Problem
A pipeline operator must find defects along hundreds or thousands of kilometers of pipe. A robot runs inside the pipe and records magnetic flux with 64 sensors around the perimeter every 2 mm. Lab staff review this data semi-manually: the report takes about two months and processing costs about 200 thousand rubles per 10 km. More than half of ruptures happen on small, short pipelines, which are inspected less often than long mainlines. Fixing a rupture costs about 3 million rubles per day.
Example from the article: a 15 km run with 64 features, 700 defects and 1400 welds labeled in the lab. The defect log was filled in by hand, and part of the labels were shifted by up to several centimeters.
Solution in brief
Treat magnetograms as images and classify them with a small CNN, after repairing the data: shift misaligned defect coordinates, fill zero stripes left by dead sensors, augment rare classes, normalize each image. Flag anomalous images and let specialists review them. Wrap the model in an offline report generator and dashboard.
Questions this case answers
- How to detect rare defects in multichannel sensor scans with a convolutional network when manual labels are unreliable and some sensors fail
- How to deliver an anomaly detector as an offline on-premise service that flags images for expert review instead of labeling everything
- How to detect defects and welds in in-line inspection magnetograms of a pipeline with a CNN
- How to deal with unreliable manual labels and shifted defect coordinates when training a defect detector
Tools and methods
- Small CNN (2 then 5 layers) — detects defect, weld or normal image
- U-Net — segments defects to measure their size
- Gradient boosting — estimates defect depth from tabular features
- Physics-aware augmentation — enlarges rare classes without breaking geometry
- Gap filling — handles stripes from failed sensors
- On-premise web service — report, dashboard, dispatch
Google or a one-time email code. No password.