脑机接口技术通过建立人脑与外部设备之间的直接交流通道，为人类提供了一种非常规的信息沟通与控制手段，已被广泛地应用于医疗、军事、娱乐等多种领域。然而，由于控制环境越来越复杂以及脑电信号本身的特性，传统的机器学习方法已不能较完善地提取生物电信号的特征信息，因而基于深度学习的生物电分析方法应运而生。深度学习方法可以挖掘信号的深层特征，从而实现复杂环境下的人机交互。本文将深度学习方法应用于生物电信号分类任务中，并将该模型应用于一款基于脑机接口的游戏控制系统中，利用人体眼电、肌电和脑电信号组成的多模态生物电信号作为控制信号，该系统可极大地提升游戏使用体验感。本文的主要内容及创新点如下：

（1）针对眼动信号分类识别过程中，对眼电信号的时序特性利用不够充分的问题，本文设计了一种基于双向长短期记忆网络与自注意力机制的眼动分类模型。利用双向长短期记忆网络提取眼动信号的前向与后向的时序特征信息，并将其输出通过自注意力机制，可自动地寻求对输出结果影响较大的部分，并分配上较高的权重，从而对关键信息进行侧重学习，提高了该模型的特征提取与分类识别的能力。对其他分类模型：SVM、CSP、阈值法、KNN、CNN，本章所提的眼动分类模型在四个方向眼动信号的分类识别方面，取得了较好的表现，平均准确率最高可达93.29%，比其他的较优分类模型CNN，准确率提升了8.42%。

（2）针对注意力脑电信号分类识别过程中，存在的两个问题：二维数据表征形式忽略了电极之间的空间拓扑关系以及脑电信号数据本身信息挖掘地不够充分，本文设计了基于3DCNN-BiGRU和注意力机制的注意力脑电分类模型。利用3DCNN提取脑电信号的空间特征，利用BiGRU提取脑电信号的时间特征。在两模型之间添加注意力机制，增强对数据本身信息的挖掘，从而提升整个模型的分类识别效果。对比其他分类模型：SVM、CNN、EEGNet、LSTM、BiLSTM，平均准确率可达93.35%，比其他的较优分类模型BiLSTM，准确率提升了3.93%。

（3）搭建基于深度学习与生物电信号的游戏控制系统，该系统由数据采集端、数据处理服务器端以及基于Unity3D平台搭建的游戏控制客户端三部分组成。人体生物电信号由数据采集端采集之后，通过数据处理服务器端对信号进行分类识别、转换以及传输等操作，并最终将原来这些生物电信号转换为游戏人物的控制指令，控制整个游戏的进程。由实验可得：受试者对游戏人物动作控制的平均准确率可达90%以上，实时性也满足相应的要求，因此该游戏控制系统能够流畅地运行。

Brain-computer interface technology provides an unconventional means of information communication and control by establishing a direct communication channel between human brain and external devices, and has been widely used in medical, military, entertainment and other fields. However, due to the more and more complex control environment and the characteristics of EEG itself, traditional machine learning methods have been unable to extract the characteristic information of EEG signals. Therefore, brain-computer interface system based on deep learning came into being. Deep learning method can excavate the deep features of EEG signals, so as to realize human-computer interaction in complex environment, with higher stability and reliability. This paper combines the deep learning method with the game system, researches and designs a game control system based on attention mechanism and deep learning, and uses the multi-modal bioelectrical signal composed of human eye electromyography, electromyography and electroencephalogram as the control signal, which can greatly improve the experience of game use. The main contents and innovations of this paper are as follows:

(1)In order to solve the problem of insufficient utilization of temporal characteristics of ophthalmic signals in the classification and recognition process of eye movement signals, this paper designs an eye movement classification model based on bidirectional long short-term memory network and self-attention mechanism. The bidirectional long short-term memory network is used to extract the forward and backward temporal feature information of eye movement signals, and output it through the self-attention mechanism, which can automatically find the part that has a greater impact on the output result and assign a higher weight, so as to focus on learning the key information, and improve the feature extraction and classification recognition ability of the model. For other classification models, such as SVM, CSP, threshold method, KNN and CNN, the eye movement classification model proposed in this chapter has achieved a good performance in the classification and recognition of eye movement signals in four directions, with an average accuracy of up to 93.29%, which is 8.42% higher than that of other superior classification models CNN.

(2)Aiming at two problems in the classification and recognition process of attention EEG signals, namely, the spatial topological relationship between electrodes is ignored in the two-dimensional data representation form, and the information mining of EEG data itself is insufficient, this paper designs an attention EEG classification model based on 3DCNN-BiGRU and attention mechanism. The spatial features of EEG signals were extracted by 3DCNN and the temporal features were extracted by BiGRU. The attention mechanism is added between the two models to enhance the information mining of the data itself, so as to improve the classification and recognition effect of the whole model. Compared with other classification models such as SVM, CNN, EEGNet, LSTM and BiLSTM, the average accuracy can reach 93.35%, which is 3.93% higher than that of other better classification model BiLSTM.

(3)Build a game control system based on deep learning and bioelectrical signals, which consists of three parts: data acquisition terminal, data processing server and game control client based on Unity3D platform. After the signal is collected by the data acquisition terminal, the human bioelectrical signal is classified, recognized, converted and transmitted through the data processing server, and finally the original bioelectrical signal is converted into the control instructions of the game characters to control the whole process of the game. According to the experiment, the average accuracy of the subjects' action control of the game characters can reach more than 90%, and the real-time performance also meets the corresponding requirements, so the game control system can run smoothly.