Machine Learning on IoT devices

The combination of IoT and Machine Learning creates opportunities to open new business areas, increase efficiency or save costs. In many cases it is necessary to run the machine learning model directly on the microcontroller unit of the IoT device instead of in the cloud.

Many IoT solutions are based on low-end sensors, which generate floods of data. The transmission of large amounts of data from an IoT device to a central system (in the cloud) requires a lot of power and increased costs. Therefore, to add value to IoT projects with Machine Learning, one needs to overcome certain challenges:

TinyML (also known as EdgeML) is a discipline, that combines Machine Learning and Embedded Systems. It allows for Machine Learning models running directly on low-power and low-cost microcontrollers of IoT devices and enables data analytics at extremely low power and low latency. This brings the advantage that less data needs to be transmitted. Instead of sending raw data, only the results of predictions need to be sent. This way, data analytics can be performed directly on the IoT device with very low latency and low power requirements.

One example: Damage detection on IoT devices

By processing and analyzing data directly on low power IoT devices, TinyML offers many new solutions. The goal in this example is to detect damages to assets like containers or freight wagons during loading operations. 

The following figure outlines the process: based on collected historical data, a Machine Learning model is trained in the cloud. Once a suitable model is found, it needs to be scaled down so it can be deployed locally on the IoT device. The deployment is done over-the-air via the IoT Platform. Only in the event the model detects actual damage, is an alarm sent to inform the respective business units.


TinyML - when you should consider it

  • Limited power for data transmission (e.g., in case of energy autonomous devices): Microcontrollers consume very little power. If the data processing is done directly on the IoT device, it enables them to run for a long time without being charged.
  • Desire for lower data transmission costs: Instead of constantly sending raw data, only relevant information needs to be sent to a central system.
  • Need for reduced dependency on mobile network coverage (e.g., for moving assets): Less data needs to be sent as results are generated directly on the IoT device of an asset, therefore less bandwidth is used.
  • Strict data privacy requirements: The data does not need to be stored in the cloud, as the models run directly on the IoT device.  
  • Low latency is important: As predictions are made via on-device inderence, data does not have to be transmitted to the cloud.



Interested in Smart Business Applications but still have some pending questions? Feel free to contact us any time. Our experts will come back to you as soon as possible.

Stefanie Pichler

Product Manager for Machine Learning & Advanced Data Analytics