Hybrid hardware-algorithm platform for EMG pattern recognition

Abstract

The effort of making an wearable sensor module, which can sense, collect and upload electrophysiology data such as electromyogram(EMG), electrocardiogram(ECG) and electroencephalogram(EEG), is keep increasing in the research area. At the same time, faster computer gives the occasion to apply and deploy machine learning algorithm (M.L) to the healthcare application. These two phenomena met together and opened a new era of the signal processing for electrophysiological data. One of the branch on the application of M.L with the EMG is to recognize patterns (EMG-PR) of human gestures. Many studies have been done to augment the accuracy of the EMG-PR. For example, there is a way to analyze and to approach with the mathematical way, a method to compare EMG-PR method, a study mixing different EMG-PR method to increase the accuracy and research applying 1 M.L technic after doing a data preprocessing. The data preprocessing is a common step for the EMG-PR and wavelet transform was proved for its efficiency. In this paper we present a programmable ASIC chip including 3 channels of EMG detectable circuits, a wavelet filter and an ADC with the resolution varying from 8 to 12 bits. EMG-PR study is generally limited in the hardware part and various studies were done using a commercial hardware material such as EMG Myo-Armband to get the EMG raw data. Here we designed an efficient ASIC chip with 3 channels including an algorithm processing. In comparison with the commercial Myo armband, the power consumption of the platform including the STM32F0 series microcontroller, Bluetooth, and the ASIC chip, is lower. The communicated data would then already be wavelet-processed data so the computational time for the data pre-processing could decrease for the input of the BP ANN.