Research On Multi-information Fusion Motion Recognition For Lower Limb Exoskeleton Robots

As an effective auxiliary or rehabilitation device,lower limb exoskeleton robots have been widely used in the field of walking assistance and rehabilitation.In order to meet the requirements of wearing comfort,man-machine coordination control must be carried out.The premise of coordinated control is to quickly and accurately identify the motion state of the wearer’s lower limbs.However,the real-time and accuracy of lower limb motion recognition is also one of the difficulties in research in this field.The topic uses multi-point flexible plantar pressure and inertial measurement unit sensors,combined with wireless transmission,to build a wearable multi-source information collection system.On this basis,the Relief F-PCA multi-source feature fusion algorithm and machine learning algorithm are proposed,and the lower limb motion recognition model is obtained,and the recognition of the lower limb motion model and speed is realized by using the constructed model.First of all,for the traditional signal acquisition system based on motion and mechanics,there are limitations such as a small number of pressure sensors,large acquisition equipment,and poor wearing comfort.This topic designs a portable 45-point pressure sensor acquisition system,which forms a multi-source information acquisition system with the IMU system developed in the laboratory.Using WIFI communication and designing host computer software to realize synchronous start of multi-sensors,high-speed data reception and storage.Secondly,data acquisition,signal denoising,feature extraction and information fusion are carried out on the multi-source information movement mode.Determine the installation position of the sensor,collect information on 9 common movement patterns,analyze multi-source information noise and perform wavelet noise reduction on plantar pressure.Use the pressure cloud map to formulate the selection rules of the pressure area,and optimize the 36 pressure areas of a single foot through the rules;and optimize the 4 channels of the IMU.18 kinds of features are extracted from the optimal signal to obtain 1512-dimensional feature data,and multi-source information feature fusion based on Relief F-PCA is adopted.Then,nine movement patterns of human lower limbs were identified and analyzed.Using KNN,BP neural network,and SVM to establish a nonlinear mapping model between feature vectors and motion states,optimize model parameters,and use 5-fold cross-validation to evaluate the generalization ability of the model.The experimental results show that feature extraction is performed using a sliding window with a size of10 sampling points(82.2ms)and a step size of 3 sampling points(24.7ms),and 139-dimensional high-weight features are selected through Relief F features,PCA feature fusion,and the constructed The FA-optimized SVM model has the best classification performance,the pattern recognition accuracy rate reaches 97.5%,and the entire movement pattern recognition time is 177.2～242.0ms.Finally,the recognition and analysis of 15 speed patterns of human lower limbs.Collect 15 speed gait data sets,and establish a nonlinear mapping model between feature vectors and motion speeds through machine learning.The final experimental results show that a sliding window with a size of 18 sampling points(147.9ms)and a step size of 6 sampling points(49.3ms)is used for feature extraction,and 66-dimensional high-weight features are selected through the Relief F feature for PCA feature fusion.The constructed FA optimized SVM model has the best classification performance,the pattern recognition accuracy rate reaches 95.9%,and the entire movement pattern recognition time is 302.9～368.3ms.