CityU researchers develop self-supervised AI framework for high-accuracy EMG sensing

Researchers from the City University of Hong Kong (CityU) have made significant progress in addressing the variability in electromyography (EMG) sensing signals caused by biological differences among users. Their groundbreaking development, known as EMGSense, is a deep learning-based framework that utilizes AI self-training techniques to achieve high sensing performance for new users. This innovation has the potential to revolutionize the field of EMG and enable the development of more accurate wearable devices for applications such as neurorehabilitation and virtual reality.

The exceptional capabilities of EMGSense were recognized at The 21st International Conference on Pervasive Computing and Communications (PerCom 2023), held in Atlanta, U.S., where it received a prestigious award. By overcoming the limitations of existing approaches, EMGSense paves the way for widespread adoption of EMG-based applications.

EMG involves the measurement of muscle electrical activity using surface electrodes placed on the skin. This sensing technique has gained significant attention in recent years, leading to the development of intelligent applications such as neurorehabilitation, activity recognition, gesture recognition, and virtual reality.

However, a fundamental challenge in EMG systems lies in addressing cross-user scenarios. The reliability of EMG signals can be heavily influenced by various biological factors, including body fat, skin conditions, age, and fatigue. Consequently, when EMG systems are utilized by different users, the time-varying biological heterogeneity can lead to significant performance degradation.

To tackle this challenge, researchers from CityU’s Department of Computer Science have introduced the first low-effort, AI-empowered domain adaptation framework: EMGSense. This framework offers high-accuracy EMG sensing for new users through the use of AI training techniques. EMGSense operates as a self-supervised system with a self-training AI strategy, effectively addressing the performance degradation caused by inter-user biological heterogeneity.

The novel framework combines advanced self-supervised techniques with a carefully designed deep neural network (DNN) structure. It leverages small-scale unlabeled data from new users along with pre-collected data from multiple existing users to train a discriminative model, enabling intelligent applications for new users. The pre-collected data is stored in the cloud and can be accessed by all new users, reducing the burden of data collection and annotation.

The key principle of the method is the shared common feature extractor, whose aim is to ensure the transferability of features. The combination of domain-specific feature extractors and classifiers are responsible for independently exploring the diversity among the deep features from different source domains. Credit: Di, D. et al,

The DNN structure of EMGSense incorporates two training stages that complement each other, enabling effective adaptation to new users without significant deployment overhead. In the first stage, user-specific features in the feature space are eliminated to facilitate easy transfer. Subsequently, AI techniques are employed to re-learn the new user’s specific biological features in that space, resulting in high-performance EMG sensing. This self-supervised approach ensures satisfactory performance for new users while minimizing effort and resource requirements.

To address the time-varying nature of EMG signals, the researchers utilized the unlabeled data collected during usage, enabling long-term robust performance.

Through a comprehensive evaluation using two substantial datasets from 13 participants, EMGSense achieved impressive results. It obtained an average accuracy of 91.9% in gesture recognition and 81.2% in activity recognition. EMGSense outperformed existing EMG-oriented domain adaptation approaches by 12.5% to 17.4% and demonstrated comparable performance to supervised-learning models.

The innovative EMGSense framework has the potential to revolutionize EMG sensing by reducing the burden of data collection and annotation, while maintaining high accuracy in a low-effort manner. It addresses the challenges associated with heterogeneity in EMG sensing and opens avenues for various cross-user applications, including clinical practice, neurorehabilitation, and human-machine interaction. Furthermore, it contributes to the advancement of smart EMG wearable devices, enabling superior performance in real-world scenarios.

The research paper detailing EMGSense’s development and findings has been published as part of the 2023 IEEE International Conference on Pervasive Computing and Communications (PerCom).

Source: City University of Hong Kong

Leave a Reply

Your email address will not be published. Required fields are marked *