RFSS: A Multi-Standard RF Signal Source Separation Dataset with 3GPP-Standardized Channel and Hardware Impairments

Abstract

Cellular standards from 2G to 5G share spectrum, and receivers see several at once. Source separation could recover the waveforms, but to our knowledge no public dataset provides per-source references for multi-standard cellular mixtures. We present RFSS (RF Signal Source Separation), 100,000 multi-source samples built from the project's own implementations of the four 3GPP physical-layer specifications (GSM, UMTS, LTE, 5G NR), with a companion file of 4,000 single-source references. Each sample mixes 2 to 4 sources drawn with replacement, at rates from 2.166 to 122.88 MHz. Every source passes through its own 3GPP Tapped Delay Line (TDL) multipath fading channel, and the impairment mode is drawn per source: 19.9 percent of the released sources receive no impairment, 29.9 percent receive a single impairment type and 50.1 percent receive all six stages, whose two frequency offsets come from one oscillator error. Sources are mixed co-channel or adjacent-channel at a fixed 2 MHz spacing, and the multi-source file is 102.5 GiB in HDF5, split 70/15/15 by index. We benchmark FastICA, Frobenius-norm NMF and three deep models (Conv-TasNet, DPRNN, an STFT-domain BLSTM) with the permutation-invariant scale-invariant signal-to-interference-plus-noise ratio (PI-SI-SINR), and report the gain over the input mixture. Each deep model uses a validation-chosen learning rate. On the two-source test split the best deep model improves PI-SI-SINR by 6.1 dB over the input, where an ideal-ratio-mask oracle reaches 8.0 dB and ICA and NMF reach -13.6 and -1.2 dB. The dataset, trained checkpoints and evaluation code are released publicly for non-commercial use.