High-voltage circuit breaker state quantity feature extraction method and system

Abstract

The invention belongs to the technical field of power equipment state monitoring and fault diagnosis, and discloses a high-voltage circuit breaker state quantity feature extraction method and system, and the method comprises the steps: extracting a local maximum value point and a local minimum value point from a denoised original monitoring signal; feature vectors are constructed and input into the trained neural network model, six-dimensional vectors are output in a Softmax output layer, and the six-dimensional vectors correspond to prediction probabilities of six orders in the optimal B-spline interpolation orders; based on the constructed envelope, adopting empirical mode decomposition to obtain a plurality of intrinsic mode function IMF components, performing Hilbert transform on each IMF component, constructing a state quantity feature set according to the instantaneous frequency and amplitude change of the analysis signal, and judging whether the high-voltage circuit breaker is in an abnormal state or not. According to the method, the vibration signals collected by the circuit breaker are subjected to multi-stage intelligent processing, high-precision and low-noise feature extraction is achieved, and fault diagnosis and state recognition of the circuit breaker can be effectively supported.

Description

Technical Field

[0001] This invention relates to the field of power system monitoring and fault diagnosis technology, specifically a method and system for extracting state variables of high-voltage circuit breakers. Background Technology

[0002] High-voltage circuit breakers are critical protective devices in power systems, bearing the important responsibility of rapidly interrupting current and preventing equipment damage in the event of a fault. Their operation is accompanied by complex nonlinear and non-stationary signals such as vibration and current, making efficient and reliable feature extraction of their state variables crucial.

[0003] Traditional signal processing methods, such as algorithms based on time-domain or frequency-domain analysis, struggle to fully address the non-stationary characteristics present in circuit breaker signals. The Hilbert-Huang Transform (HHT) has garnered significant attention due to its ability to process nonlinear and non-stationary signals. This method comprises two parts: Empirical Mode Decomposition (EMD) and Hilbert Transform. It decomposes complex signals into a series of Intrinsic Mode Functions (IMFs) and analyzes their time-frequency characteristics using instantaneous frequency and amplitude.

[0004] However, the key step in EMD (construction of upper and lower envelopes) faces technical bottlenecks in practical applications. Traditional methods use cubic spline interpolation, which easily introduces distortion at signal boundaries and leads to overfitting. B-spline interpolation, as an improvement, has been gradually adopted due to its stronger boundary control capabilities, but its performance is highly dependent on the interpolation order. Currently, most mainstream methods use manual specification or fixed strategies based on grid search, which cannot dynamically select the optimal order according to specific signal characteristics, making it difficult to meet the requirements of high robustness and online performance in practical applications. For example, the invention patent with publication number CN117591826A discloses a transformer winding evaluation method, device, and storage medium based on mode decomposition. It acquires the vibration signal of the transformer in operation, then uses the artificial fish swarm optimization method to obtain the optimal parameter combination of the time-varying filtering empirical mode decomposition algorithm, and then inputs the transformer vibration signal into the decomposition algorithm to obtain the intrinsic mode function. Based on the center frequency value of the intrinsic mode function, the high and low frequency energy of the signal is calculated, and the operating status of the transformer winding is evaluated by the high and low frequency energy. This solves the problem that traditional transformer winding status detection methods cannot monitor online and have poor timeliness.

[0005] In addition, existing models often use a single loss function (such as mean squared error or cross-entropy) during training, ignoring the structural morphology of the signal. This results in the B-spline envelope predicted by the model being inconsistent with the modal structure of the real signal, thus affecting the fault diagnosis effect. Summary of the Invention

[0006] Purpose of the invention: The present invention aims to propose a method for extracting state variables of high-voltage circuit breakers, which solves the problems of inaccurate extraction of fault state features and low fault identification accuracy in the prior art. The present invention also proposes a system for extracting state variables of high-voltage circuit breakers.

[0007] Technical solution: This invention first provides a method for extracting state variables of high-voltage circuit breakers, the method comprising:

[0008] Local maxima and minima are extracted from the denoised original monitoring signal to form basic nodes with upper and lower envelopes;

[0009] Feature vectors are constructed from the denoised original signal and extracted extreme points, and then input into the trained neural network model. A structure-preserving loss function is used to optimize the parameters of the neural network. Finally, a six-dimensional vector is output in the Softmax output layer, which corresponds to the prediction probabilities of the six orders in the optimal B-spline interpolation order. Based on the obtained optimal order, the B-spline interpolation method is used to fit the maxima and minima respectively to construct the upper and lower envelopes.

[0010] Based on the constructed envelope, several intrinsic mode functions (IMF) components are obtained by empirical mode decomposition. Hilbert transform is then performed on each IMF component, and each IMF component and the result of the Hilbert transform are used to construct an analytic signal.

[0011] Based on the instantaneous frequency and amplitude changes of the analyzed signal, a state variable feature set is constructed. Combined with the fault knowledge base, it is determined whether the high-voltage circuit breaker is in an abnormal state.

[0012] Furthermore, including:

[0013] The extraction of local maxima and minima from the denoised original monitoring signal includes:

[0014] Let the signal be x(t). Under discrete conditions, a point is considered an extreme point when the following conditions are met:

[0015] For the i-th sampling point, if the first derivative of the signal x ′ (i) = 0 and the second derivative x ″ (i) < 0, indicating a maximum value;

[0016] If the first derivative x ′(i) = 0 and the second derivative x ″ (i) > 0, which is a minimum value;

[0017] Among them, the derivative is approximately calculated by the difference method, and the extreme point interval is set to 50% of the signal period to avoid interpolation oscillation caused by over-density.

[0018] Furthermore, it includes:

[0019] Constructing a feature vector from the denoised original signal and the extracted extreme points includes:

[0020] Performing statistical analysis and frequency-domain feature extraction on the denoised signal to generate a one-dimensional feature vector. Among them, the extracted features include: the number of extreme points, the mean value of the maximum value, the standard deviation of the maximum value, the mean value of the minimum value, the standard deviation of the minimum value, kurtosis, skewness, main frequency, spectral entropy, and the above eight feature components are combined into a one-dimensional feature vector with a length of 8:

[0021] F = [f1, f2, f3,..., f8].

[0022] Furthermore, it includes:

[0023] The above eight feature components are obtained from the input one-dimensional voltage disturbance signal sequence through the following method:

[0024] Number of extreme points: By traversing the original signal sequence, each sampling point x(i) that satisfies the following conditions is an extreme point:

[0025] Maximum value: x(i) > x(i - 1) and x(i) > x(i + 1); Minimum value: x(i) < x(i - 1) and x(i) < x(i + 1)

[0026] Record the number of points that meet the conditions, denoted as feature f1;

[0027] Mean value of the maximum value and standard deviation of the maximum value:

[0028] In the set of maximum value points, let its value be the set X max = {x1, x2,..., x n}, then: Here, n is the total number of maximum value points,

[0029] The mean value of the maximum value f2 is:

[0030]

[0031] The standard deviation of the maximum value f3 is:

[0032]

[0033] Mean value of the minimum value and standard deviation of the minimum value:

[0034] Let the set of local minima be X. min ={x1,x2,...,x m}, then: m is the total number of local minima.

[0035] The minimum mean f4 is expressed as:

[0036]

[0037] The minimum standard deviation f5 is expressed as:

[0038]

[0039] Kurtosis: Let the signal sampling sequence be X = {x1, x2, ..., x...} T}, with mean μ and standard deviation σ, then:

[0040]

[0041] This is denoted as feature f6, used to measure the peak level of the signal; T is the total number of sampling points;

[0042] Skewness: Represents the symmetry of a signal, and is calculated using the following formula:

[0043]

[0044] This is denoted as feature f7, reflecting the degree of signal offset;

[0045] Dominant frequency: Perform a Fast Fourier Transform on the original signal to obtain the spectrum X(f), and select the frequency f8 corresponding to the non-zero frequency component with the largest amplitude, i.e.:

[0046]

[0047] Furthermore, including:

[0048] The method of using a structure-preserving loss function to optimize the parameters of the neural network includes:

[0049] The loss function is defined as follows:

[0050]

[0051] Among them, MSE(e pred ,e true ) represents the B-spline envelope e predicted by the model. pred With reference optimal envelope e true The mean square error between the two is used to measure the accuracy of the envelope fitting, and the calculation formula is as follows:

[0052]

[0053] Where N represents the number of signal sampling points, e pred (i) and e true (i) are the values ​​of the predicted and reference envelopes at the i-th sampling point, respectively;

[0054] DTW (IMF) pred IMF true The distance () represents the dynamic time warping distance between the IMF component sequences generated by the model prediction order and the reference IMF sequence. It is used to measure the consistency of the modal structure and is calculated using the following formula:

[0055]

[0056] Where x = IMF pred ,y=IMF true w represents the set of all feasible time-aligned paths. Each pair of points (i,j) represents the alignment point pair of x and y.

[0057] Furthermore, including:

[0058] The six-dimensional vector output by the Softmax output layer corresponds to the prediction probabilities of six orders in the optimal B-spline interpolation order, including:

[0059] The Softmax output layer is used to perform B-spline interpolation order prediction on the high-dimensional embedded feature vector output by the neural network model. Its input is the 64-dimensional intermediate feature vector output by the neural network model, and its output is the probability distribution vector of the corresponding order set {k = 3, 4, 5, 6, 7, 8}, specifically including the following:

[0060] The function expression for the Softmax output layer is as follows:

[0061]

[0062] Among them, z k P represents the original score at order k, mapped from the output vector of a fully connected layer in a neural network model after a linear transformation. k This is the probability of predicting an order of k;

[0063] The output vector is a six-dimensional probability vector [P3, P4, P5, P6, P7, P8], where each term represents the order probability that the signal is best suited for B-spline interpolation. Therefore, the following condition is met:

[0064]

[0065] Furthermore, including:

[0066] After outputting a six-dimensional vector at the Softmax output layer, which corresponds to the prediction probabilities of the six orders in the optimal B-spline interpolation order, the method further includes:

[0067] The k value corresponding to the highest probability is selected as the optimal interpolation order, i.e.:

[0068]

[0069] If the maximum probability If the value is less than the set threshold, a secondary verification mechanism or manual review will be triggered to improve the robustness of the model.

[0070] Furthermore, including:

[0071] The step of fitting the maximum and minimum points using B-spline interpolation based on the obtained optimal order to construct the upper and lower envelopes includes:

[0072] Based on the obtained optimal order k * B-spline interpolation was used to fit the maxima and minima respectively, and the upper envelope E was constructed. upper (t) and lower envelope E lower (t);

[0073] The B-spline used is defined as follows:

[0074]

[0075] Among them: B i,k *(t) indicates that the order is k * The i-th B-spline basis function, c i The corresponding control point coefficients, i∈[0,n], represent the basis function numbers involved in the interpolation, n=N c -1, where N c It is the number of extreme points currently used to construct the envelope, that is, the number of maximum points corresponding to the upper envelope and the number of minimum points corresponding to the lower envelope. Therefore, n is the total number of basis functions minus 1.

[0076] Furthermore, t in the formula is a continuous time variable in the interpolation process, while the node vector {t} i} is the basis function B i,k The set of implicit parameters upon which *(t) depends determines the support range and local properties of each basis function across the entire interval. The selection strategy for the node vector significantly impacts interpolation performance; therefore, this invention employs a non-uniformly distributed node vector {t}. i The algorithm is adaptively generated based on the distribution of actual extreme points to mitigate boundary effects and improve fitting stability and local control capabilities.

[0077] This construction method not only effectively maintains the smoothness of signal boundaries, but also has good fitting flexibility, providing a more reliable local mean estimate for subsequent empirical mode decomposition.

[0078] Furthermore, including:

[0079] The constructed envelope is used to obtain several intrinsic mode functions (IMF) components through empirical mode decomposition. A Hilbert transform is then performed on each IMF component, and each IMF component and the result of the Hilbert transform are combined to form an analytic signal, including:

[0080] Calculate the local mean based on the constructed envelope:

[0081]

[0082] The first round of residual signal is obtained by subtracting the local mean m(t) from the original signal;

[0083] Through multiple iterations, several intrinsic mode function (IMF) components c are extracted. i (t);

[0084] Each iteration continues until the IMF meets the following condition:

[0085] The number of extreme points and the number of zero crossover points in the IMF do not differ by more than 1.

[0086] The mean of the local extremum envelope approaches 0;

[0087] The residual signal energy is below the threshold or the number of IMFs reaches the preset upper limit.

[0088] For each IMF component c i (t) is subjected to Hilbert transform to obtain its instantaneous amplitude A. i (t) and instantaneous frequency f i (t);

[0089] c i (t) and its Hilbert transform Constructing analytical signals:

[0090]

[0091] Among them, the instantaneous amplitude is Instantaneous phase is The instantaneous frequency is the derivative of its phase with respect to time:

[0092]

[0093] Furthermore, including:

[0094] The fault types include overload faults, short circuit faults, ground faults, and mechanical faults.

[0095] On the other hand, the present invention also provides a high-voltage circuit breaker state quantity feature extraction system, the system comprising:

[0096] The signal processing module is used to extract local maxima and minima from the denoised original monitoring signal to form basic nodes with upper and lower envelopes.

[0097] The fitting module is used to construct feature vectors from the denoised original signal and extracted extreme points, and input them into the trained neural network model. It uses a structure-preserving loss function to optimize the parameters of the neural network. Finally, it outputs a six-dimensional vector at the Softmax output layer. This six-dimensional vector corresponds to the prediction probabilities of the six orders in the optimal B-spline interpolation order. Based on the obtained optimal order, the B-spline interpolation method is used to fit the maxima and minima respectively to construct the upper and lower envelopes.

[0098] The analytical module is used to obtain several intrinsic mode function (IMF) components based on the constructed envelope using empirical mode decomposition, and to perform Hilbert transform on each IMF component, and to construct an analytical signal from each IMF component and the result of the Hilbert transform.

[0099] The diagnostic module is used to construct a set of state variables based on the instantaneous frequency and amplitude changes of the analyzed signal, and combine it with a fault knowledge base to determine whether the high-voltage circuit breaker is in an abnormal state.

[0100] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0101] 1. Improved Signal Decomposition Accuracy: This invention introduces an intelligent order prediction module based on CNN and Softmax classifiers to adaptively optimize the order parameters of B-spline interpolation. Compared to traditional methods that rely on empirical settings or grid search, this approach better matches the local features and variation patterns of the signal, effectively reducing fitting errors during the envelope construction stage and improving the stability and resolution of IMF in empirical mode decomposition.

[0102] Specifically, an improved B-spline interpolation method is used to fit the maxima and minima of the signal to construct upper and lower envelopes, providing a foundation for subsequent Empirical Mode Decomposition (EMD). Unlike traditional methods that manually set the B-spline order or rely on grid search strategies, this invention proposes an intelligent prediction module based on Convolutional Neural Networks (CNN) and a Softmax classifier. This module adaptively determines the optimal order k based on the signal's statistical and frequency domain characteristics, improving the efficiency and accuracy of envelope construction.

[0103] 2. Enhanced structural preservation capability: By introducing a structural preservation loss function during model training, combined with envelope fitting error (MSE) and mode alignment error (DTW), the predicted envelope has higher structural consistency with the actual signal. This is especially helpful in preserving the physical characteristics of high-voltage circuit breaker vibration signals with structural features such as short-term impact and frequency drift, thereby improving the interpretability of the model.

[0104] Specifically, to further enhance the model's interpretability and its ability to preserve typical circuit breaker structural signals, this invention designs a structure-preserving loss function. It jointly optimizes the Mean Squared Error (MSE) and the intrinsic mode function (IMF) modal structure alignment error based on Dynamic Time Warping (DTW), thereby enhancing the model's ability to identify key features such as short-term impacts and frequency changes. Based on this, the EMD method is used to progressively decompose the IMF, and a Hilbert transform is performed on each IMF component to obtain its instantaneous frequency and amplitude information, thus extracting key feature variables reflecting the circuit breaker's operating state.

[0105] Therefore, combining points 1 and 2 above, the method of this invention exhibits excellent adaptability, structure preservation, and real-time performance, making it particularly suitable for processing circuit breaker vibration signals with nonlinear and non-stationary characteristics. Compared to traditional signal processing methods, the B-spline interpolation scheme combining CNN+Softmax order prediction and structure-preserving loss function optimization significantly reduces boundary effects and overfitting, improving the accuracy and robustness of signal decomposition. This method can be widely applied to online condition monitoring, fault diagnosis, and intelligent operation and maintenance of high-voltage circuit breakers and other equipment in power systems, demonstrating high engineering practical value and promising prospects for widespread application.

[0106] 3. Applicable to monitoring scenarios of various types of non-stationary power equipment: This method is not only applicable to the extraction of state variables of high-voltage circuit breakers, but also to the state monitoring and intelligent diagnosis of other power equipment with non-stationary signal characteristics, such as instrument transformers, transformers, and surge arresters. It has good scalability and promotion value. Attached Figure Description

[0107] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0108] Figure 1 This is a flowchart of the high-voltage circuit breaker state variable feature extraction method according to an embodiment of the present invention;

[0109] Figure 2 This is a schematic diagram of the processing flow for order adaptive prediction implemented by the CNN+Softmax architecture according to an embodiment of the present invention. Detailed Implementation

[0110] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0111] like Figure 1 As shown, the present invention also provides a method for extracting state variables and diagnosing faults in high-voltage circuit breakers based on improved B-splines and Hilbert-Huang transform, comprising the following steps:

[0112] a) Signal preprocessing and extreme point extraction

[0113] The raw monitoring signals collected from the high-voltage circuit breaker are denoised to remove high-frequency noise and other interference. The state variables involved include current, voltage, vibration signals, and other monitoring signals related to the circuit breaker's operating state. In this embodiment, denoising methods such as wavelet denoising and Kalman filtering can be used. Wavelet denoising uses the Daubechies wavelet basis, with a decomposition level of 5, and a soft thresholding method is selected. Then, the maxima and minima in the signal are extracted. The extraction method for extrema is based on the criterion that the first derivative is zero; that is, when the gradient (first derivative) of the signal is greater than a certain set threshold, it is determined as a maxima or minima. The interval between extrema is set to 50% of the signal period. Specifically:

[0114] In this embodiment, the raw vibration or voltage signals collected during the operation of the high-voltage circuit breaker are preprocessed to remove noise and abnormal interference, thereby improving the accuracy of subsequent analysis. Wavelet denoising or a Kalman filter can be used as the denoising method. Specifically, the wavelet denoising uses the Daubechies wavelet basis, such as db4, with a decomposition level of 5, and a soft threshold function is used to achieve a smooth transition.

[0115] After denoising, the first derivative extremum method is used to extract the maximum and minimum points of the signal. Let the signal be x(t). Under discrete conditions, a signal is considered an extremum when the following conditions are met:

[0116] For the i-th sampling point, if the first derivative x ′ (i) = 0 and the second derivative x ″ (i) < 0, indicating a maximum value;

[0117] If the first derivative x ′ (i) = 0 and the second derivative x ″ (i)>0, which is a local minimum;

[0118] The derivative is approximated using the finite difference method, and the interval between extreme points is set to 50% of the signal period to avoid excessive interpolation oscillations.

[0119] b) Feature vector construction

[0120] The denoised signal undergoes statistical analysis and frequency domain feature extraction to generate a one-dimensional feature vector. Statistical features include the signal's mean, standard deviation, skewness, and kurtosis; frequency domain features are obtained through Fast Fourier Transform (FFT) to obtain the dominant frequency and spectral entropy. The extracted features include: number of extreme points, mean of maximum value, standard deviation of maximum value, mean of minimum value, standard deviation of minimum value, kurtosis, skewness, dominant frequency, and spectral entropy. These features will be used as feature vectors input to subsequent models.

[0121] Specifically, feature vectors are constructed from the denoised original signal and extracted extreme points, which are then used as input to the subsequent CNN model. The feature dimensions include, but are not limited to, the following components:

[0122] Number of extreme points N ext This reflects the frequency of local changes in the signal;

[0123] Maximum Mean The arithmetic mean of all maxima;

[0124] Maximum Standard Deviation The degree of dispersion of the maxima;

[0125] Minimum mean μ min The arithmetic mean of all local minima;

[0126] Minimum standard deviation σ min The degree of dispersion of the minimum value;

[0127] Kurtosis It measures the sharpness of the signal and characterizes the features of impulse-type faults;

[0128] Skewness When measuring signal symmetry, an asymmetric distribution may indicate an anomaly.

[0129] The dominant frequency is the frequency component corresponding to the maximum value of the spectrum obtained through the Fast Fourier Transform.

[0130] The above 8 features constitute the input vector

[0131] c) B-spline order prediction

[0132] The constructed one-dimensional feature vector v is input into the trained convolutional neural network (CNN) to predict the optimal B-spline interpolation order k∈{3,4,5,6,7,8}. This network contains two one-dimensional convolutional layers, a max-pooling layer, a fully connected layer, and a softmax classifier to automatically predict the B-spline interpolation order. In this embodiment, the CNN model structure is as follows:

[0133] Input layer: Receives one-dimensional feature vectors with dimension 1×N; N is the number of features, such as 8.

[0134] The first convolutional layer contains 16 one-dimensional convolutional kernels with a kernel length of 3 and a stride of 1. It uses the ReLU activation function and has an output dimension of 16×(N-2).

[0135] First pooling layer: Use max pooling with a window size of 2 and a stride of 2 to reduce the output dimension to 16×(N-2) / 2;

[0136] The second convolutional layer contains 32 convolutional kernels with a kernel length of 3, the activation function is ReLU, and the output dimension is 32×((N-2) / 2-2).

[0137] Second pooling layer: window size is 2, output dimension is 32×(((N-2) / 2-2) / 2);

[0138] Flatten layer: Flattens the convolutional feature tensor into a one-dimensional vector as an intermediate representation;

[0139] Fully connected layer: Maps the flattened vector to a vector of length 64, and outputs high-dimensional intermediate features for the Softmax classifier.

[0140] The CNN continuously optimizes the convolution kernel parameters and fully connected layer weights during the training phase using the backpropagation algorithm, enabling it to effectively extract feature information related to the B-spline order.

[0141] Specifically, in this embodiment, the input layer receives a 1x8 one-dimensional feature vector.

[0142] Convolutional layer 1: 16 convolutional kernels, kernel width 3, stride 1, activation function ReLU;

[0143] Max pooling layer 1: window size is 2, stride size is 2;

[0144] Convolutional layer 2: 32 convolutional kernels, kernel width 3, stride 1, activation function ReLU;

[0145] Max pooling layer 2: window size is 2;

[0146] Flattened layer:

[0147] Fully connected layer: 64 nodes;

[0148] Softmax output layer: Using the Softmax activation function, the output is a six-dimensional vector p = [p3, p4, p5, p6, p7, p8], which corresponds to the predicted probability of each order.

[0149] The Softmax classifier in this embodiment is used to predict the B-spline interpolation order based on the high-dimensional embedded feature vector output by the CNN. Its input is the 64-dimensional intermediate feature vector output by the CNN network, and its output is the probability distribution vector of the corresponding order set {k = 3, 4, 5, 6, 7, 8}, specifically including the following:

[0150] The expression for the Softmax function is as follows:

[0151]

[0152] Among them, z k P represents the original score at order k, mapped from the output vector of a fully connected CNN layer after a linear transformation. k This is the probability of predicting an order of k.

[0153] The output vector is a six-dimensional probability vector [P3, P4, P5, P6, P7, P8], where each term represents the order probability that the signal is best suited for B-spline interpolation, satisfying:

[0154]

[0155] Order decision rule: Select the k value corresponding to the highest probability as the optimal interpolation order.

[0156]

[0157] If the maximum probability is less than a set threshold, such as 0.5, a secondary verification mechanism or manual review is triggered to improve the robustness of the model.

[0158] In this embodiment, a structure-preserving loss function is introduced during model training to simultaneously optimize prediction accuracy and signal morphology consistency. This loss function consists of two parts: first, the mean squared error (MSE) between the predicted envelope and the reference envelope; and second, the dynamic time warping (DTW) distance between the predicted intrinsic mode functions (IMFs) and the reference IMF, used to measure the alignment of the decomposed structure.

[0159] Specifically, in this embodiment, a structure-preserving loss function is introduced during model training to enhance the ability to model the physical structure of the signal. This loss function is defined as follows:

[0160]

[0161] Among them, MSE(e pred ,e true The calculation formula is as follows:

[0162]

[0163] N represents the number of signal sampling points, e pred (i) and e true (i) are the values ​​of the predicted and reference envelopes at the i-th sampling point, respectively;

[0164] Among them, DTW (IMF) pred IMF true The calculation formula is:

[0165]

[0166] Where x = IMF pred ,y=IMF true w represents the set of all feasible time-aligned paths. Each pair of points (i,j) represents the alignment point pair of x and y. When calculating the DTW distance, this invention selects only the first N main IMFs, such as N=3, and calculates the alignment error of each IMF separately and then averages the results to avoid higher-order modal errors interfering with the overall evaluation.

[0167] Here, α and β are adjustable weighting factors, reflecting the relative weights of fitting accuracy and structure preservation. The recommended default settings are α = 0.7 and β = 0.3. This combined loss function can guide the model to generate a more physically consistent envelope and decomposition structure during the training phase, adapting to the time-frequency characteristics of high-voltage circuit breaker vibration signals under different conditions.

[0168] In this embodiment, the structure-preserving loss function guides the prediction results during the training phase, ensuring that while automatically predicting the B-spline order, the model maintains the physical structural features of the circuit breaker vibration signal, such as short-term impact patterns, dominant frequency variation trends, and modal distribution patterns, thereby improving the interpretability and reliability of the signal decomposition. Furthermore, by jointly optimizing the envelope fitting error of the signal and the consistency of the modal structure, the structure-preserving loss function effectively retains physical features such as short-term impacts, dominant frequency drift, and attenuation trends in the high-voltage circuit breaker vibration signal, thereby improving the interpretability of the signal decomposition results and the accuracy of fault identification.

[0169] The final order selection strategy is as follows: the order k corresponding to the maximum value in the Softmax output probability is used as the optimal interpolation order for constructing the upper and lower envelopes; if the maximum probability is lower than the set threshold, such as the threshold being set to 0.5, a secondary confirmation mechanism is triggered, which can be combined with envelope fitting residual rescoring, prior distribution adjustment, or manual intervention to improve robustness.

[0170] In other words, in this embodiment, the selection criterion for the order k is that when the predicted probability of the optimal order k is greater than the threshold, the order is selected; otherwise, secondary verification or manual review is triggered to improve the reliability of the diagnosis.

[0171] d) Construction of B-spline envelope

[0172] Based on the optimal order k obtained in step c), * B-spline interpolation was used to fit the maxima and minima respectively, and the upper envelope E was constructed. upper (t) and lower envelope E lower (t).

[0173] The node vectors for B-spline interpolation can be uniformly or non-uniformly distributed depending on the location of the extreme points, in order to reduce boundary effects and overfitting of the envelope. In practice, B-spline basis functions are used for interpolation, where the number of nodes is determined by the number of extreme points, and the order of the B-spline interpolation is determined by the order k.

[0174] In this embodiment, the interpolation process uses the following B-spline definition:

[0175]

[0176] in:

[0177] ·B i,k *(t) indicates that the order is k * The i-th B-spline basis function,

[0178] ●c i For the corresponding control point coefficients,

[0179] ●i∈[0,n], representing the basis function numbers involved in the interpolation.

[0180] n = N c -1, where N c It is the number of extreme points currently used to construct the envelope (the upper envelope corresponds to the number of maximum points, and the lower envelope corresponds to the number of minimum points), so n is the total number of basis functions minus 1.

[0181] Furthermore, t in the formula is a continuous time variable in the interpolation process, while the node vector {t} i} is the basis function B i,k The set of implicit parameters upon which *(t) depends determines the support range and local properties of each basis function across the entire interval. The selection strategy for the node vector significantly impacts interpolation performance; therefore, this invention employs a non-uniformly distributed node vector {t}. i The algorithm is adaptively generated based on the distribution of actual extreme points to mitigate boundary effects and improve fitting stability and local control capabilities.

[0182] This construction method not only effectively maintains the smoothness of signal boundaries, but also has good fitting flexibility, providing a more reliable local mean estimate for subsequent empirical mode decomposition.

[0183] e) Empirical Mode Decomposition (EMD)

[0184] First, calculate the local mean based on the constructed envelope:

[0185]

[0186] Based on the constructed envelope, the local mean of the signal is calculated, and Empirical Mode Decomposition (EMD) is performed stepwise. The EMD method extracts the IMFs of the signal by progressively subtracting the local mean until the remaining signal meets a stopping condition. The stopping condition is typically that the signal residual is less than a preset threshold or the number of IMFs reaches an upper limit. The extracted IMF components exhibit good time-frequency locality in the frequency domain, making them suitable for state-varying analysis of high-voltage circuit breakers.

[0187] Specifically, in this embodiment, the original signal is subtracted from the local mean to obtain the first-round residual signal. Through multiple iterations, several intrinsic mode function (IMF) components are extracted.

[0188] Each iteration continues until the IMF meets the following condition:

[0189] The number of extreme points and the number of zero crossover points in the IMF do not differ by more than 1.

[0190] The mean of the local extremum envelope approaches 0;

[0191] The residual signal energy is below the threshold or the number of IMFs reaches the preset upper limit, such as 10.

[0192] f) Hilbert Transform

[0193] For each IMF component, a Hilbert transform is applied to obtain its instantaneous amplitude and instantaneous frequency. The Hilbert transform is implemented through convolution, using principal value integration to avoid the boundary problems of traditional integration. The instantaneous amplitude of each IMF component is obtained through an absolute value function, and the instantaneous frequency is calculated by the difference of the phase derivative.

[0194] Specifically, in this embodiment, for each IMF component c i (t) is subjected to Hilbert transform to obtain its instantaneous amplitude A. i (t) and instantaneous frequency f i (t). The Hilbert transform is defined as:

[0195]

[0196] c i (t) and its Hilbert transform Constructing analytical signals:

[0197]

[0198] Among them, the instantaneous amplitude is Instantaneous phase is The instantaneous frequency is the derivative of its phase with respect to time:

[0199]

[0200] g) Feature extraction and fault diagnosis

[0201] Based on the instantaneous frequency and amplitude changes of each IMF component, its time-frequency distribution characteristics, such as frequency drift rate, energy concentration, and the average and standard deviation of instantaneous frequency, are extracted to construct a state variable feature set. This set is then combined with a fault knowledge base to determine whether the circuit breaker is in an abnormal state.

[0202] Common fault types include:

[0203] Overload fault: The frequency range is abnormally low, and the amplitude is greatly increased;

[0204] Short circuit fault: instantaneous energy concentration, with a rapid change in frequency;

[0205] Grounding fault: Low-frequency harmonics are enhanced, and there is a sudden change in skewness;

[0206] Mechanical failure: Enhanced mid-to-high frequency components and increased waveform asymmetry.

[0207] On the other hand, the present invention also provides a high-voltage circuit breaker state quantity feature extraction system, the system comprising:

[0208] The signal processing module is used to extract local maxima and minima from the denoised original monitoring signal to form basic nodes with upper and lower envelopes.

[0209] The fitting module is used to construct feature vectors from the denoised original signal and extracted extreme points, and input them into the trained neural network model. It uses a structure-preserving loss function to optimize the parameters of the neural network. Finally, it outputs a six-dimensional vector at the Softmax output layer. This six-dimensional vector corresponds to the prediction probabilities of the six orders in the optimal B-spline interpolation order. Based on the obtained optimal order, the B-spline interpolation method is used to fit the maxima and minima respectively to construct the upper and lower envelopes.

[0210] The analytical module is used to obtain several intrinsic mode function (IMF) components based on the constructed envelope using empirical mode decomposition, and to perform Hilbert transform on each IMF component, and to construct an analytical signal from each IMF component and the result of the Hilbert transform.

[0211] The diagnostic module is used to construct a set of state variables based on the instantaneous frequency and amplitude changes of the analyzed signal, and combine it with a fault knowledge base to determine whether the high-voltage circuit breaker is in an abnormal state.

[0212] Other technical features of the high-voltage circuit breaker state variable feature extraction system described in this invention are similar to the corresponding high-voltage circuit breaker state variable feature extraction method, and will not be repeated here.

[0213] Throughout this description, numerous specific details, such as examples of components and / or methods, are provided to provide a complete understanding of embodiments of this application. However, those skilled in the art will recognize that embodiments of this application may be practiced without one or more of these specific details or by other devices, systems, components, methods, parts, materials, components, etc.

[0214] It should also be understood that one or more of the elements shown in the figures may be implemented in a more separate or more integrated manner, or may even be removed because they are inoperable in certain circumstances or provided because they may be useful for a particular application.

[0215] Furthermore, unless otherwise expressly stated, any arrows in the accompanying drawings should be considered illustrative only and not limiting. Additionally, unless otherwise stated, the term "or" as used herein is generally intended to mean "and / or". Where a term is anticipated to provide a separation or combination capability that is unclear, a combination of components or steps will also be considered as indicated.

[0216] The above description of the embodiments shown in this application (including the content in the specification summary) is not intended to be an exhaustive enumeration or to limit this application to the precise forms disclosed herein. Although specific embodiments and examples of this application have been described herein for illustrative purposes only, various equivalent modifications are possible within the spirit and scope of this application, as will be recognized and understood by those skilled in the art. As indicated, these modifications can be made to this application in accordance with the above description of the embodiments of this application, and such modifications will be within the spirit and scope of this application.

[0217] This document has generally described the systems and methods in detail to aid in understanding the present application. Furthermore, various specific details have been provided to offer a general understanding of the embodiments of this application. However, those skilled in the art will recognize that embodiments of this application can be practiced without one or more specific details, or using other means, systems, accessories, methods, components, materials, parts, etc. In other instances, well-known structures, materials, and / or operations have not been specifically shown or described in detail to avoid obscuring various aspects of the embodiments of this application.

[0218] Therefore, although this application has been described herein with reference to specific embodiments thereof, freedom of modification, various changes and substitutions are also within the scope of the above disclosure, and it should be understood that in some cases, certain features of this application may be adopted without departing from the scope and spirit of the invention and without corresponding use of other features. Thus, many modifications can be made to adapt a particular environment or material to the essential scope and spirit of this application. This application is not intended to be limited to the specific terminology used in the following claims and / or the specific embodiments disclosed as the best mode of carrying out this application, but this application will include any and all embodiments and equivalents falling within the scope of the appended claims. Therefore, the scope of this application will be determined only by the appended claims.

Claims

1. A method for extracting state variables of a high-voltage circuit breaker, characterized in that, The method includes: Extracting local maximum points and minimum points from the denoised original monitoring signal to form basic nodes with upper and lower envelope lines; Constructing a feature vector from the denoised original signal and the extracted extreme points, and inputting it into the trained neural network model. A structure-preserving loss function is used to optimize the parameters of the neural network. Finally, a six-dimensional vector is output at the Softmax output layer. This six-dimensional vector corresponds to the prediction probabilities of six orders in the optimal B-spline interpolation order. According to the obtained optimal order, the B-spline interpolation method is used to fit the maximum points and minimum points respectively to construct the upper envelope line and the lower envelope line; Based on the constructed envelope lines, several Intrinsic Mode Function (IMF) components are obtained by empirical mode decomposition, and each IMF component is subjected to Hilbert transform. Each IMF component and the result after Hilbert transform form an analytic signal; According to the instantaneous frequency and amplitude changes of the analytic signal, a state quantity feature set is constructed, and combined with the fault knowledge base, it is judged whether the high-voltage circuit breaker is in an abnormal state.

2. The method for extracting state variables of high-voltage circuit breakers according to claim 1, characterized in that, The extracting of local maximum points and minimum points from the denoised original monitoring signal includes: Let the signal be x(t). Under discrete conditions, it is determined as an extreme point when the following conditions are met: For the i-th sampling point, if the first derivative of the signal x ′ (i) = 0 and the second derivative x ″ (i) < 0, indicating a maximum value; If the first derivative x ′ (i) = 0 and the second derivative x ″ (i)>0, which is a local minimum; Among them, the derivative is approximately calculated by the difference method, and the extreme point interval is set to 50% of the signal period to avoid interpolation oscillation caused by over-density.

3. The method for extracting state variables of high-voltage circuit breakers according to claim 1, characterized in that, The constructing of the feature vector from the denoised original signal and the extracted extreme points includes: Performing statistical analysis and frequency-domain feature extraction on the denoised signal to generate a one-dimensional feature vector. The extracted features include: the number of extreme points, the mean value of the maximum value, the standard deviation of the maximum value, the mean value of the minimum value, the standard deviation of the minimum value, kurtosis, skewness, the main frequency, and spectral entropy. The above eight feature components are combined into a one-dimensional feature vector with a length of 8: F = [f1, f2, f3,..., f8].

4. The method for extracting state variables of high-voltage circuit breakers according to claim 3, characterized in that, The above eight feature components are obtained from the input one-dimensional voltage disturbance signal sequence through the following methods: The number of extreme points: By traversing the original signal sequence, each sampling point x(i) meets the following conditions to be an extreme point: Maximum value: x(i) > x(i - 1) and x(i) > x(i + 1); Minimum value: x(i) < x(i - 1) and x(i) < x(i + 1) Record the number of points that meet the conditions, denoted as the feature f1; The mean value of the maximum value and the standard deviation of the maximum value: Let the set of maxima be the set X. max ={x1,x2,...,x n }, where n is the total number of maxima, then: The mean value of the maximum value f2 is: The standard deviation of the maximum value f3 is: The mean value of the minimum value and the standard deviation of the minimum value: Let the set of local minima be X. min ={x1,x2,...,x m Let m be the total number of local minima, then: The mean value of the minimum value f4 is expressed as: The standard deviation of the minimum value f5 is expressed as: Kurtosis: Let the signal sampling sequence be X = {x1, x2, ..., x...} T Let the mean be μ, the standard deviation be σ, and T be the total number of sampling points. Then: Denoted as the feature f6, which is used to measure the peak degree of the signal; Skewness: Represents the symmetry of the signal, and the calculation formula is: Denoted as the feature f7, which reflects the offset degree of the signal; The main frequency: Performing a fast Fourier transform on the original signal to obtain the spectrum X(f), and selecting the frequency f8 corresponding to the non-zero frequency component with the largest amplitude, that is:

5. The method for extracting state variables of high-voltage circuit breakers according to claim 1, characterized in that, The using of the structure-preserving loss function to optimize the parameters of the neural network includes: The loss function is defined as follows: Where α and β are adjustable weighting factors, reflecting the relative weights of fitting accuracy and structure preservation, MSE(e pred ,e true ) represents the B-spline envelope e predicted by the model. pred With reference optimal envelope e true The mean square error between the two is used to measure the accuracy of the envelope fitting, and the calculation formula is as follows: Where N represents the number of signal sampling points, e pred (i) and e true (i) are the values ​​of the predicted and reference envelopes at the i-th sampling point, respectively; DTW (IMF) pred IMF true The distance () represents the dynamic time warping distance between the IMF component sequences generated by the model prediction order and the reference IMF sequence. It is used to measure the consistency of the modal structure and is calculated using the following formula: Where x = IMF pred ,y=IMF true w represents the set of all feasible time-aligned paths. Each pair of points (i,j) represents the alignment point pair of x and y.

6. The method for extracting state variables of high-voltage circuit breakers according to claim 1, characterized in that, The six-dimensional vector output by the Softmax output layer corresponds to the prediction probabilities of six orders in the optimal B-spline interpolation order, including: The Softmax output layer is used to perform B-spline interpolation order prediction on the high-dimensional embedded feature vector output by the neural network model. Its input is the 64-dimensional intermediate feature vector output by the neural network model, and its output is the probability distribution vector of the corresponding order set {k = 3, 4, 5, 6, 7, 8}, specifically including the following: The function expression for the Softmax output layer is as follows: Among them, z k P represents the original score at order k, mapped from the output vector of a fully connected layer in a neural network model after a linear transformation. k This is the probability of predicting an order of k; The output vector is a six-dimensional probability vector [P3, P4, P5, P6, P7, P8], where each term represents the order probability that the signal is best suited for B-spline interpolation. Therefore, the following condition is met:

7. The method for extracting state variables of high-voltage circuit breakers according to claim 6, characterized in that, After outputting a six-dimensional vector at the Softmax output layer, which corresponds to the prediction probabilities of the six orders in the optimal B-spline interpolation order, the method further includes: The k value corresponding to the highest probability is selected as the optimal interpolation order, i.e.: If the maximum probability If the value is less than the set threshold, a secondary verification mechanism or manual review will be triggered to improve the robustness of the model.

8. The method for extracting state variables of a high-voltage circuit breaker according to claim 7, characterized in that, The step of fitting the maximum and minimum points using B-spline interpolation based on the obtained optimal order to construct the upper and lower envelopes includes: Based on the obtained optimal order k * B-spline interpolation was used to fit the maxima and minima respectively, and the upper envelope E was constructed. upper (t) and lower envelope E lower (t); The B-spline used is defined as follows: Among them: B i,k *(t) indicates that the order is k * The i-th B-spline basis function, c i The corresponding control point coefficients, i∈[0,n], represent the basis function numbers involved in the interpolation; n=N c -1, where N c It is the number of extreme points currently used to construct the envelope, that is, the number of maximum points corresponding to the upper envelope and the number of minimum points corresponding to the lower envelope. Therefore, n is the total number of basis functions minus 1. Furthermore, t in the formula is a continuous time variable in the interpolation process, while the node vector {t} i } is the basis function B i,k The set of implicit parameters upon which *(t) depends determines the support range and local properties of each basis function throughout the entire interval.

9. The method for extracting state variables of high-voltage circuit breakers according to claim 1, characterized in that, The constructed envelope is used to obtain several intrinsic mode functions (IMF) components through empirical mode decomposition. A Hilbert transform is then performed on each IMF component, and each IMF component and the result of the Hilbert transform are combined to form an analytic signal, including: Calculate the local mean based on the constructed envelope: The first round of residual signal is obtained by subtracting the local mean m(t) from the original signal; Through multiple iterations, several intrinsic mode function (IMF) components c are extracted. i (t); Each iteration continues until the IMF meets the following condition: The number of extreme points and the number of zero crossover points in the IMF do not differ by more than 1. The mean of the local extremum envelope approaches 0; The residual signal energy is below the threshold or the number of IMFs reaches the preset upper limit. For each IMF component c i (t) is subjected to Hilbert transform to obtain its instantaneous amplitude A. i (t) and instantaneous frequency f i (t); c i (t) and its Hilbert transform Constructing analytical signals: Among them, the instantaneous amplitude is Instantaneous phase is The instantaneous frequency is the derivative of its phase with respect to time:

10. The method for extracting state variables of a high-voltage circuit breaker according to claim 9, characterized in that, The fault types include overload faults, short circuit faults, ground faults, and mechanical faults.

11. A system for extracting state parameters of a high-voltage circuit breaker, characterized in that, The system includes: The signal processing module is used to extract local maxima and minima from the denoised original monitoring signal to form basic nodes with upper and lower envelopes. The fitting module is used to construct feature vectors from the denoised original signal and extracted extreme points, and input them into the trained neural network model. It uses a structure-preserving loss function to optimize the parameters of the neural network. Finally, it outputs a six-dimensional vector at the Softmax output layer. This six-dimensional vector corresponds to the prediction probabilities of the six orders in the optimal B-spline interpolation order. Based on the obtained optimal order, the B-spline interpolation method is used to fit the maxima and minima respectively to construct the upper and lower envelopes. The analytical module is used to obtain several intrinsic mode function (IMF) components based on the constructed envelope using empirical mode decomposition, and to perform Hilbert transform on each IMF component, and to construct an analytical signal from each IMF component and the result of the Hilbert transform. The diagnostic module is used to construct a set of state variables based on the instantaneous frequency and amplitude changes of the analyzed signal, and combine it with a fault knowledge base to determine whether the high-voltage circuit breaker is in an abnormal state.

Images