How AI helps keep food and medicine safe

Every time someone says “we use AI for temperature monitoring”, it’s worth asking exactly for what. The honest answer today is that artificial intelligence solves three specific cold chain problems very well — and doesn’t yet solve three others. This article separates them.

What already works

1. Anomaly detection in the thermal curve

A fixed temperature threshold is a blind instrument. It fires once the product is already at risk and stays silent while equipment behaviour degrades inside the range.

Anomaly detection models learn the normal shape of each asset’s curve: how many compressor cycles per hour, the amplitude of each cycle, how long defrost takes, how temperature recovers after a door opening. When the shape changes — cycles get longer, recovery gets slower — the model flags it, even with temperature still sitting at 4 °C.

In practice, that is how you discover an evaporator icing up two weeks before the cold room floods.

2. Compressor failure forecasting

This is the application with the most direct return. The combination of vibration signature, inrush current and thermal behaviour contains a fairly consistent wear signal.

Compressors don’t fail out of nowhere. They get louder, draw more current at start-up, take longer to reach setpoint and run hotter. Each of those signals alone is noisy; together they form a pattern a model can recognise weeks ahead.

The value isn’t predicting the exact day of failure. It’s turning an unplanned stop — with lost load, emergency transport and a technician on call-out rates — into maintenance scheduled for Tuesday morning.

3. Translating data into operational language

This is the most underrated use. A store manager doesn’t want to look at eighteen time series. They want to know what happened last night and what to do today.

Generative models are very good at that layer: read the series, cross-reference logged events and write a summary in plain language. “Cold room 02 spent 38 minutes above range between 4am and 5am. The door was open for 22 of those minutes. The compressor compensated, but consumption rose 18% over the period. Recommendation: check the door closure and inspect the gasket.”

That isn’t sophisticated analysis — it’s translation. And it’s the difference between a dashboard people consult and one they ignore.

What doesn’t work yet

Predicting shelf life. We have seen promises of “dynamic shelf life” computed from thermal history. The physics exists and is well studied, but it depends on parameters specific to each product, packaging and batch that most operations simply don’t have. Without those, the number comes out pretty and wrong.

Deciding on its own. No model should shut down a compressor, mute an alarm or discard a batch without a person in the loop. Not out of conservatism, but because the cost of an autonomous false positive is high and decision traceability matters in an audit.

Working without good data. A predictive model trained on 30-minute readings, with connectivity gaps and badly placed sensors, learns noise. The expensive, tedious part of AI in cold chain isn’t the model — it’s the instrumentation.

The pharmaceutical case

In pharma the bar is different. The 2 to 8 °C range admits no exception, records require traceable calibration and a thermal excursion demands a formal report.

Here AI contributes less to prediction and more to consistency: identifying sensors drifting out of calibration, detecting excursion patterns tied to a shift or a specific procedure, and shortening the time between an excursion and a technical assessment. The gain is compliance and response speed, not optimisation.

How to start without fooling yourself

The order we recommend is always the same, and it is decidedly unglamorous:

1. Instrument properly. The right sensor in the right place, a short enough interval, connectivity with a local buffer so the series has no holes.
2. Log events, not just readings. Door, defrost, maintenance, load replenishment. Without context, the model confuses normal operation with anomaly.
3. Run rule-based alarms first. Fixed thresholds and time-out-of-range solve most of the losses. Only then does prediction earn its keep.
4. Then predict. With six months of clean data and annotated events, a wear model starts to be worth what it costs.

AI in cold chain is less magic and more data discipline. The good news is that data discipline alone already pays the bill.

---

Curious how this looks in practice? The interactive demo shows a compressor degrading and the maintenance window the model suggests.