Detecting and Adapting to CSI Feedback Drift in Data-Driven Compression for MIMO Systems

Deep neural networks make it possible to learn key characteristics of data without having to assume mathematically tractable models. This in turn results in an ability to compress data in a model-free way. One of the promising areas for the application of deep learning to the physical layer of communication networks is compression. Hundred-fold compression with a small loss of the CSI in massive MIMO systems has been shown to be both feasible and necessary. However, model-free and data-driven compression comes with a downside: the encoding and decoding models need to be trained on a large set of CSI arrays indicative of a wide spectrum of propagation and environmental conditions. As a result, in the early stages of the deployment of deep CSI compression models, it would be necessary to detect if and when users’ channels have drifted significantly away from the distribution of the CSI data on which the deep compression model was trained. In this paper, we present both 1) a technique for detecting harmful channel drift and 2) a lightweight scheme for fine-tuning the deep compression models to adjust to such shifts. Using public-domain synthetic channel data as well as 3GPP-compliant simulated data, we demonstrate the practicality of our proposed deep compression and detection framework. We close with recommendations for a viable implementation of the proposed drift detection by the standards bodies.