In order to solve the problem that the traditional method has limited performance under noise interference and complex working conditions, a method based on Markov transition field (MTF), residual network (ResNet) and convolutional block attention module (CBAM). The 1-D vibration signal is mapped to a 2-D MTF image to retain the temporal dependence and dynamic features. ResNet is used to extract the deep features, and CBAM is used to adaptively assign weights in the channel and spatial dimensions to enhance the expression of key information and suppress redundant interference. The experimental verification is carried out under four typical working conditions (normal, inner ring fault, outer ring fault and rolling element fault). The results show that the overall test accuracy of the model reaches 96. 67%, which is about 8%～15% higher than that of VGG, AlexNet and CNN models, and the method can maintain high diagnostic accuracy and stability in the complex operating environment of ordnance equipment.