The vibration signals of cylindrical roller bearings in the gearbox of special vehicles are usually weak, nonlinear and non-stationary. Moreover, the extracted feature quantity value is not obvious, and it is difficult to obtain a large number of typical fault samples. These problems make it difficult for roller bearings to be accurately diagnosed. In order to solve those problems, approximate wavelet packet entropy and a support vector machine are used to diagnose these roller bearings. First, four typical state vibration signals of the roller bearings are collected, including normal, outer ring wear, rolling element failure, pitting and indentation from the self-built test bench. Second, the approximate wavelet packet entropy value of four typical state vibration signals can be extracted as the input of support vector machines. The results of the support vector machine output can help determine whether the bearing is faulty or the fault type. The results show that this can effectively diagnose the typical state of cylindrical roller bearings in the gearbox of special vehicles, and provide a practical reference for the fault diagnosis of similar gearbox cylindrical roller bearings.