AI control of fruit dehydration — the drying end point should not arrive hours late.
The tunnel is continuous; the fruit is not: every lot comes in with its own size and its own moisture. iLEAN connects the tunnel's existing probes, analyzes 100% of the outfeed with vision and proposes the setpoint adjustment from data. The person validates; the batch is documented with its drying curve.
The tunnel does not wait for the lab: by the time the moisture reading lands, the batch is already out.
The drying tunnel is a continuous process; the raw material is exactly the opposite. Every lot comes in with its own size, its own initial moisture and its own variety, and all of that shifts with the season. On top of that variability, the end point is decided today with the same old tools:
- The line lead's experience — he knows "how the fruit is coming in this year" and adjusts the setpoint by eye. It works almost every time. But that experience does not transfer to the night shift, or to the next season, or to the new operator.
- Lab moisture sampling — rigorous, but with hours of delay. By the time the result comes back, the tons that went through the tunnel in the meantime are already dry: if the reading says "past the point", the margin has already been burned.
- Over-drying is expensive — every extra point of moisture evaporated is saleable weight disappearing off the scale, and the color degrades: the fruit loses the look the customer is paying for.
- Falling short is worse — a water activity (aw) above range compromises shelf life: rework if you are lucky, and if not, an entire batch under review with the customer waiting.
- The drying curve is not kept anywhere — the temperature and humidity profile each batch lived through gets lost between chart recorders and spreadsheets. When something goes wrong, reconstructing what happened is archaeology.
The result is a setpoint that swings between the fear of falling short and the cost of overshooting. Most plants systematically over-dry "to be safe" — and that safety is paid for in kilos, in color and in energy, batch after batch.
iLEAN does not change your tunnel — it closes the loop between the fruit going in, the drying profile and what comes out.
The bottleneck is not measurement (the probes are already in the tunnel); it is that the probe signal lives in the chart recorder, the incoming lot lives on the delivery note and the outfeed quality lives in the lab, with hours of delay between the three. Nobody correlates anything in real time. iLEAN is the putty that joins the three — with no construction work on the tunnel and without changing the PLC or the plant system you already have.
Connect reads the existing probes zone by zone. Edge analyzes 100% of the outfeed with vision. The Agents correlate the drying profile with the incoming lot and propose the setpoint adjustment — the person validates. Every batch closes with its drying curve in the file.
The iLEAN pieces applied to fruit dehydration control:
- Connect — hooks into the temperature and humidity probes the tunnel already has, zone by zone, with no construction work. What today ends up in a chart recorder nobody looks at becomes a live drying curve per batch. And it also captures what does not arrive as a signal: the lot on the delivery note, the variety, the incoming moisture measured at receiving.
- Edge — machine vision at the tunnel outfeed, above the belt, on 100% of production: color, size and visible defects of every piece, not just the sample the lab happened to draw. Color is the first symptom of over-drying; catching it in line means arriving on time, catching it in the sample means arriving late.
- Agents — correlate the drying profile (temperatures and humidities by zone, residence time) with the incoming lot and with what Edge detects at the outfeed, and propose the setpoint adjustment with its rationale. The person validates or discards; without validation the tunnel is not touched. Every batch closes with its own file: complete drying curve, lot of origin, aw result.
The Brain that orchestrates the agents lives in Central; critical data stays in the plant's ring 1, not in some random cloud (see the IRIS architecture and the three safety rings).
End point by experience and sampling vs. end point by data with iLEAN
| Aspect | Experience + lab sampling | With iLEAN Connect + Edge + Agents |
|---|---|---|
| End-point decision | By the line lead's eye + lab moisture hours later | Setpoint proposal from real-time data — the person validates |
| Inspection at the outfeed | Occasional visual sampling | Vision on 100% of production: color, size, visible defects |
| Over-drying | Systematic, "to be safe": saleable weight lost and color degraded | Over-drying waste down ≥30% (estimate to be validated) |
| aw out of range | Discovered in the lab with the batch already closed | Anticipated during drying — less rework |
| Lot and season variability | Fixed setpoint tuned by intuition | Incoming lot ↔ drying profile correlation, learned batch by batch |
| Batch traceability | Chart recorders and scattered spreadsheets | Per-batch file with the complete drying curve and the aw result |
Impact estimate for your plant — to be validated with your numbers.
The block below is an estimate to be validated with the specific data of your tunnel and your season. It is there so the committee has an order of magnitude; we refine it during the diagnostic.
- Dehydration plant with one or more continuous tunnels or ovens, several varieties or SKUs per season, and an end point decided today by experience plus lab moisture sampling.
- Deployment of Connect on the existing probes + a vision Edge at the outfeed + correlation Agents. No construction work on the tunnel. First value expected within a few weeks on a pilot line and a reference variety.
- Over-drying waste reduction expected at ≥30% (often more, but that is the defensible floor — estimate to be validated). Less rework from out-of-range aw, by anticipating the end point in line instead of discovering it in the lab.
- Indicative payback between 4 and 9 months, depending on tons per season, the price of the finished product and your current rework and energy costs. Estimate to be validated with your data.
And production's reasonable doubt
“What if the AI gets the setpoint wrong?” — hallucination is a problem of free generation, not of anchored tasks. Proposing a setpoint adjustment from probe signals, residence time and outfeed vision is exactly an anchored task: the AI recontextualizes one piece of data (this batch's drying profile) using another (the history of comparable lots and their aw results). In this kind of task, the best models brought error below 1.5% [1]. And even so, the person validates every proposal — without a signature the tunnel is not touched. The three safety rings are there for exactly this.
[1] OpenAI paper “Why Language Models Hallucinate”, 2025 — on the reliability of AI in anchored tasks.
What people ask about AI control of fruit dehydration
What instrumentation does the drying tunnel need?
No construction work to start with. Connect hooks into the temperature and humidity probes the tunnel already has in each zone, whatever the PLC or the plant recorder reads. If a critical zone has no probe, a standard probe is added without stopping the line. The vision at the outfeed is a compact Edge unit above the belt — camera and controlled lighting — that does not touch the tunnel's mechanics.
How does it decide the end point without waiting for the lab?
The system correlates the drying profile (temperature and humidity by zone, residence time) with the incoming lot (size, initial moisture, variety) and with what the vision system detects at the outfeed (color, size, visible defects). From that history it proposes the setpoint adjustment before the batch goes past its point. Lab moisture sampling does not disappear: it stops being the only signal — with hours of delay — and becomes the verification that feeds the model back.
Does it replace water activity (aw) measurement?
No, it complements it. aw is still measured as the release criterion, because it is what guarantees the product's shelf life. What changes is the number of surprises: by anticipating the end point during drying, fewer batches reach the lab with aw out of range, and every batch is documented in its own file with the complete drying curve and the associated analytical result.
Does it work with variability between lots and seasons?
That variability is exactly why you would use it. Size, incoming moisture and variety change from lot to lot and from season to season, which is why a fixed setpoint sometimes falls short and sometimes overshoots. The Agents learn lot by lot: they link every input to its drying curve and its final result. Faced with an atypical lot that resembles nothing in the history, the system flags it and leaves the decision to the person, instead of extrapolating silently.
Who touches the tunnel setpoint, the AI or the person?
The person. The AI proposes the setpoint adjustment with its rationale — which signal triggers it and what effect is expected — and the line lead validates or discards it. Without validation nothing is changed on the tunnel. It is the same principle as the three safety rings of the IRIS architecture: plant data stays in ring 1 and any action on the process always goes through a human signature.
Tell us your case and in 48h we'll send you the estimated ROI of this AI project for your dehydration line.
We work on the real data of your tunnel and your season, not ours. Diagnostic with no commitment.
Request estimated ROI in 48h See food industry