Scholarlink

<p>Transformers have significantly advanced deep learning by effectively long-range dependencies in sequential data. However, in (EEG) signal processing, traditional transformers struggle to capture local temporal correlations due to the permutation-invariant nature of self-attention. To address this limitation, we propose a novel hybrid deep neural network architecture that combines convolutional neural networks (CNN), recurrent layers (LSTM, GRU, or SimpleRNN), and a sequential learning transformer (SLTrans) module for EEG classification. Our model comprises two primary components: (1) a CNN module that extracts spatial features and captures short-range temporal dependencies and (2) an SLTrans module, which we enhance by incorporating LSTM/GRU/SimpleRNN layers between the multi-head attention (MHA) and the feed-forward network (FFN). This design refines the attention outputs, enabling more effective sequential feature learning and a richer representation of both short- and long-term dependencies. The recurrent layers help mitigate the transformers limitations in modeling temporal dynamics, reduce noise, and preserve local feature characteristics. We evaluate our approach on a publicly available real-world EEG dataset. The proposed models—CNN + LSTM + SLTrans, CNN + GRU + SLTrans, and CNN + SimpleRNN + SLTrans—achieve classification accuracies of 98.64%, 98.28%, and 96.73%, respectively. These results highlight the effectiveness of our fusion-based framework, demonstrating its potential for robust spatial–temporal feature extraction and high-performance EEG signals&#xa0;classification.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

SLTrans: a sequential learning transformer for extracting spatial–temporal features from EEG signals

  • 期刊论文

  • 发表日期：

    2025年9月29日

查看全文
  • Gunda Manasa,
  • Krashana D. Nirde,
  • Suhas S. Gajre,
  • Ramchandra R. Manthalkar

摘要

Transformers have significantly advanced deep learning by effectively long-range dependencies in sequential data. However, in (EEG) signal processing, traditional transformers struggle to capture local temporal correlations due to the permutation-invariant nature of self-attention. To address this limitation, we propose a novel hybrid deep neural network architecture that combines convolutional neural networks (CNN), recurrent layers (LSTM, GRU, or SimpleRNN), and a sequential learning transformer (SLTrans) module for EEG classification. Our model comprises two primary components: (1) a CNN module that extracts spatial features and captures short-range temporal dependencies and (2) an SLTrans module, which we enhance by incorporating LSTM/GRU/SimpleRNN layers between the multi-head attention (MHA) and the feed-forward network (FFN). This design refines the attention outputs, enabling more effective sequential feature learning and a richer representation of both short- and long-term dependencies. The recurrent layers help mitigate the transformers limitations in modeling temporal dynamics, reduce noise, and preserve local feature characteristics. We evaluate our approach on a publicly available real-world EEG dataset. The proposed models—CNN + LSTM + SLTrans, CNN + GRU + SLTrans, and CNN + SimpleRNN + SLTrans—achieve classification accuracies of 98.64%, 98.28%, and 96.73%, respectively. These results highlight the effectiveness of our fusion-based framework, demonstrating its potential for robust spatial–temporal feature extraction and high-performance EEG signals classification.