Parameter optimization of 3D convolutional neural network for dry-EEG motor imagery brain-machine interface.

Easing the behavioral restrictions of those in need of care not only improves their own quality of life (QoL) but also reduces the burden on care workers and may help reduce the number of care workers in countries with declining birthrates. The brain-machine interface (BMI), in which appliances and machines are controlled only by brain activity, can be used in nursing care settings to alleviate behavioral restrictions and reduce stress for those in need of care. It is also expected to reduce the workload of care workers. In this study, we focused on motor imagery (MI) classification by deep-learning to construct a system that can identify MI obtained by electroencephalography (EEG) measurements with high accuracy and a low latency response. By completing the system on the edge, the privacy of personal MI data can be ensured, and the system is ubiquitous, which improves user convenience. On the other hand, however, the edge is limited by hardware resources, and the implementation of models with a huge number of parameters and high computational cost, such as deep-learning, on the edge is challenging. Therefore, by optimizing the MI measurement conditions and various parameters of the deep-learning model, we attempted to reduce the power consumption and improve the response latency of the system by minimizing the computational cost while maintaining high classification accuracy. In addition, we investigated the use of a 3-dimension convolutional neural network (3D CNN), which can retain spatial locality as a feature to further improve the classification accuracy. We propose a method to maintain a high classification accuracy while enabling processing on the edge by optimizing the size and number of kernels and the layer structure. Furthermore, to develop a practical BMI system, we introduced dry electrodes, which are more comfortable for daily use, and optimized the number of parameters and memory consumption size of the proposed model to maintain classification accuracy even with fewer electrodes, less recall time, and a lower sampling rate. Compared to EEGNet, the proposed 3D CNN reduces the number of parameters, the number of multiply-accumulates, and memory footprint by approximately 75.9%, 16.3%, and 12.5%, respectively, while maintaining the same level of classification accuracy with the conditions of eight electrodes, 3.5 seconds sample window size, and 125 Hz sampling rate in 4-class dry-EEG MI.