Next Generation RBI Using Explainable AI (XAI): A CUI Case Study

Introduction

This article describes a next-generation, fully-quantitative, and fully-probabilistic approach to risk-based inspection (RBI) using a form of explainable AI (XAI) that improves probability of failure (POF) calculations for all damage mechanisms, provides more quantitative inspection recommendations, makes localized real-time predictions when connected to streaming sensors, learns and gets smarter with every real-world observation, and replaces the somewhat arbitrary risk target concept with a total cost minimization objective. This method goes beyond traditional RBI to provide total asset lifecycle optimization when design and operational decisions are considered in addition to inspection and maintenance decisions. However, for the sake of brevity, we limit the discussion in this article to RBI for corrosion under insulation (CUI), with a focus on the improved POF calculations. This new XAI RBI methodology constitutes a significant advancement over existing methods in API 581 and, for that reason, can be regarded as a next-generation RBI approach that addresses all the recommendations made in a recent Inspectioneering Journal article for next-generation RBI [1]. Additionally, it offers full transparency into its methods, a necessary requirement for industry acceptance. Even though the focus here is on CUI, it can be easily extended to cover all other damage mechanisms.

The XAI RBI approach presented here is rooted in physics-based, probabilistic, causal network models that mimic human expert intelligence – making good decisions when faced with difficult challenges in an uncertain world – by integrating knowledge from multiple sources (physics-based models, human experts, and field data) and getting smarter over time (improving predictions based on comparisons with real-world observations from inspections, sensor readings, etc.) [2]. The dynamic and continuously updated nature of this approach is a key differentiator over the current static methods. Also, the full transparency of this approach makes it a glass-box, as opposed to the many alternate “black-box” AI methods (that still have a role to play within this larger XAI RBI approach). The goal of XAI, and a necessary requirement for any AI method to be adopted by industry, is to foster trust, transparency, and accountability by allowing users to understand how the system works, verify its correctness, and identify potential biases and errors. 

A case study from an actual CUI pilot by one of the early adopters of this technology reveals a total potential savings of about $38 million for just one unit had this XAI RBI approach been followed from the start (which was not the case, unfortunately). These savings could have been realized by preventing 20 failures (approximately $20 million in savings) and significantly reducing inspection costs through more focused inspections (approximately $18 million in savings). It should be noted that this assessment was made by using readily available data and information and did not require any additional data collection efforts.

Separately, one of the most infamous failures in the fertilizer industry, the explosion at a European ammonia plant in 2009, was caused by CUI in an unused cold bypass line that was not properly accounted for (i.e., the bypass line was modeled using the main line’s operating temperature and not the actual pipe temperature of the bypass line where the failure occurred) [3]. The insulation was damaged, and the coating quality was poor. This allowed water to make direct contact with bare metal, a necessary precursor for CUI. The plant was shut down for 12 months, resulting in a production loss of 420,000 tons of ammonia and 920,000 tons of nitrate. Had the XAI RBI approach described in this article been used, this disaster may have been avoided.