Container glass hot-end cameras: visibility isn't the win

A lot of plant managers walk into a capex meeting believing a hot-end camera system is the fix for their reject rate. Install the hardware, watch checks and blisters drop, book the win. I've sat through enough of those meetings to tell you it doesn't play out that way. The camera is the easy part of the project. The other part, the part nobody puts in the proposal, is what actually moves the number.

A camera doesn't fix a feeder problem

Hot-end vision systems, whether it's an Xpar Vision gob temperature scanner, an Iris Inspection Machines array, a Tiama camera set or Heye International's HeyeVision bolted onto a newer machine, do exactly what they're built to do. They watch the ware coming off the blank side and flag a defect fast. What they don't do is fix the feeder tuning, the mould cooling or the swabbing cycle that actually caused it. The named defect modes point somewhere specific upstream:

• Checks, cold and hot, when the gob homogeneity band drifts past ±5-8°C in the feeder bowl
• Choke marks and thin-wall, when cavity-to-cavity weight variation creeps past roughly 2%
• Bird-swing and spikes, when plunger or blank timing drifts, usually right after a job change

Gob temperature homogeneity across the feeder bowl is commonly held within that ±5-8°C band. Push past it and you get checks and overpress before the camera even earns its keep. A vision system tells you the defect happened. It doesn't retune the feeder. On most lines I've audited, the person who needs that information finds out after the defective ware is already in the lehr at 550-580°C, wasting an anneal cycle it never should have entered.

Where the reject rate actually moves

Typical hot-end reject rate before cold-end inspection runs somewhere around 1-3% of gross production. Vendors sell hot-end cameras on shaving that down. In my experience most of the real gain doesn't come from raw catch rate at all. It comes from cutting the time it takes to diagnose why the line drifted in the first place, which is a data problem before it's a hardware problem.

But that only works if the camera data actually talks to something else. If it isn't tied into furnace batch records, cullet ratio and cold-end inspection, root-cause work never happens. Nobody sits down and links a run of checks back to a swabbing-robot cycle that drifted out of sync with the capture window (and the OEM rep who sold the array is long gone by the time someone asks why). The system just logs alarms nobody reads.

A camera that flags a defect nobody's trained to act on isn't quality control. It's an expensive alarm log nobody reads.

The first 20 minutes after a job change is where the money is

Job changes, mould swaps in plain language, are a disproportionate source of defects on every hot end I've worked. Camera value concentrates in the first 20 to 30 minutes after a job change, and that's exactly the window most ROI models built around steady-state production don't count properly. Steady state is not where the pain lives.

In 2016 I audited a five-line plant in the GCC running a mix of legacy Emhart 8-section machines, some still on relay logic from the 1980s, alongside one newer line carrying a proper hot-end camera array. Manual gob and cull inspection on paper checksheet on four of the five lines. The fifth line, with the camera, still ran the same job-change chaos as the other four, because nobody had told the gaffer he was authorised to react to an alert inside the 8 to 12 second machine cycle without calling the shift supervisor first. Twenty-three minutes. That was the average time from mould-change complete to first acceptable ware on that line. The camera didn't move that number by itself.

Train the gaffer before you train the algorithm

Cavity-to-cavity weight variation on multi-cavity IS machines needs to stay under roughly 2% to avoid choke marks and thin-wall failures. That's a feeder-tuning problem, and it has to be corrected inside the machine's cycle time, not raised at the next shift meeting. A camera alert is worthless if the hot-end operator or IS mechanic isn't trained and authorised to act on it in real time.

Not a camera problem. A training problem. Plants that skip retraining on alert response see reject rates unchanged 6 to 12 months after the install. And the 0600 handover, in my experience, misses the night-shift swabbing data on most lines I see, which is exactly the data a root-cause review needs to link a check run back to its actual cause.

This is where a systemised job change process earns its keep. Our founder, Zaid Hassoneh, ran the hot end at O-I Brisbane before cameras were standard kit, and built the Job Change Tool around the 9-stage Job Change Lifecycle for that reason, not off a vendor deck. Cross-shift variance on identical SKUs, plan through post-mortem, typically runs 30-60% in plants that haven't locked the recipe, the mould set and the named owner for each stage. Get the discipline right around it and job-change time can fall by real margins, not the flat line you get from hardware alone. A hot-end camera sitting on top of that variance just measures it faster. It doesn't remove it. That's the argument for going in vendor-neutral on the capex decision, because every OEM rep in the room has a camera to sell, not a forehearth to fix.

Europe and the Gulf are pulling the same lever from different directions

Europe adds its own pressure. Under EU ETS Phase IV the linear reduction factor rises to 4.3% a year from 2024, tightening free allocation for energy-intensive plants including container glass. FEVE puts EU container glass recycling at roughly 80%, the highest of any packaging material in the bloc, and that pushes furnaces toward higher cullet ratios and more batch-to-batch gob variability for the hot end to absorb. A camera array can see that variability arrive. Fixing it still comes down to forehearth thermal control and feeder response, not the lens.

In the Gulf it runs the other way. Vision 2030 demand and the shift back to glass after Dubai's single-use plastics ban are pushing new melting capacity, but most GCC IS-machine lines still run without integrated hot-end vision, and defect data is still a paper checksheet on the floor. Regional cullet supply is thin too, so furnaces run closer to virgin-batch chemistry, which shifts gob viscosity and colour risk onto hot-end control rather than composition. Camera capex there needs benchmarking against forehearth homogenisation and mould-cooling redesign, not bought as a standalone line item because a competitor plant has one.

None of this means skip the camera. It means don't buy it as an island. Before the capex request goes in, get a proper forming audit done on the line it's going onto, so you know whether the bigger yield sits in vision, in mould cooling or in the swabbing cycle. Build the data path into your digitalisation and reporting layer from day one too, so the alerts land somewhere a root-cause review can actually use, not a screen nobody's watching on night shift. A vendor-neutral container glass consultant will tell you that before they tell you which camera to buy.