End-to-end burst signal demodulation via adaptive masked deep learning framework

The bit error rate (BER) directly determines the quality of wireless communication transmission. Traditional demodulators are limited in operating on burst signals and exhibit poor BER performance in low signal-to-noise ratio (SNR) conditions. For real-world burst signals, symbol-by-symbol approaches fail to capture inter-symbol dependencies, and existing end-to-end frameworks cannot handle the variable output lengths required for burst signals. To address this issue, we propose an end-to-end demodulation framework based on deep learning (DL), in which detection, recognition, channel compensation, and demodulation stages were trained as a unified system, enabling the entire signal burst to be demodulated in a single operation during inference. The framework's generalization and robustness are enhanced by a proposed masking mechanism and a denoising autoencoder (DAE), respectively. The former dynamically adjusts the output bitstream length while preventing gradient flow from redundant components, and the latter compensates for channel fading effects. We further introduce a dedicated end-to-end training strategy to optimize the adaptation between these modules. Experimental results on real-world Frequency Shift Keying (FSK), Minimum Shift Keying (MSK), Phase-Shift Keying (PSK), and Quadrature Amplitude Modulation (QAM) signals demonstrate that the proposed framework achieves superior demodulation accuracy for long-sequence burst signals. Compared to existing methods, the proposed framework enables parallel demodulation, and dynamically adapts the output bit stream in terms of varying message types and lengths.