Arm Movements Recognition by Implementing CNN on Microcontrollers

Surface electromyography is a technique mainly used to detect hand movements to help patients regain control over their fingers or manipulate prosthetic arms. This body signal measuring technique is usually used with machine learning to recognize various arm movement s. However, past studies on arm movement recognitions used powerful computers that is inconvenient for patients to carry around to perform real-time sEMG signal measuring. This paper compares the performance of the two commonly used sEMG signal feature extraction methods, 1D-CNN, and 2D-CNN architectures. We first collected sEMG signals from 10 subjects. The 1D-CNN architecture reached an average recognition accuracy of 89.4% and the 2D-CNN architecture reached an average recognition accuracy of 98.9%. The 2D-CNN architecture is converted from TensorFlow file to TensorFlow Lite file and is imported into the Arduino nano 33 BLE sense microcontroller. The microcontroller is able of repeating the machine learning process with a processing time of 79-85ms and 132-135ms respectively for 1D-CNN and 2D-CNN models. In the future, it is suggested that ASIC devices with specially designed electrodes can be applied to further reduce power consumption, size, and processing time of the device to help patients regain control of their hands or to manipulate prosthetic hands to perform dangerous experiments.