- 學位論文
高斯程序於模型辨識，控制與效能評估：數據選擇以改善模型不確定性
Gaussian Process Model Based Process Identification, Control and Performance Assessment: Data Selection for Model Uncertainty Improvement
摘要
全球產業都很重視提高效率，降低運營成本因此具有重要意義。通過新的方法，以及資源的增強減少對環境和健康的影響，有很大的重要性。因此，有必要提昇現有的方法，並引入新的，突破性的技術。在化學工程的研究，程序識別，控制和性能評估是活躍的研究領域。本文旨在提出基於高斯過程 (Gaussian Process) 方法來改進序程在上述方面的運作。在高斯過程模型方法會有因為矩陣的求逆的高計算量問題。本研究提出了一遞歸更新協方差矩陣的方法。該更新方法是有選擇性，只對需要改善的區域進行更新，避免高運算的需求。此外，配合修剪過程除去冗餘數據使得模型的數據量維持精簡。目前工業使用的主要控制器仍然是PID 控制，其調整方法則是基於程序的模型。因此，該模型的完整性是非常重要的，而有關模型預測的可信度的資訊則是有利的。高斯過程模型能提供可信度的資訊，因此本研究提出了基於高斯過程模型PID控制調整。利用方差信息於控制，達到安全與性能權衡的控制。此外方差信息提供了改善模型的一個方法。控制器性能評估的目的是衡量控制系統的能力，以改善系統的運作。過去性能評估，模型被認為完整的，但在實際情況下，這可能並非如此。該模型的準確性方取決於數據的品質。其結果是，基於模型所得到的性能指標實際上的不是最佳值。因此在考慮改善目前的程序時，有必要提供操作員更多的資訊。本研究提出基於高斯過程 (Gaussian Process) 的性能評估架構，利用方差作為模型的可信度資訊，以及對應的性能指標，並判定改善的可能性。此外，該方法可應用在非線性系統的評估。本研究方法的特性會透過測試範例來呈現。
並列摘要
The global industries place a great importance in increasing efficiency to reduce operating costs. This also have a great importance in minimizing environmental and health impacts, through new industrial approaches, as well as an enhancement of resources. Therefore, there is a need to upgrade existing approaches and introduce new, breakthrough technologies. In chemical engineering research process identification, control and performance assessment are three areas of active research. This thesis aims to propose methodologies to improve the operation of process with regards to the aforementioned areas of research based on the Gaussian Process (GP) model. The GP model based method can encounter a high computation load because of the inversion of matrix. In this work, a method which recursively updates the covariance matrix is proposed. The update scheme is selective by admitting data to the region requiring improvement. This enables the model to be updated without placing a high computation demand. In addition, the process is augmented by a pruning process which removes redundant data from the model to keep the data size compact. The predominant controllers used in most process industries are still mainly PIDs and model based method is used for controller tuning. The integrity of the model is therefore very important and information on the model based prediction can be invaluable. The GP model gives the information on the reliability of the prediction. A GP model based PID tuning control has been proposed in this study. The variance information is used for control which results in a safety-performance trade-off control. In addition the variance information provides a mean for the selection of data to improve the model at successive control stage. The aim of the controller performance assessment is to determine and measure the capability of control systems in order to improve the degradation performance. Conventionally it is assumed the model is perfect but in actual situation this may not be the case. In terms of accuracy of the model it is only as accurate as the quality of the data available. The consequence is that the performance index based on the identified model is not the actual optimum. This leads to a need for the providence of trust on the performance index that provides the operators with better information when considering improving the current process. A GP model based performance assessment framework is proposed using the variance of the predictive distribution as a confidence level of the model and thus the corresponding performance index. The proposed index is used to provide a trust region on the performance index. Based on the indication of the model quality the reliability of the evaluated performance index can be indicated. It is used to provide a framework for judging whether improvement to the control structure as well as model is worthwhile. Moreover the proposed method provides a framework for nonlinear process assessment. The capabilities of the proposed methods are demonstrated through a series of case studies.