Transformer Autoencoder-Based Anomaly Detection and Intelligent Process Monitoring for Smart Water Treatment Plants

Smart water treatment plants increasingly rely on interconnected sensors and automated monitoring systems to maintain water quality, process stability, operational efficiency, and safety. Continuous generation of multivariate sensor and actuator data provides opportunities for artificial intelligence-based process monitoring and early anomaly detection. Deep learning can identify complex temporal relationships and support proactive detection of abnormal operating conditions. Conventional monitoring systems primarily depend on fixed thresholds, predefined rules, and statistical methods, which may fail to identify complex anomalies caused by sensor faults, equipment malfunctions, process disturbances, and abnormal operating events. These limitations can lead to delayed detection, false alarms, operational disruptions, and reduced process reliability. Therefore, an intelligent framework is required to learn normal process behavior and identify significant deviations in real time. This study proposes a Transformer Autoencoder-based anomaly detection framework for intelligent process monitoring in smart water treatment plants. The framework is evaluated using the Secure Water Treatment (SWaT) dataset, containing multivariate sensor and actuator measurements from a realistic water treatment testbed under normal and abnormal operating conditions. The data are normalized and transformed into temporal sequences before being processed by the Transformer Autoencoder. Its self-attention mechanism captures temporal dependencies and relationships among process variables while learning normal operating patterns. During inference, reconstruction errors are calculated to generate anomaly scores. The detection threshold is determined from normal validation data using the mean reconstruction error plus three standard deviations (μ + 3σ). Detected anomalies generate early warnings for proactive intervention. The trained model supports edge-based deployment for local real-time inference, while cloud infrastructure can optionally support training, storage, and model updates. Evaluation on the SWaT dataset achieved 98.4% accuracy, 97.8% precision, 98.1% recall, and 97.9% F1-score. The framework reduced false alarm rates by 27% compared with conventional monitoring approaches and detected abnormal conditions 8–12 minutes before critical threshold violations. The proposed framework provides an effective solution for real-time anomaly detection and intelligent process monitoring. By learning multivariate process behavior and enabling edge-based inference, it improves operational reliability, process safety, and timely decision-making. Future work will integrate digital twins and adaptive process control for autonomous water treatment operations.