关键词: 关键词: 时间序列, 注意力机制, 位置编码, 深度神经网络, 多尺度融合

Abstract: To address the problem of low classification accuracy caused by insufficient potential information interaction between time series subsequences, a time series classification method based on multi-scale cross-attention fusion in time-frequency domain called TFFormer (Time-Frequency Transformer) was proposed. First, time and frequency spectrums of the original series were divided into subsequences, and the point-value coupling problem was solved by adding positional embedding after linear projection. Then, through the improved multi-head self-attention mechanism, the long-term sequence dependency problem was solved, and the model is made to focus on more important features. Finally, a multi-scale cross-attention module was proposed to enhance the interaction between the time domain and frequency domain, so that the model can further mine the frequency information of the sequence. The experimental results show that compared with Fully Convolutional Networks (FCN), the classification accuracy of the proposed method on the Trace, StarLightCurves and UWaveGestureLibraryAll datasets reaches 100%, 97.6% and 92.3%, respectively, which is better than 99.7%, 96.7% and 90.9% of FCN. It is proved that by enhancing the information interaction between time domain and frequency domain of the time series, the model convergence speed and classification accuracy can be improved.

Key words: time series, attention mechanism, positional embedding, deep neural network, multi-scale fusion