Robust Spatiotemporal Data Recovery in Wireless Sensor Networks via Fourier-Enhanced Adaptive Manifold Latent Feature Analysis

Abstract

Wireless sensor networks (WSNs) used for environmental monitoring often suffer substantial data loss due to energy and communication constraints, demanding robust recovery of missing spatiotemporal measurements. The main challenge lies in accurately reconstructing values when the observations exhibit nonlinear temporal variations and structured noise intertwined with genuine anomalies such as pollution spikes or sensor faults. Existing low-rank matrix approximation (LRMA) methods, based on linear smoothness assumptions and static regularization, struggle to distinguish structured noise from real outliers and fail to capture multiscale temporal dynamics in real-world sensor data. To address these limitations, the adaptive latent feature analysis with Fourier embedding (ALFA-FE) framework is proposed. It replaces linear graph smoothing with dynamic FEs for multiscale temporal pattern capture and substitutes fixed $L_{1}$ / $L_{2}$ penalties with Huber-loss-based optimization for robust noise–outlier handling. Extensive experiments on four real-world environmental datasets ranging from 35 to 74 sensors and up to 8600 time steps show that ALFA-FE consistently outperforms nine representative baselines, achieving up to 36% accuracy gains and tenfold speedup while preserving robustness under severe noise and data loss. By balancing accurate signal recovery and anomaly retention, ALFA-FE enhances the reliability of environmental sensing and can support next-generation IoT deployments under harsh or resource-limited conditions.