Crack detection of plastic gears using a convolutional neural network pre-learned from images of meshing vibration data with transfer learning

In this paper, a crack detection method of plastic POM (Polyoxymethylene) gears using a deep convolutional neural network is proposed. The vibration signal is collected from an automatic data acquisition system for a gear operation test rig, with an accelerometer installed on the housing of bearings. The FFT (Fast-Fourier Transform) spectrums of the measured vibration signal are visualised as grayscale images for training input. Additionally, a high-speed camera observes cracks occurring at the root of the teeth which enable to label the visualised images for training data. A popular convolutional neural network, called VGG16 ConvNet, is employed for the classification of the crack or non-crack situation of gears. This VGG16 is pre-learned from image data of ImageNet and relearned from the created images of vibration data with the transfer learning technique. The weights of the last two layer are re-trained in the transfer learning process. The accuracy rate for learning from training data approached 99% by transfer learning. Additionally, the accuracy rate can reach 100% on classification tasks from testing data. Finally, four endurance tests of plastic gear were carried out. The proposed crack detection was carried out at the rate of 10 times per minute, and the errors of detection results are investigated by training data assessment. The experimental results show that the proposed method is sufficient to specify the crack or non-crack situation of gear during the operation test.