| 中文摘要 | 脑-机接口技术（Brain-Computer Interfaces，BCI）作为一种人机交互技术，能够不依赖于肌肉、大脑外周神经以及骨骼等人体正常的输出通路，在大脑和外界之间传递信息。稳态视觉诱发电位（SSVEP）是指由中心视网膜上具有恒定频率的重复刺激引起的连续视觉皮层反应。基于稳态视觉诱发电位的脑机接口系统具有无需训练和较高信息传输率的优点，是脑机接口系统的一个重要研究方向。最近，深度学习在模式识别领域得到了有效应用，并且，在脑电分析中具有良好的表现，也体现了在SSVEP-BCI中的应用潜力。
在本文中，共设计了两种不同的 SSVEP 数据采集实验，坐姿静态 SSVEP 实验和站姿及步姿 SSVEP 实验，分别募集了23名和15名被试。针对解码SSVEP信号的大多数方法仅限于 CCA 和一些基于 CCA 的扩展方法的局限性，本文提出了一种基于卷积神经网络的模板比较网络，用于学习 EEG 信号与 SSVEP 各刺激频率相对应模板之间的关系，以实现检测不同的目标频率。该方法能够与先验知识和空间滤波器（任务相关成分分析，TRCA）相结合，进而增强 SSVEP 的检测能力。基于模板比对网络，本文共设计了三种不同的模板，进一步对模板比对网络的性能进行了分析和比较。
通过基于三种不同模板的模板比对网络与传统 CCA，TRCA 以及现有 CNN 方法的比较，结果发现模板比对网络，在静态和站姿步姿下的 SSVEP 数据中，相比于已有方法，解码精度显著提升，并且，基于任务相关成分分析模板的比对网络分类结果表现最优（静态，91.24%，站姿步姿，0 km/h: 88.55%, 2.5 km/h: 77.12% , 5 km/h: 67.93%），此外，在混有非频率诱发脑电的数据中，三种速度下，基于任务相关成分分析模板的比对网络也呈现了最高的正确率（0 km/h: 81.74%, 2.5 km/h: 67.26% , 5 km/h: 57.18%）。
综上所述，本文所提出的基于TRCA模板的比对网络，具有较高的SSVEP识别精度，证明了深度学习方法在基于EEG的SSVEP解码中的潜在应用，本研究有望为进一步研究BCI系统解码精度提升和应用提供技术支持。 |
| 英文摘要 | As a human-computer interaction technology, Brain-Computer Interfaces (BCI) transfers information between the brain and the outside world, which is independent of the normal output channels of the human body such as muscles, peripheral nerves, and bones. Steady-state visual evoked potential (SSVEP) refers to a continuous visual cortical response caused by repeated stimuli with a constant frequency on the central retina. The brain-computer interface system based on steady-state visual evoked potential requires no training and has high information translate rate, and it is an important research field of the brain-computer interface system. Recently, deep learning has been effectively applied in the field of pattern recognition, also has a good performance in EEG analysis. Thus, deep learning shows the potential to be applied in SSVEP-BCI.
In this study, two different SSVEP experiments were designed. sitting static SSVEP experiment and standing posture and stepping SSVEP experiment recruited 23 and 15 subjects, respectively. Nowdays, most methods for decoding SSVEP are limited to the standard CCA and some CCA-based extension methods. This study proposed a template comparison network based on convolutional neural network for learning the relationship between template corresponding to the stimulation frequency and EEG signals to improve the detection of different target frequencies.This new method can be combined with prior knowledge and spatial filters (Task Related Component Analysis, TRCA) to enhance the detection capabilities of SSVEP. Based on the template comparison network, three different templates were designed, and the performance of the template comparison network was further analyzed and compared.
By comparing the template comparison network based on three different templates with the traditional CCA, TRCA and existing CNN methods, the results showed that the template comparison network, in the static and stance step SSVEP data, compared to the comparison methods, the decoding accuracies were significantly improved. Moreover, the classification results indicated that the network based on the task-related component analysis template was optimal (static, 91.24%, standing step, 0 km/h: 88.55%, 2.5 km/h: 77.12%) , 5 km/h: 67.93%) and in the data mixed with non-frequency induced EEG, the highest accuracy rate at the three speeds (0 km/h: 81.74%, 2.5 km/h: 67.26%, 5 km/h: 57.18%).
In summary, The TRCA template-based comparison network proposed in this paper has high SSVEP recognition accuracy, which proves the potential application of deep learning method in EEG-based SSVEP decoding. Furthermore, it could improve the BCI performance and provide technical support for the application of the BCI system. |
修改评论