Underfitting Autoencoder-Based Unsupervised Framework for Real-Time Channel Quality Assessment in HF-OFDM Systems

Abstract

High-frequency orthogonal frequency division multiplexing (HF-OFDM) systems demand precise and real-time channel state information (CSI) for adaptive modulation and coding (AMC), yet the rapid temporal dynamics and scarcity of labeled data inherent in ionospheric channels render conventional assessment methods inadequate. This paper proposes a novel unsupervised framework for fine-grained channel quality assessment that leverages a deliberately underfitting autoencoder (AE) coupled with $k$-means clustering. The core insight is that, by restricting the number of training epochs, the AE is maintained in an underfitting regime where the reconstruction error of pilot signals exhibits markedly different convergence rates across channel quality classes: signals traversing benign channels converge rapidly toward the pilot target, whereas those suffering severe impairments converge slowly. The per-sample in-phase and quadrature reconstruction errors are combined into a two-dimensional feature vector, which is then classified into four quality levels via $k$-means. Simulation experiments conforming to the MIL-STD-188-110 standard demonstrate that the proposed framework achieves an overall classification accuracy of $\mathbf{8 9. 2 5 \%}$-representing a $\mathbf{1 1 6. 9 \%}$ improvement over direct clustering on raw features-with an end-to-end inference latency below 2 ms, well within the coherence time of typical HF channels. These results confirm that the proposed framework provides a practical, label-free, high-precision, and low-latency CSI acquisition solution for HF-OFDM AMC systems.