;

[0010] S3: Evaluate the surface condition of the vacuum interrupter contact based on the preprocessed signal and a lightweight convolutional recurrent neural network model;

[0011] The lightweight convolutional recurrent neural network model is pre-constructed through a customized lightweight gating structure and a bidirectional quasi-recurrent neural network; the lightweight convolutional recurrent neural network model includes a three-layer convolution module, wherein the first layer sets the void rate to 2, the second layer introduces residual connection, and the third layer adds an attention mechanism SE module; the lightweight convolutional recurrent neural network model is obtained by model training through a hybrid loss function and an AdamW optimizer, L2 regularization is performed during training, and a multi-dimensional dynamic risk index and an environmental adaptation factor are introduced.

[0012] Preferably, in the step of performing noise reduction on the acquired field emission current signal, arc extinguishing chamber voltage signal, and contact stroke curve signal, an improved short-time Fourier transform with an adaptive window length is specifically used to perform noise reduction on the field emission current signal, arc extinguishing chamber voltage signal, and contact stroke curve signal, wherein the adaptive window length is dynamically adjusted according to the kurtosis coefficient, specifically as follows:

[0013]

[0014] in, is the adaptive window length, is the critical frequency, is the sampling frequency of the signal; is the kurtosis coefficient, , is the fourth-order central moment of the current frame, is the standard deviation of the current signal frame data, represents the minimum function, Represents the maximum value function.

[0015] Preferably, the de-noised signal is decomposed and reconstructed by dual-tree complex wavelet transform to obtain an enhanced multi-channel time-frequency signal. Specifically, during the decomposition and reconstruction process, the threshold is dynamically adjusted according to the frequency band energy, specifically as follows:

[0016]

[0017] in, For the Dynamic threshold of layer subband, For the Noise estimation of the layer, represents the number of sample points, For the The energy proportion of the sub-band is is the total energy of the signal, Expressed as a natural constant The natural logarithm of base .

[0018] Preferably, the enhanced multi-channel time-frequency signal is dynamically normalized using an improved K-means clustering method, specifically:

[0019]

[0020] in, is the normalized signal value, is the original input data, i Indicates the i A signal, is the signal mean of the current pre-breakdown type, is the standard deviation of the corresponding pre-breakdown type, is the scaling factor, is the offset coefficient.

[0021] Preferably, the lightweight convolutional recurrent neural network model is pre-constructed by a custom lightweight gating structure and a bidirectional quasi-recurrent neural network. Specifically, when pre-constructed, the number of hidden layer neurons is 256, and the update gate formula is:

[0022]

[0023]

[0024]

[0025]

[0026] in, is a candidate hidden state, is the weight matrix of the candidate hidden state, represents the convolution operator, is the current time step and its adjacent input sequence, is the current iteration time step, is the convolution window size, is the bias term of the candidate hidden state, represents the hyperbolic tangent function;

[0027] To update the output parameters of the gate, is the Sigmoid function, To update the gate weight, is the bias term of the update gate;

[0028] is the final hidden state, is the final hidden state at the previous moment, represents element-wise multiplication;

[0029] is the attention weight matrix, is the normalized exponential function, are the query matrix and the key matrix respectively, is the matrix transpose operation, For the dimension, Represents two positions in a sequence a , b The distance between them.

[0030] Preferably, the hybrid loss function is specifically:

[0031]

[0032]

[0033] in, Represents the overall loss function value, represents the cross entropy loss, represents the maximum mean difference loss, represents the learnable parameter matrix of the model, represents the 2-norm of W;

[0034] Represents the source domain sample characteristics, represents the sample characteristics of the target domain, and Represents the number of samples in the source domain and the target domain respectively, is the kernel function mapping, is the reproducing kernel Hilbert space.

[0035] Preferably, when the AdamW optimizer is used for model training, the learning rate adopts a triangular cycle strategy, specifically:

[0036]

[0037] in, is the learning rate at the current time, is the maximum learning rate, is the minimum learning rate; is the number of steps in a half cycle, is the current iteration time step, Indicates taking the cosine value.

[0038] Preferably, the surface condition assessment of the vacuum interrupter contact is performed based on the preprocessed signal and the lightweight convolutional recurrent neural network model, specifically:

[0039] Based on the fuzzy logic algorithm, the Gaussian membership function is used to define the membership of the input variables. The formula is:

[0040]

[0041] in, is the degree of membership; is the central value of the membership function, To control the standard deviation of the membership function shape, is the probability value of each type of pre-breakdown, is a natural constant;

[0042] Calculating rule activation strength , and merge the intercepted fuzzy sets to output the centroid value ,according to The interval determines the degree of contact surface ablation, specifically:

[0043]

[0044] in, Indicates the output center of gravity value, represents the discretization sampling points of the output fuzzy set, Score the output, For affiliation , ∑ represents summation.

[0045] A vacuum interrupter contact surface condition evaluation system based on an improved CRNN, comprising:

[0046] A signal acquisition module is used to obtain the field emission current signal, the arc extinguishing chamber voltage signal and the contact stroke curve signal during the pre-breakdown process of the vacuum interrupter;

[0047] a signal processing module for performing noise reduction on the acquired field emission current signal, arc extinguishing chamber voltage signal, and contact travel curve signal to obtain a noise-reduced signal; decomposing and reconstructing the noise-reduced signal by dual-tree complex wavelet transform to obtain an enhanced multi-channel time-frequency signal; and dynamically normalizing the enhanced multi-channel time-frequency signal by using an improved K-means clustering method to obtain a preprocessed signal;

[0048] The signal evaluation module evaluates the surface condition of the vacuum interrupter contacts based on the preprocessed signals and a lightweight convolutional recurrent neural network model;

[0049] The lightweight convolutional recurrent neural network model is pre-constructed through a customized lightweight gating structure and a bidirectional quasi-recurrent neural network; the lightweight convolutional recurrent neural network model includes a three-layer convolution module, wherein the first layer sets the void rate to 2, the second layer introduces residual connection, and the third layer adds an attention mechanism SE module; the lightweight convolutional recurrent neural network model is obtained by model training through a hybrid loss function and an AdamW optimizer, L2 regularization is performed during training, and a multi-dimensional dynamic risk index and an environmental adaptation factor are introduced.

[0050] A computer system includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above method.

[0051] Compared with the prior art, the present invention has the following beneficial technical effects:

[0052] The present invention proposes a method for evaluating the surface condition of vacuum interrupter contacts, which can intelligently identify the DC pre-breakdown waveform of the vacuum interrupter, and assist in realizing the accurate evaluation of the surface condition of the vacuum interrupter contacts. First, an improved short-time Fourier transform is used to remove noise interference, and dual-tree complex wavelet transform signal enhancement is performed to improve data quality; the improved K-means clustering method is used for normalization operation to map the data to a specific interval, making the model training more stable. Secondly, the CRNN model is improved, and the residual connection is introduced in the convolution layer to avoid the degradation problem of the model. The addition of the SE module gives the model a unique advantage in balancing local features and global dependencies; a customized lightweight gating structure is used, and the improved GRU recurrent layer is selected. The bidirectional QRNN is used to determine the number of neurons to capture the time series characteristics of the waveform and establish long-term dependencies. Finally, the AdamW optimizer was integrated for L2 regularization, decoupling weight decay and addressing the slow convergence and overfitting issues of traditional methods. A multidimensional dynamic risk index (mDRI) was established, along with an environmentally adaptive factor. Vacuum interrupter insulation performance evaluation criteria were developed, and parameter coupling corrections were performed to improve the rationality of identification. The improved CRNN algorithm leverages both CNNs to extract spatial waveform features and RNNs to process time series information. It also addresses the model degradation and overfitting issues of traditional algorithms, achieving faster convergence and higher classification performance. This provides a more accurate solution for intelligently identifying pre-breakdown types and assessing contact surface conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without paying any creative work.

[0054] Figure 1 Schematic diagram of a flow chart of a method for evaluating the surface condition of a vacuum interrupter contact in the present invention;

[0055] Figure 2 This is a flow chart of a method for evaluating the surface condition of a high-frequency inrush current ablated vacuum interrupter contact based on CRNN in Example 2 of the present invention;

[0056] Figure 3 This is a schematic diagram of a circuit for the DC voltage pre-breakdown characteristic experiment of a vacuum interrupter in Example 2 of the present invention;

[0057] Figure 4 (a), (b), and (c) are schematic diagrams of three types of DC pre-breakdown waveforms in Example 2 of the present invention;

[0058] Figure 5 This is a diagram of a feature extraction data preprocessing module in Example 2 of the present invention;

[0059] Figure 6 This is a diagram of a module for implementing contact surface condition assessment using the improved algorithm in Example 2 of the present invention;

[0060] Figure 7 This is a structural schematic diagram of a vacuum interrupter contact surface condition evaluation system based on an improved CRNN in the present invention.

[0061] Among them: 1-DC voltage source, 2-first resistor, 3-circuit protection switch, 4-first capacitor, 5-second capacitor, 6-second resistor, 7-third resistor, 8-transient voltage suppressor, 9-non-inductive resistor, 10-test vacuum interrupter, 11-displacement sensor, 12-oscilloscope. DETAILED DESCRIPTION

[0062] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations.

[0063] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention as claimed, but rather merely represents selected embodiments of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort shall fall within the scope of protection of the present invention.

[0064] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings.

[0065] The present invention is described in further detail below with reference to the accompanying drawings:

[0066] Example 1

[0067] Convolutional Neural Networks (CNNs) automatically learn features from waveform data through convolutional and pooling layers. Sliding convolution kernels over the data capture key waveform features such as local amplitude variations and frequency spikes. When processing the pre-breakdown current and voltage waveforms of vacuum circuit breakers, CNNs can autonomously discover the unique spatial characteristics of different breakdown waveforms, reducing reliance on empirical evidence. CNNs are also more robust to complex waveforms subject to interference from multiple factors, addressing the limited accuracy of complex waveform classification and improving classification accuracy and adaptability. Furthermore, the pre-breakdown current and voltage waveforms of vacuum circuit breakers exhibit time series characteristics. Recurrent Convolutional Neural Networks (RNNs) can track waveform changes in real time as the device state changes through feedback connections in the hidden layers, processing this time series information. Furthermore, CNNs enhance the dynamic adaptability of the classification model, adjusting the model output to varying operating conditions, thereby addressing the inability of traditional methods to adapt to changing device states.

[0068] The Convolutional Recurrent Neural Network (CRNN), a combination of CNN and RNN, leverages both CNN's ability to extract spatial waveform features and RNN's ability to process time series information. When processing the pre-breakdown current and voltage waveforms of vacuum circuit breakers, the CRNN simultaneously considers both the waveform's local details (such as the spatial distribution of amplitude and frequency) and its temporal variations, comprehensively describing the waveform's characteristics and enabling more accurate and efficient classification of breakdown types, thus facilitating further research into insulation performance. In deep neural networks, as the number of layers increases, the vanishing gradient phenomenon (VGV) during backpropagation is prone to occur, leading to a gradual decline in model performance during actual training. Therefore, residual connections are introduced to more efficiently update the parameters of early convolutional and recurrent layers, preventing these layers from learning useful waveform features due to VGV. This prevents model degradation and accelerates model convergence. Regularization techniques are also employed during the optimization process to prevent overfitting and enhance the model's generalization and stability. Therefore, the present invention proposes a vacuum interrupter contact surface condition assessment method based on CRNN combined with residual connection and regularization technology, which is used to achieve accurate identification and classification of three DC pre-breakdown waveforms, and then accurately evaluate the degree of ablation on the vacuum interrupter surface, thereby improving the safety and reliability of power system operation.

[0069] Pre-breakdown caused by field emission current and pre-breakdown caused by particles can be distinguished by the waveforms of the field emission current and the arc extinguishing chamber voltage. Furthermore, the travel curves can be used to determine the pre-breakdown distance between the high-voltage contact VI_H and the low-voltage contact VI_L, serving as an important reference for classifying pre-breakdown types. For pre-breakdown caused by field emission current (FEPS), before pre-breakdown occurs, the field emission current increases exponentially, and the voltage across the arc extinguishing chamber decreases exponentially. When the field emission current reaches a certain level, pre-breakdown occurs, at which point both the field emission current and the voltage across the arc extinguishing chamber rapidly drop to zero. The corresponding pre-breakdown distance is 1.0-2.0 mm. For pre-breakdown caused by particles (PPS), there is no noticeable increase in the field emission current before pre-breakdown occurs. When pre-breakdown occurs, the voltage across the arc extinguishing chamber suddenly drops to a certain value. Due to the large distance between the contacts at this time, the dielectric strength of the vacuum gap gradually recovers, causing the voltage across the arc extinguishing chamber to increase. As the distance between the contacts continues to decrease, the electric field strength between the contacts gradually increases. When the electric field strength exceeds the insulation strength of the vacuum gap, breakdown will occur again, causing the voltage at both ends of the arc extinguishing chamber to drop sharply to zero, and the corresponding pre-breakdown distance is 2.0~3.0mm. As for the pre-breakdown caused by field emission and microparticles (FE-PPS), there is a significant increase in field emission current before the pre-breakdown occurs. At the same time, after the pre-breakdown occurs, there is a voltage increase caused by insulation recovery, and the corresponding pre-breakdown distance is greater than 3mm. Therefore, the present invention proposes that the pre-breakdown type can be identified and classified based on the waveform diagram of the field emission current and the arc extinguishing chamber voltage, as well as the pre-breakdown distance reflected by the stroke curve, based on CRNN.

[0070] Specifically, such as Figure 1 As shown, the present invention provides a method for evaluating the surface condition of a vacuum interrupter contact, comprising the following steps:

[0071] S1: Acquire the field emission current signal, arc chamber voltage signal and contact stroke curve signal during the pre-breakdown process of the vacuum interrupter;

[0072] S2: performing noise reduction on the acquired field emission current signal, arc extinguishing chamber voltage signal, and contact stroke curve signal to obtain a noise-reduced signal; performing noise reduction on the field emission current signal, arc extinguishing chamber voltage signal, and contact stroke curve signal using an improved short-time Fourier transform with an adaptive window length to obtain a noise-reduced signal; performing decomposition and reconstruction on the noise-reduced signal using a dual-tree complex wavelet transform to obtain an enhanced multi-channel time-frequency signal; performing dynamic normalization processing on the enhanced multi-channel time-frequency signal using an improved K-means clustering method to obtain a preprocessed signal;

[0073] In the above-mentioned noise reduction process, the present invention uses an improved short-time Fourier transform with an adaptive window length to reduce the noise of the field emission current signal, the arc extinguishing chamber voltage signal and the contact stroke curve signal to obtain a noise-reduced signal;

[0074] S3: Based on the preprocessed signals and a pre-built lightweight convolutional recurrent neural network model, the surface condition of the vacuum interrupter contact is evaluated.

[0075] The lightweight convolutional recurrent neural network model is pre-constructed through a customized lightweight gating structure and a bidirectional quasi-recurrent neural network; the lightweight convolutional recurrent neural network model includes a three-layer convolution module, wherein the first layer sets the void rate to 2, the second layer introduces residual connection, and the third layer adds an attention mechanism SE module; the lightweight convolutional recurrent neural network model is obtained by model training through a hybrid loss function and an AdamW optimizer, L2 regularization is performed during training, and a multi-dimensional dynamic risk index and an environmental adaptation factor are introduced.

[0076] Traditional voltage and current waveform denoising methods lack learning ability and adaptability. The present invention uses an improved short-time Fourier transform (STFT) to denoise the original current and voltage waveforms. The adaptive window length is dynamically adjusted according to the kurtosis coefficient, specifically:

[0077]

[0078] in, is the adaptive window length, is the critical frequency, is the sampling frequency of the signal; is the kurtosis coefficient, , is the fourth-order central moment of the current frame, is the standard deviation of the current signal frame data, represents the minimum function, Represents the maximum value function.

[0079] And generate the following time-frequency mask:

[0080]

[0081] in, is the second-order matrix representing the time-frequency mask, is the time-frequency coefficient, that is, the complex amplitude after STFT transformation, Frequency The standard deviation of noise at ; is the time point of the corresponding signal segment, is the frequency corresponding to the time-frequency analysis.

[0082] The cumulative probability distribution of the pre-breakdown distance of the vacuum interrupter is an important parameter for studying the breakdown characteristics between vacuum contacts, which can reflect the insulation characteristics between the two contacts during the closing operation. It is the distance between the two contacts at the moment of pre-breakdown in the vacuum interrupter. The cumulative probability distribution of the cup-shaped longitudinal magnetic contact at the DC voltage level after capacitive closing follows the two-parameter Weibull distribution, as shown in the following formula:

[0083]

[0084] Where: is the cumulative probability distribution function, is the pre-breakdown distance, is the actual breakdown distance, is the shape parameter; for The eigenvalues ​​of Expressed as a natural constant The power operation.

[0085] When the gap length between the two contacts is less than a given When the Weibull distribution can show the probability of pre-breakdown of the vacuum interrupter, it can be calculated based on the experimental data. and The distribution of the pre-breakdown distances of the high-voltage side contact VI_H and the low-voltage side contact VI_L can be obtained through the cumulative probability distribution of the pre-breakdown distances.

[0086] The pre-breakdown distance of the vacuum interrupter is used as a weight to improve the selection of the window function of the Short Time Fourier Transform (STFT):

[0087]

[0088] in, represents the improved Gaussian window function, is the time variable, that is, the time axis coordinate of the signal, is the center time point of the Gaussian window, is the standard deviation of the Gaussian window (i.e. the width parameter of the window), They are respectively the pre-breakdown distances of high and low voltage side contacts, is the proportional coefficient, which can realize adaptive time-frequency analysis.

[0089] The Dual-Tree Complex Wavelet Transform (DTCWT) boasts excellent directional selectivity, near-shift invariance, anti-aliasing, and multi-resolution analysis, enabling precise signal quality improvement. This method is employed in the present invention for signal enhancement. Specifically, the DTCWT is used to decompose and reconstruct the denoised signal to obtain an enhanced multi-channel time-frequency signal. The retained signal is then subjected to a five-layer decomposition. Specifically, during the decomposition and reconstruction process, the threshold is dynamically adjusted based on the frequency band energy, as follows:

[0090]

[0091] in, For the Dynamic threshold of layer subband, For the Noise estimation of the layer, represents the number of sample points, For the The energy proportion of the sub-band is is the total energy of the signal, Expressed as a natural constant The natural logarithm of base .

[0092] The three different DC pre-breakdown types are segmented, and the enhanced multi-channel time-frequency signals are dynamically normalized using the improved K-means clustering method. The dynamic normalization formula improved by the K-means clustering method is:

[0093]

[0094] in, is the normalized signal value, is the original input data, i Indicates the i A signal, is the signal mean of the current pre-breakdown type, which is dynamically updated through real-time K-means clustering; is the standard deviation corresponding to the pre-breakdown type, which is used to scale the data distribution; the scaling factor α = 0.8 prevents the data from being too concentrated after normalization; the offset factor β = 0.1 avoids the gradient saturation problem.

[0095] When designing the model architecture of CRNN, the convolution module selects three layers for time-frequency feature extraction. The hole rate is set to 2 in the first layer. The residual connection (Residual Network, Resnet) is introduced in the second layer to solve the model degradation problem. The third layer adds the attention mechanism Squeeze-and-Excitation (SE) module to dynamically weight important features and add trip weights. The traditional Long Short-Term Memory Network / Gated Recurrent Unit (LSTM / GRU) has high computational complexity, a large number of parameters, and limited long-term processing. Therefore, the present invention customizes a lightweight gating structure and only retains the update gate to reduce the number of parameters. At the same time, a bidirectional quasi-recurrent neural network (QRNN) is adopted for time series modeling. That is, the lightweight convolutional recurrent neural network model in the present invention is pre-constructed using a customized lightweight gating structure and a bidirectional quasi-recurrent neural network. During modeling, the hidden layer is set to 256, and lightweight convolution is used for update. The implementation formula is as follows:

[0096]

[0097]

[0098]

[0099]

[0100] in, is a candidate hidden state, is the weight matrix of the candidate hidden state, Represents the convolution operator, which means performing convolution operation on the input data. is the current time step and its adjacent input sequence, is the current iteration time step, is the convolution window size, is the bias term of the candidate hidden state, represents the hyperbolic tangent function;

[0101] To update the output parameters of the gate, is the Sigmoid function, To update the gate weight, is the bias term of the update gate;

[0102] is the final hidden state, is the final hidden state at the previous moment, represents element-wise multiplication;

[0103] is the attention weight matrix, is the normalized exponential function, are the query matrix and the key matrix respectively, is the matrix transpose operation, For the dimension, Represents two positions in a sequence a, b The distance between them.

[0104] The model's classification head design first processes the 512-dimensional time series features (after mean pooling) output by a quasi-recurrent neural network (QRNN) through a fully connected layer. A Gaussian Error Linear Unit (GeLU) activation function is used for nonlinear transformation, and 30% dropout regularization is added to prevent overfitting, especially for mixed breakdown samples with limited data. A softmax classifier then maps the 512-dimensional features to three categories of breakdown probabilities. The weight matrix has a dimension of 512×3, and a temperature coefficient τ=1 to maintain the original probability distribution. Finally, the model outputs probability predictions for each category. This architecture effectively improves the model's generalization performance through dropout and dimension-appropriate design while ensuring feature expressiveness.

[0105] During the model training and optimization process, the hybrid loss function is set as follows:

[0106]

[0107]

[0108] in, Represents the overall loss function value, represents the cross entropy loss, represents the maximum mean difference loss, Represents the learnable parameter matrix of the model; the hyperparameters are set to 0.2 and 0.1 to balance the contributions of different loss terms.

[0109] Represents the sample characteristics of the source domain (training set), Represents the sample features of the target domain (test set or validation set), and Represents the number of samples in the source domain and the target domain respectively, is the kernel function mapping, represents the reproducing kernel Hilbert space.

[0110] The AdamW optimizer performs L2 regularization on the basis of Adam, can decouple weight decay and prevent overfitting, has good convergence and high stability, and uses a triangular cyclic learning rate:

[0111]

[0112] in, is the learning rate at the current time, Represents the maximum learning rate, which can control the upper limit of parameter update; Represents the minimum learning rate to ensure the final convergence accuracy; is the number of half-cycle steps, which is used to adjust the fluctuation frequency of the learning rate. represents the current iteration time step, Indicates taking the cosine value.

[0113] For the final surface ablation degree evaluation standard, the present invention adopts a fuzzy logic algorithm with strong ability to handle uncertainty and fuzzy boundaries, and designs an index system based on the voltage and current waveforms and the pre-breakdown distance value to achieve dynamic adaptive evaluation of the contact surface state. In order to improve robustness and scalability, a Gaussian membership function is selected as the membership function, and for each input variable , define the membership function of its fuzzy set as:

[0114]

[0115] in, is the membership degree, which indicates the degree of membership of the input variable under a given Gaussian distribution; is the central value of the membership function, To control the standard deviation of the membership function shape, is the probability value of each type of pre-breakdown, is a natural constant.

[0116] Then the rule activation strength is calculated. , activation intensity is the minimum value of the input membership. For each rule, its output fuzzy set is sorted by activation strength Truncate, merge all intercepted fuzzy sets, and calculate the output centroid value :

[0117]

[0118] in, Indicates the output center of gravity value, Score the output, is the degree of membership, represents the discretization sampling points of the output fuzzy set, and ∑ represents the summation.

[0119] The present invention finally judges the contact surface condition of the vacuum interrupter based on the center of gravity value. [0, 25] indicates good performance, (25, 50] indicates slight ablation, (50, 75] indicates moderate ablation, and (75, 100] indicates severe ablation.

[0120] The present invention proposes an algorithm optimization model that deeply integrates time-frequency collaborative preprocessing and lightweight CRNN, which can intelligently identify the DC pre-breakdown waveform of the vacuum interrupter and assist in the accurate evaluation of the surface condition of the vacuum interrupter contact. First, an improved short-time Fourier transform is used to remove noise interference, and dual-tree complex wavelet transform signal enhancement is performed to improve data quality; the improved K-means clustering method is used for normalization operation to map the data to a specific interval to make model training more stable. Secondly, the CRNN model is improved, and the introduction of residual connections in the convolutional layer avoids the degradation problem of the model. The addition of SE modules gives the model a unique advantage in balancing local features and global dependencies; a customized lightweight gating structure is used, and the improved GRU recurrent layer is selected. The bidirectional QRNN is used to determine the number of neurons to capture the time series characteristics of the waveform and establish long-term dependencies. Finally, the AdamW optimizer was integrated for L2 regularization, decoupling weight decay and addressing the slow convergence and overfitting issues of traditional methods. A multidimensional dynamic risk index (mDRI) was established, along with an environmentally adaptive factor. Vacuum interrupter insulation performance evaluation criteria were developed, and parameter coupling corrections were performed to improve the rationality of identification. The improved CRNN algorithm leverages both CNNs to extract spatial waveform features and RNNs to process time series information. It also addresses the model degradation and overfitting issues of traditional algorithms, achieving faster convergence and higher classification performance. This provides a more accurate solution for intelligently identifying pre-breakdown types and assessing contact surface conditions.

[0121] Example 2

[0122] In order to verify the effectiveness of the method proposed in the present invention, this embodiment uses a three-phase circuit breaker composed of two commercial 12kV vacuum interrupters of the same model connected in series as an experimental object.

[0123] Figure 2 This is a flow chart of the method for evaluating the surface condition of high-frequency inrush current ablation vacuum interrupter contacts based on CRNN in this embodiment.

[0124] During the signal acquisition process, the arc chamber voltage waveform can be measured through a voltage divider, the field emission current can be measured through a non-inductive resistor protected by a transient voltage suppressor, and the contact stroke can be recorded through a displacement sensor. Figure 3This is a schematic diagram of the experimental circuit for the DC voltage pre-breakdown characteristics of a vacuum interrupter. Signal 1 represents the travel curve signal, Signal 2 represents the interrupter voltage signal, and Signal 3 represents the field emission current signal. The oscilloscope 12 collects these three signals and is connected to the displacement sensor 11, the voltage divider (composed of the first capacitor 4, the second capacitor 5, the second resistor 6, and the third resistor 7), and the non-inductive resistor 9 protected by a transient voltage suppressor 8, respectively, to collect the travel curve waveform, the interrupter voltage, and the field emission current. The other side of the displacement sensor 11 is connected to the test vacuum interrupter 10, which in turn is connected to the circuit protection switch 3. The circuit protection switch 3 is connected in series with the first resistor 2 and then with the DC voltage source 1. The non-inductive resistor 9, protected by the transient voltage suppressor 8, is grounded at the end away from the test interrupter.

[0125] The field emission current, arc extinguishing chamber voltage, and travel curve waveforms during the closing process are collected. An improved short-time Fourier transform is used for feature extraction and noise removal, followed by signal enhancement using a dual-tree complex wavelet transform. The traditional CRNN is then subjected to L2 regularization, a bidirectional gated loop is added, and the AdamW optimizer is integrated for continuous training and optimization.

[0126] Figure 4 (a), (b) and (c) are schematic diagrams of three types of pre-breakdown waveforms, representing the voltage and field emission current waveforms of pre-breakdown caused by field emission current, pre-breakdown caused by particles, and pre-breakdown caused by field emission current and particles, respectively.

[0127] The data of the DC pre-breakdown voltage, current, and stroke curve of the vacuum interrupter are preprocessed. The feature extraction and denoising processing use the improved short-time Fourier transform to remove noise interference, and the dual-tree complex wavelet transform signal enhancement is used to improve data quality. The improved K-means clustering method is used for normalization operation to map the data to a specific interval to make the model training more stable. Data partitioning is to divide the data into training set, validation set, and test set according to a certain ratio for model training, evaluation, and testing. Figure 5 Diagram of the data preprocessing module for feature extraction.

[0128] Then, combined with the CRNN model architecture design, a three-layer convolutional module was selected, with residual connections introduced in the second layer and attention gating added in the third layer. The recurrent layer was modified from the GRU, and a bidirectional QRNN was used to determine the number of neurons to capture the time series characteristics of the waveform. A fully connected layer and a Softmax classifier were added to achieve classification output for the three types of breakdown.

[0129] During model training and optimization, a hybrid loss function was set and the AdamW optimizer was used to minimize the discrepancy between model predictions and true labels. Hyperparameters such as the learning rate and number of iterations were experimentally adjusted to find the optimal parameter combination. Model performance was monitored using a validation set, and L2 regularization was used to improve convergence and prevent overfitting.

[0130] For the final surface ablation degree evaluation standard, the present invention adopts a fuzzy logic algorithm with strong ability to handle uncertainty and fuzzy boundaries, designs an index system for the pre-breakdown current and voltage waveforms, and realizes dynamic adaptive evaluation of the contact surface state. Figure 6 Diagram of the contact surface condition evaluation module for improving the algorithm.

[0131] Example 3

[0132] In addition, if Figure 7 As shown, the present invention also discloses a vacuum interrupter contact surface condition evaluation system based on an improved CRNN, comprising:

[0133] A signal acquisition module is used to obtain the field emission current signal, the arc extinguishing chamber voltage signal and the contact stroke curve signal during the pre-breakdown process of the vacuum interrupter;

[0134] The signal processing module is used to perform time-frequency coordinated preprocessing on the acquired field emission current signal, arc extinguishing chamber voltage signal and contact travel curve signal; the time-frequency coordinated preprocessing includes:

[0135] Improved short-time Fourier transform with adaptive window length is used to reduce the noise of field emission current signal, arc extinguishing chamber voltage signal and contact stroke curve signal to obtain the noise-reduced signal.

[0136] The denoised signal is decomposed and reconstructed by dual-tree complex wavelet transform to obtain enhanced multi-channel time-frequency signal;

[0137] The enhanced multi-channel time-frequency signal is dynamically normalized using the improved K-means clustering method to obtain the preprocessed signal.

[0138] A signal evaluation module is used to evaluate the surface condition of the vacuum interrupter contact based on the preprocessed signal and a pre-built lightweight convolutional recurrent neural network model;

[0139] The lightweight convolutional recurrent neural network model is pre-constructed through a customized lightweight gating structure and a bidirectional quasi-recurrent neural network; the lightweight convolutional recurrent neural network model includes a three-layer convolution module, wherein the first layer sets the void rate to 2, the second layer introduces residual connection, and the third layer adds an attention mechanism SE module; the lightweight convolutional recurrent neural network model is obtained by model training through a hybrid loss function and an AdamW optimizer, L2 regularization is performed during training, and a multi-dimensional dynamic risk index and an environmental adaptation factor are introduced.

[0140] In addition, a schematic diagram of a terminal device is provided in one embodiment of the present invention. The terminal device in this embodiment includes: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of each of the aforementioned method embodiments are implemented. Alternatively, when the processor executes the computer program, the functions of each module / unit in each of the aforementioned device embodiments are implemented.

[0141] The computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to accomplish the present invention.

[0142] The terminal device may be a computing device such as a desktop computer, a notebook computer, a PDA, a cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0143] The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0144] The memory may be used to store the computer programs and / or modules, and the processor implements various functions of the terminal device by running or executing the computer programs and / or modules stored in the memory and calling the data stored in the memory.

[0145] If the module / unit integrated in the terminal device is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention can implement all or part of the process steps in the above-mentioned method embodiments by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When executed by a processor, the computer program can implement the steps of each of the above-mentioned method embodiments. The computer program includes computer program code, which can be in source code form, object code form, executable file, or some intermediate form. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal, and software distribution medium. It should be noted that the content of the computer-readable medium can be appropriately increased or decreased based on the requirements of legislation and patent practice in a jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media does not include electric carrier signals and telecommunication signals.

[0146] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A method for evaluating the surface condition of a vacuum interrupter contact, characterized in that: The following steps are involved: S1: Acquire the field emission current signal, arc chamber voltage signal and contact stroke curve signal during the pre-breakdown process of the vacuum interrupter; S2: Denoising the acquired field emission current signal, arc extinguishing chamber voltage signal, and contact travel curve signal to obtain a denoised signal; decomposing and reconstructing the denoised signal by dual-tree complex wavelet transform to obtain an enhanced multi-channel time-frequency signal; and dynamically normalizing the enhanced multi-channel time-frequency signal by using an improved K-means clustering method to obtain a preprocessed signal; S3: Evaluate the surface condition of the vacuum interrupter contact based on the preprocessed signal and a lightweight convolutional recurrent neural network model; The lightweight convolutional recurrent neural network model is pre-built using a custom lightweight gating structure and a bidirectional quasi-recurrent neural network. The lightweight convolutional recurrent neural network model includes a three-layer convolution module, wherein the first layer sets a void rate of 2, the second layer introduces a residual connection, and the third layer adds an attention mechanism SE module. The lightweight convolutional recurrent neural network model is obtained by model training using a hybrid loss function and an AdamW optimizer. L2 regularization is performed during training, and a multi-dimensional dynamic risk index and an environmental adaptation factor are introduced. The lightweight convolutional recurrent neural network model is pre-built using a custom lightweight gating structure and a bidirectional quasi-recurrent neural network. Specifically, when pre-built, the number of hidden layer neurons is 256, and the update gate formula is: in, is a candidate hidden state, is the weight matrix of the candidate hidden state, represents the convolution operator, is the current time step and its adjacent input sequence, is the current iteration time step, is the convolution window size, is the bias term of the candidate hidden state, represents the hyperbolic tangent function; To update the output parameters of the gate, is the Sigmoid function, To update the gate weight, is the bias term of the update gate; is the final hidden state, is the final hidden state at the previous moment, represents element-wise multiplication; is the attention weight matrix, is the normalized exponential function, are the query matrix and the key matrix respectively, is the matrix transpose operation, For the dimension, Represents two positions in a sequence a , b the distance between them; The custom lightweight gating structure only retains the update gate.

2. A method for evaluating the surface condition of a vacuum interrupter contact according to claim 1, characterized in that: In the step of performing noise reduction on the acquired field emission current signal, arc extinguishing chamber voltage signal, and contact stroke curve signal, an improved short-time Fourier transform with an adaptive window length is specifically used to perform noise reduction on the field emission current signal, arc extinguishing chamber voltage signal, and contact stroke curve signal, wherein the adaptive window length is dynamically adjusted according to the kurtosis coefficient, specifically as follows: in, is the adaptive window length, is the critical frequency, is the sampling frequency of the signal; is the kurtosis coefficient, , is the fourth-order central moment of the current frame, is the standard deviation of the current signal frame data, represents the minimum function, Represents the maximum value function.

3. The method for evaluating the surface condition of a vacuum interrupter contact according to claim 1, wherein: The de-noised signal is decomposed and reconstructed by dual-tree complex wavelet transform to obtain an enhanced multi-channel time-frequency signal. Specifically, during the decomposition and reconstruction process, the threshold is dynamically adjusted according to the frequency band energy, specifically as follows: in, For the Dynamic threshold of layer subband, For the Noise estimation of the layer, represents the number of sample points, For the The energy proportion of the sub-band is is the total energy of the signal, Expressed as a natural constant The natural logarithm of base .

4. A method for evaluating the surface condition of a vacuum interrupter contact according to claim 1, characterized in that: The enhanced multi-channel time-frequency signal is dynamically normalized using the improved K-means clustering method, specifically: in, is the normalized signal value, is the original input data, i Indicates the i A signal, is the signal mean of the current pre-breakdown type, is the standard deviation of the corresponding pre-breakdown type, is the scaling factor, is the offset coefficient.

5. The method for evaluating the surface condition of a vacuum interrupter contact according to claim 1, wherein: The hybrid loss function is specifically: in, Represents the overall loss function value, represents the cross entropy loss, represents the maximum mean difference loss, represents the learnable parameter matrix of the model, represents the 2-norm of W; Represents the source domain sample characteristics, represents the sample characteristics of the target domain, and Represents the number of samples in the source domain and the target domain respectively, is the kernel function mapping, is the reproducing kernel Hilbert space.

6. A method for evaluating the surface condition of a vacuum interrupter contact according to claim 1, characterized in that: When using the AdamW optimizer for model training, the learning rate adopts a triangular cycle strategy, specifically: in, is the learning rate at the current time, is the maximum learning rate, is the minimum learning rate; is the number of steps in a half cycle, is the current iteration time step, Indicates taking the cosine value.

7. A method for evaluating the surface condition of a vacuum interrupter contact according to claim 1, characterized in that: The surface condition assessment of the vacuum interrupter contact is performed based on the preprocessed signal and the lightweight convolutional recurrent neural network model, specifically: Based on the fuzzy logic algorithm, the Gaussian membership function is used to define the membership of the input variables. The formula is: in, is the degree of membership; is the central value of the membership function, To control the standard deviation of the membership function shape, is the probability value of each type of pre-breakdown, is a natural constant; Calculating rule activation strength , and merge the intercepted fuzzy sets to output the centroid value ,according to The interval determines the degree of contact surface ablation, specifically: in, Indicates the output center of gravity value, represents the discretization sampling points of the output fuzzy set, Score the output, For affiliation , ∑ represents summation.

8. A vacuum interrupter contact surface condition assessment system, characterized in that: A method for evaluating the surface condition of a vacuum interrupter contact, used to implement any one of claims 1 to 7, comprising: A signal acquisition module is used to obtain the field emission current signal, the arc extinguishing chamber voltage signal and the contact stroke curve signal during the pre-breakdown process of the vacuum interrupter; a signal processing module for performing noise reduction on the acquired field emission current signal, arc extinguishing chamber voltage signal, and contact travel curve signal to obtain a noise-reduced signal; decomposing and reconstructing the noise-reduced signal by dual-tree complex wavelet transform to obtain an enhanced multi-channel time-frequency signal; and dynamically normalizing the enhanced multi-channel time-frequency signal by using an improved K-means clustering method to obtain a preprocessed signal; The signal evaluation module evaluates the surface condition of the vacuum interrupter contacts based on the preprocessed signals and a lightweight convolutional recurrent neural network model; The lightweight convolutional recurrent neural network model is pre-constructed through a customized lightweight gating structure and a bidirectional quasi-recurrent neural network; the lightweight convolutional recurrent neural network model includes a three-layer convolution module, wherein the first layer sets the void rate to 2, the second layer introduces residual connection, and the third layer adds an attention mechanism SE module; the lightweight convolutional recurrent neural network model is obtained by model training through a hybrid loss function and an AdamW optimizer, L2 regularization is performed during training, and a multi-dimensional dynamic risk index and an environmental adaptation factor are introduced.

9. A computer system comprising a memory, a processor, and a computer program stored in the memory, wherein: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 7.

Images