Ralph Stephan, Stäubli Sargans
Dr. Jianyong Wen, Stäubli Sargans
Real-time industrial fabric inspection systems face challenges of a great number of pattern variations, fast and easy training processes, strongly imbalanced datasets, and even more, lack of samples from certain classes at the beginning of inspection. To solve these problems, Stäubli Sargans AG has developed an incremental model that combines machine learning and deep learning techniques.
Their system consists of two interoperating models: a “base” that models general fabric characteristics, and a second (“update”) model that is iteratively retrained during the deployed inspection process. To improve robustness and accuracy, the update-model is iteratively trained by using more samples during subsequent inspection processes.
Key advantages of this iterative approach include feasibility, applicability, resource efficiency, fast training with fewer samples, and incremental improvement required by industrial products. To enable fast implementation and verification of the algorithm on system-on-chip platforms, Stäubli has chosen Model-Based Design from model creation through to hardware-software-co-simulation, ensuring continuous and rapid improvement, conforming to changing requirements.