Machine Learning-Based Sensitivity Analysis of Geometric and Material Variables of Speed-Lock Beam-to-Upright Connections
Please login to view abstract download link
The rising demand for pallet rack systems in warehouse storage facilities has resulted in the widespread adoption of thin-walled structures fabricated from cold-formed steel. These semi-rigid assemblies are renowned for their lightweight, rapid installation, and favourable weight-to-strength ratio. Nonetheless, accurately predicting their stiffness and ultimate strength remains challenging and is the subject of ongoing research. Numerous researchers have concentrated on experimental investigations of beam-to-upright connections and numerical simulations of these joints to develop a reliable method for replicating the behaviour of the characteristic moment-rotation curve and identifying the common failure modes that affect these assemblies. However, current methodologies are either highly costly for experimental tests or computationally expensive for numerical simulations and fail to capture all the complex details within the joint. This study aims to utilize experimental and numerical data to develop a machine learning (ML) tool capable of assessing the effects of material mechanical properties and geometric characteristics of connection components—such as column thickness, beam depth, and number of tabs—on initial stiffness and ultimate moment. A comprehensive hybrid dataset is compiled from 20 distinct experimental configurations, supplemented by numerical data from validated FEM models, to train and evaluate the performance of the ML models. Hyperparameter optimization is performed and validated using experimental data retained from each analyzed configuration. The findings demonstrate that the trained models effectively capture the influence of variations in the input variables, successfully identifying the most critical factors.