Scholarlink

<p>In recent years, Wi-Fi-based human activity recognition (HAR) has achieved remarkable progress. However, models trained in specific environments often experience performance degradation when applied to unseen environments. To address this challenge, we propose a device-free cross-domain activity recognition approach that combines multi-feature fusion with federated learning (FL). Given the high heterogeneity of channel state information (CSI) across different environments, the quality of feature samples is often affected by factors such as acquisition conditions and channel interference, making it difficult for a single feature to capture complex activity patterns comprehensively. To this end, we propose an uncertainty-aware feature fusion method that predicts the true class probability (TCP) to assign adaptive fusion weights to the amplitude, phase, and Doppler features of CSI, thereby enhancing representation capability and mitigating the impact of low-quality features. Meanwhile, we employ an FL framework and optimize the aggregation strategy to address the distribution differences across environments, enhancing the generalization ability of the global model in various environments without the need for additional data collection from the target domain, thus achieving good recognition performance. Furthermore, the FL framework allows each client to train locally, thereby preserving user data privacy. We evaluate the proposed method on the Widar3.0 dataset and two self-collected datasets in cross-domain recognition tasks, achieving average recognition accuracies of 96.5%, 92.1%, and 98.2%, respectively, and outperforming the current state-of-the-art methods.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Cross-Domain Wi-Fi-Based Activity Recognition System via Federated Learning and Feature Fusion

  • 期刊论文

  • 发表日期：

    2025年9月29日

查看全文
  • Xin Wang,
  • Mingquan Shi,
  • Juming Li

摘要

In recent years, Wi-Fi-based human activity recognition (HAR) has achieved remarkable progress. However, models trained in specific environments often experience performance degradation when applied to unseen environments. To address this challenge, we propose a device-free cross-domain activity recognition approach that combines multi-feature fusion with federated learning (FL). Given the high heterogeneity of channel state information (CSI) across different environments, the quality of feature samples is often affected by factors such as acquisition conditions and channel interference, making it difficult for a single feature to capture complex activity patterns comprehensively. To this end, we propose an uncertainty-aware feature fusion method that predicts the true class probability (TCP) to assign adaptive fusion weights to the amplitude, phase, and Doppler features of CSI, thereby enhancing representation capability and mitigating the impact of low-quality features. Meanwhile, we employ an FL framework and optimize the aggregation strategy to address the distribution differences across environments, enhancing the generalization ability of the global model in various environments without the need for additional data collection from the target domain, thus achieving good recognition performance. Furthermore, the FL framework allows each client to train locally, thereby preserving user data privacy. We evaluate the proposed method on the Widar3.0 dataset and two self-collected datasets in cross-domain recognition tasks, achieving average recognition accuracies of 96.5%, 92.1%, and 98.2%, respectively, and outperforming the current state-of-the-art methods.