Research On The Decoding And Calibration Method Of Implantable Brain Machine Interface On Rhesus Monkey

Neural Interface can provide information communication between the nervous system and external devices,it is an emerging technology with great potential in restoring the sensory and motor functions of paralyzed patients.Among neural interface,the implantable brain-machine interface has been in the spotlight for years,but there are still many problems to be solved.In terms of neural signal coding and decoding,due to the non-stationary characteristics of neural signals,there may be statistical distribution differences in neural signals data from different periods of time,decoder trained on historical data cannot accurately decode current data.Therefore,the users of the implantable brain-machine interface need to repeatedly collect new samples for decoder recalibration,which brings heavy burden to them.This paper conducts research from two aspects that decoder recalibration and construction of highly robust decoder,aims to cope with the problem of distribution differences of neural signal data through efficient decoding and calibration methods and reduce the requirement for number of current samples for recalibration.Firstly,for the action potential signals,this paper proposed a PCA-based multi-source domain adaptation decoder calibration method,which used a very small number of target domain training samples.Through the domain alignment between the target domain and each source domain,each sub-classifier was constructed,and then the sub-classifiers were used to weight the integrated prediction results.The experimental results showed that this method can make full use of the motion mode information contained in the multi-source domain,and achieve better and more robust calibration performance than the single-source domain adaptation method,and can effectively reduce the user's calibration burden.Secondly,for the local field potential signals,this paper proposed a highly robust decoding method based on power spectral density and convolutional neural networks,which constructed power spectral density features through short-time Fourier transform,and combined with convolutional neural networks to further learn feature and classify.The experimental results showed that this method can achieve good and robust decoding performance,and can well adapt to the non-stationary characteristics of high frequency LFP signals.Thus,this method can provide a new idea for the application of implantable brainmachine interface decoding based on local field potential signals.Finally,for the local field potential signals,this paper proposed a calibration algorithm based on joint distribution adaptation.The algorithm can make full use of the label information of the target domain training samples to optimize the estimation of class conditional distribution probability density in domain adaptation,and optimize the selection of feature representation based on a thought of random walk.The experimental results showed that the algorithm can achieve good calibration performance under different tasks though using very few current calibration samples,thus can effectively reduce the user's calibration burden of the implantable brain-machine interface based on local field potential signals.