Vibrations-Based Machine Fault Diagnosis and Prognosis Using Convolutional Neural Networks

Abstract

This thesis addresses vibration-based machine health monitoring (MHM) by applying the fundamentals of machine learning (ML), convolutional neural networks (CNNs) and selected signal processing. The thesis first presents an exploration of the relationship between the hyperparameters of two-layer CNNs, the type of signal preprocessing used, and resulting diagnostic accuracy. For this, two popular bearing fault datasets and a gear fault dataset are used to reveal cross-domain trends. It is found that using time-frequency representations provided by the spectrogram transformation results in a reduced dependence on hyperparameter optimization and lays the foundation for the following work. Moreover, by applying ML theory and best practices, the thesis demonstrates shortcomings in currently accepted benchmarking practices to evaluate the domain adaptability of bearing fault diagnosis algorithms and proposes an alternative benchmarking framework to resolve them. A novel data preparation and transfer learning procedure that capitalizes on the use of multiple sensors and that achieves higher accuracy than state-of-the-art algorithms is demonstrated. In addition to fault diagnosis, the thesis addresses bearing health prognosis by applying CNNs to health indicator estimation using data from accelerated life testing. Several data augmentation methods adapted from other ML fields are compared. It is determined that methods proven in sound classification or image recognition fields are not guaranteed to benefit this task. Lastly, the thesis presents a 3D CNN designed for bearing health prognosis that uses a multi-sensor time-frequency input to improves upon single-sensor variants. The thesis explores the strengths, as well as the shortcomings, of CNNs for MHM, an emphasis is placed on network design, signal transformation, and experimental methodology.