Acoustic monitoring of tool wear state with limited samples based on time-series generative adversarial network

Tool wear is unavoidable for mechanical manufacturing. The accurate tool wear prediction can reduce production costs, enhance production efficiency, and facilitate high-quality production. Acoustic monitoring using microphones is a reliable, efficient, and economical solution. Nevertheless, it is challenging to obtain an adequate quantity of sound data regarding the degradation process of the tool under laboratory conditions. To address the issue of limited data size for training models, a time-based data enhancement method based on a time generative adversarial network (TimeGAN) is proposed to expand the experimental data set. A visual evaluation of the generated data indicated that the model exhibits robust learning generalization capabilities. To ascertain the efficacy of the generated data for tool wear classification, the convolutional neural network (CNN) model is trained with the generated data, and the original experimental data is employed to test the model. The findings demonstrate that this approach can markedly enhance the precision of tool wear classification and offer a novel perspective on tool wear monitoring in the context of limited data sets.