Transformer-Based Anomaly Detection for Multivariate Process Control Data (TBAD)

Abstract

Industrial Internet of Things (IIoT) environments generate large volumes of multivariate sensor data through distributed sensing infrastructures used for monitoring, control, and optimization of physical processes. Reliable anomaly detection within these high-dimensional sensor streams is critical for maintaining operational safety and system performance. However, traditional anomaly detection methods often struggle to capture temporal dependencies and inter-sensor relationships inherent in multivariate IIoT time-series data. This paper evaluates conventional and deep learning-based anomaly detection approaches—including Isolation Forest, Local Outlier Factor, One-Class Support Vector Machines, Long Short-Term Memory networks, Transformers, and Generative Adversarial Networks—on real-world industrial sensor data. We introduce a Transformer-Based Anomaly Detection (TBAD) model that learns temporal and cross-sensor dependencies directly from sensor streams, with controlled anomalies introduced for quantitative evaluation. Experimental evaluation on a real-world industrial dataset representative of IIoT data pipelines shows that TBAD achieves a Precision–Recall Area Under the Curve (PR-AUC) of 0.9737 and Receiver Operating Characteristic Area Under the Curve (ROC-AUC) of 0.9986, outperforming classical and recurrent baselines. These results demonstrate the effectiveness of attention-based architectures for scalable anomaly detection in IIoT sensor networks.