Abstract:In the current mode of railway subgrade deformation prediction after strong earthquakes, the algorithm fails to consider the nonlinear properties of the activation function, resulting in incomplete extraction of nonlinear deformation features and errors in the characteristic data. Based on this, the back-propagation (BP) neural network algorithm, incorporating the activation function, was proposed to predict the subgrade subsidence deformation after strong earthquakes. The nonlinear activation function of ReLU was compensated for by bipolar S factor, so as to optimize the BP neural network algorithm and solve the problem of characteristic number extraction in nonlinear subgrade deformations. Using the data standardized normalization method, the deviation data of all the characteristic data extracted and corrected was normalized to obtain the subgrade deformation characteristic data set. The subgrade subsidence deformation after a strong earthquake could then be predicted. Combined with the measured results, the prediction experiment was carried out using Matlab. The results showed that the proposed hybrid method can effectively predict the deformation degree of railway subgrade subsidence under horizontal earthquake action, and the error of predicted value is within acceptable limits.