The flash welding process is a mainstream to manufacture anchor chains in the shipbuilding industry. However, traditional quality control approaches only use the tension test of the whole chain. This cannot adequately guarantee the quality of each anchor chain. Also, the incurred cost of post-build repair is far more than production. There is an urgent need to design and develop new quality control methods for real-time monitoring and control of flash welding processes. In this investigation, we firstly develop a data acquisition and control system to collect sensor data pertinent to process dynamics (i.e., electrode-position and electric-current profiles) and monitor the flash welding process in real time. Then a novel spatiotemporal warping approach is proposed to quantify the dissimilarity of electric-current and electrode-position signals collected in the flash welding process. Further, we embed the resulted warping matrix into feature vectors in the low-dimensional space that preserves the dissimilarity distances. Finally, Dirichlet Process (DP) models are proposed to cluster embedded features (coordinates in low-dimensional space). Experimental results demonstrated that the proposed methodology not only effectively reveals the directional differences among flash welding profiles, but also significantly outperforms traditional clustering algorithms such as the K-means approach, i.e., 13.25%, 1.67% and 12.33% increases in the prediction performance with the use of electric-current, electrode-position and combination recordings, respectively.
|Number of pages
|Published - 2019
|47th SME North American Manufacturing Research Conference, NAMRC 2019 - Erie, United States
Duration: Jun 10 2019 → Jun 14 2019
All Science Journal Classification (ASJC) codes
- Industrial and Manufacturing Engineering
- Artificial Intelligence