Accurately forecasting the Remaining Useful Life (RUL) of critical engineering systems is essential for effective Prognostics and Health Management (PHM), ensuring reliability, safety, and cost efficiency. Conventional deep learning approaches, such as Long Short-Term Memory (LSTM) networks and Convolutional Neural Networks (CNNs), often struggle to simultaneously model long-term dependencies and intricate nonlinear relationships in sensor data. To address these challenges, we propose Transformer-KAN (TransKAN), a novel hybrid model that integrates Transformer-based temporal feature extraction with the Kolmogorov–Arnold Network (KAN) for enhanced non-linear feature mapping. By leveraging the multi-head attention mechanism, the Transformer module efficiently learns long-range dependencies and degradation patterns in time-series data, while KAN enhances non-linear representation learning by bridging approximation theory with modern machine learning. We assess TransKAN on the widely used CMAPSS dataset, comparing its performance against leading deep learning models. Experimental results highlight its superior accuracy in RUL estimation, reinforcing its effectiveness in predictive maintenance.

remaining useful life,Transformer,Kolmogorov–Arnold Network,deep learning,attention mechanism