B

Brain–Computer Interface (BCI) technology is a multidisciplinary field that merges artificial intelligence (AI) and neuroscience. It monitors brain activity through sensors and AI, facilitating direct interaction between the human brain and external devices. As an innovative neural computing technology, BCI is implemented in applications such as robotic arms [1], Virtual Reality (VR) systems [2], and wheelchairs [3].

Motor imagery (MI) is a widely adopted paradigm in BCI systems, utilizing electroencephalogram (EEG) signals generated when users mentally imagine body movements. This technique enhances BCI applications by enabling control without interference from the external environment [4]. During the capture of MI EEG signals, the cerebral cortex produces two rhythmic signals: the mu rhythm (8–13 Hz) and the beta rhythm (13–32 Hz). Additionally, the associated brain regions display Event Related Desynchronization (ERD) and Event Related Synchronization (ERS) phenomena. These phenomena are critical for understanding the neural mechanisms underlying MI-based BCI systems [5]. Machine learning plays a crucial role in the development of BCI systems, particularly in classification tasks. Several advanced algorithms, such as Artificial Neural Networks (ANNs) [6], Support Vector Machines (SVMs) [7], and Bayesian classifiers [8], have been employed. However, traditional machine learning approaches often struggle with EEG decoding because of their limited ability to extract features effectively. For instance, Common Spatial Patterns (CSP) can extract spatial features but are susceptible to noise and artifacts that may obscure critical information, often neglecting the temporal aspects of EEG signals.

With the advancement of deep learning, its applications are expanding across various fields [9]. By harnessing its strengths in processing high-dimensional data and modeling non-linear relationships, researchers are progressively utilizing deep learning methods to interpret complex MI-EEG signals [4]. Different neural network architectures are employed to enhance feature extraction and improve decoding accuracy. For example, Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and Autoencoders (AEs) are effective models in this context. Notably, Poonam et al. [10] combined Sparse Nonnegative Matrix Factorization with CNNs, achieving better performance compared to using CNNs alone. Wang et al. [11] proposed a network named IFNet, which enhances the mutual learning of spatial features between low-frequency and high-frequency bands. Liu et al. [12] introduced a weight-sharing network that integrates CNN and LSTM, reducing computational load and speeding up network training. Additionally, Phadikar et al. [13] proposed a motor imagery EEG decoding framework that maps EEG signals into AEs weight vectors, extracts multiple discriminative features, and employs SVM for multiclass classification.

EEG signals are typically analyzed for motor imagery (MI) task classification by extracting combinations of spatial, spectral, and temporal features. Among these, the integration of spatial and temporal features has attracted considerable interest from researchers. Liu et al. [14] proposed an adaptive 3D CNN that assigned significant weights to spatial channels related to motor information and temporal features, effectively reducing the impact of artifacts with EEG temporal-spatial features. Temporal Convolutional Networks (TCNs) have shown great potential for decoding temporal features [15]. Ingolfsson et al. [16] combined EEGNet and TCN architectures, achieving improved extraction of temporal-spatial features and classification performance. Zhao et al. [17] proposed a model that integrates CNNs with a novel attention mechanism within the Transformer architecture to improve the modeling of temporal-spatial feature dependencies in MI-EEG signals.

Furthermore, attention is increasingly being recognized as a factor in MI-EEG decoding. The attention mechanism allows networks to discern which parts of the data are significant and which are not, facilitating a form of perceptual and attentional behavior similar to that of humans. Liu et al. [18] introduced network that combined temporal-spatial attention blocks to enhance subject-specific MI-EEG classification. Sun et al. [19] used squeeze-and-excitation (SE) attention blocks with CNNs to improve the robustness of image feature extraction. Amin et al. [20] introduced an attention-inception CNN combined with LSTM, resulting in a model with fewer parameters and lower processing time for MI-EEG decoding. Similarly, some researchers have explored the combination of attention mechanisms and temporal convolutional networks. Zhang et al. [21] proposed AMSTCNet, which incorporates SE and Efficient Channel Attention (ECA) blocks. This architecture is designed to capture deep temporal information from the signal and dynamically fuse features at various scales. Altaheri et al. [22] utilized multi-head self-attention (MSA) in conjunction with TCN to highlight the most important information in EEG time series signals. Additionally, Liang et al. [23] proposed a model consisting of an inception block, MSA, temporal TCN, and layer fusion to improve the interpretability of the EEG decoding framework.

In this paper, aiming to enhance the model’s capability for extracting temporal-spatial features, thereby improving robustness, applicability, and generalization while advancing real-world applications of BCIs, a novel method named BANet is proposed. This method utilizes a bridge structure and multiple attention mechanisms for decoding MI-EEG signals. By integrating the Convolutional ECA block, Bridge block, and Inception-based temporal convolutional network (TCN) block, BANet effectively captures both local and global temporal and spatial features in the signal, thereby enhancing the representation of temporal-spatial features. This approach overcomes the limitations of CNNs and LSTMs in capturing global information and comprehensively understanding complex features, while also addressing the inadequate local feature extraction typically observed in Transformer structures. Experimental results indicate that the proposed network achieves outstanding decoding performance and enhances the model’s real-time deployment capabilities, highlighting its practical application potential in BCI technology. To facilitate researchers in improving the network proposed in this paper, our network settings can be found in our GitHub repository: https://github.com/woaixueximax/BANet.

The remainder of this paper is structured as follows: Section 2 introduces the proposed model and data preprocessing methods; Section 3 details the experimental setup and analyzes the results; and Section 4 provides a summary of this work.

Dataset description and preprocessing

To evaluate the decoding performance of BANet, the input to the network consists of the BCI Competition IV datasets 2a [28] and 2b [29], which involve different channels and classification tasks.

Experimental details and performance metrics

Three types of experiments are conducted, within-subject, short-train, and cross-subject, to evaluate the performance of the proposed network. The within-subject experiment is designed to evaluate the model’s decoding performance. In this experiment, the model is trained using the first session of the BCI-2a dataset and the first three sessions of the BCI-2b dataset, with testing performed on the last session of BCI-2a and the last two sessions of BCI-2b, this experiment allows for the

Conclusion

This paper presents BANet, a multi-attention network with an integrated bridge block designed for decoding MI-EEG signals. The architecture of BANet consists of three designed blocks and two attention modules, which are capable of emphasizing significant temporal and spatial features. The bridge block facilitates a comprehensive extraction of features, enabling the capture of both local and global EEG signal features. Experiments have been designed to evaluate BANet’s performance across

CRediT authorship contribution statement

Xuejian Wu: Writing – review & editing, Writing – original draft, Visualization, Resources, Methodology, Data curation, Conceptualization. Yaqi Chu: Visualization, Supervision, Methodology, Funding acquisition. Yang Luo: Validation, Supervision, Investigation. Yiwen Zhao: Supervision, Investigation, Formal analysis. Xingang Zhao: Supervision, Resources, Methodology.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.