Carrier-Frequency-Offset (CFO)-aware Deep Learning-based Orthogonal Frequency-Division Multiplexing (OFDM) Detection under a 60 GHz Clustered Delay Line Channel

Abstract

—This paper evaluates classical, end-to-end deep learning, and hybrid Orthogonal Frequency-Division Multiplexing (OFDM) receivers under a 60 GHz Third-Generation Partnership Project Technical Report 38.901 Clustered Delay Line Model A (3GPP TR 38.901 CDL-A) channel with explicit synchronization impairments. Four receivers are compared: no equalization, pilot-assisted least-squares channel estimation with Least-Squares Zero-Forcing (LS-ZF), an end-to-end regression Deep Neural Network (DNN), and a hybrid Least-Squares-aided Deep Neural Network (LS-DNN) detector. The main contribution is a Carrier-Frequency-Offset (CFO)-aware comparison of whether a simple fully connected DNN is more effective as a direct receiver replacement or as a post-equalization refinement stage. The impairment model includes frame-wise common phase error and deterministic normalized CFO. Performance is evaluated from 0−50 dB using Bit Error Rate (BER), symbol Mean Squared Error (MSE), Error Vector Magnitude (EVM), receiver runtime, and Wilson confidence intervals over 𝝐 ∈ {0, 0.01, 0.02, 0.05, 0.06, 0.10}. Results indicate that LS-ZF remains the strongest BER baseline, while the hybrid detector is BER-competitive at low-to-moderate CFO and improves symbol-fidelity metrics in several CFO-impaired cases. At 𝝐 = 𝟎. 𝟏𝟎 , all receivers degrade sharply, showing limited robustness outside the DNN CFO training range. Overall, the results suggest that deep learning is more useful as a structured post-equalization refinement stage than as a universal end-to-end OFDM receiver replacement