基于维度注意力与多尺度卷积网络的脑电分类方法研究

EEG classification based on dimensional attention and multi-scale convolutional networks

摘要:

针对脑电信号的非平稳性、时变复杂和分类准确率较低的问题,以及传统机器学习方法在提取复杂特征方面的不足,提出了一种基于维度注意力机制的多尺度时空卷积神经网络分类模型(DIMS-CNN),旨在提高分类准确率,以更好地适用于实际应用场景。首先,对数据进行带通滤波和去伪迹,并进行了降采样和通道选择等预处理；随后,将经过处理的数据输入构建的时空卷积模型中,为了进一步增强网络的特征提取能力,加入了时序和通道的多维度注意力机制以及正则化技术；对于数据量不足的问题,采用了频带互换的方法进行数据增强,且提高了模型的泛化性能。分别在HGD数据集和实验室自采集数据集上进行验证,获得了90.97%和90.21%的平均分类准确率,通过相比较其他算法,可以该文方法在分类准确率上有显著提高。

Abstract:

Aiming at the problems of non-stationarity, time-varying complexity and low classification accuracy of EEG signals, as well as the shortcomings of traditional machine learning methods in extracting complex features, a multi-scale spatio-temporal convolutional neural network classification model (DMS-CNN) based on dimensional attention mechanism is proposed to improve classification accuracy and better apply to practical application scenarios. Firstly, the data is filtered by bandpass and de-artifacts, and pre-processed by downsampling and channel selection. Then, the processed data is input into the spatiotemporal convolution model constructed. In order to further enhance the feature extraction capability of the network, the multi-dimensional attention mechanism of timing and channel and regularization technology are added. For the problem of insufficient data, the method of frequency band exchange is used to enhance the data, and the generalization performance of the model is improved. The average classification accuracy of 90.97% and 90.21% was obtained on HGD data set and laboratory self-collected data set respectively. Compared with other algorithms, the classification accuracy of this method is significantly improved.