Wistron has adopted AI technology to inspect defects of products at three of its factories and will extend such inspection to other plants, according to company senior software manger Liang Wei-quo.
The AI-based defect inspection has been adopted for SMT (surface-mount technology) and DIP (dual in-line package) lines for producing notebooks and servers mainly, Liang said.
Automatic optical inspection (AOI) may have high efficiency, but many manufacturers set AOI parameters at high levels in order to achieve high yield rates, resulting in over-sensitivity of inspection equipment, Liang said, adding manual work has to follow to double check the products for misjudged ones possibly made by AOI, Liang indicated
In a bid to attain accurate inspection without using human eyes, Wistron has used more than 150,000 images in AI-based deep learning for 15 models of defect recognition for SMT production lines, with accurate recognition taking only 0.01 second, Liang noted. For DIP production lines, more than 130,000 images have been used in AI deep learning for five models of defect recognition, Liang indicated.
The first step in using AI-based inspection is to define what defects need to be inspected and the second step is data marking, Liang said. Wistron has collected more than one million images but selected from them 150,000 that represent various defect attributes for data marking, Liang noted.
After deep learning, AI-based models of recognition came into operation and, despite high levels of accuracy in recognition, the overall performance was short of expectation, Liang said. Thus, Wistron checked the dataset of images to choose images that represented defect attributes but were not chosen originally for additional data marking to improve the overall performance, Liang indicated.
