As a result of low accuracy and susceptibility to noise interference of traditional bearing fault diagnostic algorithms, a diagnosis method combining empirical mode decomposition and convolutional neural network is proposed. First, fiber Bragg grating (FBG) is used to obtain the vibration signal of the bearing, and then empirical mode decomposition is used to decompose the signal into multiple intrinsic mode function (IMF) components. After the extraction of useful signals, based on the structural characteristics of IMF components, the IMF components are combined into a matrix and input into the improved convolutional neural network for fault classification and recognition. The results show that the proposed method can effectively identify normal, faulty, and composite faulty bearings. Furthermore, the recognition accuracy of the proposed method is greater than 91%.