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

Area of Science:

• Signal Processing
• Machine Learning for Communications
• Wireless Receiver Design

Background:

• Demodulating low signal-to-noise ratio (SNR) wireless signals is difficult due to noise, fading, and interference.
• Existing receiver chains struggle with accurate signal recovery in challenging communication environments.
• Traditional waveform metrics may not fully capture the impact of preprocessing on downstream bit error rate (BER).

Purpose of the Study:

• To evaluate neural denoising as a front-end module for improving low-SNR data-link signal demodulation.
• To assess the effectiveness of a residual denoising autoencoder (DAE) in recovering interference-free waveforms.
• To compare demodulation performance using waveform-level metrics versus bit-recovery metrics.

Main Methods:

• Implemented a residual one-dimensional denoising autoencoder (DAE) as a preprocessing step before a conventional demodulator.
• Corrupted long-sequence in-phase/quadrature (I/Q) wireless signals with various disturbances and channel effects.
• Evaluated denoised outputs using mean squared error (MSE), output SNR, bit error rate (BER), and relative bit error rate (RBER).

Main Results:

• Neural preprocessing significantly enhanced demodulation accuracy for low-SNR signals across diverse communication classes.
• The proposed DAE-based front end achieved lower relative bit error rate (RBER) compared to baseline methods.
• Waveform-level metrics were found to be insufficient for fully reflecting receiver-level bit recovery performance.

Conclusions:

• Neural denoising is a viable and effective front-end solution for improving low-SNR wireless signal demodulation.
• Demodulation-oriented evaluation is crucial for assessing neural receiver preprocessing techniques.
• Waveform restoration effectiveness must be validated through downstream bit-recovery tasks.