by Eric Lindgren, John Aldrin, and David Forsyth

Editor’s note: The following is an excerpt of a paper presented at the 30th ASNT Research Symposium in 2022. ASNT members can access the full paper at no cost here. It is also available for purchase to nonmembers at the same link.

Intelligence Augmentation for Aviation-based NDE Data

Abstract

With the increased availability of digital data from nondestructive evaluation (NDE) systems, there is a natural inquisitiveness to explore the use of statistical regression and classification methods for NDE data. A continuous issue for artificial intelligence/machine learning (AI/ML) methods is a question of how much data is required to enable training and how high of fidelity is required for such training. The challenge of relevant NDE-based data for aviation applications is not trivial. There are limited data sets as typical areas with flaws, such as fatigue cracks or corrosion, are repaired as soon as they are detected.

Another challenge with US Air Force (USAF) specific aviation NDE data is the broad range of variables that affect the data. To address the limitations of available data, the approach taken by the USAF NDE community is to integrate attributes of AI/ML with other algorithms for analysis of NDE data, plus integrating human analysis into the final decision-making process. The combination of both statistical analysis of data combined with human analysis to determine if flaws are present has been named intelligence augmentation (IA).

The USAF has a rich history of using IA to analyze large NDE data sets, typically acquired from inspections that use automated scanning to acquire data. USAF research continues in the area of IA for various applications. Future opportunities will include improved integration of models, especially as a function of their maturity through validation.

Background

With the increased availability of digital data from NDE systems,… there is a natural inquisitiveness to explore the use of statistical regression and classification methods for NDE data. These fall into three general categories, namely heuristic or rule-based, model-based, or data-based. The latter category is more commonly known as AI/ML and has two general categories of supervised and unsupervised training.

However, for AI/ML to be successful, a significant amount of data is required to provide “truth” data for either type of training. A continuous issue for AI/ML methods is a question of how much data is required to enable training and how high of fidelity is required for such training. Recent examples of potential limits of AI/ML techniques when applied to large sets, such as the decision to not use facial recognition for the Internal Revenue Service [1], have provided illustrative examples of this challenge.

Aviation-based Data

The challenge of relevant NDE-based data for aviation applications is not trivial. There are limited data sets as typical areas with flaws, such as fatigue cracks or corrosion, are repaired as soon as they are detected. For anticipated fleet-wide concerns, repairs frequently are performed proactively due to the very low acceptable risk tolerance for aviation, which further limits the amount of available data from flaws. Another contributing factor to limited data is that fractography is not always performed for repaired locations, which limits the accuracy of truth data from detected flaws.

Another challenge with USAF-specific aviation NDE data is the broad range of variables that affect the data. These include equipment, structural configurations, and flaw variability. Each of these three major classes can be further decomposed to address their influence in a classification or regression analysis to determine if a flaw is present.

An Alternative Approach that Integrates Attributes of AI/ML

To address the limitations of available data, the approach taken by the USAF NDE community is to integrate attributes of AI/ML with other algorithms for analysis of NDE data, plus integrating human analysis into the final decision-making process. As mentioned, these algorithms fall into three classes: heuristic or rules-based methods, model-based methods, or data-based methods. However, these methods are limited when applied to nuanced data or when seeking outliers in large data sets. For aviation-based NDE data, both of these parameters frequently occur, which is why the approach has included the human analysis of the data. Humans are typically much better at recognizing nuances and outliers in data sets as witnessed by the reversion to human analysis for facial recognition. The combination of both statistical analysis of data combined with human analysis to determine if flaws are present has been named IA.

Read the full paper at https://ndtlibrary.asnt.org/2022/IntelligenceAugmentationforAviationbasedNDEData

Reference

“IRS announces transition away from use of third-party verification involving facial recognition,” IRS News Release available at: https://www.irs.gov/newsroom/irsannounces-transition-away-from-use-of-third-party-verification-involving-facialrecognition

_______

Learn more about ASNT’s Ad Hoc Committee for Artificial Intelligence and Machine Learning:

The purpose of the committee is to investigate how AI/ML is presently affecting or will affect the NDT industry. For more information or to join the committee please email volunteer@asnt.org.