Removal Of Electromyogram Artifacts In A Small Number Of Channel EEG Signals

This work studies the pre-treatment of electroencephalography(EEG)in epilepsy seizure detection,which is part of biomedical signal processing.It is chanllenging since EEG is acquired by scalp electrodes,and thereby it is very weak and accompanied by interference from various artifacts.The study focused on the removal of the electromyogram(EMG)artifacts from a small number of channel EEG Signals.The main contents are as follows:Firstly,two blind source separation algorithms,independent component Analysis(ICA)and canonical correlation analysis(CCA)were introduced.The mathematical model,constraints and pretreatment process of removing EMG artifacts from EEG by the ICA and CCA were preliminarily studied,and their improved algorithm were used as the comparison algorithms for the subsequent algorithms.In addition,the evaluation method of algorithm performance was introduced.Secondly,aiming at the limitation that the traditional algorithm can't completely separate EMG artifacts with a single discrimination degree,an independent vector analysis(IVA)method is proposed to separate EMG artifacts in few channels of EEG.The IVA algorithm uses both second-order statistics and higher-order statistics to analyze the non-gaussian property and weak correlation of the EMG artifacts,which has the advantages of both ICA and CCA.The removal process,performance,advantages and disadvantages of IVA-L,IVAG and IVA-GL were developed,evaluated and compared.It is concluded that the removal of EMG artifacts by the IVA-GL algorithm could achieve better separation effect.Further,this work proposes two improvements of the IVA-GL algorithm based on its two obvious limitations,which needs a large amount of iterative computation and its iteration operation with fixed step size is easy to cause large steady-state error.(1)The optimization method combined with the principal component analysis(PCA)is used to effectively reduce the computation amount of the algorithm.For instance,to process real-time EEG data,if the data is acquired every 10 seconds,the algorithm needs to complete the processing within 10 seconds at most,which is a heavy task.If each fragment of each electrode is processed,the task is even heavier,and the fragments that not affected by the artifacts will always produce errors with the source signal after algorithm reconstruction.Due to the high energy of EEG signal fragments affected by the EMG artifact,only the high-energy sections after PCA analysis were processed,which can effectively reduce the task while protecting the collected EEG signals.(2)A step length optimization is used to reduce the separation error and improve the efficiency of the separation system.Usually,the separation effect of the algorithm needs to be verified by complex signal processing,the iterative algorithm of adaptive updating of step size is realized by means of the fastest descent method,and the nonlinear function based on joint information entropy is introduced to ensure that the improved algorithm can maintain low steady-state error and fast convergence speed.Finally,the experimental data,semi-simulated data and database data were used to evaluate,compare and analyze the performance of above algorithms.It is found that the IVAGL combined with PCA algorithm after step size optimization achieves a SNR of 55?100,a relative RMS error of 0.1?0.3,and an average EMG artifact separation rate of 98%,which compared with fixed step IVA-GL algorithm time saving 32%.While the traditional algorithm(ICA,CCA)achieves a SNR of 20?70,a relative RMS error of 0.3?0.6,and an average EMG artifact separation rate of 82?95%.And from the graph analysis and comparison of traditional algorithms(ICA,CCA),the IVA-GL combined with the PCA algorithm after the step size optimization has a relatively good artifact separation processing effect.