Anomaly detection in batch chemical processes

Abstract Fault detection and diagnosis systems (FDDS) are efficient ways for preventing accidents and helpful for supporting plant operators' decision-making. In this direction, a new anomaly detection system (AD) has been posed through a fault detection and diagnosis system. The objective is to be able to detect faults that have never happened before by using training information obtained from processes history (known faults). The methodology consists of two steps: a first fault detection stage (binary classification problem) and a subsequent diagnosis stage (multi-class problem) addressed under a multi-label (ML) approach. The FDDS has been implemented using Support Vector Machines (SVM). The problem addressed is the monitoring and diagnosis of transient operation modes, for which the FDDS has been tested in a pilot plant heat exchange system operating batch wise. Results are discussed and promising measures of diagnosis performance (F1 index) are finally reported.