ERI: Harnessing Probabilistic Deep Learning Method Integrated with Tailored Features for Enhanced Real-Time Machinery Fault Diagnosis and Prognosis

This award is funded in whole or in part under the American Rescue Plan Act of 2021 (Public Law 117-2).

This Engineering Research Initiation (ERI) grant will fund research that enables real-time quality control and accurate decision making in the operation of complex machinery, including critical components in modern manufacturing, thereby promoting the progress of science and advancing the national prosperity. Fault diagnosis and prognosis for machinery systems play an indispensable role in ensuring integrity, safety, and performance.

Due to many sources of uncertainty, reliable real-time fault monitoring and prediction are beyond current capabilities. Machine learning-based techniques show promise in overcoming the limitations of traditional measurement approaches, but state-of-the-art tools lack the ability to identify faults that were not encountered during training on previously collected data.

This project will overcome these limitations by creating a new machine-learning framework that uses probabilistic ideas to account for measurement uncertainty, is sensitive to time-varying fault signatures, and trains continuously on real-time data. The new framework will enable substantial performance enhancements in reliability, efficiency, practicality, and robustness of machinery fault diagnosis and prognosis in aerospace, transportation, and infrastructure industries.

Related software tools and curated datasets will be shared publicly in order to promote technology transfer and broad access to the research methodology. Undergraduate research opportunities will provide hands-on learning experiences and, leveraging a partnership between Michigan Tech University and three Michigan community colleges, help broaden participation in STEM of students from currently underrepresented groups and the institutions that serve them.

This research aims to make fundamental contributions to the integration of several deep-learning technologies with an algorithm for optimized sensor placement to enable the use of real-time vibration measurements for detection and prediction of machinery faults also outside of those in a given training data set, robustly to measurement noise and time-varying operating conditions, and reliably even given limited data. It will achieve this outcome by relying on a Bayesian convolutional neural network architecture that builds a probabilistic model for feature detection, a long short-term memory architecture that is sensitive to intrinsic temporal correlations characteristic of the progressive nature of faults, and a variational inference-based backpropagation optimization algorithm for real-time model updating, further facilitating generalizations to previously unseen faults.

This project will use the Shapley Additive Explanations approach to quantify the importance of signal features for fault detection, and will apply this metric to optimize sensor placement for an experimental gearbox that will be used to test and validate the algorithmic framework.

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.