This paper describes a novel approach for nonlinear signal modeling and prediction. We propose a nonlinear extension of the conventional AR model by using several linear models, each covering a subset of the whole signal data. Considering that AR modeling is conducted by associating a delay vector with the future value to be predicted, we interpret the signal as a codebook of patterns having the desired predicted value as the associated output. The pattern groups associated with each AR model are obtained by competition among the linear models as they are trained: the competition gives us the AR models as well as a classification of the data patterns. The obtained data groups are the basis for estimating a segmentation model, i.e., a classifier that, given a new data pattern, indicates what linear model should be used. The combination of the classifier and the linear models constitutes the final system. Several practical applications show the advantages of our approach. [References: 26]
展开▼
