Robustness of surface EMG classifiers with fixed-point decomposition on reconfigurable architecture

To the present day, the control of prosthetic mechanisms and the classification of surface Electro-Myography (sEMG) signals is heavily influenced by multiple factors, such as the overall quality of the signal acquired, the number and the position of electrodes, and unwanted noise or crosstalk between them. For this reason, the technological state-of-the-art is moving from the analysis of low-order features (e.g. Median Absolute Value and spectral features) that could suffer the aforementioned issues, to more complex techniques, such as Independent Component Analysis (ICA) separation Non-Negative Matrix Factorization (NMF) decomposition of signal channels. The latter hypothesizes a high-order modular control of muscles at the Central Nervous System (CNS) level built upon coordinated motor patterns (called muscle synergies), and it removes the limitations on Degrees-of-Freedom (DoF) per electrode given by standard sEMG controllers. Unfortunately, the utilisation of such techniques in prosthetics is limited by their computational complexity, which limits them to research laboratories and bulky processing systems, calling for novel algorithms and hardware implementations that achieve same the level of accuracy and speed necessary for movement control, while minimizing the power consumption and the device size. This paper presents a FPGA-based, real-time NMF processor that extracts muscle synergies from a 8-electrode sEMG recording. The execution of the NMF algorithm is entirely performed using a fixed-point architecture, in order to increase computation speed and minimize the amount of DSPs used. The NMF fixed-point decomposition is then fed to two software different classifiers, a Support Vector Machine (SVM) and a Neural Network, in order to quantify the degradation of the accuracy given by fixed-point calculations and to understand which mechanism is more robust. Preliminary results show that there are no remarkable differences between the classification results obtained using floating-point or fixed-point operations.