Demodulation-Oriented Neural Denoising for Long-Sequence I/Q Wireless Signals in Data-Link Receiver Chains

Reliable demodulation of low-signal-to-noise ratio (SNR) data-link signals is challenging when thermal noise, multipath fading, and structured interference overlap with the target waveform in time and frequency. This paper studies neural denoising as a front-end module in a receiver chain for long-sequence in-phase/quadrature (I/Q) wireless signals. A residual one-dimensional denoising autoencoder (DAE) is inserted before a conventional demodulator to recover the interference-free waveform from corrupted inputs. The denoised outputs are assessed using the waveform-level mean squared error (MSE), output SNR, and demodulated bit error rate (BER). The relative bit error rate (RBER) quantifies the residual BER penalty after denoising within the receiver transition region. Extensive experiments across three practical communication-signal classes and various disturbance/channel-effect conditions demonstrate that neural preprocessing significantly improves low-SNR demodulation accuracy. Crucially, the results show that traditional waveform metrics do not fully reflect receiver-level bit recovery. Compared with convolutional, attention-based, adversarial, and pooling-based baselines, the proposed receiver front end provides lower RBER in the operating region where demodulation is most sensitive to waveform distortion. The results support demodulation-oriented evaluation for neural receiver preprocessing and show that waveform restoration should be verified through the downstream bit-recovery task.