Intelligent Assessment of Partially-Restrained Beam-Column Joints Using Vision-Based Deep Learning and Decision Trees

The assessment of partially-restrained beam-column joints is critical for ensuring the safety and performance of steel structures, particularly under seismic and dynamic loading conditions. Traditional methods for evaluating joint behavior often rely on labor-intensive experimental setups or simplified analytical models, which may not capture the full complexity of real-world scenarios. This study proposes a novel framework that integrates vision-based deep learning (DL) and decision trees (DT) for the intelligent assessment of partially-restrained beam-column joints. A high-fidelity finite element model is developed using ANSYS to simulate the nonlinear behavior of these joints under varying loading conditions, including monotonic, cyclic, and dynamic excitations. The model incorporates material nonlinearities, geometric imperfections, and contact interactions to ensure accurate representation of joint responses. A convolutional neural network (CNN) is trained on a dataset of visual representations (e.g., strain contours and deformation patterns) extracted from the ANSYS simulations, enabling real-time classification of joint performance. Decision trees are then employed to interpret the CNN outputs and provide actionable insights into joint rigidity, failure modes, and load-carrying capacity. The proposed framework achieves over 90% accuracy in predicting joint behavior, significantly outperforming traditional assessment methods. Results are visualized through 3D deformation plots, confusion matrices, and feature importance diagrams, offering multiple perspectives for structural engineers. The study concludes that the integration of vision-based DL and DT provides a robust, scalable solution for the intelligent assessment of beam-column joints, with potential applications in structural health monitoring and retrofit design.