基于EEG的Cbi-BPNN模型在线预测"坐-站-走"姿态变换研究OA

The purpose of this paper is to put forward an online prediction method for'sit-stand-walk'posture transfor-mation,which can provides theoretical support for better control of lower limb rehabilitation robot.Firstly,the elec-troencephalogram(EEG)signals were preprocessed online by filtering and the dynamic baseline correction method.Secondly,the multi-dimensional features were extracted online using the improved directional common spatial pat-tern(DCSP),correlation coefficient(CC)calculation,curve fitting(CF),Hilbert-Huang transform(HHT)and discrete wavelet transform(DWT).Finally,the sit-stand-walk posture transformation was predicted online by the model which is cascaded by two back-propagation neural network(Cbi-BPNN).In this paper,eight subjects were recruited to carry out the sit-stand-walk posture transformation EEG signals acquisition experiment and the verifica-tion experiment of the online prediction system.Eight feature combinations and five classification models for posture transformation prediction were also compared.The experimental results show that using the hybrid features of DC-SP-CC-CF-HHT-DWT as the inputs of the posture transformation prediction model can obtain the highest offline ac-curacy of 76.09%.The model can predict human posture transformation about 200 ms before the actual movement with the highest online accuracy of 65.00%.