基于混合自编码器的脑力负荷识别

Recognition of Mental Workload Based on Hybrid Autoencoders

为了消除多通道近红外光谱信号中存在的冗余信息并提取抽象特征,构建了基于混合自编码器的脑力负荷识别模型。首先,将原始信号送入栈式自编码器中进行通道降维;然后使用卷积自编码器对降维后的信号进行无监督学习从而提取抽象特征,并将特征依次送入支持向量机、K最近邻、随机森林这3种基分类器中进行建模;最后,用软、硬投票的集成策略来提高模型对脑力负荷识别的准确性。实验结果表明,混合自编码器具有良好的通道降维和提取抽象特征的能力,该模型在脑力负荷三分类任务中的准确率可以达到95.12%,相对于同类研究准确率有明显提升。

Mental workload reflects people’s capability of processing information when performing a task,and it can be employed as an indicator of the brain effort. Recently, the assessment on mental workload has been widely studied in various tasks such as simulated flight, cognitive, and so on. The traditional methods of evaluating mental workload include subjective scale method, task performance method and physiological signal parameters method. Near-infrared spectroscopy(NIRS) have many advantages over other physiological techniques as it has better spatial resolution than EEG and better temporal resolution than fMRI. Besides, it is portable and lightweight with simple data acquisition and less exposed to electrical artifacts. In this paper, NIRS are selected to build an assessment model on mental workload. As well known, deep learning with convolutional neural networks has revolutionized signal processing through end-to-end learning due to its efficiency and convenience. In order to eliminate redundant information and extract features from multi-channel NIRS, a novel recognition model on mental workload is created based on the hybrid autoencoders. Firstly, the original signals are sent to the stack autoencoder for channel dimensionality reduction, and these processed signals are fed to the convolutional autoencoder to extract the abstract features. Then, we employ three base classifiers, i.e., the support vector machine(SVM), K nearest neighbors(KNN), random forest(RF), for building models. Finally, the integration strategy of soft voting and hard voting is applied to improve the assessment accuracy for mental workload. It is shown from the results that proper compressing signal channels can help improve the recognition accuracy of the recognition model. The best accuracy of this proposed model for classifying three levels of mental workload can reach 95.12%, which has significant improvement compared to similar studies.