Robust Fault Detection for EV Charging Piles: Integrating Fine-Tuned LLMs with Adaptive Unsupervised Learning on Real-World Data

As the penetration of renewable energy and Electric Vehicles (EV) deepens, EV Charging Piles (EVCP) have emerged as critical power electronic interfaces bridging the power grid and end-users, their reliability is increasingly challenged by high-dimensional, imbalanced data and severe label scarcity. To overcome these limitations, we propose EVCPFD-LLM, a pioneering Large Language Model-based (LLM) fault detection framework. Our approach integrates an autoencoder-enhanced feature extractor for unsupervised scenarios, LoRA-based lightweight fine-tuning for resource-constrained deployment, and a dynamic threshold mechanism for adaptive decision-making. Experimental results on real-world datasets demonstrate a 13.44% accuracy improvement and a 70.87% reduction in latency over conventional methods. These findings validate the framework's potential in securing grid stability and advancing intelligent energy management for sustainable transportation.