LLM-Based Fault Detection in Connected Vehicle Time-Series Data
Rakesh Das, Henry Griffith, Heena Rathore · 2025
Large Language Models (LLMs) have demonstrated remarkable capabilities in various domains, including time series anomaly detection. Connected vehicles (CVs) rely on this capability to ensure safe operation. While the embedding of time series data prior to LLM processing has been shown to improve performance for various tasks, this technique has not been explored for fault detection within Basic Safety Messages (BSMs), which CVs rely on for their effective operation. This paper addresses this gap by exploring the effectiveness of BSM time series embeddings for LLM-based BSM fault detection. Performance of this approach is benchmarked versus fine-tuning and various prompting strategies, such as zero-shot and chain-of-thought reasoning. The techniques assessed are validated using a modified version of the TampaCV BSM dataset to include four types of faults: drift, stuck-at, hard-over, and trend. We demonstrate that time series embedding improves fault detection performance versus the other considered techniques.