Modern industrial systems rely on condition monitoring (CM) to identify potential failures and ensure operational efficiency. Fault detection and diagnostics (FDD) aim to identify anomalies and diagnose their root causes, providing actionable insights for timely maintenance and minimizing downtime.
Challenges in Fault Detection and Diagnostics
Industrial systems often operate under varying environmental and operational conditions, which introduce challenges such as:
- Variability in Data: Sensor readings can be influenced by external factors such as temperature, humidity, and load conditions. (see this part)
- Rare Fault Occurrences: Faults are infrequent, making it difficult to collect representative datasets for training models.
- Complex Fault Patterns: Different fault types can manifest in similar ways, making fault isolation challenging.
Approaches to Fault Detection and Diagnostics
Our research explores various approaches to improve fault detection and diagnostics, including:
- Generative AI: Leveraging models like Stable Diffusion and Wasserstein GANs to generate realistic unseen anomalies conditioned on normal samples and enable domain adaptation for fault detection across different operating environments [1,2].
- Residual-Based Approaches: Comparing autoencoders and input-output models to establish a mapping between operating conditions and sensor readings, enabling unsupervised fault detection and interoperability [3]
- Contrastive Learning: Using triplet loss-based feature representation learning to achieve invariance to changing operating conditions and sensitivity to novel fault types [4].
- Graph Neural Networks: Developing dynamic graph-based models such as DyEdgeGAT to track evolving relationships between sensor signals, enhancing fault detection under varying operating conditions [5].
By integrating these approaches, our research aims to develop practical, scalable, and interpretable solutions for fault detection and diagnostics, enabling industries to move towards predictive maintenance strategies.
Papers:
[1] Sun, H., Cao, Y., & Fink, O. (2024). CUT: A Controllable, Universal, and Training-Free Visual Anomaly Generation Framework. arXiv preprint arXiv:2406.01078.
[2] Rombach, K., Michau, G., & Fink, O. (2023). Controlled generation of unseen faults for partial and open-partial domain adaptation. Reliability Engineering & System Safety, 230, 108857.
[3] Hsu, C. C., Frusque, G., & Fink, O. (2023). A comparison of residual-based methods on fault detection. arXiv preprint arXiv:2309.02274.
[4] Rombach, K., Michau, G., & Fink, O. (2021). Contrastive learning for fault detection and diagnostics in the context of changing operating conditions and novel fault types. Sensors, 21(10), 3550.
[5] Zhao, M., & Fink, O. (2024). Dyedgegat: Dynamic edge via graph attention for early fault detection in iiot systems. IEEE Internet of Things Journal.