Managing the growing data challenge in crack detection

Advances in EMAT technology provide more insight into pipeline conditions than ever before. As inspection datasets continue to grow, machine learning is helping analysts evaluate information more efficiently, enabling faster and more consistent integrity assessments while keeping expert judgement at the center of decision-making.

Author: Stephan Eule, ROSEN Group

Crack detection: A discipline built on evidence

Inspection technologies have evolved significantly over the past decade, enabling operators to collect more information about pipeline condition than ever before. Yet crack detection remains one of the most challenging areas of pipeline integrity.

Stress corrosion cracking, fatigue cracking, and seam weld anomalies create complex inspection signatures that require more than advanced tools alone. Reliable assessment depends on a combination of technology, in-field verification, engineering judgement, and decades of accumulated expertise.

EMAT technology has become a leading crack-detection solution for gas pipelines. Recent generations of EMAT tools provide higher channel counts, denser sampling, and improved signal quality, delivering a much richer view of pipeline condition. These advances increase the likelihood of identifying smaller or more subtle indications, but they also generate significantly larger datasets that must be analyzed and interpreted.

Why EMAT analysis requires specialist expertise

Unlike some other inspection technologies, EMAT does not directly measure crack depth. Analysts must interpret guided-wave behavior by evaluating reflections, attenuation patterns, transmitted signals, and interactions across multiple channels.

Different conditions can produce similar responses. Crack colonies, corrosion, manufacturing features, surface roughness, and sensor liftoff may all affect the signal in comparable ways. Distinguishing between them depends on experience, context, and a deep understanding of both the technology and the pipeline being assessed.

Sizing adds another level of complexity. While identifying a crack indication is challenging, estimating its actual depth requires even greater confidence and precision. Structured qualification processes and review procedures help ensure reliable and auditable results, but interpretation remains a task that relies heavily on expert judgement.

More data, more decisions

Improved sensor performance is a major benefit for operators. Higher-resolution data provides stronger diagnostic evidence and earlier insight into developing threats.

At the same time, richer datasets create new challenges. More signals per kilometer lead to a higher number of candidate indications to review and consequently more decisions for analysts to make. Experienced EMAT analysts remain one of the most valuable resources in the crack detection process. Building this expertise takes years, so using analyst time efficiently becomes increasingly important as data volumes grow.

Where machine learning creates value

Machine learning models, and in particular deep learning models trained on ILI signal data, are helping analysts manage this growing analytical workload more efficiently.

Rather than replacing analysts, machine-learning models can prioritize signals that require detailed review, filter benign routine cases, and support more consistent screening across large datasets. This reduces the time spent evaluating inspection data while helping analysts focus their expertise on the indications that matter most.

Several AI model applications are particularly valuable. Triage models can rank candidate indications based on their likelihood of being injurious crack-like features. Classification models can help identify non-injurious signals associated with coating disbondment, manufacturing features, or geometric effects. Machine learning can also support consistency by highlighting cases that deserve additional review and providing supplementary information for crack sizing decisions.

In all cases, the analyst retains full responsibility for the final assessment; machine learning supports the decision-making process but does not replace expert judgement.

Models depend on trusted data

The quality of a machine-learning model depends on the quality of the data behind it. For crack detection, verified field outcomes provide the most valuable foundation. Excavation findings, in-ditch inspections, laboratory investigations, and recovered pipe samples establish the relationship between inspection signals and actual feature characteristics.

The most effective models connect three elements: the original inspection signal, the analyst interpretation, and the verified field outcome. This traceability is essential for achieving reliable and defensible performance.

Trust also requires strong governance. Models must be validated using industry-relevant performance metrics, continuously monitored, and supported by transparent version control and quality processes. In a safety-critical environment, confidence depends on more than accuracy alone. It depends on evidence, traceability, and accountability.

Looking ahead

One of the most promising developments is the ability to evaluate information from multiple inspection technologies together. Combining EMAT, MFL, geometry data, and historical inspection records can reveal relationships and patterns that are difficult to assess manually across an entire pipeline system.

As inspection technologies continue to evolve, the industry challenge is no longer collecting more information. It is transforming increasingly large and complex datasets into actionable integrity insights efficiently and reliably.

The future of crack detection will be shaped by four connected advances: better sensors, trusted data, effective machine-learning tools, and highly skilled analysts. Together, these capabilities will help operators make faster, more consistent, and more confident integrity decisions while ensuring that human expertise remains at the center of the process.