Analysis Of Wind Induced Vulnerability Of Light Steel Workshop Roof Based On Neural Network

A large number of light steel plant structures along the southeast coasts of China were damaged due to typhoons,causing huge economic losses.In the first,the roof of the plant structures is often blowed.Then Rainwater enters the buildings,causing even more losses.At present,the wind induced vulnerability analysis of light-steel workshop roof is mainly based on traditional reliability analysis with data from CFD numerical simulations,or wind tunnel tests.When the size and environment of the building change,new models need to be established for re-calculations,or new wind tunnel tests to be done,which call for heavy works.In this paper,the NIST-UWO wind tunnel test data for a large number of scenarios are used as training and validation samples,and the machine learning algorithms such as support vector machine,random forest,PSO-BP(Particle Swarm Optimization Back Propagation)neural network are used to learn from the samples,to build models with strong generalization ability.With the trained models,by inputting such data as plant structure and environment characteristic parameters and wind field characteristic,the wind induced failure probability of roofs can be predicted.Firstly,a PSO-GRNN is adopted in this paper to predict the wind pressure time history at locations on roof without measurements,and then the forces in the self-tapping screws at any location are obtained.Then the region maximum method is used to determine the extreme value of forces in the self-tapping screws.Then the failure probability of roof slab of different parameter combinations is obtained by Monte Carlo simulations,in which the correlation of forces among screws is considered.The failure probabilities and the associated parameter combinations are the samples for the training of neural networks of roof vulnerability models.Sixteen parameters related to the roof panel failure are selected as input layer attributes of the network,including the building scale ratios,environmental factors,and roof panel wind pressure characteristics.The mean and mean square root of wind pressure coefficients are taken as representative of the roof panel wind pressure characteristic,and two PSO-BP neural networks are used to predict the mean and mean square root of wind pressure coefficients,respectively.Roof failure probability is selected as output layer attribute.With the prepared training data set,three classification algorithms,namelyPSO-BP neural network,Support Vector Machine and Random Forest are used to build models to project roof vulnerability under wind loading..Results show that the three models can predict the failure probability of roof slab to some extend,but the PSO-BP neural network shows the best performance,delivering the most stable and accurate predictions.