Explainable AI for prognostics with uncertainty quantification and domain knowledge enhancement
Abstract
The objective of this project is to establish a new integrated suite of explainable artificial intelligence (AI) techniques with uncertainty quantification to significantly enhance prognostics by incorporating domain knowledge. With the rapid development of sensing and communication technology, massive and heterogeneous types of data including life history of individual systems (e.g., maintenance logs) and condition monitoring signals from different sensors (e.g., pressure sensor, acoustic sensor, proximity sensor, environmental sensor), have been collected and widely available in DoD applications nowadays. These data contain valuable information regarding the system performance in real time and offer an unprecedented opportunity to empower prognostics. In addition, we have seen significant progress in AI methods, which have shown human- or super-human level performance in many tasks such as image recognition,natural language processing, gaming, etc. However, we are currently lackingadvanced AI methods that can fully leverage the available data and incorporate the domain knowledge to accurately predict the remaining useful life (RUL) with meaningful explanation and uncertainty quantification.To address these critical research gaps and challenges, four interrelated research tasks are proposed: (i) In task 1, we will propose two effective DL-based prognostic models based on event data and condition monitoring signals. In particular, a semi-parametric DL framework based on the extension of the popular cox proportional hazards model and a non-parametric flexible DL model that directly allows conditional monitoring signals and event data as inputs will be established. (ii) In task 2, we will then (1) investigate how to intelligently tune model parameters by Bayesian optimization to greatly speed up the training process and (2) provide an accurate uncertainty quantification of RUL by Bayesian neural network. (iii) With the proposed DL-based prognostic models in (i), task 3 will study how to innovatively incorporate domain knowledge through a unified loss function and further enhance the model interpretability by identifying the most impactful input variables that explain the prognostic result. (iv) In task 4,we will conduct synthetic simulation and real case studies to thoroughly test and validate the proposed methods. The potential impact of the project will be significant and transformative. First, from the methodological viewpoint, this project will significantly enrich the prognostic capability of ONR by establishing a new integrated suite of explainable DL-based techniques, which will provide a paradigm shift in predictive maintenance for military systems. Second, from the application viewpoint, this research will significantly strengthen DoD competitiveness, including reduced operation and maintenance cost, improved operational readiness, and enhanced mission understanding, reliability, and safety for warfighters. As the proposed research will be fundamental in nature, the results can be easily extended to various after-sales services and industries as well, in which predictive maintenance is critically important to the system availability and safety.Approved for Public Release.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 15, 2023
- Source ID
- N000142312495
Entities
People
- Kaibo Liu
Organizations
- Office of Naval Research
- United States Navy
- University of Wisconsin System