Condition Monitoring Optimization Part 2: Making Inference with Uncertain Data

Introduction

This article is Part 2 in a series of articles discussing condition monitoring optimization, which aims to provide a framework for quantitatively optimizing inspection scope, techniques, and intervals based on historical inspection data and subject matter expertise, while also dynamically updating the inspection plan to maximize reliability and return on investment (ROI) as new information becomes available. Using this methodology, data collected through inspection can be used to improve confidence in the asset damage state and determine when additional data is required, when inspection adds little or no value, or when corrective maintenance is required. In Part 1 of this series, entitled “Condition Monitoring Optimization: Going Beyond Traditional CML Optimization” and published in the September/October 2021 issue of Inspectioneering Journal, it was assumed that inspection coverage and techniques were sufficient to capture the true damage state of the asset.

While the above scenario is certainly ideal, it is often not representative of the real-world situations that are encountered in practice. As a specific example, inspection coverage may be limited to a fraction of the total susceptible area of an asset for a variety of reasons, such as inaccessibility. A radiographic testing (RT) scan conducted on a piping elbow, for example, generally captures only a single angle of the potential surface, leaving the inspection professional with imperfect information regarding degradation. Alternatively, even when an inspection does cover the entire area susceptible to a particular damage mechanism, the inspection technique utilized may not provide a 100% probability of detection. Consider the case of using magnetic particle inspection to identify a surface breaking crack – even when a crack is present, this technique may only have a 90% chance of detecting the crack, which leaves the inspection professional with the job of considering that damage may be present even when the inspection technique finds no evidence.

Fundamentally, the industry is faced with the challenge of making statistically meaningful inferences in the presence of limited or potentially erroneous data. The industry can combat this situation by using Bayesian statistical analysis, which combines measured inspection data with prior information derived from subject matter expertise or historical experience. This article will outline this data analysis methodology across a series of practical examples, which focus on local degradation. First, the article will examine the case of thinning on a piping circuit with limited inspection coverage. Second, the article will consider the case of local pitting on heat exchanger tubes with limited inspection data, as well as an imperfect probability of detecting damage where inspections are conducted.

Extreme Value Analysis with Limited Inspection Coverage

While, ideally, one would like to inspect 100% of the susceptible area for any type of damage, such a comprehensive inspection is often either cost-prohibitive (e.g., scanning a large surface area in its entirety) or impossible (e.g., portions of the asset are inaccessible). In such a scenario, one is faced with making an inference about an asset given limited inspection data. In the case where damage is detected using the limited inspection, one can take action to remedy the situation. However, what options exist when no evidence of damage is detected? One can assume that there is no severe damage on the asset, but must also consider the possibility that there is significant damage that has simply not been detected due to the limits in inspection coverage. Ultimately, it’s important to quantifiably answer the following question: What is a reasonable estimate of the damage state of the asset given the data that has been collected?