BENGI: A Bayesian ENGine for Insights

In the aging energy industries, pressurized fixed equipment and piping systems are exposed to harsh environments and operating conditions that promote a whirlwind of undesirable damage mechanisms. This damage may eventually result in catastrophic failure (leak or rupture) if left unchecked, with potentially significant consequences, especially if there is an environmental release of toxic or flammable fluids. Failure can also lead to loss of production, damage to nearby equipment, personnel injury, loss of life, and repair, replacement, and legal costs. To minimize the likelihood of unexpected failures, proactive inspection and maintenance plans are required. These plans require complex decision-making platforms to assist in integrating data and experience that ultimately end up in a series of decisions. From the foundational work accomplished in Phase I, we have developed an explainable AI engine for making such large-scale complex decisions. The core technology was made possible by leveraging DOE funded tensor engine libraries. This engine, called Bayesian ENGine for Insights (Bengi), is based on probabilistic cause-and-effect diagrams known as Bayesian Decision Networks. By using these networks, decisions under uncertainty are optimized in even the most complex scenarios to minimize cost, maximize profitability, and increase safety. Bayesian Networks properly blend data from various sources, such as human knowledge, mathematical models, numerical simulations, and field observations. In Phase I of this project: (1) We explored more advanced tensor contraction algorithms and data structures. In the end, we were able to integrate DOE HPC libraries into Bengi’s core tensor engine. Bengi now supports multiple backends.; (2) We built networks efficiently with a CPT library and thematic network generator. Now we can quickly generate networks on-the-fly without the use of a user-interface for systems level networks.; (3) Implemented the graph theoretic infrastructure to compute complex decision hierarchies. These algorithms were then used to implement the optimal decisions algorithm leveraging tensor solvers.; and (4) Incorporated these innovations into a user-interface, with a coherent user-experience in mind, to enable a wide array of users to have access to DOE-funded HPC software and resources. In Phase II of this project, the goal is to combine the modules that were developed in Phase I and extend, refine, verify, and commercialize those efforts as part of a complete decision-making solution for the next level of Life Cycle Management. To circumvent the limitations of traditional of Life Cycle Management and address the industry need for a financial-based automated decision support tool, we propose leveraging Bengi to develop Risk-Based Inspection Plus, a probabilistic cost-benefit analysis extension to traditional risk-based inspection that recommends optimal lifecycle decision strategies. During Phase II we will develop a market-ready the Risk-Based Inspection Plus software into a solution centered around facility-wide, financial-based inspection optimization with integrated fitness-for- service capabilities. The project is the fusion of tensor based HPC libraries and asset integrity management