MIT Develops AI Tool for Anomaly Detection in Complex Systems

SigLLM, a new method developed by MIT researchers, uses large language models (LLMs) to detect anomalies in complex time-series data without needing costly training. The way the framework work is converting time-series data into text inputs for LLM processing. This was tested through two approaches 1) Prompter where the prepared data was fed into the model and prompted to locate anomalous values and 2) Detector, where the LLM was used as a forecaster to predict the next value from a time series. The predicted value was compared to the actual value and large discrepancies suggested that the real value is likely an anomaly.

This method could be used to help flag equipment issues in wind turbines, satellites, and heavy machinery early on. While SigLLM is not yet outperforming state-of-the-art deep learning models, it shows promising results and could offer a more accessible, off-the-shelf solution for anomaly detection. Future work on this aims to improve the speed, accuracy, and explainability to better support technicians in preventing failures.