ChangShuoRadioData (CSRD) Simulation Framework

• ChangShuoRadioData (CSRD) is an open-source simulation framework that produces terabyte-scale, perfectly labeled RF data for modern AI-driven spectrum sensing.
• It integrates Transmitter, Channel, and Receiver modules in a configurable JSON-driven pipeline to simulate diverse modulation schemes and realistic RF impairments.
• The CSRD2025 benchmark, with over 25 million IQ frames and 200 TB of data, enables reproducible research and advanced cognitive radio applications.

Searching arXiv for the cited CSRD papers to ground the article in the relevant literature. ChangShuoRadioData (CSRD) is an open-source, end-to-end simulation framework and associated dataset for generating large-scale synthetic radio frequency data for spectrum sensing and wireless signal analysis. It was introduced to provide the scale, diversity, realism, and perfect labeling required by modern AI models for wireless communications, especially Large AI Models for complex tasks such as spectrum sensing. Its flagship benchmark, CSRD2025, comprises over 25,000,000 passband IQ frames, approximately 200 TB of aggregate IQ data, supports 100 modulation types spanning analog, digital single-carrier, OFDM, and OTFS, covers SISO/MISO/SIMO/MIMO antenna configurations up to $4\times 4$, and provides standardized $8:1:1$ training, validation, and test splits by frame index for reproducible research (Chang et al., 27 Aug 2025).

1. Historical placement and problem setting

CSRD was developed in response to a central bottleneck in AI-driven wireless communications: the need for vast, diverse, and realistic labeled RF corpora. The framework is explicitly oriented toward spectrum sensing and management, where multiple transmitters, realistic propagation, RF front-end nonidealities, and signal overlap must be represented jointly rather than in isolation (Chang et al., 27 Aug 2025).

An earlier stage of the same line of work appeared in the study of joint signal detection and automatic modulation classification, where a coexisting synthetic dataset was generated for a multiple-signal environment and released through the same repository. In that context, the dataset is described as CSRD, also referred to as CRML23, and was designed to address the limitation of publicly available AMC datasets that ignore the signal detection step and contain only one signal (Xing et al., 2024). That earlier formulation simulated up to $n$ independent transmitters, each using one of five digital modulations, and represented the received I/Q samples as

$$x[l] = \sum_{i=1}^n h_i\, e^{\,j\bigl(2\pi f_i\,l + \theta_i\bigr)}\, g_i[l] + w[l],$$

with Rayleigh or Rician fading, random phase offset, and AWGN (Xing et al., 2024).

CSRD2025 generalizes that earlier synthetic-data agenda to a substantially larger and more heterogeneous benchmark. The later framework moves from a narrowly defined coexisting-signal corpus to a modular platform capable of generating terabyte-scale datasets with broader waveform families, richer channel models, receiver and transmitter impairment models, and standardized metadata and annotation formats (Chang et al., 27 Aug 2025).

2. Framework architecture and generation workflow

CSRD is organized around three principal modules—Transmitter, Channel, and Receiver—combined in a highly configurable, JSON-driven pipeline that can be scaled via parallel execution to produce millions of labeled RF frames (Chang et al., 27 Aug 2025).

The Transmitter Module accepts a source message consisting of pseudo-random bits for digital waveforms or real audio for analog waveforms. It implements approximately 100 modulation schemes, including AM/SSB/VSB/FM/PM, ASK, PSK, QAM, FSK, CPM, OFDM, SCFDMA, and OTFS. Configuration parameters include pulse shaping such as RRC, symbol rate, carrier frequency, and MIMO coding such as OSTBC. A configurable impairment submodule introduces transmitter nonidealities including polynomial, Saleh, and Rapp nonlinearities, IQ imbalance, DC offset, and phase noise via MATLAB’s comm.MemorylessNonlinearity and custom spectral-mask models (Chang et al., 27 Aug 2025).

The Channel Module applies either statistical fading or site-specific ray tracing. In the statistical mode, distance-based path loss and multipath fading are modeled using Rayleigh or Rician processes with path delays $\tau_\ell$, gains $\alpha_\ell$, and a Rician $K$-factor. In the ray-tracing mode, CSRD ingests 3D building geometry from OpenStreetMap .osm files for 25 environment types and computes a channel impulse response of the form

$$H(\tau)=\sum_{\ell=1}^{L}\alpha_\ell e^{-j2\pi f_c \tau_\ell}\delta(\tau-\tau_\ell).$$

Path loss is modeled as

$PL(d)=PL_0+10\,n\log_{10}(d/d_0)+X_\sigma,$

where $n$ is the path-loss exponent and $8:1:1$0 is shadowing. Delay spread is given by

$8:1:1$1

The Receiver Module sums the channel outputs across transmitters and then adds receiver RF impairments, including IQ imbalance, DC offset, LNA nonlinearity, and AWGN

$8:1:1$2

Ground-truth SNR per link is computed before receiver impairments as

$8:1:1$3

Data generation is orchestrated by a high-level Runner. For each frame, it samples the number of transmitters and receivers, from 1 to 4 each, together with modulation types, channel parameters, impairment levels, symbol rates, and event timings in order to simulate bursty and overlapping signals. All parameters are specified in hierarchical JSON configuration files, and the MATLAB-based runner archives each frame’s IQ in a .mat file while emitting a SigMF-style JSON that records every aspect of the simulation (Chang et al., 27 Aug 2025).

3. CSRD2025 dataset specification

CSRD2025 is the reference dataset generated by CSRD using fixed random seeds and the “default” configuration. Its scale is summarized by

$8:1:1$4

yielding over 25 million frames and an aggregate IQ data size of approximately 200 TB. The dataset is described as roughly $8:1:1$5 larger than the 18 GB RadioML 2018.01A, and approximately $8:1:1$6 times larger than the widely used RML2018 dataset (Chang et al., 27 Aug 2025).

The modulation inventory comprises 100 total classes. These are partitioned into analog waveforms, digital single-carrier schemes, OFDM/SCFDMA families, and OTFS waveforms. The analog set includes AM-DSB/SSB/VSB, FM, and PM; the digital single-carrier set includes ASK/OOK, PSK from 2 to 64, QAM from 8 to 4096, FSK from 2 to 8, and CPM variants such as GMSK, MSK, and CPFSK. Class instance counts are highly nonuniform: common schemes such as 2-OOK and 2-GMSK occur at approximately $8:1:1$7 instances, whereas rare high-order QAMs occur at approximately $8:1:1$8 instances (Chang et al., 27 Aug 2025).

Antenna configurations are randomly chosen among SISO, MISO, SIMO, and MIMO up to $8:1:1$9. The standardized split is $n$0 by frame index, with split files provided for reproducibility. The framework description also states this as a standardized $n$1 training, validation, and test partition (Chang et al., 27 Aug 2025).

Dataset attribute	Specification	Notes
Total size	Over 25 million frames, $n$2 TB	Fixed random seeds, “default” configuration
Modulation classes	100 total classes	Analog, digital single-carrier, OFDM/SCFDMA, OTFS
Antenna settings	SISO, MISO, SIMO, MIMO up to $n$3	Randomly chosen
Split protocol	$n$4 by frame index	Split files provided

The earlier CRML23/CSRD release used for joint detection and AMC was much smaller and structurally different. It employed five equally probable digital modulations—BPSK, QPSK, 8-PSK, 16-QAM, and 64-QAM—with sample rate $n$5 Hz and $n$6 samples per entry, under per-signal SNRs from 12 dB to 30 dB in 2 dB steps (Xing et al., 2024). This earlier configuration is useful for understanding the lineage of CSRD, but it should not be conflated with the broader CSRD2025 specification.

4. Propagation, coexistence, and RF impairment models

A defining feature of CSRD is the simultaneous inclusion of channel realism, front-end nonidealities, and controlled spectral coexistence. This combination is central to the framework’s stated aim of bridging the Sim2Real gap in spectrum sensing (Chang et al., 27 Aug 2025).

For statistical fading, CSRD uses Rayleigh and Rician models via MATLAB’s COMM toolbox. Path delays $n$7 are log-uniform, gains $n$8 are exponentially decaying, the Rician $n$9-factor is uniform over $$x[l] = \sum_{i=1}^n h_i\, e^{\,j\bigl(2\pi f_i\,l + \theta_i\bigr)}\, g_i[l] + w[l],$$0, and Doppler shift $$x[l] = \sum_{i=1}^n h_i\, e^{\,j\bigl(2\pi f_i\,l + \theta_i\bigr)}\, g_i[l] + w[l],$$1 is uniformly derived from speeds $$x[l] = \sum_{i=1}^n h_i\, e^{\,j\bigl(2\pi f_i\,l + \theta_i\bigr)}\, g_i[l] + w[l],$$2–$$x[l] = \sum_{i=1}^n h_i\, e^{\,j\bigl(2\pi f_i\,l + \theta_i\bigr)}\, g_i[l] + w[l],$$3 m/s. For site-specific ray tracing, an internal SBR/Image-Method engine uses OSM geometry to compute up to $$x[l] = \sum_{i=1}^n h_i\, e^{\,j\bigl(2\pi f_i\,l + \theta_i\bigr)}\, g_i[l] + w[l],$$4 rays, producing

$$x[l] = \sum_{i=1}^n h_i\, e^{\,j\bigl(2\pi f_i\,l + \theta_i\bigr)}\, g_i[l] + w[l],$$5

which is then convolved with the modulated waveform. The resulting path loss and angular information are intended to reinforce realism in urban, suburban, indoor, and rural settings (Chang et al., 27 Aug 2025).

The IQ imbalance model is expressed by taking a baseband signal $$x[l] = \sum_{i=1}^n h_i\, e^{\,j\bigl(2\pi f_i\,l + \theta_i\bigr)}\, g_i[l] + w[l],$$6 and forming

$$x[l] = \sum_{i=1}^n h_i\, e^{\,j\bigl(2\pi f_i\,l + \theta_i\bigr)}\, g_i[l] + w[l],$$7

or equivalently

$$x[l] = \sum_{i=1}^n h_i\, e^{\,j\bigl(2\pi f_i\,l + \theta_i\bigr)}\, g_i[l] + w[l],$$8

where $$x[l] = \sum_{i=1}^n h_i\, e^{\,j\bigl(2\pi f_i\,l + \theta_i\bigr)}\, g_i[l] + w[l],$$9 and $\tau_\ell$0 are determined by amplitude error $\tau_\ell$1 and phase error $\tau_\ell$2. The specified impairment range includes transmitter and receiver amplitude error $\tau_\ell$3 and phase error $\tau_\ell$4 (Chang et al., 27 Aug 2025).

Phase noise is modeled through a spectral mask $\tau_\ell$5 that injects a random phase process $\tau_\ell$6, replacing the transmitted carrier $\tau_\ell$7 by

$\tau_\ell$8

The phase-noise level spans $\tau_\ell$9 to $\alpha_\ell$0 dBc/Hz at 10 kHz (Chang et al., 27 Aug 2025).

Nonlinear amplification is represented either by a memoryless polynomial,

$\alpha_\ell$1

or by the Saleh model,

$\alpha_\ell$2

CSRD specifies PA/LNA nonlinearity using Cubic, Saleh, and Rapp models with $\alpha_\ell$3 dBm, and receiver noise figure from 10 to 20 dB (Chang et al., 27 Aug 2025).

AWGN is modeled as $\alpha_\ell$4, with $\alpha_\ell$5 set by receiver noise figure and thermal temperature. Ground-truth SNR spans roughly $\alpha_\ell$6 dB to $\alpha_\ell$7 dB (Chang et al., 27 Aug 2025). A plausible implication is that CSRD is designed to cover both weak-signal and high-SNR regimes within a single corpus rather than specializing only in conventional AMC ranges.

5. Metadata, annotations, and task interfaces

CSRD stores raw .mat IQ files together with SigMF-style JSON metadata that describes every transmitter and receiver parameter, including modulation, impairments, channel taps, SNRs, timing, and bandwidth (Chang et al., 27 Aug 2025). This metadata is intended to preserve exact simulation provenance and enable perfect ground truth for supervised learning.

For time-frequency learning, the framework provides STFT-derived spectrograms, with a Hamming window by default, and COCO-format annotations. Each signal instance is mapped to a bounding box with coordinates $\alpha_\ell$8 corresponding to time extent $\alpha_\ell$9 and frequency range $K$0, together with a class label such as “4-GFSK.” Repository scripts perform batch conversion from IQ to spectrogram plus COCO JSON (Chang et al., 27 Aug 2025).

These annotation layers make CSRD directly usable for classification, detection, and localization tasks. The earlier CRML23 release illustrates the same design philosophy in a narrower setting. There, each entry in CSRD/entries/ stored iq with shape $K$1, gt_boxes with shape $K$2, and metadata fields such as snr, channel_type, k_factor, doppler, and clock_offset; train/validation/test lists were provided under splits/, and anchors.npy stored the three detector anchor bandwidths (Xing et al., 2024).

In the joint detection–classification formulation, a detector produces proposals

$K$3

where $K$4 is predicted center frequency, $K$5 is predicted bandwidth, and $K$6 is a confidence score interpreted as estimated IoU with respect to a ground-truth box. Those proposals are then mixed down, low-pass filtered, and passed to an AMC network, yielding outputs of the form $K$7 (Xing et al., 2024). This earlier proposal interface clarifies how CSRD’s metadata and frequency-localized annotations can support end-to-end pipelines rather than only standalone classifiers.

6. Sim2Real strategy, benchmarking, and research use

CSRD’s Sim2Real strategy is explicitly built around three mechanisms: realistic RF impairments with stochastic variation, site-specific ray tracing based on OSM-derived propagation, and spectral coexistence with controlled overlap percentages to simulate realistic interference (Chang et al., 27 Aug 2025). A common simplification in earlier synthetic AMC corpora is to ignore detection, coexistence, and front-end distortion; the earlier CRML23 work was already positioned against that simplification by generating a multiple-signal coexisting scenario rather than a one-signal entry format (Xing et al., 2024).

Preliminary benchmarks reported for CSRD2025 use common CNN and object-detector architectures, including ResNet-based classifiers and Faster R-CNN on spectrograms. These experiments report classification accuracies exceeding $K$8 at $K$9 dB across 100 classes and $$H(\tau)=\sum_{\ell=1}^{L}\alpha_\ell e^{-j2\pi f_c \tau_\ell}\delta(\tau-\tau_\ell).$$0 above $$H(\tau)=\sum_{\ell=1}^{L}\alpha_\ell e^{-j2\pi f_c \tau_\ell}\delta(\tau-\tau_\ell).$$1 for time-frequency localization on the test split. The same discussion states that models trained on CSRD generalize well to limited OTA data in initial trials (Chang et al., 27 Aug 2025). This suggests that the benchmark is intended not merely as a synthetic pretraining source but as a testbed for transfer from simulated RF environments to over-the-air conditions.

The earlier JDM benchmarks in the five-class coexisting setting provide a more granular view of task difficulty. Detection mAP was reported as approximately $$H(\tau)=\sum_{\ell=1}^{L}\alpha_\ell e^{-j2\pi f_c \tau_\ell}\delta(\tau-\tau_\ell).$$2 at 30 dB AWGN and approximately $$H(\tau)=\sum_{\ell=1}^{L}\alpha_\ell e^{-j2\pi f_c \tau_\ell}\delta(\tau-\tau_\ell).$$3 at 12 dB, with a drop of about 10 points under full simulated fading and offsets. AMC accuracy was modulation-dependent, with BPSK near $$H(\tau)=\sum_{\ell=1}^{L}\alpha_\ell e^{-j2\pi f_c \tau_\ell}\delta(\tau-\tau_\ell).$$4 at 30 dB and near $$H(\tau)=\sum_{\ell=1}^{L}\alpha_\ell e^{-j2\pi f_c \tau_\ell}\delta(\tau-\tau_\ell).$$5 at 12 dB, while 64QAM was near $$H(\tau)=\sum_{\ell=1}^{L}\alpha_\ell e^{-j2\pi f_c \tau_\ell}\delta(\tau-\tau_\ell).$$6 at 30 dB and near $$H(\tau)=\sum_{\ell=1}^{L}\alpha_\ell e^{-j2\pi f_c \tau_\ell}\delta(\tau-\tau_\ell).$$7 at 12 dB. The joint framework accuracy ranged from approximately $$H(\tau)=\sum_{\ell=1}^{L}\alpha_\ell e^{-j2\pi f_c \tau_\ell}\delta(\tau-\tau_\ell).$$8 at 30 dB to approximately $$H(\tau)=\sum_{\ell=1}^{L}\alpha_\ell e^{-j2\pi f_c \tau_\ell}\delta(\tau-\tau_\ell).$$9 at 12 dB, and conventional pipelines such as MF+DT and MF+SVM remained below $PL(d)=PL_0+10\,n\log_{10}(d/d_0)+X_\sigma,$0 even at high SNR (Xing et al., 2024). These results are not directly comparable to CSRD2025 because the task definitions and class sets differ, but they document the repository’s earlier benchmark lineage.

CSRD and CSRD2025 are intended for AI-driven spectrum sensing, cognitive radio, interference detection, and modulation classification research. The framework’s modularity allows custom datasets to be generated by editing JSON configurations for frequency bands, waveform libraries, and environment types. Although the 200 TB CSRD2025 instance is not directly hosted, the framework states that full reproducibility is guaranteed through public code, fixed seeds, and detailed configuration files in the GitHub repository (Chang et al., 27 Aug 2025).