Channel-aware CSI Feedback: Dataset Bias Mitigated Pre-training, Online Out-of-Distribution Detection and its Domain Adaptation
This is a repository for Channel-aware CSI Feedback: Dataset Bias Mitigated Pre-training, Online Out-of-Distribution Detection and its Domain Adaptation, a submitted manuscript to IEEE Internat of Things Journal.
The deep learning (DL)-based CSI feedback paradigm encounters significant challenges in practical wireless communication scenarios, where channel characteristics and distributions fluctuate drastically. To enhance the adaptability of the CSI feedback model, this work innovates a channel-aware CSI feedback paradigm that integrates channel awareness into CSI reporting. Firstly, the paradigm reports channel-aware information and introduces regularization in the model training phase, improving the generalizability of the pre-trained model and mitigating the bias in the mixed dataset. Secondly, the channel-aware information, through the proposed unified channel-aware out-of-distribution (OOD) detection framework in the model monitoring phase, agilely triggers model updates and distils online in-distribution (ID) data for efficient model updating. Finally, the distribution-aligned channel-aware information and synthesized online-deployment-specific features are integrated into a transductive-based hybrid domain adaptation (HDA) scheme for channel-aware CSI feedback in the model updating phase, effectively adapting the model to new scenarios with significant data drift while avoiding the catastrophic forgetting of prior knowledge. The thorough experiments verify the superior performance and adaptability of the proposed channel-aware CSI feedback paradigm in the model training, monitoring and updating phases.
- Once the manuscript is ready for early access, more details about this work will be provided.
This subsection first verifies the bias mitigation and performance of the channel-aware CSI feedback paradigm pre-trained on mixed datasets under different setups. The efficiency of the channel-aware CSI feedback to mitigate dataset bias of the offline pre-train dataset collected from 5 multiple channel configurations is evaluated, and the downlink channels are collected in a perfect downlink channel estimation mode with the setting of the following table.
This evaluation first compares the proposed channel-aware CSI feedback paradigm with baseline CsiNet and TransNet in terms of learnable parameters and Floating Point Operations per Second (FLOPs), as shown in the following table.
This subsection assesses the accuracy of the proposed unified channel-aware out-of-distribution detection framework with Softmax-score and Energy-based score functions with two online datasets setups from the IndoorHall-5GHz and the SemiUrban-300MHz, respectively, which simulate an open-world model deployment. The downlink channel is collected from the COST2100 model with channel configurations enumerated in the following table.
The TPR 95% performance of the unified channel-aware out-of-distribution detection with two setups of online deployment datasets is enumerated in the following table.
This subsection evaluates the proposed HDA scheme for online model updates. The channel-aware CSI feedback model will be updated under conditions of limited accessible training data and a restricted number of trainable iterations using an online dataset with significant data drift, which is the set of distilled ID samples from the proposed channel-aware unified out-of-distribution detection framework in the simulating online deployment dataset. The online test dataset configuration is enumerated in the following table.
The updated model with the proposed HDA is compared with conventional inductive-based online learning and the related domain adaptation approach on the channel reconstruction error and channel-aware accuracy of the prior knowledge, and the channel reconstruction error and similarity of the present knowledge. The evaluation result is presented in the following table.
