PFI-TT: Physics-based Deep Transfer Learning for Predictive Maintenance of Industrial and Agricultural Machinery

Iowa State University, Ames IA

Investigators

Abstract

The broader impact/commercial potential of this Partnerships for Innovation - Technology Translation (PFI-TT) project is to provide the industrial and agricultural sectors with a practical and scalable solution for proactively predicting and preventing the failures of rotating machinery. An unexpected failure of a rotating machine may incur high maintenance and downtime costs, reduce customer satisfaction in a produced good, and cause human injuries and fatalities. These consequences not only impact the end user of the rotating machine, but also the machinery manufacturer by tarnishing its reputation and potentially impacting its competitive advantage. This PFI-TT project will accelerate the commercialization of deep learning in predictive maintenance to provide more accurate failure predictions than current solutions and is easily deployed across different types of machines and equipment. By making predictive maintenance practical and scalable, this project is expected to produce major advancements in developing machine systems that are more reliable and safer, as well as incurring lower maintenance and downtime costs than existing systems. Ultimately, the economic competitiveness of the U.S. industrial and agricultural sectors will be enhanced based on more reliable, safer and lower-cost rotating machinery. The proposed project will create a cost-effective, easy-to-implement, easy-to-scale Industrial Internet of Things (IIoT) platform for remotely monitoring machine health and predicting when and where maintenance actions need to be taken. The core of the proposed IIoT platform is a new deep learning solution that exploits the concepts of physics-based learning, transfer learning and online learning. To date, deep learning approaches to diagnostics/prognostics have been mostly relying on large volumes of training data and largely in isolation from the underlying physics of component faults. This project will overcome these limitations by integrating physics-based modeling and data-driven transfer learning. The resulting solution does not simply use run-to-failure data to train a deep learning model. Instead, training data is used, in conjunction with known physics and previously learned knowledge, to achieve more accurate predictions than possible from using training data alone. Additionally, the solution offers the capability of online learning that may lead to a paradigm shift in machinery prognostics toward unit-specific learning and prediction. The proposed deep learning solution has the potential to make predictive maintenance practical and scalable, thereby significantly promoting the wide-scale adoption of this maintenance strategy. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.