Bayesian Optimization Driven Data-physics Hybrid Digital Twin Self-evolution Modeling Method for Tool Wear Prediction

Self-evolution is a critical capability for Digital Twin (DT) to maintain high fidelity amidst the stochastic degradation of complex equipment. However, achieving this capability within data-physics hybrid models faces a challenging ”stability-plasticity” dilemma, which requires balancing physical consistency with adaptive learning from non-stationary sensor streams. To address this, we propose a novel framework that synergizes a physics-guided generative deep neural network architecture with a surrogate-assisted evolutionary optimization strategy. First, we develop a Physics-Guided Lightweight Temporal Convolutional Transformer (PGLT-Transformer). By replacing the conventional Transformer encoder with a Temporal Convolutional Network (TCN) module and incorporating Grouped-Query Attention (GQA), this architecture embeds physical features directly into a compact deep learning structure, ensuring both interpretability and computational efficiency. Second, we formulate the self-evolution of the hybrid model as an expensive black-box optimization problem. A Bayesian Optimization (BO)-driven surrogate engine is introduced to co-optimize the physics-loss regularization and the neuron re-initialization ratio for continual learning. This mechanism efficiently navigates the non-convex search space with minimal function evaluations, overcoming the prohibitive costs of standard evolutionary algorithms. Experiments on a real-world CNC machine tool wear dataset demonstrate that the framework significantly outperforms state-of-the-art methods. The results validate that combining physics-informed modeling with surrogate-assisted optimization provides a trustworthy and generalizable pathway for the self-evolution of industrial digital twins.