PAOLTransformer: Pruning-adaptive optimal lightweight Transformer model for aero-engine remaining useful life prediction

Aero-engines are core equipment in aerospace field, and their remaining useful life (RUL) prediction is a critical aspect in spacecraft monitoring and maintenance. Transformer model demonstrates remarkable performance in this domain, but its construction heavily relies on a sizeable number of parameters and the excessive model redundancy may adversely affect its prediction performance. To address this issue, a pruning-adaptive optimal lightweight Transformer (PAOLTransformer) is proposed. The method employs norm information to evaluate the contribution of each element within the model to the outputs and subsequently utilizes structured pruning to eliminate unimportant redundant elements. The pruning procedure is executed automatically by seeking out the optimal compression rate via reinforcement learning with a reward score combining the accuracy and efficiency of RUL prediction. Experimental analysis on the C-MAPSS aero-engine dataset shows that PAOLTransformer significantly improves the performance metrics, including a 7% reduction in prediction error and a 33% reduction in computational complexity compared to the standard Transformer. It follows that the proposed model can achieve an optimal equilibrium between model performance and pruning rate. Furthermore, PAOLTransformer outperforms several advanced models in predicting long and complex time series. Therefore, this study holds significant implications for the implementation of preventive maintenance strategies for aero-engines.