Enhancing interoperability in IoT platforms using small language models
Loading...
Date
2026-07-08
Authors
Σακελλαρίου, Δήμος
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Modern cities depend more and more on data from sensors, vehicles, traffic systems, and public services. The problem is that all this data needs to be shared between many different applications, but each manufacturer uses its own format. This is exactly the issue that FIWARE addresses: it is an open-source framework, supported by the European Commission, that provides standard building blocks for smart city applications, so that systems from different vendors can work together.
The core of FIWARE is NGSI-LD, a standard way to describe real-world things, like sensors and parking spots, as digital entities with properties and relationships. On top of NGSI-LD, FIWARE provides the Smart Data Models catalogue, which is a community-maintained collection of ready-made entity definitions for many areas: parking, lighting, agriculture, water management, energy, and others. The idea is that if every developer uses the same Smart Data Models, then an application written for one city can be reused in another with little effort, which is the foundation of true interoperability and scalability across smart city deployments.
In practice, this is not so simple. Most commercial smart parking systems today are closed and proprietary. The biggest reason is the cost of integrating new sensors: every manufacturer sends data in a different format, and someone has to write code to translate that format into the canonical Smart Data Model. This has to be done for every new sensor, which makes it impossible for open systems to grow and limits the scalability of open systems by slowing the adoption of FIWARE in real deployments.
The main part of the thesis is a smart parking application built on FIWARE and currently running with real LoRaWAN parking sensors at the University of Patras campus. The system uses Orion-LD as the Context Broker, FIWARE IoT Agents for sensor integration, and ChirpStack for the LoRaWAN network. A web application allows users to search for parking spots, see availability in real time, check details for each lot, and get navigation help. The data inside the system follows the Smart Data Models (Device, ParkingSpot, OffStreetParking), which means the platform is interoperable with any other FIWARE-compliant system.
To make the platform scalable, that is, to make it possible to add new sensors and new parking lots without rewriting code, the thesis introduces a Dynamic Sensor Registry, which allows new sensors or new parking lots to be added to the running system through a single HTTP call, without modifying any code or resetting any service. This is a crucial architecture feature and the main mechanism for scalability in the platform: once a new source is registered, the backend automatically includes it in every subsequent user request. A supporting backend layer handles the practical side of this, querying the Registry, calling each registered NGSI-LD source, and combining the results into one consistent response for the frontend. To verify that this is a solid solution, a Mock Orion API was created to provide simulated parking data in NGSI-LD format. After registering it in the Dynamic Registry, its parking lots appeared immediately in the application, which shows that adding a new data source can be done with minimal effort.
Even with the Dynamic Registry in place, one part of the onboarding process remains manual: every new sensor needs a parser to translate its raw payload into Smart Data Model schema. The Registry makes it easy to plug in a new source once the parser exists, but it does not write the parser itself. This means that for every new vendor, a developer must still study the device's documentation, the payload structure, and write a “translation” code specific to the sensor. Recent progress in Large Language Models raises a natural question: can Large Language Models help with this last manual step, by reading a few sample payloads and helping with the translation process? If LLMs can handle this work reliably, the cost of onboarding a new vendor could drop, and FIWARE-based platforms could scale without significant proportional engineering effort. The third part of the thesis investigates exactly this question.
The last part is a study about whether Large Language Models (LLMs) could help reduce the manual work of writing parsers for new sensors. Eight small open-source models, were tested with three different prompting strategies on two parking sensors that represent different difficulty levels. A main finding is that the success of an LLM does not really depend on the size of the payload or the size of the model, but on how similar the sensor's field names are to the Smart Data Model names. When the names are similar, the models can translate the payload almost perfectly. When the names are different, even the better models fail in specific ways, which the thesis describes as four failure modes. Adding a small vocabulary hint to the prompt fixes most of the problem for some models.
Overall, the thesis offers a working FIWARE-based smart parking application, an architectural pattern that brings true scalability to multi-vendor onboarding, and an exploratory study of how LLMs could further reduce the integration cost in the future. The goal is to show one practical and fully open-standards-based path that other smart city applications can reuse.
Description
Keywords
FIWARE, NGSI-LD, Smart data models, Smart parking, Parking sensors, Temperature, Batteries, Data translation, Small language models, Interoperability, Scalability, Artificial intelligence