THz UM-MIMO system channel estimation algorithm based on deep residual block fixed-point network

To mitigate the channel estimation challenges induced by hybrid near-far field and beam squint effects in THz ultra-massive MIMO systems, a deep learning-based FPN-OAMP-SRLG algorithm was proposed. A feature extraction network SRLG was constructed by developing a deep residual block (BSRB) with coordinate attention and partial channel shift, along with a gated linear self-attention module (SARG). The channel estimation problem was formulated as an image restoration task through integration with the FPN-OAMP framework. The algorithm utilized pilot information, estimated via the least squares method, as input features and recovered channel state information through iterative linear and nonlinear estimators. Simulation results demonstrate that the proposed algorithm achieves high-precision THz channel estimation, exhibiting fast convergence and robust performance.