Model Learning for Cloud-Edge Digital Twins

Experiment description

MoLe aims to facilitate the development and use of digital twins for smart factories, so factory stakeholders can enjoy all the benefits of digital twin technologies with little effort.

In order to achieve this, we have connected the following technologies to the MIDIH platform:

• A data translator to offer better flexibility in data formats, capable of translating data on the fly, without adding any significant delay in data delivery.
• The Scorpio Broker, to quickly ingest digital twin’s data.
• Knowledge Infusion, a new method to combine ML with external domain knowledge, reducing the effort needed to create simulation and prediction models for digital twins.
• FogFlow for definition of serverless functions, that encapsulate ML models and other algorithms, instantiating them on-demand and seamlessly running them on the Edge and Cloud.

By combining these technologies, we have created a Digital Twin Execution Framework that allows factory stakeholders to easily implement Digital Twins of machines and processes. We have tested our approach at the MIDIH didactic factory at Politecnico di Milano, supporting the process of manufacturing and assembly. The image below, shows the architecture we have implemented during the MoLe experiment.

Technical impact

During MoLe, we have successfully implemented the data translator and made it available as open source software to the community, being a part of the Scorpio Broker. The translator has a worst-case performance of 1ms, easily keeping up with all the data from the production line in real-time. Thanks to Scorpio, it is now possible to access data of the didactic factory in real-time by consuming it from an online endpoint.

In addition, we have implemented the software architecture to allow the easy creation of digital twins. To test this architecture, we have created two digital twins of two stations in didactic factory: the Front Cover Magazine and the Press Station. These digital twins are able to monitor all the data from the real twin (each station) and infer their state and energy consumption. We have used Knowledge Infusion to support the monitoring of the station’s state with an accuracy of 82.33%, while we used pure ML for predicting energy usage, obtaining a MAE of 4% in our predictions.

Economical / Business impact

The NGSI-LD Translator is published as open source software on the Github repository of the Scorpio Broker. We will promote it towards the FIWARE community as an integration tool for NGSI-LD. We plan to reuse and expand the Translator in future projects to integrate further data sources into NGSI-LD. Overall, the NGSI-LD Translator is a generic, programmable tool for capturing JSON-based data and convert it to NGSI-LD. This is a common task during the Knowledge Acquisition phase on any AI system.

The insights gained in the MoLe experiment were shared with various business groups within NEC, especially the Smart Industry team but also to a new team targeting Smart Agriculture. As the topic of Digital Twin is gaining importance for Smart Industries/Smart Agriculture, we will explore the approach further with our respective business units for taking advantage of the Digital Execution Framework developed during this experiment.