| dc.contributor.author | Webber, Jonathan R. | en_US |
| dc.date.accessioned | 2014-10-21T12:34:13Z | |
| dc.date.available | 2007 | |
| dc.date.issued | 2007 | en_US |
| dc.identifier.other | AAINR27652 | en_US |
| dc.identifier.uri | http://hdl.handle.net/10222/54931 | |
| dc.description | Since controllers play a vital role in industrial processes, it is important to have methods that monitor controller performance and diagnose the reasons when performance is poor. This thesis deals with the inherent uncertainty involved in measuring the controller performance. Since a common cause of poor performance in model-based controllers is model error, this thesis also proposes several closed-loop methods for diagnosing the presence of model error. | en_US |
| dc.description | The estimates of the statistical properties of the minimum variance controller performance index need to be made by using small samples since in practice the process data may only be stationary over short intervals. Residual based bootstrapping is used to estimate these statistical properties. It is demonstrated that accurate confidence intervals can be obtained by using small samples in case of normal and non-normal innovations. The broader applicability of the bootstrapping approach is also demonstrated by estimating the sampling distribution of the closed loop settling time performance index. An experimentally determined sampling distribution for the minimum variance performance index shows that the index can have strong non-normal behavior. The distribution estimated by the proposed bootstrap method is shown to capture the overall features of the experimental distribution. Also, a bootstrap method is proposed which can capture the effects of model-order uncertainty on the sampling distribution of the performance index. | en_US |
| dc.description | While bootstrapping of the controller performance index addresses the detection of performance issues, several closed loop methods are also proposed to diagnose (and in some cases correct) model-plant mismatch. An iterative method is proposed for univariate systems to detect and correct gain and dead-time mismatch for models of arbitrary order. This method is also used to correct for time constant mismatch in first order plus dead time models. A multivariate cross-correlation method is presented to help detect which specific models in model-based controllers are mismatched. A partial control method is developed which utilizes the patterns contained in the closed loop step response of the prediction error to set point changes to assist in the determining which models are mismatched in multivariate model-based controllers. A sieve bootstrap method is developed to estimate the confidence bands for the impulse response functions determined via closed loop correlation analysis. Further the method is used to estimate the confidence intervals for the process gains. This method can help detect the presence of statistically significant model mismatch, and also the extent of the mismatch. | en_US |
| dc.description | Thesis (Ph.D.)--Dalhousie University (Canada), 2007. | en_US |
| dc.language | eng | en_US |
| dc.publisher | Dalhousie University | en_US |
| dc.publisher | | en_US |
| dc.subject | Engineering, Chemical. | en_US |
| dc.title | Bootstrap estimation of controller performance indices and detection of model mismatch. | en_US |
| dc.type | text | en_US |
| dc.contributor.degree | Ph.D. | en_US |
