2026 International Conference on Sensing, Measurement & Data Analytics in the era of Artificial Intelligence (ICSMD 2026)

Signal-to-noise ratio (SNR) estimation in radio frequency (RF) loopback testing is biased by the composite colored noise whose PSD straddles the inverse-square-frequency and inverse-frequency regions, and is further complicated by the coherent reference-clock phase noise and the fractional-bin spectral leakage. We adopt the in-band SNR definition, namely the signal power divided by the noise power within the signal band, since out-of-band noise is removed by the receiver filter and does not compete with the signal, and propose three complementary components on the Welch PSD: (1) a curvature-adaptive robust LOESS envelope that preserves sharp in-band spectral features; (2) a leakage-kernel deconvolution that recovers the signal power from the fractional-bin leakage; and (3) a Whittle-likelihood multi-frequency estimator that jointly fits an inverse-square-frequency plus inverse-frequency plus white-noise model across K carriers with K-fold variance reduction. The deconvolution and the Whittle fit are solved jointly by alternating a projected Landweber update with a frequency-dependent noise-model fit, which keeps the estimate unbiased down to SNR = −20 dB where the scalar-noise-floor baselines fail. As shown in the simulations over three experiments, the adaptive LOESS keeps the mean error within ±0.2 dB across the −10 to 30 dB SNR range, the Whittle estimator within [−0.14, −0.08] dB, and the adaptive window delivers a 3 to 11 dB SNR accuracy gain over the fixed-span baseline in the in-band spur scenario, where the joint estimator is biased by about 3 dB and the conventional full-band average fails.

Adaptive-window LOESS, coherence modeling, iterative deconvolution, RF loopback, SNR estimation.