Adaptive weighted generative adversarial network with attention mechanism: A transfer data augmentation method for tool wear prediction.

The on-machine monitoring of tool wear is of great significance to improve the machining efficiency and reliability of CNC machine tools. Although numerous approaches have been developed for condition monitoring of milling tool, two major problems still exist: (a) The actual manufacturing data is complicated due to the different combinations of the cutting speed, cutting feed, and depth of cut. (b) Abnormal data collection is extremely expensive and difficult to obtain. To fully explore the transferable wear-related information from multi-source data, an attention-based cross-domain generative adversarial network is developed in this study. First, an adaptive weighted feature selection network based on attention mechanism is established to extract shared features from source and target domain. Second, an auxiliary classifier generative adversarial network is introduced to utilize the shared features for transfer data augmentation. Finally, a new objective function of generative adversarial network's discriminator is built with correlation alignment regularization term to further utilize the wear-related information for improving the accuracy of tool wear prediction. Experiments are conducted on a machine tool to verify the effectiveness of the proposed cross-domain adaptive generation adversarial network based on the attention mechanism (CDAGAN) method. The results show that the proposed method can identify different tool wear states and capture the specific tool wear frequency feature on multi-machining parameters. • A new adaptive weighted generative adversarial network named CDAGAN is proposed. • The method can extract wear-related transferable features from the frequency spectrum for multi-working conditions tool wear prediction. • A spectrum similarity evaluation method is proposed to assess the data augmentation quality. [ABSTRACT FROM AUTHOR]