Research on Tool Wear State Prediction Model Based on Knowledge Distillation

Aiming at the problem of high annotation cost and insufficient cross-case generalisation capability faced by tool wear state prediction in industrial scenarios, a model optimisation method based on knowledge distillation is proposed. By constructing a bi-directional gated recurrent unit (BiGRU-GRU) teacher model and a student model of Transformer architecture, the response-based KL scatter distillation, feature-based Euclidean distance distillation and cross-entropy supervised training are jointly adopted, and a multi-objective joint loss function is designed to achieve knowledge migration. The experimental results show that knowledge distillation generally improves the generalisation performance of the Transformer-FC, Transformer-GRU, and Transformer-BiGRU student models, which increases the F1 value of the Transformer-GRU model from $\mathbf{6 9. 0 9 \%}$ to $\mathbf{7 2. 3 6 \%}$.