Abstract: Aiming at frequent elevator door faults and failure of traditional regular inspection to realize all-weather monitoring, this paper proposes a non-contact monitoring scheme combining multi-modal sensing and edge AI computing. Under normal working conditions without manual intervention, the system accumulated and analyzed roughly 6 000 complete elevator operation cycles. Using real data, we adopt the YOLO vision model to precisely locate landing and car door image anomalies. Meanwhile, with acoustic RMS extraction, ambient brightness thresholding and K-means spatial clustering, a joint discriminant system is built for high-risk defects including excessive door gap, slow door movement, mechanical noise and floor positioning errors. Field tests verify that the designed system delivers stable anti-interference and robustness under complex conditions, offering a viable path for data-driven safety supervision research on special equipment.