AI-Driven Automation for Safety Monitoring in the Chemical Industry: A Review of Hazard Detection, Anomaly Detection, and Predictive Risk Assessment

Chemical and petrochemical plants work with hazardous materials and complex equipment, so small changes in operating conditions can sometimes signal a serious problem. These early signs may appear across several sensors and may not immediately trigger a conventional alarm. For this reason, researchers have explored artificial intelligence (AI) for gas-leak detection, visual hazard recognition, process anomaly detection, fault diagnosis, predictive maintenance, and risk assessment. This review discusses research across these areas, including industrial sensing, deep learning, explainable models, anomaly detection, foundation models, digital twins, and AI-assisted hazard analysis. The reported results are promising within the settings in which they were tested. For example, one pipeline-leak study reports 96.88% classification accuracy, a Tennessee Eastman Process study reports 92% prediction efficiency on an external test set, and a run-level diagnosis study reports macro-F1 values near 0.95 [12], [15], [16]. These figures come from different tasks and test designs, so they should not be treated as directly comparable. Across the studies, recurring concerns include dependence on simulated benchmarks, limited testing of difficult faults, scarce labelled incident data, changes in operating conditions, false alarms, and a lack of validation across multiple sites. The review discusses a layered architecture that connects plant sensors and control infrastructure with validated AI models, evidence-based alerts, and operator decisions. AI may strengthen monitoring, but it should not replace independent protective systems or the judgment of qualified process-safety personnel. Field validation, uncertainty reporting, cybersecurity, traceability, and ongoing oversight remain essential.