Real-World IQ Dataset for Automatic Radio Modulation Recognition under Multipath Channels
Description
This dataset contains real-world complex baseband (IQ) radio signal samples for training and evaluating machine learning models in automatic modulation recognition (AMR). It includes seven modulation types (BPSK, QPSK, QAM, GMSK, OFDM, NBFM, and WBFM) captured at 2.4 GHz under both clean (line-of-sight) and multipath propagation conditions across signal-to-noise ratio (SNR) levels from 20 dB to 30 dB. Signals are segmented into fixed-length frames of 1024 IQ samples and stored in HDF5 format. Each frame is annotated with modulation type, channel condition, and SNR value. The dataset is suitable for benchmarking AMR performance, robustness analysis under realistic channel impairments, and reproducible research in wireless signal processing and cognitive radio. A baseline convolutional neural network (CNN) is provided, achieving approximately 84% classification accuracy on the test set.
Files
Steps to reproduce
1. Load the provided HDF5 files containing IQ frames and corresponding labels. 2. Normalize each IQ frame by its average signal power. 3. Use the provided training, validation, and test splits, or alternatively create custom splits if required. 4. Train a convolutional neural network (CNN) using IQ frames of length 1024 samples. 5. Evaluate the trained model on the test set using classification accuracy and confusion matrix metrics. All experiments were conducted using TensorFlow 2.x. Baseline training and evaluation scripts are available upon request.
Institutions
- Ural'skij federal'nyj universitet imeni pervogo Prezidenta Rossii B N El'cina Institut radioelektroniki i informacionnyh tehnologijSverdlovskaa ooblast', Ekaterinburg