Patch-Based Time-Series Anomaly Detection with Cross-Variable Attention

Modern manufacturing processes are influenced by smart factories, which collect data from multiple sensors in real time. However, detecting anomalies is challenging due to their irregular and complex patterns, often resulting in unreliable labeling that depends on the engineer's expertise. Moreover, in these processes, where sensors are interconnected, fluctuations in one variable can potentially influence subsequent variables. The proposed method uses cross-variable attention to reflect the relationship between variables at different time points. It also utilizes contrastive learning to extract a representation of only normal data that can be distinguished from anomalies without labels. Experimental results on multivariate time series data demonstrate that the proposed method outperforms existing models in anomaly detection.