Scenario-Aware Framework for DL-Based CSI Feedback With Unified Training, Monitoring, and Updating

The deep-learning-based (DL-based) channel state information (CSI) feedback faces significant challenges in open-world wireless communication systems, where CSI data are characterized by multiscenario diversity and high dynamics. These properties introduce distribution bias and distribution shift issues: the former leads to overfitting during model training, while the latter causes model degradation during deployment and catastrophic forgetting during updates. To address these challenges across the artificial intelligence (AI) lifecycle, this work proposes a unified scenarios-aware CSI feedback framework that operates throughout the training, monitoring, and updating phases. It includes: bias-resilient model training, which introduces CSI scenario awareness to enhance CSI reconstruction while mitigating overfitting; learnable-free out-of-distribution (OOD) detection for model monitoring, which identifies distribution shifts and enables efficient updates via in-distribution (ID) samples filtering; and forgetting-resistant model updating via a hybrid domain adaptation (HDA) strategy, which retains knowledge of known scenarios while improving CSI reconstruction in unseen scenarios. Extensive experiments demonstrate the superiority of the scenario-aware CSI feedback framework: it achieves up to 4.65 db normalized mean squared error (NMSE) improvement over its prototype in random bias setups dataset, enables responsive OOD detection using both Softmax and energy-based confidence functions with an average gain of 1 dB after updates on filtering ID samples, and facilitate adaptation to unseen scenarios (up to 0.30 similarity improvement) while preserving known knowledge (above 97.6% scenario-aware accuracy) even under significant scenario shifts.