Development of Surface Electromyogram Operated Exoskeleton for Lower Limb

Abstract

Surface electromyography (sEMG) is a technique used in human-machine interface, rehabilitation, and exoskeleton control. In recent years, cost reductions and increased availability of essential technology have made sEMG a more realistic alternative for the development of the exoskeleton. sEMG based exoskeletons can identify the user's intentions and enable cooperative interaction, but sEMG base control strategy in assistive devices is currently unclear since each individual's muscles and joint forces are unique. This may eventually increase the time and cost of developing a real-time design exoskeleton. Biomechanical simulation tools and arti cial intelligence are being used to save time and cost to improve the design. This thesis presents a human biomechanics analysis for knee joints in AMS, interfacing of design exoskeleton with human musculoskeletal, Arti cial Intelligence (AI) techniques for prediction and classi cation. These are brought together to set a new baseline for optimizing the real-time development of sEMG based exoskeleton. The thesis work includes the design of a 3D lower limb exoskeleton in SolidWorks CAD software. The sEMG signals were obtained from lower limb muscles and converted into muscle force using the Hill muscle model. In addition, a statistical comparison of experimentally generated muscle forces with human musculoskeletal muscle forces was performed in AMS. ANOVA t-test was conducted on validation datasets. A new approach for predicting muscle force was developed using a Machine Learning (ML) approach called a force simulator. These models were trained and tested using dataset 500 human musculoskeletal generate using python tools in AMS, where human musculoskeletal utilized to perform squatting movement during inverse dynamic analysis. To predict knee muscle force, four di erent ML models were trained and tested on datasets. The random forest-based ML model outperforms the other models: Neural Network, Generalized Linear Model, Decision Tree in terms of mean square error (MSE), coe cient of determination (R2), and Correlation (r) for the musculoskeletal datasets. The MSE of predicted vs actual muscle forces obtained from the random forest model for Biceps Femoris, Rectus Femoris, Vastus Medialis, Vastus Lateralis was 19:92; 9:06; 5:97; 5:46, and Correlation was 0:94; 0:92; 0:92; 0:94, and R2 was 0:88; 0:84; 0:84, 0:89 for the test datasets, respectively. Particle Swarm Optimization-Long Short Term Memory (PSO-LSTM) was used to classify three di erent movements (Flexion, Extension, Ramp Walking) based on sEMG signal and predict the results of sEMG signal and knee angle. After that, random vii LSTM was used to validate the results to explore the e ectiveness of PSO-LSTM. Four knee muscles sEMG signals, namely biceps femoris (BF), vastus medialis (VM), rectus femoris (RF) and semitendinosus (ST), and knee angle, were used as model inputs. RMSE, r, and R2 were taken as evaluation parameters to identify the model's robustness for predicting SEMS signal and knee angle. The average RMSE value for an extension, exion and ramp walking for both PSO-LSTM and random LSTM model was (80%; 10:16%); (133:33%; 30:96%) and (116:66%; 19:48%), respectively. The average 'r' value for an extension, exion, ramp walking for both PSO-LSTM and random LSTM model was more by (4:65%; 4:04%); (6:60%; 3:03%) and (3:57%; 7:07%), respectively. The average R2 value of sEMG for an extension, exion, ramp walking for both PSO-LSTM and random LSTM model was higher by (5:88%; 4:08%); (5:05%; 4:04%) and (5:33%; 8:16%), respectively. The PSO-LSTM model was used to classify three di erent movements (Flexion, Extension, Ramp Walking) with an accuracy of 98:58%. It was observed that the proposed PSOLSTM model shows better capability than the random LSTM model for both inputs (sEMG signals, knee joint angle). The prototype of the knee was developed in the laboratory and operated through sEMG signal for three activities (Flexion, Extension, Ramp Walking) to actuate the linear actuator. The PSO-LSTM model helped was used to validate the results of three di erent activities.