X Vision · Technology
How a camera over the cell becomes a register in your PLC
An X Vision instrument watches a surface, measures it continuously, and writes the answer into a Modbus holding register. All of that happens on the camera, on your network. This is the long version of that sentence, written for the metallurgist working out whether it suits the circuit and the integrator who has to put it in a tag database.
Signal path
Four stages, and none of them leave site
-
01
The surface
The camera looks at the thing an operator currently judges by eye: the froth carrying over the lip of a flotation cell, or the foam sitting on a process tank. It is a hard subject. The surface moves, the light changes through the day, and no two cells look alike. Most of the engineering is in reading it reliably rather than reading it at all.
-
02
The camera
One housing on a bracket over the launder or the tank, with the sensor and the processor inside it. It bolts on wherever there is a clear view of the surface. Power and a network drop are the rest of it. There is no sampling rig and nothing that touches the process. X Float reads the froth as a three-dimensional surface, so it has height and depth rather than a flat picture.
-
03
Inference on the device
The measurement is computed continuously, in the camera. Nothing is uploaded and nothing waits on a server. The instrument runs on a network with no route to the internet.
-
04
The register
The result lands as Modbus TCP holding registers. Your control system reads them like any other process tag. There is nothing for the PLC to learn and no new protocol for IT to approve.
No cloud. No GPU rack.
Calibration
What it takes to calibrate a froth measurement to one cell
A bubble diameter in millimetres is only a number once the camera knows how many millimetres a pixel is worth at the surface it is looking at. That depends on the mount height, the angle, and where the froth sits in the frame. Move the camera 300 mm and the same froth reads differently. The geometry gets measured at the install rather than assumed from a datasheet.
Picking out the bubbles is the harder half. What counts as one bubble depends on how the froth is lit, how wet it is, and what the ore is doing that week. A model tuned on one cell’s froth will mis-read another cell’s, and it fails quietly: it still returns a bubble size, and the number is wrong.
So the model is trained on your cell, not shipped with the box. We take footage from the cell the camera is going to watch, label it against what your metallurgist calls a good froth and a bad one, and calibrate the states to your process. The process-state classifier does the same thing. “Boiling” and “stalled” mean something different on a rougher than on a cleaner, and the only definition worth writing into a register is the one your operators already use.
If the measurement you want has never been made before, that is the job we want. The request does not join a roadmap behind forty other plants. It gets built, or we tell you it can’t be.
The cost of that is time on your footage. A standard install commissions in hours, because the metrics already exist and only the geometry and the thresholds are yours. A measurement nobody has made before takes longer: it needs footage across enough conditions to be worth trusting, and you get an estimate of how much before the work starts.
On the device
A number you can argue with
Most of what the instrument reports is measured rather than predicted. The number comes out of the picture, so you can hold it up against the frame it came from and see for yourself whether it is right. When a metallurgist asks why a reading moved overnight, there is a frame to point at.
An operator who cannot sanity-check a reading stops trusting it by the second week, and a measurement nobody trusts changes nothing on the plant. So the instrument is built to be checkable first. How it is built is the part we sell, and that goes to the integrator in a specification rather than onto this page.
Grade is the case worth being plain about. Froth appearance correlates with grade at a given cell, and we will fit that correlation against your assays if you want it. It is a correlation calibrated to your site, not an assay, and the instrument does not report grade as a live measurement.
When it cannot see, because of a splash on the lens or a light that went out, it publishes no number at all. Your control system is told, and can stop acting on the reading until the camera can see again. Write your logic to expect that and a bad view costs you nothing.
Interface
What your control system has to speak
This is the part an integrator needs and the part we keep off the open web. The point list describes what the instrument measures closely enough that it goes to the person wiring it up rather than to everyone with a browser. Below is what your control system needs to support. The list itself arrives with the specification.
Two connection roles are possible, and the one you want depends on your control system. In server mode the camera listens and your control system polls it, the way a master polls any other slave. That suits most PLCs and SCADA packages, and nothing writes unsolicited into your controller. X Foam ships that way by default. In client mode the camera opens the connection and writes into holding registers on your control system, which then runs the Modbus server. X Float works this way today.
| Protocol | Modbus TCP |
|---|---|
| Register model | Holding registers |
| Roles | The camera writes into your control system, or serves registers for your master to poll |
| Reconnection | Automatic, with nothing for anyone to restart |
| Optional analog | 4–20 mA drive, for a site with no control system in the loop |
| Full specification | Point list, scaling and failsafe semantics are issued to your integrator under NDA |
The point list, with its scaling and what each point does when the camera cannot see, goes to your integrator as a specification document. Ask and we will send it under NDA, cut down to the readings your site is turning on.
The instrument measures and recommends. Actuation and interlocks stay in your PLC, and so does the safe fallback on loss of camera data. If the camera stops reporting, your control logic should fall back safely rather than keep acting on an old froth reading.
Commissioning
What we need from your site
A mount point with a clear view of the surface, power, a network drop, and, if the camera is writing into your control system, its IP address, port, unit ID, and confirmation that the register block is free. That is the whole list.
Mounting takes minutes. Commissioning takes hours, with one of us on site or connected in, aiming the camera and calibrating it against your process. There is no shutdown and no sampling rig to build. Nothing needs routine maintenance in normal operation, and when the cell you care about changes, the camera unbolts and goes on another one.
Questions
Common questions
- How does a froth camera measure bubble size?
- It picks out the individual bubbles in the image and converts their size to millimetres, using a scale set when the camera was mounted. That is why the mount geometry is measured at the install instead of assumed: a mount angle guessed wrong is a bubble size wrong every day after. What X Float derives from the bubbles, and how that is written to the registers, goes to your integrator with the interface specification.
- Does the instrument need cloud connectivity?
- No. Inference runs on the camera and the measurement goes straight to your control system. Commissioning takes a few hours with one of our engineers on site or connected in, and after that the instrument sits on a network with no route out. Nothing leaves site.
- Can my SCADA poll the camera directly?
- On X Foam, yes: the camera can run as a Modbus TCP server on your control LAN for your master to poll. X Float currently works the other way round, opening the connection itself and writing into holding registers on your control system. Tell us which your equipment needs and we will confirm what is supported before you commit to anything.
- What happens to the registers if the camera loses its view?
- The camera publishes nothing it cannot stand behind. When it loses its view, your control system knows to stop trusting the reading until the view comes back. The exact behaviour on loss of view is written into the interface specification your integrator receives.
- How long does calibration to my cell take?
- For metrics that already exist, hours. The camera geometry is measured at the mount and the thresholds are set against your process. A measurement that has not been built before needs footage from your cell across a range of conditions, and you get an estimate of how much before the work starts.
- Is it a black box?
- No. Most of what it reports is measured from the picture, so any number can be checked against the frame it came from, and a metallurgist asking why it moved gets a reason. Learning is used where measurement alone does not reach, and anything site-calibrated is fitted against your own assays. How it is put together goes to your integrator with the specification.
Send us a photo of the cell.
That is usually enough to say whether it will work and what it would measure. If it looks right, a trial runs on your own froth or foam before anything is committed.