Split-BKNet: A Decomposed Bussgang Gain Learning for Robust State Estimation with 1-Bit Observations

The severe nonlinearity of 1-bit quantization poses a significant challenge in state estimation. While the Bussgang-aided KalmanNet (BKNet) handles imperfect models, its monolithic learning of the Bussgang gain suffers from training instability in complex dynamic environments. This letter proposes Split-BKNet, which decomposes the gain learning into two dedicated modules for ‘state error covariance’ and ‘observation innovation mapping’. This split architecture enhances training stability and robustness, and simulations demonstrate that it consistently outperforms BKNet under model mismatch and heterogeneous noise conditions.