Keeping Engineering Evidence Visible in AI-Assisted Pipeline Integrity

Machine learning and AI can support pipeline integrity by classifying inspection signals, ranking anomalies, identifying patterns, forecasting degradation and helping engineers prioritize review. These capabilities, however, do not remove dependence on the quality of the underlying engineering evidence.

This white paper develops a series-specific AI Integrity Evidence Trace for the use of AI and machine learning in pipeline-integrity applications. The framework keeps the evidence chain visible from the original engineering task through data, labels, model development, validation, uncertainty, human review and post-deployment monitoring.

The trace focuses on six connected assurance elements:

data suitability — whether input data is relevant, representative, sufficiently complete and appropriately controlled;

label and ground-truth quality — what physical or engineering evidence supports the target values used for learning or evaluation;

model validation — whether performance has been demonstrated using meaningful task-specific metrics and independent evidence;

operating-domain control — where the model is valid and how out-of-domain cases are identified;

human review — how consequential outputs can be challenged, interpreted, overridden and adjudicated; and

lifecycle monitoring — how data drift, model drift, updates, unexpected failure modes and revalidation needs are managed after deployment.

The analysis is informed by publicly available work from PRCI, DNV, NIST and ROSEN, including PRCI research on applying machine learning to pipeline integrity management, DNV-RP-0665 Assurance of Machine Learning Applications, DNV-RP-0671 Assurance of AI-Enabled Systems, the NIST AI Risk Management Framework, and attributed industry material concerning AI-supported integrity analytics.

A central proposition is that an AI result used in pipeline integrity should be treated as a bounded analytical claim rather than self-validating ground truth. A model may generate a probability, ranking or prediction with high numerical confidence while still being limited by weak labels, incomplete physical verification, distribution shift, unrepresentative training data or an operating environment outside its validated domain.

Ground truth is therefore treated as a first-class engineering issue. Verification digs, inspection records, NDT results, expert adjudication and historical outcomes can provide valuable labels, but each has its own limitations. Label provenance and uncertainty should remain visible, and expert-labelled data should not automatically be treated as equivalent to direct physical measurement.

The paper also emphasizes that validation should resemble actual deployment. Random train/test splits can produce optimistic results when closely related observations from the same assets, inspection campaigns or operating regimes appear in both datasets. Validation should therefore consider separation by asset population, geography, vintage, inspection technology, threat type, operating regime or time period where appropriate.

Human–AI collaboration is treated as part of the assurance system rather than an afterthought. Engineers should be able to understand the relevant model evidence, uncertainty, applicability boundaries and low-confidence or out-of-domain conditions. Overrides, disagreements and adjudications should themselves become evidence for model improvement and failure-mode discovery.

Finally, the paper addresses MLOps and continuous assurance. Long-lived pipeline assets experience changes in operating conditions, inspection technologies, sensor systems, datasets and software. Production AI therefore requires model and version control, drift monitoring, rollback capability, documented change triggers, performance monitoring and periodic revalidation.

The AI Integrity Evidence Trace is a pipeline-specific practitioner synthesis, not a new AI standard. It does not replace established DNV, NIST, PRCI, regulatory, operator or engineering assurance processes, and it does not evaluate or endorse any proprietary AI model.

The scope is deliberately limited to publicly discussable asset information, inspection, analytics, assurance and human engineering decision support. AI recommendations are not treated as authority to execute physical actions, and automated pipeline control remains outside the scope of this work.

This paper forms No. 23 of the From Project to Live Asset — Independent Practitioner Research Series, Phase III, extending the wider steel-pipe lifecycle research programme from contextualized data and bounded digital-twin claims into evidence-aware AI-assisted integrity analysis.

Author-written abstract, reproduced from the authoritative Zenodo DOI record.