Combining physical modeling, neural processing, and likelihood testing for online process monitoring

Two basic approaches can be taken to model physical systems. One approach derives a model structure from the known physical laws. However, obtaining a model with the required fidelity can be difficult if the system is not well understood. A second approach is to employ a black-box structure to learn the implicit input-output relationships from measurements in which no particular attention is paid to modeling the underlying processes. Clearly, one should exploit whatever physical insight one has into the properties of the system. This paper describes a method that draws on the respective strengths of each of these two approaches. The technique combines known first principles knowledge derived from physical modeling with measured input-output mappings derived from neural processing to produce a computer model of a dynamical process. The technique is used to diagnose operational changes of mechanical equipment by statistically comparing, using a likelihood test, the predicted model output for the given measured input with the actual process output. Experiments with a peristaltic pump were conducted and results are presented here.