Fractional-order multiscale attention feature pyramid network for time series classification

For time series classification, it is a key problem needed to be solved that the deep learning methods do not consider the relationships between different feature layers in neural networks. Additionally, the deep learning methods are slightly inadequate for the feature-learning ability of single-channel time series data. To address these problems, we propose the fractional-order multiscale attention feature pyramid network(FO-MAFPN) to extract feature sequences of different scales from multichannel time series and fuse features of different levels. First, FO-MAFPN forms multichannel time series using a convolutional layer based on fractional-order to increase sample diversity. Subsequently, it uses a pyramid network based on the multiscale attention mechanism to combine low-level detail and high-level abstract semantic information. This pyramid network adaptively learns channel-attention scores and promotes effective feature sequences while suppressing invalid feature sequences. Finally, FO-MAFPN classifies time series by fusing the feature sequences of a long short-term memory layer with the output of the feature pyramid network. The results of experiments conducted on 85 datasets from the UCR archives indicate the superiority of the FO-MAFPN model for time series classification problems.