어텐션 기반 오토인코더를 이용한 제조 공정 데이터 이상 탐지

Abstract

This study proposes an Attention-based Autoencoder for anomaly detection by learning complex interactions among variables in manufacturing process data. Conventional machine learning and autoencoder-based approaches often fail to fully capture nonlinear relationships and contextual dependencies among variables, resulting in limited reconstruction accuracy. The proposed model vectorizes each continuous variable using column embedding, learns contextual interactions through a Transformer encoder, and generates a global latent representation via trainable weighted pooling, which is then reconstructed using an MLP autoencoder. Experimental results demonstrate that the proposed model achieves higher accuracy (93.35%) and AUROC (97.79%) compared to simple MLP autoencoders and existing Transformer-based models, highlighting the effectiveness of attention mechanisms in enhancing anomaly detection performance for manufacturing process data.