Extremely Rare Anomaly Detection Pipeline in Semiconductor Bonding Process With Digital Twin-Driven Data Augmentation Method

With advancements in precise semiconductor manufacturing processes, a new category of anomalies has increasingly emerged. However, due to the probability of an abnormal occurrence during the semiconductor bonding process being less than 1 in 10 million, conventional statistical methods and supervised learning-based neural networks face significant limitations in detecting these anomalies. To address this, several data augmentation techniques have been proposed, yet they fail to ensure the similarity of the augmented time-series data. In response, this study proposes a time-series data augmentation method using digital twins to address the extreme class imbalance problem and presents a pipeline that incorporates this method with an autoencoder-based anomaly detection approach. A robotic arm for the bonding process of nonductile materials was designed to closely mimic the actual process, reflecting the physical properties of the robotic arm, nonductile materials, and particles. The effectiveness of this approach was validated by applying the optimized anomaly score threshold derived from the augmented data to detect anomalies in the actual manufacturing process. This study not only presents an anomaly detection method capable of selecting the most representative patterns from numerous normal samples for comparison with abnormal data but also offers valuable insights into addressing the challenge of detecting extremely rare anomalies.