Advancing Myoelectric Prothesis Control via EMG-Force with Transfer Learning and via Wireless Sensor Time Synchronization

Wang, He

Etd

Advancing Myoelectric Prothesis Control via EMG-Force with Transfer Learning and via Wireless Sensor Time Synchronization

Public Deposited

The surface electromyogram (EMG) signal can be collected when muscle tissue receives motor commands. EMG has been widely used in fields including myoelectric prothesis control and human rehabilitation. This dissertation initially focused on a simplified method for optimal estimation of the time-varying standard deviation of EMG (EMGσ), comparing four whitening techniques under different conditions. Then, a study of EMG-force modeling when limited calibration data are available was investigated. Bidirectional-Long Short Term Memory (Bi-LSTM) neural networks utilizing transfer learning were applied to both control subjects and limb-absent subjects and compared to a conventional “baseline” linear regression method. Finally, two time synchronization and data alignment methods for a distributed wireless EMG/bioelectric signal acquisition system were introduced and compared using the Bluetooth Low Energy (BLE) protocol. The wireless system is intended to replace socket-based wired EMG electrodes to support the evolving osseointegration of prosthetics. The first part of this thesis was a study of simplified optimal estimation of EMGσ. In order to facilitate the broader use of EMG signal whitening, which is not commonly used due to its complexity, four different whitening approaches were studied in an EMG-force task. We analyzed the performances of subject-specific whitening, universal IIR whitening, high- pass Butterworth whitening, and first difference whitening by using force-varying and constant-force contractions from 64 subjects who completed constant-posture tasks about the elbow by tracking a random target. We found that root difference of squares (RDS) processing was necessary for noise removal from the EMG of constant-force tasks. From the force-varying tasks, we first found that sampling at 4096 Hz has significantly lower EMG-force error compared to 2048 Hz and 1024 Hz. At the sampling rate of 4096 Hz, we found subject-specific whitening achieved the lowest average error [4.74% maximum voluntary contraction (MVC) EMG-force error], followed by universal IIR whitening (4.83 % MVC), followed by high-pass Butterworth whitening (4.89 % MVC), then a first difference whitening filter (4.91 % MVC) — but no significant difference was found between these approaches, yet all had lower error than the unwhitened method (5.5 % MVC). The first difference method needed no calibration (prior information), hence is an excellent whitening choice which can simplify the processing. The second part of this thesis was a study of EMG-force modeling using Bi-LSTM models with transfer learning and limited training data. It is expected that lower error EMG-force models can improve the performance of myoelectric control. However, limited data from limb-absent subjects makes it hard to design a generative deep neural network. We pre-trained Bi-LSTM models using control subjects’ data and applied transfer learning with training durations of 40, 20, 10, 5 and 2.5 s in intra-subject, cross-subject and cross-number of DoF scenarios, and compared to a “baseline” linear regression method. For limb-absent subjects, the performance of Bi-LSTM models with transfer learning did not differ as a function of training duration. Thus, minimum calibration durations can be used. Also, Bi-LSTM models with transfer learning always outperformed the baseline method whenever a significant difference was found. In general, the Bi-LSTM model with transfer learning provided lower-error EMG-force estimation with limited training data. The third part of this thesis introduced and compared two time synchronization and data alignment methods, denoted SDA and LIDA, based on the BLE wireless transmission protocol. Existing time synchronization methods are not suitable for multi-channel bioelectric data acquisition systems which require high precision, low latency, and high throughput—such as wireless electrodes for myoelectric prosthesis controllers. These two methods were designed to meet the requirements and were implemented in the application layer of the BLE protocol, which makes them transferable between different platforms and not affected by the rapid development of the BLE protocol. Testing results found that LIDA always had significantly lower time alignment error than SDA. The lowest average absolute alignment error of LIDA was 189.9 ± 204.7 μs, which is well below one sample period for commonly acquired bioelectric signals.

Creator