Parameter Estimation of Nearly Non-Stationary Autoregressive Processes

摘要

Autoregressive modeling of noise data can be used for the detection ofmalfunctioning in a nuclear reactor at an early stage by spotting increases in the prediction error at the occurrence of an anomaly. An autoregressive model depends on a limited number of parameters, which are estimated from measured noise data. Several methods are available to estiamte these parameters. Usually these methods approximately yield the same parameter estimates. If the characteristic poles of the autoregressive process are located closely to the unit circle, it will exhibit a pseudo-periodic, almost non-stationary behaviour. In that case the Yule-Walker estimation technique should not be used to determine the autoregressive model, as it will lead to a large residual and prediction error variance as well as to incorrect results for the autoregressive parameters. In this report the underlying mathematical circumstances are shown that cause the Yule-Walker technique to be unreliable.