Towards Segmentation and Labelling of Motion Data in Manufacturing Scenarios

Abstract

There is a significant interest to evaluate the occupational exposure that manufacturing operators are subjected throughout the working day. The objective evaluation of occupational exposure with direct measurements and the need for automatic annotation of relevant events arose. The current work proposes the use of a self similarity matrix (SSM) as a tool to flag events that may be of importance to be analyzed by ergonomic teams. This way, data directly retrieved from the work environment will be summarized and segmented into sub-sequences of interest over a multi-timescale approach. The process occurs under 3 timescale levels: Active working periods, working cycles, and in-cycle activities. The novelty function was used to segment non-active and active working periods with an F1-score of 95%. while the similarity function was used to correctly segment 98% of working cycle with a duration error of 6.12%. In addition, this method was extended into examples of multi time scale segmentation with the intent of providing a summary of a time series as well as support in data labeling tasks, by means of a query-by-example process to detect all subsequences.