Screw Tightness Detection System Based on Deep Machine Learning

Abstract

Industrial processes make heavy use of screws and nuts to join wood and make connections with metals, which essentially form the foundation for creating complex structures. It is important to have effective screw tightness to manufacture stable industrial components, this makes it critical to have methods of detecting defects in screws. However, in many circumstances, including severe settings, hazardous industrial situations, and the field of rotary equipment, it is difficult for humans to manually detect screw tightness and flaws. Therefore, a smart classifier is required to identify screw defects quickly and accurately while presenting the least amount of risk. In this work we applied deep machine learning convolutional neural networks (CNN) for automatic detection of flaws in a screw where three conditions are detected; tight screw, loose screw, and missing screw, in which serious actions will take place once loose or missing screw is detected. We examined three CNN models which are SqueezeNet, GoogLeNet, and AlexNet. Experimental results show that GoogleNet obtained the highest classification accuracy of 99.8%.