Mechanical fault diagnosis of high voltage circuit breaker using multimodal data fusion

Signal acquisition

This study focuses on the 12kV spring-operated high-voltage circuit breaker model ZW32-12F. Vibration signals at different operating positions of the spring-operated mechanism were collected using TX9R033-2 vibration sensors. As shown in Fig. 1, the vibration sensors were installed at various positions on the operating mechanism and the chassis using strong magnets attached to the bottom of the sensors. By deploying multiple vibration sensors, vibration signals from different positions on the operating mechanism and the chassis were obtained. Additionally, acoustic sensors were used to collect sound signals during the sampling process, although these signals were not used for fault diagnosis. The collected signals were transmitted to a data storage unit via an NI 9234 data acquisition card, with the sampling frequency controlled by edge computing devices set at 25.6 kHz. Finally, the collected data were uploaded to a cloud server. This study captures the vibration signals of the operating mechanism and the chassis during HVCB disconnection by setting a threshold to determine the start time of the vibration signals.

Convolutional attention network

To effectively detect and classify faults in high-voltage circuit breakers, our approach involves a comprehensive analysis of collected data from various locations. These signals are pivotal in determining the operational state of HVCB over time. To achieve this, we have developed a specialized neural network model, termed CANet, designed to extract and process features from these varied inputs and ultimately categorize the different types of faults using a multi-layer perceptron (MLP).

The methodology begins with the selection of vibration signals of a fixed length, ensuring a uniform input structure. These signals are then downsampled at a predetermined frequency to standardize the data and reduce noise, making them suitable for deep learning processing. The downsampled signals are then concatenated, creating a composite signal that combines the characteristics of each modality. This composite signal is fed into the CANet for in-depth feature extraction.

CANet primarily consists of a one-dimensional convolution module and an attention module. In the one-dimensional convolution process, filters slide along the time series, engaging in point-wise multiplication with different parts of the series to effectively capture local dependencies and features. By employing various filters, the one-dimensional convolution can extract a diverse range of features from the raw time series. The attention mechanism enables the model to focus more on the parts that are most critical for the current task while processing information. When handling time series data, the attention mechanism is particularly useful in identifying the information most relevant to the current task. Compared to traditional convolutional networks, the attention mechanism offers a more dynamic way of extracting features, as it can adaptively adjust its focus on different parts of the series based on the task requirements and contextual changes.

The architecture of CANet is designed to handle the complexity of multi-point data. The input tensor $\in R^{L\times 4}$, where L represents the signal length, is first subjected to one-dimensional convolution. This process is pivotal for extracting and expanding features from the vibration data. The convolutional layer not only captures the inherent characteristics of each signal type but also begins the process of blending these features to form a comprehensive fault signature. The vibration signals after feature extraction are shown in Fig. 2.

Following initial feature extraction, the data are fed into the convolutional attention (CA) module for further processing. This module employs both one-dimensional convolution and attention modules to extract spatial and temporal features, offering a more nuanced understanding of the data. The convolutional component of the CA module works to fuse features across different channels, effectively synthesizing information from vibration signals collected at various locations. The attention module then comes into play, focusing on crucial segments of the data sequence. It identifies patterns and correlations over time, bringing out subtle but significant anomalies indicative of potential faults.

Finally, the processed data are passed through a final one-dimensional convolutional layer. This layer integrates the temporal information and prepares the data for fault classification. The MLP module then takes over, applying its powerful classification capabilities to accurately categorize the faults in HVCB.

CA module

As shown in Fig. 3, the CA module we designed utilizes one-dimensional convolutional modules and attention modules to effectively extract features from input data across temporal and spatial dimensions, enhancing the interaction between local and global features. Initially, the one-dimensional convolutional module captures local features of the time-series signal, expanding the feature space to facilitate the exchange of information across channels. After convolution, the MLP is used to further enrich the feature tensor, introducing non-linearity to uncover more complex data characteristics. Following this, the application of the attention module allows us to focus on crucial parts of the sequence, capturing global correlations within the time-series signal.

We employ a multi-head attention module to adaptively learn different features. The tensor is passed through three separate FFN modules, generating Query, Key, and Value respectively. We first multiply Query and Key, and then apply a Softmax function to generate the corresponding heatmap. Following this, we multiply the heatmap with Value. Finally, we concatenate the results along the channels and project them through an FFN for mapping. The specific expression formula is as follows:

(1) $$\begin{array}{rl}MultiHead(Q,K,V)& =Concat(hea{d}_1,...,hea{d}_h)W^o\\ hea{d}_i& =Attention(Q{W}_i^Q,K{W}_i^K,V{W}_i^V)\\ Attention(Q_i,K_i,V_i)& =softmax\left({\displaystyle \frac{Q_i{K_i}^T}{\sqrt{d_k}}}\right)V_i\end{array}$$where $h$ denotes the number of heads, $hea{d}_i$ represents the output of the i^th head, and ${\mathrm{W}}^{\mathit{o}}$ is the output transformation matrix. ${\mathrm{W}}_{\mathrm{i}}^{\mathrm{Q}}$, ${\mathrm{W}}_{\mathrm{i}}^{\mathrm{K}}$, ${\mathrm{W}}_{\mathrm{i}}^{\mathrm{V}}$ are the transformation matrices for query, key, and value, respectively, and ${\mathrm{d}}_{\mathrm{k}}$ is the dimensionality of the key tensor.

Ultimately, the features outputted by the convolutional layer are combined with those enhanced by the attention module through residual connections. This not only preserves the original information but also enhances the expressiveness of features, providing robust support for deep learning and fault detection within the model.

Implementation details

Following the installation method described in ‘Signal Acquisition’, we collected a total of 42,468 data entries. A total of 200 vibration signals were collected during the high-voltage circuit breaker operation. Among them, there were 17 sets of vibration signals with contact faults, nine sets with mechanism seizure, 23 sets with bolt loosening, and 10 sets with spring fatigue. Due to the limited number of faulty vibration signal data, it is necessary to augment the data. The method involves overlaying faulty vibration signals with normal vibration signals and adding random noise to generate a completely new faulty vibration signal. Similarly, multiple faulty signals can be overlaid to augment the dataset. To test the performance of our models, we employed the Monte Carlo cross-validation method. We randomly selected 80% of the continuous data from the entire dataset and split it into a training set and a test set in a 4:1 ratio. Each model was trained five times following the aforementioned method, and we saved the weights of the model that performed best on the test set. The average accuracy of each model was calculated and used as a key indicator to measure model performance.

The entire process was conducted in a Python environment, using the PyTorch framework for building, training, and testing the models. We only employed one CA modules to construct CANet. The output dimension of the first convolutional layer is set to 128, with a kernel length of five. In the CA module, the input and output feature dimensions of the first convolution operation are both 128, with a kernel length of five. The MLP is configured with a three-layer structure, where the input and output dimensions are both 128, and the hidden layer dimension is 256. The number of heads in the attention module is eight. In the final convolutional layer, the input dimension is 128, and the output dimension is 256, with a kernel length of 50. The ultimate MLP is also configured with a three-layer structure, where the input dimension is 256, the output dimension is four, and the hidden layer dimension is 512.

During the training process, we trained the model for 100 epochs, set the learning rate to 1e-4, and fixed the batch size at 256. We set the length of the time series data to 50 and used a GTX 2080Ti graphics card for training the models.

Evaluation metrics

To evaluate the performance of our constructed CANet in detecting four types of faults in high-voltage circuit breakers, we utilize accuracy, precision, recall, and the F1-score (F1 measure) as metrics to measure the performance of CANet on our custom-built dataset. The following sections will provide a detailed introduction to these evaluation metrics.

Accuracy refers to the proportion of samples that are correctly classified out of the total number of samples. Precision and Recall are the most commonly used metrics in evaluating classification tasks. Precision is defined as the proportion of true positive samples among all samples predicted as positive. Recall, from the perspective of the original samples, is defined as the probability of samples that are truly positive being predicted as positive. The mathematical definitions of these metrics are as follows:

(2) $$\begin{array}{r}Precision={\displaystyle \frac{TruePositive}{TruePositive+FalsePositive}}\\ Recall={\displaystyle \frac{TruePositive}{TruePositive+FalseNegative}.}\end{array}$$

To take into account both Precision and Recall, the F1-score is commonly used as a measure. The F1-score is the harmonic mean of Precision and Recall, designed to balance the two metrics. It is defined as:

(3) $$F1=2\cdot {\displaystyle \frac{Precision\cdot Recall}{Precision+Recall}.}$$

Experiments

In our dataset, we randomly allocated 80% of the data for training and testing purposes, and conducted fivefold cross-validation. The CANet produces an output tensor of length four, representing four types of faults: contact fault, mechanism seizure, bolt loosening and spring fatigue. In this study, we adapted existing models—LSTM, GRU, TIMESNet (Wu et al., 2022) and Anomaly Transformer (Xu et al., 2022)—by modifying their source code to accept our data for fault detection. We then conducted a comparative analysis across these models and CANet, evaluating their performance on four types of mechanical faults in HVCB in terms of precision, recall, and F1-scores. The result is as shown in Table 1. The results demonstrate that CANet slightly outperforms Anomaly Transformer in terms of overall recall and F1-score, indicating a superior ability to detect all relevant cases and maintain a balance between precision and recall. While Anomaly Transformer exhibits a marginally higher average precision, suggesting fewer false positives, the differences between the two methods are relatively minor.

The classification results of CANet are presented in Table 2. In Table 2, we present the accuracy, recall, precision, and F1-score of CANet across fivefold cross-validation. Averaging the results of the fivefold cross-validation, we obtained an accuracy of 0.942, a recall of 0.972, a precision of 0.970, and an F1-score of 0.971. The high accuracy indicates that CANet performs well across the entire dataset. The high recall and precision demonstrate that the model effectively identifies and classifies positive samples with minimal misclassification of negative samples as positive. An F1-score of 0.971 reflects a good balance between precision and recall, showcasing the model’s excellent overall performance.

In Table 3, we present the average detection performance metrics of CANet for four types of faults under fivefold cross-validation. The faults include contact fault, mechanism seizure, bolt loosening and spring fatigue. We list the Precision, Recall, and F1-score for the detection of these four faults. Among these faults, CANet demonstrates the best detection performance for spring fatigue, with Precision, Recall, and F1-score all approaching one. The detection performance for the other three faults is also high, with Precision and Recall exceeding 0.95, and F1-scores close to or exceeding 0.95.

In Table 4, we have listed the F1-scores for four types of faults in each round of the fivefold cross-validation. “C. F1-score” represents the F1-score in detecting contact fault, “M. F1-score” for mechanism seizure, “B. F1-score” for bolt loosening, and “S. F1-score” for spring fatigue. The F1-scores for all four types of fault detection are generally very high, indicating that the model possesses a high level of accuracy and reliability. The model’s performance is relatively stable across different rounds, with only minor fluctuations occurring in a few instances. However, these fluctuations are minimal, demonstrating the model’s robustness in detecting various types of faults. For the mechanism seizure, the F1-scores are slightly lower in some rounds (dropping to as low as 0.91), which could suggest that this type of fault is more challenging to detect compared to others, or that the dataset features for this part are not sufficiently distinct.

In Tables 5 and 6, we have listed the precision and recall for four types of faults in each round of the fivefold cross-validation. CANet achieves a Precision and Recall of over 0.95 in detecting these four types of faults. This indicates that CANet is both accurate and comprehensive in its fault detection capabilities, particularly excelling in the detection of spring fatigue. The minor variations in Precision and Recall across different data splits suggest that the model’s performance is stable and not significantly affected by specific data partitions, demonstrating good robustness.