Active Sensing-Driven Deep Learning for Robust XL-MIMO Hybrid Precoding
Abstract
Near-field extremely large-scale multiple-input multiple-output (XL-MIMO) channels are characterized by spherical wavefront propagation and spatial non-stationarity. These characteristics cause significant performance degradation in angular-domain codebook-based schemes due to basis mismatch, while simultaneously introducing more effective degrees-of-freedoms (EDoFs). To fully exploit these EDoFs, this paper proposes an active sensing framework for time-division duplex point-to-point XL-MIMO systems to enable efficient hybrid precoding without explicit channel state information. By exchanging multi-round ping-pong pilots, the receiver aggregates low-dimensional observations into a continuous hidden state using a closed-form continuous-time neural network, achieving robust temporal feature extraction under noise. Based on the learned state, a Chebyshev polynomial-based Kolmogorov-Arnold network provides an expressive mapping from the hidden state to beamformers. Furthermore, QR decomposition, adaptive attention fusion, and an off-diagonal suppression objective are jointly integrated to stabilize multi-beam learning and encourage alignment with the channel’s dominant singular vectors. Simulation results demonstrate that the proposed framework maintains near-optimal spectral efficiency even at a low signal-to-noise ratio of −10 dB or under out-of-distribution channel conditions, while the learned beams exhibit minimal power leakage and strong interpretability.