| The active movement of the earth’s plates has caused frequent earthquakes in recent years,bringing huge economic losses and casualties to China.The ground motion attenuation relationship provides important theoretical support for works of disaster prevention and damage mitigation such as structural seismic design and seismic risk assessment,which have great significance in reducing the losses caused by earthquakes.At present,the traditional method and artificial intelligence method are mainly used in the ground motion attenuation relationship pre dictions.Both methods have certain limitations.Limited by statistical regression,the traditional method has low accuracy in prediction models.The artificial intelligence method lacks the physical meaning,resulting in models with low robustness.This p aper proposes a method for the ground motion attenuation relationship predictions based on the hybrid of artificial intelligence technology and physical equations;this method absorbs empirical formulas which have physical meanings in the neural network while maintaining a high accuracy by using the artificial intelligence method,which therefore improves the generalization ability of the predictive model.The main research scenarios of this paper are as follows:(1)A hybrid method of ground motion attenuation relationship combining artificial intelligence and the physical equation is proposed,and its implementation steps are determined.This method uses the Pytorch framework to build a neural network.The traditional ground motion attenuation relationship is adopted when constructing the network;physical meaning is added to the neural network.The regression coefficients in the neural network are set as variables with gradient updates.They are iteratively updated with the training of the neural network u ntil the optimal ones are obtained.Eventually,a ground motion attenuation relationship with strong robustness and high accuracy is constructed.(2)The ground motion attenuation relationship of the CB14 model based on the hybrid method is established.Compared with the traditional prediction model,the hybrid method is found to have a better performance in several indicators.The attenuation curve of the hybrid method can fairly reflect the objective laws of ground motion attenuation relationships,such as large earthquake saturation,near-field saturation,and inelastic attenuation.The inter-event residuals and intra-event residuals are distributed densely and symmetrically.The standard deviations of the inter-event and intra-event residuals and the total residual of the hybrid methods are lower than those of the CB14 model.The ground motion attenuation relationship established by the hybrid method has higher accuracy and stronger robustness compared with the one established by the traditional method.(3)The ground motion attenuation relationship of the Sichuan-Yunnan region is established by the hybrid method.Establishing a ground motion database of the Sichuan-Yunnan region by collecting historical ground motion records in that region.Using the hybrid method and the database to predict the ground motion attenuation relationship in the Sichuan-Yunnan region.The results demonstrate that the residual distribution of the ground motion attenuation relationship is symmetrically dense.The standard deviation of inter-event,intra-event and total residuals are lower than those of the traditional method,indicating a higher accuracy.Most of the ground motion data points fall between the attenuation curves of M_W=4.5 andM_W=7.5,which indicates that the attenuation relationship established by the hybrid meth od proposed in this paper is suitable for the Sichuan-Yunnan region. |
