Toward generalizable robot vision guidance in real-world operational manufacturing factories: A Semi-Supervised Knowledge Distillation approach

Zhao，Zizhou1; Lyu，Junyu2; Chu，Yinghao3; Liu，Ke1,2; Cao，Daofan1; Wu，Changning2; Qin，Longjun4; Qin，Shiwei5

2024-04-01

10.1016/j.rcim.2023.102639

Robotics and Computer-Integrated Manufacturing   影响因子和分区

0736-5845

The complexity and diversity of scenarios, along with the presence of environmental noise in factory settings, pose significant challenges to the implementation of deep learning-based vision-guided robots for smart manufacturing. In response to these challenges, we introduce a novel Semi-Supervised Knowledge Distillation (SSKD) framework that has been extensively validated and deployed across numerous real-world production lines. The proposed SSKD framework combines the advantages of semi-supervised learning and knowledge distillation to offer optimization for the majority of deep learning models. Experiments conducted in real-world factory settings demonstrate that the SSKD framework significantly enhances the performance of deep learning models, reducing inference time from 185 ms to 45 ms and improving generalizability across different working environments, achieving recall and precision values that exceed 99.5% and 92.6%, respectively, achieved a remarkable 200% improvement in labor efficiency. Our innovative SSKD framework provides a reliable and scalable solution for enhancing manufacturing productivity and product quality. The success of this approach in transforming vision-guided robotic systems for smart manufacturing highlights its potential for broader industry adoption. The SSKD framework offers a reliable and scalable solution for enhancing manufacturing productivity and product quality. Our results underscore the potential of this innovative approach to transform vision-guided robot systems in smart manufacturing, making it an attractive candidate for widespread adoption in the industry. We are proud to report that, as of the end of 2022, the SSKD framework has been successfully implemented in 50 robots – a more than ten-fold increase from the initial 4 in 2020 – resulting in an annual yarn production capacity exceeding 100,000 kg. This accomplishment underscores the practical impact and effectiveness of the SSKD framework in real-world production lines.

Domain adaptation Knowledge distillation Semi-supervised learning Smart manufacturing Vision-based robot guidance

SCI

英语

第一

ENGINEERING

2-s2.0-85172670679

Scopus

1.Department of Chemistry,Southern University of Science and Technology,Shenzhen,China
2.School of Innovation and Entrepreneurship,Southern University of Science and Technology,Shenzhen,China
3.Department of Systems Engineering,City University of Hong Kong,Hong Kong
4.School of Environment and Energy,Peking University Shenzhen Graduate School,Shenzhen,518055,China
5.AIATOR,Shenzhen,China

Zhao，Zizhou,Lyu，Junyu,Chu，Yinghao,et al. Toward generalizable robot vision guidance in real-world operational manufacturing factories: A Semi-Supervised Knowledge Distillation approach[J]. Robotics and Computer-Integrated Manufacturing,2024,86.

Zhao，Zizhou.,Lyu，Junyu.,Chu，Yinghao.,Liu，Ke.,Cao，Daofan.,...&Qin，Shiwei.(2024).Toward generalizable robot vision guidance in real-world operational manufacturing factories: A Semi-Supervised Knowledge Distillation approach.Robotics and Computer-Integrated Manufacturing,86.

Zhao，Zizhou,et al."Toward generalizable robot vision guidance in real-world operational manufacturing factories: A Semi-Supervised Knowledge Distillation approach".Robotics and Computer-Integrated Manufacturing 86(2024).