Accurate modeling between physical devices and digital twins in industrial digital twins requires synchronizing massive amounts of sensing data. This results in difficulty ensuring real-time virtual-physical mapping. To address this issue, this paper proposes a four-layer bidirectional closed-loop architecture suitable for digital twins, consisting of a device, federated, cloud, and application layers. We develop a distributed federated bucketized decision tree algorithm utilizing histogram-based gain for global aggregation based on this architecture. Additionally, we design a pruning algorithm with adaptive gradient-weight redistribution to accelerate model convergence. Experimental results on the dataset demonstrate that the proposed federated aggregation model achieves accuracy improvements of 10%, 11%, and 22% compared to baseline methods. It also achieves convergence improvements of 15%, 20%, and 27%. Moreover, the proposed model maintains additional communication overhead within 300 mJ, even under the worst channel conditions.