Robust unsupervised outlier detection in IoT using contrastive learning-driven autoencoders

Outlier detection plays a critical role in ensuring the reliability, security, and operational stability of Internet of Things (IoT) systems, where data streams are often high-dimensional, heterogeneous, and unlabeled. In this paper, we propose a robust unsupervised outlier detection framework that integrates autoencoder-based representation learning with contrastive loss to enhance discriminative capability in latent space. The proposed model jointly optimizes reconstruction loss and contrastive loss, enabling the autoencoder not only to accurately reconstruct normal patterns but also to enforce clear separation between normal and anomalous samples in the learned feature space. To further improve detection robustness, we introduce an adaptive threshold determination algorithm that identifies outliers based on statistical analysis and percentile-based modeling of reconstruction errors, eliminating the need for manually tuned thresholds. This dual-thresholding strategy allows the proposed approach to adapt effectively to varying data distributions and operational conditions. Extensive experiments conducted on the Statlog (Landsat Satellite) and UNSW-NB15 datasets demonstrate that the proposed method consistently outperforms traditional unsupervised and deep learning-based baselines in terms of precision, recall, and F1-score. The results confirm that incorporating contrastive learning significantly improves anomaly separability and detection accuracy, making the proposed framework a reliable and scalable solution for outlier detection in real-world IoT environments.