Research Proposes Unsupervised Outlier Detection Method for IoT Using Autoencoders

Outlier detection is used to maintain reliability, security and operational stability in Internet of Things systems that generate high-dimensional, heterogeneous and unlabeled data streams. A research paper published today outlines an unsupervised framework that integrates autoencoder-based representation learning with contrastive loss.

The model jointly optimizes reconstruction loss and contrastive loss so the autoencoder reconstructs normal patterns while creating clearer separation between normal and anomalous samples in the latent feature space. The method also incorporates an adaptive threshold determination algorithm.

This algorithm uses statistical analysis and percentile-based modeling of reconstruction errors, removing the requirement for manually set thresholds. The dual-thresholding approach is designed to adjust to different data distributions and operating conditions.

The paper reported that the framework was tested on the Statlog (Landsat Satellite) and UNSW-NB15 datasets. According to the results, the proposed method recorded higher precision, recall and F1 scores than traditional unsupervised techniques and other deep learning baselines.

The inclusion of contrastive learning improved anomaly separability in the tested data. The authors stated that the overall framework offers a reliable and scalable option for outlier detection in real-world IoT environments. The article was received on February 21, 2026, accepted on April 27, 2026 and published on May 8, 2026.