In short: A vision inspection system on one production line is a solved problem in UK manufacturing, with cameras, lighting and a trained model catching defects a tired operator misses at 3am. Scaling from one line to a whole plant is where programmes stop, because every additional camera has historically meant another armoured cable run through a live factory. The fix is to put the inference at the camera and carry results, clips and model updates over a network the plant controls.
Key Takeaways
- The first line proves it, the fortieth kills it — a vision inspection cell pays back within months on the worst-performing line, then the plant-wide rollout meets a cable run, a conduit and a production shutdown for every camera after it.
- Raw machine vision does not belong on a radio link — a 5 MP camera at 60 frames per second produces roughly 300 MB per second uncompressed, so the sensible split is inference at the edge and only results, compressed clips and model updates across the network.
- Redeployable cameras need licensed spectrum — a vision cell that can be wheeled to a new line for a new product run is worth more than a bolted-down one, and that mobility only works on coverage engineered for the whole building rather than patched access point by access point.
In a nutshell
What is a vision inspection system, and what does it actually catch?
A vision inspection system is a camera, controlled lighting and a classifier placed over a production line so that every unit is checked rather than sampled. The classifier may be a traditional rules-based machine vision tool measuring dimensions and presence or absence, or a trained neural model judging surface finish, seal integrity, label placement, weld quality and the long tail of defects that resist being written down as a tolerance.
The appeal is straightforward. Manual visual inspection on a fast line is genuinely hard work, accuracy drops across a shift, and the defects that escape are found by the customer. An inspection cell checks 100% of units at line speed, applies the same judgement at 3am as at 9am, and produces a record of every decision. For a food plant under BRCGS Issue 9, a pharma line under FMD serialisation rules, or an aerospace supplier under AS9100, that record is not a bonus feature; it is the audit trail.
UK adoption has been helped along by the Made Smarter Adoption programme, which has spent several years match-funding exactly this kind of project for small and mid-sized manufacturers, and by the demonstration work at the Manufacturing Technology Centre in Coventry and the National Manufacturing Institute Scotland. The technology is not the risky part any more.
Why do vision inspection system rollouts stall after the first line?
Because the first camera and the fortieth camera are different projects, and only the first one is priced properly.
The pattern repeats across UK plants. Quality picks the line with the worst customer complaint record, an integrator installs a cell, the false-reject rate settles down after a few weeks of tuning, and the payback lands somewhere between six and eighteen months. Everyone agrees it should go everywhere. Then the rollout is scoped and three costs appear that were invisible in the pilot.
Firstly, the cabling. Industrial machine vision has historically been wired for good reason: GigE Vision cameras want a dedicated gigabit link, often with Power over Ethernet, run in armoured conduit from the camera position back to a cabinet. On a live factory floor that means tray work, isolations and a production window to install it, and the installed cost per camera position frequently exceeds the cost of the camera. Across forty positions it turns a compelling case into a capital request nobody wants to sign. This is the same arithmetic that stops condition monitoring programmes at the pilot stage, and brownfield plants do not get to start again the way greenfield reshoring projects do.
Secondly, the cameras want to move. A cell that inspects one product on one line is useful. A cell that can be repositioned for a new SKU, moved to the line that is currently struggling, or wheeled out for a short contract run is considerably more useful, and hard-wiring removes that option. Plants that have hard-wired forty positions tend to discover they have also fixed their inspection strategy for a decade.
Thirdly, nobody owns the layer underneath. The pilot belonged to quality. The rollout needs IT for the network, OT for the line controls, production for access windows and finance for a multi-year case. Programmes die in that gap more often than they die from a technical failure.
Can factory WiFi carry inspection cameras?
It can, for a handful of cameras in a clean radio environment, provided you compress at the camera. In most plants it cannot, and the reason is worth setting out precisely.
Start with the raw numbers, because they defeat any wireless technology if you get the architecture wrong. A 5 megapixel monochrome camera running at 60 frames per second generates in the order of 300 MB per second, or around 2.4 Gbps, uncompressed. That is why machine vision has 10 Gigabit Ethernet and CoaXPress standards rather than making do with the gigabit link. No private 5G network, and no Wi-Fi 6 deployment, is going to carry that, and anyone who tells a plant manager otherwise is selling something.
What actually travels is much smaller. Modern inspection cells run inference at the camera or on a small GPU box in the cell, so what leaves is a pass or fail decision, a few kilobytes of measurement data, and a compressed image or short clip of the rejects. A plant rejecting 2% of units at 60 units per minute is producing a handful of JPEGs a minute, not a video wall. Add periodic model updates pushed down to the cells, remote access for the integrator, and the occasional live stream when an engineer is diagnosing a tuning problem, and the traffic profile is modest but demanding: bursty, uplink-heavy, and intolerant of the multi-second stalls that a busy factory WLAN produces.
That profile is where factory WiFi struggles, for the reasons we have set out in detail on why warehouse WiFi fails. Unlicensed 2.4 and 5 GHz spectrum on a factory floor is shared with scanners, AGVs, contractors' hotspots and a decade of accumulated devices; metal racking and moving plant reshape coverage between the survey and the shift; and a roaming handover that drops a barcode scan is annoying, whereas one that drops a reject signal before the diverter fires is a quality escape. We have written a question-by-question comparison of Wi-Fi 6 and private 5G for plant managers, and inspection cells sit firmly in the category where the licensed option earns its keep.
How does private 5G change the camera-per-line arithmetic?
It replaces a per-camera cable cost with a one-off building cost, and it makes redeployment free.
A private 5G network for a factory hall is a small number of indoor radios and a local core, operating under Ofcom's Shared Access licence framework. A Low Power licence in the n77 band, 3.8 to 4.2 GHz, costs £80 per 10 MHz channel per site per year, which is a rounding error against the hardware. The licensed power levels are what allow a handful of cells to fill a large shed where WiFi needs twenty access points fighting each other from the roof steel.
Once that coverage exists, an inspection cell becomes a trolley: camera, light, edge box, a 5G module, and power. Commissioning is a SIM and a config rather than a cable pull and a shutdown. The marginal cost of camera position forty-one is the hardware plus an afternoon. More usefully, the cell can be moved to the line that needs it this quarter, which is the flexibility that hard-wiring quietly took away.
Two further properties matter on a factory floor. The network can prioritise traffic classes, so a reject decision is not queued behind a firmware download, which is a guarantee unlicensed spectrum cannot make. And the data stays on site: with a local core and local breakout, inspection images never leave the building unless the plant decides they should, which removes an argument with IT security that has stalled more than one cloud-based vision project in a defence or pharma supply chain.
The honest trade-offs. A private network is a larger up-front commitment than adding four access points, and for a plant with six cameras in a benign radio environment, WiFi with compression at the camera may genuinely be enough. High-rate raw imaging still wants a cable, and always will. The case strengthens with the size of the building, the amount of metal, the number of cameras, and above all the number of other applications, AGVs, wash-down-proof handhelds, condition monitoring sensors, cleanroom tablets, that will share the same infrastructure. Vision inspection rarely justifies a network on its own. It is normally the second or third application on the list that tips the case over.
What does a realistic UK rollout look like?
Start from defect cost rather than coverage. Rank lines by the cost of the defects that escape them, including customer complaints, retailer penalties, scrap and rework, and let the top slice define phase one. Most plants find a small number of lines carry most of the pain.
Design the network once for the whole building, even if phase one lights up two lines. Resurveying and re-architecting per phase is where the budget doubles, and it is the mistake we see most often in staged deployments.
Specify cells as redeployable from the start. Trolley-mounted, battery or single-plug power, wireless backhaul, with mounting points designed in at each line position rather than a cell welded into place. The plant that can move its inspection capacity is the plant that keeps using it after the first product change.
Agree the false-reject budget before installation, in writing. An inspection system tuned to catch everything will also stop the line for things that do not matter, and a cell that operators have learned to override is worse than no cell at all. More technically successful installations get switched off for this reason than for any hardware fault.
Put a name against the network layer. Quality owned the pilot and the integrator owned the cell, but the network needs an owner in IT or engineering with a budget line and a service level, because every subsequent application will sit on top of it.
Where should a plant start?
With one number and one walk. The number is your annual cost of defects that reached a customer, including the complaints, the credits and the time your quality team spent on them. The walk is around the plant with your quality lead, marking the positions where you would put a camera if cabling were free. Those two artefacts are the business case, and in most plants the second list is considerably longer than anyone expected.
We build and manage private 5G networks for UK manufacturers who need the factory floor to go wireless without hiring a telecoms team to run it, sized and priced for mid-sized plants rather than multinational flagships. If your vision inspection programme is stuck on one line because the next forty cameras cost more to cable than to buy, talk to us.