TOPOLOGICAL PRESERVATION TECHNIQUES FOR NONLINEAR PROCESS

MONITORING

M. Thomas and J.A. Romagnoli Department of Chemical Engineering Louisiana State University

Baton Rouge, LA, 70804
Recent advances in computing power and data storage have created an explosion of information, a trend that has not been ignored by chemical engineers. Modern computerized plants can build large historical data bases documenting plant performance and operations events. The challenge becomes leveraging â??big dataâ? techniques for retrieving important information and developing useful models of plant dynamics to assist operators.
The goal of this research is to develop process monitoring technology capable of forming a useful model from large stores of accumulated process data to create a basis for fault detection and diagnosis. There is demand for new techniques for the monitoring of non-linear topology and behavior, and this research presents a topological preservation method for process monitoring using Self Organizing Maps (SOM). The novel architecture proposed adapts SOM to the full spectrum of process monitoring tasks including fault detection, fault identification, fault diagnosis, and soft sensing.
Specifically, fault detection is performed using a SOM-Gaussian mixture model (GMM) to model the probability density function of classes of data. For fault identification, a new method is introduced using an observationâ??s residuals from the SOM fit to the normal region. Fault diagnosis is done by creating a map for each fault, known as multiple self-organizing maps (MSOM), and associating an observation with the map it most closely resembles. The architecture with multiple maps allows new faults to be included without directly affecting previously characterized faults.
The proposed methodology is applied to the Tennessee Eastman process (TEP) simulation. Previous studies of the TEP have considered particular step-change faults where the root causes of the disturbance are generally limited to one variable. Here, the focus is on characterizing more challenging faults such as random variations, sticky valves and slow drift in kinetics to

effectively illustrate the advantage afforded by SOM. Results indicate that MSOM is able to improve upon linear distance preservation techniques such as PCA and the more standard SOM based approach in process monitoring tasks. This presentation will also briefly describe the results of ongoing studies on an industrial application