Abstract
The paper proposes a three-step computational strategy for detection of changes in the performance of machines and complex systems. Once a distinct change is discovered, the information is further used for diagnosing the cause of the system malfunction.At first, a small number of (usually three-four) parameters are selected, with physical cause-effect relations between them. Then, the first step of the computational procedure is to model the relationships between these parameters from two different operation periods (two different data sets). Here a special two-stage learning strategy for creating normalized radial-basis (NRBF) network models is proposed.Then, in the second computational step, a subtraction of the outputs from these models is made on a preliminary defined grid in the input space of parameters. These data are further used as new input information for generation of fuzzy rules by a specially proposed algorithm. The generated fuzzy rules show the amount of the difference in the system performance in the different regions of the input space.A given threshold can be used to detect the areas with unacceptable changes, which is an evidence for deterioration of the system performance.Data from a real performance of a diesel engine are used in the paper to illustrate the whole procedure for performance evaluation and to show its practical applicability.
