Unsupervised Trace-SPC Anomaly Detection Solution Based on VAE Model

Wafer manufacturing is a complex, expensive, and time-consuming process that involves multiple steps. By closely monitoring the corresponding variations in process parameters, potential anomalies and faults in the process can be timely detected and identified, effectively improving equipment utilization and product yield. However, the anomaly detection methods currently in use rely on extracting feature values from sensor data and analyzing the statistical information of these features. This approach fails to identify complex anomaly patterns, and constructing meaningful feature values can be challenging. In this paper, we propose an unsupervised anomaly detection solution based on the Variational Autoencoder (VAE) model. We optimized the training strategy of the VAE model to address the issue of posterior collapse. By reconstructing the sensor data using the VAE model, we can effectively convert anomalous patterns into normal ones. The experimental results confirm the effectiveness of the model across various sensor datasets, demonstrating the capability of the proposed solution to accurately identify abnormal patterns within the sensor data.