Infrared thermal imaging is the mainstream technology used in power plants for fault detection. But the conventional methods have trouble with analyzing the operating information for massive infrared images and ignore the performance degradation mechanism, resulting in a rough diagnosis. To tackle the two problems, a fine diagnosis method based on spatiotemporal feature extraction and Shapley additive explanation attribution clustering algorithm is proposed. Considering the spatiotemporal characteristics and process knowledge, the fine degradation model including the normal stage, attention stage, warning stage, and abnormal stage is established, which provides a basis for the predictive maintenance of power equipment. The broad learning system with incremental learning capability is employed to solve the model adaptation problem caused by the increase in equipment types and fault number, which realizes the rapid update of the model.