D4.4 - Edge data processing and service certification – Final version

This report details the final architecture and tools developed within the SEDIMARK project for edge processing and services certification. The work focuses on providing the foundational components for managing the entire lifecycle of data and AI assets, from their creation and processing at the edge to their certification and exchange in the marketplace. The key contributions establish a framework for AI-driven modules that can be deployed at the data source, adhering to MLOps principles while managing complex edge-cloud interactions.

A primary achievement of this work is the development of an edge processing framework designed for resource-constrained environments. Key innovations include:

• WebAssembly (WASM) on MCUs: A secure and sandboxed architecture was implemented to allow user-defined code to run on low-power microcontrollers, enabling flexible and frequent updates without compromising the core firmware's stability.
• Fine Timestamping and Energy Optimization: A novel algorithm was developed to calibrate the on-device Real-Time Clock (RTC) by compensating for temperature-induced drift. This significantly improves data timestamp accuracy and dramatically reduces the need for energy-intensive network synchronizations, extending the operational battery life of edge devices to meet a target of over four years.
• Edge-Cloud Orchestration: An analysis of open-source tools led to the selection and deployment of platforms like Mage.ai and Apache NiFi to manage distributed data processing flows between edge devices and the cloud.

To address the challenges of managing AI models in a diverse ecosystem, the project has established a comprehensive MLOps strategy and a solution for model interoperability. By adopting MLFlow, SEDIMARK provides a standardized framework for the entire machine learning lifecycle. A critical innovation is the use of Keras Core to create framework-agnostic model descriptions. This allows models to be defined once and then seamlessly trained or used for inference across different backends like TensorFlow, PyTorch, and JAX, which is essential for fostering collaboration in federated learning scenarios where participants may use different tools.

Finally, to build a foundation of trust within the marketplace, a multi-faceted conformity evaluation Service has been designed. This service provides validation for all marketplace assets:

• Data Assets are certified for conformance with standards like NGSI-LD and Smart Data Models, ensuring interoperability.
• Service Assets are validated for API compliance, leveraging and contributing to the official ETSI NGSI-LD Test Suite.
• AI Model Assets undergo a two-fold assessment, verifying not only their performance against quantitative KPIs but also their trustworthiness based on principles of fairness, transparency, and security, in alignment with emerging regulations like the EU AI Act.

Together, these advancements in edge computing, MLOps, and certification provide the core technical infrastructure for a robust, transparent, and efficient decentralized data marketplace.

Deliverable D4.4 can be downloaded from here.