Study on learning efficient stroke representations in clocked sheet metal processing: theoretical and practical evaluation

Clocked manufacturing processes such as sheet metal forming and cutting processes pose a challenge for process monitoring approaches due to inaccessibility of tool components and high production rates which make direct measurement of the physical process conditions unfeasible. Auxiliary data such as force signals are acquired and assessed, often still relying on control and run charts or even visual control in order to monitor the process. The data of these signals are high-dimensional and contain a large amount of redundant information. Therefore, the processing of such signals focuses on compressing information into as few variables as possible that still represent the important information for the manufacturing process. Due to repeatability in clocked sheet metal processing, the data generated consist of a series of time series of the same operation with varying physical conditions due to wear and variations in lubrication or material properties. In this paper two major research objectives are identified: (i) the theoretical evaluation of representation learning methods in context of clocked sheet metal processing, and the connection with (ii) the practical evaluation of the learned representations with a given use case to track the wear progression in series of strokes. The contribution of this paper is the comparison of varying time series representation learning techniques and their performance evaluation in a theoretical and practical scenario.