Enhancing the Evaluation of Fault Detection Models in Smart Agriculture Using LLM Agents for Rule-Based Anomaly Generation