The traditional convolutional neural network (CNN) has the limitations of insufficiently extracting features from the original vibration signals and requiring a larger sensory field to fully capture the temporal correlation of the signals in the process of extracting the features when recognizing the fault types. A dilated convolution and enhanced multi-scale adaptive feature fusion model (DC-MAFFM) was proposed considering the inherent multi-scale characteristics of bearing vibration signals. The signal features were extracted using the large receptive field of the dilated convolution, and the residual connection was introduced to reduce the information loss on the convolution layer, so as to effectively filter the noise in the signal. An improved multi-scale feature extraction module was designed to capture complementary diagnostic features at different scales, meanwhile, the different-scale feature fusion was performed at each layer to fully learn the high-frequency and low-frequency features of the signal. The proposed feature adaptive fusion module was used to adaptively assign weights to the features at different scales to enhance the ability of discriminative feature learning. Verification was carried out on two bearing datasets, and results showed that the proposed model had strong diagnostic ability under noise and variable working conditions. In the case of strong noise, the fault diagnosis accuracy reached 88.08% and 75.56%, respectively, which demonstrated that the DC-MAFFM had a significant advantage over other methods.