P‐96: Importance of Data Structure Change by Applying Domain Knowledge in Industrial Tabular Data

Abstract

Recently, AI has penetrated various fields in industrial settings. For example, the display industry is also actively researching AI applications, and the most advanced field is using inspection data. A measurement sensor from a processing facility or data from an inspection facility is used to detect defects at the display manufacturing site. However, in the case of raw material defects, the effects of minor defects can accumulate and cause abnormalities in subsequent processes, which are very difficult to detect. In this case, the accumulated data is analyzed, but there is no exact standard, and there is a limit to subjective judgment by the inspector. To solve this issue, we applied DBSCAN clustering, clustering the X and Y axes individually, coordinates with binning, and normalization as the preprocessing method to maximize the discriminative power of the defect. Moreover, for the final defect detection, an AI structure that detects defects was completed using a structure that combines DBSCAN and Isolation Forest. We minimized detection errors by using the clustering function of DBSCAN as a noise removal effect. As a result, it was possible to develop and mass‐produce ML with an F1 score of 0.998 for defect detection, which was difficult to automate in the past. AI performed this comparison, objectifying the existing engineer's resource‐based work and shortening the progress time.