28 October 2021 in IEEE Transactions on Systems, Man, and Cybernetics: Systems
Brain–computer interface (BCI) based on motor imagery electroencephalogram (EEG) has been widely used in various applications. Despite the previous efforts, the remained major challenges are effective feature extraction and the time-consuming calibration procedure. To address these issues, a novel multiattention adaptation network integrating the multiple attention mechanism and transfer learning is proposed to classify the EEG signals. First, the multiattention layer is introduced to automatically capture the dominant brain regions relevant to mental tasks without incorporating any prior knowledge about the physiology. Then, a multiattention convolutional neural network is employed to extract deep representation from raw EEG signals. Especially, a domain discriminator is applied to deep representation to reduce the differences between sessions for target subjects. The extensive experiments are conducted on three public EEG datasets (Dataset IIa and IIb of BCI Competition IV, and High Gamma dataset), achieving the competitive performance with average classification accuracy of 81.48%, 82.54%, and 93.97%, respectively. All the results outperform the state-of-the-art algorithms demonstrate the effectiveness and robustness of the proposed method. Importantly, we confirm that it is easier and more appropriate to transfer the information from local brain regions than from the whole brain. This enhances the transfer ability of deep features and, hence, it improves the performance of BCI systems.
4 October 2022 in IEEE Transactions on Neural Systems and Rehabilitation Engineering
Electroencephalography-based Brain Computer Interfaces (BCIs) invariably have a degenerate performance due to the considerable individual variability. To address this problem, we develop a novel domain adaptation method with optimal transport and frequency mixup for cross-subject transfer learning in motor imagery BCIs. Specifically, the preprocessed EEG signals from source and target domain are mapped into latent space with an embedding module, where the representation distributions and label distributions across domains have a large discrepancy. We assume that there exists a non-linear coupling matrix between both domains, which can be utilized to estimate the distance of joint distributions for different domains. Depending on the optimal transport, the Wasserstein distance between source and target domains is minimized, yielding the alignment of joint distributions. Moreover, a new mixup strategy is also introduced to generalize the model, where the inputs trials are mixed in frequency domain rather than in raw space. The extensive experiments on three evaluation benchmarks are conducted to validate the proposed framework. All the results demonstrate that our method achieves a superior performance than previous state-of-the-art domain adaptation approaches.
13 April 2022 in IEEE Transactions on Cognitive and Developmental Systems
The applications of myoelectrical interfaces are majorly limited by the efficacy of decoding motion intent in electromyographic (EMG) signal. Currently, EMG classification methods often rely substantially on hand-crafted features, or ignore key channel and inter-feature information for classification tasks. To address these issues, a multi-scale feature extraction network (MSFEnet) based on channel-spatial attention is proposed to decode EMG signal for the task of gesture recognition classification. Specifically, we fuse the spatio-temporal characteristics of EMG signal with different scales. Then, we construct the feature channel attention module and the feature spatial attention module to capture more key channels features, and more key spatial features. To evaluate the efficacy of proposed method, extensive experiments are conducted on two public datasets: Ninapro DB2 and CapgMyo DB-a. An average accuracy of 86.21%, 90.77%, 92.53% and 98.85% has been achieved in Exercise B, C, D of Ninapro DB2 and CapgMyo DB-a, respectively. The experimental results demonstrate that MSFEnet is more capable of extracting temporal and spatial fused features. It performs well in generalization and has higher classification accuracy compared with the state-of-the-art methods.
19 October 2022 in IEEE Transactions on Instrumentation and Measurement
As a worldwide disease, major depressive disorder (MDD) severely damages patients’ mental health. It is of great significance of detecting MDD accurately in providing necessary guidance for physicians. Here, a novel complex network-based graph convolutional network (CN-GCN), is developed to detect MDD. First, multichannel electroencephalogram (EEG) signals are decomposed into several frequency bands. Then, a multilayer brain network is constructed via a phase-locking value (PLV), where each layer corresponds to a specific frequency band. Aiming at accurately identifying brain states, the CN-GCN is developed, with multilayer brain network as input. Moreover, power spectral density (PSD) is applied for refining node-level rhythm features. Such structure of CN-GCN allows learning the node features based on the topology connections of the brain network. The proposed framework shows the state-of-the-art (SOTA) detection accuracy of 99.29% on a public MDD dataset. Our work confirms the validity on integrating complex network and GCN in multichannel EEG signal analysis and contributes to identifying complex brain states better.
20 February 2023 in IEEE Transactions on Neural Networks and Learning Systems
A robust decoding model that can efficiently deal with the subject and period variation is urgently needed to apply the brain–computer interface (BCI) system. The performance of most electroencephalogram (EEG) decoding models depends on the characteristics of specific subjects and periods, which require calibration and training with annotated data prior to application. However, this situation will become unacceptable as it would be difficult for subjects to collect data for an extended period, especially in the rehabilitation process of disability based on motor imagery (MI). To address this issue, we propose an unsupervised domain adaptation framework called iterative self-training multisubject domain adaptation (ISMDA) that focuses on the offline MI task. First, the feature extractor is purposefully designed to map the EEG to a latent space of discriminative representations. Second, the attention module based on dynamic transfer matches the source domain and target domain samples with a higher coincidence degree in latent space. Then, an independent classifier oriented to the target domain is employed in the first stage of the iterative training process to cluster the samples of the target domain through similarity. Finally, a pseudolabel algorithm based on certainty and confidence is employed in the second stage of the iterative training process to adequately calibrate the error between prediction and empirical probabilities. To evaluate the effectiveness of the model, extensive testing has been performed on three publicly available MI datasets, the BCI IV IIa, the High gamma dataset, and Kwon et al. datasets. The proposed method achieved 69.51%, 82.38%, and 90.98% cross-subject classification accuracy on the three datasets, which outperforms the current state-of-the-art offline algorithms. Meanwhile, all results demonstrated that the proposed method could address the main challenges of the offline MI paradigm.
29 May 2023 in IEEE Transactions on Instrumentation and Measurement
Gas-liquid two-phase flow is pervasive in industrial processes, and accurate measurement of its parameters can effectively improve industrial efficiency. Void fraction, a key parameter in gas-liquid two-phase flow, plays a crucial role in mixture density determination and flow structure analysis. In this paper, a novel multi-task learning method is proposed to predict the gas void fraction. Firstly, we conduct vertical upward gas-liquid two-phase flow experiments to obtain fluid signals relying on the four-sector distributed conductance sensor. Then, we pair the measurement of gas void fraction and its classification and design a Co-Attention based Cross-Stitch Network (CACSNet) to process these two closely related tasks simultaneously. In CACSNet, two parallel embedding branches are adopted to extract task-specific features. There are three feature extractors in each branch to capture in-depth representations of temporal and spatial. To learn an optimal combination of these task-specific features and generate shared representations, we propose the Cross-stitch Embedding with Attention Modules (CEAM), which is incorporated the Channel Co-Attention Module (CCAM) and the Temporal Co-Attention Module (TCAM). Finally, we compare our CACSNet with single-task baseline and other competitive methods, which proves that our framework achieves superior performance in gas void fraction prediction. We also conduct detailed model ablation and parameter analysis to illustrate the efficiency of our proposed structure.