In order to figure out the matter that the traditional construction manners of rolling bearing degradation index extremely rely on manual selection features,an end-to-end construction method of rolling bearing performance degradation indicator based on convolutional neural networks was put forward.In this method,two nodes were set at the output end of Softmax,respectively representing the normal state and the failure state.Normalized vibration amplitude spectrum of the healthy state was taken as the training sample,and the rolling bearing performance degradation index was constructed based on the probability of normal state.The effectiveness and superiority of this method were verified by application in different experiment data sets and comparison with other indices.