Refinery Reliability Reinvented: An In-House System for Process Anomaly Detection

Process reliability is the foundation of safe and profitable operations in petroleum refining. However, traditional anomaly detection frameworks—built on static alarm limits and manual surveillance—struggle to capture subtle, multivariate deviations in real-time process data. As refining operations evolve to become more dynamic and complex, there is an increasing need for adaptive, data-driven systems that can autonomously identify anomalies, explain their causes, and help mitigate them before they impact operations. This paper presents the end-to-end design, development, and deployment of an industrial-scale Process Anomaly Detection System (PADS), developed entirely in-house by Hindustan Petroleum Corporation Limited (HPCL). The system's objective is to enhance plant reliability and operational discipline by proactively detecting anomalous process behavior and instrumentation faults using unsupervised learning techniques. Unlike conventional systems that depend on labeled failure data or fixed thresholds, PADS dynamically learns multivariate relationships among process parameters to determine what constitutes normal operation—continuously updating its internal reference states as the plant evolves. At its core, PADS employs a hybrid suite of unsupervised statistical learning techniques including Moving Window Principal Component Analysis (MWPCA), Hierarchical PCA (HPCA), and Slow Feature Analysis (SFA). MWPCA enables adaptive modeling in non-stationary environments by recalibrating reference baselines over sliding windows. HPCA introduces a physical hierarchy into the anomaly detection framework, facilitating isolation of unit-level or inter-unit issues. SFA detects slow-developing degradation trends such as catalyst deactivation, fouling, or instrument drift, often missed by conventional alarms. Once a deviation is detected through Q² and T² metrics, the system uses Reconstruction-Based Contribution (RBC) analysis to pinpoint contributing variables and offer root cause visibility. The system architecture comprises real-time integration with the refinery's Distributed Control System (DCS), a Python-based backend for inference and data processing, and a Django-powered industrial-grade web dashboard for intuitive visualization and operator interaction. This complete solution was built using open-source tools and deployed across more than 20 process units in HPCL's Mumbai and Visakhapatnam refineries. Field application of PADS has demonstrated several operational gains, including early detection of process deviations, valve instability, and sensor drift. Additionally, operators have reported increased trust in digital diagnostics owing to the transparency and interpretability provided by contribution plots and anomaly evolution timelines. PADS represents not just a software solution but a fundamental shift in how refineries approach anomaly detection. It replaces reactive alarm-handling with proactive, explainable insights, enabling a structured reliability framework at scale. Moreover, the in-house development allowed HPCL to tailor every aspect of the system—model tuning, hierarchy definition, frontend workflows—to the refinery's specific operating context. This paper highlights how unsupervised learning, when embedded into the daily fabric of plant operations and developed by domain experts themselves, can significantly elevate operational reliability and advance refinery digitization without dependency on external vendors or costly proprietary tools. By bridging the gap between data science and field operations, PADS sets a precedent for refinery-wide anomaly detection systems that are interpretable, scalable, and resilient to real-world process variability.