StorAIge Insights
Research Project StorAIge
Pushing the Limits of Edge AI
Artificial intelligence is increasingly moving from the cloud to the ‘edge’ – directly into the devices and sensors that operate in the field. While this shift opens up enormous possibilities for industries, it also presents a significant challenge: how can advanced AI be run on microcontrollers with only a few kilobytes of memory?
The EU-funded StorAIge project was established to address this very issue, uniting leading European partners in the development of memory-efficient Edge AI solutions for applications spanning industrial automation to renewable energy. At ShiraTech-Knowtion, we played a pivotal role in bringing this vision to life.
StorAIge, short for ‘Embedded Storage Elements on Next MCU Generation Ready for AI at the Edge’, is a large-scale European research initiative focused on emerging memory technologies and AI models that can run directly on constrained hardware. The goal is to enable predictive, reliable and energy-efficient edge systems that reduce latency, save bandwidth and improve autonomy. With dozens of partners across Europe, the StorAIge project addressed the entire innovation chain, from new semiconductor materials to real-world application demonstrators.
Our Contribution
Our focus was on developing machine learning algorithms optimized for resource-limited microcontrollers, which serve as a core building block for true Edge AI.
We implemented and refined a Hoeffding-Tree classification model, specifically adapted to:
- Operate with extremely low memory requirements while maintaining high accuracy
- Incrementally update its knowledge in real time, allowing it to adapt to changes in operating conditions (known as concept drift)
- Perform real-time classification of complex movement and vibration patterns from sensor data
- Integrate into embedded systems without sacrificing speed or reliability
This work was backed by a feature extraction pipeline combining statistical, time-domain, and frequency-domain methods, all tailored for microcontrollers with very limited resources.
Real-World Applications
Logistics
Smarter Pallet Tracking
In collaboration with Microsensys GmbH, we developed a system to classify forklift movements during pallet transport using 3-axis accelerometer data. The model distinguished between vertical motion, horizontal motion, and pauses, enabling better safety monitoring and operational insight. The result was a classification accuracy of 92.71%, achieved on hardware with only kilobytes of RAM.
Renewable Energy
Wind Turbine Monitoring
Together with Endiio engineering GmbH and ZF Friedrichshafen AG, we adapted our algorithms to monitor gearbox bearings in wind turbines. Using vibration sensor data, we applied anomaly detection techniques such as the Isolation Forest to identify early signs of wear or damage. The result was a detection accuracy of 84–91%, enabling earlier maintenance interventions and reduced downtime.