Minimum Redundancy Maximum Relevancy-Based Multiview Generation for Time Series Sensor Data Classification and its Application

The extraction and ensemble of more diverse time series features has been continuously focused on by univariate time series classification (TSC) algorithm development. However, the feature stability is considered rarely by TSC algorithms, resulting in decreased algorithmic robustness to disturbed data and reduced classification performance. This article proposes a minimum redundancy maximum relevancy-based multiview generation (MRMR-MVG) TSC algorithm utilizing an ensemble feature selection architecture, improving the feature diversity and stability to benefit classification ability. Specifically, the raw series is mapped to the various series spaces containing different information, and the dilated convolution features are extracted from multiscale to increase feature diversity efficiently. Then, based on mutual information, extracted features are recombined via splitting and concatenation to generate three initial views that maximize correlation with labels. To further optimize the views, the objective function, which represents the feature correlation within the view, is minimized by exchanging allelic features between views based on the greedy strategy, improving the view features diversity and stability. Finally, the predicted outputs of three views are quickly ensemble via hard voting to get the final output label in the multiview ensemble. The effectiveness and advancement of the proposed MRMR-MVG algorithm are verified by comparative experiments on public UCR archive and fault diagnosis application on real excavator sensor datasets.