Research On Steady State Detection And Extraction Methods For Multivariable Complex System

In research of a process system, the steady state is the most basic and common assumption, which will determine the following methods for modeling, control and optimization of the system. With the rapid development of the process industry, the process system and the production methods tend to be more complicated. And the process system objects involved are high dimensional, nonlinear, time-varying. So how to find the time interval in which a complex process system is steady and obtain the data of the corresponding steady state seems to be important in process control research.In this paper, we will solve this problem from the point of view of data-driving. Combined with the traditional methods of steady state detection and the idea of statistical learning algorithm, a study on the multi variable complex process steady states detection and how to obtain data of the corresponding steady state is carried out. Based on the characteristics of the complex process object, principal component analysis, clustering and kernel density estimation algorithm which are common used in the field of statistical learning and data mining, are applied in the steady state detection. And those algorithms are performed on data from a distillation column simulation model and a power plant boiler unit as examples of algorithm verification and performance analysis.The main content of the article and its innovations are as following:Based on the improvement of F-like test algorithm on the single variable steady-state detection, a new improved F-like test algorithm based on PCA is proposed, which extends the application of the algorithm to the multi variable situation. The examples show that the PCA based method is more resistant to interference than the existing approach.Based on the fact that the data distribution is almost Gaussian when system is steady, a Gaussian mixture model is proposed to model the distribution of multiple steady states. So steady state detection can be solved by clustering. Based on this, a method of steady state detection combined hierarchical clustering and time series is proposed. And validity of the algorithm is verified by the simulink process data, which also showed our method's convenience for it doesn't need to select the key variables. Through the experiments of the data in different dimension, our algorithm complexity won't increase with the dimensional increasing, which is to say the clustering algorithm can avoid the curse of dimensionality.To obtain the data of the corresponding steady states when steady states are detected, the kernel density estimation is introduced, which is based on the estimation of the probability distribution of the data, and can obtain multiple steady states in one data set while avoiding the interference of the non-steady state data. In addition, PCA dimension reduction and reconstruction to reduce the complexity of the KDE algorithm when dealing with high dimensional data set is introduced. Improved algorithm is performed on simulation data and the results show that the PCA dimension reduction data is still satisfied with the accuracy requirements while reducing the computational complexity. Based on all above, steady state detection, steady states extraction and computational complexity reduction are performed on real power plant boiler unit data, which shows the validity of the algorithms on real process data.