TL;DR
This paper introduces a novel, unsupervised framework for diagnosing faults in high-voltage circuit breakers (CBs) using vibration and acoustic data. By combining a convolutional autoencoder for fault detection with explainable AI (XAI) for diagnostics, the system can identify and interpret potential issues without requiring pre-labeled fault data, making it more practical for real-world applications.
What Problem Does This Paper Address?
Current methods for monitoring the health of high-voltage circuit breakers (CBs) often rely on physical parameters like gas pressure, which only capture a limited range of potential failures and typically require the CB to be offline. While non-intrusive methods like vibration and acoustic sensing offer continuous monitoring capabilities, existing diagnostic approaches are largely supervised. This means they require extensive datasets with known fault types, which are difficult and expensive to obtain in real-world scenarios where faults are rare and diverse. The paper addresses this gap by proposing an unsupervised framework that can detect, segment, and diagnose faults without prior knowledge of specific fault labels, making it more applicable to operational environments.
Methods
The proposed framework consists of three main steps:
- Unsupervised Fault Detection: A Convolutional Autoencoder (CAE) is trained solely on healthy data. When presented with new data, the CAE reconstructs the input. A high reconstruction error (residual) indicates a deviation from the healthy state, signaling a potential fault.
- Unsupervised Fault Segmentation: The latent space features extracted by the CAE from the detected faulty samples are then fed into various clustering algorithms (K-means, OPTICS, SOM). These algorithms group similar faulty samples into distinct clusters, effectively segmenting different fault types without explicit labels.
- XAI-guided Fault Diagnostics: To interpret these clusters and provide actionable insights, an explainable AI (XAI) approach, specifically Integrated Gradients (IG), is employed. A classifier is trained on the latent space features using the cluster assignments as pseudo-labels. IG is then used to generate attribution maps, highlighting the features (e.g., specific time-frequency regions in the spectrograms) that contribute most to a sample being assigned to a particular cluster. These maps help domain experts identify potential fault types.
Datasets
The framework was validated using an experimental dataset collected from a high-voltage CB under laboratory conditions. This dataset included data from healthy operations and artificially introduced fault conditions related to spring and damper mechanisms. The paper does not mention any newly introduced benchmarks.
Key Insights and Ideas
- Unsupervised approach is feasible: The research demonstrates that it's possible to achieve effective fault detection, segmentation, and diagnostics in complex systems like CBs without relying on pre-labeled fault data.
- XAI bridges the gap: Integrating XAI into an unsupervised framework is crucial for making the detected clusters interpretable. It allows domain experts to understand why a fault is flagged and what specific features are indicative of a particular issue.
- Latent space is key: The CAE's ability to compress sensor data into a meaningful latent space is fundamental for both effective fault detection and subsequent clustering.
- Sensor redundancy: While multiple sensors (vibration in different directions and microphone) provide better results, there's significant redundancy. This suggests that careful sensor selection or using a single multi-directional sensor could be a practical compromise.
Key Results
- High detection rate: The CAE achieved a high fault detection rate, correctly identifying approximately 98.21% of faulty samples with a false negative rate of only 1.79%.
- Effective segmentation: K-means clustering achieved high scores (ARI > 0.91, homogeneity > 0.91, completeness > 0.91, v-measure > 0.91) for fault segmentation, indicating well-separated clusters. OPTICS and SOM also showed promising results, with SOM achieving a homogeneity score of 0.9564.
- XAI provides interpretability: The XAI-generated diagnostics matrices successfully highlighted distinguishing features for different fault conditions, aiding domain experts in identifying potential fault types.
- Sensor contribution: Using all three accelerometers and one microphone yielded the best segmentation performance. However, individual accelerometers (especially axial and vertical) also provided comparable results, suggesting that cost-effective solutions are possible.
Limitations and Future Work
- Data scarcity in online settings: While the framework is designed for unsupervised learning, the performance in online settings with limited new fault data can still be challenging, as demonstrated by the merging of some fault clusters.
- Domain gap: The study was conducted on a single CB in a laboratory setting. Future work should investigate the transferability of the framework to different CBs of the same type, different manufacturers, or even different CB types.
- Distinguishing fault severity: The authors suggest future research could focus on distinguishing between different severities of the same fault type and understanding how these severities evolve over time within the clustering space.
- Optimal clustering selection: The paper notes that selecting the optimal clustering method for real-world applications is application-dependent and may require user exploration and combination of methods.
Summary updated