The increasing interdependence of transmission and distribution networks calls for system-wide voltage state estimation. Nevertheless, measurements remain fragmented across operators, and data sharing is often restricted, leaving each entity able to observe and supervise only a subset of system variables. This paper addresses this issue by formulating state estimation as a regression problem with incomplete labels and heterogeneous inputs, using a physics-based integrated transmissiondistribution model as a structured data generator. Steady-state AC power-flow simulations produce diverse operating scenarios with measurement noise and operator-level data partitioning; a residual multilayer perceptron then predicts real and imaginary voltage components using a masked-loss formulation that enables training despite missing target states, with the entire procedure carried out in a federated learning framework so that raw data remain local to each operator. Results demonstrate that the proposed representation and training strategy achieve accurate reconstruction of global voltage states despite fragmented observability and noisy inputs, confirming the feasibility of privacypreserving collaborative state estimation in multi-operator power systems. ... Read more