To study the tread surface hardness difference that caused by wheel diameter difference of wheelsets, an investigation on corresponding factors of wheelset hardness was done. A group of data including wear rate, wheel diameter, wheel rim hardness and tread hardness was collected to serve as the dataset for neural network. Improved extreme learning machine (ELM) based on particle swarm optimization (PSO) algorithm was chosen to be the main method, trained by dataset and used to predict the tread surface hardness difference. The result shows that PSO-ELM is able to describe the changing trend of tread surface hardness difference, and reaches the best correlation level compared to the original ELM and BP network. Finally, the trained network was applied to analyze the relation between tread surface hardness difference, revealing that the rolled steel is a less sensitive material than the casted steel when meeting hardness or diameter difference.