June 5, 10:15am
Location: Magnolia

Cost-Benefit analysis for justifying investment in Predictive Maintenance: mapping it to observability, performance targets and robustness of AI models

The operational safety, cost and criticality of complex, industrial assets like aircraft engines, gas/wind turbines, locomotives, nuclear reactors have historically led to the design of conservative sustainment and maintenance strategies for these assets. However, such strategies: are often wasteful due to remaining component-life that is discarded, they curtail revenue by reducing effective availability of assets, and lead to high O&M (operations and maintenance) costs due to unnecessary time, effort and cost expended. In cases like the nuclear industry, a key barrier to adoption as well as future growth is the extremely high capital and O&M costs. Additionally, the dire demands of climate change are more regularly driving the need to employ many of these industrial assets within paradigms that are more efficient, frugal, and less wasteful. Predictive analytics, and maintenance strategies devised using them, have been identified to be a key capability for addressing some of these challenges. When designed correctly, preventive maintenance (PM) strategies make use of sensing, monitoring, and predictive analytics to reliably assess condition of asset-components in real-time, thus enabling cost-efficient, sustainment workflows in support of the asset, that optimally trade off sustainment costs, waste, and revenue, while retaining same levels of operational safety. However, the additional investments required to stand up the PM paradigm can itself become a barrier to the adoption and deployment of PM strategies in the industry. Like most investments, an assessment of the return on investment (RoI) is often a critical factor that can serve towards helping make them. However, a cost-benefit analysis of a PM paradigm for a complex machinery with thousands of components can be a non-trivial activity – the information resolution required to assess conclusive benefits might often not be available or uncertain, while a very coarse analysis might make it hard for the operator to believe in the outcome. Multiple factors of uncertainty arise from not knowing how much of the relevant phenomena is observable with the sensors in place, not knowing the true entitlement of the predictive analytics driving the PM, and how effective the analytics will need to be, in terms of their accuracy and reliability, to drive decisions that are overall cost-reducing. At a secondary level, errors caused by predictive analytics lead to newer costs from suboptimal decisions, and thus one needs to account for those as well. In this talk, we introduce a cost-benefit framework we developed at GE to support strategy development for PM related to the new BWRX300 nuclear reactor. The framework is generalizable to enable similar cost-impact evaluations for PM evaluation scenarios specific to other industries. Most importantly, it allows one to assess the limits of cost-impact (entitlement) from usage of predictive maintenance, while additionally deriving requirements for PHM algorithm performance, which traditionally have not been rigorously connected to high level metrics. The goal is to drive such analysis to help estimate likely cost reductions from deploying PM and to derive required performance targets, both in terms of accuracy and reliability of the predictive analytics, for the cost reductions to manifest.This work was funded in part by the Advanced Research Projects Agency-Energy (ARPA-E), U.S. Department of Energy, under GEMINA program Award Number DE-AR0001290.

Dr. Naresh Iyer is a Principal Scientist in the AI and Machine Learning group at GE Research. He has over 20 years of experience in the research and application of machine learning to a variety of industry problems, including asset life prognostics, surrogate modeling, multi-objective optimization and decision making under uncertainty. He has developed solutions for a diverse range of industrial applications using methods in supervised, unsupervised, semi-supervised learning and evolutionary soft computing. Prior to joining GE in 2001, Dr. Iyer graduated with a Ph.D. in Artificial Intelligence from the Ohio State University. 

His recent research interests include robust machine learning, sequential decision making, active learning, adversarial machine learning and generative design. He is currently engaged in research dealing with the extension of epistemology to studying “knowability”, interpretability as well as general and adversarial robustness of machine learning and AI models. This includes implementation of epistemically justifiable machine learning models for various self-supervised and supervised, industrial applications. Dr. Iyer is deeply interested in the full potential of additive manufacturing and the ability of machine learning to realizing that potential, including deep generative machine learning approaches for process monitoring, in situ control and defect identification in LPBF printed parts. In the recent past, Dr. Iyer has been the technical lead on a program targeting the development of large scale, onboard deep learning solution for automated defect characterization and in situ control in Additive manufacturing machine. More recently, he was a lead contributor on an ARPA-E program targeting generative design to improve manufacturability of an additively manufactured part. He is currently a co-PI on a DOE-program targeting the design of intelligent inspection strategies for large, additively manufactured aerostructures, using process monitoring data. He is also a technical contributor on an active FAA-funded program targeting the formal verification and validation of Deep Learning models. Dr. Iyer has over 30 peer-reviewed publications and 45 filed patents.