STOCHASTIC THRESHOLD MODELING FOR LINEAR AND NONLINEAR SYSTEM MONITORING AND DIAGNOSIS (DYNAMIC DATA, NONPARAMETRIC, PARAMETRIC ON-LINE ADAPTIVE MODELING, ARMA)

This thesis presents both the theory and applications of a threshold-type adaptive modeling algorithm for the identification of parameters in systems containing multi-valued nonlinearities. Hysteretic restoring force and mechanical backlash in a single-degree-of-freedom dynamic system are represented by a class of threshold-type nonlinear autoregressive moving average models with unknown parameters identifiable by the nonlinear least squares methods. All the possible nonlinearities are searched by means of a systematic signal processing approach.; The choices of the numbers, values, lags and orders of a threshold model were here investigated in detail; it was determined which parameters are identifiable and to what accuracy. In order to force the nonlinear system to exceed the yield deflection point and enter the plastic region, sufficiently large input excitations were selected. The effects of measurement noise on the parameter estimates were studied by means of numerical experiments. Indetification was successful with 20% additive measurement noise in the univariate (self-limiting), and bivariate sequences. Five synthetic processes (nonlinear spring, nonlinear damping, saturation, hysteresis and deadband) and a real data set (Canadian lynx cycle) were employed to demonstrate the effectiveness of the present approach.; The application of Threshold Nonlinear Dynamic Data System (TNLDDS) methodology to the real process--drill wear prediction and monitoring--was carried out for an automatic tool replacement system. The present methodology offers, especially for multi-leveled nonlinear systems, a significant improvement in prediction accuracy over a linear modeling approach.; Advances in computer technology have made the manipulation of time series more feasible and practical. With the aid of digital computation and on-line algorithm development, the Dynamic Data System (DDS) modeling technique has become more attractive for problems in several fields of engineering and science. This thesis enhances the feasibility of such an approach and extends it to deal with multi-valued nonlinear problems and it has made possible a unified treatment of on-line system monitoring and signature analysis.