Challenge
In the semiconductor post-process packaging stage, precision drilling or cutting microscopic holes in wafers with CO2 laser drills often creates a bottleneck. Equipment anomalies lead to significant downtime, impacting productivity. While data is collected to monitor these drills, tracking numerous data points across multiple machines using only basic rule conditions (e.g., max/min values) makes proactive and effective anomaly detection difficult.
Approach
A monitoring system was built to proactively detect anomalies in CO2 laser drills using a deep learning-based anomaly detection model that learns the normal data distribution. This model analyzes the difference between predicted and actual values of individual data points in complex patterns, providing a single indicator to determine overall abnormality.
Value Delivered
The deep learning-based anomaly detection model, capable of analyzing complex patterns, detects anomalies in 10 CO2 laser drills a month in advance with a detection rate of about 93%. A model-based interpretation algorithm (AIX, Explainable AI) offers summary insights into the key sensors where anomalies occurred. The model is continuously retrained with a training adequacy determination algorithm and is designed to be robust, drawing inferences even if some sensors fail, ensuring stable operations through machine learning operations (MLOps).
