Abstract: The reliable operation of mining machinery equipment is crucial for ensuring production safety and economic efficiency. Addressing the characteristics of scarce abnormal data and abundant normal operation data in mining environments, this paper designs and implements a predictive maintenance system based on Siamese LSTM reconstruction. The system constructs data pairs between normal and abnormal samples to train a Siamese LSTM network that learns the normal operation patterns of equipment and achieves anomaly detection based on reconstruction errors. The system adopts a weakly supervised learning strategy to effectively reduce data annotation costs and designs an adaptive threshold adjustment mechanism to improve the accuracy of anomaly detection. In the application validation at a large-scale concentrator, the system achieves a fault detection accuracy of 92.8% and a recall rate of 91.7%, effectively reducing unplanned equipment downtime and validating the effectiveness and practicality of the system.