Abstract: The existing multi-scale convolutional network models lack a fault discrimination mechanism and ignore the influence of different convolution scales on the extracted features of the model. Furthermore, most of the neural network models based on current signals for bearing fault diagnosis are poor in interpretability. Aiming at solving these problems, a bearing fault diagnosis method based on Multi-dilated Convolutional neural network with Adaptive Feature Selection (MCAFS) is proposed. Firstly, the demodulation techniques are used to suppress the interference of the fundamental frequency component of the original current signal, and then the fast Fourier transform (FFT) is employed to convert the demodulated signal from the time domain to the frequency domain. Secondly, the shallow features are extracted from the frequency domain through a standard convolutional neural network. Next, multi-scale features are selected from shallow features using parallel dilated convolution (PDC) blocks with different kernel scales. Then, an improved multiscale dilated convolutional attention (MCA) module is proposed to adaptively select the size of the convolution scale and extract the deep features. Finally, the spatial attention module (SAM) is introduced to visualize the attention distribution of the input frequency signal, which further improves the interpretability of the network. The proposed network model is verified by the current signals of the Paderborn rolling bearing dataset. The experimental results show that the proposed network model can adaptively select the convolution scale, effectively locate the key information of the input data, and provide certain interpretable significance for the feature extraction process of current signal-based bearing fault diagnosis.