Abstract: Model ensemble methods train multiple basic models and use a certain rule to aggregate the output of the basic models for time series classification.However,they mainly focus on two aspects.The first one is which model is chose as the basic mo-del.And the Second one is how to increase the difference and the diversity of the basic models.They all ignore the exploration of aggregation rules.Aiming at this problem,Transformer feature fusion network for time series classification(TFFN) is proposed.TFFN have two key components,dual Transformer encoder decoder(Dual TED) and Transformer encoder head(TEH).Dual TED leverage attention module to fuse the basic feature into more discriminative fusion features.Transformer encoder head,a sample-distribution-aware classifier,is adopted to classify time series more accurately.Experiments show that TFFN achieves state-of-the-art results on multiple mainstream time series classification datasets.

Key words: Time series classification, Model ensemble, Transformer, Feature fusion, Deep learning