A Mobile Switched Attention Network for Defects Classification on Co-Fired Piezoelectric Actuators

Zhuolu Li, Jianbo Yu, Zhenhua Wang, Xiaofeng Yang, Zhi Li

Abstract: Efficient feature extraction is vital for assessing classification performance in deep learning (DL) networks. While attention mechanisms enhance feature extraction, their fixed structures limit performance across diverse tasks. In this article, a mobile switched attention network (MSAN) is proposed for the classification of surface defects on co-fired piezoelectric actuators (CPEAs), which employ weakly supervised training with only image-level labels. The MSAN utilizes a novel training strategy with a switching attention block (SAB), allowing dynamic adjustments to the attention block’s structure in real time using switch cells. By employing this adaptive approach, the focus on key feature information is enhanced, and the model parameters are optimized, thereby significantly improving the feature extraction ability. Meanwhile, a weight-sharing testing strategy is implemented, allowing the simple configuration network to utilize the optimal parameters for accurate decision making. Additionally, a data cleaning module (DCM) and a visualization method are introduced to mitigate external environmental noise during network training and increase network interpretability, respectively. Experiments are conducted to validate the effectiveness of the proposed MSAN.