Physics-guided high-value data sampling method for predicting milling stability with limited experimental data

Accurate milling stability prediction is necessary for selecting chatter-free machining parameters to ensure the machining quality. With the development of machine learning techniques, data-driven methods have demonstrated powerful modelling capabilities for stability prediction. However, the significant performance of data-driven modelling usually requires a large labelled training dataset consisting of stable and unstable experimental data, which is expensive and time-consuming for metal-cutting scenarios. Therefore, how to design an experimental parameter set to build the experimental labelled dataset which is small but can provide sufficient support for data-driven stability prediction has been a critical problem and has received increasing attention. Existing research samples the experimental parameters by the grid or the boundary method, which inevitably brings lots of low-value data points for model training. To address this, this paper proposes a Physics-Guided High-Value (PGHV) data sampling method to reduce the required experiments for data-driven stability prediction. A novel value function is designed based on the physics information of milling dynamic stability to quantify the potential contribution of different experimental parameters. The optimal experimental parameter set can then be determined by maximising the dataset value. After that, the experimental labelled dataset can be constructed by performing cutting experiments under the sampled experimental parameters. Finally, the stability prediction model can be obtained by the data-driven modelling method with the experimental labelled dataset. Experimental verification shows that the proposed method can reduce the number of experiments by more than 60% compared to the existing sampling methods.