The contact state of the pantograph-catenary sliding electrical contact has a direct impact on the smooth running of the locomotive. In order to judge whether the pantograph-catenary sliding electrical contact fails under certain working conditions, an improved sparrow search algorithm based on balanced data set training is proposed to optimize the kernel extreme learning machine failure diagnosis model. Firstly, the locomotive operation is simulated by the grinding experiment of the metal-impregnated carbon sliding plate and the copper wire, and the current-carrying stability coefficient and off-line rate with the sliding speed, pressure fluctuation amplitude, pressure fluctuation frequency and contact current are obtained. Secondly, the minority samples in the experimental data are expanded by the adaptive comprehensive oversampling method, and the generated balanced data set is used to train the kernel extreme learning machine failure diagnosis model. At the same time, the parameters of the model are optimized by the improved sparrow algorithm. For the shortcomings of the basic sparrow search algorithm, the chaotic mirror initialization strategy, the rotation search strategy and the Cauchy cross mutation strategy are applied to the position update of the sparrow, and then the improved sparrow search algorithm is obtained. The improved algorithm is simulated and tested by the test function, and the results show that the algorithm has better stability and convergence accuracy. Finally, by comparing the model proposed in this paper with other diagnostic models, it further verifies the effectiveness of the model and the superiority of the improved algorithm under unbalanced dataset.