Disentangled Representation Learning Empowered CSI Feedback Using Implicit Channel Reciprocity in FDD Massive MIMO
Wei Xu, Jie Wu, Shi Jin, Xiaohu You, Zhaohua Lu
Abstract
Channel state information (CSI) compression and feedback is a common way of acquiring the CSI at the transmitter in frequency division duplex (FDD) massive multiple-input multiple-output (mMIMO) systems due to the lack of channel reciprocity. However, implicit reciprocity potentially exists in the bi-directional channels of an FDD system because they in fact share physically the same propagation paths. We propose to leverage this implicit reciprocity in FDD mMIMO systems to minimize the feedback overhead and enhance the CSI recovery with uplink channel information at the transmitter. To achieve this, we develop a disentangled representation (DR) learning enabled neural network (NN), named DrCsiNet, to realize the selective CSI compression feedback with the assistance of uplink CSI. The proposed DrCsiNet successfully extracts the information of reciprocity implicitly shared between the downlink and uplink channels in FDD mMIMO, while it simultaneously extracts selective information from the downlink CSI excluding the implicit reciprocity component for compression feedback. We conduct extensive simulations to evaluate the performance of the proposed DrCsiNet against existing methods under various setups. Results demonstrate remarkable performance gains of DrCsiNet for CSI recovery and evidence a strong generalization ability across various network structures. These findings validate the efficacy of disentangling implicit CSI reciprocity embedded in uplink CSI for enhancing the downlink CSI recovery in FDD mMIMO.