Toward a robust approach to multivariate time series anomaly detection

Background: Anomaly detection in semiconductor manufacturing is vital for maintaining yield and reducing costs, particularly in high-volume production where inspections are time-consuming and expensive. Aim: We aim to develop a robust, unsupervised deep learning framework for multivariate anomaly detection, addressing limitations in current fault detection and classification systems. Approach: We propose a transformer-based model integrating aggregated z-normalization to mitigate distribution drift and peaks-over-threshold for adaptive thresholding, ensuring reliable performance across varying datasets. Results: The framework achieved an F1 score of 0.9827 on semiconductor datasets, with high precision (0.9866) and minimal false alarms, validated through extensive ablation studies. Conclusions: The proposed solution is scalable and adaptable for industrial environments, with future work focused on improving the detection of single-spike anomalies and borderline cases to enhance yield and operational reliability.