Robustness-enhanced Myoelectric Control with GAN-based Open-set Recognition

TL;DR

This work tackles instability in myoelectric control caused by EMG signal variability and unseen gestures by introducing a GAN-based open-set recognition framework. A CNN-based gesture classifier is complemented by a GAN discriminator that can identify and reject unknown actions, with thresholding guided by ROC-AUC to ensure reliable open-set handling. The method achieves 97.6% accuracy on known gestures and reduces Active Error Rate by 23.6% when unknown gestures are rejected, tested on Ninapro DB1 and self-collected data, while remaining lightweight enough for edge deployment. The study discusses integration with confidence-based rejection and transfer learning to further enhance cross-domain robustness and real-world usability, suggesting a practical pathway toward robust, open-set aware myoelectric systems.

Abstract

Electromyography (EMG) signals are widely used in human motion recognition and medical rehabilitation, yet their variability and susceptibility to noise significantly limit the reliability of myoelectric control systems. Existing recognition algorithms often fail to handle unfamiliar actions effectively, leading to system instability and errors. This paper proposes a novel framework based on Generative Adversarial Networks (GANs) to enhance the robustness and usability of myoelectric control systems by enabling open-set recognition. The method incorporates a GAN-based discriminator to identify and reject unknown actions, maintaining system stability by preventing misclassifications. Experimental evaluations on publicly available and self-collected datasets demonstrate a recognition accuracy of 97.6\% for known actions and a 23.6\% improvement in Active Error Rate (AER) after rejecting unknown actions. The proposed approach is computationally efficient and suitable for deployment on edge devices, making it practical for real-world applications.

Figures (14)

• Figure 1: Three phases of EMG-based signal control flow for human motion detection: signal acquisition, pattern recognition and pre-execution. The techniques/approaches focus on known classes only are categorized as Close scenario, while the ones focusing on both known and unknown classes are categorized as Open scenario. The pre-execution phase involves explorations like confidence-based rejection. In Open scenario, the focus shifts to unknown class detection and rejection.
• Figure 2: The Framework for GAN-based EMG Signal Pattern Recognition.
• Figure 3: known and unknown class, test and train set, GAN and CNN training and testing.
• Figure 4: The architecture of CNN-based classifier. The input to the CNN is an $N_{Channel} \times N_{Sample~points}$ matrix representing the EMG signals. The input is processed through two convolutional layers, Conv1 and Conv2, each comprising 32 $3\times3$ convolutional kernels. Padding is set to 1 to maintain consistent input and output dimensions for each convolutional layer. The output of Conv2 is flattened and passed to the fully connected layer, FC1. The number of neurons in the final layer is set to match the number of known classes, $N_{known}$.
• Figure 5: The GAN consists of two competing networks during the training process: the generator and the discriminator. The generator takes random Gaussian noise of size $1 \times N_{Hidden}$ as input and produces an output vector of length $N_{Known}$. The input to the discriminator has a size of $(1 \times N_{Known})$, and its output is a single numerical value between 0 and 1, representing the classification judgment. The fully connected layers FC1 and FC2 in the discriminator consist of 128 and 64 neurons, respectively.