Parallel Optimization Of Blind Source Separation Algorithm Based On Multi-core Platform

At present,with the continuous development of devices collecting brain wave signal,brain computer interface research has become more practical,and some brain controlled wheelchairs and other equipment based on brain computer interface have been put into the market.However,in the process of EEG signal collection,it is inevitable to dope with noise signals,which come from inherent noise of mechanical equipment or interference signals of surrounding environment.Blind source separation algorithm can extract the original signal from the hybrid signal without knowing the original signal.Although the separation effect of these algorithms has been proved to be very ideal,the speed of these algorithms is difficult to meet the real-time requirements in today's brain computer interface system.The main research of this paper is based on the multi-core hardware platform to optimize the current mainstream blind source separation algorithm by utilizing parallel computing.The optimized algorithms are aimed at improving the efficiency of each algorithm and laying the foundation for the futher use of the algorithm in the real-time system.Besides,the algorithms do not change the original separation principle and separat ion effect of each algorithm.In this paper,three mainstream blind source separation algorithms in the field of EEG processing are selected for parallel optimization,which are SOBI algorithm,InfoMax algorithm and FastICA algorithm.According to each different algorithm,the mathematical calculation principle is analyzed,and the multi-core processor is adapted from the aspects of algorithm principle and implementation by combining the high-performance computing programming technology.Parallel optimization measures such as task parallel,thread level parallel,data and instruction level parallel are adopted to speed up the calculation process of the algorithm.In addition,according to the characteristics of NUMA architecture of multi-core hardware platform,the memory access optimization of algorithms is carried out in runtime.In order to facilitate researchers to use the parallel blind source separation algorithms optimized in this paper,a parallel ICA tool is designed and implemented.The tool supports three kinds of blind source separation parallel algorithms,and provides a variety of programming language interfaces.What users need to do is to specify the input data and the running algorithm.And then the parallel ICA tool can automatically adapt to the current running hardware platform in an optimal way.Through the performance test and analysis of this paper,the parallel optimized SOBI algorithm,infomax algorithm and FastICA algorithm have achieved 39 times,3.2 times and 4.9 times speedup respectively on Intel's multi-core platform.