Convolutional AutoEncoders for Anomaly Detection in Semiconductor Manufacturing

Semiconductor manufacturing, characterised by its complex processes, demands efficient anomaly detection (AD) systems for quality assurance. This study extends from previous work utilising unsupervised Convolutional AutoEncoders for AD in Semiconductor batch manufacturing by applying the technique to a novel dataset supplied by a local Semiconductor Manufacturer. Our method uses an approach that employs 1-dimensional Convolutional Autoencoders (1d-CAE) to improve AD performance and interpretability through the numerical decomposition of reconstruction errors. Identifying anomalies this way allows engineering resources to explain anomalies more effectively than traditional methods. We validate our approach with experiments, demonstrating its performance in accurately detecting anomalies while providing insights into the nature of these irregularities. The experiments also demonstrate the impact of training setup on detection capability, outlining an efficient framework for determining an optimal hyperparameter set-up in an industrial dataset. The proposed unsupervised learning approach with AE reconstruction error improves model explainability, which is expected to be beneficial for deployment in semiconductor manufacturing, where interpretable and trustworthy results are critical for solution adoption by process engineering teams.