Hybrid-Learning-Based Operational Visual Quality Inspection for Edge-Computing-Enabled IoT System
Yinghao Chu,Daquan Feng,Zuozhu Liu,Zizhou Zhao,Zhenzhong Wang,Xiang-Gen Xia,Tony Q. S. Quek
DOI: https://doi.org/10.1109/jiot.2021.3107902
IF: 10.6
2021-01-01
IEEE Internet of Things Journal
Abstract:Deep learning-enhanced Internet of Things (IoT) plays a pivot role in advancing the transformation toward smart manufacturing, and an essential component in many smart manufacturing IoT systems is the quality inspection. However, challenges, such as expensive data labeling, innumerable types of defects, and high costs for iterative optimization, hinder the industrial applicability of previous visual surface quality inspection methods. In this article, we present an edge-computing-enabled IoT system based on an innovative hybrid learning method for visual surface quality inspection using only few labeled data and minimum iterative optimization efforts. Our hybrid learning method first employs a deep neural network to synthesize global representations of real-world industrial images, which are subsequently analyzed via an unsupervised clustering algorithm for anomaly detection. Besides, enhancement strategies, such as fine-tuning and data augmentation, are proposed to improve the robustness against the noisy data set and support low-cost inference in multiple edge devices for manufacturing operation. On a holdout data set collected from real-world factories, our method achieves classification accuracies between 90% and 98%, outperforming the benchmark method by 7%-12%. Moreover, this hybrid learning method demonstrates the effectiveness in detecting new types of surface defects and achieves test recalls between 86% and 97%, outperforming the benchmark method by 11%-34%.