Abstract:

The application of deep learning methods to fNIRS(functional Near-Infrared Spectroscopy) has become a research hotspot in the field of brain computer interface, but less available data limits the performance of deep learning models. A method for generating fNIRS original signal is proposed based on DEVAE-GAN(Dual-Encoder-Variational Autoencoder-Generative Adversarial Network) in this study. In this method, the pre-processed fNIRS signals are converted into time and space representations, input into a dual encoder to extract time and space information, splice two pieces of information, and send to the decoder to generate samples. In order to verify its effectiveness, experiments are conducted on public data sets of mental load tasks, and different numbers of generated samples are extended to the training data set, and the enhanced data set is used to train the deep neural network. Compared with multiple baseline generation models, the proposed method generates the highest sample quality, and the average classification accuracy of all subjects after using this method is 95.86%, which is increased by 0.91% when compared with the original data set. The experimental results show that the proposed method can effectively learn the distribution of raw data of mental load task fNIRS, generate high-quality samples, and improve the performance of deep learning models.

Key words: fNIRS, DEVAE-GAN, data generation, mental load task, deep learning, generative models, brain computer interface, deep neural network