Abstract
Control systems for prosthetic hands, actuated with muscle-like actuators face challenges in achieving natural and intuitive movement. The utilization of electromyography (EMG) has the potential to control wearable devices and establish a connection between the human body and smart devices. Although this holds a great promise, there are still challenges to be addressed. Accurately replicating control signals from nerve and muscle sources is significant due to the complexity of signals and the high noise-to-signal ratio. This must be effectively addressed to ensure precise control of the robotic hand, prior to implementing artificial intelligence for the creation of smart devices. To address these challenges, we utilized EMG dataset comprising 8 channels, having 15 different hand motions, with 24 trials (8 individuals with 3 repeated measurements), and 20 seconds of data for each trial. The dataset was sampled at a rate of 4000 samples per second. We used Generalized Eigendecomposition (GED) technique to isolate and distinguish various sources of naturally occurring signals in the human arm, simultaneously reducing the noise ratio. Subsequently, these processed signals are then employed to train machine learning models for movement recognition, imitation learning, and adaptive control. Our machine learning algorithms trained using GED achieved exceptional accuracy in recognizing unprocessed EMG signals. For individual data, we achieved 100% accuracy, while for combined training and testing data from multiple individuals, the accuracy reached over 83%. Notably, our approach outperformed convolutional neural network models on the same dataset in terms of both speed and accuracy. These results highlight the effectiveness of employing multivariate signal processing techniques, such as GED, in improving accuracy, reducing overfitting, and enhancing the robustness and reliability of predictions. These findings have significant implications for optimizing the control of prosthetic hands.
