Fred Loquasto III

Ph.D. Candidate  

Department of Chemical Engineering

University of California, Santa Barbara, CA, 93106

loquasto@engineering.ucsb.edu

Abstract

Model Predictive Control (MPC) systems are widely used in the petrochemical industry for difficult multiple input, multiple output (MIMO) control problems.  MPC provides improved control performance compared to a multi-loop system by accounting for hard and soft constraints and the interactions between controlled and manipulated variables.  Due to the expense of the MPC system itself, plant operating and utility costs, and the potential loss of revenue for poor product quality, effective methods of monitoring MPC system performance are desired, together with techniques for diagnosing problems.  Although performance assessment techniques of MIMO controllers have recently been reported in the research literature, the most popular methods are based on the “minimum variance” control benchmark.  But MPC systems are not designed to provide minimum variance control; they are designed to produce optimal constrained control.  Thus, techniques that assess MPC system performance using appropriate benchmarks are required.

There are essentially four types of situations that will make an MPC system perform poorly: 1) a change in the plant dynamics, 2) large, frequent disturbances, 3) instrumentation or equipment failure, and 4) poor controller design or tuning.  For convenience, we will use the term “fault” to refer to any situation that results in poor MPC performance.  Thus, if the MPC system is monitored in order to detect the presence of faults, one is assessing the performance of the MPC system fairly; if no fault exists, the system is operating as it was designed, which may or may not result in acceptable performance.  Once a fault condition is indicated, it is desirable to diagnose the source of fault in order to correct the situation and prevent the same problem from occurring in the future.  In this research, the task of monitoring and diagnosing MPC systems has been formulated as a pattern recognition problem.  Patterns based on features of the process operating data, or the process data itself, are analyzed to detect faults or to diagnose the cause of the faulty condition. 

Two methods have been evaluated in order to help characterize the pattern recognition.  Both are based on utilizing the dynamic model that has been identified during plant tests and is used in the MPC calculation.  The first method involves the use of artificial neural networks (ANNs), while the second uses principal component analysis (PCA).  For the ANN approach, neural networks are trained with features extracted from normal and faulty operating data simulated using the process models.  Then current data are presented to the network(s) and the fault condition of the system is indicated by the network outputs.  In the PCA approach, a template, PCA model for each fault type is constructed from simulated data.  Then, a PCA model of the current data is compared to the templates using the PCA similarity factor.  A high degree of similarity between the current data and a given fault template indicates the fault condition.  Both methods have shown promising preliminary results on a simulation of the Wood-Berry distillation column operating under MPC.  General fault detection has been performed with a success rate greater than 95%, while the diagnosis task achieved an accuracy of almost 90%.