The high variability in fatigue life of metal rubber (MR) components presents a substantial challenge for conventional prediction models, which often fail to account for individual specimen differences, resulting in diminished predictive accuracy. To address this limitation, this study conducts systematic fatigue life testing on MR specimens and extracts 20 statistical features from both temporal and spectral domains to characterize fatigue degradation processes. Through comprehensive evaluation of feature monotonicity, noise robustness, and temporal correlation characteristics, we implement principal component analysis (PCA) for dimensional reduction to optimize input data quality. Building upon this foundation, we develop a novel prediction framework utilizing temporal convolutional networks (TCN) that effectively incorporates real-time operational data. The proposed TCN-based model demonstrates enhanced capability in capturing individual component variations and mitigating prediction errors induced by lifespan dispersion. Experimental results highlight the critical role of feature selection and dimensionality reduction in model optimization, with the final implementation achieving a 10.90% global prediction error. This research establishes a novel methodology for fatigue life prediction and maintenance optimization of MR components, offering significant improvements over traditional approaches through synergistic integration of multidimensional feature analysis and deep learning techniques.
Chen et al. (Sat,) studied this question.