Streaming Fixed-Point FPGA Preprocessing and Feature Preparation for QCM Sensor Signals on a SoC Platform

This paper presents an FPGA-based streaming preprocessing pipeline for Quartz Crystal Microbalance (QCM) sensor signals on a Zynq-class system-on-chip (SoC). In electronic-nose applications, QCM frequency-shift streams are often noisy and offset-dependent, requiring conditioning before data-driven analysis. The proposed AXI-Stream/DMA pipeline implements baseline shifting, fixed-point exponential moving average (EMA) smoothing with a shift-based coefficient α=1/4, and interval-based gradient extraction with L=8, all in programmable logic with initiation interval II=1. On-chip inference is intentionally excluded to keep the IP lightweight and independent of the downstream ML model, enabling reuse across deployments. Numerical evaluation against a software reference shows that 9-13 fractional bits preserve fidelity with low error and marginal resource increase, with gradient features being the most quantization-sensitive. The pipeline produces stable, analysis-ready features suitable for downstream feature aggregation and edge artificial-intelligence pipelines.