Abstract: Aiming at the problem of frequent failures in elevator door systems and the difficulty of traditional periodic inspections in achieving 24/7 monitoring, this paper proposes a non-contact elevator door system monitoring scheme integrating multimodal sensing and edge AI computing. Under the regular working condition without manual intervention, the system accumulated and parsed about 6000 complete elevator action cycles. Based on real data, this paper introduces the YOLO vision model to accurately locate image anomalies of landing doors and car doors. Meanwhile, by utilizing acoustic audio root mean square （RMS） extraction, ambient brightness threshold determination, and K-means spatial clustering, a joint discrimination system is established for high-risk hazards such as excessive door gap, slow opening and closing, mechanical noise, and floor positioning errors. Through field measurement, it is verified that the system designed in this paper shows stable anti-interference and operational robustness under complex working conditions, which explores a feasible research direction for data-driven research on safety supervision of special equipment.