© 2019 SPIE. This paper shows crack detection systems based on deep neural networks, which analyze meshing vibration of plastic gears. A gear operating test rig has an acceleration sensor attached on a bearing housing and a high-speed camera. The meshing vibration of plastic gears during operation was measured and teeth images that enable us to decide whether cracks exists were captured. After transferring the meshing vibration data in the time domain to the frequency domain by FFT, the amplitude and phase information of the meshing vibration was converted to image data. According to the images from the high-speed camera, the imaged vibration data were separated to two classes, with or without crack, as the training data for deep neural networks. Furthermore, two convolutional neural networks, 4 layers and 16 layers were constructed for classification of crack existence or non-existence, and the systems were learned from the labeled data set. In the training, the random weighting functions of the convolution were prepared, and the number of images were 350 and the number of epoch was 125. The learning of the 4 layers convolutional neural network was finished appropriately, however, the learning of the 16 layers convolutional neural network did not progress at all. Then, the transfer learning method was used for the 16 layers convolutional neural network. The transfer learning of the 16 layers convolutional neural network was finished appropriately, and the accuracy at 125 learning steps reached to 97.2%.