Put the Camera on the Crane If You Want Closed-Loop Combustion Control

"If the camera can see what's on the grate, why is an operator still setting the air?" I get a version of that question on most closed loop combustion control projects, usually within ten minutes of someone watching a flame-image demo. The straight answer is that the camera watching the fire and the dampers feeding it can't share a loop that tight. By the time a model can see a channeling bed, the fuel that caused it went through the feed ram a minute or two ago, and the air correction was due before it arrived.

Which is why the loop worth building runs the other way round. Classify the fuel on the way in, and trim on the fire only to catch what you got wrong.

A crane camera sees the fuel before the oxygen probe does

Observation points on a mass-burn line sit in a fixed order: tipping floor, bunker, crane grab, feed hopper, ram, drying zone, burning zone, then the flue-gas duct where the oxygen and CO analyzers live. Every step adds dead time. Grab to burning zone runs a few minutes on a reverse-acting Martin grate, more when the hopper is holding a deep column [engineering estimate]. Those analyzers sit at the far end of the chain, and they're also the only signals a conventional air-trim controller works from.

That ordering is the whole opportunity. A computer vision feedstock classification model over the crane has those minutes as warning: this grab is mattress foam over wet kitchen organics, the next two are baled commercial cardboard. Primary air split, grate speed per zone and feed stroke length can all move before the oxygen dips rather than after. It's the part of the Optimal Waste Intelligence stack I've spent the most hours on, and it has the same shape as predictive-maintenance work, where the whole trick is reading a bearing's vibration spectrum weeks before the bearing admits to anything.

Precision is easy; recall is where the model lies to you, and on a crane that asymmetry works in your favor. Getting the class exactly right on every grab isn't the job. Keeping the running estimate of moisture and heating value unbiased across a five-grab window is, and missing one plastic bale in ten costs you almost nothing if the misses scatter. A classifier that's right four times in five with symmetric errors beats one that's right nine times in ten and reads baled film as paper every single time, because the second one walks your air ratio steadily in one direction and the trim loop absorbs the error until it can't.

You also need far less camera than the vendors quote. One good frame per grab, consistently lit, beats thirty frames a second that nobody has the bandwidth to move off the crane gantry.

Flame images tell you where you already are

The flame-image literature deserves better than its reputation among operators. Zhang, Sun, Tang and Pei reported 96.16 percent accuracy on combustion-state recognition from incinerator flame images in Sustainability in April 2025, using a Res-Transformer with attention modules grafted onto a ResNet backbone. Figures in that band are now routine in the municipal incineration control literature, and several groups working on Chinese grate furnaces have published above 95 percent on their own plant's image sets.

But accuracy on held-out flame images isn't control value. A four-state output - normal, partial, channeling, smothering - is a diagnosis, and the experienced operator on shift usually reaches the same diagnosis from the same video feed. What the model adds is consistency at four in the morning, and a labeled signal stream you can regress against steam flow and CO. Worth having. Not a loop, though, and anyone demonstrating it as adaptive combustion control is counting on nobody in the room asking about dead time.

Feedstock vision is the harder of the two problems, and the public benchmarks say so plainly. According to the ZeroWaste dataset its authors published in 2022, the detectors they trained on 4,661 annotated frames from a single-stream recycling line in Massachusetts missed far more than they caught. The strongest fully supervised segmentation model on that footage reached 52.5 mIoU and Mask R-CNN managed 22.8 AP, at a time when street-scene segmentation had long since passed eighty. Waste is deformable, occluded, overlapping and lit badly, and a bunker grab under a crane is a worse scene than a conveyor belt.

So treat a published flame-state number as a lower bound on what your own plant will cost you in labeling, not an upper bound on what the technique can do.

Between the grab and the hopper, your signal gets mixed

Most of the engineering budget on these projects goes somewhere unglamorous: reconciling what the crane saw with what the furnace is burning. The hopper holds several grabs' worth of material and mixes them vertically, badly and unevenly. A per-grab classifier output has to be convolved with a residence model of the hopper and the ram before it becomes a usable feedforward signal, and no residence model I've worked with has been good. We fit the delay empirically instead (grab timestamps, hopper level, ram stroke count, and the lag between a known-odd grab and the oxygen response), rather than deriving it from hopper geometry, which flatters itself.

Tune the inputs before you touch the architecture, every time. Swapping a ResNet for a transformer moves classification accuracy a couple of points. Getting the time alignment right between the vision stream and the process historian moves the control result far more, and a waste intelligence software layer that only reports classes, without timestamping them against the DCS clock, can't support a closed loop at all.

One qualifier: that holds for mass burn with a deep bunker. On a bubbling fluidized bed fed from a prepared RDF line, the mixing is better, the dead time is shorter and the feedforward argument weakens considerably, because the bed itself is doing the averaging you were trying to model.

The averaging window sets your controller's deadline

Every objective function written for closed loop combustion control has to live inside the operating-practice standards. For a mass burn waterwall unit, per 40 CFR 60.53b, carbon monoxide is capped at 100 ppmv corrected to 7 percent oxygen, measured as a four-hour block average. A refuse-derived fuel stoker gets a looser 150 ppmv, but on a twenty-four-hour average. The same section caps unit load at ten percent above the maximum load the unit demonstrated, and holds the particulate control device inlet to seventeen degrees C above its demonstrated maximum.

Those two averaging windows ask for different controllers, and that distinction gets lost in nearly every pitch I've read. A four-hour block will absorb a ninety-second CO spike without blinking; what it won't absorb is a controller that trades a small, steady CO rise for better steam flow. On a twenty-four-hour average, the mean is the only thing that counts, and the transient behavior everyone likes to demo barely registers. A reward function built to minimize excursion counts, running on a unit whose binding limit is a mean, does exactly what you asked and puts your permit at risk anyway.

Four things I want settled before anyone writes that objective function:

1. Which averaging window the binding limit uses, and whether the objective therefore penalizes peaks or means.
2. Which actuators the model may write to, each with an explicit rate limit and a bounded range.
3. What the loop does when the camera goes blind - fogged lens, crane parked, bunker lighting failed, classifier confidence collapsed - and how many seconds it takes to notice.
4. Who owns the fallback controller, and whether every handoff in either direction is logged with a reason code an auditor can read.

The fourth one gets waved through most often and bites hardest. A closed loop you can't audit is a liability, not an asset, because the first time a regulator asks why the unit ran rich for forty minutes on a Tuesday, "the model decided to" is not an answer anybody accepts.

The interlock layer isn't yours to tune

NFPA 85 covers stoker-fired boilers burning municipal waste and refuse-derived fuel from 12.5 million Btu per hour upward, and the burner management system it describes is a safety function sitting apart from the process control you're optimizing. Your model writes setpoints into the DCS (Emerson, Honeywell, Siemens, pick your flavor) inside a bounded, rate-limited envelope. It takes no part in flame supervision, purge logic or any trip. Anyone pitching real-time reactor control that reaches into that layer has either not read the code or is hoping you haven't.

And now the unpopular part. Most of what gets sold as closed loop combustion control is a rule engine with a CNN bolted on the front, and the rules are doing the real work. That isn't a scandal; a well-tuned rule base driven by a decent feedforward signal is a better plant than what most operators run today. But it changes what you should pay for. Ask to see the rule base. Ask what the system does when model confidence drops under threshold and the rules take over, because on a bad feedstock week that's the mode you'll be living in.

Pyrolysis takes away the variable you were hiding behind

On a grate, excess air is a forgiving knob. Get the fuel estimate wrong and the trim loop buys you time at a small efficiency cost. A pyrolysis reactor gives you no such margin, because there's no excess air to trim in the first place. The handles are feed rate, wall or bed temperature, residence time and recycle-gas routing, and the reactor's thermal time constant is long enough that temperature makes a poor manipulated variable for anything fast.

So AI process control for pyrolysis comes out on the feed side almost by default: screw speed, charge mass, moisture-driven preheat. That makes the feedstock classifier load-bearing rather than decorative. It's the sensor the loop is built on, and the one place in these thermal conversion systems where a vision model genuinely closes something. How much of this is even available to you depends on the reactor: whether your feedstock lets you choose a reactor type at all determines whether you have a continuous feed to modulate or a batch charge you can only weigh.

Chlorides are the other reason to classify hard on the way in. A vision model that flags PVC-heavy loads before they're charged protects product slate and reactor metallurgy in a way no downstream analyzer can, since by the time the gas chromatograph tells you, the chloride is already in the oil.

Where closing the loop isn't worth the integration

All of this assumes a bunker. A small modular starved-air unit with direct ram feed off the tipping floor, held to fifty ppmv CO on a four-hour window, has no observation point far enough upstream to hand a feedforward model anything useful; by the time the waste is in frame, it's committed. The same goes for a single-customer industrial feed, where a plant burning one shredder's production scrap has little variance left to predict and the vision stack buys nothing but a dashboard. Plenty of plants are also locked out by their own controls - an older DCS with no write path exposed, or a vendor combustion package whose performance warranty lapses the moment an external system touches a setpoint - and that commercial obstacle has stopped more of these projects than model accuracy ever has.

Disclosure: I build the control-side models on the OWI platform, so I'm hardly a neutral party on the question of where a vision model belongs.

Sources & Notes

• The CO limits and averaging times, the load cap and the particulate-device temperature margin all sit in 40 CFR 60.53b, the operating-practice standards for municipal waste combustors. Read the table rather than a summary of it: the averaging window moves with combustor type, and it changes controller design.
• Flame-state recognition accuracy comes from Zhang, Sun, Tang and Pei, "M3RTNet: Combustion State Recognition Model of MSWI Process", Sustainability 17(8), April 2025.
• For segmentation and detection baselines on real waste footage, Bashkirova and colleagues' ZeroWaste dataset (CVPR 2022) is public, and it's the benchmark to hold a vendor's demo reel against.
• NFPA 85, the Boiler and Combustion Systems Hazards Code, is where the burner-management boundary gets drawn; Chapter 10 is the part covering stokers that fire municipal waste. Check which edition your jurisdiction has adopted, because it often isn't the current one.