(314d) Interpretable Fault Detection in Chemical Processes Using Large Language Models | AIChE

(314d) Interpretable Fault Detection in Chemical Processes Using Large Language Models

Authors 

Khan, A., Purdue University
Chen, H., Purdue University
Constante-Flores, G., Purdue University
Process operators of the chemical industry face the risk of safety hazards from different sources, e.g., equipment failure, switching raw material compositions, disturbance from extreme weather, human operating errors, etc. Failing to detect the abnormal operating conditions that cause safety hazards can lead to disastrous accidents. Early detection and identification of safety hazards can prevent economic and life losses. This research addresses the vital need for interpretable anomaly detection using artificial intelligence. Traditional methods have excelled in detecting anomalies [1-3] but often fall short in translating these findings into actionable insights for operational personnel. Our study bridges this gap by innovatively integrating Principal Component Analysis (PCA) with Large Language Models (LLMs), focusing on both detection and interpretability of process safety data.

The methodology introduces a refined approach to anomaly detection using PCA. It effectively identifies operational deviations, and through the analysis of T² contributions [3], it enables precise identification of plausible features contributing to these anomalies. Following the detection phase, the application of LLMs plays a pivotal role in deciphering the complex interrelations of these identified features. By employing causal graphs [2] and process schematics in conjunction with LLM analysis, our approach provides a comprehensive understanding of the interactions and dependencies among various process elements. This not only aids in pinpointing the root causes of anomalies but also enhances the overall interpretability of the data, making it more actionable for process engineers and safety managers. The LLM can provide recommendations on the strategies to mitigate the fault, thus preventing similar fault to occur again.

We used the Tennessee Eastman Process (TEP) to demonstrate our proposed method. The OpenAI GPT-4 model is the LLM used in the study. The study's results indicate that the synergistic application of PCA and LLMs significantly elevates the interpretability of process data, a key element in transforming complex analytics into practical tools for process safety management. The impact of this research extends broadly, proposing a shift in the approach to process safety within the chemical processing sector. By elevating the clarity of data analysis, this research establishes a new benchmark for operational safety, potentially shaping future safety practices and methodologies within the chemical industry and beyond.

[1] Venkatasubramanian, V., Rengaswamy, R., Yin, K., & Kavuri, S. N. (2003). A review of process fault detection and diagnosis: Part I: Quantitative model-based methods. Computers & chemical engineering, 27(3), 293-311.

[2] Suresh, R., Sivaram, A., & Venkatasubramanian, V. (2019). A hierarchical approach for causal modeling of process systems. Computers & Chemical Engineering, 123, 170-183.

[3] Chiang, L. H., Russell, E. L., & Braatz, R. D. (2000). Fault detection and diagnosis in industrial systems. Springer Science & Business Media.

[4] Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., ... & McGrew, B. (2023). Gpt-4 technical report. arXiv preprint arXiv:2303.08774.

Topics