| The intelligent powered knee prosthesis is a new type of prosthesis that combines bionic and electronic technology.It mimics the motion trajectory of the healthy human knee joint,and recognizes the locomotion intent of the human body through the real-time sensing data changes embedded in the prosthesis.However,there are still some problems in this field,such as sensor redundancy,high cost of prosthetic devices,complex structure of prediction models,low accuracy of human locomotion intent recognition,and inability to be applied to real-time prediction of multilocomotion scenes in life.In order to solve the above problems,this paper uses the intelligent powered knee prosthesis independently developed by our team,Bionic Knee VT 2.0,to build a multi-source sensing system,which only includes a six-axis IMU,an axial load cell and a knee joint angle encoder.The sensor data from 8 patients with above-knee amputation was collected for 9 daily locomotion modes,resulting in the successful construction of a steady-state locomotion database for above-knee amputation patients.The proposed model aims to estimate the user’s locomotive intent at a high level.It is combined with mid-level controllers and a low-level controller to achieve hierarchical control of an intelligent powered lower-limb prosthesis.In order to conduct experiments on human locomotion intent prediction,the subject wears an intelligent powered knee prosthesis to walk in both laboratory environments and complex daily life scenes.The main research contents and conclusions of this paper are as follows:(1)The study recruited a total of 8 patients,consisting of 7 males and 1 female,who had undergone above-knee amputations.Data was collected using a multi-source sensing system integrated into the Bionic Knee VT 2.0.This system recorded data from 8 channels of sensors,which included three-axis acceleration,three-axis angular velocity,knee joint angle,and vertical ground reaction force(v GRF).The data was collected for 9 different daily locomotion modes,such as low-,mid-,and fast-speed level-ground walking,ramp ascent/descent,stair ascent/descent,and sit/stand.The objective of the study was to develop a locomotion database specifically for steadystate locomotion in patients with above-knee amputation The purpose was to create a locomotion database for steady-state locomotion in patients with above-knee amputation.(2)This paper examines the impact of data segmentation window,classifier location,and feature selection methods on classification outcomes.The experimental findings indicate that the hybrid Gray Wolf Optimization and Particle Swarm Optimization(BGWOPSO)feature selector is employed to choose time domain features extracted from the 200 ms time window following foot contact with the ground.By utilizing the selected 23 features,it is determined that the K-Nearest Neighbor(KNN)classifier achieves the highest classification accuracy of 93.33%among the eight machine learning classifiers.(3)This paper presents a proposed model for recognizing human locomotion intents.The model achieves a high accuracy of 96.66% ± 0.68% in recognizing 9daily locomotion modes using only 7 features,including the mean value of Y-axis acceleration.By utilizing parameters obtained from offline model training,a real-time human motion intention prediction system is developed.The improved KNN algorithm not only reduces the running time and memory usage of the prediction system,but also enhances the classification accuracy.The shortest real-time prediction time achieved is only 9.8ms.(4)To address the challenges associated with the complex structure and large parameter requirements of existing classification models,the authors propose an enhanced Extreme Learning Machine(ELM)classifier for recognizing the intent of human locomotion.This classifier is specially tailored for intelligent powered knee prostheses and can accurately identify 9 daily locomotion modes using a minimal set of 282 parameters.The improved ELM classifier utilizes the output coefficient and intercept of the Logistic Regression(LR)classifier to effectively reduce the number of hidden layer nodes,ensuring high classification accuracy.Additionally,a hybrid optimization algorithm,which combines Grey Wolf Optimization and Slime Mould Algorithm(GWO-SMA),is introduced to optimize the hidden layer bias of the enhanced ELM classifier.The numerical results demonstrate the success of the proposed model in accurately recognizing nine daily motion modes.With 5-fold cross-validation,the model achieves an impressive accuracy of 96.57% while maintaining a real-time prediction time of only 2 ms.(5)In this paper,we propose a real-time human locomotion intent prediction system that seamlessly switches between locomotion modes.This system aids subjects in smoothly transitioning between multiple locomotion states in various environments,including a laboratory,an indoor staircase scene,and an outdoor campus scene.Subject TF02 walked a total of 111 steps(56 steps on the prosthetic side)in the outdoor campus scene,completing the task in just 89 seconds.The experiment demonstrates that the intelligent powered knee prosthesis effectively assists patients with above-knee amputations in restoring their daily mobility.Furthermore,the real-time human locomotion intent prediction system achieves mode recognition in just 1 ms and successfully sends mode switching instructions to the prosthesis,allowing seamless transitions between locomotion states.The proposed real-time human locomotion intent prediction system aims to address the issue of the complex structure of the prediction model,which hinders its application in real-time prediction of complex scenes involving multiple locomotions in daily life.The results indicate that the implementation of human locomotion intent recognition models in an intelligent powered knee prosthesis can assist above-knee amputees in restoring their ability to walk in various locomotion modes,thereby enhancing their quality of life and overall well-being. |
