A high-throughput and small ADALINE based adaptive noise canceller for 50-hz noise in surface Electromyography on an FPGA

This paper presents a design of an adaptive noise canceller for 50-Hz noise in surface Electromyography (SEMG). An adaptive linear neural network (ADALINE) filter without reference signal from the outside was designed into three formats: a 32-bit floating-point format, a 32-bit fixed-point format, and a 16-bit fixed-point format. In additions, the 16-bit fixed-point format was optimized for area using the resource-sharing technique that considers the interconnect complexity. The design and the simulation were performed by using Matlab Simulink and Xilinx Accel DSP Toolbox to find the most optimized circuit on Xilinx Spartan-3 (XC3S400-4TQ144). The filter efficiency was measured by the correlation coefficient and the normalized mean square error (NMSE). From the results, the 10-tap ADALINE implemented in the 16-bit Q0.15 fixed-point format can achieve the throughput of 1.656 MSPS and the area of three multipliers, 1831 LUTs, and 858 flip-flops, whereas the signal quality is the same as the others.