Feature Selection-Based Multiview Concentration for Multivariate Time Series Classification and Its Application

In this article, a feature selection-based multiview concentration (FS-MVC) algorithm is proposed for multivariate time series classification (MTSC). The data-driven excavator fault diagnosis task is used as an application case of the MTSC algorithm. The three steps of FS-MVC comprise multidimensional feature extraction, multiview concentration, and multiview ensemble. In multidimensional feature extraction, the multivariate time series (MTSs) are mapped to multiple spaces by various series transformations to extract dimension-dependent and dimension-independent features. A new time series pooling feature is extracted, which denotes the power of proportion of positive values (PPPV) in the feature map generated by the convolution operation. The ensemble-group feature selection framework is introduced into the MTSC task to remove redundant features and generate multiviews through feature vector concentration in multiview concentration. In multiview ensemble, each view feature is input into the corresponding classifier, and then, the final prediction label is obtained by the hard voting ensemble. Feature diversity is improved via the series transformations and dimension-dependent and dimension-independent PPPV features, while feature stability is enhanced significantly by the ensemble-group feature selection framework. Finally, the superiority of FS-MVC over other state-of-the-art MTSC algorithms is demonstrated in both contrast experiments on the public UEA MTS and the practical excavator datasets.