Abstract: Unsafe driving behaviors significantly contribute to accidents in open-pit mining operations. Conventional monitoring systems often fail to capture the temporal continuity and semantic ambiguity of such behaviors. To address these challenges, an event-level abnormal driving behavior recognition method based on a vision-language model (VLM) is proposed. The method integrates Temporal-Aware Low-Rank Adaptation (T-LoRA) with event-level supervised fine-tuning to enable efficient vertical-domain customization of the VLM toward complex open-pit operating environments. By explicitly enhancing temporal modeling of visual sequences, the proposed method enables accurate event-level recognition of abnormal driving behaviors, precise localization of their onset and duration, and the generation of interpretable semantic descriptions. Experiments conducted on a real-world mining truck dataset demonstrate that the proposed approach achieves an event-level F1-score of 0.923, while maintaining a mean absolute error for temporal localization below 0.8 s. These results indicate that the proposed method effectively captures continuous behavioral patterns and improves both recognition accuracy and temporal precision, providing a reliable solution for intelligent safety monitoring. Furthermore, the system supports proactive intervention through event-level early warnings, contributing to safer and more efficient mining operations.