Machine learning based screw drive state detection for unfastening screw connections

Abstract

The electrification of the transport sector, limited primary materials, and the resulting need for a circular economy drives the automated disassembly of End of Life battery systems. Environmental influences and improper maintenance may cause stripped screw drives. Industrial screw drivers and nut runners evaluate unfastening operations in binary form as OK and NOK (Not OK). However, deeper evaluations such as the detection of stripped screw drives are necessary to decide between destructive and non-destructive disassembly and the resulting tool selection. This paper presents a novel supervised learning based approach for detecting stripped screw drives during the positive locking phase of unfastening operations by using intrinsic data (torque of the screw driver), where the bit or nut are inserted inside the screw drive. The basis of this work is an experimental data set with more than 1,000 unfastening operations from a manual and robot-based test bench, which includes four screw drive types: Torx, Phillips, external hexagon, and internal hexagon. A comparison between multilayer perceptrons (MLPs) and convolutional neural networks (CNNs) have been examined using the experimental data set. Additionally, the reliability and accuracy of the proposed approach have been examined. The findings from this paper demonstrate the feasibility of CNN-based detection for a revolution number lower than one, which is one revolution better than the state of the art techniques, a screw size and screw drive independent detection of stripped screw drives. The CNN model gives an accuracy of 0.98 to predict the screw drive state.