IQ-EVNet: An Eigenvalue-Aware End-to-End Framework for SNR Estimation

Accurate estimation of the signal-to-noise ratio (SNR) plays a vital role in improving the performance and reliability of modern communication systems. The eigenvalue vector (EV) derived from the covariance matrix (CM) of in phase and quadrature (IQ) signals provides a compact yet discriminative statistical representation of signal characteristics. Motivated by this property, we first propose an EV-based SNR estimation method that leverages eigenvalue representations as compact network inputs, effectively reducing feature dimensionality and improving estimation robustness and accuracy under low-SNR conditions. Building upon this foundation, we further propose an IQ-EVNet framework, which directly learns eigenvalue representations from IQ sequences and performs SNR estimation based on the learned features. This design enables end-to-end SNR estimation without explicit CM construction or eigenvalue computation. This design preserves the discriminative capacity of EV features while significantly improving inference efficiency. Simulation results demonstrate that the proposed IQ EVNet achieves notably higher estimation accuracy than IQ based methods and approaches the performance of the EV-based method, offering an excellent trade-off between accuracy and computational efficiency.