Real-World IQ Dataset for Automatic Radio Modulation Recognition under Multipath Channels
Description
This dataset contains real-world complex baseband (IQ) radio signal samples intended for training and evaluation of machine learning and deep learning models for automatic modulation recognition (AMR). The dataset includes seven modulation types: BPSK, QPSK, QAM, GMSK, OFDM, NBFM, and WBFM. Signals were captured under both clean (line-of-sight) and multipath propagation conditions and generated across multiple signal-to-noise ratio (SNR) levels ranging from 20 dB to 30 dB. All 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 (clean or multipath), and SNR value. The dataset is suitable for benchmarking modulation classification performance, robustness analysis under channel impairments, and reproducible research in wireless signal processing and cognitive radio. Baseline deep learning experiments using a convolutional neural network (CNN) are provided, demonstrating classification accuracy of approximately 83% on the test set.
Files
Steps to reproduce
1. Load the HDF5 dataset files containing IQ frames and corresponding labels. 2. Normalize each IQ frame by its average signal power. 3. Split the dataset into training, validation, and test subsets. 4. Train a convolutional neural network (CNN) using the provided frame length of 1024 IQ 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. The baseline CNN architecture and training scripts are available upon request.