Memory-Efficient Artificial Intelligence Framework for Real-Time Multivariate Anomaly Detection

The semiconductor industry is rapidly expanding, driven by growing demands in artificial intelligence (AI), cloud computing, the internet of things, and the electrification of automotive and infrastructure. As manufacturing processes grow more complex and dynamic, accurate real-time anomaly detection becomes critical for maintaining yield and reliability. However, conventional statistical methods lack adaptability in dynamic process conditions, while deep learning models are typically too resource-intensive for edge deployment. Herein, we present a lightweight yet robust AI-based anomaly detection framework for real-time, sensor-specific semiconductor manufacturing. Our framework incorporates a four-layer convolutional neural network trained on unsupervised dual-segment inputs with hierarchical layer-wise objectives. Using only 3,783 parameters for 49 input features, the framework achieves 99.90% accuracy and an F1 score of 82.14%, reducing false alarms by up to 10.3 times and 4.0 times compared to the convolutional autoencoder control and industrial statistical model baseline, respectively. We further introduce a computationally efficient method for generating labeled virtual data to enable sensor-wise adaptive thresholding under imbalanced data conditions. Together, these innovations enable accurate and robust real-time fault detection in environments with limited computing resources, and provide a scalable foundation for diverse edge AI applications.