Circuit breaker fault assessment method based on multi-domain information fusion and deep learning

Abstract

The application is a circuit breaker fault evaluation method based on multi-domain information fusion and deep learning. First, the original vibration signals of the contact under different working conditions are collected, and the fault-related vibration signal segments are extracted from the original vibration signals. Then, the frequency domain and time-frequency domain information of the fault-related vibration signal segments are extracted, and an RGB image is generated through multi-domain information fusion. Then, the improved auxiliary classification generative adversarial network is used to generate fault samples. Finally, a fault evaluation model is constructed, which includes a shared layer and two task layers connected with the shared layer. The shared layer includes an A-type residual block, two B-type residual blocks and a double attention module, and the two task layers are used for fault type and fault degree classification respectively. The expanded data set is used to train the model, and the trained fault evaluation model is used for fault type and fault degree evaluation. The problems of data set class imbalance and signal analysis limitation in circuit breaker fault diagnosis are solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of contact system fault assessment of circuit breakers, and particularly relates to a circuit breaker fault assessment method based on multi-domain information fusion and deep learning. BACKGROUND

[0002] As an important electrical equipment, the universal circuit breaker is mainly used in low-voltage distribution networks. When it is in an operating state, various faults may occur, thereby causing a great threat to the safety of the entire distribution network, and even leading to casualties of operating personnel in severe cases. Considering that the mechanical fault of the circuit breaker is an evolution process from slight to severe, it is of great significance to simultaneously assess the fault type and degree of the universal circuit breaker to ensure the safe and reliable operation of the entire low-voltage distribution network.

[0003] During the opening and closing process of the circuit breaker, the dynamic and static contacts will collide violently, and the vibration signals generated thereby contain rich information of the contact system and its operating mechanism, and can represent the operating state of the circuit breaker. Therefore, the vibration signals are used for fault diagnosis. Wang Xiaoming et al. (Wang Xiaoming, Zhou Ke, et al. Mechanical fault diagnosis method of high-voltage circuit breaker based on identity mapping CNN[J]. High Voltage Technology, 2021, 47(10): 3657-3663.) proposed a mechanical fault diagnosis method of high-voltage circuit breaker based on identity mapping convolutional neural network (CNN), which autonomously extracts fault feature information and improves the recognition effect of mechanical fault of high-voltage circuit breaker. However, this method is for the single-target task of fault diagnosis, and since the fault severity is accumulated slowly after the mechanical fault occurs, an effective fault diagnosis method should include fault classification and degree assessment. In addition, since the fault type and degree are related to each other, multi-task learning (MTL) can be used to learn the shared features of multiple related tasks and use the useful information across tasks to improve the generalization performance of the model. For example, Sun Shuguang et al. (Sun Shuguang, Zhang Tingting, et al. Fault degree assessment of circuit breaker based on continuous wavelet transform and MTL-SEResNet[J]. Chinese Journal of Scientific Instrument, 2022, 43(06): 162-173.) used the two-dimensional time-frequency diagram of the vibration signal as the input and realized the fault diagnosis and degree assessment of the low-voltage circuit breaker by using the multi-task learning sharing mechanism. Deep learning has achieved good results in simultaneously solving multiple fault diagnosis tasks, but there are still the following deficiencies in the research on fault diagnosis of low-voltage circuit breakers: (1) Big data dependency: In industrial practice, the normal data generated by the circuit breaker is usually more than the fault data, thereby forming a class-unbalanced data set. In the class-unbalanced data set, there are disadvantages such as low fault recognition accuracy and poor generalization ability. (2) Single signal analysis: Most of the existing methods start from a single analysis domain, and the fault information contained is limited, so the recognition degree between different faults is poor, and the diagnosis accuracy is low. SUMMARY

[0004] In view of the deficiencies of the prior art, the technical problem to be solved by the present application is to provide a circuit breaker fault evaluation method based on multi-domain information fusion and deep learning.

[0005] The technical scheme adopted by the present application to solve the technical problem is as follows:

[0006] A circuit breaker fault evaluation method based on multi-domain information fusion and deep learning, characterized in that the method comprises the following steps:

[0007] Firstly, a fault simulation test is performed on a circuit breaker contact system to collect original vibration signals of the contact under different working states;

[0008] Secondly, a double-threshold method based on TKEO is used to extract a fault-associated vibration signal segment from the original vibration signals;

[0009] Thirdly, frequency domain and time-frequency domain information of the fault-associated vibration signal segment is extracted, and the time domain, frequency domain and time-frequency domain information of the fault-associated vibration signal segment is fused to generate an RGB image;

[0010] Fourthly, the Wasserstein distance is used as a cost function and a gradient penalty is introduced to improve an auxiliary classification generative adversarial network; the RGB image is input into a generator of the improved auxiliary classification generative adversarial network to generate a fault sample and realize expansion of a data set;

[0011] Fifthly, a fault evaluation model is constructed, the fault evaluation model comprising a shared layer and two task layers connected with the shared layer; the shared layer comprises an A-type residual block, two B-type residual blocks and a double attention module, the A-type residual block being located between the two B-type residual blocks, and the output of each residual block being spliced with the input of the shared layer to serve as the input of the next residual block; one of the task layers is used for fault type classification and comprises a global average pooling layer and a classification layer, and the other task layer is used for fault degree classification and comprises an A-type residual block, a B-type residual block, a global average pooling layer and a classification layer; the A-type residual block and the B-type residual block are both ResNeXt residual blocks, and the double attention module comprises position attention and channel attention;

[0012] The expanded data set is used to train the model, and the trained fault evaluation model is used for fault type and fault degree evaluation.

[0013] Further, the A-type residual block comprises eight parallel branches, each branch comprising a first convolutional layer and a second convolutional layer, a batch normalization layer and an activation layer being connected after each convolutional layer, the outputs of all branches being shortcut-connected with the input of the A-type residual block to obtain the output of the A-type residual block;

[0014] The B-type residual block consists of eight parallel branches, each of which includes a third and a fourth convolutional layer. Each convolutional layer is followed by a batch normalization layer and an activation layer. The input of the B-type residual block passes through the fifth convolutional layer, the batch normalization layer, and the activation layer simultaneously, and then is quickly connected to the outputs of all branches to obtain the output of the B-type residual block.

[0015] The dual attention module includes a positional attention module and a channel attention module. The input of the dual attention module is processed by three convolution operations to obtain three feature maps F. a F b and F c The three feature maps are reshaped to obtain three feature maps F'. a 、F' b and F' c Feature map F' b and feature map F' a After multiplying by the transpose, and then passing through the classification layer, we obtain the spatial attention map M; the transpose of the spatial attention map M and the feature map F' c After multiplication, and then reshaping, the feature map M is obtained. S Feature map M S After multiplying by the weight coefficient of the positional attention module, it is added to the input of the dual attention module to obtain the output of the positional attention module;

[0016] The input to the dual attention module undergoes three reshaping operations to obtain three feature maps F1, F2, and F3. The transposes of feature maps F2 and F3 are multiplied together and then passed through a classification layer to obtain a channel attention map Q. The channel attention map Q is multiplied with the transpose of feature map F1 and then reshaped to obtain feature map Q'. s Feature map Q s After multiplying by the weight coefficient of the channel attention module, the result is added to the input of the dual attention module to obtain the output of the channel attention module.

[0017] Furthermore, in step 3, the fault-related vibration signal segments are subjected to Fast Fourier Transform and Discrete Cosine S-Transform respectively to obtain frequency domain and time-frequency domain information; the time domain, frequency domain and time-frequency domain information of the fault-related vibration signal segments are converted into grayscale images respectively, and all grayscale images are fused in the channel dimension to obtain RGB images.

[0018] Furthermore, the auxiliary classification generative adversarial network includes a generator and a discriminator. The input of the generator passes through a fully connected layer and a reshaping operation, and then passes through four deconvolutional layers in sequence. Each deconvolutional layer is followed by a batch normalization layer and an activation layer. The discriminator includes four convolutional layers and one fully connected layer. Each convolutional layer is followed by a batch normalization layer, an activation layer, and a Dropout layer.

[0019] Compared with the prior art, the essential features of this invention are:

[0020] Considering the imbalance of datasets and the limitations of signal analysis in circuit breaker fault diagnosis, this paper first extracts fault-related vibration signal segments from the original vibration signals and synthesizes color images by fusing multi-domain information to enrich the fault characterization information. Secondly, an improved ACGAN is used to expand the samples and balance the dataset. Finally, a fault assessment model is constructed to evaluate the fault type and severity. The fault assessment model incorporates multi-task learning and a dual attention mechanism, and key aspects are improved to enhance network performance. Experimental results show that the method of this invention has higher classification accuracy and operational efficiency.

[0021] The significant advancement of this invention compared to existing technologies is:

[0022] (1) Fault-related vibration signal segments are extracted from the original vibration signals of the circuit breaker, and the frequency domain and time-frequency domain information of the segments are extracted. RGB images are obtained by fusing multi-domain information. Since the fault information contained in a single analysis domain is limited, and the fault information features in different analysis domains are different, the fused RGB image contains more effective features, enriches the fault information of the circuit breaker, makes different fault features more separable, and improves the evaluation accuracy. In addition, deep learning is mostly image-oriented, and feature extraction methods such as two-dimensional convolution in deep learning are more compatible with two-dimensional images. Therefore, using RGB images as input to the fault evaluation model can improve the evaluation efficiency.

[0023] (2) To improve the generalization ability of the model, the evaluation criterion for evaluating the similarity between real samples and generated samples in ACGAN is replaced by JS divergence with Wasserstein distance with gradient penalty. The WACGAN-GP model expands the sample by generating effective fault sample data, thereby solving the class imbalance problem and improving the robustness of the subsequent fault assessment model.

[0024] (3) The balanced dataset formed by mixing the original samples and the generated samples is used as the input of the fault assessment model. The front-end ResNeXt uses grouped convolution to mine weak fault information. While avoiding the gradient vanishing phenomenon caused by network depth, it also reduces the number of hyperparameters. Then, the weights of the channel and spatial dimensions are learned through the dual attention module, so as to distinguish important features. Finally, the sharing mechanism of multi-task learning is used to realize the mutual learning of the model on the fault type and degree assessment task.

[0025] (4) Compared with other circuit breaker fault assessment methods, the method of the present invention effectively improves the fault classification performance by learning the features of multi-domain information fusion images and by taking advantage of the information transfer capability between fault types and degrees unique to network structures. The accuracy rates of fault type and degree classification reached 99.83% and 99.54%, respectively. Attached Figure Description

[0026] Figure 1 This is an overall flowchart of the present invention;

[0027] Figure 2 A flowchart for generating an RGB image;

[0028] Figure 3 This is a schematic diagram of the fault assessment model.

[0029] Figure 4(a) is a schematic diagram of the structure of the type A residual block;

[0030] Figure 4(b) is a schematic diagram of the structure of residual block B;

[0031] Figure 5 This is a schematic diagram of the dual attention module.

[0032] Figure 6(a) shows the signal waveform during the circuit breaker tripping process;

[0033] Figure 6(b) shows the signal waveform during the circuit breaker closing process;

[0034] Figure 7(a) is a schematic diagram of the calibration results of the starting point of the fault-related vibration signal segment under normal conditions;

[0035] Figure 7(b) is a schematic diagram of the calibration results of the starting point of the fault-related vibration signal segment under different phases in item A;

[0036] Figure 7(c) is a schematic diagram of the calibration results of the starting point of the fault-related vibration signal segment under different phases in item B;

[0037] Figure 7(d) is a schematic diagram of the calibration results of the starting point of the fault-related vibration signal segment under different phases in item C;

[0038] Figure 7(e) is a schematic diagram of the calibration results of the starting point of the fault-related vibration signal segment under the false closing state;

[0039] Figure 7(f) is a schematic diagram of the calibration results of the starting point of the fault-related vibration signal segment under the condition of incomplete circuit breaker tripping;

[0040] Figure 8 The loss curves of the generator and discriminator during WACGAN-GP training, as well as the effect diagram of the generated samples;

[0041] Figure 9A similarity comparison graph between real samples and generated samples;

[0042] Figure 10 A graph showing the change in fault classification error rate under the different loss proportions for different fault types;

[0043] Figure 11(a) shows the ROC curve of the fault degree;

[0044] Figure 11(b) shows the ROC curves for different fault types. Detailed Implementation

[0045] The technical solution of the present invention will be further described below with reference to the accompanying drawings and this embodiment, but it is not intended to limit the scope of protection of the present invention.

[0046] This invention provides a circuit breaker fault assessment method (hereinafter referred to as the method) based on multi-domain information fusion and deep learning, comprising the following steps:

[0047] The first step is to build a circuit breaker fault test platform to conduct fault simulation tests on the circuit breaker contact system. The test content includes fault type and fault degree, and collect the original vibration signals of the contacts under different operating conditions.

[0048] The second step is to extract fault-related vibration signal segments from the original vibration signal using the dual-threshold method based on the Teager-Kaiser Energy Operator (TKEO).

[0049] 2-1 The original vibration signal is framed using a window function of predetermined length. The expression for the framed processing is as follows:

[0050] o i (n)=w(n)*u((i-1)*inc+n) (1)

[0051] In the formula, u(·) is the original vibration signal, w(n) is the Hanning window function, and o i (n) represents the vibration signal at the nth sampling point of the i-th frame, n = 1, 2, ..., L, i = 1, 2, ..., v, L is the frame length, inc is the frame movement step size, and v is the total number of frames;

[0052] 2-2 For vibration signal o i (n), the result of its TKEO operation is:

[0053] ψ(o i (n))=[o i (n)] 2 -o i (n-1)o i (n+1) (2)

[0054] In the formula, o i (n-1), o i (n+1) represent the vibration signals at the (n-1)th and (n+1)th sampling points of the i-th frame, respectively;

[0055] 2-3 Fault-related vibration signal segments were extracted from the original vibration signal using the dual-threshold method.

[0056] The third step is to extract the frequency domain and time-frequency domain information of the fault-related vibration signal segments, and then fuse the time domain, frequency domain, and time-frequency domain information of the fault-related vibration signal segments through multi-domain information fusion to generate an RGB image.

[0057] 3-1 Fast Fourier Transform (FFT) and Discrete Cosine S-Transform (DCST) are performed on the fault-related vibration signal segments to obtain the frequency domain and time-frequency domain information of the fault-related vibration signal segments, while the fault-related vibration signal segments themselves are time domain information.

[0058] 3-2 The information from the three analysis domains of the fault-related vibration signal segment is converted into grayscale images of size m×m×1, respectively. 2 The length of the fault-associated vibration signal segment; the pixel matrix of the grayscale image is represented as:

[0059]

[0060] In the formula, (j,k) represents a pixel, j=1…,m represents a row of the grayscale image, k=1,…,m represents a column of the grayscale image, h(·) represents the signal strength, and Max(h) and Min(h) represent the maximum and minimum values ​​of the signal strength, respectively.

[0061] 3-3 The grayscale images from the three analysis domains are fused along the channel dimension to generate an RGB image of size 64×64×3 pixels. See [link / reference]. Figure 2 .

[0062] The fourth step is to improve the Auxiliary Classifier Generative Adversarial Network (ACGAN) to obtain the improved Auxiliary Classifier Generative Adversarial Network (WACGAN-GP model). The Auxiliary Classifier Generative Adversarial Network includes a generator and a discriminator. The RGB image generated by fusing multi-domain information is input into the generator to generate fault samples and realize sample expansion.

[0063] 4-1 The Wasserstein distance is used instead of the JS divergence in ACGAN as the cost function to measure the distance between generated samples and real samples. The distance calculation formula is shown in Equation (4):

[0064]

[0065] In the formula, ∏(P r ,P g P represents the set of all possible joint distributions of the generated sample and the real sample. r P g Let denot the true sample distribution and the generated sample distribution, respectively; let inf denote the maximum lower bound; and let γ denote the set ∏(P) r ,P g Let E be a joint distribution of the real and generated samples, where x is the real sample, y is the generated sample, and ||xy|| is the distance between the real and generated samples. (x，y)～γ [||xy||] represents the expectation between samples from two distributions;

[0066] 4-2 As can be seen from equation (5), the essence of the Wasserstein distance is to reduce the distance between the real sample x and the generated sample y. However, it still suffers from slow convergence speed and gradient explosion. Therefore, a gradient penalty (GP) term is introduced to replace weight clipping and to impose a Lipschitz constraint on the discriminator. The gradient penalty calculation formula is as follows:

[0067]

[0068] In the formula, λ is the gradient penalty coefficient. For the true sample distribution P r and the generated sample distribution P g The joint distribution Indicates sample Follows joint distribution Expectations Let ||·||2 be the gradient of the discriminator, and ||·||2 be the L2 norm.

[0069] 4-3 After adding gradient penalty to ACGAN, the loss functions of the discriminator D and the generator G are expressed by equations (6) and (7), respectively:

[0070]

[0071]

[0072] in, This indicates that sample x follows a distribution P. r Expectations Indicates sample Follows distribution P g Expectations Indicates sample Follows distribution P g Expectations Indicates sample Follows distribution P g Expectations This means by using x and The random sample obtained by linear interpolation, ε is a random number following U(0,1), P(Y=y|S real S represents the conditional probability distribution of class labels. real Let D(·) represent the real sample, and D(·) be the discriminator function.

[0073] The generator takes 100-dimensional random noise Z and labels C as input. After passing through a fully connected layer and a reshaping operation, the input passes through four deconvolutional layers. Each deconvolutional layer is followed by batch normalization and a ReLU activation function to avoid overfitting and accelerate training, ultimately generating an image of size m×m×3. The discriminator D consists of four convolutional layers and one fully connected layer. Each convolutional layer is followed by batch normalization, a LeaKyReLU activation function, and a Dropout operation to reduce the model's computational parameters and alleviate overfitting. The discriminator takes an image of size m×m×3 as input and outputs the probability that the image is the true class and its class.

[0074] The fifth step involves selecting a multi-task-based Hard sharing mechanism as the basic framework to construct a fault assessment model. This model comprises a sharing layer and two task layers. The sharing layer is used for feature extraction, and the two task layers are used for fault type and fault severity classification, respectively. (See also...) Figure 3 Fault samples are processed through Conv2D, Batch Normalization (BN), ReLU, and MaxPool before entering the shared layer. The shared layer consists of one Type A residual block (Resblock_a), two Type B residual blocks (Resblock_b), and a dual attention module (DANET). The Type A residual block is located between the two Type B residual blocks. The output of each residual block is concatenated with the input of the shared layer before being fed into the next residual block. The task layer for fault type classification adds a global average pooling layer (GAP_1) and a Softmax classification layer to the shared layer. The task layer for fault severity classification adds one Type A residual block, one Type B residual block, a global average pooling layer (GAP_2), and a Softmax classification layer to the shared layer.

[0075] Both Type A and Type B residual blocks are ResNeXt residual blocks, which are essentially grouped convolutions, meaning the input is grouped along the channel dimension. As shown in Figure 4(a), a Type A residual block includes eight parallel branches, each consisting of a first convolutional layer (Conv1) and a second convolutional layer (Conv2). Each convolutional layer is followed by a batch normalization layer (BN) and an activation layer (ReLU). The outputs of all branches are shortcut-connected (SC) to the input of the Type A residual block to obtain the A... The output of the B-type residual block is shown in Figure 4(b). The B-type residual block includes eight parallel branches, each of which includes a third convolutional layer (Conv3) and a fourth convolutional layer (Conv4). Each convolutional layer is connected to a batch normalization layer and an activation layer. The input of the B-type residual block passes through the fifth convolutional layer (Conv5), the batch normalization layer, and the activation layer simultaneously, and then is quickly connected to the output of all branches to obtain the output of the B-type residual block. The kernel size of the first to fifth convolutional layers is 3×3, and the stride is 1.

[0076] Assume x a and x b Given the inputs to type A residual block Resblock_a and type B residual block Resblock_b respectively, the output y a and y b It is represented by equations (8) and (9);

[0077]

[0078]

[0079]

[0080]

[0081] In the formula, Indicates the first type of residual block A. q The input and output of each branch, Indicates the first type of residual block in type B q The inputs and outputs of each branch, w1 to w5 are the weights of the first to fifth convolutional layers respectively, b1 to b5 are the biases of the first to fifth convolutional layers respectively, and σ r It is the ReLU activation function;

[0082] The dual attention module is used to optimize feature selection. It is composed of a positional attention module (PAM) and a channel attention module (CAM) fused in parallel. See [link to documentation]. Figure 5 The input to the dual attention module is the output of the second B-type residual block in the shared layer. C represents the number of channels in the feature map, H and W represent the width and height of the feature map, and R represents the feature space. The positional attention module is used to capture the spatial dependencies between any two locations of the feature map and assign different weights. The input of the dual attention module is processed by three convolution operations to obtain three feature maps F. a F b and F c The three feature maps are reshaped to obtain three feature maps F. a '、F b 'and F c ';Feature map F b 'and feature map F a After multiplying by the transpose of ', and then passing through a Softmax layer, we obtain the spatial attention map M; the transpose of the spatial attention map M and the feature map F c After multiplication, a reshaping operation is performed to obtain the feature map M. S Feature map M S After multiplying by the weight coefficients of the positional attention module, and then adding them to the input of the dual attention module, the feature map E is obtained. P This is the output of the position attention module; the specific formula is as follows:

[0083]

[0084] F′ a ,F′ b ,F′ c =reshape([F a ;F b ;F c (5)

[0085]

[0086]

[0087]

[0088] In the formula, f 1×1 This represents a convolution operation with a kernel size of 1×1 and a stride of 1. `reshape` represents the reshaping operation, and `softmax` represents the normalization operation. This indicates multiplication, T indicates transpose, and α is the weight coefficient for positional attention, with an initial value of 0;

[0089] The channel attention module is responsible for integrating the relevant features between all channel mappings, thereby selectively emphasizing interdependent channel mappings. The input of the dual attention module undergoes three reshaping operations to obtain three feature maps F1, F2, and F3. The transposes of feature maps F2 and F3 are multiplied together and then passed through a softmax layer to obtain the channel attention map Q. The channel attention map Q is multiplied with the transpose of feature map F1 and then reshaped to obtain feature map Q. s Feature map Q s After multiplying by the weight coefficients of the channel attention, and then adding them to the input of the dual attention module, the feature map E is obtained. C This refers to the output of the channel attention module; the specific formula is as follows:

[0090]

[0091]

[0092]

[0093]

[0094] In the formula, β is the weight coefficient of channel attention, with an initial value of 0;

[0095] The final output E∈R of the dual attention module C×H×W for:

[0096] E = E C +E P (13)

[0097] The sixth step is to train and optimize the fault assessment model, and then use the trained fault assessment model to evaluate the fault type and fault severity.

[0098] 6-1 The improved auxiliary classification generative adversarial network is trained using RGB images. During training, the Adam optimizer is used to optimize the parameters. The batch size is 64 and the learning rate is 0.0005. The trained generator is used to generate fault samples to expand the dataset.

[0099] 6-2 80% of the samples were randomly selected from the expanded dataset as the training set, and the remaining 20% ​​as the test set to train the fault assessment model. The Adam optimizer was used during training, with a learning rate of 0.001, and 32 samples were processed per iteration, with a maximum of 200 iterations. The total loss of the model was calculated using the cross-entropy loss function, as shown in the following formula:

[0100] Loss=δLoss1+(1-δ)Loss2,0<δ<1 (14)

[0101] In the formula, δ is the proportion of the failure type classification loss, and Loss1 and Loss2 are the failure type classification loss and failure severity classification loss, respectively.

[0102] Example 1

[0103] The circuit breaker fault assessment method based on multi-domain information fusion and deep learning in this embodiment has the following specific steps:

[0104] The first step mainly studies the mechanical faults generated in the contact system of the DW15-1600 universal circuit breaker during the opening and closing process. The working states include six states: normal, phase A out of phase (NA), phase B out of phase (NB), phase C out of phase (NC), false closing (FC), and incomplete opening (IO). Vibration signals are measured using an LC0159 accelerometer installed on the crossbeam of the circuit breaker. Vibration signals are sampled through a USB-7648A data acquisition card at a sampling frequency of 20kHz and a single vibration sampling duration of 250ms.

[0105] Various faults were simulated by adjusting the opening distance and overtravel of the three-phase contacts. Due to the mechanical structure of the circuit breaker, the opening distance and overtravel of the contacts could only be adjusted in 0.5mm increments, with a maximum of 5.0mm, when simulating faults and their severity. Therefore, this embodiment simulated fault data at eight distances: 0.5mm, 1.0mm, 1.5mm, 2.0mm, 2.5mm, 3.0mm, 3.5mm, and 4.0mm. Forty fault samples were generated for each distance, with faults below 1.5mm classified as minor faults, 1.5-2.0mm as general faults, 2.5-3.0mm as severe faults, and 3.5-4.0mm as fatal faults. The sample data for various operating conditions are shown in Table 1.

[0106] Table 1 Dataset

[0107]

[0108] The second step is to extract fault-related vibration signal segments from the original vibration signal using the dual-threshold method based on TKEO.

[0109] The signal waveforms of the circuit breaker's opening and closing processes are shown in Figures 6(a) and (b). The opening process includes the contact state voltage, the current of the opening coil, and the contact vibration signal. The closing process includes the contact state voltage, the current of the closing coil, and the contact vibration signal. During the circuit breaker's opening process, t3 is the moment when the contact state voltage jumps and the moving main contact begins to move, representing the start of the opening action. For the closing process, T3 is the moment when the contact state voltage jumps and the moving arc contact collides with the stationary arc contact. Since this invention addresses faults in the contact system's opening and closing operation process, T3 and t3 are selected as the starting points of the vibration signal segments associated with the closing and opening faults, respectively.

[0110] The Hamming window length was set to 40 and the frame shift to 20. The starting point of the associated vibration signal segments for different faults was calibrated using the TKEO-based dual-threshold method. The starting point calibration results for various faults are shown in Figures 7(a) to (f), where T′3 and t′3 represent the start times of the effective sub-events after calibration. The dual-threshold calibration process is illustrated using Figure 7(a) as an example: For the TKEO of the vibration signal, the upper and lower thresholds of the dual thresholds are determined by testing to be L2 and L1, respectively, and L2 > L1. The TKEO range greater than L2 is searched, i.e., segment AB in the figure, which must be the interval where the fault-associated vibration signal segment is located. Then, the search proceeds from A and B towards both ends, and based on L1, C can be accurately determined as the initial starting point of the fault-associated vibration signal segment. Considering the attenuation time of the vibration signal and the need for subsequent frequency domain and time-frequency domain feature extraction, 4096 data sampling points are extracted after T′3 and t′3 as the fault-associated vibration signal segment.

[0111] The third step is to extract the frequency domain and time-frequency domain information of the fault-related vibration signal segments, and generate an RGB image by fusing multi-domain information.

[0112] Fault signals have different characteristics in different analysis domains. By fusing information from multiple domains, the differences between fault signals can be more fully reflected. RGB images are used as input to the fault identification model.

[0113] The fourth step involves generating effective fault sample data using the WACGAN-GP model to expand the dataset. The Adam parameters are optimized with a batch size of 64 and a learning rate of 0.0005. After the WACGAN-GP model is trained, a generator is used to generate fault samples to balance the dataset. The loss curves of the generator and discriminator during WACGAN-GP training, as well as the performance of the generated samples at each stage, are shown below. Figure 8 As shown, it can be observed that the loss curves of both the discriminator D and the generator G tend to stabilize and converge near 0 as the number of iterations increases, indicating that the discriminator D and the generator G have reached Nash equilibrium. Simultaneously, the generated samples gradually approximate the real samples.

[0114] To evaluate the quality of the generated samples, the difference hashing (dHash) algorithm, cosine similarity, and Pearson correlation coefficient (PCC) were used to determine the similarity between the generated samples and the real sample images. Higher values ​​for these three evaluation metrics indicate a stronger linear relationship between the two variables. This section evaluates 100 generated samples for each of the 20 fault categories listed in Table 1. Figure 9 As shown, the average similarity of each generated sample across the three metrics is higher than 0.8, indicating that all generated samples are highly correlated with real samples of the same category. Therefore, the proposed WACGAN-GP can generate high-quality samples to address the problem of imbalanced sample data in fault diagnosis.

[0115] The proposed WACGAN-GP was compared with other GAN models and classic data augmentation methods, namely image transform (GT), DCGAN, WGAN-GP, and SA-ACGAN. The average accuracy and F1-score were used as metrics after 10 iterations. Table 2 shows the test results for different data augmentation methods. It can be seen that although GT can expand the training set, it cannot provide further effective knowledge for subsequent diagnostic models, thus limiting its ability to improve accuracy. Using GAN models, however, significantly improves both recognition accuracy and F1-score. Our proposed WACGAN-GP performs best, improving diagnostic accuracy by 2.24%, 2.82%, and 1.43% compared to other GAN models in fault category tasks, and by 2.2%, 2.71%, and 1.43% in fault severity, respectively. This demonstrates that different images generated by WACGAN-GP can provide a wider range of features for the classifier.

[0116] Table 2. Diagnostic results under different data augmentation methods

[0117]

[0118] The fifth step is to establish a fault severity assessment model. The model is trained using the training set, with a batch size of 32, a maximum number of iterations of 200, and the Adam optimizer used with a learning rate of 0.001. The loss of the fault assessment model is calculated using the cross-entropy loss function. Figure 10 As shown, when α = 0.925, the classification error rates for both fault type and fault severity reach their lowest levels, meaning that the diagnostic capabilities of both tasks are optimal. Based on the above analysis, the fault type classification loss weight in this embodiment is set to 0.925.

[0119] Using the trained fault assessment model, fault type and severity are assessed based on the test set. To avoid randomness in the results, 10 experiments were conducted using the test set, and the average value was taken as the diagnostic result. Figures 11(a) and (b) show the ROC (receiver operating characteristic) curves for fault type and fault severity, respectively. As can be seen from the figures, the ROC curves for fault type and fault severity under this method are very close to the upper left corner (TPR=1 and FPR=0). Furthermore, the average AUC values ​​of this model reach 99.88% and 99.47%, respectively, indicating that this method can accurately classify fault type and fault severity even with imbalanced data.

[0120] The advantages of the proposed model under the multi-task Hard mechanism are discussed, with the balanced dataset used as the model input. Four network structures were set up: MTL-SEResNet, HCNN, FD, and DI. FD is the fault category network branch structure of MDA-ResNeXt, and DI is the degree evaluation branch structure of MDA-ResNeXt. Table 3 shows that compared to the single-branch structures of FD and DI, the Hard multi-task mechanism enhances generalization ability, thus further improving the classification accuracy of the proposed method. Secondly, compared with multi-task models based on the Hard mechanism such as HCNN and MTL-SEResNet, the MDA-ResNeXt model proposed in this invention uses improved residual blocks, which can maximize the extraction of fault feature information contained in vibration signals and reduce model complexity, thereby achieving higher classification accuracy and shorter running time.

[0121] Table 3 Classification performance of different methods

[0122]

[0123] Finally, this method is compared with recently published circuit breaker fault diagnosis methods. These include two machine learning methods: Random Forest (RF) and Grey Wolf Optimization-Support Vector Machine (GSO-SVM), and two deep learning methods: 2D-CNN and 2D-ResNet. The input data types and model structure settings for the machine learning methods are consistent with the original paper. For the deep learning methods, 2D-CNN and 2D-ResNet are used to replace the methods described in this paper. Figure 3 Each residual block in the dataset is used as input, and the rest is consistent with the model settings of this invention. In general, shallow machine learning methods lag behind deep learning methods on both evaluation metrics. This is attributed to the powerful deep feature extraction capabilities of deep learning. Meanwhile, the proposed method outperforms 2D-CNN and 2D-ResNet on both evaluation metrics, achieving fault type and severity classification accuracies of 99.83% and 99.54%, respectively.

[0124] Compared with existing technologies, considering the imbalance of datasets and the limitations of signal analysis in circuit breaker fault diagnosis, this invention proposes a circuit breaker fault assessment method based on multi-source information fusion and deep learning. First, based on the segmentation of vibration signals associated with contact system faults, multi-domain information is fused to synthesize color image samples to enrich fault characterization information. Second, an improved ACGAN sample augmentation method is used to balance the dataset. Finally, the balanced dataset is fed into a constructed MDA-ResNeXt model for fault severity assessment. This model introduces multi-task learning and a dual attention mechanism, and key aspects are improved to enhance network performance. Fault assessment case studies show that, compared with other methods, the method in this invention has higher classification accuracy and operational efficiency, and is suitable for fault assessment of universal circuit breaker contact systems.

[0125] Any aspects not covered in this invention are applicable to existing technologies.

Claims

1. A circuit breaker fault assessment method based on multi-domain information fusion and deep learning, characterized in that, The method includes the following steps: The first step is to conduct a fault simulation test on the circuit breaker contact system and collect the original vibration signals of the contacts under different operating conditions; The second step is to extract fault-related vibration signal segments from the original vibration signal using the dual-threshold method based on TKEO. The third step is to extract the frequency domain and time-frequency domain information of the fault-related vibration signal segments, and then fuse the time domain, frequency domain, and time-frequency domain information of the fault-related vibration signal segments to generate an RGB image. The fourth step involves improving the auxiliary classification generative adversarial network by using Wasserstein distance as the cost function and introducing gradient penalty; then, RGB images are input into the generator of the improved auxiliary classification generative adversarial network to generate fault samples, thereby expanding the dataset. The fifth step is to construct a fault assessment model, which includes a shared layer and two task layers connected to the shared layer. The shared layer includes an A-type residual block, two B-type residual blocks, and a dual attention module. The A-type residual block is located between the two B-type residual blocks, and the output of each residual block is concatenated with the input of the shared layer to serve as the input of the next residual block. One task layer is used for fault type classification, including a global average pooling layer and a classification layer. The other task layer is used for fault severity classification, including an A-type residual block, a B-type residual block, a global average pooling layer, and a classification layer. Both the A-type and B-type residual blocks are ResNeXt residual blocks, and the dual attention module includes positional attention and channel attention. The type A residual block includes eight parallel branches, each branch including a first convolutional layer and a second convolutional layer. Each convolutional layer is followed by a batch normalization layer and an activation layer. The outputs of all branches are fast-connected to the inputs of the type A residual block to obtain the output of the type A residual block. The B-type residual block consists of eight parallel branches, each of which includes a third and a fourth convolutional layer. Each convolutional layer is followed by a batch normalization layer and an activation layer. The input of the B-type residual block passes through the fifth convolutional layer, the batch normalization layer, and the activation layer simultaneously, and then is quickly connected to the outputs of all branches to obtain the output of the B-type residual block. The model is trained using the expanded dataset, and the trained fault assessment model is used to assess fault type and fault severity.

2. The circuit breaker fault assessment method based on multi-domain information fusion and deep learning according to claim 1, characterized in that, The dual attention module includes a positional attention module and a channel attention module. The input of the dual attention module is processed by three convolution operations to obtain three feature maps F. a F b and F c The three feature maps are reshaped to obtain three feature maps F. a '、F b 'and F c ';Feature map F b 'and feature map F a After multiplying by the transpose of ', and then passing through the classification layer, we obtain the spatial attention map M; the transpose of the spatial attention map M and the feature map F c After multiplication, a reshaping operation is performed to obtain the feature map M. S Feature map M S After multiplying by the weight coefficient of the positional attention module, it is added to the input of the dual attention module to obtain the output of the positional attention module; The input to the dual attention module undergoes three reshaping operations to obtain three feature maps F1, F2, and F3. The transposes of feature maps F2 and F3 are multiplied together and then passed through a classification layer to obtain a channel attention map Q. The channel attention map Q is multiplied with the transpose of feature map F1 and then reshaped to obtain feature map Q'. s Feature map Q s After multiplying by the weight coefficient of the channel attention module, the result is added to the input of the dual attention module to obtain the output of the channel attention module.

3. The circuit breaker fault assessment method based on multi-domain information fusion and deep learning according to claim 1 or 2, characterized in that, In step 3, the fault-related vibration signal segments are subjected to Fast Fourier Transform and Discrete Cosine S-Transform respectively to obtain frequency domain and time-frequency domain information; the time domain, frequency domain and time-frequency domain information of the fault-related vibration signal segments are converted into grayscale images respectively, and all grayscale images are fused in the channel dimension to obtain RGB images.

4. The circuit breaker fault assessment method based on multi-domain information fusion and deep learning according to claim 1, characterized in that, The auxiliary classification generative adversarial network includes a generator and a discriminator. The generator's input passes through a fully connected layer and a reshaping operation, and then sequentially through four deconvolutional layers. Each deconvolutional layer is followed by a batch normalization layer and an activation layer. The discriminator includes four convolutional layers and one fully connected layer. Each convolutional layer is followed by a batch normalization layer, an activation layer, and a Dropout layer.

Images