Process Monitoring and Data Mining with Chemical Process Historical Databases

Abstract

Modern chemical plants have distributed control systems (DCS) that handle normal operations and quality control. However, the DCS cannot compensate for fault events such as fouling or equipment failures. When faults occur, human operators must rapidly assess the situation, determine causes, and take corrective action, a challenging task further complicated by the sheer number of sensors. This information overload as well as measurement noise can hide information critical to diagnosing and fixing faults. Process monitoring algorithms can highlight key trends in data and detect faults faster, reducing or even preventing the damage that faults can cause. This research improves tools for process monitoring on different chemical processes. Previously successful monitoring methods based on statistics can fail on non-linear processes and processes with multiple operating states. To address these challenges, we develop a process monitoring technique based on multiple self-organizing maps (MSOM) and apply it in industrial case studies including a simulated plant and a batch reactor. We also use standard SOM to detect a novel event in a separation tower and produce contribution plots which help isolate the causes of the event. Another key challenge to any engineer designing a process monitoring system is that implementing most algorithms requires data organized into “normal” and “faulty”; however, data from faulty operations can be difficult to locate in databases storing months or years of operations. To assist in identifying faulty data, we apply data mining algorithms from computer science and compare how they cluster chemical process data from normal and faulty conditions. We identify several techniques which successfully duplicated normal and faulty labels from expert knowledge and introduce a process data mining software tool to make analysis simpler for practitioners. The research in this dissertation enhances chemical process monitoring tasks. MSOM-based process monitoring improves upon standard process monitoring algorithms in fault identification and diagnosis tasks. The data mining research reduces a crucial barrier to the implementation of monitoring algorithms. The enhanced monitoring introduced can help engineers develop effective and scalable process monitoring systems to improve plant safety and reduce losses from fault events.