Variance-Guided Three-Channel Wi-Fi-BLE Fusion With Multi-Objective Domain Adaptation for Indoor Localization

Abstract

Indoor multimodal fingerprinting is constrained by limited feature expressiveness and by distribution shift induced by environmental dynamics such as pedestrian interference. This letter proposes a domain adaptive WiFi-BLE fusion framework with three coupled designs. First, WiFi and BLE received signal strength measurements are encoded as variance-guided three-channel fingerprints, which preserves modality-specific spatial structure while enabling instance-level adaptive fusion without additional trainable parameters. Second, a compact convolutional extractor with residual efficient channel attention (ECA) recalibrates channel responses at a negligible computational cost. Third, a multi-objective domain adaptation module combines adversarial alignment, first and second order statistical regularization, and exponential moving average consistency learning, followed by progressive target domain refinement at deployment. Under pedestrian interference, the full system attains an average error of 4.12 m, a median of 3.93 m, and a 95th percentile error of 7.74 m. Ablations show that removing the domain adaptation module raises the average error by 31.6%, and removing ECA alone raises it by another 8.0%.