The deployment of autonomous vehicles, aircraft, and cooperative multi-agent fleets in human-shared environments makes the reliability of their state estimation a critical requirement. Beyond an accurate estimate, the estimator must certify at every time step that the true state lies within a known bounded domain, since a single loss of this guarantee may cause a collision or an undetected fault. Probabilistic filters such as the Kalman filter and its extensions are efficient but describe the estimation error only in statistical terms, without certifying that the true state is enclosed. Set-membership estimation provides this guarantee by propagating bounded sets, the zonotope being the most prominent representation for its balance between accuracy and efficiency; it remains limited, however, by the growth of the generator count and by the cost and conservatism of the intersection operation. More broadly, model-based methods depend on an accurate model, which may be only approximate or unavailable, whereas data-driven methods improve accuracy and apply without a model but offer no formal inclusion guarantee. This thesis combines these complementary families. The first contribution revisits the core zonotopic operations through a hyperplane-based formulation used both as a contractor for the intersection of two zonotopes and as an order-reduction operator that bounds the generator count; integrated into the Zonotopic and Gaussian Kalman Filter under an aggressive reduction order, it preserves the coverage guarantee with substantially fewer generators. The second and third contributions are two complementary hybrid filters that bridge the model-based and data-driven families. In the first, a model-based prediction is corrected by a GRU-based neural stage, and the corrected estimate centers a constrained zonotope propagated by Jacobian linearization along the trajectory. In the second, a GRU predicts the state and an Unscented Kalman filter corrects it, the corrected estimate again centering a propagated constrained zonotope. In both, the linearization is confined to the set-propagation step, so the inclusion guarantee is preserved.