ESTABLISHING THE FOUNDATION TO ROBOTIZE COMPLEX WELDING PROCESSES THROUGH LEARNING FROM HUMAN WELDERS BASED ON DEEP LEARNING TECHNIQUES

As the demand for customized, efficient, and high-quality production increases, traditional manufacturing processes are transforming into smart manufacturing with the aid of advancements in information technology, such as cyber-physical systems (CPS), the Internet of Things (IoT), big data, and artificial intelligence (AI). The key requirement for integration with these advanced information technologies is to digitize manufacturing processes to enable analysis, control, and interaction with other digitized components. The integration of deep learning algorithm and massive industrial data will be critical components in realizing this process, leading to enhanced manufacturing in the Future of Work at the Human-Technology Frontier (FW-HTF). This work takes welding manufacturing as the case study to accelerate its transition to intelligent welding by robotize a complex welding process. By integrate process sensing, data visualization, deep learning-based modeling and optimization, a complex welding system is established, with the systematic solution to generalize domain-specific knowledge from experienced human welder. Such system can automatically perform complex welding processes that can only be handled by human in the past. To enhance the system\u27s tracking capabilities, we trained an image segmentation network to offer precise position information. We incorporated a recurrent neural network structure to analyze dynamic variations during welding. Addressing the challenge of human heterogeneity in data collection, we conducted experiments illustrating that even inaccurate datasets can effectively train deep learning models with zero mean error. Fine-tuning the model with a small portion of accurate data further elevates its performance