SF-CNN-Transformer with spectral cumulative-amplitude for rolling bearing RUL prediction

Abstract Bearing remaining useful life (RUL) prediction is central to prognostics and health management (PHM), and its accuracy depends on informative degradation features and effective temporal modeling. To address the weak trendability of conventional features and redundancy in high-dimensional representations, this study proposes a hybrid deep network, SF-CNN-Transformer, based on a spectral cumulative amplitude feature (SF). SF is constructed by cumulatively aggregating spectral amplitudes, yielding a monotonic and smooth degradation feature representation that enhances degradation interpretability and suppresses noise. The SF is first processed by a 1D-CNN for deep representation learning and dimensionality reduction, and then fed into an improved Transformer integrated with a convolutional autoencoder (CA-Transformer) to strengthen feature discriminability and model long-term temporal dependencies over the full degradation trajectory. Experiments on the PHM2012 dataset show that the proposed method achieves consistently better performance than the considered baselines and representative state-of-the-art methods. Compared with the baseline model, RMSE and MAE decreased by 15% and 14.58% respectively, and R 2 increased by approximately 2.22%, indicating that the accuracy and stability of bearing RUL prediction have been improved.