Transformer-KF Aided Sparse mmWave Channel Estimation for Complex Dynamic Scenarios

Millimeter-wave (mmWave) MIMO systems provide large bandwidth and high array gain, but reliable channel estimation is challenged by high-dimensional arrays, severe path loss, blockage, and dynamic dominant-path switching. Sub-6 GHz channels are easier to observe and can provide out-of-band angular priors for mmWave sparse recovery. However, Root-MUSIC angle observations may contain outliers under low SNR, blockage, and path switching. This paper proposes a Sub-6 GHz aided sparse mmWave channel estimation method. Sub-6 GHz pilots are first processed by least-squares channel estimation and Root-MUSIC angle observation. A gated residual Transformer-Kalman filter (TF-KF) then tracks the angular prior, and the resulting angle and confidence are transferred to confidence-aided AWP-AMP for mmWave sparse channel recovery. The proposed TF-KF does not replace the Kalman filter; instead, it learns Root-MUSIC observation confidence, bounded residual compensation, and measurement-covariance scaling while preserving the recursive KF structure. In a representative fixed-seed test trajectory, TF-KF reduces the average AoA/AoD RMSE from 14.903 deg for conventional KF to 2.937 deg. In sparse recovery, the TF-KF prior with a strong fixed threshold coefficient lambda = 0.5 obtains an average NMSE of -1.5647 dB, almost identical to the learned threshold-calibration result of -1.5646 dB. These results indicate that the dominant performance gain comes from reliable confidence-aware angular priors, while AMPParamNet mainly acts as a lightweight threshold-calibration module in the current setting.