For the clustering time-varying sensor network systems with uncertain noise variances, according to the minimax robust estimation principle, based on the worst-case conservative system with conservative upper bounds of noise variances, applying the optimal Kalman filtering, the two-level hierarchical fusion timevarying robust Kalman filter is presented, where the first-level fusers consist of the local decentralized robust fusers for the clusters, and the second-level fuser is a global decentralized robust fuser for the cluster heads. It can reduce the communication load and save energy resources of sensors. Its robustness is proved by the proposed Lyapunov equation method. The concept of robust accuracy is presented, and the robust accuracy relations of the local, decentralized, and centralized fused robust Kalman filters are proved. Specially, the corresponding steady-state robust local and fused Kalman filters are also presented, and the convergence in a realization between the time-varying and steady-state robust Kalman filters is proved by the dynamic error system analysis method. A simulation example shows correctness and effectiveness. Figure 1. The two-level hierarchical clustering sensor network.A sensor network system consists of a number of sensor nodes distributed over a spatial region. Each node has sensing, communication, and computation capabilities, which is also called an agent. Usually, the number of sensor nodes is very large, and these nodes distribute in a remote area, so the cost of transmitting is higher than computation; hence it may be beneficial to partition the sensor nodes into groups (clusters) to reduce the communication load and save energy resources of sensor nodes, using the nearest neighbor rules [12], as shown in Figure 1. Each cluster has a cluster head (local fusion center) and a number of member nodes. Each member node sends its data or local estimator to the corresponding cluster head, and each cluster head sends its data or fused estimator to the base station (global fusion center) to obtain the global fused estimator. This is a hierarchical strategy with two levels, which decomposes a large sensor network into separate zones within which data processing and aggregation can be carried out locally.There exist two kinds of fusion methods, centralized and distributed fusion methods, depending on whether raw data are used directly for fusion or not [13]. The former can give the globally optimal state estimation by directly combining local measurement data to obtain an augmented measurement equation, but its disadvantage is that it requires a larger computational burden. The latter includes the distributed state fusion and distributed measurement fusion methods [14]. The distributed state fusion method can give the globally optimal or suboptimal state estimation by combining or weighting the local state estimators [15][16][17][18][19] whose advantages are that it can facilitate fault detection and isolation, and can reduce the computational burden. The distributed measurement fu...