Abstract: Aiming at the difficult monitoring problems of equipment status under complex environments such as low illumination, high dust and others in coal mine downholes, constructs an automatic diagnostic system that integrates video enhancement, spatiotemporal feature extraction, and self evolutionary learning. Methodologically, a two-step type image enhancement strategy is adopted to optimize image quality, and a feature extraction network integrating lightweight 3D convolution and spatiotemporal attention is built to strengthen the feature capture capability of transient faults; Model dynamic updates are achieved and a small sample incremental learning mechanism is introduced through an edge-cloud collaborative architecture, data upload and retraining are triggered based on fault confidence degree to improve model adaptability and scalability. Experimental verifications in multiple scenarios such as belt conveyor, roadheader, emulsion pump station and others indicate that the accuracy and real-time performance of this method for fault identification are better than the comparison scheme, with a inference latency of less than 40 ms and a false alarm rate and missed alarm rate are all controlled within 3%, meeting the requirements of downhole deployment, this system has good environmental adaptability and engineering feasibility.