A LiDAR + camera pair on the belt continuously classifies and measures the bulk material crossing the conveyor — ore grade, contamination, blend ratio. Feeds the blending model, prevents off-spec batches, gives the control room a real material balance instead of an inferred one. Custom-built on top of OWL EYE® Volume Flow.
Material flow AI is a perception pipeline that runs on top of a conveyor-belt scanner (LiDAR + camera) to continuously classify and analyse the material crossing the belt — ore grade, contamination type, blend ratio — and feeds the result into your blending model, your control room, your material balance.
The pain is simple. A belt scale gives you mass per unit time. A volume scanner — like our OWL EYE® Volume Flow — gives you volume and density. Neither tells you what is on the belt. And in cement, sugar, fertiliser, recycling, ore processing and grain handling, that is exactly the question that drives blending, off-spec prevention and contract billing. The control room ends up inferring composition from feeder set-points and lab samples that arrive hours after the shift is over.
Our approach starts at the belt cross-section. A 2D LiDAR profiles the geometry of the material flow, a colour camera captures surface signature, and reflectivity at the belt adds a third channel. A per-section classification model — trained on your material categories (ore grades, contamination types, blend components, off-spec patterns) — labels every scan. Results aggregate into per-minute, per-shift and per-day flow reports, and the same stream drives real-time alarms when contamination spikes or the blend drifts off target.
This is not an off-the-shelf SaaS. Every plant has its own feed sources, its own blend recipe and its own control-room stack. We treat every engagement as a discovery + fixed-scope build, descending from our hub service Industrial Perception AI. Often layered on top of an existing OWL EYE® Volume Flow or Stockpile installation — same sensor, additional software layer.
Built on the same perception stack that powers OWL EYE® Volume Flow and Stockpile in production at industrial sites today.
A 2D LiDAR for geometry plus a colour camera for material signature, mounted over the belt or at a transfer point. Often the same sensor pair already installed for OWL EYE® Volume Flow — one mechanical install, two pipelines. IP65+ housings, dust-tolerant optics, runs through the full shift in mill, mine and recycling-plant conditions.
The classifier model labels every belt cross-section against your material categories — ore grades, contamination types, blend components, off-spec patterns. Models retrained on your labelled data, not on a public benchmark. Aggregated to per-second, per-minute, per-shift and per-day flow records ready for the control room and the historian.
Classification results write into the blending PLC, the historian and the ERP material balance — OPC UA, MQTT, REST. The blender sees real composition instead of feeder set-points. Contamination spikes and off-spec mixes trigger control-room alarms before the bad batch reaches the silo or the customer.
We keep the architecture boring on purpose. Three loosely coupled stages, each one independently testable, each one swappable when the sensor stack or the recipe changes. Built on the same stack we use across all our perception work: PCL, Open3D, OpenCV, PyTorch.
The 2D LiDAR profiles the belt at 10–50 Hz, producing overlapping cross-sections of the material flow. The colour camera adds a synchronised, lens-calibrated frame of the same patch of belt. Belt-speed input from the existing encoder ties geometry to mass and aligns frames into a continuous longitudinal record. Raw data writes to local storage with a per-section UUID for audit trail and re-labelling.
The point cross-section goes through ground removal, voxel down-sampling and a PointNet-style classifier trained on your material categories. The colour frame goes through a CNN trained on the same labels. Outputs fuse into one per-section prediction — typically 85–95 % classification accuracy on production data, depending on how clean your material categories actually are — then aggregate to per-second, per-minute, per-shift and per-day rollups.
Per-second output: class label and confidence, mass per category, contamination flags. Writes into the blending PLC, the DCS, the historian and the ERP material balance over OPC UA, MQTT or REST — whatever the plant already runs. The blending model now reads real composition instead of inferring it from feeder set-points and delayed lab samples.
All three stages run on an industrial PC at the belt cabinet. No cloud dependency, no external API, no licence dial-home. The code is yours at handover.
Every belt scan gets a structured record: class label, confidence, estimated mass per category, contamination flags, audit image. Published continuously over OPC UA or MQTT, archived to the historian, and ready for cross-link with belt-scale and feeder records.
Material composition rolled up per minute, per shift and per day — comparable across feeders, belts and product lines. Same temporal model we use on stockpile monitoring — applied to belt flow. Drops straight into the production-quality dashboard and the contract-billing record.
When contamination spikes or the blend drifts off target, the system raises a control-room alarm before the bad batch reaches the silo, the rail wagon or the customer. Routed into the existing DCS alarm list — no new operator screen to learn, no parallel dashboard to ignore.
Every plant has its own feed sources, its own blend recipe and its own control-room stack. A generic classifier solves the generic case; your case is not the generic case.
Your blend, your contamination definitions, your off-spec patterns. We calibrate the classifier on your material under your lighting and dust conditions, and we re-calibrate when the feed source or recipe changes. A cement raw-mix classifier and a copper-ore-grade classifier share the same architecture and nothing else.
If you already run OWL EYE® Volume Flow on the belt, the AI layer reuses the same sensor — one mechanical install, two pipelines. Volume and mass continue to flow into the belt-scale and inventory record; classification flows into the blender and the contamination alarm. No second portal, no second cabinet.
We integrate with your blending PLC, your historian, your ERP material balance — OPC UA, MQTT, REST, analog outputs all supported. No generic "let's send you a CSV every hour". The blender reads classification at control-loop rates; the ERP reads aggregated composition per shift; the historian keeps the per-section record for audit.
You own the source code, the model weights and the labelled dataset at handover. We document the system, train your team, and walk away clean. No black box, no monthly per-tonne licence, no service contract you can't exit. See our FAQs on IP and engagement model for the standard terms.
The engagement-model questions we hear from every plant considering a custom perception build on top of OWL EYE® Volume Flow. Need something more specific to your belt? Ask us.
Sensor setup, sample point clouds, your material categories, your blender interface — anything you have. We come back within two business days with an honest first assessment and a fixed-price scope for the Discovery workshop.