Signal recovery and bit error rate prediction in high-speed fiber optic communication based on improved transformer and performance-feedback-driven WDM configuration-embedding refinement

In high-speed dense wavelength division multiplexing (WDM) fiber optic systems, cross-channel nonlinear coupling and dynamic link configurations induce significant non-stationarity in signal distortion mechanisms and in the statistical characteristics of bit error rate (BER), making it difficult for existing methods to achieve stable signal recovery and reliable BER prediction within a unified model. This paper proposes the Performance-Constrained WDM-Aware Transformer Network (PC-WATNet), a joint modeling framework that embeds WDM configuration as a learnable variable jointly refined with encoder weights through gradient signals from both symbol recovery and BER prediction objectives. Frequency-dependent dispersion and cross-channel nonlinear coupling constraints are incorporated into self-attention computation, and a dual-output decoding structure completes symbol-level recovery and BER prediction within a shared feature space, with BER prediction errors propagating gradients back to reshape shared representations. Under multi-scenario simulations, PC-WATNet achieves an error vector magnitude of 7.12% and an effective signal-to-noise ratio of 22.95 dB under center-channel conditions, with a BER prediction mean squared error of 1.35 × 10 −4 under joint modeling conditions. These results indicate that the proposed framework provides a unified solution for the collaborative modeling of signal recovery and performance prediction in complex WDM fiber optic systems under simulation-based conditions.