Research On Spike Real-time Processing In Invasive Brain-Machine Interface

Invasive Brain-Machine Interface (BMI) obtains the relationships between actual movements and neural signals which are recorded from cerebral cortex by signal processing technology, and provides users with communication and control channels to communicate with the outside world directly. The spike detection and sorting, neural decoding algorithm are very important and challenging issues for the invasive BMI. In current invasive BMI, it is hard to achieve real-time detecting and sorting of multichannel spikes in parallel. The huge information brought from multichannel signals obstructs the real-time processing of neural decoding algorithm, which restricts the development of invasive BMI. Therefore, it is of the theoretical significance and practical value to carry on an in-depth study as how to complete the real-time processing of invasive BMI’s spikes in parallel.This thesis, based on comprehensive survey of the existent spike-based BMI system, proposes spike-based BMI signals’real-time processing algorithms. The beneficial and explorative researches are done on the following:A wavelet analysis based algorithm for spike detection is proposed, and the lifting scheme of wavelet is adopted for reducing the computational burden due to the complexity of convolution in traditional wavelet transform. The proposed algorithm that utilizes the wavelet analysis for improving the signal’s SNR and detects the spikes by the thresholding method, is presented. The real-time detecting of multichannel spikes is realized by the hardware-based method, and the multichannel spikes’detecting is done by the parallel modules.A probabilistic neural network based algorithm for spike sorting is presented. The network can be constructed expediently by loading the train data, and then spikes are sorted by this trained network. The hardware-based method with the architecture of parallel computation is proposed according to the parallelism of probabilistic neural network. The Look-up table and COordinate Rotation DIgital Computer (CORDIC) mixed method which improves the accuracy and lower the resource utilization, is proposed for computing the activation function in the neural network.A hardware-based method for real-time calculation of neural decoding algorithms is proposed, which includes Kalman filter and general regression neural network. In the implementation of Kalman filter, a novel processing element is proposed for multi function calculation through different configurations, the array that is composed of processing elements could speed up the calculation of matrix. The implementation of general regression neural network could be completed by reusing the architecture of probabilistic neural network, which speeds up the calculation and finishes real-time computation.Since the real-time neural decoding cannot be completed due to the huge information brought by multichannel signals, a hardware-based invasive BMI system is built with a lifting wavelet based algorithm for spike detection, a probabilistic neural network based algorithm for spike sorting, and hardware-based implementations of neural decoding algorithms. The prototype of neural signal real-time processing with the library of neural signal processing algorithm is constructed.