Exploring The Cause Of Abnormal Pattern In Complex Systems Based On Dimension Reduction

The Complex system records a large number of high-dimensional data, and thehigh-dimensional features of the data tend to exhibit a high degree of coupling and thestrong association of linear or non-linear phenomena, and even include extraneousnoise. At the same time, with the increasing of the possibility of the abnormal state inthe production process, abnormal patterns in complex systems happen frequently.Timely diagnosis and to exclude abnormal patterns in the system have become themost important problem that engineering staff and research scholars should solve theproblem without delay. To solve this problem, the way to explore the cause ofabnormal pattern in complex systems based on dimension reduction is given in thisarticle, which comes from the point of view of the model section in pattern recognition.Firstly, in order to solve the linear redundancy and noise problems of high-dimen-sional features in complex systems, an original feature selection method forclassification models based on Partial Least Squares (PLS) and False NearestNeighbors (FNN) is established as follow: the original feature similarity measure isestablished by FNN; the use of the principal component space obtained by PLS featureextraction meets the requirements of establishing the FNN similar measure, whichrequires the space has orthogonal property; the original feature similarity measure isgiven by combining with PLS and FNN, and the similarity measure portrays theimportance of the original feature to the category labels, then gets the method toexplore the cause of abnormal pattern in the linear models. In order to verify theeffectiveness of this method, the simulation of three classification models of complexsystems in linear models is give through using the above method in order to select thekey features that cause the exception category and the accuracies of classification ofthe Support Vector Machine (SVM) classifier. The experimental results have shownthat the features selected are the essential features and SVM has high accuracies ofclassification, thus this indicates that the method is appropriate and effective.Secondly, in order to solve the nonlinear redundancy and noise problems ofhigh-dimensional features in complex systems, an original feature selection methodbased on Kernel Partial Least Squares (KPLS) and FNN is established as follow:kernel function method transforms the nonlinear problem into a linear problem, thenthe nonlinear feature selection method based on KPLS+FNN is given by using PLS+FNN linear feature selection methods, in order to explore the cause of abnormal patternin the nonlinear models. Similarly, the simulation of three complex system models in the nonlinear models and the experimental results also show that the method isappropriate and effective.Finally, the classic control process—Tennessee Eastman(TE) process is used todo the empirical research. After selecting three kinds of abnormal patterns in TEprocess, KPLS+FNN nonlinear feature selection method is selected to explore thecause of abnormal pattern, due to its nonlinear models. We also have compared theexperimental results with the results existed.The corresponding simulations and the results of empirical research show that thePLS+FNN feature selection method under linear model and KPLS+FNN featureselection method under non-linear model are able to select out features related thecategory, so the methods is applicable and effective. The work provides a way toexplore the key cause of abnormal pattern in the complex system.