The goal of time series anomaly detection in Cyber-Physical Systems (CPS) is to maintain the reliability and security of environmental sensors. Due to the lack of labeled data in real-world CPS scenarios, most existing anomaly detection methods utilize unsupervised learning to capture the normal behavior of unlabeled time series. However, constructing reliable anomaly detection models for CPS is challenging, as they must operate accurately in the context of complex system dynamics and uncertain sensor noise levels. To address this issue, we propose a robust attention-based contrastive representation learning method designed to capture discriminative and robust representations, specifically tailored to solve challenges posed by noisy and dynamic sensor data in CPS. In particular, we introduce the graph convolutional network (GCN) to learn embeddings for all variables, effectively capturing the time-varying correlations between variables in MTS and enhancing the representations’ robustness. Furthermore, we leverage the feature graphs of variables constructed by GCN to perform multi-scale attention-based contrastive representation learning. Finally, through minimizing the discrepancies between original and GCN-enhanced samples across multiple attention scales, our model learns a robust and discriminative representation.