Identification And Its Applications For Non-uniformly Sampled-data Systems

The non-uniformly sampled-data (NUSD) systems with irregular sampling intervalsfor the inputs and/or outputs are a class of more general multirate sampled-data systems.Due to hardware limitations, economic considerations and environmental impacts, suchNUSD systems can be widely found in petroleum, chemical, food, medicine and otherprocess industries. However, since traditional identifcation methods and control theoryare designed mainly for single-rate sampled-data systems, research on identifcation andcontrol of the NUSD systems is still relatively sparse. Therefore, this thesis focuses ondevelopment of new identifcation algorithms for the NUSD systems and investigates theirapplications in the inferential control design and the soft sensor development, which hasimportant theoretical signifcance and practical values. The major contribution of thisthesis includes:(1) Identifcation of the input NUSD systems under diferent noise interferences is stud-ied. First, a gradient-based iterative algorithm is derived for the output error modelof the input NUSD systems with white noise. The proposed algorithm not only hasbetter identifcation performance than the auxiliary model based stochastic gradi-ent algorithm, but also has lower computational cost than the least squares basediterative algorithm. Furthermore, a fltering based recursive least squares algorithmis derived for the Box-Jenkins model of the input NUSD systems with colored noise.The proposed algorithm interactively estimates the parameters of the system modeland the noise model; thus has low computational load and high estimation accuracy.(2) For the synchronous input-output NUSD systems with colored noises, the Box-Jenkins model is derived based on the lifting technique, and an auxiliary modelbased multi-innovation generalized extended stochastic gradient algorithm is pre-sented for the model identifcation. The proposed algorithm makes full use of theavailable data and improves the identifcation performance via innovation expan-sion. Considering that the lifted state-space model involves causality constraintproblem and the corresponding transfer function model is quite complicated withtoo many parameters, neither one is convenient for identifcation and control ofthe synchronous input-output NUSD systems. Therefore, a novel transfer functionmodel is derived by introducing a time-varying backward shift operator, and anauxiliary based least squares algorithm is developed for its identifcation.(3) An Expectation-Maximization (EM) based identifcation algorithm is developed forthe output NUSD systems with uncertain sampling delays, where the uncertain delays are considered as the hidden states, and the parameters of the underlyingfast-rate fnite impulse response model are estimated along with the delays. Fur-thermore, two algorithms are proposed to recover the approximated fast-rate outputerror model by using the least square algorithm and the EM algorithm, respectively.The simulation examples and experimental results illustrate that the proposed al-gorithm can overcome the negative impacts of uncertain delays on the identifcationperformance, and provide parameter estimates with high precision.(4) For the input NUSD systems with fast non-uniformly updated inputs and slowlysampled outputs, the mathematical relationship between the transfer function modelof the sampled outputs and the non-uniform missing outputs is derived. Basedon this relationship and using the identifed model of the sampled outputs, themodels of the missing outputs can be estimated. Then by using the estimatedmodel to predict the output at the next non-uniform sampling instant, an inferentialadaptive control algorithm is proposed for the input NUSD systems according tothe minimum variance criterion. By using the auxiliary variables, the proposedinferential control scheme is further extended to the input NUSD systems disturbedby unknown colored noises.(5) For a practical output NUSD system, i.e., the once-through steam generator, softsensors based on hybrid modeling technique are developed to predict its steamqualities. First, simplifed frst-principle models are derived according to the en-ergy balance equations, thus the steam qualities can be estimated online throughtemperatures, fows and other process variables. Then, to improve the adaptabilityof the soft sensors, of-line lab analyses of steam qualities are adopted to rectifythe frst-principle model predictions via bias compensation; and to obtain high-performance soft sensors, the predication error method is applied to optimize themodel parameters and the weighting factor in the bias update equation. Further-more, online outlier detection and rejection are considered to ensure the reliabilityof the developed soft sensors.