Intelligent Predictive Maintenance (IPdM) has garnered significant attention from the industrial sector due to its immense potential in reducing unexpected downtime and maintenance costs of industrial equipment. Existing IPdM methods have limited perception capabilities for vast equipment status sequence data, and they cannot predict future equipment health status in a timely and effective manner. There is considerable room for optimization in terms of model accuracy, adaptability, and interpretability. In this paper, we propose a Time Series Perception and Prediction (TPP) method based on self-attention for immediate IPdM. This method consists of two submodules: the Health Status Perception module (TP1) and the Time Series Prediction module (TP2). Specifically, by predicting a period of time series through TP2 and inputting it into TP1 for perception, we can achieve the prediction of future health status. Employing a refined attention framework, we’ve identified key temporal relationships in time series data to inform both long-term and immediate maintenance strategies. To validate this method, we introduce a case study on predictive maintenance of the stator cooling system during the operation of a linear induction motor. The results show that the TPP scheme has good predictive accuracy and can effectively reduce the number of unplanned interruptions for this industrial equipment.