A new method for bearing remaining useful life prediction based on dynamic wavelet and physical information constraints

Accurate prediction of bearing remaining useful life (RUL) in mechanical systems represents a pivotal challenge for enabling reliable operational maintenance. In recent years, RUL prediction methods fusing physical mechanisms with deep learning have shown significant potential, but still exhibit significant limitations in terms of dynamic adaptation when dealing with strong nonlinear degradation feature extraction of bearings. In addition, model stability is not always guaranteed due to the lack of deeply fused physical information. To address these issues, this study introduces a hybrid RUL prediction framework that synergizes dynamic signal processing with physics-informed learning. Initially, a dynamically differentiable wavelet decomposition module is proposed, which supersedes conventional wavelet bases by implementing an adaptive filter bank architecture with dynamic parameter tuning, enabling multi-scale dynamic analysis of vibration signals. The module incorporates a dual-channel enhancement mechanism to amplify localized low-frequency features while integrating frequency-domain gated filtering to mitigate high-frequency noise interference. Subsequently, a bidirectional adaptive graph convolutional network is developed to model spatiotemporal interdependencies within multi-scale degradation features, employing learnable adjacency matrices to dynamically characterize evolving correlation patterns during bearing deterioration. Finally, a dual-constrained physical information loss function is introduced, which enhances the stability and accuracy of RUL prediction by integrating physical principles into model training. Experimental results on two publicly available datasets demonstrate the significant effectiveness and superiority of the proposed method.