Sensor-based Human–Process Interaction in Discrete Manufacturing

Abstract

The rise of Industry 4.0 and the convergence with business process management provide new potential for the automatic gathering of process-related sensor information. In manufacturing, information about human behavior in manual assembly tasks is rare when no interaction with machines is involved. We suggest technologies to automatically detect material picking and placement in the assembly workflow to gather accurate data about human behavior and flexible support of human–process interaction. The detection of material picking is achieved by using background subtraction in combination with scales. For placement detection, two approaches are tested: image classification using convolutional neural networks and object detection using Haar wavelets. The detected fine-grained worker activities are then correlated with a hybrid model of the assembly workflow using the business process model and notation and case management model and notation, enabling the measurement of production time (time per state) and quality (frequency of error) on the shop floor as an entry point for conformance checking and process optimization. The approach has been evaluated in a quantitative case study recording the assembly process 30 times in a laboratory setup within 4 h. Under these conditions, the classification of assembly states using a neural network provides a test accuracy of 99.25% on 38 possible assembly states. Material picking based on background subtraction has been evaluated in an informal user study with six participants performing 16 picks each, providing an accuracy of 99.48%. The suggested method offers a promising approach to easily assess fine-grained timings and error rates of assembly steps which can be used to optimize the corresponding process.