国家专用集成电路设计工程技术研究中心知识库

随着可穿戴设备的普及和移动互联网的发展，生理信号的采集和获取变得更为便捷。生理信号分类是根据人体物理现象判断用户状态或意图的研究，该研 究在健康监测、情绪判断和人机交互领域都有广泛的应用前景，因此开展该研究具有重要的应用价值。 生理信号由多模态信号组成，其时间、空间和频率特征复杂，并且存在着信噪比较低和个体差异性大等特点，这对生理信号模式分类带来了挑战。本文利用 生理信号开展了睡眠分期、情绪识别和人机交互等三项研究工作。 本文的主要研究内容和贡献如下：

1. 基于 Transformer 编码器的生理信号睡眠分期 针对睡眠分期任务中生理信号存在长时依赖关系的问题，本文提出了端到 端自动睡眠分期模型 SleepZzNet，该模型包含多个卷积神经网络和 Transformer 编码器，分别用于提取波形特征和长时序依赖的睡眠转换特征。在 SleepEDF 数 据集上，分类精度达到了 86.10%，超过了当前最先进模型方法。此外，我们招募 受试者采集其睡眠时的生理信号数据，实验证明我们的模型有良好的泛化能力。

2. 基于时空注意力机制的脑电信号情绪识别 针对情绪识别任务中生理信号存在时间、空间、频率特征复杂的问题，我们 提出 VitEmoNet。该模型使用时空注意力机制，用于提取频率、空间、时间多维 度的特征。Transformer 编码器模型用于特征融合和模式分类。在 SEED 数据集 上精度可达到 94.26%，与其它先进模型持平。

3. 基于眼动信号的人机交互系统 本章提出了一种基于眼动信号的人机交互系统，该系统从多个方位上采集 用户的眼动信号，使用分类器判断用户的视线方向类别，得到对应的控制策略， 对与人机交互设备连接的输入设备进行移动操作。该系统在实现简单有效的人 机交互的同时，可减少软件和硬件成本。

With the popularity of wearable devices and the development of mobile Internet, the collection and acquisition of physiological signals have become more convenient. Physiological signal classification is the study of judging the user's state or intention based on the physical phenomena of the human body. The study has a wide range of application prospects in the fields of health monitoring, emotional judgment and human-computer interaction, so it has important application value to carry out this research.

Physiological signals consist of multimodal signals with complex temporal, spatial and frequency characteristics, and are characterized by low signal-to-noise ratio and high individual variability, which pose a challenge for physiological signal pattern classification. This paper discusses each of the three tasks in using physiological signals for sleep staging, emotion recognition and human-computer interaction.

The main research and contributions of this paper are as follows.

Sleep staging of physiological signals based on Transformer encoder

​      To address the problem of long-time dependence of physiological signals in sleep staging tasks, this paper proposes SleepZzNet, an end-to-end automatic sleep staging model, which contains multiple convolutional neural networks and Transformer encoders for extracting waveform features and long time-dependent sleep transition features, respectively. A classification accuracy of 86.10\% was achieved on the SleepEDF dataset, surpassing the current state-of-the-art modeling methods. In addition, we recruited subjects to collect their physiological signal data during sleep and experimentally demonstrated that our model has good generalization ability.

EEG signal emotion recognition based on spatio-temporal attention mechanism

​      For the problem of complex temporal, spatial, and frequency features of physiological signals in emotion recognition tasks, we propose VitEmoNet. the model uses a spatio-temporal attention mechanism for extracting features in multiple dimensions of frequency, space, and time. the Transformer encoder model is used for feature fusion and pattern classification. The accuracy can reach 94.26\% on the SEED dataset, which is on par with other advanced models.

Human-computer interaction system based on eye-movement signals

​      This chapter proposes an HCI system based on oculogram signals, which collects the user's oculogram signals from multiple orientations, uses a classifier to determine the user's directional category of vision, and obtains a corresponding control strategy for moving the input device connected to the HCI device. This system can reduce software and hardware costs while achieving simple and effective human-computer interaction.