Research on Knowledge- and Data-Driven Collaborative Gear Degradation State Prediction Model Based on KDPINN

Because of the complex and multifactorial nature of gear degradation, purely data-driven prediction models are often limited by their closed box characteristics, resulting in low reliability and limited interpretability. To overcome these limitations, a knowledge- and data-integrated gear degradation state prediction model, termed the knowledge data physical information neural network (KDPINN), is proposed. This model capitalizes on the strengths of physics-informed neural networks (PINNs) in solving partial differential equations (PDEs) by modeling the gear degradation rate as a first-order PDE. The architecture consists of a two-branch neural network: one branch learns the data-driven mapping from the healthy to degraded gear states, thereby predicting the degradation trajectory; the other branch incorporates domain knowledge by solving the degradation-related differential equation, effectively learning the degradation rate as a form of explicit knowledge. This design enables a synergistic integration of data and knowledge, using the degradation rate as the core knowledge representation to jointly drive the prediction process. Experimental results demonstrate that the proposed knowledge- and data-driven collaborative prediction model outperforms conventional data-driven models in terms of prediction accuracy and robustness. Furthermore, to mitigate the decline in performance when transferring the model from a source domain to various target domains, a weight fine-tuning strategy is introduced. This strategy significantly improves the model’s adaptability and predictive performance in cross-domain scenarios, thereby validating the effectiveness and generalizability of the proposed approach.