A dual-path time–frequency integration architecture applied to remaining useful life prediction of air turbine starter bearing

Abstract The air turbine starter is a critical device for starting aircraft engines, in which rolling bearings play a core role. Remaining useful life (RUL) prediction methods primarily focus on time-domain signals and overlook the potential of the frequency domain, restricting their ability to comprehensively capture complex bearing degradation characteristics. This paper introduces a dual-path time–frequency integration (DPTFI) architecture to efficiently fuse time and frequency features. Firstly, the DPTFI architecture applies a structural adjustment strategy to enrich frequency information while improving computational efficiency. Subsequently, the preprocessed data are fed into dual paths independently; the frequency-domain path uses a frequency weighting mechanism to extract global spectral patterns, while the time-domain path employs a multi-layer perceptron to capture dynamic temporal features. Finally, a bidirectional cross-attention mechanism fuses features from both paths to enhance degradation trend representation, further refining the modeling process to predict the RUL. The experimental results demonstrate that the proposed method achieves reductions in mean absolute error by 17.46%, 21.69%, 33.16%, 36.59%, and 50.19%; root mean square error by 25.19%, 29.41%, 36.58%, 40.71%, and 50.17%; and score by 5.97%, 23.65%, 36.16%, 40.72%, and 49.49% compared to the adaptive time–frequency ensembled network (ATFNet), joint time–frequency domain transformer (JTFT), time-series mixer (TSMixer), TimesNet, and convolutional neural network models, respectively. The effectiveness of the proposed method is validated through experimental data from the RUL prediction of civil aircraft bearing components, showcasing its ability to achieve high-precision predictions of bearing RUL.