Abstract:In order to achieve accurate prediction of rail wear on small radius curve of heavy-duty railway, a prediction model for rail wear combining improved particle swarm optimization with BP neural network and the GM(1,1) model is proposed. First, the calculation method of background value of is optimized according to the principle of integration, and preliminary prediction results of wear are obtained based on the improved model. Then, the weights and thresholds in network structure nodes of the BP model is optimized by using the IPSO algorithm, and the residual values of the preliminary prediction results are corrected. Finally, the optimized two models are combined to construct a prediction model for rail wear in small radius curve of heavy-duty railways based on GM(1,1)-IPSO-BP model. Taking a curve with 400 m radius on a Heavy Duty Railway bridge as a case study, applicability analysis of the method in this article is analyzed using long-term wear monitoring data. The results show that the GM(1,1)-IPSO-BP model overcomes the influence of nonlinear and random characteristics of wear data on the calculation results, and its prediction accuracy is superior to using GM (1,1) and IPSO-BP alone. The GM (1,1) model with optimized background values has more reliable prediction accuracy. Prediction accuracy and calculation speed of the BP network optimized by the IPSO algorithm is improved. The relative error between the predictive value calculated with this model and the detection value is less than 4%, and the absolute error of abrasion value is not more than 0.4 mm. Through this method, the development law of rail wear can be accurately obtained. The research results could provide reference for the precise maintenance and scientific use of small radius curve rail in heavy haul railways.