Screw defect detection system based on AI image recognition technology

Abstract

In the past ten years, smart manufacturing has been widely discussed and gradually introduced into various manufacturing fields. Since Germany proposed the concept of “Industry 4.0” in 2011, it has been spreading and feverish all over the world. For Industry 4.0, information digitization, intelligent defect detection and database platform management are their main core technologies. Aiming at a large screw industry manufacturing field in central and southern Taiwan, this paper proposes a screw defect detection system based on AI image recognition technology to detect damage to the nut during the “molding” process in the screw production process, and it is determined whether the inspected screw passes the inspection. The recognition result is given as shown in Figure 1. This paper uses 500 non-defective screw samples and 20 defective screw samples provided by the screw factory. The above samples collected real-time images through the sampling structure designed in this article, and we adopt Microsoft Corporation's ML.NET suite to model AI images, and uses the following four deep learning models: ResNetV2 50, ResNetV2 101, InceptionV3, MoblieNetV2 for learning; in the process of learning, this article divides the data set into three types of data sets (one is the unknown set that is not used for training but mixed with correct and defective samples, and the other is used for post-training verification of mixed samples with correct and defective samples. The third is a training set for training a mixture of correct and defective samples) This arrangement is used for subsequent verification models; after training, a PC-based screw defect detection system is implemented as shown in Figure 2; finally, with Detect screw defects in the form of instant photography. After the experiment, in 1,000 repeated tests, the success rate of defect detection reached 97%, while the false positive rate was only 2%.