Convolutional neural networks for industrial inspection using sound measurement data

Abstract

This thesis proposes an inspection method using a convolutional neural network (CNN) to automate industrial inspection using sound measurement data. We first consider the industrial inspection problem as a classification problem in machine learning to automate inspection. Given the sound measurement data of normal and defective samples for rotating machines, which can be inspected with sound measurement data, we train a classifier that use the CNN, which is a kind of deep learning. In general, it is difficult to obtain large amounts of data for learning in industrial inspection problems. To overcome the lack of training data, we use transfer learning. In addition, Greedy layer-wise supervised training method is proposed to improve the performance in transfer learning. As an example of industrial inspection using sound measurement data, we conduct the inspection of the electric motor used in the drones by the inspection method presented above. Given the sound measurement data of the electric motors, we perform several experiments to show the performance of our algorithm. Our inspection algorithm using the CNN shows better detection of defective motors than the inspection using conventional classification method in machine learning. Especially, the algorithm using the CNN is a kind of end-to-end learning, and it shows excellent performance without manually extracting the features adequate to the given data. Therefore, it is applicable to various inspection fields using sound measurement data without deep understanding of the given data.