In this information age, semiconductors are in use everywhere. China has now become one of the largest manufacturers and consumers of semiconductors in the world. Nexperia Hong Kong Limited, as the Asian headquarter of one of the largest semiconductor manufacturing companies in the world has its main manufacturing bases located in China and produces billions of semiconductor devices to meet global demand.
One challenge is, the occurrence of abnormal devices is rare, but it does add up considering the enormous number of devices produced. Hundreds or thousands of human operators check the large amount of devices produced by checking the device images for possible defects. In the assembly line of semiconductors, the proportion between passed and defected images are extremely skewed. The key is to detect as many defected images as possible, while not sacrificing too many passed ones. Moreover, even well-trained human operators might make the occasional mistake as they get tired or bored. So it is important and necessary to develop accurate and robust methods to identify defective semiconductor devices exploiting advanced techniques in machine learning and artificial intelligence.
In this project, in order to detect abnormal semiconductor devices with a large collection of human-labelled data, a computer aided anomaly detection system based on state-of-the-art data analytics and machine learning techniques is designed. For this purpose, some "Nexperia ImageNet" datasets were collected, consisting of human labeled images of semiconductor devices, even with defect location information. Then various models that have been successfully used in large scale image classifications are applied to such datasets. For example, the popular ResNet and VGG pretrained on ImageNet are fine-tuned in datasets, enhanced with attention models for defect localization. The objective is to control the false discovery (positive) rate while keeping the true positive rate as close to 1 as possible. In addition, through such a computer aided anomaly detection system, we are able to discover those wrong labels provided by original human operators and improve the datasets. Systems with the realistic application scenarios are tested. Finally, the goal of this project is to improve the quality and efficiency of the current anomaly detection system in industry.
Ongoing project information website: https://github.com/huangkaiyikatherine/nexperia
A Kaggle competition website: https://www.kaggle.com/c/semi-conductor-image-classification-first
- Professor, Department of Mathematics
- Associate Professor, Department of Chemical and Biological Engineering
HKUST celebrated the opening of The Big Data for Bio Intelligence Laboratory which is dedicated to designing data analytic solutions for big data in biology...