Cold-Chain Integrity Recorder

Background

A pharmaceutical distributor shipping temperature-controlled product across Australia relied on single-use chart recorders. A recorder only revealed an excursion after the pallet arrived, by which point the consignment was already at the receiving dock and the product was written off rather than recovered.

The problem

What made it hard.

The recorder had to run fourteen days on a single primary cell while sampling often enough to catch short excursions, which meant the radio had to stay off almost all the time. Simple threshold alarms produced constant false positives - a door opening at a cross-dock looks identical to a failing compressor for the first several minutes. Distinguishing them needed pattern recognition, but there was no power budget for a general-purpose processor.

Approach

What we built.

We designed a sealed logger around an STM32U5 with a 6 µA/MHz low-power core and ran a small temporal convolutional model in TensorFlow Lite Micro directly on the MCU. The model reads a rolling window of temperature, humidity and 3-axis accelerometer data and classifies the thermal signature - door event, transient ambient, or genuine refrigeration failure - in 4 KB of RAM. Only a genuine-failure classification wakes the NB-IoT modem, so the radio duty cycle stays under 0.1%. Everything else is logged locally and offloaded over BLE when the pallet reaches a gateway. The result is a device that spends 99.4% of its life asleep and still catches a compressor fault within eleven minutes.