Case study: AI cameras at an agricultural plant

Return on investment

40Operator hours/week freedPreviously spent on manual monitoring of the valves.

4Hours of downtime avoided/weekFaster, more consistent auger corrections keep the line running.

98%Classification accuracyReached in production, matching the standard an operator would apply by eye.

Expansion now ready

The removal of the limiting factor on operation allows the whole plant to scale.

Background

A complex, 24/7 production process

Aumkar Plantations in Malaysia operate a 24/7 pyrolysis process, taking agricultural residues and converting them into carbon removals, biochar and green energy. The operation is large and requires multiple pieces of equipment to work in unison to optimise outputs. Previously, level sensors were relied upon to monitor throughputs, but reactions to issues were too slow, resulting in interruptions to production.

Two engineers beside the rotary valve and feed equipment at the agricultural plant

The challenge

Keeping every valve filled without a human watching it

Keeping each rotary valve correctly filled is critical to smooth, continuous operation. Too little feedstock underfills the valve, too much overfills it, and either state requires a fill-time correction on the augers. Historically, this has relied entirely on manual visual inspection: an operator watches a live camera feed above each valve, classifies what they see into one of six fill states every 90 seconds, and manually adjusts auger fill time in small ±2 second increments.

01

One person, two feeds, around the clock

The manual process ties up one operator continuously monitoring two camera feeds, on a line that never stops.

02

Inconsistent, fatigued judgement calls

Classifying fill state by eye, every 90 seconds, introduces the possibility of inconsistent or fatigued judgement calls.

03

A ceiling on how far production could scale

Relying on manual monitoring slows Aumkar's ability to scale the operation to more equipment and more sites.

Mainspring solution

A computer vision fill-state classifier

Reading fill state from the existing cameraMainspring built a computer vision classification model to automatically read the fill state of each rotary valve from the existing overhead camera feed, replacing the manual visual check.
The same six categories an operator usesThe model classifies each image into the same six categories an operator would use: Empty, Underfilled, Filled, Almost full, Overfilled, or Unreadable.
Straight into the PLCThat classification feeds directly into the PLC, triggering the ±2 second auger adjustment automatically.

See it in actionPick a screen, use the arrows, or tap the cards behind

01The camera the operators already watched02Fill state read automatically, every cycle03Every training image checked by a person04Tested class by class before it goes live

01 / 04

02 / 04

03 / 04

04 / 04

01

The camera the operators already watched

No new hardware went in. The overhead camera above each rotary valve, the same feed an operator used to keep an eye on, became the model’s only input. Every capture is timestamped at source, so each read can be traced back to the exact frame it came from.

Existing cameraNo new hardwareTimestamped frames

carboflow.onnu.com/rv-monitor/camera

The camera the operators already watched

The raw hopper feed: what an operator used to watch by eye, all shift.

02

Fill state read automatically, every cycle

The model outlines the material, detects the auger, and returns a fill level and a class with a confidence score on every read. It is the same judgement an operator made, now made on a fixed cycle and passed straight to the PLC.

Fill levelAuger detectedConfidence scored

carboflow.onnu.com/rv-monitor/live

Fill state read automatically, every cycle

Normal fill at 78%, called with 96% confidence, without anyone watching the screen.

03

Every training image checked by a person

Before anything trains the model, each image is annotated and reviewed: fuel level, auger, feed consistency, hopper structure and obstructions. Reviewers correct the labels, add notes, and only then send the image to training.

Human-reviewedFive classesSent to training

carboflow.onnu.com/rv-monitor/datasets/review

Every training image checked by a person

One review job of 48 hopper images, each labelled and checked before it is used.

04

Tested class by class before it goes live

Each model is tested on images it has not seen, with accuracy, precision, recall and F1 reported for every fill state rather than as one headline number. The weakest classes stay visible, so the team knows where more examples are needed.

94.3% accuracyPer-class F1Live test

carboflow.onnu.com/rv-monitor/models/test

Tested class by class before it goes live

Strongest on filled and empty; most errors sit between overfilled and underfilled.

The outcome

From manual checks to automated, real-time classification

Following full deployment across both rotary valves and integration with the PLC, the manual visual monitoring process has been replaced with automated, real-time fill-state classification driving auger adjustments directly.

Classification accuracy in production has reached 98%.

Operators are no longer required to continuously monitor the valves and sensors; their attention has been freed up for higher-value tasks.

In the first 12 weeks of operations, just 2 unplanned interventions were required.

Lighting variability across the 24/7 operating cycle has been addressed via supplementary LED lighting, with no measurable drop in accuracy across day/night conditions.

Work with us