Studies of Surface Electromyogram-to-Joint-Torque Relationship in Two Upper-Limb Joints via a Transfer Learning Approach

Wang, Haopeng

Etd

Studies of Surface Electromyogram-to-Joint-Torque Relationship in Two Upper-Limb Joints via a Transfer Learning Approach

Público Deposited

The surface electromyogram (sEMG) measures the electrical activities of muscles and is often related to muscle force/joint torque. sEMG has many applications including clinical biomechanics, exercise science, and the focus of this dissertation: myoelectric prostheses control. sEMG-based myoelectric prostheses control can often be categorized into two types of schemes: classification/pattern recognition-based control and regression-based control. Classification-based control schemes take various sEMG features as input and generate categorical outputs, usually patterns of motion. Advanced sEMG classifiers such as neural networks-based classifiers can have over 95% classification accuracy. However, typically only one motion can be selected at a time from the classifier. Human limb movements are often simultaneous and proportional, thus, regression-based control can better replicate natural movements. Using sEMG signals to estimate force is the most commonly used regression-based myoelectric control scheme. Traditionally, sEMG to force modeling has been done in a subject/condition-specific manner for single-use models due to inter-subject variability and other confounding factors such as electrode shift. Single-use sEMG to force models are often believed to have high accuracy but are burdensome to calibrate. In this dissertation, we attempt to generalize sEMG to force modeling across subjects and two upper-limb joints (elbow extension-flexion, hand-wrist with four degrees-of-freedoms (DoF) (extension-flexion, radial-ulnar, pronation-supination, open-close), and aim to improve model training efficiency and overall model performance via transfer learning. Two parameter-based transfer approaches are studied. The first one is a system identification model calibrated by regularized least-squares regression and ensemble learning. The second one is deep learning models using neural networks and fine-tuning. In the first study, we created a generic EMG to force model that captures the dynamics and leaves only gain for each channel to calibrate. The fit parameters from each EMG channel form an FIR low-pass filter with similar cut-off frequencies but different gains. We normalized each such filter to a DC gain of one, and took the ensemble median to form a generic EMG to force model for each DoF in hand-wrist, and for elbow and wrist joint (all DoFs combined), respectively, with various dynamic model orders. We then applied the generic EMG to force models in cross-subject, cross-DoF, and cross-joint scenarios, comparing their performances with subject-specific models and with each other. We found that even in some cases when there is a statistical difference (e.g. generic vs subject-specific in elbow), the strength of this difference should be negligible in real-life applications. Thus, even though there are still factors such as muscular effort level that can introduce variance in gain selection for EMG to force models, the EMG to force dynamics are likely to be similar across subjects, DoFs, and even joints. In the second study, we pre-trained neural network models from the elbow joint and fine-tuned model parameters on hand-wrist data to improve sEMG to torque estimation accuracy and reduce training cost. Deep neural networks (DNNs) usually require a large amount of training data to build robust models. With inspiration from the first study and combined with a relatively large available elbow dataset at hand, we explored cross-joint transfer learning with DNNs. We first evaluated four neural network structures (multilayer perceptron (MLP), convolutional neural networks (CNN), long short-term memory network (LSTM), CNN-LSTM concatenated (C-LSTM)) with different sliding window lengths. We found that a feedforward structure with a 391 ms sliding window length had similar performance to that of a recurrent structure, thus, the time dependency between sEMG and torque is most likely short-term. Then, we aggregated 65 subjects’ data from the elbow dataset to obtain pre-trained models for each DNN architecture separately. To investigate the effect of transfer learning, we compared models trained from scratch (i.e., DNN model parameters were randomly initialized) versus models fine-tuned with a smaller learning rate based on the pre-trained model (i.e., pre-trained model parameters were used as the starting point). We found that transfer learning improved model performance, and when compared to the least-squares regression models from the first study, we reduced the required training data duration by half. This dissertation also covers several other collaborative projects including simplified optimal EMG amplitude estimation, real-time myoelectric prostheses control, and EMG to force with transfer learning in limb-absent subjects.

Creator