Neural Signal Processing For Invasive Brain-Machine Interface Based On Rats

Technology of Brain-Computer Interface aims to find and construct a novel information interaction approach not depending on peripheral nervous system. Brain-Computer Interface comes to a new research and developing climax as the revolutionary of chronic implantable electrodes array, and the neural signal processing methods and neural decoding algorithms has been a new center of the Technology of Brain-Computer Interface. Research presented in this thesis provide novel neural signal processing approaches for action potential and local field potential based Brain-Computer Interface system.Neuronal action potential detection methods are presented, time-domain energy operator, matched filter and improved matched filter, their performances are compared with each other and traditional threshold crossing approach. The result shows that the improved matched filter has highest accuracy and stability. And neuronal action potential sorting methods are also implemented and evaluated, including template matching, clustering and feature reduction method, where we find that the template matching has highest sorting accuracy.Afterwards, the correlation between rat's pre-limb movement and activities of neuronal action potentials is evaluated. Four neuronal firing patterns are found and we find that neurons with similar firing pattern usually locate closely. More, we implement three neural decoding algorithms, Optimal Linear Estimation, Kalman Filter and Probabilistic Neural Network, on three rats that electrode array are implanted. Results show that improved Probabilistic Neural Network performs better than other neural decoders, which has highest correlation coefficient and least mean square error with actual pre-limb movement.In addition, the correlation between Local Field Potential (LFP) and rat's pre-limb movement is also analyzed. The correlation is easily found and obvious in the results and more than 50% features (three bands:0-10Hz,10-45Hz and 55-200Hz) that extracted from LFP has high (larger than 0.3) correlation coefficient with rat's pre-limb movement. More, these features are used to decode rat's pre-limb movement and the best result is found by Optimal Linear Estimation, whose correlation coefficient with its pre-limb movement is 0.8889, which is better than the decoding result based on neuronal action potentials.In conclusion, this thesis presents neural signal processing and neural decoding methods to find the correlation between neural signals and rat's pre-limb movement. Results show that both neuronal action potentials and local field potential are available for neural decoding with accuracy higher than 0.85.