The primary problem solved in rotating machinery fault diagnosis is how to effectively extract fault features from the vibration signals with noise. To extract fault features accurately, this study proposes a multi-scale parallel convolutional neural network fault recognition algorithm, which can carry out feature fusion. The above method combines empirical feature extraction (e.g., fast Fourier transform) to enrich feature information, which can effectively implement deep learning. The effectiveness and reliability of the method are verified through example studies on JNU, SEU and PU rolling bearing experimental data sets. The algorithm has the higher classification capability and diagnostic accuracy compared with four common deep learning algorithms.