Motor Imagery EEG Recognition Based on Scheduled Empirical Mode Decomposition and Adaptive Denoising Autoencoders

Artificial definition is mainly used as the pretraining in existing research on motor imagery (MT), which costs time on adjusting parameters. Deep-learning method like convolutional neural networks (CNNs) used in motor imagery classification would cause the problem of local convergence because of its random initialization strategy. To address these limitations, we proposed a method by reconstructing MI signal by using empirical mode decomposition (EMD) and apply the reconstructed signal in a denoising auto encoder (DAE) trained by an adaptive learning rate strategy. In this model, DAE obtains its noise level adaptively according to the annealing principle and chooses an adaptive learning rate to solve the local minimums in non-convex network. Both offline and online experiments are carried out to evaluate its performance. The corresponding results show that the proposed method can achieve higher recognition accuracy and better convergence speed, compared with other mainstream methods