Toward Standardized IoT Ontologies Using a Machine Learning-Based Framework for Seamless Data Exchange

Summary

This study presents a machine learning-based framework for real-time IoT ontology alignment, enabling seamless data exchange across heterogeneous systems. By integrating semantic modeling and adaptive optimization, the approach enhances interoperability, reduces latency, and achieves high accuracy. Validated in real-world settings, it offers a scalable, standardized IoT integration solution.

Abstract

The increasing heterogeneity of Internet of Things (IoT) devices has led to significant challenges in achieving real-time interoperability and seamless data exchange. Existing IoT ecosystems often operate using diverse data models, communication protocols, and semantic representations, resulting in fragmented systems that hinder integration. To address this problem, we propose a unified framework that employs machine learning-based ontology alignment for standardized, adaptive IoT integration. The hypothesis guiding this research is that combining semantic modeling with intelligent optimization techniques can significantly improve the consistency and efficiency of data exchange across heterogeneous IoT environments. The proposed framework integrates real-time data stream processing, semantic similarity analysis, and adaptive ontology mapping to dynamically align device ontologies. Using simulated and real-world environments, including smart homes and healthcare systems, the framework was tested against key performance metrics such as accuracy, latency, and interoperability rate. Results demonstrate that the proposed method achieves a high ontology alignment accuracy of 97%, reduces latency to under 20 ms, and maintains over 95% interoperability among diverse device types. The findings confirm that the integration of machine learning algorithms with semantic modeling significantly enhances the performance, scalability, and adaptability of IoT systems. The framework successfully addresses semantic inconsistencies and supports dynamic device onboarding without manual intervention. This study presents a robust and scalable solution for IoT interoperability, offering real-time, intelligent ontology alignment that is adaptable to evolving devices and data standards. This work contributes to the development of next-generation IoT architectures capable of supporting standardized, efficient, and automated communication across diverse applications.

Introduction

The Internet of Things (IoT) is rapidly evolving into a core infrastructure for smart environments, connecting a wide array of heterogeneous devices that operate across diverse domains such as healthcare, smart cities, agriculture, and industrial automation^1,2,3. These devices generate large volumes of data and rely on semantic understanding to communicate meaningfully^4,5,6,7. However, the lack of a standardized semantic structure has emerged as a key barrier to seamless data exchange^8,9,10,11,12. Diverse ontologies, varying protocols, and inconsistent data models limit interoperability, making it difficult for devices to collaborate efficiently in real time. This semantic fragmentation often leads to misinterpretations, increased latency, and integration failures that compromise system scalability and performance^5,12,13. Therefore, there is a critical need for a unifying approach that can standardize semantic representations while adapting to the dynamic nature of IoT environments^6,10,11,12.

This study proposes a novel machine-learning-based framework for real-time ontology alignment to address these interoperability challenges^4,8,9,11,14. The framework combines advanced semantic modeling, adaptive ontology mapping, and data stream integration^9,10,13 to dynamically align IoT ontologies. Leveraging machine learning techniques for ontology alignment can significantly enhance semantic consistency and integration efficiency in heterogeneous IoT systems^8,9,14,15. Unlike traditional rule-based methods^{4,5,7,8,9,10,13}, the proposed approach employs weighted similarity measures, structural and lexical mappings, and optimization algorithms that adapt to incoming data^15,16,17. Experimental validation was conducted in both simulated and real-world environments, including smart homes and healthcare systems. The results demonstrate that the framework achieves an ontology alignment accuracy of 97%, reduces average latency to under 20 ms, and maintains interoperability rates above 95% across device types⁸.

The adaptability of the proposed framework further distinguishes it from existing solutions^9,13,15. As new devices join the network, the system automatically updates semantic relationships without requiring manual reconfiguration^{5,6,8,9,13,15}. This is made possible through the integration of learning-based optimization that minimizes semantic mismatches and computational overhead^8,9. The system also exhibits improved performance in memory and CPU usage, making it suitable for resource-constrained environments^1,2,3,11,12. By addressing semantic inconsistency, latency, and scalability in a unified architecture, this work lays the foundation for a standardized, intelligent, and future-ready IoT communication model^13,18.

The goal of this study is to develop and validate a machine-learning-based framework for real-time ontology alignment to facilitate semantic interoperability in heterogeneous IoT environments. The study includes the following phases: (1) Ontology collection and analysis, (2) Semantic modeling and alignment, (3) Machine learning-based optimization, and (4) Real-world validation in simulated and live IoT contexts. The experimental framework was tested using public datasets, simulated environments, and live deployments in smart home and healthcare environments. A schematic representation of a machine learning based IoT ontology alignment framework is provided in Figure 1.

Figure 1: Machine learning-based IoT ontology alignment framework. This figure presents a high-level view of the proposed framework that integrates machine learning techniques to align ontologies across heterogeneous IoT systems. The pipeline includes semantic similarity assessment, optimization using loss minimization, and iterative refinement for enhanced interoperability. Please click here to view a larger version of this figure.

Protocol

This research did not involve human or vertebrate subjects or tissue sampling. All experiments were performed in compliance with institutional computational research guidelines at J. C. Bose University of Science & Technology, YMCA, Faridabad.

Ontology collection and evaluation
Public ontologies relevant to healthcare, smart homes, and industrial monitoring were obtained from established repositories, including Linked Open Vocabularies (LOV) and domain-specific portals, in RDF/OWL formats^1,2,3. Each ontology was inspected in an ontology editor (for example, Protégé) and programmatically parsed to extract class hierarchies, object and data properties, and associated metadata in accordance with RDF/OWL specifications^1,2.

For pairwise comparisons, lexical similarity between class labels was computed using a normalized edit-distance based string similarity⁴. Structural similarity was derived by traversing subclass and superclass relations and computing overlap ratios of local neighborhoods. Instance similarity was assessed by detecting shared or compatible instances across ontologies and verifying datatype compatibility. A combined score was defined as Equation (1):

S_total=w1S_lexical+w2S_structural+w3S_instance     Equation (1)

with initial weights w1 = 0.5, w2 = 0.3, w3 = 0.2, and a threshold selected from a held-out validation subset. Candidate correspondences with combined scores at or above the threshold were logged in a CSV register with ontology identifiers, entity IRIs, component scores, and the decision outcome for downstream alignment.

Semantic modeling and data integration
To convert raw IoT streams into ontology-compliant representations, a unified ontology was constructed that covers static metadata (e.g. Student, Course, Device) and dynamic observations (e.g. Sensor, Observation, Timestamp), aligning with RDF/OWL best practices¹⁰ and established IoT Semantic^5,6,13. Semantic transformation was defined by the mapping function:

D_semantic=f_map (D_raw, O)     Equation (2)

where O denotes the applied ontology and f_map materializes triples with stable IRIs. Transformed data were stored as RDF and validated for consistency using SPARQL queries; a SPARQL endpoint was configured to support semantic queries across sources¹⁹. The resultant RDF artifact was archived as semantic_data.rdf, and a synchronized tabular export (a normalized wide table) was produced for learning tasks, including a data-provenance column that indicates feature origin and any imputation or scaling applied.

Ontology alignment using machine learning
A hybrid approach combined rule-based similarity signals with supervised learning to select high-confidence correspondences^14,15. The training set comprised 2,500 manually annotated ontology-element pairs (match or non-match) to create a balanced dataset. Features included the three similarity components together with auxiliary statistics such as label-length difference. A decision-tree ensemble (Random Forest) with 100 trees, maximum depth 10, class_weight = "balanced", and a fixed random seed (42) was trained following established practice¹⁴. The learning objective was formalized as Equation (3):

    Equation (3)

where y_i is the ground truth label, y_i^{^} is the predicted label, and λ represents the regularization parameter.

After probability calibration, pairs with predicted probability ≥ 0.50 were retained, and one-to-one constraints were enforced to avoid many-to-one mappings. The final mapping file, including source and target IRIs, confidence scores, and feature contributions, was serialized as alignment_results.json; the trained model artifact was stored as alignment_model.pkl. Model quality was reported on a held-out test set using accuracy, precision, recall, F1-score, and a confusion matrix with 95% confidence intervals computed by five-fold cross-validation^14,15.

Deployment and real-world evaluation
A simulated environment representing 1,000 heterogeneous devices was created to stress-test ingestion and alignment, followed by real-world streams from smart-home and healthcare settings^5,11. Ontology-aligned data were integrated into a cloud middleware layer for real-time ingestion and querying^11,12. New devices were automatically onboarded by inspecting incoming data headers and applying dynamic semantic classification using the trained alignment model^14,15.

System performance was evaluated along four axes: (i) alignment quality (the classification metrics above), (ii) interoperability rate (successful cross-source interactions as a percentage of attempted interactions), (iii) end-to-end latency defined as Equation (4), and (iv) resource utilization (CPU and memory).

L= T_response−T_request     Equation (4)

Configuration files, logs, and outputs were archived with stable file names and a README to facilitate independent replication¹⁹. The completed framework consistently enabled semantic interoperability in heterogeneous IoT settings; data outputs, performance logs, and configurations are bundled to support reproducibility and external validation^11,12,19.

Implementation details for replication
The pipeline was implemented in Python (version 3.10 or later) using rdflib and owlready2 for RDF/OWL handling, scikit-learn for model training and evaluation, and standard data libraries for preprocessing^8,14,15. Random seeds were fixed, library versions were pinned in a requirements.txt file, and a canonical directory structure was used: ontologies/, data/raw/, data/processed/, models/, and results/. File names referenced in text are exact: semantic_data.rdf, alignment_results.json, and alignment_model.pkl¹⁹.

Research design
This mixed-methods, computational study integrates semantic modeling with a supervised ontology-alignment module to improve interoperability across heterogeneous IoT sources. The protocol proceeds as follows in brief: public ontologies are collected and parsed; raw IoT tables are mapped to a unified RDF/OWL schema¹⁰; candidate correspondences are generated from lexical, structural, and instance signals^15,16; a compact decision-tree ensemble is trained on a manually labeled set to accept one-to-one matches at a fixed threshold^14,15; aligned graphs/tables are fused; and performance is assessed by correspondence-level metrics (accuracy, precision, recall, F1, confusion matrix) and system behavior (interoperability rate, end-to-end latency, CPU/memory)^14,15. A rules-only weighted-similarity method serves as the baseline¹⁶. Random seeds and library versions are fixed, and artifacts use stable names (semantic_data.rdf, alignment_results.json, alignment_model.pkl) to support replication. Figure 2 illustrates the overall framework for Unified Ontology Design in IoT environments.

Figure 2: Unified ontology design in IoT environments. The figure illustrates the core architectural layers of the unified ontology, showcasing semantic modeling, AI integration, and domain knowledge abstraction for seamless IoT data interoperability. The framework supports both cross-domain and domain-specific ontological structures. Please click here to view a larger version of this figure.

Data collection
Data collection was performed in two distinct phases.

Ontology Evaluation: A critical analysis of existing IoT ontologies and semantic models was conducted to identify gaps in current integration strategies. Publicly available IoT datasets and domain-specific ontologies were examined to evaluate structural and semantic alignment limitations^5,6,10,11.

Trial Endorsement: Real-world IoT systems (e.g., smart homes, healthcare applications, industrial automation) were utilized to collect data for validating the framework. Additionally, simulated datasets were generated to assess the system's performance under controlled and repeatable conditions^8,9,11,12.

Data analysis
The analysis phase employed a combination of semantic, statistical, and AI-based techniques to validate the proposed framework.

Practical aligners include lexical string matchers using normalized Levenshtein distance(NLD) for labels/synonyms¹⁶; structural graph based methods that propagates or aggregate similarity Flooding that propagate similarity over subclass/superclass neighborhoods^12,15; logic-aware/coherence-preserving techniques that combine matching with reasoning to keep mappings consistent^15,19; instance-based/probabilistic approaches that leverage evidence from shared or compatible instances¹⁷; and hybrid/ensemble matchers that fuse multiple signals and rules^12,14,15. Recent neural variants employ contextual sentence embeddings to score candidate similarities^12,14.

Ontology Alignment: High-level alignment algorithms were used to resolve semantic discrepancies among heterogeneous IoT ontologies. These algorithms utilized heuristic rules and machine learning (ML) methods to compute accurate mappings and maintain consistency constraints^15,17. Figure 3 presents the ontology alignment process for achieving interoperability among IoT devices.

Figure 3: Ontology alignment process for IoT device interoperability. This diagram demonstrates the alignment process that resolves syntactic and semantic mismatches between devices in a heterogeneous IoT ecosystem. It emphasizes the role of lexical, structural, and instance-based similarity metrics used in the alignment algorithm. Please click here to view a larger version of this figure.

Semantic modeling: An AI-driven semantic modeling strategy was applied to develop a unified ontology capturing both domain-specific and cross-domain semantics. This model integrates static and dynamic IoT data streams and conforms to RDF/OWL best practices and established IoT semantics^5,6,10,13, supporting real-time applications.

Performance evaluation: The experimental setup evaluated key performance metrics, including interoperability rate, latency, system efficiency, and scalability ^11,12. Comparative analysis was conducted against baseline integration techniques to demonstrate the flexibility and resilience of the proposed solution^12,14,15,16. Figure 4 showcases the computational flow and critical components of the ontology alignment and semantic modeling processes.

Figure 4: Semantic model for real-time IoT applications. This figure highlights the semantic modeling strategy that integrates static device configurations and dynamic sensor data streams into the unified ontology. The architecture supports real-time semantic enrichment for context-aware decision-making. Please click here to view a larger version of this figure.

Key Equations
Equation for ontology alignment using similarity metrics: To align heterogeneous IoT ontologies, the aggregate similarity is computed as a weighted blend of lexical, structural, and instance-based signals as per Equation 1 above, where Stotal is the aggregate similarity between two ontology elements, Slexical is similarity from lexical matching (e.g., normalized string distance), Sstructural is similarity based on underlying relationships (e.g., parent-kid hierarchy), Sinstance is Comparability based on instance data (e.g., overlap in usage examples), and w1, w2, w3 are weight coefficients, adjusted based on the application context.

Equation for machine learning-based optimization for ontology mapping: The enhancement of mappings can be formulated as an expense minimization problem using a misfortune capability:

where L is Misfortune capability for mapping prediction, yi is ground truth label for the ith mapping, y^i is predicted label for the ith mapping, ||Θ|| is regularization term to prevent overfitting, λ is regularization parameter.

Equation for semantic modelling of IoT data streams: Semantic enrichment is expressed as an ontology-based data transformation as per equation 2 above, where D_semantic is the semantically annotated dataset, D_raw is the raw IoT data, O is the ontology used for annotation, and f_map is the mapping function derived from the unified ontology.

Equations for performance metrics (interoperability and latency) 
Interoperability index (percentage of successful cross-source interactions):

I_index = (successful interactions / total interactions) x 100      Equation (5)

Latency (reproduced here for completeness) is computed as per equation 4 above, where I_index is the interoperability index (in %), L is latency (ms), T_response is the response timestamp, and T_request is the request timestamp.

Equation for scalability analysis of the IoT ecosystem
Computational complexity with respect to the number of devices and ontology elements:

C = O(n · m)     Equation (6)

where C is the computational complexity, n is the number of IoT devices, and m is the number of ontology elements.

Results

Ontology collection and evaluation
Ontology analysis revealed substantial inconsistencies across domain-specific IoT ontologies in terms of class hierarchy, semantic labels, and data property definitions. These inconsistencies were more pronounced between healthcare and smart home datasets, demonstrating a 28% structural mismatch rate. The identification of these variations validated the initial hypothesis that lack of standardization impairs interoperability across IoT environments. These mismatches...

Discussion

The developed machine learning-based framework demonstrates its effectiveness in addressing semantic interoperability challenges in heterogeneous IoT environments. Through a structured protocol integrating semantic modelling, machine learning-based ontology alignment, and cloud-based middleware deployment, the system achieved high ontology alignment accuracy and consistent data integration across varied devices.

Critical protocol steps
Several steps within the proposed p...

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
Cloud-based Middleware Platform	Open-source / Proprietary (e.g., Firebase)	N/A	Facilitates real-time data ingestion and storage.
Input Ontologies	Public Repositories (e.g., LOV)	N/A	Domain-specific OWL/RDF ontologies for IoT environments.
Machine Learning Library	Open-source (e.g., scikit-learn)	N/A	Used for supervised classification model training.
Network Simulation Tool	Open-source / Commercial (e.g., NetSim)	N/A	Generates simulated heterogeneous IoT device datasets.
Ontology Editing Software	Open-source (e.g., Protégé)	N/A	Used for ontology parsing, editing, and visualization.
Programming Environment	Open-source (e.g., Python)	N/A	Implements machine learning models and data processing.
Raw IoT Data Streams	Public / Custom Dataset Sources	N/A	CSV or JSON files containing raw IoT device data.
RDF Output Files	Generated In-study	N/A	RDF/XML files representing semantically enriched IoT data.
Semantic Parsing Library	Open-source (e.g., RDFLib)	N/A	Converts IoT data into RDF triples for semantic modeling.
SPARQL Query Engine	Open-source	N/A	Validates RDF data consistency using SPARQL queries.