Action Recognition Of Waist Surface Emg Signals Based On Stacking Ensemble Learning

Aiming at the problem that a single model can not accurately extract the data features of different actions,which leads to difficulties in recognition,a waist surface EMG signal recognition model based on stacking integrated learning is proposed.Firstly,collect surface electromyographic signals of the lumbar spine muscles under six different movements.Secondly,preprocess the raw signals containing noise,extract matrix counting features within the threshold range,and finally design a stacking integrated recognition algorithm for lumbar motion verification.The research content of this article is mainly reflected in the following three aspects:(1)In the face of problems such as difficulty in collecting EMG signals from the human waist and insufficient data sets,experiments were conducted to collect EMG signals from the waist.Firstly,the isometric contraction experiment was conducted to reduce the signal differences among different volunteers;Secondly,six basic movements were designed based on waist movements: bending forward,bending right,bending left,bending right,twisting left,and twisting right.The EMG signals of twelve volunteers under these six movements were collected;Then,the EMG signal is homogenized and segmented based on short-term energy;Finally,the LMS filtering algorithm is used to denoise the signal.(2)To address the limitations of information redundancy in extracting time-frequency features of EMG signals,an improved energy kernel feature extraction algorithm-matrix counting based on threshold range-is proposed to improve the efficiency of EMG signal classification in the field.Experiments have been carried out in terms of operation time and classification accuracy.Through comparison,it has been proved that the threshold matrix counting method for feature extraction is effective and can improve both operational efficiency and classification accuracy.(3)A waist EMG signal model based on Stacking ensemble learning is proposed.This model introduces the Stacking ensemble learning framework into EMG signal recognition,using KNN,RF,XGboost,Light GBM algorithms as base learners,and decision tree models as meta learners to construct an integrated model.The training set is divided using a 50 fold cross validation method to reduce the over fitting problem in the repeated training process of the integrated model.According to the experiment,the average accuracy rate of the model for six waist movements is 94.98%,which is about 6% higher than the single model.The recognition speed can be within 150 ms,meeting the requirements of real-time recognition.Finally,the host computer communication page is designed to analyze the signal quality and signal classification when patients collect EMG signals.