PSI - Issue 19

Aaron Stenta et al. / Procedia Structural Integrity 19 (2019) 27–40 Stenta and Panzarella / Structural Integrity Procedia 00 (2019) 000–000

28

2

Keywords: Probalistic fatigue apprach; Bayesian Decision Networks ;

1. Introduction Regardless of the damage mechanisms an asset is susceptible to, critical decisions must be made at every stage in the life cycle management process. These include materials selection, design and construction, operating conditions, inspection planning, damage identification, analysis, maintenance, replacement, retirement, et cetera. Each of these decisions has an associated cost on one side as well as a benefit on the other. For example: (i) if the cost of inspection and repair were zero, then one would simply choose to always inspect and repair everything that was found, (ii) if the cost of Inconel 625 were as cheap as carbon steel, then one would use it more frequently in corrosive service(s), and (iii) if there were no consequence of failure, one would run every piece of equipment right up until it failed. Of course, these are not realistic possibilities, and this shows that more practical decision strategies must be found that provide the proper balance between cost and benefit when performing life cycle management (LCM). All decisions that affect the return on investment (ROI), throughout the equipment’s entire life-cycle, must be considered in order to use the approach described here, which provides the maximum possible advantage. The computational engine for ProbFat is a Bayesian Decision Network (BDN) that incorporates knowledge from multiple sources (OEM data, field data, experimental data, expert knowledge, models, supplemental software, sensors, etc.) for the purpose of optimizing life-cycle decision strategies. These optimal decision strategies address questions such as: (i) where to monitor, (ii) when and where to inspect, (iii) what monitoring/inspection techniques to use, (iv) what should be done if damage is detected (continue as normal, perform a repair, adjust storage methods, or plan for a replacement), and (v) what can pro-actively be done to prevent future damage. At each point in time, as new knowledge becomes available, all decision strategies are reevaluated, and the optimal strategy is identified as the one that still maximizes overall return (return = benefit – cost). Bayesian AI methods have been used to solve similar problems in engineering, data mining, volcano monitoring, and autonomous driving applications, all of which rely on continuously processing large volumes of data and accounting for inherent uncertainty, Ayello (2014), Cannavo (2017), Hackl (2013), Heckerman (1997), Jain (2012), McAllister (2017), Shabarchin (2016), Stenta (2017), and Tan (2017). Some supplementary benefits of the ProbFat BDN include its ability to: (i) recommend optimal engineering decisions even with incomplete, missing or uncertain data (e.g., an uncertain damage state), (ii) provide insight and predictive reasoning throughout all stages of development, regardless of the quality of the data, and (iii) allow data to be integrated from a wide-range of sources including expert knowledge and experience, laboratory experiments, in-field observations, and physical, mathematical and numerical models. Even though fatigue is inherently uncertain, an extensive collection of literature has been published over the past half a century that has led to advanced technology development, experimental and in-field data, scientific discoveries, and elaborate codes, standards, and industry wide best practices. The ProbFat BDN incorporates and properly balances all these disparate sources of information as well as any future research results and data as they become available. Due to inherent uncertainty, it may not be feasible to exactly know the time-varying fatigue damage rates at all locations in an asset. Fortunately, these rates may not need to be known beyond some degree of accuracy to make the sorts of lifecycle decisions that are needed. Hence, one needs to better understand how uncertainty in these rates affects the decisions that must be made and to always make the best decision regardless of the current state of knowledge. If more accurate damage rate predictions emerge (empirical or theoretical), they can be incorporated into the proposed BDN framework. However, it is important to know what to do now, with just the current, yet imperfect, state of knowledge that exists today. Throughout the remainder of this paper we walk through probabilistic Bayesian Decision Networks as a method for optimizing life-cycle decisions. This approach is demonstrated for multiple fatigue applications, i.e. both low cycle and high-cycle fatigue. Incorporating such tools into daily operations increases awareness and promotes pro active decision-making, giving facilities the ability to properly manage risk, decrease expenditures and improve reliability.