CustomMOD-2026.a: A Dataset for Geometry–Physics Guided Manifold Feature Fusion in Automatic Modulation Classification

CustomMOD-2026.a Dataset for Automatic Modulation Classification (AMC) The CustomMOD-2026.a dataset is designed for Automatic Modulation Classification (AMC) tasks, providing high-fidelity synthetic radio signals that closely simulate real-world communication environments. This dataset offers a diverse set of modulation schemes, enabling the evaluation and development of advanced modulation recognition algorithms, particularly those based on deep learning and advanced feature engineering techniques. Key Features: Modulation Types (17 Total): Includes a wide range of modulation schemes such as PSK (BPSK, QPSK, 8PSK), QAM (16QAM, 32QAM, 64QAM, 128QAM, 256QAM), APSK (16APSK, 32APSK), FSK (2FSK, 4FSK), CPM (MSK, GMSK), and differential modulation types (OQPSK, π/4-DQPSK). Signal-to-Noise Ratio (SNR) Range: Covers a broad SNR range from -20 dB to +30 dB with a 2 dB step size, providing varied channel conditions from low to high SNRs. High-Fidelity Channel Simulation: Signals are subjected to realistic channel impairments including Additive White Gaussian Noise (AWGN), random phase and frequency offsets, and root-raised cosine (RRC) pulse shaping for simulating bandwidth-limited channels. Data Format: Each signal sample is standardized to 128 samples, with each point represented as a complex I/Q pair stored as floating-point arrays in an HDF5 file. The dataset is structured with: X: Signal data (samples, 128, 2) Y: One-hot encoded modulation labels Z: Corresponding SNR values Dataset Size: 17 modulation types 26 SNR levels (from -20 to +30) 1,000 samples per modulation/SNR combination Total of 442,000 samples Design Goals & Applications: CustomMOD-2026.a is aimed at bridging the gap between simple idealized datasets and complex real-world signals. It is particularly suitable for training and testing deep learning-based modulation recognition models (such as CNN, RNN, and Transformer). The dataset supports research in advanced features, few-shot learning, and model robustness under controlled SNR and dataset conditions. It provides a standardized platform for developing new feature fusion and model integration strategies. This dataset will be invaluable for researchers and engineers working on advanced AMC techniques, providing a rich and controlled environment to test and refine modulation recognition algorithms.