Future of machine diagnostics in smart manufacturing environments: a study of deep neural networks transfer learning for fault diagnosis applications

Abstract

Intelligent fault diagnostics involves instrumenting machinery with sensors, and interpreting data to predict faults. We applied deep learning algorithms, called Deep Convolutional Neural Networks (DCNNs), to analyze raw vibration sensor data and accurately determine the source and magnitude of the fault, if any. The neural network proposed for this study was based on past research, but expanded to include 12 fault classes. The goals of this study were to obtain improved classification accuracy and improve the ability of the network to natively transfer training from one operating condition to others. In addition, we explored how to make the network architecture more efficient and how to understand and represent the 1st layer of the DCNN output.