In order to improve the operation and maintenance level of shearer, this paper proposes a predictive maintenance hierarchical framework, which decomposes the operation and maintenance tasks into four levels : component-level diagnosis, system-level prediction, multi-source fusion evaluation and structured rapid diagnosis, and establishes task-model adaptation mapping. Based on the CWRU bearing data set, the high noise environment in the mine is simulated, and the performance of the four models of 1D-CNN, LSTM, CNN-LSTM and LightGBM is compared. The experimental results show that the accuracy of 1D-CNN component-level diagnosis in a clean environment is 99.91% ; in terms of noise robustness, LSTM has the best comprehensive score (0.755 0), followed by CNN-LSTM (0.491 4), both of which are significantly better than pure CNN and LightGBM, revealing the key role of temporal modeling ability in noise robustness. This study provides a framework design and model selection basis for predictive maintenance systems in high-noise industrial scenarios.