SNR-Anchored Representative Data for Robust INT8 Automatic Modulation Classification

Edge automatic modulation classification (AMC) benefits from full-integer inference, but post-training quantization (PTQ) estimates activation ranges from a small representative set whose physical-condition coverage is rarely reported. We study how representative-set signal-to-noise ratio (SNR) composition affects INT8 AMC. Two compact one-dimensional convolutional networks were trained with three independent seeds and evaluated on RadioML 2016.10A and RadioML 2016.04C. Each model was quantized with 1,024 training-only records selected using balanced-SNR, low-SNR-only, and low-SNR-heavy distributions. Low-SNR-only calibration reduced mean accuracy by 2.31 and 5.94 percentage points on RadioML 2016.10A, and by 14.56 and 14.69 points on RMS-normalized RadioML 2016.04C, for the lightweight and standard networks respectively. Layer-level integer-tensor analysis linked the loss to underestimated late-layer ranges and high-SNR boundary saturation. Replacing only 1% of the low-SNR calibration set (10 of 1,024 records) with high-SNR anchors recovered 3.97-14.74 percentage points relative to low-SNR-only calibration and matched balanced calibration within 0.05 points on average. Five calibration draws per trained model showed the recovery was not tied to one lucky sample. An unnormalized negative control on RadioML 2016.04C failed, establishing that sparse anchors require controlled input-power scaling. These results identify SNR coverage as a deployment variable and provide a simple representative-set safeguard for integer-only AMC.