A Neural Network Evaluation of the Ultimate Resistance of Plate Girders Subjected to Patch Loading (Technical Note)

Webs of rolled and built-up beam and plate girders could be subjected to local in-plane compressive patch loads. Examples are wheel loads, loads from purlins and roller loads during construction. The behavior of plate girders under patch loading, represents complex stability and elasto-plastic problems. It is influenced by many different factors and even the increasing number of experimental results and laborious theoretical work could not give complete insight into the problem. Some empirical and semi-emprical formulas have been established, but significant errors still present in the current design formulas. As an alternative solution, numerical methods, such as finite element have been used to model the problem but they still present significant difference when compared to experimental results. This study considers the use of artificial neural networks (NN) to predict the ultimate resistance of plate girders subjected to patch loading. In this work, multilayer perceptrons with backpropagation are used. This choice is mainly due to their adaptive structure and to the efficient learning algorithms nowadays available. The training and testing patterns of the proposed neural network system are based on well established experimental results taken from literature. Among these 200 tests, 50 tests are used as the test data set and remaining ones used as the training set for NN training. Each training data sample is composed of the eight geometrical and material parameters and by the experimental ultimate load. The performance of an NN model mainly depends on the network architecture and parameter settings. One of the most difficult tasks in NN studies is to find this optimal network architecture, which is based on the determination of numbers of optimal layers and neurons in the hidden layers by a trial and error approach. In this study the MATLAB toolbox is used for NN applications. To produce reasonable results, the available experimental data divided in two classes according to their ultimate load capacity. Backpropagation algorithm was used to train the first and second class neural networks. It is observed that all networks presented error values below 11% and no higher than 4.35% for majority of them. The predictions of the proposed NN model are compared with experimental values. It is obvious that the proposed NN model learned well to map the relationship between the ultimate path resistance and its geometric and mechanical properties. One of the advantages of the proposed NN model is its wide range of input parameters, which enables the NN model to be used in practical applications. The division of neural training data into two load classes provided better training and better suitability to distinct characteristics of the problem. The well trained NN model could be also used to conduct parametric studies. The results of the proposed NN are compared with current design formulae and are found to be considerably more accurate. The proposed NN presents a maximum error lower than 11% while existing design formulas errors are over 20%.