Pipeline & Corrosion Vision Inspection AI

What problem does this solve?

Aerial and robotic inspection has solved the collection problem and created a review problem. A single drone survey of a pipeline corridor, a tank farm or an above-ground installation produces thousands of high-resolution frames; a crawler run inside a line produces hours of video. The people qualified to interpret that material are the same small group of integrity and coating inspectors who are already committed to field work, so imagery accumulates and is reviewed selectively, in a hurry, and inconsistently between reviewers. The value of flying more often disappears if nobody can look at the results.

Location is the second recurring failure. A photograph of corrosion is only actionable if someone can say which line, which chainage or kilometre post, which support or which weld it belongs to. In practice the imagery arrives with GNSS coordinates that are good to a few metres, no linear referencing, and file names that encode the flight rather than the asset. Findings then have to be manually matched to the pipeline route in GIS before a repair crew can be sent, and that matching effort often exceeds the review effort.

The third problem is that findings do not survive into the systems that matter. A defect noted in an inspection report is not the same thing as an anomaly in the integrity management register with a severity grade, a re-inspection interval and a work order. Reports get filed as PDFs, the next survey starts from scratch, and comparison over time — which is the entire point of periodic inspection — becomes impossible because the same defect was recorded differently by two reviewers three years apart.

How the solution works

The workflow is built in the correct sequence: georeference first, detect second, review third, integrate fourth. Imagery from drones, crawlers, thermal cameras, rope-access and handheld capture is tied to the asset through GNSS, IMU, odometer readings and the pipeline's linear referencing, so every frame has a chainage or an asset identifier before any model runs. AI models then screen the imagery for corrosion and rust staining, coating and insulation breakdown, damaged supports and clamps, thermal anomalies suggesting insulation failure or leakage, and changes in the right-of-way such as new construction, excavation, vegetation encroachment or unauthorised crossings — comparing against the previous survey where one exists.

The output is a prioritised queue of candidate findings, each with an image, a location, a severity indication and a confidence value, not a finished inspection verdict. An inspector confirms, grades or rejects each candidate, and only confirmed findings become anomalies in the integrity management system and features in GIS, with work orders raised for repair, recoating or right-of-way clearance. Inspector decisions are captured as labels, so accuracy improves over successive surveys. Swedish Technology delivers this as custom development against the operator's own GIS, integrity and maintenance platforms, with the models, pipelines and integration code handed over.

1. 1
   Input Pipeline route and linear referencing data, asset register, previous inspection findings, the survey plan, and the defect taxonomy the operator already uses in its integrity procedures.
2. 2
   Capture Drone or fixed-wing survey of the corridor and above-ground installations, crawler or robotic capture inside lines and vessels, thermal imaging for insulation and leak indications, plus handheld or rope-access imagery for close-up work — with GNSS, IMU and odometry logged throughout.
3. 3
   Georeferencing & alignment Frames are positioned along the route using linear referencing, matched to assets, welds, supports and previous survey positions, and where relevant orthorectified or assembled into a corridor mosaic so the same location can be compared across years.
4. 4
   Processing / detection Models segment corrosion and rust staining, coating and insulation defects, structural damage on supports and clamps, thermal anomalies, and land-cover or object changes in the right-of-way, each with a confidence value and an extent measurement in image or ground units.
5. 5
   Severity scoring & prioritisation Findings are scored using the operator's own criteria — defect extent, location criticality, proximity to population or crossings, coating class and time since last inspection — to produce a review queue ordered by risk rather than by flight sequence.
6. 6
   Inspector review A qualified inspector confirms, grades or rejects each candidate in a review interface showing the image, the map location and the history of that position. Nothing reaches the integrity register without this step.
7. 7
   Integration & action Confirmed findings are written to GIS as located features and to the integrity management system as anomalies, and generate work orders for repair, recoating, NDT verification or right-of-way clearance in the maintenance system.
8. 8
   Reporting & learning Corridor-level reporting on coating condition, encroachment trends and repair backlog; inspector decisions returned as training labels; model versions recorded against every finding for auditability.

Reference architecture

Imagery is heavy, findings are light, and the integrity record is the system of truth. The layers are separated so that survey volume never destabilises the operational systems.

Deployment options: Corridor imagery of national infrastructure is sensitive in most GCC jurisdictions, so the reference deployment is on-premise or in a customer-controlled UAE-region private cloud, with no imagery leaving the operator's environment. Air-gapped operation is supported: survey media is ingested locally and model updates are delivered as signed offline packages.

Key capabilities

Integrations

Findings are only useful inside the operator's existing integrity, spatial and maintenance systems. All integrations are engineered as custom development against the customer's own platforms and licences.

Industry use cases

Implementation approach

1. 1
   Scoping and defect taxonomy (1–2 weeks) Agree which defect types are in scope, using the operator's existing integrity taxonomy and severity grades rather than inventing new categories that will not map to the anomaly register.
2. 2
   Imagery assessment Review existing survey data for ground sample distance, overlap, blur, exposure and positional quality; state which defect types are detectable at that quality and what capture standard would be needed for the rest.
3. 3
   Capture standard definition Define flight altitude, overlap, camera and thermal settings, crawler speed and lighting, and the positioning data to be logged, so surveys are comparable between campaigns and vendors.
4. 4
   Georeferencing pipeline build Ingest, quality screening, linear referencing to route and chainage, asset matching and alignment with previous surveys — this is usually the largest engineering item and the one most often underestimated.
5. 5
   Model development and evaluation Train on labelled imagery from the operator's own assets; report detection and false-positive rates per defect type against inspector-labelled ground truth on a held-out segment.
6. 6
   Review workflow and severity rules Build the inspector review interface, configure severity scoring with the integrity team, and define who confirms what and at which authority level.
7. 7
   Integration build GIS feature publication, integrity management anomaly creation, CMMS work orders, document filing and reporting, each with an agreed field-level interface contract.
8. 8
   Pilot survey and acceptance Run a complete cycle on one segment or site: capture, detection, review, integration and repair dispatch, with acceptance measured on inspector agreement and time saved per survey.
9. 9
   Roll-out and hand-over Extend to the full programme, train inspectors in Arabic and English, hand over models, datasets, pipelines and source, and set the retraining cadence to follow each survey campaign.

Security & deployment

Imagery of pipeline corridors and energy infrastructure is handled as sensitive throughout: ingest, storage, processing and review all take place inside the operator's environment, with no outbound connectivity required and air-gapped operation supported. Access to imagery, findings and export functions is role-based and logged, and exports carry watermarks and an audit record where the operator requires it. Every finding stores the model version, the confidence value, the source frame reference and the identity and timestamp of the inspector who confirmed or rejected it, so an integrity decision can be reconstructed years later during an audit or incident investigation. Retention rules for raw imagery are agreed explicitly, since full-resolution corridor imagery grows quickly and carries its own security classification.

Limitations & prerequisites

• Visual and thermal inspection assesses surfaces and appearance only. It cannot measure wall thickness or detect internal metal loss; ultrasonic, radiographic or in-line inspection remains the source of truth for remaining wall and defect depth.
• Corrosion under insulation is usually invisible until it produces an external symptom. Thermal and visual screening narrows where to strip and inspect; it does not confirm the condition beneath the cladding.
• Buried pipeline cannot be inspected visually. For buried sections these methods address the right-of-way, surface indications and exposed crossings, not the pipe itself.
• Detection quality is bounded by ground sample distance, blur, exposure, shadow, glare on coated or reflective surfaces and, for corridor surveys, by season and vegetation state. A survey flown for one purpose is often unsuitable for another.
• Georeferencing accuracy limits everything downstream. Without RTK or good odometry and a maintained linear referencing system, findings arrive with locations too coarse for a crew to act on directly.
• Change detection needs comparable surveys. Different altitudes, cameras, seasons or times of day generate large numbers of apparent changes that are not real changes.
• The system produces candidates, not verdicts. Grading a defect and deciding on repair remains the responsibility of a qualified inspector, and regulatory regimes generally require that; the AI layer reduces the volume that requires expert attention rather than replacing it.
• Rare defect types will have few training examples in any single operator's history, so early performance on those classes is weak and improves only as inspector-labelled findings accumulate over successive campaigns.

Manual imagery review vs rule-based image processing vs AI screening vs NDT

These methods answer different questions. AI screening changes how much imagery a human has to look at; it does not change what visual inspection is capable of measuring.

The workable combination is AI screening to prioritise, inspectors to confirm and grade, and NDT targeted at what the screening and the risk model together identify as worth measuring.

FAQ

Related topics & solutions

• Industrial Vision Inspection with AI
• Predictive Maintenance AI for Industrial Assets
• Site Safety Computer Vision (PPE, Zones, Vehicles)
• GeoAI Object Detection with ArcGIS
• Road Inspection & Pavement Condition Intelligence in GIS
• Smart Inspection
• RASM ÔÇô Digital Twin
• Document Management & Correspondence System