Brain-computer Interface Competition Research Articles

Multiclass classification of motor imagery tasks based on multi-branch convolutional neural network and temporal convolutional network model.

Motor imagery (MI) is a cognitive process wherein an individual mentally rehearses a specific movement without physically executing it. Recently, MI-based brain-computer interface (BCI) has attracted widespread attention. However, accurate decoding of MI and understanding of neural mechanisms still face huge challenges. These seriously hinder the clinical application and development of BCI systems based on MI. Thus, it is very necessary to develop new methods to decode MI tasks. In this work, we propose a multi-branch convolutional neural network (MBCNN) with a temporal convolutional network (TCN), an end-to-end deep learning framework to decode multi-class MI tasks. We first used MBCNN to capture the MI electroencephalography signals information on temporal and spectral domains through different convolutional kernels. Then, we introduce TCN to extract more discriminative features. The within-subject cross-session strategy is used to validate the classification performance on the dataset of BCI Competition IV-2a. The results showed that we achieved 75.08% average accuracy for 4-class MI task classification, outperforming several state-of-the-art approaches. The proposed MBCNN-TCN-Net framework successfully captures discriminative features and decodes MI tasks effectively, improving the performance of MI-BCIs. Our findings could provide significant potential for improving the clinical application and development of MI-based BCI systems.

A time segment adaptive optimization method based on separability criterion and correlation analysis for motor imagery BCIs

Motor imagery (MI) plays a crucial role in brain-computer interface (BCI), and the classification of MI tasks using electroencephalogram (EEG) is currently under extensive investigation. During MI classification, individual differences among subjects in terms of response and time latency need to be considered. Optimizing the time segment for different subjects can enhance subsequent classification performance. In view of the individual differences of subjects in motor imagery tasks, this article proposes a Time Segment Adaptive Optimization method based on Separability criterion and Correlation analysis (TSAOSC). The fundamental principle of this method involves applying the separability criterion to various sizes of time windows within the training data, identifying the optimal raw reference signal, and adaptively adjusting the time segment position for each trial’s data by analyzing its relationship with the optimal reference signal. We evaluated our method on three BCI competition datasets, respectively. The utilization of the TSAOSC method in the experiments resulted in an enhancement of 4.90% in average classification accuracy compared to its absence. Additionally, building upon the TSAOSC approach, this study proposes a Nonlinear-TSAOSC method (N-TSAOSC) for analyzing EEG signals with nonlinearity, which shows improvements in the classification accuracy of certain subjects. The results of the experiments demonstrate that the proposed method is an effective time segment optimization method, and it can be integrated into other algorithms to further improve their accuracy.

Riemannian transfer learning based on log-Euclidean metric for EEG classification.

Brain computer interfaces (BCI), which establish a direct interaction between the brain and the external device bypassing peripheral nerves, is one of the hot research areas. How to effectively convert brain intentions into instructions for controlling external devices in real-time remains a key issue that needs to be addressed in brain computer interfaces. The Riemannian geometry-based methods have achieved competitive results in decoding EEG signals. However, current Riemannian classifiers tend to overlook changes in data distribution, resulting in degenerated classification performance in cross-session and/or cross subject scenarios. This paper proposes a brain signal decoding method based on Riemannian transfer learning, fully considering the drift of the data distribution. Two Riemannian transfer learning methods based log-Euclidean metric are developed, such that historical data (source domain) can be used to aid the training of the Riemannian decoder for the current task, or data from other subjects can be used to boost the training of the decoder for the target subject. The proposed methods were verified on BCI competition III, IIIa, and IV 2a datasets. Compared with the baseline that without transfer learning, the proposed algorithm demonstrates superior classification performance. In contrast to the Riemann transfer learning method based on the affine invariant Riemannian metric, the proposed method obtained comparable classification performance, but is much more computationally efficient. With the help of proposed transfer learning method, the Riemannian classifier obtained competitive performance to existing methods in the literature. More importantly, the transfer learning process is unsupervised and time-efficient, possessing potential for online learning scenarios.

Optimization of transfer learning based on source sample selection in Euclidean space for P300-based brain-computer interfaces.

Event-related potentials (ERPs), such as P300, are widely utilized for non-invasive monitoring of brain activity in brain-computer interfaces (BCIs) via electroencephalogram (EEG). However, the non-stationary nature of EEG signals and different data distributions among subjects create significant challenges for implementing real-time P300-based BCIs. This requires time-consuming calibration and a large number of training samples. To address these challenges, this study proposes a transfer learning-based approach that uses a convolutional neural network for high-level feature extraction, followed by Euclidean space data alignment to ensure similar distributions of extracted features. Furthermore, a source selection technique based on the Euclidean distance metric was applied to measure the distance between each source feature sample and a reference point from the target domain. The samples with the lowest distance were then chosen to increase the similarity between source and target datasets. Finally, the transferred features are applied to a discriminative restricted Boltzmann machine classifier for P300 detection. The proposed method was evaluated on the state-of-the-art BCI Competition III dataset II and rapid serial visual presentation dataset. The results demonstrate that the proposed technique achieves an average accuracy of 97% for both online and offline after 15 repetitions, which is comparable to the state-of-the-art methods. Notably, the proposed approach requires <½ of the training samples needed by previous studies. Therefore, this technique offers an efficient solution for developing ERP-based BCIs with robust performance against reduced a number of training data.

Recruiting neural field theory for data augmentation in a motor imagery brain-computer interface.

We introduce a novel approach to training data augmentation in brain-computer interfaces (BCIs) using neural field theory (NFT) applied to EEG data from motor imagery tasks. BCIs often suffer from limited accuracy due to a limited amount of training data. To address this, we leveraged a corticothalamic NFT model to generate artificial EEG time series as supplemental training data. We employed the BCI competition IV '2a' dataset to evaluate this augmentation technique. For each individual, we fitted the model to common spatial patterns of each motor imagery class, jittered the fitted parameters, and generated time series for data augmentation. Our method led to significant accuracy improvements of over 2% in classifying the "total power" feature, but not in the case of the "Higuchi fractal dimension" feature. This suggests that the fit NFT model may more favorably represent one feature than the other. These findings pave the way for further exploration of NFT-based data augmentation, highlighting the benefits of biophysically accurate artificial data.

A Cascade xDAWN EEGNet Structure for Unified Visual-Evoked Related Potential Detection.

Visual-based brain-computer interface (BCI) enables people to communicate with others by spelling words from the brain and helps professionals recognize targets in large numbers of images. P300 signals evoked by different types of stimuli, such as words or images, may vary significantly in terms of both amplitude and latency. A unified approach is required to detect variable P300 signals, which facilitates BCI applications, as well as deepens the understanding of the P300 generation mechanism. In this study, our proposed approach involves a cascade network structure that combines xDAWN and classical EEGNet techniques. This network is designed to classify target and non-target stimuli in both P300 speller and rapid serial visual presentation (RSVP) paradigms. The proposed approach is capable of recognizing more symbols with fewer repetitions (up to 5 rounds) compared to other models while possessing a better information transfer rate (ITR) as demonstrated on Dataset II (17.22 bits/min in the second repetition round) of BCI Competition III. Additionally, our approach has the highest unweighted average recall (UAR) performance for both 5 Hz ( 0.8134±0.0259 ) and 20 Hz ( 0.6527±0.0321 ) RSVP. The results show that the cascade network structure has better performance between both the P300 Speller and RSVP paradigms, manifesting that such a cascade structure is robust enough for dealing with P300-related signals (source code is available at https://github.com/embneural/Cascade-xDAWN-EEGNet-for-ERP-Detection).

Ensemble-based feature engineering mechanism to decode imagined speech from brain signals

Speech impairments, resulting from brain injuries, mental disorders, or vocal abuse, substantially affect an individual’s quality of life and can lead to social isolation. Brain–Computer Interfaces (BCIs), particularly those based on EEG, offer a promising support mechanism by harnessing brain signals. Owing to their clinical efficacy, cost-effective EEG devices, and expanding applications in the medical and social spheres, their usage has surged. This study introduces an ensemble-based feature engineering mechanism to pinpoint the optimal brain rhythm, channel subset, and feature set for accurately predicting imagined words from EEG signals via machine learning models. Leveraging the 2020 International BCI competition dataset, we employed bandpass filtering, channel wrapping, and ranking methods were applied to discern suitable brain rhythms and features associated with imagined speech. Subsequent application of kernel-based principal component analysis enabled us to compress the feature space dimensionality. We then trained various machine learning models, among which the kNN model excelled, achieving an average accuracy of 73% in a 10-fold cross-validation scheme ,surpassing 18% higher than the existing literature. The Gamma rhythm was identified as the most predictive of imagined speech from EEG brain signals. These advancements herald a new era of more precise and effective BCIs, poised to significantly improve the lives of those with speech impairments.

M-FANet: Multi-Feature Attention Convolutional Neural Network for Motor Imagery Decoding.

Motor imagery (MI) decoding methods are pivotal in advancing rehabilitation and motor control research. Effective extraction of spectral-spatial-temporal features is crucial for MI decoding from limited and low signal-to-noise ratio electroencephalogram (EEG) signal samples based on brain-computer interface (BCI). In this paper, we propose a lightweight Multi-Feature Attention Neural Network (M-FANet) for feature extraction and selection of multi-feature data. M-FANet employs several unique attention modules to eliminate redundant information in the frequency domain, enhance local spatial feature extraction and calibrate feature maps. We introduce a training method called Regularized Dropout (R-Drop) to address training-inference inconsistency caused by dropout and improve the model's generalization capability. We conduct extensive experiments on the BCI Competition IV 2a (BCIC-IV-2a) dataset and the 2019 World robot conference contest-BCI Robot Contest MI (WBCIC-MI) dataset. M-FANet achieves superior performance compared to state-of-the-art MI decoding methods, with 79.28% 4-class classification accuracy (kappa: 0.7259) on the BCIC-IV-2a dataset and 77.86% 3-class classification accuracy (kappa: 0.6650) on the WBCIC-MI dataset. The application of multi-feature attention modules and R-Drop in our lightweight model significantly enhances its performance, validated through comprehensive ablation experiments and visualizations.

Multiscale Spatial-Temporal Feature Fusion Neural Network for Motor Imagery Brain-Computer Interfaces.

Motor imagery, one of the main brain-computer interface (BCI) paradigms, has been extensively utilized in numerous BCI applications, such as the interaction between disabled people and external devices. Precise decoding, one of the most significant aspects of realizing efficient and stable interaction, has received a great deal of intensive research. However, the current decoding methods based on deep learning are still dominated by single-scale serial convolution, which leads to insufficient extraction of abundant information from motor imagery signals. To overcome such challenges, we propose a new end-to-end convolutional neural network based on multiscale spatial-temporal feature fusion (MSTFNet) for EEG classification of motor imagery. The architecture of MSTFNet consists of four distinct modules: feature enhancement module, multiscale temporal feature extraction module, spatial feature extraction module and feature fusion module, with the latter being further divided into the depthwise separable convolution block and efficient channel attention block. Moreover, we implement a straightforward yet potent data augmentation strategy to bolster the performance of MSTFNet significantly. To validate the performance of MSTFNet, we conduct cross-session experiments and leave-one-subject-out experiments. The cross-session experiment is conducted across two public datasets and one laboratory dataset. On the public datasets of BCI Competition IV 2a and BCI Competition IV 2b, MSTFNet achieves classification accuracies of 83.62% and 89.26%, respectively. On the laboratory dataset, MSTFNet achieves 86.68% classification accuracy. Besides, the leave-one-subject-out experiment is performed on the BCI Competition IV 2a dataset, and MSTFNet achieves 66.31% classification accuracy. These experimental results outperform several state-of-the-art methodologies, indicate the proposed MSTFNet's robust capability in decoding EEG signals associated with motor imagery.

Improved motor imagery classification using adaptive spatial filters based on particle swarm optimization algorithm.

As a typical self-paced brain-computer interface (BCI) system, the motor imagery (MI) BCI has been widely applied in fields such as robot control, stroke rehabilitation, and assistance for patients with stroke or spinal cord injury. Many studies have focused on the traditional spatial filters obtained through the common spatial pattern (CSP) method. However, the CSP method can only obtain fixed spatial filters for specific input signals. In addition, the CSP method only focuses on the variance difference of two types of electroencephalogram (EEG) signals, so the decoding ability of EEG signals is limited. To make up for these deficiencies, this study introduces a novel spatial filter-solving paradigm named adaptive spatial pattern (ASP), which aims to minimize the energy intra-class matrix and maximize the inter-class matrix of MI-EEG after spatial filtering. The filter bank adaptive and common spatial pattern (FBACSP), our proposed method for MI-EEG decoding, amalgamates ASP spatial filters with CSP features across multiple frequency bands. Through a dual-stage feature selection strategy, it employs the Particle Swarm Optimization algorithm for spatial filter optimization, surpassing traditional CSP approaches in MI classification. To streamline feature sets and enhance recognition efficiency, it first prunes CSP features in each frequency band using mutual information, followed by merging these with ASP features. Comparative experiments are conducted on two public datasets (2a and 2b) from BCI competition IV, which show the outstanding average recognition accuracy of FBACSP. The classification accuracy of the proposed method has reached 74.61 and 81.19% on datasets 2a and 2b, respectively. Compared with the baseline algorithm, filter bank common spatial pattern (FBCSP), the proposed algorithm improves by 11.44 and 7.11% on two datasets, respectively (p < 0.05). It is demonstrated that FBACSP has a strong ability to decode MI-EEG. In addition, the analysis based on mutual information, t-SNE, and Shapley values further proves that ASP features have excellent decoding ability for MI-EEG signals and explains the improvement of classification performance by the introduction of ASP features. These findings may provide useful information to optimize EEG-based BCI systems and further improve the performance of non-invasive BCI.

A multiscale feature fusion network based on attention mechanism for motor imagery EEG decoding

The decoding of motor imagery (MI) electroencephalogram (EEG) is an essential component of the brain–computer interface (BCI), which can help patients with motor impairment communicate directly with the outside world through assistive devices. The key to motor imagery electroencephalogram (MI-EEG) classification is to extract multiple temporal, spatial, and spectral features, to obtain more comprehensive and representative information. However, current deep learning methods must fully consider the depth of temporal features and multi-spectral knowledge in EEG and often ignore the temporal or spectrum dependence in MI-EEG. In addition, the lack of effective feature fusion methods can lead to information redundancy, which affects decoding performance. To solve the above problems, this paper proposes a novel MI-EEG decoding method, named multi-scale feature fusion network based on attention mechanism (MSFF-SENet). Firstly, the multi-scale spatio-temporal module (MS-STM) and the multi-scale temporal module (MSTM) extract spatial and high-dimensional temporal features from the original signal. Then, the power spectral density estimation (PSD) convolution module (PSD-Conv module) acquires the multi-spectral features of the MI-EEG signal. Secondly, the feature fusion module fuses spatio-temporal and multi-spectral features to generate integrated feature mappings and establish the dependencies between different features. Finally, we conducted a visual analysis of the results, explaining the neural activity patterns of various motor imagery tasks in different frequency ranges and revealing the potential relationship between body movement and changes related to brain activity. The experimental results show that the classification accuracy of this model in BCI Competition IV 2a (BCI IV 2a) and High Gamma (HGD) datasets is 85.37% and 96.60%, respectively, which is superior to the most advanced methods.

One-Dimensional Convolutional Multi-branch Fusion Network for EEG-Based Motor Imagery Classification

The Brain-Computer Interface (BCI) system based on motor imagery (MI) is a hot research topic nowadays, which can control external devices through the brain and has a wide range of applications in rehabilitation, gaming, and entertainment. Due to the non-smooth, non-linear, and low signal-to-noise ratio of the MI EEG signal, it is challenging to accurately decode the MI task intention. A new end-to-end deep learning method is proposed to decode raw MI EEG signals without preprocessing, such as filtering and feature reinforcement. The 1D convolution is used to learn the time-frequency features in MI signals, and a four-branch fusion network is used as the main body to add a 1D CNN-AE block and 1D SE-block to enhance the algorithm's performance. Experiments on two publicly available datasets demonstrate that our proposed algorithm outperforms the current state-of-the-art methods. It achieves 86.11% and 89.51% on the BCI Competition IV-2a and the BCI Competition IV-2b datasets, respectively, and a 6.9% improvement in the generalizability test. The proposed data enhancement method can effectively alleviate the overfitting of the algorithm and improve the decoding performance. Further analysis shows that 1D convolution can effectively extract the features associated with the MI task.

ETCNet: An EEG-based motor imagery classification model combining efficient channel attention and temporal convolutional network

Brain-computer interface (BCI) enables the control of external devices using signals from the brain, offering immense potential in assisting individuals with neuromuscular disabilities. Among the different paradigms of BCI systems, the motor imagery (MI) based electroencephalogram (EEG) signal is widely recognized as exceptionally promising. Deep learning (DL) has found extensive applications in the processing of MI signals, wherein convolutional neural networks (CNN) have demonstrated superior performance compared to conventional machine learning (ML) approaches. Nevertheless, challenges related to subject independence and subject dependence persist, while the inherent low signal-to-noise ratio of EEG signals remains a critical aspect that demands attention. Accurately deciphering intentions from EEG signals continues to present a formidable challenge. This paper introduces an advanced end-to-end network that effectively combines the efficient channel attention (ECA) and temporal convolutional network (TCN) components for the classification of motor imagination signals. We incorporated an ECA module prior to feature extraction in order to enhance the extraction of channel-specific features. A compact convolutional network model uses for feature extraction in the middle part. Finally, the time characteristic information is obtained by using TCN. The results show that our network is a lightweight network that is characterized by few parameters and fast speed. Our network achieves an average accuracy of 80.71% on the BCI Competition IV-2a dataset.

Dynamic Convolution With Multilevel Attention for EEG-Based Motor Imagery Decoding

Brain-computer interface (BCI) is an innovative technology that utilizes artificial intelligence (AI) and wearable electroencephalography (EEG) sensors to decode brain signals and enhance the quality of life. EEG-based motor imagery (MI) brain signal is used in many BCI applications including smart healthcare, smart homes, and robotics control. However, the restricted ability to decode brain signals is a major factor preventing BCI technology from expanding significantly. In this study, we introduce a dynamic attention temporal convolutional network (D-ATCNet) for decoding EEG-based motor imagery signals. The D-ATCNet model uses dynamic convolution and multilevel attention to enhance the performance of MI classification with a relatively small number of parameters. D-ATCNet has two main blocks: dynamic and temporal convolution. Dynamic convolution uses multilevel attention to encode low-level MI-EEG information and temporal convolution uses shifted window with self-attention to extract high-level temporal information from the encoded signal. The proposed model performs better than the existing methods with an accuracy of 71.3% for subject-independent and 87.08% for subject-dependent using the BCI Competition IV-2a dataset.

IMH-Net: a convolutional neural network for end-to-end EEG motor imagery classification

As the main component of Brain-computer interface (BCI) technology, the classification algorithm based on EEG has developed rapidly. The previous algorithms were often based on subject-dependent settings, resulting in BCI needing to be calibrated for new users. In this work, we propose IMH-Net, an end-to-end subject-independent model. The model first uses Inception blocks extracts the frequency domain features of the data, then further compresses the feature vectors to extract the spatial domain features, and finally learns the global information and classification through Multi-Head Attention mechanism. On the OpenBMI dataset, IMH-Net obtained 73.90 ± 13.10% accuracy and 73.09 ± 14.99% F1-score in subject-independent manner, which improved the accuracy by 1.96% compared with the comparison model. On the BCI competition IV dataset 2a, this model also achieved the highest accuracy and F1-score in subject-dependent manner. The IMH-Net model we proposed can improve the accuracy of subject-independent Motor Imagery (MI), and the robustness of the algorithm is high, which has strong practical value in the field of BCI.

EEG-FMCNN: A fusion multi-branch 1D convolutional neural network for EEG-based motor imagery classification.

Motor imagery (MI) electroencephalogram (EEG) signal is recognized as a promising paradigm for brain-computer interface (BCI) systems and has been extensively employed in various BCI applications, including assisting disabled individuals, controlling devices and environments, and enhancing human capabilities. The high-performance decoding capability of MI-EEG signals is a key issue that impacts the development of the industry. However, decoding MI-EEG signals is challenging due to the low signal-to-noise ratio and inter-subject variability. In response to the aforementioned core problems, this paper proposes a novel end-to-end network, a fusion multi-branch 1D convolutional neural network (EEG-FMCNN), to decode MI-EEG signals without pre-processing. The utilization of multi-branch 1D convolution not only exhibits a certain level of noise tolerance but also addresses the issue of inter-subject variability to some extent. This is attributed to the ability of multi-branch architectures to capture information from different frequency bands, enabling the establishment of optimal convolutional scales and depths. Furthermore, we incorporate 1D squeeze-and-excitation (SE) blocks and shortcut connections at appropriate locations to further enhance the generalization and robustness of the network. In the BCI Competition IV-2a dataset, our proposed model has obtained good experimental results, achieving accuracies of 78.82% and 68.41% for subject-dependent and subject-independent modes, respectively. In addition, extensive ablative experiments and fine-tuning experiments were conducted, resulting in a notable 7% improvement in the average performance of the network, which holds significant implications for the generalization and application of the network.

A novel framework for classification of two-class motor imagery EEG signals using logistic regression classification algorithm.

Robotics and artificial intelligence have played a significant role in developing assistive technologies for people with motor disabilities. Brain-Computer Interface (BCI) is a communication system that allows humans to communicate with their environment by detecting and quantifying control signals produced from different modalities and translating them into voluntary commands for actuating an external device. For that purpose, classification the brain signals with a very high accuracy and minimization of the errors is of profound importance to the researchers. So in this study, a novel framework has been proposed to classify the binary-class electroencephalogram (EEG) data. The proposed framework is tested on BCI Competition IV dataset 1 and BCI Competition III dataset 4a. Artifact removal from EEG data is done through preprocessing, followed by feature extraction for recognizing discriminative information in the recorded brain signals. Signal preprocessing involves the application of independent component analysis (ICA) on raw EEG data, accompanied by the employment of common spatial pattern (CSP) and log-variance for extracting useful features. Six different classification algorithms, namely support vector machine, linear discriminant analysis, k-nearest neighbor, naïve Bayes, decision trees, and logistic regression, have been compared to classify the EEG data accurately. The proposed framework achieved the best classification accuracies with logistic regression classifier for both datasets. Average classification accuracy of 90.42% has been attained on BCI Competition IV dataset 1 for seven different subjects, while for BCI Competition III dataset 4a, an average accuracy of 95.42% has been attained on five subjects. This indicates that the model can be used in real time BCI systems and provide extra-ordinary results for 2-class Motor Imagery (MI) signals classification applications and with some modifications this framework can also be made compatible for multi-class classification in the future.

Bayesian Opportunities for Brain–Computer Interfaces: Enhancement of the Existing Classification Algorithms and Out-of-Domain Detection

Bayesian neural networks (BNNs) are effective tools for a variety of tasks that allow for the estimation of the uncertainty of the model. As BNNs use prior constraints on parameters, they are better regularized and less prone to overfitting, which is a serious issue for brain–computer interfaces (BCIs), where typically only small training datasets are available. Here, we tested, on the BCI Competition IV 2a motor imagery dataset, if the performance of the widely used, effective neural network classifiers EEGNet and Shallow ConvNet can be improved by turning them into BNNs. Accuracy indeed was higher, at least for a BNN based on Shallow ConvNet with two of three tested prior distributions. We also assessed if BNN-based uncertainty estimation could be used as a tool for out-of-domain (OOD) data detection. The OOD detection worked well only in certain participants; however, we expect that further development of the method may make it work sufficiently well for practical applications.

Classification algorithm for EEG-based motor imagery using hybrid neural network with spatio-temporal convolution and multi-head attention mechanism.

Motor imagery (MI) is a brain-computer interface (BCI) technique in which specific brain regions are activated when people imagine their limbs (or muscles) moving, even without actual movement. The technology converts electroencephalogram (EEG) signals generated by the brain into computer-readable commands by measuring neural activity. Classification of motor imagery is one of the tasks in BCI. Researchers have done a lot of work on motor imagery classification, and the existing literature has relatively mature decoding methods for two-class motor tasks. However, as the categories of EEG-based motor imagery tasks increase, further exploration is needed for decoding research on four-class motor imagery tasks. In this study, we designed a hybrid neural network that combines spatiotemporal convolution and attention mechanisms. Specifically, the data is first processed by spatiotemporal convolution to extract features and then processed by a Multi-branch Convolution block. Finally, the processed data is input into the encoder layer of the Transformer for a self-attention calculation to obtain the classification results. Our approach was tested on the well-known MI datasets BCI Competition IV 2a and 2b, and the results show that the 2a dataset has a global average classification accuracy of 83.3% and a kappa value of 0.78. Experimental results show that the proposed method outperforms most of the existing methods.

Transformer-based ensemble deep learning model for EEG-based emotion recognition

Emotion recognition is one of the most important research directions in the field of brain–computer interface (BCI). However, to conduct electroencephalogram (EEG)-based emotion recognition, there exist difficulties regarding EEG signal processing; moreover, the performance of classification models in this regard is restricted. To counter these issues, the 2022 World Robot Contest successfully held an affective BCI competition, thus promoting the innovation of EEG-based emotion recognition. In this paper, we propose the Transformer-based ensemble (TBEM) deep learning model. TBEM comprises two models: a pure convolutional neural network (CNN) model and a cascaded CNN-Transformer hybrid model. The proposed model won the abovementioned affective BCI competition’s final championship in the 2022 World Robot Contest, demonstrating the effectiveness of the proposed TBEM deep learning model for EEG-based emotion recognition.