ABiESN-MTCN: A Novel Time Series Classification Method

Echo State Networks (ESNs), a highly efficient variant of recurrent neural networks (RNNs), have been demonstrated significant success in time series classification tasks. How-ever, traditional ESNs are limited to capture local patterns from past time steps, which hinders their capacity to effectively model long-term dependencies. To overcome this limitation, we propose a novel time series classification framework, the Attention-enhanced Bidirectional Echo State Network integrated with Multiscale Temporal Convolution Network (ABiESN-MTCN). This model consists of two main components: an encoder and a decoder. In the encoding phase, the time series data is fed into a Bidirectional Echo State Network (BiESN), which generates bidirectional echo states for each time step. These states are further refined using a self-attention mechanism to better capture long-term dependencies. In the decoding phase, we employ a Multiscale Temporal Convolutions Network (MTCN) with the max-over-time layer to extract the most discriminative features. Finally, the model uses a fully connected layer and a softmax layer for classification. Experimental results on 65 UCR datasets demonstrate that ABiESN-MTCN outperforms MLP, FCN, ResNet, and EMAN.