Internal defects inspection of arc magnets using multi-head attention-based CNN.

• A novel data augmentation strategy named the overlapping sampling augmentation (OSA) technique is designed to generate large amounts of arc magnet data. • The multi-head attention strategy is introduced into the arc magnet defect detection model for the first time to highlight features that play an important role in defect detection. • A comprehensive experimental scheme including single category samples, multi categories samples, small samples, and the coexistence of noise and insufficient data is designed to verify the generalization and robustness of the proposed method from multiple perspectives. The results show that the arc magnet internal defect classification method has good industrial applications prospects. • Construct an acoustic-based detection system regarding arc magnet internal defect. Arc magnets are the key components of various motor machinery, and their internal defects detection is extremely significant for maintaining system performance and ensuring operational safety. In this paper, an end-to-end improved convolutional neural network (CNN) model based on multi-head attention is presented, where features that play a more important role in defect detection could be efficiently highlighted. In addition, owing to the characteristics of strong parallel working ability in multi-head attention, the training process is greatly accelerated. Meanwhile, to meet the requirements of the model on the amount of data, a data augmentation method is designed accordingly. Then, the performance of the constructed framework is verified in different test scenarios. Experiment results demonstrate that the presented approach owns superior inspection performance based on relatively fewer model parameters compared to other existing methods, even under the small sample, intense noise, and the coexistence of noise and insufficient data. [ABSTRACT FROM AUTHOR]