CrossNetMIMO‐Transformer for Joint Channel Estimation and Localization in Massive MIMO

ABSTRACT

Massive MIMO in integrated sensing and communication introduces a dimensionality curse that limits deep‐learning based joint channel estimation and localization, as convolutional architectures fail to capture long‐range inter‐antenna dependencies. This letter compares three architectural generations to address this challenge. Generation I extends a SISO CrossNet to MIMO, Generation II introduces task‐decoupled residual branches and Generation III employs angular‐delay (AD) domain tokenization with a dual‐branch transformer. Compared with Generation I, the proposed architecture reduces P90 error by 88.81% while requiring only 0.236 GFLOPS, achieving a 17.4 higher localization efficiency per GFLOP. These results demonstrate that AD sequence modelling substantially improves localization accuracy while reducing arithmetic inference cost relative to the convolutional baselines.