A Fast Defect Detection Method for Band Saw Machines with Deep Learning and Image Processing Technique

• Main Authors: Yu-Yi Chiang, 江宥儀
• Other Authors: 吳俊霖
• Format: Others
• Language: zh-TW
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/cgi-bin/gs32/gsweb.cgi/login?o=dnclcdr&s=id=%22107NCHU5394036%22.&searchmode=basic

碩士 === 國立中興大學 === 資訊科學與工程學系所 === 107 === 　　In recent years, the technology of factory automation have become the main current in today '' s world. As the low-birth rate, the problem of population decline and the minimum wage increase occur one after another. Many industries have begun to combine artificial intelligence to produce instead of the previous manual detection methods. Whether it is service industry, manufacturing industry, technology industry, etc., automation technology is one of the key points in nowadays. 　　Both deep learning and image processing have their own advantages and disadvantages. For example, the operation speed of the machine is dominated by image processing, and deep learning is dominant in the case of multi-shape change or multi-species detection. Therefore, we combine these two technologies and apply them to the factory''s band saw machine inspection, hoping to effectively reduce the cost of manual inspection and successfully implemented defect warning. 　　In this study, since the band saw machine cuts the material in the high-speed mode, we need to use a camera with a high FPS to capture the images. Then use deep learning to automatically detect the defect, but because the machine will spray the cutting fluid to cool down and spray off the defect residue,so it is possible to detect some residual water droplets. We use image processing method to filter the false detection to improve accuracy. Calculate all the number of defects detected in each of the first ten screens and transfer them to the PC-Based controller to evaluate the quantity and give different levels of warning. The experimental results show that the accuracy of the proposed system achieves higher than 91%. It demonstrates the effectivess and efficient of the proposed method.