RGB High-Pressure Leak Detection

← Back to Model Catalog

The Problem

High-pressure processes — pipelines, wellheads, compressor stations, chemical reactors, and aerospace ground systems among them — carry an inherent risk of loss-of-containment events. When a leak occurs under pressure, it often escapes as a visible, high-velocity jet of liquid and/or vapor rather than a slow drip or pool.

These jetting leaks can develop and escalate quickly. Traditional SCADA and process instrumentation is often tuned to detect pressure or flow deviations only after a leak has grown significant enough to disturb the broader process — by which point valuable response time has already been lost.

FloLink's Approach

We built a lightweight computer vision base model purpose-trained to recognize the visual signature of a high-pressure jetting leak — the plume geometry, motion pattern, and dispersion characteristics of escaping liquid or vapor — from a standard RGB camera. No thermal, hyperspectral, or specialty imaging hardware is required.

In testing, the model has detected jetting leaks at their inception — in many cases before pressure or flow-based SCADA monitoring registers a deviation — which can translate directly into critical extra time to respond, depending on process conditions and the severity of the event.

• Detects jetting liquid and/or vapor leaks from ordinary RGB camera feeds — daytime or nighttime
• Inference tested at under 0.5 seconds on common on-chip camera hardware, enabling real-time alerting at the edge
• Trained to be robust to environmental false triggers — including water droplets on the lens, rain, mist, glare, and other conditions that can visually resemble a leak

This is an AI-generated image of an imaginary industrial facility with a leak through a process flange gasket, used to test the model against a realistic leak scenario. Any similarity to any real industrial plant is coincidental and an artifact of the generative model used.

The bounding box above reflects an actual output of our current base model on this test image — no client footage, past or present, was used to produce this example.

Built In-House — No Client Data Involved

Consistent with how every FloLink model is built, this leak detection base model was developed entirely in-house using our own proprietary training algorithms and internally sourced and synthetically generated imagery. It has not been trained or fine-tuned on any client's proprietary footage.

This base model is already a strong performer out of the box. For a specific client site, we typically fine-tune it further using that client's own camera feeds and site conditions to maximize accuracy and minimize false alarms. As with all our deployments, once that fine-tuning begins, the resulting model becomes the exclusive property of that client and never leaves their premises. Read our full model ownership & data policy →