Research On De-noising Of Surface Electromyography And Evaluation Of Muscle Fatigue

The lower-limb rehabilitation robots plays a crucial role in helping humans restore and rebuild the motor function of their lower limbs.During rehabilitation,muscle fatigue is an unavoidable response of the relevant muscles due to continuous exercise.Moderate muscle fatigue can stimulate an increase in the level of motor function,whereas excessive muscle fatigue can lead to secondary damage to the lower limbs.Therefore,the evaluation of muscle fatigue is essential to ensure the effectiveness and safety of rehabilitation training.Surface electromyography(sEMG)signals provide a reliable basis for evaluating muscle fatigue because they are highly correlated with muscle function and limb movement patterns,and are completely non-invasive and non-destructive.However,sEMG signals are inevitably mixed with noise during the acquisition process,making them necessary to de-noise in practical applications.Thus,it is of great importance and clinical value to study the de-noising method of the sEMG signal,to extract effective features of muscle fatigue,and to evaluate the level of fatigue.This paper conducts research in these relevant aspects,presenting the main research findings as follows.(1)An optimized complementary ensemble empirical mode decomposition(OCEEMD)method is presented for de-noising sEMG signals.While the traditional complementary ensemble empirical mode decomposition(CEEMD)method has certain advantages in de-noising non-stationary sEMG signals in time-domain,it often suffers from the mode mixing of intrinsic mode functions(IMFs),leading to poor de-noising results.To address this issue,an OCEEMD-based method is presented to suppress the mode mixing of the IMFs and to achieve a good de-noising effect.The method involves several steps: first,the least-squares mutual information is embedded in the CEEMD process to detect the correlation of the IMFs.Then,a chaotic quantum particle swarm optimization algorithm is used to find the optimal solution of the amplitude of the added auxiliary white noise so that the signal is adaptively decomposed with the least amount of added noise.Additionally,a correction function is employed to further suppress the mode mixing of IMFs.Next,the signal-dominant IMFs are reconstructed to obtain the de-noised signal.Finally,the presented de-noising method is compared with the CEEMD and complete ensemble empirical mode decomposition with adaptive noise(CEEMDAN)methods,demonstrating superior de-noising performance.(2)A novel method is presented for de-noising sEMG signals based on the improved variational mode decomposition(IVMD).Traditionally,variational mode decomposition(VMD)has been used in frequency domain to de-noise signals.However,VMD-based methods often suffer from the mode mixing of IMFs and the problem that the signaldominant IMFs are not "pure" enough.To address these issues,an IVMD-based method for de-noising sEMG signals is presented that not only obtains IMFs without mode mixing but also further improves the de-noising ability through its improved threshold wavelet transform.The method involves several steps: first,a three-dimensional logisticsine coupled mapping function is constructed and combined with the fruit fly optimization algorithm to form a tree-dimensional logistic-sine chaotic fruit fly optimization algorithm(3D LSCFOA),in order to improve global optimization performance in threedimensional space.Then,the combination of the three key parameters in the VMD decomposition is optimized using the 3D LSCFOA to adaptively decompose IMFs without mode mixing at appropriate parameter combination settings.Next,the signaldominant IMFs are further de-noised by integrating muscle fatigue features and an improved threshold wavelet method.The "pure" IMFs are reconstructed to obtain the denoised sEMG signal.Finally,the presented method is compared with the local mean decomposition,singular spectrum analysis,VMD,and other relevant methods for processing both simulated and actual sEMG signals.The results show that IVMD has the decomposition performance of suppressing mode mixing and achieving a good de-noising effect.(3)Multi-scale envelope spectral entropy(MSESEn)is presented as a new feature of muscle fatigue.Currently,most muscle fatigue features are extracted from a single scale,resulting in an insignificant representation of muscle fatigue.To address this issue,a new feature of muscle fatigue called MSESEn is presented,which has high sensitivity and stability in characterizing and evaluating muscle fatigue processes.The feature takes into account not only the potentially complex potential-generating sequence patterns and correlation scales of the sEMG signal but also the rich envelope spectral information in the IMFs.Three experiments are performed,including MSESEn values in the range of [1,20] scales,the comparison of MSESEn performance of sEMG signals before and after de-noising,and the comparison of MSESEn with approximate entropy and sample entropy for muscle fatigue characterization.The experimental results verify the effectiveness and reliability of the MSESEn.(4)A method based on the adaptive grey wolf optimizer and extreme gradient boosting(AGWO-XGBoost)for evaluating muscle fatigue levels is presented.Muscle fatigue level evaluation is often converted into level classification.However,improperly selected muscle fatigue features or an XGBoost model with inappropriate parameters tend to result in poor classification accuracy.To address this issue,an evaluation method based on AGWO-XGBoost is presented to significantly improve the classification accuracy of muscle fatigue.First,some muscle fatigue features are extracted from the de-noised sEMG signals,and the redundant features are removed based on principal component analysis to provide a reliable sample set for the evaluation model.Then,an AGWO algorithm is proposed to optimize the parameters of XGBoost to form an AGWOXGBoost model.Finally,two types of experiments are conducted,including the comparison of the classification performance of single-feature input,multi-feature input without selection,and selected multi-feature input,as well as the classification comparison between the support vector machine,deep neural network model,and XGBoost models.The results show that the presented method has high classification accuracy and efficient operation in evaluating muscle fatigue levels.