LEAD - Framework for efficient time-series anomaly detection on large scale data using LLMs

Akash Chandrayan, Amir ZIDI, Matthew Reimherr, Anis Mjirda, Abhinav Pradhan

Keywords: anomaly detection, llm, time-series, zero-shot, statistics

TL;DR: LEAD is a two-stage anomaly detection framework that combines lightweight statistical filtering with batch prompting in LLMs to improve efficiency and accuracy of LLM-only anomaly detection pipelines.

Abstract: Condition-based monitoring (CBM) is essential for maintaining high machine uptime in industrial settings. While existing CBM solutions effectively use time-series data (e.g., thermal, vibration, amperage, etc.), these can be enhanced with LLMs to integrate domain knowledge and generate interpretable summaries. However, LLMs often incur higher latency and cost than traditional methods. We thus propose LEAD (LLM Enabled Anomaly Detection), a two-stage framework. The first stage acts as a screening step, detecting soft anomalies using lightweight statistical methods. The second stage leverages LLMs and batches multiple time series in the same prompt to generate final anomalies. We show that the combination of statistical filtering and batching leads to a more efficient and accurate anomaly detection pipeline. Applying LEAD to industrial motor amperage data improves precision from 27% (unsupervised deep learning) and 39% (LLM-only) to 72%, while reducing latency 14 fold and token usage by 12 fold compared to an LLM-only baseline. Lastly, we demonstrate that LEAD’s accuracy gains from statistical filtering and batching hold even on public datasets.

Submission Number: 54