Pump Reliability & Predictive Maintenance (PlusSmart) · 9 min read · Updated November 2026
AI Cameras for Leak, Overflow and Safety Detection: What Works at the Edge
Cameras already watch most plants; nobody watches the cameras. Edge AI turns them into tireless operators that spot a dripping seal, an oil pool, a tank about to overflow, smoke, a missing helmet or a person in a red zone — and raise the alarm in SCADA within seconds. What edge AI can and cannot do, how to specify cameras for it, and where it fits beside vibration and oil monitoring.
Condition monitoring sees inside the machine; it does not see the oil pool under it, the seal weeping onto the floor, the relief valve venting, the sump overflowing or the operator working without eye protection. Those are visual events, and plants have always relied on patrols to find them. PlusVision runs detection models on the camera itself or on an edge station next to it, on existing IP-camera streams or new industrial cameras, and sends alerts — not video — to the right people and to SCADA, historian or DCS through a REST API. The video stays on site; the plant gets an event log instead of a recording nobody reviews.
Why edge rather than cloud
| Criterion | Streaming to cloud or central servers | Edge AI (PlusVision) |
|---|---|---|
| Bandwidth | Continuous full video | Only alerts and short clips |
| Alert latency | Seconds to minutes, link dependent | Near-instant on site |
| Remote sites, weak networks | Hard to deploy | Works offline, syncs when connected |
| Security and privacy | Video leaves the plant | Video stays; only events shared |
| Infrastructure | GPU servers or cloud cost | Lightweight Linux on camera or edge station |
What it detects, and what it does not
Process and equipment: pipe, flange and mechanical-seal leaks; oil leaks and pools; relief-valve discharge; overheating, smoke and fire; tank overflow, blocked sumps, slurry spills and slippery areas; reading analogue gauges and valve positions; verifying lockout isolation. Safety: PPE (helmet, glasses, hearing protection, harness at height), hot-work conditions (extinguisher present), confined-space entry, operator absence or fatigue, slips and falls, people under suspended loads, vehicle–pedestrian separation. Security: perimeter and red-zone intrusion, muster-point headcount, smoking in prohibited areas, abandoned objects. What it does not do: replace certified safety systems (light curtains, gas detectors, fire detection to code) or guarantee detection in conditions the camera cannot see. It complements patrols and certified systems and makes both more effective.
Specifying cameras for AI
- Pixels on target. The object to recognise — an oil drop, a helmet, a gauge needle — must cover enough pixels; gauge reading needs a close camera or a long focal length, PPE detection needs people at a reasonable distance. Decide the detection task first, then the lens and position.
- Lighting. Avoid backlight; plan for day–night transitions; dark bays need IR or additional lighting. Leak detection on a wet floor at night is the classic failure.
- Environment. Housing IP rating for dust, rain and chemicals; washdown-rated where cleaned; Ex-certified enclosures in hazardous areas; sunshades and wipers outdoors.
- Network. Cameras and edge stations in a segregated OT zone with only the connections to SCADA opened — see OT network segmentation. The REST API endpoint is the one conduit.
- Existing cameras. Most IP cameras with RTSP streams can be used; the model runs on an edge station. New PlusSmart cameras include the compute.
Deployment in four steps
- Define goals and locations. Safety, equipment reliability or process efficiency — pick the highest-value areas: pump bays with seal-leak history, tank farms with overflow risk, kiln and furnace platforms for PPE, loading areas for vehicle–pedestrian conflict.
- Choose hardware. Existing cameras plus edge station, or new cameras; set view angle, lighting and distance per task.
- Integrate and tune. Connect to SCADA/historian/DCS via REST; tune models to each area's lighting, background and normal activity; define alert routing per shift.
- Operate and improve. Track precision (true alerts versus false) per camera and retune; alerts go to the person who can act on that shift.
Where it fits with condition monitoring
A seal leak detected by camera is the visual end of a chain that vibration and drive data saw earlier; the camera closes the loop when a pump has no sensors or when the leak is the first sign (static equipment, flanges). Combined with i-ALERT and the Analytics ranking, the camera's events become another input to the asset list — a pump that leaks and vibrates moves up — and for safety the camera covers what no machine sensor can. Start with two or three locations where the concern is clear, measure the alert quality for a month, then extend.
Need engineering help with this?
Our engineers apply this to real projects every week — selection, specification, commissioning. Send your question or a part number.
Quick answers
Can we use our existing CCTV cameras?
Usually yes. PlusVision runs on existing RTSP video streams via an edge station; camera position, lens and lighting may need adjustment for the detection task.
Does it need a strong internet connection?
No. Processing is at the edge; only alerts and relevant clips are sent, and the system keeps working offline and syncs later.
Can an AI camera replace safety patrols or certified safety systems?
It makes patrols more effective by watching fixed locations continuously, and it complements certified systems; it does not replace light curtains, gas or fire detection required by code.