An embedded neural network approach to nonlinear system identification with application to model-based machinery diagnostics.

摘要

In model-based machinery diagnostics the machineries to be monitored are treated as dynamic systems and the models of the systems are utilized to extract the diagnostic information. To apply a model-based strategy to the monitoring of machinery, a model of the machinery must first be established. Most machineries are nonlinear systems to some extent. Often, it is the problem of how to represent nonlinearities in the system that presents the most difficulties in modeling the system.; Neural networks have been used as a powerful tool for modeling nonlinear dynamic systems in the form of black-box models. But black-box models can not reveal insights into the systems, so this type of models only has limited use in model-based diagnostics.; In this dissertation, a methodology for modeling nonlinear dynamic systems using embedded neural networks within the frameworks of physical models of the systems is developed in the interest of model-based machinery fault diagnostics.; Neural networks' capability of approximating commonly-encountered nonlinearities in mechanical systems is investigated. The algorithms for tuning the embedded neural networks and simultaneously estimating the initial states of the systems are derived. Some examples of system identification, including a piezoelectric actuator with complex hysteresis, are presented to demonstrate the effectiveness of the proposed approach.; The methodology is then applied to a challenging machinery diagnosis problem—condition monitoring of turbine rotors. The proposed approach simultaneously deals with the task of detecting a transverse crack in a rotor shaft, determining the location of the crack, and estimating the size of the crack. The method of incorporating the effect of crack into the finite element models of the rotors using embedded neural networks is established. Two monitoring schemes are proposed, two identification algorithms are evaluated, and the sensitivity and accuracy of the proposed approach are demonstrated.