Motion Acceleration Estimation Of Human Knee Joint Using Mechanomyogram Signal

As the number of disabled elderly and physically handicapped people increases,there is a growing demand for physical function training and movement aids equipment.Domestic and foreign scholars and institutions have researched on wearable power assistance robots to provide movement assistance and rehabilitation training for the above-mentioned people,improve their self-care ability and reduce the burden of their family,in order to adapt to the rapid development of social aging and the growing number of physically disabled people.With the development of human-computer interaction technology,wearable power assistance robots are gradually changing from passively accepting instructions from users to actively identifying and understanding human motion intentions.More and more wearable power assistance robots use human biological signals to recognize the motion intentions.However,there are still some bottlenecks in how to continuously and reliably recognize the human motion intentions,resulting in need for the further improvements in the flexibility and adaptability of robot motion control.In this thesis,the human knee joint is taken for the research object,and the Mechanomyography(MMG)signal of the thigh muscles are used to estimate the Linear Acceleration(ACC)of the knee joint motion.The signal processing,feature extraction and ACC estimation problems have been researched in this process.Combining the latest domestic and foreign research,a regression model using the convolutional neural network-long short-term memory(CNN-LSTM)neural network based on the multi-channel MMG signals has been construc ted.The model is used to estimate the ACC value of the knee joints motion for the movement control of the wearable power assistance robots for the lower limbs,which provides new ideas for their practical application.The main research contents and innovations of this thesis are as follows.1.In order to remove motion artifacts from MMG signals,in this thesis a MMG signal processing method based on Multivariate Variational Mode Decomposition(MVMD)has been proposed.The multi-channel MMG signals of the same muscles have been decomposed by the bandwidth-limited intrinsic mode function(BIMF),and the BIMF components have been selected by combining the frequency and energy distribution of the MMG signal to finally form the MMG signal filtered from motion artifacts and other noise.Compared with the Multivariate Empirical Mode Decomposition(MEMD)algorithm,the Instantaneo us Frequency(IF)of the BIMF components of each channel decomposed by the MVMD algorithm is more stable,and the degree of mutual overlap of the components is significantly reduced.The total power of the filtered MMG signals have been increased by 50%,and the noise such as motion artifacts can be filtered out while the energy of MMG signals can be maximally retained,which can improve recognizability of the MMG signal.2.To solve the problem of MMG data feature extraction,in this thesis an automatic extraction features method based on Convolutional Neural Network(CNN)has been proposed and the CNN architecture for the estimating of ACC has been designed.A sliding window approach has been used to transform multi-channel MMG timing data into a sequence input of two-dimensional array to the CNN model,which has been used to maximize the mining of MMG data features.Compared with the Back Propagation(BP)neural network model based on the time-frequency domain features of MMG data,the CNN model has higher accuracy in estimating ACC values,and the average correlation coefficient(R)between the estimated and expected values is above 91%.Compared with the time-frequency domain features of MMG data,the performance of automatically extracted features based on CNN is better,which can enhance the generalization ability of the recognition algorithm.3.To solve the problem of the knee joint motion ACC estimation,in this thesis a method using CNN-LSTM network model based on the MMG data ha s been proposed.The multi-channel MMG data has been input to the CNN model to automatically extract the features,and then the features have been input to the LSTM network model in time series for the ACC estimating.Compared with the LSTM network model constructed from the time-frequency domain features of MMG data,the CNN-LSTM network model has higher accuracy and better generalization ability for the ACC estimation,with R-mean of over 97%.This approach can facilitate the application of the ACC value of the knee joint for the movement control of the lower-limb wearable power assistance robots in the engineering practice.