Energy
Cutting Nuisance Alarms by 94% at Unmanned Sites
Substation Perimeter Classifier
94%
reduction in nuisance alarms across 60 sites
1.2 GB
monthly backhaul per site, down from ~40 GB
AgriTech / Post-Harvest Processing
Grain Protein Sorting Classifier
NIR Grading at 12 Tonnes an Hour on a Receival Belt
AgriTech
Industry
10 months
Duration
3 → 8
TRL
4
Disciplines
Background
A regional grain handler graded incoming wheat by pulling a spear sample per truck and running it through a benchtop NIR analyser. One number represented a thirty-tonne load, so trucks were paid and binned on an average that hid substantial in-load variation, and high-protein grain was routinely blended into a lower-value bin.
The problem
What made it hard.
Grading continuously on the belt meant taking a spectrum every few hundred milliseconds through moving product with varying depth, dust and moisture — conditions the benchtop calibration was never built for. The site had no climate-controlled space for instrumentation, and the diverter had to receive a decision fast enough to act while the affected grain was still upstream of the gate.
Approach
What we built.
We mounted a ruggedised NIR head over the receival belt behind a sapphire window with an air knife to keep it clear, and paired it with an i.MX 8M Plus running the chemometric model on its NPU. Rather than port the benchtop calibration directly, we rebuilt it against belt-condition spectra with moisture and bed-depth as explicit covariates, which removed most of the drift the original showed in trials. Protein estimates land every 250 ms and feed a short predictive buffer that fires the diverter with the correct lead time for belt speed. Every estimate is logged against truck ID so the handler can reconcile payment against measured distribution rather than a single spear sample.
Outcome
What it measured.
12 t/h
continuous grading throughput
250 ms
protein estimate interval on the belt
+4.10
AUD realised value per tonne
In-load grading resolved variation the spear sample never saw — 18% of loads contained a segment at least half a protein band above their assigned grade. Diverting those segments lifted realised value per tonne by 4.10 AUD across the first receival season. Grading disputes with growers fell sharply once distributions replaced single-point readings.
Our role
Optical mounting and air-knife design; chemometric recalibration; NPU inference pipeline; diverter control integration.
Technologies
NIR spectrometeri.MX 8M Plus NPUChemometrics / PLSC++EtherCATInfluxDB
Gallery
Inside the build.
Figure 1 — Protein distribution histogram for a single truck load, showing the high-protein segment the spear sample missed..
Figure 2 — Sensor head with sapphire window and air knife, mounted on the belt gantry..
Figure 3 — Diverter timing diagram.
Figure 4 — Receival office screen reconciling load payment against measured distribution..
Next case study
Energy
Tell us what you’re building.
Send the constraint that worries you most — a latency budget, a power budget, a certification date. We’ll tell you straight whether we’re the right team.