Graph-Conditioned Active Sensing for Failure-Recovery Placement in Sparse Spatiotemporal Sensor Networks

Abstract

Sparse spatiotemporal sensor networks are vulnerable to spatially structured monitoring failures, in which geographically related sensors become unavailable while remaining observations must support budget-limited recovery decisions. Existing work has primarily addressed spatiotemporal forecasting, greenfield sensor placement, or large-domain monitoring design; less attention has been given to recovery-oriented placement in already deployed sparse networks. This article presents Failure-Recovery Graph-Aligned Learning (FR-GAL), a decision-aligned virtual deployment framework for evaluating recovery placement under structured monitoring failures. Each failure task partitions a retained network into context sensors, recoverable candidate sensors, and hidden target sensors. A graph-conditioned covariance surrogate estimates target-region uncertainty, established acquisition rules are evaluated under a common greedy reveal-and-update protocol, and acquisition scores are compared with realized reductions in target-region reconstruction error. The framework is evaluated on a real regulatory fine particulate matter (PM2.5) monitoring network using 2024 California observations, with 2023 data used as a temporal stress test. In the primary 2024 clustered-failure experiment, target-region uncertainty-reduction rules, represented by DeltaVar and JointMI, produced the strongest recovery pattern. DeltaVar achieved target-region root mean squared error values of 3.928 and 3.806 at budgets of five and ten selected sensors and showed strong oracle-gain alignment, including Spearman correlation of 0.659 and normalized discounted cumulative gain at rank five of 0.872. Additional empirical-missingness, regional-network, outdoor-temperature sensing-domain, reconstruction-backbone, surrogate-family, topology, and temporal analyses delimit the supported claim: FR-GAL is supported most strongly for spatially structured failure recovery, not for arbitrary missingness or universal sensor placement.