High-voltage circuit breaker fault identification method, system, equipment and medium

Abstract

The invention relates to the technical field of high-voltage circuit breaker fault identification, and discloses a high-voltage circuit breaker fault identification method, system and device and a medium, and the method comprises the steps: converting a multi-state vibration signal of a high-voltage circuit breaker into a multi-state signal time-frequency diagram; performing data expansion on the multi-state signal time-frequency diagram, and dividing the multi-state signal time-frequency diagram into a training set and a test set; extracting a time-frequency graph feature vector of the training set, and performing parameter optimization on the fault classifier according to the time-frequency graph feature vector to obtain an optimized classification model; calculating verification model parameters according to the optimization classification model; and constructing a target fault prediction model according to the verification model parameters, performing fault prediction on the to-be-diagnosed high-voltage circuit breaker signal by using a target fault diagnosis model, and calculating a corresponding feature importance thermodynamic diagram and a target fault evaluation factor. According to the method, the accuracy of target fault category prediction can be improved, the fault reason is directly positioned through the feature importance thermodynamic diagram, and accurate identification of high-voltage circuit breaker fault detection is realized.

Description

Technical Field

[0001] This invention relates to the field of high-voltage circuit breaker fault identification technology, specifically a high-voltage circuit breaker fault identification method, system, equipment, and medium. Background Technology

[0002] High-voltage circuit breakers (HVCBs) are core protection and control devices in power systems. Mechanical faults (such as stuck operating mechanisms, spring fatigue, and loose bases) are the main components of faults, and timely and accurate fault diagnosis is crucial to ensuring the safety of the power grid.

[0003] Existing fault diagnosis methods for high-voltage circuit breakers mainly revolve around three stages: "signal acquisition - feature extraction - fault identification." However, three core problems still exist in practical applications: First, the vibration signals of high-voltage circuit breakers are nonlinear, non-stationary, and transient, making signal feature extraction difficult. For example, during the closing process, the vibration signals generated by coil energization, spring energy storage, and contact collisions are superimposed and affected by on-site electromagnetic interference and temperature and humidity changes, resulting in a low signal-to-noise ratio. Traditional methods (such as wavelet packet decomposition and empirical mode decomposition) rely on manually designed features (such as energy entropy and peak factor), which are difficult to capture complex time-frequency domain correlation features. For example, the vibration signal of spring fatigue faults has dispersed energy in the low-frequency range (0-20kHz), but the manually extracted "low-frequency energy ratio" feature is easily confused with the base loosening fault, resulting in low identification accuracy.

[0004] Secondly, with the application of deep learning technology, Convolutional Neural Networks (CNN) and Vision Transformers (VT) have become mainstream methods: CNN extracts time-frequency texture features through local convolutional kernels, but has weak global feature capture capabilities (such as being unable to correlate the causal relationship between "high-frequency vibration during contact closure" and "coil undervoltage"); VT captures long-range dependencies through self-attention mechanisms, but has a "black box" defect, which cannot explain why the model determines the fault category, making it difficult for maintenance personnel to trust the diagnostic results, and even more difficult to trace the fault mechanism based on the results (such as the correlation between spring pre-compression and vibration characteristics).

[0005] Finally, the cost of collecting fault samples for high-voltage circuit breakers is high (it requires simulating fault conditions, which may affect equipment lifespan), and in actual engineering, there are usually fewer than 100 sets of samples for each type of fault. Traditional deep learning models (such as CNN and VT) require large-scale data training and are prone to overfitting with small samples. While data augmentation methods (such as pruning and flipping) can expand the dataset, existing methods do not incorporate the timing of circuit breaker actions (such as the division of the closing stage). The augmented data may lose key fault features, resulting in limited improvement in the model's generalization ability.

[0006] In summary, how can we address the problem of low accuracy in high-voltage circuit breaker fault detection caused by the difficulty of feature extraction, poor interpretability of results, and limited improvement in model generalization in existing technologies? Summary of the Invention

[0007] The technical problem to be solved by this invention is how to accurately identify faults in high-voltage circuit breakers.

[0008] The present invention solves the above-mentioned technical problems through the following technical means: This invention provides a method for fault identification of high-voltage circuit breakers, comprising: Acquire multi-state vibration signals of the high-voltage circuit breaker and convert the multi-state vibration signals into a multi-state signal time-frequency diagram; The time-frequency graph of the multi-state signal is extended to obtain an extended dataset, which is then divided into a training set and a test set. The time-frequency graph feature vector of the training set is extracted using a pre-built dual-branch feature extraction model. The parameters of the pre-built fault classifier are optimized based on the time-frequency graph feature vector to obtain an optimized classification model. Calculate the multi-dimensional validation parameters of the test set based on the optimized classification model, and determine the validation model parameters based on the multi-dimensional validation parameters. Based on the verification model parameters, a target fault prediction model is constructed. The target fault diagnosis model is then used to predict the fault of the high-voltage circuit breaker signal to be diagnosed, obtain the target fault category, and calculate the feature importance heatmap and target fault evaluation factor corresponding to the high-voltage circuit breaker signal to be diagnosed.

[0009] Optionally, converting the multi-state vibration signal into a multi-state signal time-frequency diagram includes: The wavelet transform coefficients are obtained by performing continuous wavelet transform on each vibration signal in the multi-state vibration signal using a preset wavelet function. Construct a time-frequency diagram of each vibration signal based on the wavelet transform coefficients; By compiling the time-frequency diagrams of the signals, a multi-state signal time-frequency diagram corresponding to the multi-state vibration signal is obtained.

[0010] Optionally, the step of data augmentation of the time-frequency graph of the multi-state signal to obtain an augmented dataset includes: Adversarial learning is performed on each signal time-frequency plot in the multi-state signal time-frequency plot to obtain an adversarial learning time-frequency plot; Each of the signal time-frequency maps is cropped at multiple scales to obtain the cropped time-frequency map corresponding to the signal time-frequency map; The multi-state signal time-frequency diagram is mixed and masked to obtain the mixed time-frequency diagram corresponding to the signal time-frequency diagram; The extended dataset is obtained by combining the multi-state signal time-frequency map, the adversarial learning time-frequency map, the cropped time-frequency map, and the hybrid time-frequency map.

[0011] Optionally, the step of extracting the time-frequency graph feature vector of the training set using a pre-built dual-branch feature extraction model includes: Temporal pruning is performed on each extended data in the training set to obtain a time-frequency plot subsequence; The temporal correlation features of the time-frequency graph subsequences are extracted based on the residual network branches in the pre-trained model; The attention weights in the feature transformation encoder of the pre-trained model are calculated using the temporal correlation features, and the transformation features of each of the extended data are calculated based on the attention weights. The temporal correlation features are flattened to obtain flattened correlation features, and the transformed features are averaged pooled to obtain pooled features. The flat correlation features and the pooling features are concatenated to obtain the time-frequency plot feature vector.

[0012] Optionally, calculating the attention weights in the feature transformation encoder of the pre-trained model using the temporal correlation features includes: The feature transformation encoder is used to embed and encode each of the extended data to obtain an encoded feature vector; The encoded feature vector is linearly mapped to obtain the query matrix, the key matrix, and the value matrix. Attention weights are calculated using the query matrix, the key matrix, the value matrix, and the temporal correlation features.

[0013] Optionally, calculating the multi-dimensional verification parameters of the test set based on the optimized classification model includes: Extract test data features from each extended data in the test set, and extract a feature subset set of the test data features; The optimized classification model is used to predict faults in the feature subset set and the test data features to obtain the probability of the test fault category. The test accuracy of the optimized classification model is calculated based on the probability of the test fault categories. A Shapu value matrix is ​​constructed based on the probability of the test fault categories, and the fault assessment factor of the extended data is calculated based on the Shapu value matrix. Calculate the Pearson correlation coefficient between the fault assessment factor and the preset standard fault assessment factor; By combining the test accuracy, the Shapu value matrix, and the Pearson correlation coefficient, multi-dimensional verification parameters are obtained.

[0014] Optionally, calculating the fault assessment factor of the extended data based on the Shapu value matrix includes: Obtain the state time-frequency diagram under a preset normal state, and construct the normal state average amplitude spectrum matrix based on the state time-frequency diagram; The test fault category is determined based on the probability of the test fault category corresponding to the extended data, and the average amplitude spectrum matrix of the fault state of each of the extended data is constructed based on the test fault category. A comprehensive information matrix for normal state is constructed based on the normal state average amplitude spectrum matrix, and a comprehensive information matrix for fault state is constructed based on the fault state average amplitude spectrum matrix and the Shapu value matrix. The fault assessment factor is calculated based on the normal state comprehensive information matrix and the fault comprehensive information matrix.

[0015] To address the above problems, the present invention also proposes a high-voltage circuit breaker fault identification system, the system comprising: The vibration signal conversion module acquires multi-state vibration signals from a high-voltage circuit breaker and converts the multi-state vibration signals into a multi-state signal time-frequency diagram. The data expansion module applies data expansion to the time-frequency graph of the multi-state signal to obtain an expanded dataset, and divides the expanded dataset into a training set and a test set. The parameter optimization module is used to extract the time-frequency map feature vector of the training set using a pre-built dual-branch feature extraction model, and optimize the parameters of the pre-built fault classifier based on the time-frequency map feature vector to obtain an optimized classification model. The verification model parameter determination module is used to calculate the multi-dimensional verification parameters of the test set based on the optimized classification model, and determine the verification model parameters based on the multi-dimensional verification parameters. The fault prediction module is used to construct a target fault prediction model based on the verification model parameters, use the target fault diagnosis model to predict the fault of the high-voltage circuit breaker signal to be diagnosed, obtain the target fault category, and calculate the feature importance heatmap and target fault evaluation factor corresponding to the high-voltage circuit breaker signal to be diagnosed.

[0016] The present invention also provides a processing device, characterized in that it includes at least one processor and at least one memory communicatively connected to the processor, wherein: the memory stores program instructions executable by the processor, and the processor can execute the above-described method for identifying high-voltage circuit breaker faults by calling the program instructions.

[0017] The present invention also provides a computer-readable storage medium, characterized in that the computer-readable storage medium stores computer instructions, the computer instructions causing the computer to execute the above-described method for identifying high-voltage circuit breaker faults.

[0018] The advantages of this invention are: This invention converts the multi-state vibration signals of high-voltage circuit breakers into multi-state signal time-frequency maps and expands the data in these maps to capture more subtle key features, expand their diversity, and avoid data distribution limitations. The expanded data is then divided into a training set and a test set. The training set is used to optimize the parameters of a pre-built fault classifier, resulting in an optimized classification model. This avoids overfitting and underfitting, effectively improving the accuracy of subsequent fault identification. Multi-dimensional validation model parameters are calculated based on the optimized classification model and used to determine the validation model parameters, further evaluating the model's practical application effect and ensuring its generalization ability. Based on the parameters of the validation model, a target fault prediction model is constructed. The target fault diagnosis model is then used to predict the faults of the high-voltage circuit breaker signals to be diagnosed, thereby obtaining the target fault category. This can improve the accuracy of the target fault category prediction. Furthermore, the characteristic importance heatmap and target fault evaluation factor corresponding to the high-voltage circuit breaker signals to be diagnosed are calculated, which can directly locate the fault cause, improve the interpretability of the target fault diagnosis model, and achieve accurate identification of high-voltage circuit breaker fault detection. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating a high-voltage circuit breaker fault identification method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a process for data expansion of a multi-state signal time-frequency diagram in one embodiment of the present invention; Figure 3 This is a flowchart illustrating a method for extracting time-frequency graph feature vectors from a training set according to an embodiment of the present invention. Figure 4 A functional module diagram of a high-voltage circuit breaker fault identification system provided in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Reference Figure 1 The diagram shown is a flowchart illustrating a high-voltage circuit breaker fault identification method according to an embodiment of the present invention. In this embodiment, the high-voltage circuit breaker fault identification method includes: S1. Obtain the multi-state vibration signal of the high-voltage circuit breaker and convert the multi-state vibration signal into a multi-state signal time-frequency diagram.

[0022] In this embodiment of the invention, the multi-state vibration signal is the closing vibration signal of the high-voltage circuit breaker under different states. For example, the closing vibration signal under normal operation and the closing vibration signal under each fault state can be collected. Each state can collect 200 sets of data.

[0023] For example, methods for simulating fault states of high-voltage circuit breakers include, but are not limited to: simulating link failure by adjusting the distance between the link contact and the bottom; simulating undervoltage failure by reducing the control circuit voltage; simulating base loosening by loosening the base screws; simulating overvoltage failure by increasing the control circuit voltage; simulating mechanism jamming by adding foreign objects to the link shaft; simulating spring fatigue by reducing the spring precompression by preset lengths (e.g., 3mm, 6mm, 9mm); and simulating contact wear by grinding and thinning the moving contact surface (e.g., 0.2mm). Multi-state vibration signals are obtained by collecting signal data under different fault states.

[0024] In detail, a multi-state signal time-frequency diagram is a visualization chart obtained by performing continuous wavelet transform on each vibration signal in a multi-state vibration signal. In this diagram, the horizontal axis represents time, the vertical axis represents frequency, and the color intensity represents the energy of the frequency component at the corresponding moment.

[0025] Specifically, converting the multi-state vibration signal into a multi-state signal time-frequency diagram includes: The wavelet transform coefficients are obtained by performing continuous wavelet transform on each vibration signal in the multi-state vibration signal using a preset wavelet function. Construct a time-frequency diagram of each vibration signal based on the wavelet transform coefficients; By compiling the time-frequency diagrams of the signals, a multi-state signal time-frequency diagram corresponding to the multi-state vibration signal is obtained.

[0026] In this embodiment of the invention, the wavelet function is the mother wavelet function for wavelet transform. The wavelet function is used to convolve each vibration signal to obtain the wavelet transform coefficients corresponding to each vibration signal.

[0027] Specifically, the wavelet function can be a generalized Morse wavelet, and the wavelet function is expressed as:

[0028] in, Describing the wavelet function, Represents the imaginary unit. This indicates the preset frequency bandwidth parameter. This represents the preset symmetry parameter. This represents the modified Bessel function of the first kind. Represents the natural constant. Represents a time variable.

[0029] Furthermore, the wavelet transform coefficients are obtained by performing a continuous wavelet transform on each vibration signal in the multi-state vibration signal using the following formula:

[0030] in, Represents the wavelet transform coefficients. This represents the preset scaling parameter of the wavelet function. Represents variables over time Changing vibration signals, Represents the conjugate of wavelet functions. This represents the preset translation parameters of the wavelet function. A differential element representing a time variable.

[0031] In detail, the scaling parameter can affect the wavelet width of the miniscale function and the frequency characteristics of the corresponding signal, while the translation parameter can affect the wavelet width of the miniscale function and the frequency characteristics of the corresponding signal.

[0032] In this embodiment of the invention, the wavelet transform coefficients are a two-dimensional matrix composed of specific time variables and wavelet transform scales. The number of columns in the wavelet transform coefficient matrix is ​​perfectly matched with the number of sampling points of the vibration signal, ensuring a one-to-one correspondence in the time dimension.

[0033] Furthermore, the amplitude of the wavelet coefficient corresponding to each (time-frequency) position in the wavelet transform coefficients is mapped to the gray value or color depth of the image, thereby constructing a signal time-frequency diagram for each vibration signal, which can intuitively represent the distribution of the signal's time and frequency.

[0034] For example, the horizontal axis of the signal time-frequency graph covers the entire circuit breaker closing cycle from 0 to 0.15 seconds, and the time period can be divided into the coil energizing stage. Spring energy storage stage Contact closing stage Braking phase The last four sub-stages cover a frequency range of 0-60kHz on the vertical axis, and are divided into low, medium and high frequency bands according to 0-20kHz, 20-40kHz and 40-60kHz; the color depth corresponds to the amplitude of the wavelet coefficients, reflecting the magnitude of the signal energy.

[0035] S2. Perform data expansion on the time-frequency graph of the multi-state signal to obtain an expanded dataset, and divide the expanded dataset into a training set and a test set. In this embodiment of the invention, data expansion involves augmenting the time-frequency graph of a multi-state signal to form a richer, enhanced dataset.

[0036] Specifically, refer to Figure 2 As shown, the data extension of the time-frequency graph of the multi-state signal to obtain the extended dataset includes: S21. Perform adversarial learning on each signal time-frequency diagram in the multi-state signal time-frequency diagram to obtain an adversarial learning time-frequency diagram; S22. Perform multi-scale cropping on each of the signal time-frequency maps to obtain the cropped time-frequency map corresponding to the signal time-frequency map; S23. Perform a mixing mask on the multi-state signal time-frequency diagram to obtain the mixed time-frequency diagram corresponding to the signal time-frequency diagram; S24. The multi-state signal time-frequency map, the adversarial learning time-frequency map, the cropped time-frequency map, and the hybrid time-frequency map are combined to obtain the extended dataset.

[0037] In this embodiment of the invention, adversarial learning utilizes an unsupervised pre-trained exemplar discrimination method to expand the information in each time-frequency plot. Specifically, the unique time-frequency texture of the signal time-frequency plots for different fault states can be learned through an unsupervised pre-trained model.

[0038] In detail, a slightly transformed version (such as a small time axis shift or a slight frequency amplitude adjustment) is generated for each signal time-frequency plot. This transformed version, along with examples of signals in the same state from a pre-defined example library, is used as a positive sample. The transformed version, along with examples of signals in different states from the example library, is used as a negative sample. For example, the transformed version of the signal time-frequency plot under normal operating conditions, along with the transformed version of the time-frequency plot under normal operating conditions from the example library, are positive samples, while the transformed version, along with the transformed version under fault conditions, are negative samples. These positive and negative samples are input into a pre-built bi-branch feature extraction model. The Transformer branch in the model learns to extract feature vectors from the time-frequency plot. The loss value is calculated using a pre-trained loss function. The Transformer branch can reduce the feature distance between each signal time-frequency plot and the positive samples, and increase the feature distance with the negative samples. Therefore, through adversarial learning, subtle feature differences in the signal time-frequency plot can be learned, expanding the data feature space that the subsequent bi-branch feature extraction model can learn.

[0039] The loss value can be calculated using the loss function shown below:

[0040] in, Indicates the loss value. Indicates the first Time-frequency diagram of a signal, This represents the total number of time-frequency diagrams of the signal. Indicates the first Positive samples corresponding to the time-frequency plot of each signal Indicates the first One negative sample, This represents the cosine similarity function.

[0041] In another embodiment of the present invention, multi-scale cropping of the signal time-frequency map involves cropping each signal time-frequency map at different resolution scales, such as 16×16, 32×32, or 64×64 pixels, such that each cropped time-frequency map contains at least one key stage feature of the signal time-frequency map.

[0042] Furthermore, the mixing mask is used to mix the time-frequency diagrams of signals with similar fault states using a binary mask matrix. The similar fault states can be preset, for example, undervoltage faults and overvoltage faults are similar fault states.

[0043] Specifically, the time-frequency diagram of the multi-state signal can be masked using the following formula:

[0044] in, Represents a mixed time-frequency diagram. This represents a preset binary mask matrix. Indicates the first Signal time-frequency diagram under fault-like conditions This represents the matrix multiplication operation. Indicates the first Time-frequency diagram of signals similar to those in similar fault states.

[0045] Preferably, after mixing the mask, the image labels for the mixed time-frequency graph are as follows:

[0046] in, This indicates the image label corresponding to the mixed time-frequency plot. This indicates the preset mixing ratio. , These represent the time-frequency diagrams of the signal, respectively. , The corresponding image tags.

[0047] In this embodiment of the invention, the fault category corresponding to each signal time-frequency map can be used as an image label, and the signal time-frequency maps of similar fault categories can be masked and mixed according to a preset mixing ratio (such as 0.3) to obtain a mixed time-frequency map.

[0048] In this embodiment of the invention, data expansion can capture more subtle key features in the signal time-frequency graph, expand the diversity of the signal time-frequency graph, avoid the limitations of data distribution, and effectively improve the robustness, generalization ability and feature mining efficiency of the subsequent dual-branch feature extraction model.

[0049] Furthermore, the extended dataset is divided into a training set and a test set using a pre-defined partitioning ratio. For example, a 4:1 partitioning ratio yields a training set for training and a test set for calculating multi-dimensional validation parameters.

[0050] In this embodiment of the invention, dividing the extended dataset into a training set and a test set can evaluate the generalization ability of the optimized feature extraction model, avoid overfitting and underfitting, and effectively improve the accuracy of subsequent fault diagnosis.

[0051] S3. Extract the time-frequency map feature vector of the training set using the pre-built dual-branch feature extraction model, and optimize the parameters of the pre-built fault classifier based on the time-frequency map feature vector to obtain an optimized classification model.

[0052] In this embodiment of the invention, the pre-built dual-branch feature extraction model includes a Transformer branch and a Residual Network (ResNet) branch.

[0053] The Transformer branch includes multiple feature extraction stages, each consisting of a block embedding layer and a TF-E (Transformer encoder) component. Different block embedding parameters and Transformer encoder parameters result in different dimensions of the output feature map. For example, in stage 1, the block embedding parameter is set to 4, corresponding to a 4×4 feature map size and 64 output channels, transforming a 256×256 time-frequency map into a 64×64×64 feature map. In stage 2, the block embedding parameter is set to 2, and the output channels are set to 128, transforming the feature map output from stage 1 into a 32×32×128 feature map. In stage 3, the block embedding parameter is set to 2, and the output channels are set to 256, transforming the feature map output from stage 2 into a 16×16×256 feature map. In stage 4, the block embedding parameter is set to 2, and the output channels are set to 512, transforming the feature map output from stage 3 into an 8×8×512 feature map. Furthermore, the TF-E (Transformer Encoder) component is constructed by stacking multiple sub-layers of self-attention and feed-forward neural networks, with residual links and layer normalization added after each sub-layer, to extract the transformation features of each extended data in the input extended dataset.

[0054] In this process, each extended data point is positionally encoded to generate a feature vector. An attention mechanism is then used to perform multi-head attention calculation on the feature vector to generate attention features. Finally, a feedforward neural network is used to perform local nonlinear optimization on the attention features to supplement the detailed information of the attention features.

[0055] Preferably, the local features extracted by the ResNet branch can be used as the attention bias of the attention mechanism to update the attention weights.

[0056] In this embodiment of the invention, the ResNet branch sequentially includes convolutional layers, max pooling layers, multiple residual blocks, global average pooling layers, and fully connected layers. For example, the convolutional layers can use 7×7 convolutional kernels with a stride of 2 and 64 output channels; the max pooling layers use 3×3 pooling kernels with a stride of 2; each of the four residual blocks contains two 3×3 convolutional layers with channels of 64, 128, 256, and 512 respectively; and the fully connected layers have 512 output channels.

[0057] Preferably, before extracting the time-frequency map feature vector of the training set using the pre-built dual-branch feature extraction model, the Transformer branch can be pre-trained using the aforementioned method of generating adversarial learning time-frequency maps, for example, with 100 pre-training rounds. The ResNet branch can be pre-trained using a preset dataset, with parameter optimization performed according to the stochastic gradient descent algorithm, where the hyperparameters can be: learning rate set to 1e-3, and momentum set to 0.9.

[0058] Specifically, refer to Figure 3 As shown, the step of extracting the time-frequency map feature vector of the training set using a pre-constructed dual-branch feature extraction model includes: S31. Perform time-series pruning on each extended data in the training set to obtain a time-frequency graph subsequence; S32. Extract the temporal correlation features of the time-frequency graph subsequence based on the residual network branches in the pre-trained model; S33. Calculate the attention weights in the feature transformation encoder of the pre-trained model using the temporal correlation features, and calculate the transformation features of each of the extended data based on the attention weights; S34. Flatten the temporal correlation features to obtain flattened correlation features, and perform average pooling on the transformation features to obtain pooled features. S35. Perform feature concatenation on the flat correlation features and the pooling features to obtain the time-frequency map feature vector.

[0059] In this embodiment of the invention, the time-series trimming of the extended data is performed according to the operation stages of the high-voltage circuit breaker. For example, the extended data is segmented into time-frequency diagram subsequences according to four stages: coil energizing stage, spring energy storage stage, contact closing stage, and braking stage.

[0060] Furthermore, the ResNet branch processes the time-frequency map subsequence sequentially through convolution, max pooling, residual blocks, global average pooling layers, and fully connected layers to obtain temporal correlation features.

[0061] In this embodiment of the invention, the temporal correlation features are used as the attention bias of the attention mechanism in the feature transformation encoder, so that the attention weight can be calculated based on the temporal correlation features.

[0062] Specifically, the step of calculating the attention weights in the feature transformation encoder of the pre-trained model using the temporal correlation features includes: The feature transformation encoder is used to embed and encode each of the extended data to obtain an encoded feature vector; The encoded feature vector is linearly mapped to obtain the query matrix, the key matrix, and the value matrix. Attention weights are calculated using the query matrix, the key matrix, the value matrix, and the temporal correlation features.

[0063] The attention weights are calculated using the following formula:

[0064] in, Indicates attention weights. express Activation function Represents the query matrix. Represents the key matrix, Represents a value matrix, This represents the preset weight parameters. Indicates temporal correlation features, Indicates transpose. Representing the query matrix The matrix dimension.

[0065] Furthermore, in each stage of the Transformer branch, the value vector (V) is weighted and aggregated according to the attention weight to obtain the attention features. Each stage then uses a feedforward neural network to perform linear transformation and nonlinear activation on the attention features to obtain the feature vector of each stage. By sequentially passing through multiple feature extraction stages, the transformation features of each extended data are obtained.

[0066] In this embodiment of the invention, flattening the temporal correlation features involves converting the temporal correlation features into a one-dimensional feature vector, and average pooling involves converting the converted features into a one-dimensional feature vector, thus obtaining one-dimensional flattened correlation features and pooled features. The flattened correlation features and pooled features are then concatenated to obtain a fused time-frequency map feature vector.

[0067] Furthermore, the fault classifier is a Softmax classifier based on the Softmax function. The Softmax classifier uses the time-frequency map feature vector to predict the high-voltage circuit breaker faults corresponding to the signal time-frequency map, and obtains the probability of each high-voltage circuit breaker fault. The parameters of the Softmax classifier are optimized based on the predicted probabilities to obtain an optimized feature extraction model with more accurate predictions.

[0068] Specifically, the step of optimizing the parameters of the pre-constructed fault classifier based on the time-frequency map feature vector to obtain an optimized classification model includes: The fault classifier is used to predict the fault probability by analyzing the time-frequency graph feature vector. Calculate the cross-entropy loss value of the fault classifier based on the fault probability; The parameters of the fault classifier are optimized based on the cross-entropy loss value to obtain an optimized classification model.

[0069] In this embodiment of the invention, a fault classifier is used to linearly transform each time-frequency map feature vector to calculate the original score of each signal time-frequency map belonging to each fault category. Then, the probability of the original score is calculated by normalizing it using the Softmax function, thus obtaining the probability of the time-frequency map feature vector being classified into each fault category.

[0070] Furthermore, the cross-entropy loss value of the fault classifier is calculated using the following formula:

[0071] in, This represents the cross-entropy loss value. Indicates the first The true fault category label corresponding to each time-frequency plot feature vector , This indicates the total number of fault category labels. This represents the total number of eigenvectors in the time-frequency plot. Indicates the first Each time-frequency plot feature vector is classified into fault category labels. The probability of failure.

[0072] Each fault category is individually encoded to obtain a label for each fault category. The actual fault category label can be the fault state corresponding to the collected closing vibration signal. For example, if the time-frequency graph feature vector X is obtained by extracting the time-frequency graph feature vector from the vibration signal collected under the undervoltage fault state, then the actual fault category label corresponding to the time-frequency graph feature vector X is... This is a unique thermal encoding for undervoltage fault conditions.

[0073] Furthermore, the parameters of the fault classifier are optimized using the cross-entropy loss value until the cross-entropy loss value is less than a preset loss threshold, thus obtaining an optimized classification model. Specifically, the Adam W (Adam - Weight Decay) optimizer, with a learning rate of 1e-4, weight decay of 1e-5, and 200 training epochs, can be used to directly decay the parameters in the fault classifier, resulting in an optimized classification model.

[0074] In this embodiment of the invention, by pre-training the dual-branch feature extraction model between extracting time-frequency map feature vectors and optimizing the fault classifier, the dual-branch feature extraction model can learn the underlying general features of a large amount of training data, reducing its dependence on feature scene data and improving training efficiency. Furthermore, by specifically optimizing the parameters of the fault classifier using the training set, it can better adapt to fault category prediction and improve the accuracy of subsequent target fault category calculation.

[0075] S4. Calculate the multi-dimensional verification parameters of the test set based on the optimized classification model, and determine the verification model parameters based on the multi-dimensional verification parameters.

[0076] In this embodiment of the invention, the multi-dimensional verification parameters are used to calculate the accuracy of the optimized feature extraction model from different dimensions. The first dimension is the average recognition rate and confusion matrix of the test set, ensuring that the average recognition rate is not lower than a first threshold (99.2%). Simultaneously, the cross-recognition rate of similar faults such as undervoltage and overvoltage faults is verified, ensuring that the cross-recognition rate is not lower than a second threshold (95%). The second dimension is the interpretability verification of the optimized classification model, including evaluating the consistency of SHAP (SHapley Additive exPlanations), for example, ensuring that the overlap rate of key features in 10 tests for the same fault category is not lower than 80%; it also includes evaluating the correlation of fault assessment factors, calculating the Pearson correlation coefficient between the fault assessment factor value and the preset fault severity, ensuring that the correlation coefficient is not lower than a second threshold (e.g., 0.85). Specifically, the step of calculating the multi-dimensional verification parameters of the test set based on the optimized classification model includes: Extract test data features from each extended data in the test set, and extract a feature subset set of the test data features; The optimized classification model is used to predict faults in the feature subset set and the test data features to obtain the probability of the test fault category. The test accuracy of the optimized classification model is calculated based on the probability of the test fault categories. A Shapu value matrix is ​​constructed based on the probability of the test fault categories, and the fault assessment factor of the extended data is calculated based on the Shapu value matrix. Calculate the Pearson correlation coefficient between the fault assessment factor and the preset standard fault assessment factor; By combining the test accuracy, the Shapu value matrix, and the Pearson correlation coefficient, multi-dimensional verification parameters are obtained.

[0077] In this embodiment of the invention, the test data features can be extracted using the above-mentioned dual-branch feature extraction model, including features extracted by the Transformer branch and the Residual Network (ResNet) branch. The second transformation feature extracted by the Transformer branch and the second temporal correlation feature extracted by the Residual Network (ResNet) branch are converted into a one-dimensional vector and then concatenated to obtain the test data features.

[0078] For example, the second transformation feature output by the Transformer branch is an 8×8×512 feature map, which is converted into a 1×1×512 vector through adaptive average pooling. The second temporal correlation feature output by the ResNet branch is a 4×4×512 feature map, which is flattened into a one-dimensional vector, resulting in a one-dimensional vector of 4×4×512=8192. The one-dimensional feature vectors of the second transformation feature and the second temporal correlation feature are concatenated to obtain the test data feature with a dimension of 8704.

[0079] Furthermore, the test data features contain 8704 feature values. A subset of features is selected from these 8704 features to form a set, which includes an empty set and the entire set. These feature subsets are then sequentially input into the optimized classification model to obtain the probability of the test fault category corresponding to each feature subset.

[0080] Specifically, the prediction accuracy of the test set is calculated using the probability of the test fault category corresponding to each test data feature. At the same time, the cross-identification rate between preset similar fault categories is verified, such as the cross-identification rate between undervoltage fault and overvoltage fault. The prediction accuracy and cross-identification rate are used as the prediction precision.

[0081] In this embodiment of the invention, constructing the Shapu value matrix based on the probability of the test fault category includes: Calculate the Shapu value corresponding to each test feature in the test data features based on the probability of the test fault category; The Sharp value is mapped to the time-frequency points corresponding to the test data features to obtain the Sharp value matrix.

[0082] Specifically, the Shapu value corresponding to each test feature in the test data features is calculated using the following formula:

[0083] in, Indicates test features The corresponding SAP value, Indicates the characteristics of the test data. This represents the factorial of the number of features in the test data. This represents any subset of features in the set of feature subsets corresponding to the test data features. The factorial represents the number of feature subsets. To represent factorial calculation, Representing a feature subset Overlay test features a subset of Representing a feature subset Superimposed features The corresponding test failure category probability, Feature subset The corresponding probability of the test fault category.

[0084] In this embodiment of the invention, each feature value in the test data features represents a test feature. The probability of the test fault category after superimposing the feature subset of each test feature value can reflect the contribution of each test feature to the probability of the test fault category.

[0085] Furthermore, the test data is a set of 8704 high-dimensional features output by the Transformer and ResNet branches. Each feature value corresponds to a specific attribute of "time-frequency - time-series - global correlation". Therefore, the SAP value of each test feature can be decomposed into the independent contribution of each time-frequency point through reverse mapping, and the SAP value is assigned to the corresponding time-frequency point, finally forming a SAP value matrix with dimension p*q (p=60 for frequency dimension, q=150 for time dimension).

[0086] In this embodiment of the invention, calculating the fault assessment factor of the extended data based on the Shapu value matrix includes: Obtain the state time-frequency diagram under a preset normal state, and construct the normal state average amplitude spectrum matrix based on the state time-frequency diagram; The test fault category is determined based on the probability of the test fault category corresponding to the extended data, and the average amplitude spectrum matrix of the fault state of each of the extended data is constructed based on the test fault category. A comprehensive information matrix for normal state is constructed based on the normal state average amplitude spectrum matrix, and a comprehensive information matrix for fault state is constructed based on the fault state average amplitude spectrum matrix and the Shapu value matrix. The fault assessment factor is calculated based on the normal state comprehensive information matrix and the fault comprehensive information matrix.

[0087] Specifically, time-frequency maps of multiple normal states are obtained. For example, Fourier transform is performed on 200 sets of time-frequency maps of normal states to calculate the average amplitude at each time-frequency point, forming a normal state average amplitude spectrum matrix. Each matrix element in the normal state average amplitude spectrum matrix is ​​multiplied by the corresponding element in the preset normal state Shapu value matrix to obtain each matrix element in the normal state comprehensive information matrix.

[0088] For example, based on the test data characteristics of the extended data, the test fault category is determined, and the time-frequency graphs corresponding to multiple preset test fault categories are obtained to construct the average amplitude spectrum matrix corresponding to the test fault category.

[0089] Specifically, the comprehensive fault information matrix can be calculated and constructed using the following formula:

[0090] in, Indicates the test fault category as Fault comprehensive information matrix The Middle Line number Column matrix elements, Represents the first value in the Shapu value matrix. Line number Column matrix elements, This represents the first element in the preset time-frequency coordinate product matrix. Line number Column matrix elements, Indicates the test fault category as The fault state average amplitude spectrum matrix of the first Line number Columns of matrix elements.

[0091] In this embodiment of the invention, the fault category is calculated using Euclidean distance. The difference between the fault condition and the normal condition is used to obtain the fault assessment factor.

[0092] Furthermore, the Pearson correlation coefficient is calculated using the fault assessment factor for each test data, and it is determined whether the Pearson correlation coefficient is greater than the second threshold (e.g., 0.85). This allows for the prediction of the correlation of the optimized classification model, effectively improving the prediction accuracy of the optimized classification model.

[0093] Specifically, the Pearson correlation coefficient is calculated using the following formula:

[0094] in, Represents the correlation coefficient. Indicates the fault assessment factor. This represents the mean of the fault assessment factors in the test set. Fault assessment factors that indicate the preset severity of a fault. The mean of the fault assessment factor, representing the severity of the fault.

[0095] In this embodiment of the invention, the test accuracy, SAP value matrix, and Pearson correlation coefficient are collected, and the average recognition rate, cross recognition rate, consistency of the SAP value matrix for the same fault, and Pearson correlation coefficient in the test accuracy are judged in turn to determine whether they simultaneously meet the requirements. If they do not meet the requirements, the parameters of the classification model and the dual-branch feature extraction model are adjusted and optimized until the multi-dimensional verification parameters simultaneously meet the first threshold and the second threshold, so as to obtain the verification model parameters of the optimized classification model and the dual-branch feature extraction model when the first threshold and the second threshold are simultaneously met.

[0096] In this embodiment of the invention, by verifying the model parameters, the actual application effect can be evaluated after training the optimized classification model and the dual-branch feature extraction model, ensuring the generalization ability of the model and avoiding overfitting.

[0097] S5. Construct a target fault prediction model based on the verification model parameters, use the target fault diagnosis model to predict the fault of the high-voltage circuit breaker signal to be diagnosed, obtain the target fault category, and calculate the feature importance heatmap and target fault evaluation factor corresponding to the high-voltage circuit breaker signal to be diagnosed.

[0098] In this embodiment of the invention, the parameters in the dual-branch feature extraction model and the optimized classification model are replaced according to the verification model parameters to obtain the target fault diagnosis model.

[0099] Furthermore, the high-voltage circuit breaker signal to be diagnosed refers to the vibration signal collected during the operation of the high-voltage circuit breaker and used to determine the health status of the equipment, such as the closing vibration signal.

[0100] Specifically, the step of using the target fault diagnosis model to predict the fault in the high-voltage circuit breaker signal to be diagnosed, and obtaining the target fault category, includes: Convert the high-voltage circuit breaker signal to be diagnosed into a time-frequency diagram of the signal to be diagnosed. The diagnostic features of the signal to be diagnosed are extracted from the time-frequency graph of the signal to be diagnosed using the target fault diagnosis model. The fault classification prediction is performed on the features of the signal to be diagnosed to obtain the target fault category.

[0101] In this embodiment of the invention, the dual-branch feature extraction model in the target fault diagnosis model is used to extract the features of the signal to be diagnosed and input into the optimized classification model in the target fault diagnosis model. This allows for prediction of the features of the signal to be diagnosed, thereby obtaining the target fault category. The specific method is as described above for extracting time-frequency map feature vectors and fault prediction, and will not be repeated here.

[0102] Furthermore, the SAP value matrix corresponding to the signal of the high-voltage circuit breaker to be diagnosed is calculated. The SAP value matrix is ​​then decomposed into SAP values ​​corresponding to each time-frequency point through reverse mapping, resulting in a feature importance heatmap of the high-voltage circuit breaker signal to be diagnosed. For example, if the high-voltage circuit breaker signal is diagnosed as an overvoltage fault, the SAP value in the 20-40kHz mid-frequency band during the spring energy storage stage is greater than 0.04; this can locate key time-frequency regions, such as "abnormalities in the 20-30kHz frequency band during the contact closing stage".

[0103] In detail, by using the above method to calculate the target fault assessment factor, it can be determined whether the target fault category is a difficult-to-identify fault and requires emergency repair. For example, if the value of the target fault assessment factor is greater than 18, it is recommended to repair within 24 hours; if it is between 12 and 18, it is recommended to repair within 72 hours; and if it is less than 12, it is recommended to plan the repair.

[0104] In this embodiment of the invention, the fault category of the high-voltage circuit breaker signal to be diagnosed can be accurately identified through the target fault diagnosis model. At the same time, the cause of the fault can be located based on the feature importance heatmap and the target fault assessment factor, thereby improving the interpretability of the target fault diagnosis model and realizing the accurate identification of high-voltage circuit breaker fault detection.

[0105] like Figure 4 The diagram shown is a functional block diagram of a high-voltage circuit breaker fault identification system provided in an embodiment of the present invention.

[0106] The high-voltage circuit breaker fault identification system 100 of this invention can be installed in a processing device. Depending on the functions implemented, the high-voltage circuit breaker fault identification system 100 may include a vibration signal conversion module 101, a data expansion module 102, a parameter optimization module 103, a verification model parameter determination module 104, and a fault prediction module 105. The module described in this invention can also be called a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, and are stored in the memory of the electronic device.

[0107] In this embodiment, the functions of each module / unit are as follows: The vibration signal conversion module 101 acquires the multi-state vibration signal of the high-voltage circuit breaker and converts the multi-state vibration signal into a multi-state signal time-frequency diagram. The data expansion module 102 applies data expansion to the time-frequency diagram of the multi-state signal to obtain an expanded dataset, and divides the expanded dataset into a training set and a test set. The parameter optimization module 103 is used to extract the time-frequency map feature vector of the training set using a pre-built dual-branch feature extraction model, and optimize the parameters of the pre-built fault classifier based on the time-frequency map feature vector to obtain an optimized classification model. The verification model parameter determination module 104 is used to calculate the multi-dimensional verification parameters of the test set according to the optimized classification model, and determine the verification model parameters according to the multi-dimensional verification parameters. The fault prediction module 105 is used to construct a target fault prediction model based on the verification model parameters, use the target fault diagnosis model to predict the fault of the high-voltage circuit breaker signal to be diagnosed, obtain the target fault category, and calculate the feature importance heatmap and target fault evaluation factor corresponding to the high-voltage circuit breaker signal to be diagnosed.

[0108] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for fault identification of a high-voltage circuit breaker, characterized in that, include: Acquire multi-state vibration signals of the high-voltage circuit breaker and convert the multi-state vibration signals into a multi-state signal time-frequency diagram; The time-frequency graph of the multi-state signal is extended to obtain an extended dataset, which is then divided into a training set and a test set. The time-frequency graph feature vector of the training set is extracted using a pre-built dual-branch feature extraction model. The parameters of the pre-built fault classifier are optimized based on the time-frequency graph feature vector to obtain an optimized classification model. Calculate the multi-dimensional validation parameters of the test set based on the optimized classification model, and determine the validation model parameters based on the multi-dimensional validation parameters. Based on the verification model parameters, a target fault prediction model is constructed. The target fault diagnosis model is then used to predict the fault of the high-voltage circuit breaker signal to be diagnosed, obtain the target fault category, and calculate the feature importance heatmap and target fault evaluation factor corresponding to the high-voltage circuit breaker signal to be diagnosed.

2. The high-voltage circuit breaker fault identification method as described in claim 1, characterized in that, The process of converting the multi-state vibration signal into a multi-state signal time-frequency diagram includes: The wavelet transform coefficients are obtained by performing continuous wavelet transform on each vibration signal in the multi-state vibration signal using a preset wavelet function. Construct a time-frequency diagram of each vibration signal based on the wavelet transform coefficients; By compiling the time-frequency diagrams of the signals, a multi-state signal time-frequency diagram corresponding to the multi-state vibration signal is obtained.

3. The high-voltage circuit breaker fault identification method as described in claim 1, characterized in that, The data extension of the time-frequency graph of the multi-state signal to obtain an extended dataset includes: Adversarial learning is performed on each signal time-frequency plot in the multi-state signal time-frequency plot to obtain an adversarial learning time-frequency plot; Each of the signal time-frequency maps is cropped at multiple scales to obtain the cropped time-frequency map corresponding to the signal time-frequency map; The multi-state signal time-frequency diagram is mixed and masked to obtain the mixed time-frequency diagram corresponding to the signal time-frequency diagram; The extended dataset is obtained by combining the multi-state signal time-frequency map, the adversarial learning time-frequency map, the cropped time-frequency map, and the hybrid time-frequency map.

4. The high-voltage circuit breaker fault identification method as described in claim 1, characterized in that, The step of extracting the time-frequency map feature vector of the training set using a pre-built dual-branch feature extraction model includes: Temporal pruning is performed on each extended data in the training set to obtain a time-frequency plot subsequence; The temporal correlation features of the time-frequency graph subsequences are extracted based on the residual network branches in the pre-trained model; The attention weights in the feature transformation encoder of the pre-trained model are calculated using the temporal correlation features, and the transformation features of each of the extended data are calculated based on the attention weights. The temporal correlation features are flattened to obtain flattened correlation features, and the transformed features are averaged pooled to obtain pooled features. The flat correlation features and the pooling features are concatenated to obtain the time-frequency plot feature vector.

5. The high-voltage circuit breaker fault identification method as described in claim 4, characterized in that, The step of calculating the attention weights in the feature transformation encoder of the pre-trained model using the temporal correlation features includes: The feature transformation encoder is used to embed and encode each of the extended data to obtain an encoded feature vector; The encoded feature vector is linearly mapped to obtain the query matrix, the key matrix, and the value matrix. Attention weights are calculated using the query matrix, the key matrix, the value matrix, and the temporal correlation features.

6. The high-voltage circuit breaker fault identification method as described in claim 1, characterized in that, The calculation of the multi-dimensional verification parameters of the test set based on the optimized classification model includes: Extract test data features from each extended data in the test set, and extract a feature subset set of the test data features; The optimized classification model is used to predict faults in the feature subset set and the test data features to obtain the probability of the test fault category. The test accuracy of the optimized classification model is calculated based on the probability of the test fault categories. A Shapu value matrix is ​​constructed based on the probability of the test fault categories, and the fault assessment factor of the extended data is calculated based on the Shapu value matrix. Calculate the Pearson correlation coefficient between the fault assessment factor and the preset standard fault assessment factor; By combining the test accuracy, the Shapu value matrix, and the Pearson correlation coefficient, multi-dimensional verification parameters are obtained.

7. The high-voltage circuit breaker fault identification method as described in claim 6, characterized in that, The calculation of the fault assessment factor of the extended data based on the Shapu value matrix includes: Obtain the state time-frequency diagram under a preset normal state, and construct the normal state average amplitude spectrum matrix based on the state time-frequency diagram; The test fault category is determined based on the probability of the test fault category corresponding to the extended data, and the average amplitude spectrum matrix of the fault state of each of the extended data is constructed based on the test fault category. A comprehensive information matrix for normal state is constructed based on the normal state average amplitude spectrum matrix, and a comprehensive information matrix for fault state is constructed based on the fault state average amplitude spectrum matrix and the Shapu value matrix. The fault assessment factor is calculated based on the normal state comprehensive information matrix and the fault comprehensive information matrix.

8. A high-voltage circuit breaker fault identification system, characterized in that, The system includes: The vibration signal conversion module acquires multi-state vibration signals from a high-voltage circuit breaker and converts the multi-state vibration signals into a multi-state signal time-frequency diagram. The data expansion module applies data expansion to the time-frequency graph of the multi-state signal to obtain an expanded dataset, and divides the expanded dataset into a training set and a test set. The parameter optimization module is used to extract the time-frequency map feature vector of the training set using a pre-built dual-branch feature extraction model, and optimize the parameters of the pre-built fault classifier based on the time-frequency map feature vector to obtain an optimized classification model. The verification model parameter determination module is used to calculate the multi-dimensional verification parameters of the test set based on the optimized classification model, and determine the verification model parameters based on the multi-dimensional verification parameters. The fault prediction module is used to construct a target fault prediction model based on the verification model parameters, use the target fault diagnosis model to predict the fault of the high-voltage circuit breaker signal to be diagnosed, obtain the target fault category, and calculate the feature importance heatmap and target fault evaluation factor corresponding to the high-voltage circuit breaker signal to be diagnosed.

9. A processing device, characterized in that, The method includes at least one processor and at least one memory communicatively connected to the processor, wherein the memory stores program instructions executable by the processor, and the processor can execute the method as described in any one of claims 1 to 7 by invoking the program instructions.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that cause the computer to perform the method as described in any one of claims 1 to 7.

Images