FDACNet: Enhancing time-series classification with fuzzy feature and integrated self-attention and temporal convolution

Abstract

Time-series classification is crucial in time series analysis and holds significant importance in real-world scenarios. Applying self-attention and temporal convolution techniques is paramount when dealing with time series data. The self-attention mechanism enables the capture of correlations between different time steps in a sequence, thereby facilitating the handling of long-term dependencies. Meanwhile, temporal convolution is designed explicitly for processing time series data, effectively capturing temporal dependencies through convolutional layers. The integration of the two technologies plays a pivotal role in time series analysis, enabling accurate temporal classification. This paper proposes a novel net with fuzzy features and integrated self-attention and temporal convolution, denoted as FDACNet. The proposed net introduces two key components: FD-FE for fuzzy dominated feature extraction, and ATCmix for integrating self-attention and temporal convolution. FD-FE captures trend information by defining gradient relationship between time points within a time series sample. On the other hand, ATCmix combines convolution and self-attention to reduce parameters and enhance efficiency in handling time-series data. Finally, the proposed method is evaluated on twenty datasets and compared against twelve other state-of-the-art approaches. Experimental results demonstrate the superior classification accuracy of the proposed model, showcasing a 5.2% and 7.1% enhancement in average accuracy compared to the state-of-the-art convolution-based and transformer-based methods ModernTCN and iTransformer.