Predicting the thermal structural behaviour of steel pallet rack connections using machine learning

Unlike traditional bolted or welded steel connections, the beam-to-column connection in steel pallet racks is established using a unique beam end connector with punched tabs, which is welded to the end of the pallet beam. This design leads to complex structural behaviour, particularly at elevated temperatures due to reduced steel properties, thermal expansion, and fire-induced failure modes. Design standards recommend experimental testing for each connector type, as a general analytical model is lacking. However, such testing is often costly and time-consuming. Recent advancements in machine learning offer a promising alternative to traditional testing methods in civil engineering. This study explores the application of machine learning techniques to predict the moment-rotation behaviour of steel pallet rack beam-to-column connections under elevated temperatures. The methods investigated include Linear Regression, Ridge Regression, Ridge Cross-Validation Regression, Logistic Regression, Least Shrinkage Regression, Least Absolute Shrinkage and Selection Operator with Cross-Validation Regression, and Localised Linear Support Regression. The dataset comprised 63 experimental tests, varying in column thickness, beam depth, and number of tabs in the beam end connector, conducted at temperatures ranging from 25 °C to 700 °C. Model evaluation employed metrics such as mean-square error, Pearson correlation coefficient, and coefficient of determination. Results indicate that linear regression effectively predicts moment-rotation behaviour, closely aligning with experimental data. While other machine learning techniques performed well in moment prediction, their accuracy in rotation forecasting was limited. This research enhances the accuracy of structural behavior predictions and offers a cost-effective alternative to physical testing for steel pallet rack designers. Future research recommendations are provided based on these findings.