Indoor movement data analytics based on machine learning for manufacturing assembly line

Abstract

The paper presents an analysis on how machine learning models can be used to extract information from indoor movement data in an assembly line. An Indoor Positioning System (IPS) is commonly used to track workers, moving objects and vehicles at indoor venues where GPS is not effective. Movements happen within an assembly line as certain sets of work activities are performed (e.g., delivering assembled units from one work bench to another, reworking at a work bench and transporting a trolley from one place to another) and anything beyond that can be considered as anomalous.A computer assembly line was simulated and the movement trajectories of five workers were generated to perform the following activities: working at their own station, moving between different workstations, helping at another work station, working at storage and moving to a storage zone. We then developed a set of machine learning models to investigate how accurately these behaviours can be detected.The machine learning pipeline for indoor tracking data involves:(a) segmenting worker trajectories using Self-Supervised Learning (SSL). SSL is a machine learning process where a model is trained to learn one part of the input data from another part without the use of manually acquired labels. SSL is used to find the change points of the trajectories (i.e., used as segment boundaries) by training a model to predict the future interval of a trajectory from its temporally adjacent past window.(b) classifying trajectory segments (i.e., sub-trajectories) into different categories of worker behaviour. We investigated different supervised machine learning models to learn from the labelled segments.(c) estimating the state of the assembly floor based on the inferred worker behaviour. The state provides a wholistic view of the factory floor including the collective behaviour of workers. We used a clustering approach to identify the dominant states (joint activity) of the factory floor.We received some promising results from our analysis. We present the methods and results in detail in the paper.