# Can AI support risk-based inspection?

> Artificial intelligence (AI), mainly machine learning, can support parts of risk-based inspection (RBI), such as risk screening. Peer-reviewed studies also show it predicting corrosion rates from data, but results depend on data quality and do not extend reliably beyond the data used. Responsibility for inspection decisions stays with the duty holder and competent people.

- Label: Fact
- Type: Explainer
- Published: 2026-11-14T11:53:52+09:00
- By: Jun Zhong
- Researched and drafted by: Beacon's AI newsroom
- Checked by: Jun Zhong on 2026-11-14
- Topics: LNG, Refining, Upstream and offshore
- Canonical: https://beacon.winda.ai/library/ai-in-rbi/

## Why does it matter?

In a 2019 peer-reviewed study, a machine learning model sorted 1,581 groups of piping and pressure vessels into low or higher RBI risk with about 92% accuracy, yet its authors advised that engineers re-check every item it rated low risk [1]. The reason was the cost of error: calling a medium- or high-risk item low risk can lead to under-inspection [1]. The model also learned from risk categories produced by software based on API RP 581, so it reproduced that method's output rather than replacing the method [1].

## Where can machine learning fit in an RBI program?

| RBI task | What peer-reviewed work reports | Limits the authors report |
|---|---|---|
| Screening equipment into risk categories | Ensemble models (gradient boosting, random forests) classified items with about 92% accuracy and varied less than human appraisers in a 20-item comparison [1] | Unbalanced data; may only carry over to data produced by the same RBI method; ensemble models are harder to interpret [1] |
| Predicting corrosion rates | A 2022 review of 34 machine learning works in 19 papers found models performed better when exposure time was an input [2] | Corrosion data are often incomplete, noisy and mixed in origin; models extrapolated poorly beyond the range of their training data [2] |
| Interpreting results | Rule extraction and argument-based learning are being developed to make models easier to understand [2] | Translating a model's internal workings into formulas is difficult [2] |

Sources: Rachman and Ratnayake (2019) [1]; Coelho et al. (2022) [2].

RBI itself is described in API RP 580 (4th edition, August 2023), which sets minimum and recommended program elements, and API RP 581 (3rd edition, April 2016, with two addenda), which gives quantitative calculation procedures [4]. Neither catalog entry mentions machine learning [4].

## What did the screening study find?

The study by Rachman and Ratnayake used an existing detailed RBI assessment of three onshore and four offshore oil and gas production and processing units, covering more than 200 pressure vessels and 3,000 piping lines [1]. Each of its 1,581 records stood for a group of items with similar features, labeled low risk or medium-to-high risk [1].

- **Inputs.** Ten fields an RBI team already holds: material, component type, diameter, age, post-weld heat treatment, insulation type, operating pressure and temperature, fluid, and corrosion allowance [1].
- **Results.** Gradient boosting reached 92.33% accuracy and random forests 92.18%; the difference was not statistically significant [1].
- **Against people.** On 20 sample items, the best model scored 95% accuracy against 80% for the best human appraiser, though it still misclassified one item [1].
- **Limits.** Only 23% of records were low risk, and the authors stated the model should not be relied on alone [1].

## What limits do reviews and regulators point to?

The 2022 review in npj Materials Degradation found that in one study the models were accurate only for low corrosion rates because high-rate data were scarce [2]. Small datasets sometimes left no separate test data at all, and only one of the reviewed papers shared its code [2]. The authors also noted that corrosion data are often published only in figures, not in machine-readable tables [2].

The US National Institute of Standards and Technology (NIST) lists validity and reliability first among the characteristics of trustworthy AI [3]. In the framework's terms, a model that is right on the data it learned from but loses accuracy on new data, or in a setting it was not trained for, adds to the risk of using AI at all [3].

## Who is responsible for the inspection decision?

The UK Health and Safety Executive (HSE) states that its goal-setting law applies whatever technology is used, and that uses of AI affecting workplace health and safety need a risk assessment and controls [5].

In the US, the process safety management standard requires inspection and testing procedures to follow recognized and generally accepted good engineering practices, and inspection frequency to be consistent with good engineering practice and operating experience [6]. The standard places these duties on the employer and does not mention AI tools [6].

## Where integrity problems show up

| Decision | Data it depends on | What goes wrong without it | In the Library |
|---|---|---|---|
| Rating an item low risk | Complete, consistent equipment records | A medium- or high-risk item rated low can be under-inspected; the authors had engineers re-check low-risk calls [1] | [Qualitative vs semi-quantitative vs quantitative RBI](/library/qualitative-vs-quantitative-rbi/) |
| Predicting a corrosion rate | Thickness histories that include high-rate cases | A model trained mostly on low rates was accurate only for low rates [2] | [How is machine learning used to predict corrosion rates?](/library/machine-learning-corrosion-prediction/) |
| Using a model at another plant | Data produced the same way, under similar conditions | Results may carry over only to data from the same RBI method [1]; poor generalization raises AI risk [3] | Machine learning vs physics-based models for asset integrity |
| Keeping data usable | Structured, machine-readable inspection records | Corrosion data locked in figures cannot be reused for modeling [2] | [What is an inspection data management system (IDMS)?](/library/inspection-data-management-system/) |

Sources: Rachman and Ratnayake (2019) [1]; Coelho et al. (2022) [2]; NIST AI 100-1 (2023) [3].

## Sources

1. [Machine learning approach for risk-based inspection screening assessment (A. Rachman, R. M. C. Ratnayake), Reliability Engineering and System Safety 185, 518-532, doi:10.1016/j.ress.2019.02.008 (accepted manuscript), University of Stavanger open archive (Brage); journal published by Elsevier](https://uis.brage.unit.no/uis-xmlui/bitstream/handle/11250/2984494/Machine_Learning_manuscript_revised.pdf?sequence=5)
2. [Reviewing machine learning of corrosion prediction in a data-oriented perspective (L. B. Coelho, D. Zhang, Y. Van Ingelgem, D. Steckelmacher, A. Nowé, H. Terryn), npj Materials Degradation 6, 8, doi:10.1038/s41529-022-00218-4, npj Materials Degradation (Nature Portfolio), open access](https://www.nature.com/articles/s41529-022-00218-4), 2022-01-26
3. [Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1, US National Institute of Standards and Technology (NIST)](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf)
4. [2025 Publications Catalog (entries for RP 580 and RP 581) (RP 580, 4th edition, August 2023; RP 581, 3rd edition, April 2016, with Addendum 1 (April 2019) and Addendum 2 (October 2020)), American Petroleum Institute (API)](https://www.api.org/-/media/files/publications/2025-catalog/2025_pubs%20catalog.pdf)
5. [HSE's regulatory approach to artificial intelligence (AI), UK Health and Safety Executive (HSE) Media Centre](https://press.hse.gov.uk/2026/06/12/hses-regulatory-approach-to-artificial-intelligence-ai/), 2026-06-12
6. [29 CFR 1910.119, Process safety management of highly hazardous chemicals, paragraph (j) Mechanical integrity, eCFR, US Office of the Federal Register](https://www.ecfr.gov/current/title-29/section-1910.119)

Published by Winda Beacon (https://beacon.winda.ai).
