Pole-mounted circuit breaker mechanical fault prediction method and system based on feature analysis

Abstract

The present application belongs to the technical field of power equipment state monitoring, and particularly relates to a mechanical fault prediction method and system for pole-mounted circuit breakers based on feature analysis, which comprises the following steps: aligning and segmenting the obtained circuit breaker vibration time series data, mapping each segment of sampling points into a data embedding vector, generating a time-frequency dual-component position code in combination with the main frequency in the frequency domain and superimposing it on the embedding vector, obtaining a query matrix, a key matrix and a value matrix through linear transformation; generating a non-Gaussian metric and constructing a gating matrix, fusing the value matrix to obtain a weighted value matrix, and simultaneously constructing a block composite kernel matrix with bandwidth adjusted according to the main frequency by using double kernels; calculating the feature angle of the query and key vectors, dynamically determining the angle threshold in combination with the non-Gaussian metric and generating a sparse mask, fusing to obtain a kernel attention matrix to complete feature extraction, and inputting the prediction network to obtain the fault prediction result. The present application can improve the accuracy and operational stability of mechanical fault prediction for pole-mounted circuit breakers.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring technology. More specifically, this invention relates to a method and system for predicting mechanical faults in pole-mounted circuit breakers based on feature analysis. Background Technology

[0002] As the most numerous and widely distributed switching equipment in the distribution network, pole-mounted circuit breakers (POPs) play a crucial role in load switching and rapid isolation of faulty sections. The operational reliability of their mechanical structure directly determines the continuity and security of the power supply. The vibration signals generated during circuit breaker operation contain rich state information about the mechanism's actions. State assessment and fault diagnosis based on vibration time-series analysis have become the mainstream technical approach for online monitoring of POPs-mounted circuit breakers. Since the opening and closing actions of circuit breakers are transient impact processes, the vibration signals exhibit both non-stationarity and strong nonlinearity. Impact components are superimposed and intertwined with background noise. Effectively extracting the time-frequency correlation characteristics characterizing early deterioration of the mechanism is a key research focus and challenge in this field.

[0003] Current research focuses on vibration signal fault diagnosis for pole-mounted circuit breakers. For example, Chinese patent application CN121346910A discloses a method and system for diagnosing mechanical faults in pole-mounted circuit breakers. This scheme simultaneously acquires dynamic force-displacement signals, high-frequency acoustic emission signals, and broadband vibration signals during the opening and closing process of the operating mechanism. A blind source separation algorithm is used to decouple the acoustic emission and vibration signals into impact source components and friction source components. Envelope spectrum impact features are then extracted from the impact source components, and friction noise energy features are calculated from the friction source components. Finally, the dynamic force-displacement curve fitting results are combined to distinguish and demarcate mechanical wear and lubrication deterioration faults. This scheme, through multi-source signal coordination and component-level feature extraction, demonstrates good distinguishing ability for specific types of mechanical faults.

[0004] However, this scheme still has significant limitations in practical applications. At the feature representation level, signal processing focuses on component decoupling, with time-domain and frequency-domain analyses remaining relatively independent. It fails to deeply integrate the time-domain waveform of the vibration signal with the segmented dominant frequency variation characteristics, resulting in insufficient mining of joint time-frequency information and inadequate feature representation of the sequence position dimension. Regarding anti-interference and adaptability, the scheme primarily focuses on diagnosing two types of faults: mechanical wear and lubrication deterioration. It lacks targeted feature enhancement and suppression mechanisms for the transient impact interference and non-stationary noise commonly found in circuit breaker vibration signals, making it difficult to adapt to the prediction needs of various fault types under complex on-site conditions. At the feature association modeling level, the feature extraction stage still relies mainly on traditional signal processing methods such as envelope spectrum analysis and energy calculation. It has not established a feature association model that considers both local transient detail associations and global long-range trend similarities, limiting its ability to capture long-sequence signal associations and restricting the improvement of fault prediction accuracy. Summary of the Invention

[0005] To address the problems of low accuracy in predicting mechanical faults and insufficient operational stability in existing pole-mounted circuit breakers, this invention provides solutions in the following aspects.

[0006] In a first aspect, the present invention provides a method for predicting mechanical faults of pole-mounted circuit breakers based on feature analysis, comprising: S1, using a synchronous trigger signal to align and segment the acquired vibration time-series data of the circuit breaker, mapping the sampling points in each data segment to a data embedding vector sequence; extracting the frequency domain dominant frequency features of the data segment and combining them with a time-coded vector to generate a position code, which is then superimposed on the data embedding vector, and then linearly transformed to generate a query matrix, a key matrix, and a value matrix; S2, extracting the higher-order statistical moments of the data embedding vector within a local time window to generate a non-Gaussian metric, and constructing a gating matrix accordingly, and fusing the gating matrix with the value matrix to obtain a weighted value matrix; Using a first kernel function and a second kernel function, the local correlation and global similarity between the query vector and the key vector in the query matrix and the key matrix are calculated respectively, and a block composite kernel matrix with bandwidth parameters adjusted according to the frequency domain dominant frequency features is constructed; S3, the feature angle between the query vector and the key vector is calculated, and the angle threshold is determined based on the non-Gaussian metric at the corresponding time step. A sparse mask matrix is ​​generated based on the comparison result between the feature angle and the angle threshold; the block composite kernel matrix and the sparse mask matrix are fused into a kernelized attention matrix, and the weighted value matrix is ​​weighted using the kernelized attention matrix to extract features, and then input into the prediction network to obtain the circuit breaker fault prediction result.

[0007] This invention first aligns and segments vibration time-series data to generate embedding vectors, then fuses the dominant frequency and time information in the frequency domain to construct a dual-component position code. A query, key, and value matrix is ​​generated through linear transformation to complete the feature initialization for time-frequency fusion. Secondly, a non-Gaussian gated weighted matrix is ​​constructed based on local higher-order statistical moments. Simultaneously, a block-based composite kernel matrix with bandwidth adaptive to the dominant frequency is constructed using a dual-kernel architecture, taking into account both local details and global correlations. Finally, a sparse mask is generated through feature angles and dynamic thresholds, and the resulting kernelized attention matrix is ​​fused to complete feature aggregation. This is then fed into the prediction network output, enriching the time-frequency feature representation, suppressing noise interference, and improving the accuracy and efficiency of fault prediction.

[0008] Preferably, extracting the frequency domain dominant frequency feature of the data segment and combining it with the time coding vector to generate a position code, and then superimposing it on the data embedding vector, includes: performing a frequency domain transformation on the data segment to obtain an amplitude spectrum, and extracting the frequency component with the largest amplitude as the frequency domain dominant frequency feature;

[0009] A frequency coding vector is generated based on the frequency domain dominant frequency feature, and then combined with a time coding vector generated based on the time step sequence number to generate the position code. The position code is added element by element to the corresponding data embedding vector to obtain the superimposed data embedding vector.

[0010] Preferably, the higher-order statistical moments of the data embedding vector within the local time window are extracted to generate a non-Gaussian metric, and a gating matrix is ​​constructed accordingly. The gating matrix is ​​then fused with the value matrix to obtain a weighted value matrix, including: calculating the mean absolute value of the third-order standard moments of the data embedding vector in each feature dimension within the local time window as skewness, and the mean of the fourth-order standard moments as kurtosis. The non-Gaussian metric is obtained by weighted summation of the skewness and the kurtosis. The non-Gaussian metric on the time series is mapped to a preset interval through a preset activation function to generate a one-dimensional nonlinear gated vector. The nonlinear gated vector is then expanded in the feature dimension and multiplied element-wise with the value matrix to generate the weighted value matrix.

[0011] This invention calculates and weights the skewness and kurtosis of each feature dimension within a local window to obtain a non-Gaussian metric. This metric is then mapped using an activation function to generate a gating matrix, which is broadcast-expanded and multiplied element-wise with the value matrix. This method utilizes higher-order statistical moments to identify non-Gaussian impulse components in the signal, highlights fault-related impulse features through gating and weighting, suppresses feature components corresponding to stationary background noise, and increases the proportion of effective fault information in the value matrix.

[0012] Preferably, the step of using a first kernel function and a second kernel function to calculate the local correlation and global similarity of query vectors and key vectors in the query matrix and key matrix, respectively, and constructing a block composite kernel matrix with bandwidth parameters adjusted according to the frequency domain dominant frequency characteristics, includes: calculating local bandwidth parameters and global bandwidth parameters based on the frequency domain dominant frequency characteristics; for query vectors and key vectors belonging to the same data segment, calculating the Euclidean distance and substituting it into a Gaussian kernel function with local bandwidth parameters as the first kernel function to obtain the local correlation, and constructing a diagonal block; for query vectors and key vectors belonging to different data segments, calculating the Manhattan distance and substituting it into a Laplace kernel function with global bandwidth parameters as the second kernel function to obtain the global similarity, and constructing an off-diagonal block; concatenating the diagonal block and the off-diagonal block in chronological order to generate the block composite kernel matrix.

[0013] Preferably, generating a sparse mask matrix based on the comparison result of the feature angle and the angle threshold includes: obtaining the feature angle between the query vector and the key vector; If the feature angle is less than the angle threshold, a preset retention weight value is assigned to the corresponding position in the sparse mask matrix. If the feature angle is greater than or equal to the angle threshold, then a zero value is assigned to the corresponding position in the sparse mask matrix to generate the sparse mask matrix.

[0014] By using the angle threshold to filter highly correlated feature pairs and remove redundant associations with excessively large feature differences, the attention weights are made sparse. This not only reduces the amount of ineffective computation and improves the efficiency of operation, but also filters out false associations caused by noise and enhances the targeting of feature aggregation.

[0015] Preferably, determining the included angle threshold based on the non-Gaussian metric at the corresponding time step includes: obtaining the non-Gaussian metrics of the query vector and the key vector at the corresponding time steps, and calculating the arithmetic mean of the two as a joint non-Gaussian feature; Obtain a preset base angle threshold, and add the product of the joint non-Gaussian feature and the preset constant scaling factor to the base angle threshold to obtain the adjusted candidate threshold. The candidate threshold is subject to boundary constraints, which limit it to a preset range between the minimum and maximum included angle thresholds, and the included angle threshold is then output.

[0016] By taking the mean of the non-Gaussian metrics at the corresponding time steps of the query vector and key vector as a joint feature, candidate thresholds are generated by combining the base threshold and scaling factor, and then the final angle threshold is output after boundary constraints. This method dynamically adjusts the angle threshold according to the degree of non-Gaussian impact of the signal; the stronger the impact, the looser the threshold, which can retain more impact-related feature associations. At the same time, boundary constraints prevent the threshold from going out of bounds, ensuring the stable operation of the mechanism.

[0017] Preferably, the step of generating the query matrix, key matrix, and value matrix through a linear transformation includes: initializing mutually independent learnable query weight matrix, key weight matrix, and value weight matrix, and setting corresponding initial bias vectors; The data embedding vector after superimposing the position encoding is multiplied by the query weight matrix, key weight matrix and value weight matrix respectively, and the corresponding initial bias vector is added for affine transformation to generate a query matrix, key matrix and value matrix with the same feature dimensions as the data embedding vector.

[0018] Preferably, the step of aligning and segmenting the acquired circuit breaker vibration time-series data using a synchronous trigger signal, and mapping the sampling points in each data segment into a data embedding vector sequence, includes: acquiring circuit breaker vibration time-series data and the trip coil current signal as the synchronous trigger signal; calculating the sequence difference using the rising edge of the trip coil current signal; and locating the maximum difference point as the alignment zero point to complete the alignment segmentation. The aligned and segmented circuit breaker vibration time series data is divided into multiple data segments using a sliding window, and local vibration features are extracted using a one-dimensional convolutional network. The one-dimensional sampling points are then mapped to a high-dimensional space to form the data embedding vector.

[0019] Preferably, the weighted matrix is ​​weighted using the kernelized attention matrix to extract features, which are then input into the prediction network to obtain circuit breaker fault prediction results, including: The weighted extracted features are flattened along the time dimension and input into a prediction network consisting of a bidirectional long short-term memory network and a multilayer perceptron cascade. By predicting the probability distribution vector of different mechanical states of the circuit breaker output by the network, the category index corresponding to the maximum probability value is extracted to determine the mechanical fault type, which is used as the circuit breaker fault prediction result.

[0020] By flattening the attention-weighted features along the time dimension, the input is given to a prediction network consisting of a bidirectional long short-term memory network and a multilayer perceptron. The network outputs the probability distribution of various mechanical states and takes the category corresponding to the highest probability as the prediction result. The bidirectional long short-term memory network is used to capture the temporal bidirectional dependency relationship, and the multilayer perceptron is used to complete the mapping from features to fault categories. The probabilistic diagnostic results are output to achieve the classification and discrimination of various mechanical fault states.

[0021] Secondly, the present invention provides a feature-based mechanical fault prediction system for pole-mounted circuit breakers, including a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned feature-based mechanical fault prediction method for pole-mounted circuit breakers is implemented.

[0022] By adopting the above technical solution, the above-mentioned feature analysis-based method for predicting mechanical faults of pole-mounted circuit breakers is generated into a computer program and stored in a memory for loading and execution by a processor. This allows for the creation of a terminal device based on the memory and processor, facilitating its use.

[0023] The technical solution of the present invention has the following beneficial technical effects: This invention employs a kernelized attention mechanism combining time-frequency dual-component position encoding, non-Gaussian metric-gated weighting, block-based composite kernel matrix, and dynamic sparse masking to perform deep feature extraction and fault prediction on the vibration time-series signal of pole-mounted circuit breakers. It integrates the dominant frequency features in the frequency domain into the position encoding process, enhancing the joint time-frequency representation capability of the vibration signal. Non-Gaussian gating is generated based on higher-order statistical moments within a local window, and the value matrix features are dynamically weighted, effectively suppressing non-stationary impact interference and highlighting effective feature components related to faults. Block-based dual-kernel differentiated modeling is used, taking into account both local detail correlation and global trend similarity. Combined with dynamic threshold sparse masking, feature aggregation focuses more on highly correlated information, overcoming the shortcomings of existing technologies such as insufficient time-frequency information mining, inadequate anti-interference capability, and single time-series correlation modeling. This improves the accuracy and operational stability of circuit breaker mechanical fault prediction and better adapts to the equipment condition monitoring needs under complex on-site conditions. Attached Figure Description

[0024] Figure 1 The flowchart shows a method for predicting mechanical faults in pole-mounted circuit breakers based on feature analysis. Figure 2 This is a schematic diagram illustrating the relationship between non-Gaussian metrics and the variation of gate values. Figure 3 This is a schematic diagram comparing the performance of ablation experiments. Detailed Implementation

[0025] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0026] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0027] As attached Figure 1 As shown in the figure, this embodiment provides a method for predicting mechanical faults of pole-mounted circuit breakers based on feature analysis. The method is divided into three stages, and the specific implementation of each stage is described in detail below.

[0028] S1. Map the data to generate a query matrix, a key matrix, and a value matrix.

[0029] This step mainly involves the acquisition and preprocessing of vibration data, as well as the generation and superposition of time-frequency position codes and the linear transformation of the query matrix, key matrix, and value matrix.

[0030] First, vibration data is acquired and aligned. A piezoelectric accelerometer is deployed on the surface of the pole-mounted circuit breaker housing to collect vibration timing signals during operation in real time, which are then transmitted to the industrial control computer via a data acquisition card. Simultaneously, the trip unit coil current signal is acquired as a synchronization trigger signal, and sequence alignment is performed using the rising edge of the current signal. Specifically, the alignment operation involves calculating the sequence difference of the coil current signal, locating the maximum difference point using the argmax function (corresponding to the instant of a sharp current rise), and using this point as the alignment zero point to achieve precise synchronization of multiple batches of signals.

[0031] Next, the aligned vibration time-series data is segmented and embedded. A sliding window method with a fixed window length and sliding step size is used to divide the long sequence into multiple data segments of equal length. In one specific parameter setting, the system sampling frequency is 10kHz, the window length is 2048 sampling points, corresponding to a time span of approximately 0.2048 seconds, the sliding step size is 512 sampling points, and the window overlap rate is 75%. Assuming the total signal contains 10,000 sampling points, it can be divided into 16 time-series data segments.

[0032] The one-dimensional sampling points within each data segment are mapped to a fixed-dimensional high-dimensional vector, i.e., a data embedding vector. In one implementation, each data segment is input into a linear projection layer, and the high-dimensional mapping is completed through a linear transformation. The calculation formula is as follows:

[0033] In the formula, This represents the input sampling point matrix, where each row corresponds to a sampling time and each column represents the single-channel vibration amplitude. The weight matrix is ​​a learnable matrix. These are the bias vectors, which are obtained through iterative optimization during model training. The output data embedding vector matrix has rows corresponding to time steps and columns corresponding to feature dimensions. In this embodiment, a single sampling point within each data segment is a one-dimensional vibration amplitude, which is linearly projected and mapped to a 256-dimensional embedding vector. Therefore, the weight matrix... The dimension is Bias vector The dimension is A data segment contains 2048 sampling points, which, after linear projection, yield a 2048×256 data embedding vector matrix. In another implementation, a one-dimensional convolutional network can be used to extract local vibration features. The convolutional kernel performs local feature extraction and high-dimensional mapping on the sampling points within the data segment to form the data embedding vector, thereby enhancing the ability to capture local waveform features of the vibration signal.

[0034] To overcome the shortcomings of traditional position coding, which only reflects time sequence and cannot characterize the frequency domain features of vibration signals, this embodiment introduces time-frequency dual-component position coding. A Fast Fourier Transform (FFT) is performed on each data segment to obtain its frequency domain amplitude spectrum. The frequency component corresponding to the maximum amplitude is extracted as the dominant frequency value of that data segment. For example, if the frequency component with the largest amplitude in the amplitude spectrum of a data segment after FFT is 850Hz, then the dominant frequency value of that data segment is 850Hz.

[0035] A frequency encoding vector is generated based on the dominant frequency value, and a sinusoidal time encoding vector is generated based on the time step sequence number. The two are then added together to obtain a time-frequency dual-component positional encoding. The frequency encoding vector is generated as follows: an initial frequency encoding vector is generated based on the dominant frequency value using a sine function, and then it is expanded to the same feature dimension as the data embedding vector, i.e., 256 dimensions, using a multilayer perceptron. This multilayer perceptron includes an input layer, hidden layers, and an output layer. All layers are fully connected and non-linear activation functions are applied. Its calculation process can be represented as follows: ; ; In the formula, This is the initial frequency encoding vector generated by mapping the main frequency value using a sine function; It is a nonlinear activation function, i.e., a modified linear unit function; and These are the weight matrices from the input layer to the hidden layer and from the hidden layer to the output layer, respectively. and For the corresponding bias vector, all four types of parameters are optimized together during model training; The hidden layer outputs features. This is the expanded 256-dimensional frequency-coded vector. The generated location coding matrix is ​​then added element-wise to the data embedding vector matrix, which also has a dimension of 2048×256, to achieve the fusion of location information and content features.

[0036] Finally, a linear transformation is performed on the position-encoded data embedding vector to generate the query matrix, key matrix, and value matrix required for subsequent attention calculations. Specifically, three independent learnable weight matrices are initialized as the query weight matrix, key weight matrix, and value weight matrix, respectively. Each weight matrix has a size of 256×256, and the parameters are initialized using a Xavier uniform distribution to ensure gradient stability in the early stages of training. Matrix multiplication is performed between the position-encoded embedding vector matrix and these three weight matrices, and the corresponding bias vectors are added to the results. Each bias vector has 256 dimensions and an initial value of 0, generating the query matrix, key matrix, and value matrix in sequence. The feature dimensions of these three matrices are all maintained at 2048×256, consistent with the input embedding vector. During model training, the weights and bias parameters are updated using the backpropagation algorithm combined with the AdamW optimizer, with an initial learning rate set to 0.001.

[0037] S2. Calculate the features to obtain the weighted matrix and the block composite kernel matrix.

[0038] This step mainly performs gated weighting based on non-Gaussian metrics and completes the construction of a block kernel matrix based on composite kernel functions.

[0039] Impact components and fault symptoms in vibration signals often exhibit non-Gaussian statistical characteristics that deviate from a normal distribution. To capture this non-Gaussianity and dynamically adjust the feature weights in the value matrix accordingly, this embodiment quantifies the non-Gaussianity of the signal by calculating higher-order statistical moments within a local time window.

[0040] Specifically, a local time window containing 21 time steps is formed by taking 10 time steps before and after each time step as the center. Within this window, the third and fourth standard moments are calculated for the data in each feature dimension. The third standard moment, or skewness, measures the symmetry of the data distribution; the fourth standard moment, or kurtosis, measures the steepness of the data distribution. The skewness of the time step is calculated by averaging the absolute values ​​of the third standard moments for all feature dimensions; the kurtosis of the time step is calculated by averaging the fourth standard moments for all feature dimensions. For example, at a certain time step, the skewness can be calculated as 0.8 and the kurtosis as 3.5. The skewness and kurtosis are then weighted and summed according to preset weights to obtain the non-Gaussian metric characterizing the local non-Gaussian distribution. The preset weighting coefficients can be set according to the prior characteristics of the fault signal. In one example, the skewness coefficient is set to 0.4 and the kurtosis coefficient is set to 0.6, resulting in a non-Gaussian metric of 2.4 for that time step.

[0041] The principle for determining the weighting coefficients is as follows: Kurtosis is more sensitive to the characteristics of impact faults than skewness, so kurtosis is given a larger weight. The specific coefficient values ​​can be determined by using a grid search method on the validation set, with the fault diagnosis accuracy as the optimization objective, and searching for the optimal combination within the range of skewness coefficient from 0.3 to 0.5 and kurtosis coefficient from 0.5 to 0.7.

[0042] The non-Gaussian metrics of all time steps are sequentially input into the Sigmoid activation function, mapping them to the interval between 0 and 1, generating a one-dimensional gated vector of length 2048. The value of this gated vector reflects the impact strength of the signal at each location: at the transient location of a strong fault, the non-Gaussian metric is larger, and the Sigmoid output approaches 1; at the location of stationary noise, the non-Gaussian metric is smaller, and the output approaches 0. This one-dimensional gated vector is then broadcast-expanded along the feature dimension, i.e., duplicated and stacked into a 2048×256 matrix, and element-wise Hadamard multiplication is performed with the value matrix to obtain a weighted matrix. This operation amplifies and preserves features rich in impact fault information, while effectively attenuating stationary background noise features. Figure 2 As shown in the figure, the non-Gaussian metric changes with the vibration time sequence. Within the stable vibration range, the non-Gaussian metric maintains a small periodic fluctuation, while a significant peak appears at the time step with significant impact characteristics. The gating value is generally in the range of 0 to 1, and rises synchronously with the increase of the non-Gaussian metric, intuitively presenting the logic of the gating weight generated by the non-Gaussian metric through the activation function mapping.

[0043] The sigmoid activation function was chosen for gating because its output range is strictly limited to 0 to 1, and it has a monotonic, smooth S-shaped curve characteristic within its domain. This allows it to continuously and non-linearly map non-Gaussian metrics to gating coefficients, resulting in gating values ​​close to 1 for time steps with strong impact features and close to 0 for time steps with stable noise. This achieves adaptive soft weighting of features and avoids information loss caused by hard thresholding binarization.

[0044] The query matrix and key matrix corresponding to all data segments are concatenated along the time dimension. Based on this, a block-based composite kernel matrix is ​​constructed to balance the local detail correlations within data segments and the global trend similarity between data segments within the same attention framework. The bandwidth parameter of the kernel function is adaptively determined based on the dominant frequency value of the data segment to which the query vector belongs, in order to adapt to the dynamic changes of different frequency components.

[0045] First, obtain the main frequency value of each data segment. Calculate the local bandwidth parameters of the Gaussian kernel function according to the inverse proportional relationship. and the global bandwidth parameter of the Laplace kernel function In one example, if the local inverse proportional constant coefficient is set to 8500 and the global inverse proportional constant coefficient is set to 4250, then for the main frequency... The data segment is 850Hz. , The inverse proportionality constant coefficient is determined by calculating the distribution range of the dominant frequency values ​​in each data segment of the statistical sample data, combined with the expected bandwidth parameter magnitude. Specifically, assuming the dominant frequency values ​​are typically distributed in the range of hundreds to thousands of Hz, the expected local bandwidth parameter... For frequencies in the range of 1 to 10, with a bandwidth parameter of approximately 8.5 corresponding to a main frequency of 1000Hz and approximately 17 corresponding to a main frequency of 500Hz as references, the local inverse proportionality constant coefficient is determined to be 8500 through linear fitting. Similarly, the global inverse proportionality constant coefficient of 4250 is determined according to a similar principle based on the expected global bandwidth parameter range. Higher main frequencies correspond to smaller bandwidth parameters, resulting in a more concentrated response range for the kernel function, thus adaptively adjusting the correlation measurement scale for signals in different frequency bands.

[0046] For query vectors that belong to the same data segment With key vector Calculate the Euclidean distance between the two and substitute it into the local bandwidth parameter of the data segment. The Gaussian kernel function is constructed, and the local correlation degree is obtained. The formula is expressed as:

[0047] In the formula, For the first section The query vector and the first The local correlation degree values ​​corresponding to each key vector; For the first The query vector corresponding to each time step. For the first The key vector corresponding to each time step; Represents the query vector With key vector The Euclidean norm between them; To query the local bandwidth parameter of the data segment containing the vector, it is calculated from the dominant frequency value of the data segment through an inverse proportional mapping. The combination of local correlation degrees of all vector pairs within the segment constitutes the diagonal blocks on the main diagonal of the block composite kernel matrix.

[0048] For query vectors belonging to different data segments With key vector Calculate the Manhattan distance between the two and substitute it into the global bandwidth parameter of the data segment where the query vector is located. The global similarity is obtained by constructing the Laplace kernel function, and the formula is expressed as follows:

[0049] In the formula, For the first section The query vector and the first The global similarity values ​​corresponding to each key vector; Represents the query vector With key vector The Manhattan distance between them is the sum of the absolute differences in each dimension. To query the global bandwidth parameter of the data segment containing the vector, it is calculated from the dominant frequency value of that data segment through an inverse proportional mapping. The global similarity combination of all inter-segment vector pairs constitutes the off-diagonal blocks of the block composite kernel matrix.

[0050] Finally, the diagonal and off-diagonal blocks are concatenated according to the time step order of the data points to generate a complete block composite kernel matrix that simultaneously contains local details and global trends. This matrix uses a Gaussian kernel within segments to capture smooth local variations, and a Laplace kernel between segments to capture long-range global similarities. Furthermore, the correlation response range of different frequency bands is adjusted by a bandwidth parameter that is adaptive to the dominant frequency.

[0051] S3. Combine sparse mask matrix features to obtain prediction results.

[0052] This step mainly involves generating sparse masks based on dynamic angle thresholds, fusing kernelized attention matrices, and finally predicting fault classification.

[0053] To suppress redundant feature associations, highlight key information, and reduce computational load, this embodiment introduces a dynamic sparse masking mechanism based on feature angles and non-Gaussian metrics.

[0054] First, calculate any query vector in the query matrix and key matrix. and key vector The characteristic angle between The calculation formula is:

[0055] In the formula, For the first The query vector and the first The angle between each key vector in the feature space, in radians; This represents the dot product of the query vector and the key vector. and Let L2 norms be the L2 norms of the two vectors respectively; This is the inverse cosine function, used to convert cosine similarity into an angle value. This feature angle reflects the directional difference between two vectors in a high-dimensional feature space; the smaller the angle, the higher the similarity between the two.

[0056] Then, an angle threshold is dynamically generated based on the non-Gaussian metric at the corresponding time step to determine the strength of the association. For the query vector... and key vector At each time step, obtain the non-Gaussian metric value for each vector, and calculate the arithmetic mean of the two values ​​as the joint non-Gaussian feature of the vector pair. For example, if the non-Gaussian metric of the query vector is 2.4 and the non-Gaussian metric of the key vector is 1.8, then the joint non-Gaussian feature is 2.1. A preset base angle threshold is set, for example, taking... The candidate threshold is obtained by multiplying the base angle threshold by the scaling factor (approximately 0.523 radians) and the joint non-Gaussian features by the scaling factor, for example, by a scaling factor of 0.15.

[0057] Basic angle threshold The radian selection is based on the following: in the feature space, the included angle is less than... The two vectors are typically considered to have significant directional consistency, which is used as the benchmark criterion under non-impact interference conditions. The selection principle for the constant scaling factor 0.15 is: when the joint non-Gaussian features vary within a typical range, the adjustment range of the included angle threshold remains within a reasonable range, avoiding excessive shift of the threshold due to fluctuations in the non-Gaussian metric. The above two parameters can be determined by performing parameter sensitivity analysis on the validation set. Boundary constraints are applied to the candidate threshold, limiting it to between the preset minimum included angle threshold and the maximum included angle threshold, where the minimum included angle threshold can be set to... The maximum included angle threshold can be set to radians (approximately 0.261 radians). Radius (approximately 1.047 radians); minimum included angle threshold The purpose of setting the radian is to prevent the attention weight matrix from becoming too sparse due to an excessively small threshold, thus losing effective information. The maximum included angle threshold... The purpose of setting the radian is to prevent the sparse constraint from failing and too much noise from being introduced due to an excessively large threshold; the result after output constraint is used as the final included angle threshold.

[0058] Next, a sparse mask matrix is ​​constructed using the included angle threshold. For each pair of query vectors and key vectors, their feature angle is calculated. And compare it with the corresponding included angle threshold. If If the angle is less than the included angle threshold, it indicates a significant correlation between the two, and a value of 1 is assigned to the corresponding position in the mask matrix to retain this attention weight; if... If the angle is greater than or equal to the included angle threshold, it indicates a weak correlation between the two vectors or noise interference; therefore, a value of 0 is assigned to the corresponding position, and the weight is reset to zero. This generates a sparse mask matrix that retains only the strongly correlated vectors. For example, for a pair of vectors... If the angle is 0.65 radians, which is less than the included angle threshold of 0.838 radians, then a value of 1 is assigned to the corresponding position in the mask matrix; if the other pair of vectors... If the angle is 1.12 radians, which is greater than the included angle threshold of 0.838 radians, then the corresponding position is assigned a value of 0.

[0059] The kernelized attention matrix is ​​obtained by element-wise multiplying the block-based composite kernel matrix and the sparse mask matrix. This matrix has two advantages: firstly, the block-based composite kernel matrix provides multi-scale capture capabilities for local detail correlations and global long-range trends; secondly, the sparse mask eliminates noise interference and weakly correlated connections, making the attention distribution more focused. As a preferred implementation, the kernelized attention matrix can be normalized row by row so that the sum of the weights in each row is 1, to ensure the scale stability of the output features.

[0060] Next, a weighted summation operation is performed on the weighted matrix using a kernelized attention matrix to aggregate the final deep feature representation. This feature is then flattened along the time dimension and fed into a prediction network composed of a bidirectional long short-term memory network and a multilayer perceptron cascaded together. The bidirectional long short-term memory network captures the forward and backward long-term dependencies of the sequence, while the multilayer perceptron performs further nonlinear transformations and classification mappings on the extracted temporal features. The prediction network ultimately outputs a probability distribution vector with the same dimension as the number of fault categories. The category index corresponding to the maximum probability value is taken as the fault prediction result, which can identify specific states such as mechanical loosening, core jamming, and normal operation.

[0061] To verify the effectiveness of the technical solution in this embodiment, an ablation experiment was conducted based on a real high-voltage circuit breaker vibration signal dataset. The dataset covers five types of state signals: normal operation, mechanical loosening, core jamming, transmission jamming, and base deformation. The experimental hardware platform uniformly used a graphics processing unit (GPU) for model training and inference. The model input sampling frequency was 10kHz, the window length was 2048 sampling points, the optimizer was AdamW, the initial learning rate was set to 0.001, the batch size was set to 64, and a total of 100 training cycles were performed. The evaluation metrics used were fault diagnosis accuracy, F1 score, and single-sample inference latency.

[0062] Five models were set up for comparative analysis in the experiment. The baseline model only uses absolute temporal position encoding and the standard global self-attention mechanism, achieving an accuracy of 88.5%, an F1 score of 87.2%, and an inference latency of 45ms. The first variant model, which removes the temporal-frequency dual-component position encoding and degenerates into a baseline temporal encoding, achieved an accuracy of 93.2%, an F1 score of 92.5%, and a latency of 32ms, showing a slight decrease in performance compared to the complete model. The second variant model, which removes the gated matrix weighting module based on non-Gaussian metrics, achieved an accuracy of 94.1%, an F1 score of 93.6%, and a latency of 32ms, also lower than the complete model. The third variant model, which removes the block composite kernel matrix and the sparse corner masking mechanism, achieved an accuracy of 95.8% and an F1 score of 95.2%, but the inference latency increased to 48ms. The complete model in this embodiment integrates all the above technical features, achieving an accuracy of 97.6% and an F1 score of 97.3%, with an inference latency reduced to 30ms, making it the best performing model among all comparison models. Figure 3 As shown, the trends of the three performance indicators of each group of models are consistent with the test results above, which can intuitively reflect the comprehensive advantages of the complete technical solution in fault identification effect and inference efficiency.

[0063] This invention also discloses a feature-based mechanical fault prediction system for pole-mounted circuit breakers, comprising a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement a feature-based mechanical fault prediction method for pole-mounted circuit breakers according to the present invention.

[0064] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting mechanical faults in pole-mounted circuit breakers based on feature analysis, characterized in that, include: S1. The acquired circuit breaker vibration time-series data is aligned and segmented using a synchronous trigger signal, and the sampling points in each data segment are mapped to a data embedding vector sequence. The frequency domain dominant frequency feature of the data segment is extracted and combined with the time coding vector to generate a position code, which is then superimposed on the data embedding vector. A query matrix, a key matrix, and a value matrix are generated through linear transformation. S2. The higher-order statistical moments of the data embedding vector within a local time window are extracted to generate a non-Gaussian metric, and a gating matrix is ​​constructed accordingly. The gating matrix is ​​then fused with the value matrix to obtain a weighted value matrix. Using the first kernel function and the second kernel function, the local correlation degree and global similarity between the query vector and the key vector in the query matrix and the key matrix are calculated respectively, and a block composite kernel matrix with bandwidth parameter adjusted according to the frequency domain main frequency feature is constructed; S3, the feature angle between the query vector and the key vector is calculated, the angle threshold is determined based on the non-Gaussian metric at the corresponding time step, and a sparse mask matrix is ​​generated according to the comparison result between the feature angle and the angle threshold; The block composite kernel matrix and the sparse mask matrix are fused into a kernelized attention matrix. The kernelized attention matrix is ​​then used to weight the weighted matrix to extract features, which are then input into the prediction network to obtain the circuit breaker fault prediction result.

2. The method for predicting mechanical faults of pole-mounted circuit breakers based on feature analysis according to claim 1, characterized in that, Extracting the frequency domain dominant frequency feature of the data segment and combining it with the time coding vector to generate a position code, and then superimposing it on the data embedding vector, includes: performing a frequency domain transformation on the data segment to obtain an amplitude spectrum, and extracting the frequency component with the largest amplitude as the frequency domain dominant frequency feature; A frequency coding vector is generated based on the frequency domain dominant frequency feature, and then combined with a time coding vector generated based on the time step sequence number to generate the position code. The position code is added element by element to the corresponding data embedding vector to obtain the superimposed data embedding vector.

3. The method for predicting mechanical faults of pole-mounted circuit breakers based on feature analysis according to claim 1, characterized in that, Extracting the higher-order statistical moments of the data embedding vector within a local time window to generate a non-Gaussian metric, and constructing a gating matrix accordingly, and fusing the gating matrix with the value matrix to obtain a weighted value matrix, including: calculating the mean absolute value of the third-order standard moments of each feature dimension of the data embedding vector within the local time window as skewness, and the mean of the fourth-order standard moments as kurtosis; The non-Gaussian metric is obtained by weighted summation of the skewness and the kurtosis. The non-Gaussian metric on the time series is mapped to a preset interval through a preset activation function to generate a one-dimensional nonlinear gated vector. The nonlinear gated vector is then expanded in the feature dimension and multiplied element-wise with the value matrix to generate the weighted value matrix.

4. The method for predicting mechanical faults of pole-mounted circuit breakers based on feature analysis according to claim 1, characterized in that, The step involves using a first kernel function and a second kernel function to calculate the local correlation and global similarity between query vectors and key vectors in the query matrix and key matrix, respectively, and constructing a block-based composite kernel matrix with bandwidth parameters adjusted according to the frequency domain dominant frequency characteristics. This includes: calculating local bandwidth parameters and global bandwidth parameters based on the frequency domain dominant frequency characteristics; for query vectors and key vectors belonging to the same data segment, calculating the Euclidean distance and substituting it into a Gaussian kernel function with local bandwidth parameters as the first kernel function to obtain the local correlation, and constructing a diagonal block; for query vectors and key vectors belonging to different data segments, calculating the Manhattan distance and substituting it into a Laplace kernel function with global bandwidth parameters as the second kernel function to obtain the global similarity, and constructing an off-diagonal block; and concatenating the diagonal and off-diagonal blocks in chronological order to generate the block-based composite kernel matrix.

5. The method for predicting mechanical faults of pole-mounted circuit breakers based on feature analysis according to claim 1, characterized in that, Generate a sparse mask matrix based on the comparison result of the feature angle and the angle threshold, including: obtaining the feature angle between the query vector and the key vector; If the feature angle is less than the angle threshold, a preset retention weight value is assigned to the corresponding position in the sparse mask matrix. If the feature angle is greater than or equal to the angle threshold, then a zero value is assigned to the corresponding position in the sparse mask matrix to generate the sparse mask matrix.

6. The method for predicting mechanical faults of pole-mounted circuit breakers based on feature analysis according to claim 1, characterized in that, The determination of the included angle threshold based on the non-Gaussian metric at the corresponding time step includes: obtaining the non-Gaussian metrics of the query vector and the key vector at the corresponding time step, and calculating the arithmetic mean of the two as a joint non-Gaussian feature; Obtain a preset base angle threshold, and add the product of the joint non-Gaussian feature and the preset constant scaling factor to the base angle threshold to obtain the adjusted candidate threshold. The candidate threshold is subject to boundary constraints, which limit it to a preset range between the minimum and maximum included angle thresholds, and the included angle threshold is then output.

7. The method for predicting mechanical faults of pole-mounted circuit breakers based on feature analysis according to claim 1, characterized in that, The step of generating the query matrix, key matrix, and value matrix through linear transformation includes: initializing mutually independent learnable query weight matrix, key weight matrix, and value weight matrix, and setting the corresponding initial bias vector; The data embedding vector after superimposing the position encoding is multiplied by the query weight matrix, key weight matrix and value weight matrix respectively, and the corresponding initial bias vector is added for affine transformation to generate a query matrix, key matrix and value matrix with the same feature dimensions as the data embedding vector.

8. The method for predicting mechanical faults of pole-mounted circuit breakers based on feature analysis according to claim 1, characterized in that, The step of aligning and segmenting the acquired circuit breaker vibration time-series data using a synchronous trigger signal and mapping the sampling points in each data segment into a data embedding vector sequence includes: acquiring circuit breaker vibration time-series data and the trip coil current signal as the synchronous trigger signal; calculating the sequence difference using the rising edge of the trip coil current signal; and locating the maximum difference point as the alignment zero point to complete the alignment segmentation. The aligned and segmented circuit breaker vibration time series data is divided into multiple data segments using a sliding window, and local vibration features are extracted using a one-dimensional convolutional network. The one-dimensional sampling points are then mapped to a high-dimensional space to form the data embedding vector.

9. The method for predicting mechanical faults of pole-mounted circuit breakers based on feature analysis according to claim 1, characterized in that, The weighted matrix is ​​weighted using the kernelized attention matrix to extract features, which are then input into the prediction network to obtain circuit breaker fault prediction results, including: The weighted extracted features are flattened along the time dimension and input into a prediction network consisting of a bidirectional long short-term memory network and a multilayer perceptron cascade. By predicting the probability distribution vector of different mechanical states of the circuit breaker output by the network, the category index corresponding to the maximum probability value is extracted to determine the mechanical fault type, which is used as the circuit breaker fault prediction result.

10. A feature-based mechanical fault prediction system for pole-mounted circuit breakers, characterized in that, include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement the feature analysis-based method for predicting mechanical faults in pole-mounted circuit breakers according to any one of claims 1-9.

Images