Explainable Prediction of Remaining Useful Life for Turbofan Engines Using Kolmogorov–Arnold Network

This paper proposes a Remaining Useful Life (RUL) prediction framework combining Kolmogorov–Arnold Network (KAN), LSTM and BiLSTM to improve prediction accuracy and interpretability for turbofan engines. The proposed framework uses KAN to perform nonlinear mapping of sensor and operating-condition inputs before temporal modeling. LSTM and BiLSTM then extract the temporal degradation features, with the final forward and backward hidden states combined for RUL prediction. In terms of interpretability, this work analyzes sensor importance within the KAN representation through structural attribution. In addition, t-SNE feature-space visualization, SHAP attribution of the complete prediction model and XGBoost analysis are incorporated to provide complementary evidence for interpretation. Fixed-model sensor-retention experiments further examine whether sensors ranked highly by KAN preserve predictive information. Experimental results demonstrate favorable prediction accuracy and low variability across repeated training runs on the four C-MAPSS subsets. Under matched parameter counts, placing KAN before temporal modeling reduces FD001 mean RMSE by 23.83% relative to placing it in the regression stage. Further analysis identifies recurring pressure, flow, rotational speed and temperature variables, whose relative contributions vary across degradation stages and operating conditions. The KAN-ranked Top-10 sensors achieve a lower RMSE than the mean of equal-size random sensor sets and the low-ranked sets on all four subsets under fixed-model retention. The proposed method provides an integrated approach to prediction and sensor-level interpretation, offering support for turbofan-engine health assessment and maintenance planning.