Automatic Modulation Recognition via Autoencoder-Driven Amplitude–Phase Interaction Enhancement

Xiaoya Zuo, Zhanxu Cui, Rugui Yao, Ye Fan, Margulan Ibraimov

Computers, 2026

Full text Open access: Gold Metadata (CSL-JSON)
DOI 10.3390/computers15090567
Data confidence 85 high
How this score is calculated

It is built from checkable properties of the record itself: whether the article has a DOI, how many independent sources confirmed its data, whether the required fields are filled in, and how recent the latest confirmation is. Disagreement between sources and a missing identifier lower it. A value of 80 or above means the record was published without further checking; 60 to 79 means it is published at reduced confidence, with a manual check advisable. A record verified by an editor by hand is never pushed below 90 by a recalculation. This scores our data about the article, not the article itself.

Abstract

Automatic Modulation Recognition (AMR) is a key technology for spectrum sensing and signal demodulation in cognitive radio systems. Most existing AMR methods use raw In-phase/Quadrature (I/Q) data or direct transformations of I/Q data as neural network inputs, while the nonlinear interaction between signal amplitude and phase is not explicitly modelled. This limitation can constrain recognition performance in complex channels and high-order modulation schemes. To address this issue, an AMR method based on Amplitude/Phase (A/P) interaction enhancement is proposed. An Amplitude–Phase Interaction Autoencoder (APIAE) with a residual transition module is designed to learn amplitude–phase interaction features in the polar coordinate domain through unsupervised reconstruction, and the learned interaction features are concatenated with the raw A/P data to form an enhanced three-channel representation for downstream classifiers. Experiments on a subset of RadioML2018.01A show that all five evaluated backbone networks benefit from the proposed representation. Across four independent runs, GRU-Enhanced attains a mean best accuracy of 98.72±0.20%, CLDNN achieves a 7.39% relative gain in overall mean accuracy, and the recognition rates of 16QAM and 64QAM remain markedly improved at 4 dB. The classifier-side input adaptation increases the parameter count by less than 0.5% for all evaluated backbones, while the standalone APIAE front-end contains approximately 7.35 million parameters. The additional end-to-end single-sample inference latency relative to the raw backbone does not exceed 12 μs in the tested environment.