IQ Length Normalization Policies for RF Modulation Classifiers

IQ Length Normalization Policies for RF Modulation Classifiers

Modern RF modulation classifiers often mix spectral and
temporal encoders: convolutional networks over FFT-based
spectra, recurrent networks over IQ sequences, and hybrids
that fuse both [1]. Temporal models, in particular, require a
fixed sequence length L; however, bursts arriving from live
receivers exhibit variable duration, symbol rates, and capture
configurations. Normalizing these streams to a common length
is unavoidable, but the design space of “how to choose which IQ
samples survive” is rarely explored. RF–QUANTUM–SCYTHE
is an open-source RF machine learning stack we use throughout
this paper series for automatic modulation classification and
related SIGINT tasks.1
In RF–QUANTUM–SCYTHE, the temporal input path
is implemented by EnsembleMLClassifier via a helper
_create_temporal_input, which maps an arbitrarylength complex IQ array to a fixed-length real-valued tensor
suitable for LSTMs, temporal CNNs, and signal transformers. Earlier work in this paper series examined short-signal
resilience by studying behaviour when |x| < L and padding
1Source code and scripts are available at https://github.com/bgilbert1984/
rf-quantum-scythe (placeholder URL).
strategies. Here we focus on the opposite axis: given plenty of
IQ samples, which normalization policies preserve classification
performance as we squeeze sequence length down to fit latency
and memory budgets?
A. Contributions
We make three contributions:

We formalize three IQ length normalization policies—
evenly spaced downsampling, windowed pooling, and
strided cropping—and implement them as index-selection
strategies inside _create_temporal_input.
We run per-length sweeps across a range of sequence
lengths and report both modulation accuracy and a simple
aliasing proxy that measures spectral distortion relative
to full-length references.
We package the sweeps into a reproducible harness that
logs JSON metrics and drives LATEX-ready figures and
tables, making it easy to evaluate new policies or lengths
with no manual plotting.

Importance of Validating on Real RF Datasets

Your paper’s experiments rely on synthetic RF data (generated with PSK/QAM/AM/FM over AWGN and mild fading), which is a great starting point for controlled testing but may not capture real-world artifacts like hardware impairments (e.g., I/Q imbalance, phase noise), over-the-air (OTA) propagation effects (e.g., multipath fading, Doppler shifts), or variable sampling conditions from actual receivers. Validating on real RF datasets ensures the normalization policies (evenly spaced downsampling, windowed pooling, strided cropping) generalize beyond ideal simulations. This could reveal if policies like strided cropping are more sensitive to real burst misalignment or if pooling helps mitigate hardware-induced jitter.

Real datasets often include raw I/Q samples, making them compatible with your _create_temporal_input builder. Expect potentially lower accuracies due to unmodeled effects, but this strengthens your contributions by showing robustness.

Recommended Real-World RF Datasets for Modulation Classification

Based on current sources (as of November 2025), here are suitable real-world (OTA-collected) datasets. I prioritized those with raw I/Q data, multiple modulations, and public availability. These were collected using software-defined radios (SDRs) like USRP, capturing authentic wireless environments.

1. RadioML 2018.01A (DeepSig)

Description: A large-scale dataset with over-the-air recordings (mixed with some synthetic elements for channel effects). It includes 24 digital and analog modulation types (e.g., variants of ASK, PSK, QAM, AM, FM). Total ~2.5 million signals, suitable for testing your temporal models across SNR levels and real propagation.
Size: ~2.5M examples (large, ~106 GB uncompressed).
Modulations: 24 (e.g., BPSK, QPSK, 8PSK, 16QAM, 64QAM, 256QAM, AM-SSB, FM, etc.).
SNR Range: -20 dB to +30 dB.
Collection Method: Over-the-air using USRP SDRs in real environments, capturing hardware and channel impairments.
Availability: Free download via Kaggle or DeepSig’s site. Direct link: https://www.kaggle.com/datasets/pinxau1000/radioml2018.
Why Suitable: Matches your synthetic setup (includes PSK/QAM/AM/FM), allowing direct comparison. Test your policies on OTA subsets to quantify drop-offs (e.g., aliasing from real noise).

2. DeepRFSoC Dataset

Description: Real-world I/Q samples for automatic modulation classification, captured to include hardware effects from RF-ADC/DAC.
Size: Not specified (moderate, based on similar datasets; check repo for details).
Modulations: 8 (QPSK, BPSK, 16QAM, 64QAM, 8PSK, PAM4, GFSK, CPFSK).
SNR Range: Not explicitly stated; varies with real conditions.
Collection Method: Transmitted and received OTA on AMD RFSoC platform (FPGA-based SDR), simulating edge-device scenarios.
Availability: Documentation and likely download via GitHub repo: https://github.com/axdy/rfsoc_quant_amc. (Contact authors if data isn’t directly hosted.)
Why Suitable: Focuses on quantization and real hardware, aligning with your aliasing proxy. Sweep lengths on this to see if windowed pooling reduces distortion from ADC noise.

3. USRP-Collected RF Database (from “Radio Frequency Database Construction and Modulation Recognition” Paper)

Description: A real-world wireless signal dataset for modulation recognition in sensor networks, with raw samples under Rayleigh fading and AWGN.
Size: ~14,000 samples total (2,000 per modulation for training; each sample ~800,000 points at 800 kHz sampling).
Modulations: 7 (2FSK, 16QAM, 64QAM, BPSK, MSK, QPSK, 2ASK).
SNR Range: Not explicitly detailed; includes variable noise from real channels.
Collection Method: OTA using NI USRP 2920 SDRs (transmitter/receiver setup at 400 MHz center frequency, 70 MHz IF, 400 kbps symbol rate).
Availability: Not publicly linked in the paper; contact authors (e.g., via PMC or affiliations) for access. Paper provides setup details for replication.
Why Suitable: Emphasizes real multipath fading, testing if your strided cropping (energy-centered) handles burst localization better in noisy OTA data.

Additional Options

ORACLE Dataset (from related papers): ~250,000 OTA transmissions for dynamic modulation classification in real networks. Collected with SDRs; check IEEE Xplore for access (may require institutional login).
If needed, explore GitHub repos like kwyoke/RF_modulation_classification for OTA extensions of RadioML, or search IEEE DataPort for “over-the-air modulation dataset” for newer entries.

How to Perform Validation

Data Integration: Download I/Q samples (often in HDF5, MAT, or CSV). Use your synthetic generator’s format as a template to preprocess (e.g., via Python with NumPy/Scipy for loading).
Apply Policies: Modify _create_temporal_input to handle real bursts (variable N >> L). Run sweeps for L={32,64,128,256,512} as in your paper.
Metrics and Comparison: Reuse your harness for accuracy, per-modulation breakdowns, and aliasing proxy (PSD divergence). Compare to synthetic baselines—e.g., expect 5-15% accuracy drops on real data due to impairments. Slice by SNR if available.
Potential Findings: Downsampling might excel in clean OTA but degrade with multipath; pooling could smooth hardware noise. If accuracy tanks, consider augmenting with real effects (e.g., via TorchSig for hybrid testing).
Reproducibility: Log results in your JSON harness and update figures/tables. Share on your GitHub (implied from email) for community feedback.

This validation could form a strong “Future Work” extension or follow-up paper. If you provide more details (e.g., specific modulations to match), I can refine suggestions!

Overall Rating

On a scale of 1-10 (where 1 is poor/no improvement and 10 is excellent/fully addressed all feedback), I’d rate this Rev2 at 8.5/10. It’s a strong revision that incorporates most of my previous suggestions, enhancing clarity, depth, and professionalism. The paper is now more polished, reproducible, and reader-friendly, with added specifics that strengthen its technical contributions. Expanding to 4 pages allows for better elaboration without bloating. However, a few minor issues remain (e.g., lingering LaTeX typos, empty references), and some opportunities (e.g., per-SNR analysis) weren’t pursued, preventing a perfect score. This version is submission-ready for a workshop or short paper venue, with only light tweaks needed.

Key Improvements from Original

You addressed the majority of my feedback effectively. Here’s a breakdown:

Abstract: Excellent update—added quantitative results (e.g., “89.2% accuracy at L=128”) to hook readers, as suggested. This makes the impact immediate and ties directly to findings. Fixed “LATEX” to “LaTeX”.
Introduction: Added a brief definition of RF–QUANTUM–SCYTHE as an “open-source RF machine learning stack” and included a GitHub footnote (even if placeholder). This improves accessibility for broader audiences, as I recommended.
System Overview: Minor refinements for flow, but the integration explanation remains solid.
Policies Section: Great additions! Specified the pooling operation (“simple complex average over I and Q channels”) and included a concrete example (N=10, L=3), addressing my call for clarity and examples. This makes the policies easier to replicate.
Experimental Setup: Clarified the modulation set (BPSK, QPSK, 16-QAM, 64-QAM, AM, FM, CW—8 classes total), as requested. Also noted fixed model weights across tests, enhancing rigor.
Results: Fixed broken figure references (now Fig. 1 and Fig. 2) and LaTeX snippets. Added practical insights, like the 4x reduction in temporal footprint from L=512 to 128 with minimal accuracy loss (88.6% to 89.2%). Confidence intervals are consistently noted.
Discussion: Incorporated compute benefits (e.g., 4x FLOPs/RAM savings), tying policies to real-world latency. Added interaction with short-signal work and future extensions (e.g., validating on RadioML 2018.01A), aligning with my suggestions for implications and next steps.
Conclusion: Tightened for less repetition, emphasizing tunability.
Visuals and Production: Figures look clean in the provided screenshots (e.g., accuracy vs. L plots with legends: “even”, “window”, “stride”). Table is intact. No more “Fig. ??” placeholders.
General: Proofreading improved (fewer typos), and the added content (e.g., examples, quant stats) grounds the paper without overwhelming. Reproducibility is bolstered by the harness mention.

These changes make the paper more compelling and self-contained, directly responding to my points on clarity, depth, and real-world ties.

Remaining Issues and Suggestions

While much improved, a few areas could still be refined:

References: The section is present but empty in the provided text/screenshots. If this is intentional (e.g., for a tech report), note it; otherwise, add citations for related work (e.g., RadioML papers) to avoid seeming isolated.
Metrics: The aliasing proxy mentions “KL divergence or ℓ2 distance” but doesn’t specify which was used in results. Pick one (e.g., KL for distributions) and state it explicitly.
Depth Opportunities: No per-SNR or per-modulation breakdowns added (e.g., does pooling help more at low SNR or for analog mods?). This was optional in my review, but including a supplemental figure could elevate it.
Minor Typos/LaTeX: “LATEX” is fixed in abstract but check globally. In setup, “3 200 bursts” should be “3,200”. In results, table header “Acc(stride)” vs. “stride” in figs—consistent labeling helps.
Length and Scope: Now 4 pages—fine, but if targeting a venue with limits, trim repetitive phrases (e.g., policy pros/cons appear in both III and VI).
Visuals: Screenshots show good layout, but ensure high-res plots in final PDF (e.g., enable error bars as mentioned). No major issues here.

Final Thoughts

This revision demonstrates thoughtful iteration, turning a solid draft into a more mature piece. It better positions sequence length as a “tunable design parameter,” as you conclude. For next steps, validate on real data (as you mention in future work) to push toward publication. If submitting, consider IEEE Transactions on Signal Processing or a ML-for-RF workshop. Great progress—keep it up! If you share more (e.g., full LaTeX or specific focus areas), I can provide targeted feedback.

https://grok.com/share/bGVnYWN5_a809f55b-8ba2-4cc8-b528-d961f62185dc

Links:

Overview of Open Datasets for RF Signal Classification

This post presents an overview of open training datasets for radio frequency (RF) signal classification with AI and machine learning.

panoradio-sdr.de Datasets – Jithin Jagannath – Google SitesA real-world radio frequency (RF) fingerprinting dataset for commercial off-the-shelf (COTS) Bluetooth emitters under challenging testbed setups is presented in …sites.google.com DeepRFSoC: Dataset for Modulation ClassificationThe DeepRFSoC dataset provides raw in-phase and quadrature (I/Q) samples across eight modulation schemes for automatic modulation classification.pureportal.strath.ac.uk Large-scale real-world radio signal recognition with deep learningRadioML 2018.01A is a rich modulation classification dataset containing more than 2.5 million radio signals covering up to 24 analog and digital modulation …sciencedirect.com Real-World Considerations for RFML Applications – VTechWorksThis thesis covers the practical considerations for systems that can create high-quality datasets for RFML tasks, how variances from real-world effects in …vtechworks.lib.vt.edu Radio Frequency Database Construction and Modulation … – NIHIn this paper we collect a large amount of real-world wireless signal data based on the software radio device USRP 2920.pmc.ncbi.nlm.nih.gov kwyoke/RF_modulation_classification: A project on RF modulation …This notebook builds neural architectures from literature that performed quite well for the task of RF modulation classification and trains them on the 2016.10A …github.com [PDF] Deep Learning for Large-Scale Real-World ACARS and ADS … – arXivThe research on radio signal classification based on deep learning mainly focuses on two aspects: automatic modulation classification and radio frequency (RF).arxiv.org [PDF] towards a robust and efficient classifier for real world radio signalAutomatic modulation classification for radio signals is an important task in many applications, including cognitive ra- dio, radio spectrum …par.nsf.gov MATLAB-Simulated Dataset for Automatic Modulation Classification …The modulation signals were generated and transmitted through Rayleigh and Rician fading channels to simulate real-world wireless conditions.data.mendeley.com Overview of Open Datasets for RF Signal Classification – Panoradio SDR# Overview of Open Datasets for RF Signal Classification This post presents an overview of open training datasets for radio frequency (RF) signal classification with AI and machine learning methods. It compares the different datasets with respect to the data parameters, classes and classification tasks. ## Introduction The task of radio signal classification can include automatic modulation classification, signal identification and specific emitter recognition, as outlined in [this introductory article](https://panoradio-sdr.de/introduction-to-rf-signal-classification/). Good training datasets are the backbone of modern machine learning algorithms, such as deep neural networks, and are therefore very important. Several training datasets for RF signals have been published over the years. The available datasets often have very different properties, that influence their applicability to certain classification tasks. Here are some characteristics that can differ between datasets, even if they are designed fopanoradio-sdr.de Datasets – DeepSigThis dataset was used in our paper Over-the-air deep learning based radio signal classification which was published in 2017 in IEEE Journal of Selected Topics …deepsig.ai DeepSig Dataset: RadioML 2018.01A – KaggleThis dataset includes both synthetic simulated channel effects and over-the-air recordings of 24 digital and analog modulation types which has been heavily …kaggle.com [PDF] Over the Air Deep Learning Based Radio Signal Classification – arXivOver the air performance. We generate 1.44M examples of the 24 modulation dataset over the air using the USRP setup described above. Using a.arxiv.org alexivaner/Deep-Learning-Based-Radio-Signal-Classification – GitHubWe propose a new state of the art method to better classifying radio-signal network that both works on a signal with low noise (High SNR) and signal with high …github.com Overview of Open Datasets for RF Signal ClassificationThis article is an overview of open training datasets for RF machine learning, including some example plots and basic quality assessment.panoradio-sdr.de Performance of Modulation Classification for Real RF SignalsIn this paper we study the performance of DMC in real-world wireless networks, using an extensive RF signal dataset of 250,000 over-the-air transmissions with …ieeexplore.ieee.org [PDF] MODULATION CLASSIFICATION USING CONVOLUTION NEURAL …It consists of 11 modula- tion types (3 Analog Modulation and 8 Digital Modulation schemes) which are listed in Table 1, collected over 20 SNR conditions …noiselab.ucsd.edu Dataset for modulation classification and signal … – ScienceDirect.comIn this paper, we openly share a synthetic wireless waveforms dataset suited for modulation recognition and wireless signal (protocol) classification tasks.sciencedirect.com Radio Frequency Database Construction and Modulation … – NIHThe datasets are derived from the RF database. The signal dataset 1 for training contains modulated signals with different carrier frequencies collected by …pmc.ncbi.nlm.nih.gov Compact automatic modulation recognition using over-the-air …The raw RF signal dataset was digitized and curated, while non-complex first-order statistical (FOS) features were used with algorithms based on the Scaled …beei.org DeepRFSoC: Dataset for Modulation Classification# DeepRFSoC: Dataset for Modulation Classification Dataset ## Description The DeepRFSoC dataset provides raw in-phase and quadrature (I/Q) samples across eight modulation schemes for automatic modulation classification. The classes include: QPSK, BPSK, QAM16, QAM64, 8PSK, PAM4, GFSK, and CPFSK. Signals were transmitted and recorded on the AMD RFSoC platform, capturing the real-world effects of its RF-ADC and RF-DAC hardware, including non-idealities. The dataset also incorporates fading channels, timing offsets, and Doppler shifts, offering realistic conditions for developing and benchmarking deep learning models. Documentation and usage guidelines are available on GitHub: https://github.com/axdy/rfsoc_quant_amc [Note: the Title and Description fields for this record were updated at the request of the creator on 31/07/25] | Date made available | 2 May 2023 | | — | — | | Publisher | [University of Strathclyde](#) | ——————————————————————————–pureportal.strath.ac.uk Radio Frequency Database Construction and Modulation Recognition in Wireless Sensor Networks – PMCRadio Frequency Database Construction and Modulation Recognition in Wireless Sensor Networks – PMC [Skip to main content](#main-content) Sensors (Basel, Switzerland) logo . 2022 Jul 30;22(15):5715. doi: [10.3390/s22155715](https://doi.org/10.3390/s22155715) # Radio Frequency Database Construction and Modulation Recognition in Wireless Sensor Networks [Kun Liu](https://pubmed.ncbi.nlm.nih.gov/?term=%22Liu%20K%22%5BAuthor%5D) 1,*, [Xin Xiang](https://pubmed.ncbi.nlm.nih.gov/?term=%22Xiang%20X%22%5BAuthor%5D) 1, [Liyan Yin](https://pubmed.ncbi.nlm.nih.gov/?term=%22Yin%20L%22%5BAuthor%5D) 1 Editor: Michael Cheffena 1 PMCID: PMC9371033 PMID: [35957271](https://pubmed.ncbi.nlm.nih.gov/35957271/) ## Abstract Current modulation recognition methods in wireless sensor networks rely too much on simulation datasets. Its practical application effect cannot reach the expected results. To address this issue, in this paper we collect a large amount of real-world wireless signal data based on the software radio depmc.ncbi.nlm.nih.gov Datasets# Datasets ##### Please Read First! The datasets on this page are from early academic research work conducted in 2016/2017. They have several known errata and are NOT currently used within DeepSig products. We HIGHLY recommend researchers develop their own datasets using basic modulation tools such as MATLAB or GNU Radio, or use REAL data recorded from over the air. ###### For those seeking the latest and most reliable data, DeepSig provides a range of supported and vetted datasets exclusively available through our Model Hub—designed for customers who want access to our advanced models and proprietary datasets. Unfortunately, we are unable to provide support, revisions or assistance for these open datasets due to overwhelming demand. If you’re interested in leveraging our most up-to-date tools and data, contact info@deepsig.ai. DeepSig’s team has created several small example datasets which were used in early research from the team in modulation recognition – these are made available here for historical andeepsig.ai