Vibration-based plastic-gear crack detection system using a convolutional neural network - Robust evaluation and performance improvement by re-learning

摘要

This paper evaluates the sensitivity of a proposed crack detection method of POM (Polyoxymethylene) gears using a deep convolutional neural network. The vibration signal was collected from an automatic data acquisition system for endurance tests of gears. The fast Fourier transform (FFT) of the measured vibration signals generated grayscale images for training input. A high-speed camera captured cracks at the tooth root, and the length of cracks was computed as a damage index for training labels. A convolutional neural network (CNN), called VGG16 ConvNet, which has 1000 classes in output, firstly was pre-learned from image data of ImageNet and then the weights of two layers, which were close to the output layer, was relearned from the created images of meshing vibration data with the transfer learning technique. The output layer was modified to fit two classifications problem related to the cracked or non-cracked situation of gears. The accuracy rate for the recognition of the gear fault reached 100%. However, the remained problem is whether the performance of the developed system is susceptible to the change of the working condition of gear, such as high rotational speed and torque, or not. Hence, the robustness of the crack detection performance of the developed system was investigated. The endurance tests of gears under some test conditions, such as high-low rotational speed and/or torque, were carried out to collect the different vibration signals. The accuracy rate of gear failure classification under various working condition was judged and the factors affected on the performance of the developed system under working condition changing was discussed. The results showed that the developed system learned from one testing condition incapably perform in varied testing conditions. In other words, the developed system must be learned from diversity data for a superior effectuation. In this case, our interest is to uncover how many experiments and images for re-learning are required in each experiment for better performance. The investigation of the re-learning in this paper showed the required number of images was 200 and a single endurance test under each condition was enough if an appropriate number of images were obtained.