Leveraging Large Language Models for Fault Detection in Connected Vehicles
Rakesh Das, Henry Griffith, Heena Rathore · 2025
Large Language Models (LLMs) have demonstrated exceptional capabilities across various tasks, including anomaly detection in time-series data, which is even critical for ensuring the safety, reliability, and cybersecurity of connected vehicles (CVs). CVs rely heavily on sensors to perform essential operations, yet these sensors often lack robust safety mechanisms, leaving them vulnerable to adversarial attacks that can lead to incorrect measurements. This study investigates the potential of LLMs in detecting sensor faults in CVs through fined tuning along with various prompting strategies, including zero-shot and chain-of-thought techniques, and compares their performance against traditional machine learning models. Using the TampaCV BSM dataset, four types of faults—Drift, Stuck-at, Hard-Over, and Trend—are introduced to evaluate the fault-detection capabilities of LLMs. While initial results indicate that LLMs underperform compared to traditional models, significant improvements are observed through fine-tuning with fault-initiated datasets.