Toward a robust approach to multivariate time series anomaly detection

Anomaly detection is essential in the semiconductor industry to ensure yield and minimize costs. It becomes particularly critical in high-volume manufacturing (HVM), where measurement processes are inherently time-consuming and expensive. Proactive anomaly detection prevents defect propagation, reducing yield losses and avoiding cost increases from delayed detection. While Fault Detection and Classification (FDC) systems are widely employed, they often fail to capture complex relationships within multivariate datasets and subtle shifts within specifications. This paper presents a robust, unsupervised deep learning framework for multivariate anomaly detection in dynamic industrial environments. By combining Aggregated z-normalization to address distribution drift and Peaks-Over-Threshold (POT) for adaptive thresholding, the proposed method effectively generalizes across diverse datasets. The model, based on a Transformer architecture, achieved an F1 score of 0.9879 in detecting anomalies across semiconductor manufacturing datasets, demonstrating high precision (0.9995) with minimal false alarms. Ablation studies further validate the importance of normalization and thresholding in improving robustness. This framework provides a scalable and efficient solution, aligning with the stringent requirements of the semiconductor industry. Future work will focus on enhancing the detection of single-spike anomalies and borderline cases to further improve yield and ensure reliable deployment in real-world industrial settings.