Leak Detection for Industrial Equipment & Pipes

Leaks are expensive in proportion to how long they go unnoticed. A weeping valve is a housekeeping issue; the same valve three weeks later is an unplanned shutdown, a slip hazard, a product loss and possibly a reportable environmental release. The trouble is that the people who would spot it early rarely walk past it — pump stations, valve manifolds and tank farms are inspected on a schedule, and a leak that starts the day after an inspection has days to develop. Leak detection from video closes that interval, by watching the same assets continuously instead of periodically.

Automated inspection sits within a broader machine-vision practice; Machine vision covers how imaging systems are used for industrial inspection, and ISO maintains the standards framework that industrial asset management programmes are typically built against.

The Three Failure Signatures

Video does not detect “a leak” in the abstract. It detects visible consequences, and there are three that matter.

• Dripping — a recurring, localised change at a fixed point on a pipe, flange or seal. The signature is periodicity rather than volume, which is what distinguishes a drip from a shadow.
• Flowing — a continuous vertical streak or a spray cone from a pressurised line. Easier to see, but often mistaken for steam or condensate, so the classifier has to be trained against those.
• Pooling — liquid accumulating on the floor or bund. This is the one that gets missed longest, because it develops slowly, and it is also the one most likely to become a slip or a containment incident.

Most deployments get value from pooling detection first, because it is the cheapest to configure and the most consequential when ignored.

Where to Point the Cameras

Coverage choice matters more than model choice here. The high-yield positions are the places where a leak is both likely and consequential:

1. Pump seals and mechanical seals — the single most common leak source in most plants, and one where early detection genuinely prevents failure.
2. Valve manifolds and flange clusters — many potential leak points in a small area, ideal for one well-placed camera.
3. Tank farms and bunded areas — where containment, not just loss, is the concern.
4. Heat exchangers and condensate returns — where steam, water and product leaks look similar and need careful tuning.
5. Loading and unloading racks — transfer operations concentrate spill risk into short windows.

A practical note on lighting: an outdoor manifold at night with poor illumination will produce unreliable results regardless of algorithm. Most sites either add modest permanent lighting at the manifold or accept reduced sensitivity out of hours. The alternative — thermal imaging — is excellent at spotting temperature anomalies but is a different sensor entirely, and the two are complementary rather than interchangeable.

Configuration and the Steam Problem

The hardest part of commissioning is teaching the system what not to report. Steam, condensate, rain on a lens, shadow movement and routine washdowns all look like liquid to a naive model. The controls that work are: a region of interest drawn tightly to the asset, a minimum-duration requirement so transients are ignored, and a reference baseline of the scene under normal operation captured across a full production cycle.

Our edge devices run 2–128 channels at 1–256 TOPS, and the same box carries the full library of 198+ algorithms — so a camera watching a manifold can simultaneously run smoke and flame detection or support factory safety rules without extra hardware.

Who This Is For

Refining and petrochemical sites use it for containment and environmental compliance. Chemical and pharmaceutical plants use it where product loss and washdown risk overlap. Food and beverage operations care about slip hazards and hygiene. Water and wastewater utilities apply it to pumping stations that are inspected infrequently. The common thread is a large number of potential leak points and a small number of people available to look at them.

For the on-site view, read industrial safety system.

FAQ

Can it tell water from oil? It can often separate them by appearance — sheen, colour, flow behaviour and how the liquid sits on a surface are all visual cues — but you should treat fluid identification as a confidence indicator rather than a certified measurement. Where fluid type matters for the response, confirm it with the operator’s process knowledge.

How slow a drip can it catch? That depends almost entirely on how many pixels the drip occupies. A drip at close range from a well-placed camera is comfortably detectable; the same drip on a pipe thirty metres away in a wide shot is not. This is a camera-placement question, and it is why we recommend a placement review before hardware selection.

Do we need new cameras? Usually not. The analytics consume existing RTSP streams from a separate edge box, so your cameras, cabling and VMS stay in place. Detection continues with no internet connection, because inference is on-device and video never has to leave the plant.

What does it cost to deploy? One-time hardware sized by channels and compute, with no per-camera monthly subscription — so adding cameras later does not add a recurring bill. Most sites validate accuracy on their own assets with a Starter Kit before committing to a plant-wide rollout.

Tell Us Your Scenario. Send us a list of your highest-consequence leak points and we will tell you how many cameras they actually need. Validate on your own footage with a Starter Kit, or contact us to scope the manifold coverage.