Given that steel is one of the main structural materials in civil engineering, it is important to understand its mechanical properties and failure mechanisms. L-shaped laminated steel elements, known as angles, are commonly bolted to other metal elements to make connections in steel structures. When these elements are connected by only one of its legs and are subjected to axial tension, the failure of the angle can be caused by rupture of the section where the holes are located. In this case, there is influence of complex phenomena such as shear lag, therefore a reduction factor is applied to the resistance of the section. The general theme of this work is the use of Artificial Neural Networks (ANN) as an approach to the study of the load capacity of cold-formed steel bolted angles connected by one leg and under axial tension. Although this class of techniques do not provide a simple regression equation, they are powerful tools in Machine Learning (ML), having the ability to model any arbitrarily complex nonlinear problem. In this work, a dataset is built with samples from different numerical and experimental works, collecting steel angles geometric parameters, tensile strength of the material and ultimate capacity regarding net section failure. The data-driven models are trained with 80% of the data, tested with 20%, validated with 5-fold cross-validation and their hyperparameters are tuned with Bayesian Optimization. Feature generation, selection, and importance algorithms are implemented, hoping to achieve more accurate, less complex and more interpretable models. ML models predictions are compared with the ones given by the equations of Eurocode-3, Brazilian NBR 14762, American Iron and Steel Institute (AISI) and Australasian AS/NZS Standards. The comparison shows mostly superior results in the ANN models, proving the competitiveness and effectiveness of Machine Learning techniques and data-driven modeling.
Read full abstract