An identification scheme combining first principle knowledge, neural networks, and the liklihood function

An identification scheme is described for modeling uncertain systems. The method combines a physics-based model with a nonlinear mapping for capturing unmodeled physics and a statistical estimation procedure for quantifying any remaining process uncertainty. The technique has been used in predictive maintenance applications to detect operational changes of mechanical equipment by comparing the model output with the actual process output. Tests conducted on a peristaltic pump to detect incipient failure are described. The inclusion of unmodeled physics and a statistical representation of uncertainties results in lower false alarm and missed detection rates than other methods.