Unsupervised Feature Aligned Domain Adaptation for WiFi-Based Human Activity Recognition

While WiFi-based human activity recognition (HAR) learning models can achieve good accuracy in one environment (domain), they typically experience a substantial drop in accuracy in new environments where spatial, human, and other physical factors have changed. This work proposes an unsupervised domain adaptation approach for WiFi-based HAR that integrates feature alignment with domain adversarial neural networks (termed FA-DANN). We evaluate FA-DANN on two publicly available multi-environment WiFi datasets: GJWiFi and OPERAnet. Our baseline HAR model, built with a CNN and Attention-based BiLSTM (CNN-ABiLSTM) experiences an anticipated drop when tested in new environments with an average $\mathbf{F 1}$-score of $\mathbf{1 4. 3 4 \%}$ across both datasets and rises to an average $\mathbf{F 1}$-score of $\mathbf{8 5. 5 7 \%}$ across both datasets after applying FA-DANN, with traditional validation and an average F1-score of 76.48% across both datasets using Leave-One-Subject-Out Cross Validation (LOSOCV). We demonstrate that FA-DANN outperforms the state-of-the-art domain adversarial neural network (DANN) approach across both datasets by an average increase in F1-score of 19.98% using traditional validation and 17.67% using LOSOCV. Our work is a significant step in addressing the challenge of creating WiFi-based HAR models that can transfer to new environments / new subjects without the need to collect expensive labeled data for each new scenario.