Anomaly Classification and Key Segment Detection in Semiconductor Manufacturing Processes Using Image-Based Siamese Networks

Semiconductor manufacturing processes are a fundamental component of modern industry and technology. Anomalies in this process degrade product quality and reliability, necessitating swift responses. To address this, various studies have been conducted on anomaly detection and efficient management using artificial intelligence algorithms. However, there are limitations that fail to consider the characteristics of actual process data, such as irregular patterns, numerous missing values, and varying data lengths. In this study, we propose to overcome these limitations by using image data with scatter plots applied to actual process data, employing a siamese network-based anomaly classification model based on the similarity with normal segments. Additionally, we apply gradient-weighted class activation map (Grad-CAM) to the siamese network model for identifying the main causes of abnormal segments. The validity and applicability of the proposed method have been demonstrated using data obtained from actual semiconductor manufacturing processes. The proposed method shows superiority in all comparative models. Furthermore, it has been confirmed that the main cause of the abnormal segment aligns with the abnormal judgment criteria for a specific segment of the field engineer.