EffiE: Efficient Convolutional Neural Network for Real-Time EMG Pattern Recognition System on Edge Devices

With the advancement of deep learning (DL) technologies, applying DL methods to processing surface electromyographic (sEMG) signals for movement intent recognition has gained increasing interest in the research community. Compared to conventional non-DL methods commonly used for EMG pattern recognition (PR), DL algorithms have the advantage of automatically extracting sEMG features without the cumbersome manual feature engineering step and are especially effective in processing sEMG signals collected from 1-dimentional (1D) or 2D sensor arrays. However, a key challenge to the deployment of DL methods in sEMG-controlled neural-machine interface (NMI) applications (e.g., myoelectric controlled prostheses) is the high computational cost associated with DL algorithms (e.g., convolutional neural network (CNN)) since most NMI applications need to be implemented on resource-constrained embedded computer systems and have real-time requirements. In this paper, we designed and implemented EffiE - an efficient CNN for real-time EMG PR system on edge devices. The development of the EffiE system integrated several strategies including a deep transfer learning strategy to adaptively and quickly update the pre-trained CNN model based on the user's newly collected data on the edge device, and a deep learning quantization method that can dramatically reduce the memory consumption and computational load of the CNN model without sacrificing the model accuracy. The proposed EffiE system has been implemented on a Sony Spresense 6-core microcontroller board as a working prototype for real-time NMIs. The embedded NMI prototype has integrated input/output interfaces as well as efficient memory management and precise timing control schemes to achieve real-time DL-based myoelectric control of a bionic arm using hand gestures. We released all the source code at: https://github.com/MIC-Laboratory/EffiE