(345f) Data Mining and Monitoring for Industrial Scale Chemical Processes

Within the last decade, modern chemical plants have begun to log large amounts of process data every day. Great potential exists in identifying patterns and knowledge emergent from the data that could enhance operators’ abilities to respond to changes in the state of the process, however a gap exists between the amount of data and humans’ ability to process all the knowledge and patterns it contains. To answer this problem, “big data” algorithms have been created to automatically organize and interpret data from chemical plants. Even with the assistance of advanced computing techniques, challenges remain in isolating the relevant data for the specific monitoring problem.

 This research seeks to develop efficient and scalable methods for leveraging data mining algorithms for process monitoring problems. A major challenge in industrial process monitoring and fault detection is the large amount of data from normal operations and the comparatively small amount of interesting data from faulty operations. Our approach utilizes the Self-Organizing Map (SOM) for its topology preservation and data representation abilities to isolate the key variables and patterns that emerge from data to facilitate the separation of faulty data from normal conditions. The goal is to achieve a data-based model for fault detection and diagnosis to provide helpful process information to operators during fault events.

 The techniques described are applied to case studies on large industrial scale reactors including a continuous flash tower and a batch polymerization reactor. Both cases present different challenges in processing the data, characterizing the region of normal operations, and detecting faulty data. In order to create a robust monitoring scheme, it is critical that process variations be taken into account. Variations in the data arise from many different factors including feed grade, outside weather conditions, and the condition of the equipment. Through sampling techniques, manipulation of the process variables themselves, and with the assistance of a limited amount of process knowledge, standard process monitoring techniques can be made more effective and reliable. The results show how SOM can be usefully applied to chemical processes to find trends hidden in the data and to accurately detect disturbances and process events.