Abstract
Complex machinery contains critical regions, such as revolute joints-ball bearings-journal bearings, that are prone to damage initiation and growth. If not detected early, damage in critical local regions leads to premature failure. The overall complexity of an integrated system limits developed classical methods from detecting early damage in complicated local areas. A pure experimental data environment could provide solutions given the broad impact of machine learning. Here an interesting idea is introduced to support a machine learning framework for damage detection in local critical regions. The vibration field developed in a local area surrounding a ball bearing support of a lab flexible shaft-rotor system was measured by a set of accelerometers to form a dataset environment. It was used as an experience for machine learning by a deep convolutional neural network adapted from the AlexNet architecture. Our main result is the casting of a solid mechanics prediction problem into a classification problem and eventually computing a solution by a deep machine learning technique. Technology innovations improve computer speed, data storage media, and graphics processing units. These factors are turning existing machine learning techniques into state-of-the-art prediction tools that can be adapted and developed to exploit large volumes of vibration data for diagnostics. Data-driven predictive-diagnostics results in improved condition monitoring of the complex machinery system with economic gains form estimated low-cost maintenance and energy savings. Classical condition monitoring techniques cannot learn from experiences predictive models of the dynamics-diagnostics of onboard ship and aircraft machinery operating under varying environmental conditions.
