Nonlinear filtering techniques for failure detection in dynamic systems.

摘要

Developing on-line methods for detecting and locating process malfunctions is an important goal towards full automation of systems. Model-based methods, which check the consistency of the observations using the known functional relationships among the process variables seem to have the potential of early detection of slowly developing failures in complex systems.; This dissertation deals with nonlinear filtering as a method for failure detection in dynamic processes. The failure detection problem is presented here as an Identification Problem, where no jumps between the possible operating modes are assumed, i.e. there is only uncertainty with regard to which is the present mode. The identification filter consists of parallel Kalman filters, each tuned to a different system mode of operation, whose estimates parametrize the conditional probability equations. An extension of the filter is derived for the case of different measurement noise coefficients in different modes for the continuous-discrete case. In the continuous-time case with different measurement noise intensities the likelihood function is shown to be ill-defined as the induced measures become singular.; The average performance of the continuous-descrete as well as the continuous-time identification filters are studied. It is shown that on the average and after a sufficiently long time a correct decision is expected for any decision threshold level.; An alternative identification filter structure is derived using the maximum-likelihood estimation philosophy. The filter reduces to parallel Kalman filters, feeding the state estimates to a maximum likelihood generator which then chooses a set of indicator functions to maximize the total likelihood.; Some aspects of interpreting a sequence of decisions, choosing a decision threshold and reinitializing the filter are discussed qualitatively using simulation examples. Furthermore, a new approach based on scaling of the likelihood functions is presented. Scaling is shown to be equivalent to choosing a threshold level for the conditional probabilities.