A TCN-KANformer-MDN model for engine remaining useful life prediction

Abstract Accurate prediction of the remaining useful life of aircraft engines is critical to the safety of aircraft operations. Targeting the two problems of modeling complex nonlinear degradation relationships and quantifying prediction uncertainty, this paper proposes a novel RUL prediction model that integrates a temporal convolutional network (TCN), KANformer, and a mixture density network (MDN). In the proposed framework, the TCN is first employed to extract multi-scale temporal dependency features from the degradation sequences. Subsequently, the KANformer is introduced to perform nonlinear transformations on the extracted features, leveraging its learnable basis functions to enhance the model’s capacity for characterizing complex degradation patterns. Finally, the MDN outputs a probability density distribution of the RUL, thereby enabling quantitative uncertainty quantification of the predictions. Experimental validation on the NASA C-MAPSS dataset demonstrates that, in comparison with benchmark methods such as Cau-DA-LSTM, the proposed TCN-KANformer-MDN model achieves approximately 10% improvement in both prediction accuracy and uncertainty quantification metrics, fully corroborating its effectiveness and robustness under complex operating conditions.