Innovative Transformer-Driven Remaining Useful Life (RUL) Prediction Enhanced by Adaptive Multi-Scale Feature Engineering

Mechanical equipment often undergoes remaining life prediction (RUL) to ensure reliable, efficient, and optimal performance. A vast amount of industrial measurement data can significantly enhance the effectiveness of data-driven methods for RUL prediction. This study utilizes the Transformer architecture to predict the RUL of the FEMTO-ST bearings dataset, generated from the PRONOSTIA platform, an experimental platform for accelerated bearing degradation testing. Th proposed method introduces four key improvements to the Dual Aspect Self-Attention Transformer (DAST) framework. It is called Multi-scale Feature DAST(MFDAST) and includes the following improvements: (1) multi-scale feature extraction for enhanced performance, (2) an advanced attention mechanism, (3) the use of a Health Index (HI) for precise degradation tracking, and (4) model optimization through a genetic algorithm. The results show that the study comprehensively analyzes global and local features within extensive datasets by augmenting the DAST model with a multi-scale feature encoder layer. This methodological advancement reduces RMSE by 50%, outperforming traditional Recurrent Neural Network (RNN) approaches.