摘要
This paper investigates robust distributed cooperative tracking of maneuvering targets in degraded sensor networks. We propose TIMM–RWIC, a resilient two-layer hierarchical framework that integrates a node-level Transformer-enhanced interacting multiple model (TIMM) filter and a network-level reliability-weighted information consensus (RWIC) module. At the node level, TIMM leverages a Transformer architecture to extract short-window temporal features, replacing traditional likelihood-based model probability updates with data-driven posterior estimation, thereby improving maneuvering-mode discrimination and reducing local estimation errors. At the network level, RWIC dynamically evaluates node reliability based on posterior confidence, measurement dropout rates, and neighbor-referenced residuals, enabling adaptive consensus weighting and robust information fusion under sensor degradation. Theoretical analysis of estimation error bounds is provided based on stochastic process theory. Extensive numerical simulations and UAV flight experiments demonstrate the effectiveness of the proposed method against measurement dropouts, noise bursts, and partial sensor failures. Compared to the classical distributed baseline DEIF-IMM, the proposed TIMM–RWIC reduces position and velocity ARMSEs by 81.5% and 78.8%, respectively. Note to Practitioners—This work addresses the practical challenge of robustly tracking maneuvering targets in distributed multi-agent networks, where sensor degradation and intermittent communication are common. Traditional target tracking methods often struggle under sensor dropouts or noise bursts, leading to performance degradation and inaccurate estimates. To address this, we propose TIMM–RWIC, a two-layer framework designed for implementation on distributed agents. At the node level, TIMM uses a Transformer to extract short-window temporal features and perform data-driven model-probability updates, improving maneuver discrimination under degraded measurements. At the network level, RWIC assigns adaptive consensus weights using posterior confidence, measurement dropout statistics, and neighbor-referenced residuals so that unreliable nodes are automatically down-weighted rather than contaminating the network estimate. Our experiments, conducted on a cooperative target-tracking experimental platform with UWB sensors, show that the proposed method improves tracking accuracy and system robustness under degraded conditions.