As a core component of motors, bearings primarily serve functions such as supporting and guiding shafts, reducing friction in equipment, and connecting different components. Predicting the remaining life of bearings is crucial for system health management. However, single sensor signals often fail to comprehensively describe the potential degradation mechanisms of the system. This paper proposes a novel approach for predicting the remaining life of motor bearings based on the multi-head attention mechanism and long shortterm memory neural network. Firstly, Mahalanobis distance is used to determine the starting point of bearing performance degradation by dividing the entire life cycle of rolling bearings into normal and degradation phases. Secondly, an Autoencoder is employed to automatically extract vibration signal features, which are subsequently fused with motor current and bearing temperature signal to construct a multi-source information feature matrix. Subsequently, the multi-head attention mechanism and long short-term memory network dynamically select features with high relevance, thereby improving the accuracy of the remaining life prediction. Finally, the model is validated using experimental data, and the results show that the proposed model has higher accuracy.