Low temperature scientific freezers are a critical piece of equipment in academic research laboratories and at commercial biotechnology and biopharmaceutical companies. Such freezers are used to house biological samples and expensive reagents — these contents can be the product of hundreds of hours of lab work and many thousands, or even millions, of dollars worth of supplies. Given the precious contents, it can be a major problem when a freezer fails, which all freezers will eventually do if given enough time in service.
Machine learning enthusiasts Adi Azulay and Jen Fox gave this problem some thought, and devised a solution that they think can help labs to get ahead of any freezer problems before they actually occur. They used a low-power microcontroller and Edge Impulse to monitor scientific freezers and predict when a maintenance problem may be in the early stages. By sending an alert to lab members, the system allows action to be taken while the freezer’s contents are still at a nice, cool 60 degrees below zero Celsius.
The hardware of this prototype device includes an Adafruit HUZZAH32 development board, sporting a 240 MHz dual-core Tensilica LX6 microcontroller and an onboard WiFi radio. An Adafruit MCP9600 I2C Thermocouple Amplifier was added to record temperature information as low as 200 degrees below zero Celsius. These components were plugged into a breadboard, and wired together for communication over I2C.
The developers decided to use a machine learning model to predict maintenance issues, so the next order of business was to collect real world data from a freezer. This requires data to be collected over a longer period of time, so Azure IoT Hub and Azure Storage were leveraged to collect that longitudinal data. After setting up the environment in Azure, firmware was designed and loaded onto the HUZZAH32, which was installed in a freezer, and then left to periodically send temperature information to the cloud.
After sufficient data had been recorded, it was downloaded, then transferred over to Edge Impulse via the data acquisition tool. An anomaly detection impulse was designed and trained that is capable of analyzing temperature observations and determining if they are outside of what is considered to be normal. Once the model had been validated against a dataset that had not been used during training, it was exported as an Arduino library, which was then deployed directly to the HUZZAH32 board via the Arduino IDE.
Should the model recognize an anomaly, the board will send a message to Azure IoT Hub, where that information can be seen by a user. This data can also be integrated with a service such as IFTTT to send an email or text message for more immediate notification of a potential issue. Between the simple, point-and-click interface of Edge Impulse and the sub-50 dollar price point of the hardware, this device has the potential to be a real life-saver in laboratories everywhere — finding a potential problem before it becomes a real problem is always the best course of action.
Take a look at Azulay and Fox’s full project write-up for the source code and additional details.
Want to see Edge Impulse in action? Schedule a demo today.