Stroke is the second leading cause of death and disability [1], [2]. It commonly occurs in the elderly and is prone to hemiparesis, affecting patients' quality of life and burdening on their families [3]. Clinical studies have shown that the Brain-Computer Interface (BCI) system can help stroke patients perform active rehabilitation of motor functions, thus obtaining better rehabilitation results than traditional treatments [4]. The electroencephalogram (EEG) is one of the most important tools for observing brain activity. The BCI system can decode human intentions in the EEG and achieve direct control of external devices (prosthesis, wheelchair, etc.) without relying on the superficial nerve and muscle tissue. BCI systems can be divided into implantable BCI and non-implantable BCI. Implantable BCI is not affected by artifacts such as electrooculography (EOG), electromyography (EMG), and electrocardiography (ECG) and can obtain high-resolution EEG signals. However, there is a crucial problem many respondents will not accept the placement of EEG devices in the brain [5]. Non-implantable BCIs are non-invasive, with electrodes placed along the scalp to acquire weak electrical signals generated by neural activity in the brain. It has the advantages of ease of use, low cost, and low risk. Still, it has disadvantages, such as poor spatial resolution and susceptibility to external interference, which can be significantly reduced with technological advances. Therefore, non-implantable BCIs are more popular among researchers and patients.

Motor imagery (MI) EEG signals are brain activity recorded when subjects imagine or perform movements of the left hand, right hand, mouth, foot, etc. The physiological basis for MI task classification is the event-related desynchronization/synchronization (ERD/ERS) phenomenon in the brain's motor cortex when subjects perform motor imagery [6]. MI signals are easily accessible and do not depend on external stimuli and, therefore, are often used as paradigms for BCI systems.

Accurate decoding of MI task intent in EEG signals is the key to the wide application of BCI. Many researchers use machine learning methods for MI classification, such as typical spatial pattern (CSP) to extract spatial feature vectors associated with motion intent and support vector machine (SVM) to classify these feature vectors to achieve decoding motion intent in EEG signals. Machine learning methods require complex preprocessing to decode MI signals. The extracted features often need to be selected manually, which requires researchers to have particular expertise, significantly limiting the application of machine learning in BCI systems. The emergence of deep learning methods has solved the above problems to a certain extent.

Deep learning methods for decoding MI EEG signals can be divided into two categories. The first category: requires preprocessing the raw EEG signal and feature reinforcement. For example, Mahamune et al. [7] used the CSP method to preprocess MI EEG signals. The Continuous Wavelet Transform (CWT) method to extract time-frequency features and transform 1D temporal EEG signals into 2D time-frequency maps, and Convolutional Neural Networks (CNN) to classify these time-frequency maps. Wu et al. [8] used the band-pass filtered EEG data as the training data for linear SVM to generate a 2×500 feature matrix and CNN to classify them, achieving better results than the traditional SVM methods. The second category: only simple preprocessing (filtering, etc.) or no processing is done on the original EEG signal, and feature extraction and MI intent classification are done using deep learning methods, also known as end-to-end deep learning methods. For example, Dose et al. [9] proposed an end-to-end convolutional neural network with raw EEG signals as input, 30×1 convolutional kernels to extract features on the time domain, and 1×NEEG convolutional kernels to extract features on the channels. Zhang et al. [10] proposed an EEG-inception end-to-end neural network with six EEG-inception blocks to learn various features and subtly build Residual-blocks to solve the gradient disappearance problem caused by too many layers and improve the robustness and classification performance of the network.

End-to-end neural networks do not have to perform tedious feature extraction and extract the overall optimal features. It is only necessary to design the network structure and input the raw EEG signal. The robustness and generalization of the network are generally better than machine learning methods. Future BCI systems will also heavily adopt end-to-end deep learning methods. The small amount of publicly available MI data results in networks with few convolutional layers, typically no more than four. Most deep learning methods only target decoding two-class MI tasks and fewer study four-class tasks. The classification accuracy of end-to-end deep learning methods that decode raw EEG signals could be much higher.

To address the above problems, we propose a 1D convolutional multi-branch fusion network for an end-to-end EEG classification algorithm. The novelty of the method lies in four aspects. First, a new end-to-end EEG classification algorithm is proposed to decode the raw EEG signals of four-class MI tasks with better classification accuracy than the state-of-the-art machine learning and deep learning methods. Second, the number of convolutional layers in the network reaches eight, much larger than that of general EEG classification algorithms. Third, the data enhancement method proposed in this paper can effectively improve the network's performance, and the method is simple and easy to implement. Fourth, the experimental results show that 1D convolution is suitable for extracting features in the original EEG signal.