This research aims to enhance predictive maintenance and inspection planning in urban construction projects. Recent advances in graph neural networks and graph transformer architecture have demonstrated significant potential for modeling complex lifecycle processes of building systems. However, most existing approaches remain predominantly data-driven and lack integration of physics-informed modeling and real-time data, which limits their applicability in large-scale urban environments. This research addresses this gap by proposing an approach for managing heterogeneous big data in urban construction projects, enabling prediction of the technical condition and inspection needs of structural elements. The core contribution is the development of a physics-informed heterogeneous graph transformer model that integrates domain-specific physical knowledge into the learning process through physics-based features and regularization mechanisms. The results confirm that all validation criteria are simultaneously satisfied: the difference between training and validation accuracy remains within the threshold (=0.05); The overall classification accuracy exceeds 92.06%; area under ROC curve above 0.8; F1-score is above 0.8 for all major classes; Physics-alignment error is lower than 0.15; and a strong Spearman correlation is observed between model predictions and physics-based indicators. The novelty of the proposed approach lies in the development of a physics-informed graph learning paradigm which enables the integration of structural mechanics, degradation processes, and heterogeneous data sources within a unified predictive framework.

This article presents a new multi-objective model that optimizes Kafka configuration to minimize end-to-end latency while quantifying independent parameter influence, interaction effects and sensitivity to local parameter changes. The proposed model addresses a challenging problem of selecting the configuration to prevent overloading while maintaining high availability and low latency of Kafka cluster. The study proposes an algorithm to implement this model using an adaptive optimization strategy that combines gradient-based and derivative-free search methods. This strategy enables a balance between convergence speed and global search capabilities, which is critical when dealing with the nonlinear parameter space characteristic of large-scale Kafka deployments. Experimental evaluation demonstrates 99% accuracy of the model verified against a trained XGBRegressor model and tested across multiple optimization strategies. The experimental results show that alternative configurations can be selected to meet secondary objectives-such as operational constraints - without significantly impacting latency. In this context, the designed multi-objective model serves as a valuable tool to guide the configuration selection process by quantifying and incorporating such secondary objectives into the optimization landscape. The proposed multi-objective function could be adopted in real time applications as a tool for Kafka performance tuning.