An aero-engine remaining useful life prediction method based on deformable convolutional residual attention enhanced Kolmogorov-Arnold networks with prediction advance and shape constraints

Improving the practicality and reliability of remaining useful life (RUL) prediction requires achieving more accurate, earlier, and more relevant forecasts. However, existing methods are limited by the weak feature extraction ability of fully connected layers and the inability of conventional loss functions to balance early prediction and trend consistency. To address these challenges, this paper proposes a prediction model base on deformable convolutional residual attention (DCRA) modules with Kolmogorov–Arnold networks (KAN). An early prediction and shape constraint loss function (ESLF) is further designed to enhance model reliability. Specifically, the DCRA module adaptively extracts multi-scale degradation features through deformable convolution kernels, while KAN serves as a regressor to improve feature mapping. Moreover, the main spline function of KAN reflects the occurrence of degradation, thereby improving model interpretability. The proposed ESLF is asymmetric and improves the similarity between predicted and actual RUL sequences. Comparative and ablation experiments conducted on the CMAPSS and N-CMAPSS datasets verify the superior prediction performance of the proposed method. The average RMSE and Score are reduced by 5.19 % and 9.28 % on CMAPSS, and by 4.63 % and 9.47 % on N-CMAPSS, respectively. These results demonstrate the effectiveness and applicability of the proposed approach in practical health management scenarios.