Temporal Contrastive Representation Learning for Unsupervised Anomaly Detection in High-Dimensional Cloud Environments

Abstract

This paper proposes an unsupervised anomaly detection method based on contrastive learning to address challenges in cloud computing environments, such as high data dimensionality, complex structure, and lack of labels. The method segments raw time series monitoring data into subsequences using a sliding window mechanism and applies various data augmentation strategies to construct positive and negative sample pairs, guiding the model to learn discriminative embeddings without supervision. A temporal attention mechanism is integrated to capture key dynamic features in the sequence, enhancing the model's ability to represent long-term dependencies and local fluctuations. Anomaly scores are calculated by measuring similarities in the embedding space, enabling an efficient detection process without the need for labels. The method is evaluated on a cloud monitoring dataset across different augmentation strategies, parameter settings, and temporal modeling configurations. Experimental results show that it outperforms several recently published unsupervised models in F1 Score, AUC, and KS Score, demonstrating its effectiveness and engineering adaptability in handling high-dimensional dynamic data within cloud platform scenarios.