A Lightweight Microscopic Metal Fracture Classification Method With Structure Pruning and Knowledge Distillation for Embedded Devices

Microscopic metal fracture (MF) recognition is essential for modern intelligent fault diagnosis. Existing deep learning (DL) methods have achieved high recognition performance but are hard to apply on industrial resource-limited devices due to being overparameterized. Although model pruning methods can handle the above problem to some extent, lightweight models still suffer from feature redundancy and suboptimal performance. To address the above difficulty, we propose a lightweight microscopic MF classification (LMMFC) method for embedded devices via three components work synergistically: First, a hybrid pruning rule (HPR) is designed to enhance the effective pruning ratio by measuring channel importance and feature similarity. Second, an automatic search unit (ASU) is introduced for the automated exploration of distance functions through the Gaussian process (GP) surrogate-assisted model. Third, a progressive knowledge distillation fine-tuning (PKDFT) strategy is given to improve the performance of lightweight models by making full use of intermediate information during the pruning process. Experiments on the MFSEM image dataset and the publicly available dataset show that the proposed LMMFC not only performs optimally in MF recognition tasks but also has superior generalization performance. The ablation study verifies the parameter sensitivity and the role of each module. A visualization study shows the model’s internal changes at different pruning stages. The application case study illustrates the relationship between recognition results and fractographic control. This work demonstrates the potential of using lightweight DL-based methods for microscopic MF classification tasks. In addition, it achieves deployment and implementation on embedded devices to meet industrial application demands.