Graduate Research Lightning Round

Complex engineering systems (CESs), such as nuclear power plants, are critical to a wide range of industries and utilities; as such, it is important to be able to monitor their system health and make informed decisions on maintenance and risk management practices. However, currently available system-level monitoring approaches either ignore complex dependencies within a CES or are intended for simpler systems. The gap found over the systematic integration of prognostics and health management (PHM) and probabilistic risk assessment (SIPPRA) for CESs needs to be closed through the development of techniques and models that consider the causal factors and operational context of the system when generating health assessments.

My dissertation research is the result of a concentrated study to address one of the challenges of system-level SIPPRA for CESs: how to appropriately segment continuous operational timelines over meaningful time periods to create the discrete time-series data used for health model inputs. This research studies how different time scales and discretization approaches impact the predictions of system health and operational status generated from a dynamic Bayesian Network (DBN), a model that is increasingly used for causal-based inferences and system-level assessments. This work identifies and defines different time segmentation strategies and performs a comparison of models built from these methods based on simulated nuclear data from a sample scenario of a sodium fast reactor (SFR) experiencing an accident event. By understanding how these discretization methods impact the model’s health assessment and other model characteristics, future risk models can be developed to provide a more meaningful assessment of a system’s health, from which decision-makers can make more informed decisions.

The main strength of quantitative risk assessment (QRA) is to enable risk management of engineering systems. Bayesian Networks (BNs) have become widely popular in QRA for their ability to incorporate multiple data sources, model multivariate joint probability distributions, and provide a causal structure representing an analyst's knowledge of a system's behavior. These capabilities have enabled QRA practitioners to use associative causal reasoning to support risk management, answering questions such as "How does new evidence x about X change the probability of observing event Y?" This enables deeper insights into which factors drive risk. However, associative reasoning provides no information on how intervening on those factors (through risk management decisions) changes risk. BNs can tackle this issue through more complex causal reasoning based on interventions and counterfactuals. Intervention reasoning, namely "How does doing X = x change the probability of event Y occurrence?" can be used to calculate the causal effect of risk management decisions on risk. Moreover, counterfactual reasoning, namely "If X had been x, would the event Y have been avoided?" can test the effect of risk management decisions on known undesired past events. These causal insights are highly relevant as they acknowledge and confirm risk managers' role in a system's safety.

In this research, we propose to expand the use of BNs in QRA to support intervention and counterfactual reasoning, thereby enhancing risk-informed decision support for engineering systems. In order to do this, we establish the mathematical background and methods to model interventions and counterfactuals. In addition, their use is illustrated for two case studies on pipeline excavation damage. The capabilities of incorporating interventions and counterfactuals to the current BN methods for QRA is demonstrated and validated on a thorough risk assessment for natural gas pipeline excavation damage in the US. The impact of the proposed research is a first-of-its-kind causal reasoning approach on QRA to enable risk management.

Probabilistic Risk Assessment (PRA) is one of the technologies that is used to inform the design, licensing, operation, and maintenance activities of nuclear power plants (NPPs). A PRA can be performed by considering the single hazards (e.g., earthquake, flood, high wind, landslide) or by considering multi-hazards (e.g., earthquake and tsunami, high wind and internal fire). Single hazards PRA has been thought sufficient to cover the analysis of the severe accident until the Fukushima Daiichi NPP accident in 2011. Since then, efforts have been made to consider multi-hazards as well; thus, multi-hazard PRAs are starting to be seen as being indispensable for NPPs. Aside from the changing frequency of global and local natural hazards, other reasons to be highlighted are that the number and diversity of NPPs will probably increase. Moreover, advanced reactors are getting close to becoming a reality by designing them with passive safety systems, smaller, standardized, and even transportable to make them cheaper across the design, licensing construction, and operation stages. Thus, multi-hazards should be addressed in any future full-scope PRA. Although I found a few studies discussing multi-hazards, a general framework for multi-hazards PRA is still missing. I argue that the starting point for any multi-hazard PRA general framework should be the Advanced Non-LWR Licensing Basis Event Selection (LBE) Approach and Probabilistic Risk Assessment Standard for Non-Light Water Reactor (non-LWR) Nuclear Power Plants. For PRA, history has shown us the path forward before, with Three Mile Accident being seen as one milestone to understand the necessity of PRA. The Fukushima Daiichi NPP Accident is another milestone in the development of PRA, showing the need for performing multi-hazard PRA for the current and future NPPs.

Since the Fukushima Daiichi accident in 2011, there has been a growing need for the development of alternative nuclear reactor fuel systems with enhanced accident tolerance, referred to as accident tolerant fuels (ATFs). ATF can tolerate active cooling loss in the reactor core longer compared to the currently used UO2-Zr fuel system. It is expected to delay or prevent core damage by providing additional coping time for accident mitigation during accidents involving loss of core cooling. Before the commercial deployment, ATF safety benefits need to be assessed to identify its viability for future use. The effect of the increased safety margins depends on accident progression; thus, age-related degradation of components that affects the plant response during accidents should be considered for a more accurate evaluation of ATF safety advantages. This study assesses the safety benefits of ATF during accidents and how they change over time in consideration of component degradation.

Human Reliability Analysis (HRA) is an integral component in probabilistic risk analysis (PRA) in multiple risk-critical industries. Human-machine teams (HMTs) maintain an integral role in the safe operation of large systems, such as nuclear reactors, and will continue to do so even under increased automation. As a result, it is critical to develop methods capable of appropriately characterizing HMT performance. This performance does not take place in a vacuum but in a complex, dynamic and highly-entangled environment. However, current HRA methods are largely centered on the notion of independence in both HMT actions and contextual performance influencing factors (PIFs).

To close this gap, I propose to undergo a comprehensive overhaul of the conceptualization and treatment of dependency in HRA. This project will analyze how dependency is treated in current HRA methods, as well as how these methods conceptualize the idea of dependency and communicate that to the end-user. This research will put forward a single, complete, and appropriate definition for dependency as a general concept and rectify confusion surrounding the terminology and mathematics used for multiple dependency-relevant aspects of HRA. Secondly, this work will identify specific relationship archetypes (dependency idioms) which are likely to be common in HRA contexts and form the basis of a causal model of dependency in HRA. This project will then lay out an exhaustive framework for identifying and modeling dependencies using Bayesian Networks (BNs). Finally, a full mathematical framework for quantifying the effects of dependency relationships will be developed which uses BNs to evaluate the numerical effects of dependency relationships on conditional probabilities in HRA models.

Supply chain (SC) shortages are defined as the deficit of units that occur when the demand on a SC exceeds its capacity. The increase in demand may have a fixed value or a range of values. As such, in a SC model the probabilities with which the increase in demand will occur will also have a probability of occurrence associated with it. Notable examples of effects of dynamically increasing demand include drug manufacturing shortages, food shortages, and power shortages.

Demand-Capacity interference theory is used to quantify the probability with which demand will exceed capacity, given that demand and capacity are random variables described by appropriate probability distribution functions. The overlap between the two distributions is called interference and represents the probability of shortage. In this work we are developing a methodology to quantify the likelihood of shortages caused by non-linear demand on a SC using interference theory.

The results of this work help us develop the risk profile of a SC subjected to a distribution of demand values. Using the risk profile stakeholders of the SC can make informed decisions stocking inventory, including redundant capacity, and incorporating diverse supply sources.

With the recent developments in nuclear technology, the use of SMR type reactors is expected to become more widespread. SMRs are the designs that are economic, reliable, equipped with passive safety systems, and have modular construction features. There are also aimed to increase the load-following capacity of a power plant by deploying multi-units at the same site. Upon the Fukushima-Daiichi nuclear power plant accident, the need for detailed safety analyzes of multi-unit nuclear reactors has emerged. It is important to carry out these analyzes meticulously, as the shared systems used by the reactors can lead to unexpected results in case of an accident. Therefore, it is critical to perform multi-unit probabilistic safety analyzes (PSA) for SMRs. In the literature, it is seen that there are a few multi-unit traditional and dynamic PSA studies for commercial nuclear reactors, and PSA methods have not yet been applied to SMRs to assess a multi-unit model.

In this context, this dissertation is aimed to do multi-unit dynamic PSA for SMRs. The SMR design to be used in the study was considered as an integrated pressurized water reactor. Dynamic PSA allows obtaining more realistic time-dependent results by combining probabilistic safety and deterministic analysis. The first phase which aims to make a multi-unit reactor deterministic model is to be completed by using RELAP/SCDAPSIM code. In the second phase, a digital twin of the reactor model is to be created by using neural networks in order to reduce the time required to complete the analysis, which is the biggest problem encountered in such analyzes. Also, neural networks would help the number of accident cases to be investigated increase. In the next phase, the dynamic PSA module will be integrated into this system, which creates dynamic event trees that contain accident sequences and their probabilities. In the last phase of the study, as an initiating event, a seismic induced accident will be investigated, and the results will be evaluated for the integrated pressurized water reactor design with twelve reactor units at the same site. Therefore, machine learning techniques offer an opportunity to analyze more data with less computational time compared to the existing nuclear safety software.