A portal-mounted LiDAR + camera pair distinguishes ferrous scrap from cooled slag in real time — per truck, per delivery, per heat. Feeds the EAF charge-mix planner, reduces sorting labour on the yard, cuts off-spec heats before they get charged. Custom-built for your scrap categories and your plant integration.
Scrap-slag classification AI is a portal-mounted perception pipeline that distinguishes ferrous scrap from cooled slag in real time on a steel-mill yard, using fused LiDAR and camera data to feed the EAF charge-mix planner.
The pain is well-known to every EAF operator: scrap deliveries arrive contaminated with cooled slag, refractory debris, oxidised fines and the occasional non-ferrous lump. When this mix gets graded the wrong way, the charge bucket carries the wrong mass of actual iron — the heat ends up under- or over-charged, the energy balance drifts, fluxes have to be adjusted on the fly, and in the worst case the tap goes off-spec. The cost shows up as wasted electricity per tonne, extra refractory wear, and slag pits that fill faster than the schedule predicted.
Our approach fuses two sensors at the yard entrance. A 3D-LiDAR scans the load geometry as the truck rolls under a portal — bulk density signature, surface texture, void structure. A high-resolution colour camera adds material colour, oxidation tint, and visible refractory or slag signatures. A custom classifier — trained on your scrap categories (shred, bushelling, plate-and-structural, heavy melt, turnings) and your slag types — returns a per-truck breakdown: tonnage estimate per category, slag fraction, contamination flags. The result writes directly into the charge-planner or MES over OPC UA, REST or MQTT.
This is not an off-the-shelf SaaS. Every steel mill has its own scrap sources, its own slag pit, its own EAF charge model. We treat every engagement as a discovery + fixed-scope build, descending from our hub service Industrial Perception AI. Typically four to twelve weeks from contract to live tool, with the code and the model handed over to you at the end.
Built by the same team that delivers our LiDAR wagon-detection pipeline — six lid centroids, sub-5 cm error, two-second response.
One LiDAR plus one calibrated colour camera mounted on a steel portal at the scrap-yard entrance or weighbridge. Trucks drive through at normal speed — the system captures load geometry and surface imagery on the fly, no stopping required. IP65+ housings, dust-tolerant optics, runs in fog, snow and night light.
The fusion classifier separates the load into your scrap categories — shred, bushelling, plate-and-structural, heavy melt, turnings — plus cooled slag, refractory debris and unknown contamination. Per-truck tonnage estimate per category, segmentation mask for visual audit. Models retrained on your labelled data, not on a public benchmark.
Classification results write into the charge-planner or MES over OPC UA, REST or MQTT — same shift, before the bucket is filled. The charge model sees real category masses instead of operator estimates. Off-spec deliveries are flagged before they enter the mix; aging slag-contaminated trucks get routed back to the supplier with the evidence attached.
We keep the architecture boring on purpose. Three loosely coupled stages, each one independently testable, each one swappable when the sensor stack changes. Built on the same stack we use across all our perception work: PCL, Open3D, OpenCV, PyTorch.
The LiDAR delivers a 3D point cloud of the full truck load — typically a million-plus points per pass at portal speed. The camera adds a synchronised colour image, lens-calibrated and time-stamped. Raw data writes to local storage with a per-truck UUID for audit trail and re-labelling.
The point cloud goes through ground removal, voxel down-sampling and a PointNet-style classifier trained on your scrap categories. The colour image goes through a CNN trained on the same labels. The two heads are fused into a single per-voxel prediction — typically 85–95% classification accuracy on production data, depending on how clean your scrap categories actually are.
Per-truck output: estimated tonnage per category, slag fraction, contamination flags, audit images. Writes into the EAF charge planner over OPC UA — or into the MES, the ERP, the yard-management system, whatever the plant already runs. The charge model now reads real category masses instead of operator estimates.
All three stages run on an industrial PC at the portal cabinet. No cloud dependency, no external API, no licence dial-home. The code is yours at handover.
Every delivery gets a structured record: scrap-category mass breakdown, slag fraction, contamination flags, audit images, supplier ID, weighbridge ticket cross-link. Drops into the supplier-quality dashboard, into ERP, and into the yard-acceptance workflow.
Across many trucks, the yard accumulates an aging map: which bay holds which scrap mix, how long it has sat, how much slag contamination is sitting in each pile. Same temporal model we use on stockpile monitoring — applied to scrap categories.
For every planned heat, the system proposes a charge mix that hits the target chemistry within typical fines and slag tolerances — drawing from real yard inventory instead of assumed averages. The metallurgist still approves; the planner does the bookkeeping.
Every steel mill has its own scrap sources, its own EAF charge model and its own existing yard sensors. A generic classifier solves the generic case; your case is not the generic case.
A shred pile in one mill looks different from a shred pile in another — different shredder, different feedstock, different oxidation history. We calibrate the model on your material under your lighting and dust conditions, and we re-calibrate when the supplier mix changes.
We bring the labelling tooling and the engineering hours. You bring the operator who actually knows what's in each truck. The result is a labelled dataset that belongs to you and that you can keep using as the model evolves — no vendor lock-in on the training data.
Most steel-mill yards already have weighbridges, RFID gates, ANPR cameras and overhead cranes. We fuse the classification output into what's already there — we do not replace your yard-management system. OPC UA, REST, MQTT and analog outputs all supported.
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-truck 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 steel-mill customer considering a custom perception build. Need something more specific to your yard? Ask us.
Sensor setup, sample point clouds, your scrap categories, your charge-planner 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.