http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
최인호(In-Ho Choi),정동휘(Dong-Hwi Jung),김성우(Sung-Woo Kim) 대한전자공학회 2021 대한전자공학회 학술대회 Vol.2021 No.6
Existing semiconductor defect inspection has relied on measurement between process steps or destruction inspection, in which only some wafers can be inspected according to production efficiency issues. However, due to an issue that leads to a large loss of some wafers that were not inspected, it was possible to monitor the value of a facility sensor capable of inspecting the entire wafer. In recent quality control, this sensor data is also added along with the existing inspection and is managed through statistical verification. In addition, verification methods such as SPC, EWMA, APC, FDC, and VM are used to verify that specifications are satisfied with real-time time series values without considering the order of time series sensor values. As defects increase as the process refines, the sensor value trend fluctuates according to the time series even within the set specification of the sensor value. At this time, there is a need for a method capable of detecting an abnormal pattern by calculating the time-series correlation of the pattern for each wafer. Pre-treatment was necessary for this method, but until now, only manual verification, which is verified after manual monitoring by humans, has been in progress. Therefore, in this study, in order to improve the shortcomings of these existing verification methods, the problem of data transmission timing was solved by using the similarity measurement method and the deep learning method by using the time series pattern data collected in a specific process. Several examples of real-time time-series abnormal pattern detection verified in the process were included. In addition, through the process data with little bad label data, we compared the performance of various methods and proposed an optimal method that can be applied.