Classification Of Nonlinear Time Series Using Functional-coefcient Regression

In raw reeling silk industry, engineers often need to measure and compare thequality of cocoon filaments according to their sizes. However, cocoon filament is ex-tremely thin, so it is difcult to measure the size series of cocoon filament. Moreover,size series of cocoon filament is usually non-stationary series with length less than30.These features make it difcult to provide a suitable model for individual size series.In this paper, we will establish a unified functional coefcient auto-regressionmodel for size series of cocoon filaments from a common population. By definingappropriate objective function, we utilize kernel function and least square method tomake local linear estimation for observations of size series of cocoon filament.Based on the above model and estimation method, we also propose a methodsimilar with EM and aim to classify size series of cocoon filament from diferent popu-lations. The proposed method is not only suitable for series with various length, it canalso deal with multiple classification problem. Several simulation examples and a realapplication both show the good performance of the classification method.