A DataOps Toolbox Enabling Continuous Semantic Integration of Devices for Edge-Cloud AI Applications

copyright © Cefriel –All rights reserved 0 32 96 255 220 6 47 60 182 0 83 255 40 130 255 94 170 240 166 199 240 91 101 135 0 32 96 42 212 138 255 181 6 #1 #2 #3 #4 #5 50 172 189 64 211 204 104 235 228 157 240 234 214 240 235 #1 #2 #3 #4 #5 Primary 5 3 14 2 5 3 1 4 2 GRAPHS Sequence of use CEFRIEL IDENTITY 0 32 96 255 40 78 255 220 6 00 00 00 Text Title 1 Title 2 TEXT Secondary Sequence of use Highlight BACKGROUND RAG A DataOps Toolbox Enabling Continuous Semantic Integration of Devices for Edge-Cloud AI Applications Mario Scrocca1, Marco Grassi1, Alessio Carenini1, Darko Anicic2, Jean-Paul Calbimonte3, Irene Celino1 In-Use Track November 4th, 2025 International Semantic Web Conference 2025, Nara, Japan 1 2 3 copyright © Cefriel –All rights reserved 0 32 96 255 220 6 47 60 182 0 83 255 40 130 255 94 170 240 166 199 240 91 101 135 0 32 96 42 212 138 255 181 6 #1 #2 #3 #4 #5 50 172 189 64 211 204 104 235 228 157 240 234 214 240 235 #1 #2 #3 #4 #5 Primary 5 3 14 2 5 3 1 4 2 GRAPHS Sequence of use CEFRIEL IDENTITY 0 32 96 255 40 78 255 220 6 00 00 00 Text Title 1 Title 2 TEXT Secondary Sequence of use Highlight BACKGROUND RAG I N T R O D U C T I O N 2 Enable AI Applications across Heterogeneous Devices Problem ‣data interoperability issues (e.g., different protocols/data formats) ‣deployment of devices across different environments (resource/deployment constraints, performance/scalability requirements, etc.) Contributions ‣Continuous Semantic Integration (CSI) approach based on Semantic Web technologies ‣DataOps toolbox to support the interoperability of static nodes’ information and runtime data exchanges ‣How the DataOps toolbox effectively supports three use cases in the industrial, automotive, and health domains Photo by Eric Krull on Unsplash Goal: Support low-code AI apps development for the collaboration of different devices across edge and cloud copyright © Cefriel –All rights reserved 0 32 96 255 220 6 47 60 182 0 83 255 40 130 255 94 170 240 166 199 240 91 101 135 0 32 96 42 212 138 255 181 6 #1 #2 #3 #4 #5 50 172 189 64 211 204 104 235 228 157 240 234 214 240 235 #1 #2 #3 #4 #5 Primary 5 3 14 2 5 3 1 4 2 GRAPHS Sequence of use CEFRIEL IDENTITY 0 32 96 255 40 78 255 220 6 00 00 00 Text Title 1 Title 2 TEXT Secondary Sequence of use Highlight BACKGROUND RAG USE CASES 3 Low-Code Edge Intelligence for Smart Factory Objectives ‣Enable low-code configuration of production lines for order-specific product assembly ‣Support flexible reprogramming to meet changing manufacturing demands Devices: Demonstrator in the Siemens Autonomous Factory Lab (AFL) considering three macro-modules: assembly, mini-backbone (transport/distribution), and shipping. Challenges ‣To enable the discovery and low-code programmability of heterogeneous devices, there is the need for structured and interoperable device descriptions ‣OPC UA standard can not be fully expressed using Web of Things (WoT) and needs dedicated transformation and discovery https://tangerexperience.siemens.cloud/AFL/ copyright © Cefriel –All rights reserved 0 32 96 255 220 6 47 60 182 0 83 255 40 130 255 94 170 240 166 199 240 91 101 135 0 32 96 42 212 138 255 181 6 #1 #2 #3 #4 #5 50 172 189 64 211 204 104 235 228 157 240 234 214 240 235 #1 #2 #3 #4 #5 Primary 5 3 14 2 5 3 1 4 2 GRAPHS Sequence of use CEFRIEL IDENTITY 0 32 96 255 40 78 255 220 6 00 00 00 Text Title 1 Title 2 TEXT Secondary Sequence of use Highlight BACKGROUND RAG 4 Smart Vehicle to Infrastructure Objectives: ‣Develop intelligent AI-enabled traffic management applications with minimal programming. ‣Support real-time computation of traffic indicators from heterogeneous data (cameras, lidars, radars, and public transport feeds) and use them to define adaptive traffic control rules and send signals to connected vechicles. Devices: Demonstrator in Helsinki’s Mobility Lab corridor comprising 6 intersections, 17 radars, 21 edge units and 2 (now 3!) connected vehicles. Challenges: ‣Heterogeneous data streams to be harmonised and fused at runtime ‣Edge computation at each intersection for low-latency decision-making ‣Minimal resource consumption on edge hardware USE CASES V2X Safety Green Extension copyright © Cefriel –All rights reserved 0 32 96 255 220 6 47 60 182 0 83 255 40 130 255 94 170 240 166 199 240 91 101 135 0 32 96 42 212 138 255 181 6 #1 #2 #3 #4 #5 50 172 189 64 211 204 104 235 228 157 240 234 214 240 235 #1 #2 #3 #4 #5 Primary 5 3 14 2 5 3 1 4 2 GRAPHS Sequence of use CEFRIEL IDENTITY 0 32 96 255 40 78 255 220 6 00 00 00 Text Title 1 Title 2 TEXT Secondary Sequence of use Highlight BACKGROUND RAG USE CASES 5 Smart Health Rehabilitation Objectives: ‣Support home-based physiotherapy (neck rehabilitation post-surgery or injury recovery) through AI-driven activity tracking ‣Enable low-code personalization of exercises by the physiotherapist Devices: Tested in collaboration with PhysioLab at HES-SO by providing wearables sensors (3 for each patient) and a tablet for exercise execution Challenges: ‣Support remote exercise definition and monitoring by the physiotherapist ‣Matching and configuring heterogeneous devices (e.g., personal devices) with low capabilities Heterogeneous Sensors Tablet copyright © Cefriel –All rights reserved 0 32 96 255 220 6 47 60 182 0 83 255 40 130 255 94 170 240 166 199 240 91 101 135 0 32 96 42 212 138 255 181 6 #1 #2 #3 #4 #5 50 172 189 64 211 204 104 235 228 157 240 234 214 240 235 #1 #2 #3 #4 #5 Primary 5 3 14 2 5 3 1 4 2 GRAPHS Sequence of use CEFRIEL IDENTITY 0 32 96 255 40 78 255 220 6 00 00 00 Text Title 1 Title 2 TEXT Secondary Sequence of use Highlight BACKGROUND RAG 6 Continuous Semantic Integration Heterogeneous group of devices that should collaborate to implement an AI application across edge and cloud copyright © Cefriel –All rights reserved 0 32 96 255 220 6 47 60 182 0 83 255 40 130 255 94 170 240 166 199 240 91 101 135 0 32 96 42 212 138 255 181 6 #1 #2 #3 #4 #5 50 172 189 64 211 204 104 235 228 157 240 234 214 240 235 #1 #2 #3 #4 #5 Primary 5 3 14 2 5 3 1 4 2 GRAPHS Sequence of use CEFRIEL IDENTITY 0 32 96 255 40 78 255 220 6 00 00 00 Text Title 1 Title 2 TEXT Secondary Sequence of use Highlight BACKGROUND RAG 7 Continuous Semantic Integration Description of the devices and their capabilities copyright © Cefriel –All rights reserved 0 32 96 255 220 6 47 60 182 0 83 255 40 130 255 94 170 240 166 199 240 91 101 135 0 32 96 42 212 138 255 181 6 #1 #2 #3 #4 #5 50 172 189 64 211 204 104 235 228 157 240 234 214 240 235 #1 #2 #3 #4 #5 Primary 5 3 14 2 5 3 1 4 2 GRAPHS Sequence of use CEFRIEL IDENTITY 0 32 96 255 40 78 255 220 6 00 00 00 Text Title 1 Title 2 TEXT Secondary Sequence of use Highlight BACKGROUND RAG 8 Continuous Semantic Integration Enable an interoperable description [1] of devices and their capabilities. Support the API-based discoverability of nodes. [1] SmartEdge Schema https://w3id.org/smartedge/smartedge-schema copyright © Cefriel –All rights reserved 0 32 96 255 220 6 47 60 182 0 83 255 40 130 255 94 170 240 166 199 240 91 101 135 0 32 96 42 212 138 255 181 6 #1 #2 #3 #4 #5 50 172 189 64 211 204 104 235 228 157 240 234 214 240 235 #1 #2 #3 #4 #5 Primary 5 3 14 2 5 3 1 4 2 GRAPHS Sequence of use CEFRIEL IDENTITY 0 32 96 255 40 78 255 220 6 00 00 00 Text Title 1 Title 2 TEXT Secondary Sequence of use Highlight BACKGROUND RAG 9 Continuous Semantic Integration Enable the low-code specification of recipes for the collaboration of nodes with different capabilities Interoperable representation of recipes [2] [2] Recipe Model https://w3id.org/smartedge/recipe-model copyright © Cefriel –All rights reserved 0 32 96 255 220 6 47 60 182 0 83 255 40 130 255 94 170 240 166 199 240 91 101 135 0 32 96 42 212 138 255 181 6 #1 #2 #3 #4 #5 50 172 189 64 211 204 104 235 228 157 240 234 214 240 235 #1 #2 #3 #4 #5 Primary 5 3 14 2 5 3 1 4 2 GRAPHS Sequence of use CEFRIEL IDENTITY 0 32 96 255 40 78 255 220 6 00 00 00 Text Title 1 Title 2 TEXT Secondary Sequence of use Highlight BACKGROUND RAG D A T A O P S P I P E L I N E S A C R O S S U S E C A S E S 16 Low-Code Edge Intelligence for Smart Factory Semantic Model for Capabilities A recipe orchestrating low-code skills for a new product Skills discovery Production stations IT/OT Integration A new product to be manufactured Real product DataOps pipelines: ‣Insertion and retrieval of OPC UA nodes descriptions in the KG: OPC UA → RDF (lifting) and RDF → OPC UA (lowering) ‣Endpoint to support skills discovery of available devices. Adds support for OPC UA devices and is aligned with the WoT TD Discovery (Domus TDD). Validation results: ‣AFL devices successfully described in the KG and made available for discovery ‣LLM-enabled skill discovery leverage the KG and removes dependency on SPARQL queries ‣Technicians could define and execute new recipes for production workflows adopting a low-code approach via Mendix copyright © Cefriel –All rights reserved 0 32 96 255 220 6 47 60 182 0 83 255 40 130 255 94 170 240 166 199 240 91 101 135 0 32 96 42 212 138 255 181 6 #1 #2 #3 #4 #5 50 172 189 64 211 204 104 235 228 157 240 234 214 240 235 #1 #2 #3 #4 #5 Primary 5 3 14 2 5 3 1 4 2 GRAPHS Sequence of use CEFRIEL IDENTITY 0 32 96 255 40 78 255 220 6 00 00 00 Text Title 1 Title 2 TEXT Secondary Sequence of use Highlight BACKGROUND RAG D A T A O P S P I P E L I N E S A C R O S S U S E C A S E S 17 Smart Vehicle to Infrastructure DataOps pipelines: ‣Harmonize multimodal sensor data and public transport feeds into RDF using ASAM OpenX and SOSA ontologies ‣Fusion logic to correct sensor inaccuracies (e.g., tram misidentified as trucks) ‣Deployment via templates of pipelines on edge units (RSUs) and cloud environments to test performance and scalability Validation results: ‣Deployed in 6 intersections across Helsinki’s Mobility Lab corridor ‣RDF stream generated is processed to compute indicators ‣Edge deployment: Average pipeline latency <100 ms, memory usage ≈100 MB per edge unit. Successfully met expectations for 10 msg/sec at intersections (4 KB/msg typical load). ‣Laboratory benchmark: processed up to 400 requests/sec, average <7 ms per message (60 KB payloads). Full description of testing activities in D3.3 at https://www.smart-edge.eu/deliverables/ copyright © Cefriel –All rights reserved 0 32 96 255 220 6 47 60 182 0 83 255 40 130 255 94 170 240 166 199 240 91 101 135 0 32 96 42 212 138 255 181 6 #1 #2 #3 #4 #5 50 172 189 64 211 204 104 235 228 157 240 234 214 240 235 #1 #2 #3 #4 #5 Primary 5 3 14 2 5 3 1 4 2 GRAPHS Sequence of use CEFRIEL IDENTITY 0 32 96 255 40 78 255 220 6 00 00 00 Text Title 1 Title 2 TEXT Secondary Sequence of use Highlight BACKGROUND RAG D A T A O P S P I P E L I N E S A C R O S S U S E C A S E S 18 Smart Health Rehabilitation DataOps pipelines: ‣Exercise transformation to the Recipe Model ontology ‣Device-capability matching for exercise execution considering available devices ‣Runtime conversion of recipe metadata for the coordinator ‣Karavan low-code editor used to define pipelines ‣Cloud deployment of pipelines as a mediator service invoked by local orchestrators running on the edge Validation results: ‣Exercise definition successfully tested by physiotherapy lecturer ‣Involvement of students wearing devices and performing exercises under the supervision of the physiotherapist for the validation of the exercise movements recognition copyright © Cefriel –All rights reserved 0 32 96 255 220 6 47 60 182 0 83 255 40 130 255 94 170 240 166 199 240 91 101 135 0 32 96 42 212 138 255 181 6 #1 #2 #3 #4 #5 50 172 189 64 211 204 104 235 228 157 240 234 214 240 235 #1 #2 #3 #4 #5 Primary 5 3 14 2 5 3 1 4 2 GRAPHS Sequence of use CEFRIEL IDENTITY 0 32 96 255 40 78 255 220 6 00 00 00 Text Title 1 Title 2 TEXT Secondary Sequence of use Highlight BACKGROUND RAG D A T A O P S T O O L B O X E V A L U A T I O N 19 Evaluation User Evaluation +Pipeline configuration: low-code configurability and reusability of components in the DataOps toolbox +Deployment templates: possibility of deploying the same pipeline in different deployment environments with minimum effort -Learning curve: Some familiarity with Apache Camel required, mapping rules are easier to write for developers Uptake & Impact Evaluation ‣Continuous Semantic Integration approach validated across heterogeneous devices on real deployments in different domains ‣Siemens will continue investigating lowcode programmability of machines (including OPC-UA devices) as a key assets for industrial innovation ‣Helsinki’s mobility stakeholders plan to adopt the approach for edge-based semantic data fusion ‣The rehabilitation prototype is used in HES-SO physiotherapy education, with ethical approval pending for larger trials Significance Evaluation ‣CSI approach: Semantic Web technologies and WoT specification ensures strong interoperability foundations ‣Flexibility: complex schema and data transformations can be handled (e.g., as demonstrated for OPC UA) copyright © Cefriel –All rights reserved 0 32 96 255 220 6 47 60 182 0 83 255 40 130 255 94 170 240 166 199 240 91 101 135 0 32 96 42 212 138 255 181 6 #1 #2 #3 #4 #5 50 172 189 64 211 204 104 235 228 157 240 234 214 240 235 #1 #2 #3 #4 #5 Primary 5 3 14 2 5 3 1 4 2 GRAPHS Sequence of use CEFRIEL IDENTITY 0 32 96 255 40 78 255 220 6 00 00 00 Text Title 1 Title 2 TEXT Secondary Sequence of use Highlight BACKGROUND RAG 20 Lessons learned ‣Stakeholders recognize the value of the CSI approach, but need to gain expertise in Semantic Web technologies. Low-code approaches can lower the barrier and Large Language Models (LLMs) can further streamline the user experience for certain tasks (e.g., devices discovery) ‣Mapping rules and integration logic can often be associated with complex requirements that can not be addressed automatically. The DataOps toolbox trades off a fully declarative approach to provide additional flexibility and customizability [1] ‣Difficulty of finding standardised vocabularies across different domains. Adds an additional step for implementing the proposed approach, e.g., requiring ontology engineering expertise ‣Balancing performance with sustainability is crucial, and we focused on enabling the instrumentation of DataOps pipelines to support their observability and efficient resource management. [1] M. Grassi et al. "typhon-rml: Modularised Declarative Knowledge Graph Construction for Flexible Integrations and Performance Optimisation" copyright © Cefriel –All rights reserved 0 32 96 255 220 6 47 60 182 0 83 255 40 130 255 94 170 240 166 199 240 91 101 135 0 32 96 42 212 138 255 181 6 #1 #2 #3 #4 #5 50 172 189 64 211 204 104 235 228 157 240 234 214 240 235 #1 #2 #3 #4 #5 Primary 5 3 14 2 5 3 1 4 2 GRAPHS Sequence of use CEFRIEL IDENTITY 0 32 96 255 40 78 255 220 6 00 00 00 Text Title 1 Title 2 TEXT Secondary Sequence of use Highlight BACKGROUND RAG D A T A O P S T O O L B O X 21 Thank You! Any questions? MARIO SCROCCA Knowledge Engineer Cefriel marioscrocca mario.scrocca @cefriel.com Discover the DataOps toolbox on GitHub at: github.com/cefriel Key Takeaways •The CSI approach leverages Semantic Web technologies to empower low-code AI application development across edge and cloud devices •The DataOps toolbox provides reusable components to build and deploy pipelines independently from the scenario requirements •Validation in use cases confirmed the approach’s adaptability and realworld relevance Future work •Investigate the role of LLMs in supporting the proposed CSI approach (e.g., minimising the effort required to configure DataOps pipelines). •Define reusable DataOps pipelines as components compliant with the WoT architecture Read more about the SmartEdge project at: smart-edge.eu