A Multi-Scale Feature Fusion Transformer for Multivariate Time Series Classification
Yuzhuo Xu, Ningning Jiang · 2024
In today's society, multivariate time series classification has wide-ranging applications, such as action recognition, health monitoring, and so on. In recent years, with the breakthrough progress of transformer networks based on self-attention mechanism in various fields, their powerful long sequence modeling ability has also shown good performance in time series classification tasks. However, present transformer networks utilized for multivariate time series classification predominantly encode the values of all variables at each timestamp as inputs to the encoder, aimed at computing dependencies between temporal steps. This practice often neglects certain dynamic features inherent in time series data, such as periodicity. In response to this limitation, the present study introduces a multi-scale feature fusion transformer (MFformer), a model that employs the Fast Fourier Transform (FFT) to ascertain periods with the most significant influence and divides the original time series data into subsequences according to these periods for input into the transformer network, thereby capturing dependencies among periods. To extract temporal features within these periods, MFformer replaces the feadforward layer in the original transformer with two convolutional layers. Concurrently, MFformer adopts a flow-attention mechanism in lieu of the traditional self-attention mechanism to achieve a lower temporal complexity. Finally, a gating mechanism is employed to integrate features from subsequences of different scale. The efficacy of the proposed approach is demonstrated through validation on eight publicly available datasets.