ROVON: An Ontology for Supporting Interoperability for Underwater Robots

Abstract

Underwater robotics produces diverse and complex streams of sensor, image, video, and navigational data under challenging environmental conditions, creating obstacles for seamless integration and interpretation. This paper introduces ROVON (Remotely Operated Vehicle Ontology), a semantic framework designed to enhance interoperability and reasoning in underwater operations. While ROVON is conceptually scalable to large, heterogeneous datasets, its validation in this study focuses on controlled underwater inspection data collected for pipeline applications. ROVON enables the representation and analysis of multimodal underwater data by semantically annotating raw sensor feeds, enforcing data integrity, and leveraging knowledge graphs to convert disparate inputs into actionable insights. The ontology demonstrates how a structured semantic approach facilitates advanced analysis that improves decision-making, supports proactive maintenance strategies, and enhances operational safety. The proposed framework was validated through a controlled pipeline inspection scenario.

1. Introduction

Underwater robots, also known as subaquatic vehicles, are devices designed to adapt to underwater environments and work independently of direct human control. Their uses range from oceanographic research vessels to underwater infrastructure support, environmental monitoring, and underwater photography. Underwater robots are typically categorized into two general categories, namely, Remotely Operated Vehicles (ROVs) and Autonomous Underwater Vehicles (AUVs). AUVs are pre-programmed to interact with their environment without direct human intervention, in contrast to ROVs, which are remotely controlled by an operator [1]. Given the complexity and magnitude of the operating environment for underwater robots, it is often necessary to deploy a group of robots to accomplish a task beyond the capabilities of a single robot [2].

Inspecting and maintaining a network of underwater pipelines that transport oil and natural gas from offshore drilling platforms to onshore facilities is a prime example of a complex operation requiring multiple robots with specialized capabilities. These robots must work in close coordination to navigate extreme underwater conditions, such as high pressure, low temperature, and limited visibility. For instance, inspection robots equipped with high-resolution cameras, sonar systems, and laser scanners are dispatched to perform detailed visual and sonar inspections of the pipeline. Meanwhile, repair ROVs with specialized tools are deployed to handle maintenance tasks, including cleaning the pipeline, applying protective coatings, and conducting minor repairs. By utilizing different types of robots, each with its unique capabilities, the process becomes more efficient and safe, ensuring comprehensive inspection and upkeep of the pipelines [3]. In such environments, seamless communication and data exchange between robots and human operators are critical.

Underwater robotic systems frequently exhibit diverse hardware and software configurations, leading to significant challenges in interoperability. Disparities in sensor types, data formats, and control algorithms make inter-system communication and collaboration challenging. Additionally, the intricate nature of the underwater environment further exacerbates these technical hurdles. The data obtained during underwater operations is often diverse, incomplete, and noisy, making it difficult to accurately interpret sensor readings, reconstruct environmental conditions, maintain reliable communication, and develop robust autonomous decision-making algorithms. These data complexities are rooted in the broad spectrum of sensors used, complex environmental conditions, and the substantial volume of generated data. Addressing issues related to data interoperability, integration, and standardization is essential for effective management [4].

Although certain communication protocols for underwater robotics, such as acoustic modem standards (e.g., JANUS) [5], do exist, they primarily address the physical and network layers of data transmission rather than providing standardized data formats or semantic interoperability. The lack of widely adopted, domain-level standards for representing and interpreting heterogeneous data remains a significant barrier to efficient interaction and collaboration in multi-robot systems. These interoperability challenges are compounded by the dynamic nature of the underwater environment, requiring robotic systems to adapt in real time. Data collection, interpretation, and transmission discrepancies often lead to inefficiencies, particularly in multi-robot operations such as deep-sea exploration or pipeline inspections. When data lacks uniform formatting or is not easily understandable across different systems, task coordination between robots and even human operators becomes laborious and inefficient. This absence of cohesion underscores the necessity for improved solutions that ensure more efficient data management and real-time coordination [6,7].

An ontological approach is proposed in this work to address the interoperability issues associated with underwater robotic systems. Robots with different sensors and software can interpret and share information consistently through a shared semantic framework that defines standardized data formats and communication interfaces. This work introduces ROVON, a domain-specific ontology designed to capture underwater robotic concepts and their relationships as related to capabilities, sensor modalities, operational constraints, and mission workflows. This enables heterogeneous robots to collaborate effectively on complex tasks such as pipeline inspections and deep-sea missions, enabling coordination and decision-making that can operate in near real time at the reasoning layer. In this study, data integration was performed offline to validate the framework, while the same architecture supports streaming data ingestion for real-time operation [8]. The proposed ontology also enhances human–robot interaction, supports long-term data reuse, and enables consistent task execution across diverse platforms [9]. The proposed ontology uses Basic Formal Ontology (BFO)as the top-level ontology.

The remainder of this paper is organized as follows: Section 2 reviews prior work on interoperability in underwater robotics and ontology-driven systems. Section 3 introduces background on ontologies and knowledge graphs used in this study. Section 4 describes the ROVON ontology—its class structure, key entities and relationships, and how we model both continuants and occurrents. Section 5 demonstrates ROVON in a pipeline-monitoring scenario, including multi-source data collection, semantic mapping with Ontotext/Refine (RDF mappings), and visualization in GraphDB. Section 6 details the semantic reasoning framework implemented on RDFox, covering quantitative thresholds, qualitative inferences, and a hybrid rule set. Section 7 and Section 8 conclude the paper and outline directions for future work.

2. Related Work

The use of ontologies in underwater robotics, particularly in the operations of ROVs and Autonomous Underwater Vehicles (AUVs), has become a critical area of study because underwater missions are often very data-intensive and involve vast amounts of heterogeneous and multimodal datasets coming from disparate sources. This section reviews the related works that focus on the development and use of ontologies for communication, navigation, and operational decision-making in underwater vehicles. It explores how these frameworks address unique challenges encountered in marine environments, such as dynamic conditions, information heterogeneity, and high autonomy requirements. Additionally, this paper discusses major contributions in the field, identifies gaps that have not been adequately addressed, and suggests potential directions for future research. Special attention is given to advancements in ontology, which significantly facilitate the integration of diverse datasets, enhance the efficiency of missions, and improve interactions between humans and machines during underwater operations.

The SWARMs Ontology [8] presents an integrated ontology system that is designed to foster collaboration and interoperation between different kinds of underwater robotic systems, including Autonomous Underwater Vehicles (AUVs) and ROVs. They applied this approach to a case study on chemical pollution monitoring, showing how operational complexities are effectively handled by the ontology. Nevertheless, the experiments conducted by the authors were limited in scope as they dealt only with specific situations without considering a wider range of applications in underwater robotics. However, it remains unclear whether these systems would perform well beyond their tested environments. Hence, there is a need for testing them under various real-world scenarios to establish their efficacy across different operational settings. Moreover, although the ontology accommodates certain levels of vagueness, the ontology’s capability to deal with partial information and ambiguities of typical deep-sea mission challenges must be strengthened further. These drawbacks point out requirements, indicating that such models should evolve toward more holistic solutions addressing diverse marine challenges.

Martínez et al. [10] developed an automated mission management system, the Missions and Task Register and Reporter (MTRR), intended to maximize the efficiency of autonomous underwater vehicles (AUVs) within the SWARMs project. Using a decentralized hierarchical control pattern, this system incorporates virtualization of the AUV’s planning capabilities, making it a crucial step toward integrating both modern and legacy systems into operations. AUVs have been enhanced in their ability to adapt to dynamic underwater environments and exogenous events that may disrupt pre-planned missions as a result of the MTRR system being validated through trials that have proven its effectiveness. However, the study reveals limitations in the scalability of the proposed system when managing extensive missions involving multiple AUVs. Interfacing with semantic repositories may result in a bottleneck in data processing speed. This identifies a need for further optimization to ensure seamless scalability and robustness in diverse operational scenarios.

The research by Miguelanez et al. [11] presents a semantic knowledge framework for Autonomous Underwater Vehicles (AUVs) that aims to enhance their situation awareness through a complex hierarchical ontology system. The framework demonstrates potential in controlled laboratory settings by improving mission flexibility and operational independence. However, its integration into the real world may face challenges. Ontological structures proposed for advanced AUVs may not be readily supported by current proprietary or legacy software architectures. As significant modifications to existing systems are typically resource-intensive and technically challenging, this adaptability gap and integration complexity represent major obstacles. Moreover, the experiments were limited to laboratory conditions, simplifying environmental variables while failing to capture the full range of unpredictable natural underwater settings. As a result of these methodological restrictions as well as their narrow scope, more modular ontologies are urgently needed, which can be easily adjusted to meet different needs. In order to validate how well this approach works under various dynamic real-world conditions, extensive field trials must be conducted.

Yao et al. [12] aimed to increase situation awareness in unmanned underwater vehicles (UUVs) by combining an ontology-based system with Bayesian networks to facilitate decision-making under uncertainty, especially in the maritime environment. It has proven effective at integrating and processing different types of relevant information while also appraising threats when tested under controlled conditions. However, a major drawback is its ability to be scaled up for use in larger, more complicated operational areas. There is no clear information about how well it would perform if applied within diverse and unpredictable marine settings where real-world missions occur. These environments often contain massive amounts of data flowing from various sources, which are not similar to anything found in a laboratory setting. Therefore, further studies are needed to determine if such systems are robust enough for broad maritime operations and to evaluate their efficiency for broader applications.

Li et al. [13] developed a mixed reasoning model that uses ontology, rule-based logic, and multi-entity Bayesian Network (MEBN) reasoning to enhance the decision-making ability of underwater robots under uncertain conditions. The goal of this new method is to make sure different kinds of AUVs and ROVs, as well as other autonomous systems, work together efficiently while also helping them understand each other’s ‘language’ more easily when in complex environments like those found in the ocean. However, despite its potential for effectively addressing intricate problems, one major methodological drawback lies in how complex it is and the amount of computational power required for running real-time operations. This means that such a high level of intricacy could slow down responsiveness times vital during dynamic tasking in underwater scenarios where quick decisions must be made constantly. Furthermore, bringing many logical systems into one may complicate things even further, making them hard to set up, maintain, and scale for use across diverse operations. Hence, there is still a need to further simplify and optimize these frameworks so they can function effectively in real-world situations; thus increasing overall capability levels of autonomy among unmanned underwater vehicles operating within complex marine environments.

Tenorth et al. [14] elaborated knowledge-processing techniques necessary for robots to effectively perform tasks based on abstract instructions. These activities include identifying missing data and filling information gaps, reasoning about relevant sources of truth, and integrating different data types into consistent knowledge bases. The KnowRob system was introduced as a knowledge processing framework designed to reduce the semantic gap between robot control systems and the abstract instructions common in AI applications. Despite its complexity, which facilitates advanced inference mechanisms and data integration necessary for conducting complex object manipulation tasks, this system may not see widespread deployment due to its high computational demands, which can hinder fast decision-making in a time-critical environment.

Toberg et al. [15] discussed the importance of commonsense knowledge for enhancing cognitive robotics. The study reviewed applications of commonsense knowledge to robot functionality in real-world environments, with an emphasis on generalization to unfamiliar settings. To examine how commonsense supported robots’ interpretation of and response to dynamic situations, the authors systematically analyzed 52 studies, focusing on household and service robots. Despite these insights, the reviewed systems often did not capture the full complexity of real-world dynamics. Moreover, the integration of commonsense knowledge into cognitive robotics remained emergent, with challenges related to terminological consistency across research communities. Overall, the review highlighted the need for further development to ensure cognitive robots could robustly understand and navigate complex, real-world situations.

The paper by Bouguerra et al. [16] presents an approach to enhance robot plan execution monitoring reliability using semantic knowledge to derive implicit expectations about the effects of actions. This method addresses the gap between simple instructions and the complex data robots need to execute these instructions. The technique allows robots to identify discrepancies between expected and observed outcomes, significantly improving plan execution reliability. Two implementations of this approach are demonstrated: one that assumes deterministic actions and reliable sensing, and another that accommodates uncertainties in action effects and sensing. The validation includes simulations and real robot experiments, showcasing the potential of semantic knowledge in improving execution monitoring. However, the complexity and computational demands of the system could limit its deployment in dynamic environments where quick decision-making is critical. Further optimization and simplification are suggested to enhance its practical application across varied operational settings.

The study by Thomsen et al. [17] explores the integration of sensor data with natural language using ontologies to enhance data interpretation in complex scenarios. The research is centered on improving the utility of sensor data for decision-making processes, in particular by developing an ontology-based framework. This system consolidates diverse data forms into a coherent framework, enhancing environmental input interpretation. Even though the study uses an innovative approach, it shows a limitation: the ontology struggles to integrate linguistic information, limiting its ability to offer a comprehensive interpretation of the data. Additionally, while the framework demonstrates potential in controlled environments, its scalability and effectiveness in diverse, real-world settings remain untested. This points to a need for further development to ensure the system’s robustness and adaptability to operational scenarios that require high levels of data fusion and flexibility.

Yin et al. [18] focused on developing a situation reasoning module for autonomous underwater vehicles (AUVs) operating under uncertain conditions. This module aims to enhance AUV decision-making capabilities by improving their ability to assess and respond to dynamic environments and unforeseen events. Their study highlights the integration of ambiguous object detection, recognition, and an ontology model to effectively manage and utilize information processed during AUV missions. The implementation is demonstrated through simulations and marine experiments, which confirm the module’s effectiveness in improving mission execution and response strategies under uncertain conditions. However, the study acknowledges a limitation in its approach, particularly in the scaling and adaptation of the system to more complex or varied operational areas beyond the tested scenarios. This points to a need for ongoing development to ensure the robustness and flexibility of the system for broader applications in diverse marine settings.

Chen et al. [19] proposed a Practical Finite-Time Observer-Based Adaptive Backstepping Super-Twisting Sliding Mode Controller to enhance the motion stability and trajectory tracking accuracy of deep-sea hydraulic manipulators operating under uncertain and disturbed conditions. Their work focuses on improving motion-level control and robustness through adaptive observation and finite-time convergence strategies. While this approach effectively mitigates dynamic disturbances, it primarily addresses control precision at the actuator level and does not consider semantic coordination or higher-level data interpretation. This distinction highlights the complementary role of ROVON, which focuses on integrating heterogeneous sensory and contextual data to support reasoning-driven decision-making and interoperability across underwater robotic systems.

Chen et al. [20] proposed a parameter identification framework for open-frame underwater vehicles using numerical simulation combined with a quantum particle swarm optimization (QPSO) algorithm. Their approach integrates the Dynamic Fluid–Body Interaction (DFBI) model in STAR-CCM+ to accurately estimate both inertial and hydrodynamic resistance parameters under four degrees of freedom. Experimental validation demonstrated a mean square error below 0.2%, outperforming traditional optimization methods such as PSO and GA. Although this work significantly improves model precision and computational efficiency for dynamic modeling, it remains primarily focused on hydrodynamic identification and controller design rather than higher-level knowledge integration or reasoning. ROVON complements such developments by addressing semantic interoperability and information representation, enabling data from vehicle models, sensors, and mission contexts to be meaningfully interpreted and shared across underwater robotic systems.

In conclusion, this literature review has illustrated significant advancements and persistent challenges in the use of underwater robotics ontologies. Despite recent advances in the field, there are limited formal ontologies available that are designed at the outset for data integration in underwater robotic applications. ROVON, as a formal ontology that has been specifically designed for underwater robots, aims to integrate and harmonize diverse data from images, sensors, and robotic operations within underwater environments. This integration is crucial for enhancing underwater vehicles’ operational efficiency and interoperability due to generating high-quality datasets that use standardized semantics and vocabulary. ROVON fills the gap in the literature by providing a means for formalizing the knowledge in the domain of underwater robotics with the objective of enhancing communication and interoperability between human operators and robotic systems. Such capabilities are vital for addressing diverse operational challenges and improving marine activities’ sustainability and economic viability. By filling critical gaps in current research, ROVON seeks to advance autonomous underwater systems and establish more reliable maintenance protocols, ultimately contributing to safer and more effective marine operations.

3. Ontology and Knowledge Graph

An ontology provides a formal representation of the concepts in a domain and the relationships among them. Ontologies support interoperability by offering a unifying semantic framework for aligning and harmonizing disparate datasets. By using a common vocabulary and concept scheme, ontologies enable streamlined data exchange among software systems. Owing to their logic-based formalism, they also support automated reasoning, data consistency checking, and the resolution of semantic ambiguities [21]. An ontology typically consists of the following components:

• Classes: The primary concepts or entities within a domain, described at varying levels of detail.
• Relations: The types of associations or connections that can exist among those classes [22].
• Formal Axioms: Logic-based constraints on classes and properties that ensure consistency and coherence across the ontology [23].

Most formal ontologies use the Web Ontology Language (OWL). OWL builds on the Resource Description Framework (RDF) and RDF Schema (RDFS) to provide a description logic–based framework for defining classes, properties, and axioms, enabling automated reasoning, consistency checking, and advanced ontology validation.

When datasets are aligned with ontologies, they form a data structure known as a knowledge graph, in which real-world entities (ontological instances) are represented as nodes and the semantic relationships between them as labeled edges. This graph-based model enables rich, interconnected representations of information that support advanced querying, inference, and visualization across heterogeneous data sources. Knowledge graphs are commonly represented either as RDF-based triples or as property graphs. While RDF organizes data in triples of subject, predicate, and object, property graph models store attributes as key–value pairs linked to each node and edge [23]. The knowledge graphs generated in this work use the RDF triple structure.

A key advantage of RDF knowledge graphs over property graphs is their built-in, standards-based semantics, which enable automatic inference and interoperability. Because RDF triples use globally unique URIs (Uniform Resource Identifiers) and a formally defined data model (RDF Schema and OWL), standards-compliant RDF stores and OWL reasoners can automatically derive new facts (via reasoning), validate consistency, and integrate disparate datasets without custom code capabilities that property graph models, which lack a unified semantic standard, must implement manually [24].

4. ROVON Ontology

ROVON is developed to provide a formal representation of the domain of underwater robotics. Similar to any other ontology, ROVON consists of a set of entities (classes) and relationships (properties) between those entities. Web Ontology Language (OWL) is the ontology modeling language used to represent ontological classes and relationships. ROVON can support knowledge-intensive applications in the underwater robotics domain in the following ways:

• Standardization: ROVON, as an open-source ontology, provides a standardized vocabulary with explicit semantics. This would be used to ensure a shared understanding of various concepts across stakeholders in an ROV application and use cases.
• Data Integrity: ROVON can be used to specify the completeness and consistency of data that must be present (interpreting ontologies as Integrity Constraints) for the integrity of the traceability system.
• Semantic Integration and Mediation: ROVON can be used as a global model for the data. ROVON can also act as a global model for querying data over heterogeneous systems and for integrating the results.
• Symbolic Reasoning for History Exploration, Discovery and Construction: ROVON could be used to support data exploration and “what if” queries to discover important relationships and fill in the missing information. This utility is particularly important for underwater environments where low-quality data and missing data are common.

ROVON is developed according to the Industrial Ontologies Foundry (IOF) technical principles and guidelines [25]. The IOF is an international community of government, industry, and academia that was formed with the vision of increasing the adoption of ontologies in the manufacturing sector. The IOF ontologies use BFO (Basic Formal Ontology) [26] as their top-level ontology (TLO). BFO has been widely used in the biological domain for integrating disparate ontologies or data models and developing interoperable ontologies for biological applications. The IOF ontologies also use a common mid-level ontology named IOF Core. The IOF Core ontology consists of generic constructs and patterns that are reusable across domain-level industrial ontologies. Using a TLO in conjunction with a mid-level ontology streamlines the ontology development process through providing pre-validated design patterns and also results in designing more reusable and principle-based ontologies. BFO categorizes entities under two broad classes, namely, continuant and occurrents. The next sections discuss various ROVON constructs under these two categories.

Ontology requirements are typically established at the outset of development by defining a set of Competency Questions (CQs)—the specific queries the ontology must be able to answer. Since ROVON is designed primarily for underwater pipeline inspection, its CQs are crafted to reflect the domain’s key concerns, such as detecting pipeline defects, identifying the root cause of defects, and correlating sensor readings with pipeline features. Below are some example CQs for ROVON:

• What was the orientation of the robot when image XYZ was taken?
• Which water portions have pH values above 8.0 and/or sodium concentrations outside the safe range?
• Which images depict five or more hole features, indicating a severe structural defect?
• Which pipe segments are inferred to have mixed or compound defects (e.g., both good and bad weld annotations, or co-occurring crack and corrosion features)?

4.1. ROVON Continuants

In BFO, continuants are entities that persist through time while maintaining their identity, regardless of changes in their internal states or external conditions. Continuants are essential for representing stable, enduring entities. This is in contrast to processes or events, which occur over time and persist differently. The hierarchy of the “continuant” class is shown in Figure 1.

4.1.1. Independent Continuants

Independent continuants do not depend on other entities for their existence. Examples of independent continuants in ROVON include objects such as the robot and its components such as the camera and the sensors. In ROVON, the class Underwater Robot is a sub-class of Robot, which in turn is a sub-class of IOF:Material Artifcat. A material artifact is defined as an object that is deliberately created to have a certain function. An ROV has multiple functions that have to be modeled in the ontology. An underwater robot can have several components, such as a body, propeller, motor, and camera, that are material artifacts themselves and are connected to one another through parthood relationships. Another example of a continuant in ROVON is the Portion of Water class. The reason portion of water is a key class in ROVON is that the operating environment of the robot is filled with water, and therefore, it is imperative to capture the data about different parameters (or qualities) of the water portion, including temperature, acidity, and hardness, as they provide contextual metadata for various robotic explorations. The portion of Water is a sub-class of BFO: Material Entity, which is defined as an independent continuant that, at all times at which it exists, has some portion of matter as a continuant part. Water parameters depend on the portion of water for their existence, and therefore, they are referred to as Specifically Dependent Continuants in the BFO. An immaterial entity refers to a type of entity that does not have a material form or physical substance. For example, spatial regions, such as cube-shaped regions of space, are a type of immaterial entity. BFO:Site is an important sub-class of the immaterial entity whose boundaries either, partially or wholly, coincide with the boundaries of one or more material entities or have locations determined in relation to some material entity. For example, the interior or a trunk of a car is a type of site bounded by the walls of the trunk. The operation area of an underwater robot is a type of site with boundaries located at a certain distance from the center of mass of the robot. The operational area of the underwater robot contains some portion of water, and it is the location in which robotic processes take place. Figure 2 shows the taxonomy of some of the critical continuant classes in ROVON. The hierarchy of the “Independent continuants” class is shown in Figure 3.

4.1.2. Dependents Continuants

A dependent continuant is an entity that depends on other independent continuants for its existence. For example, the weight or size of an ROV are the properties (qualities) that cannot exist on their own and depend on the robot for their existence. If an entity depends on a specific continuant, then the entity is a type of specifically dependant continuant or SDC. Some specifically dependent continuants need other processes for their realization. Role and Function are two types of realizable entities in BFO. For example, the imaging function of a robot is realized during the imaging process, or the role of a robot as an inspector is realized during the inspection of an underwater pipeline. Figure 4 shows the class structure for SDC in ROVON. Generically dependent continuant or GDC typically refers to a type of entity that exists dependently on multiple instances of other entities, and not necessarily any specific one. A common example could be a specific type of information or pattern that could be replicated across multiple instances—such as a digital image or a software program—which remains the same across different physical instances. Essentially, all information entities are regarded as GDCs. Measurement Information Content Entity (ICE) is an IOF class that is defined as an informational content that is the result of measuring a set of attributes (specifically dependent continuant or process characteristic or temporal region) belonging to the entity (independent continuant or process or process boundary) the informational content is about. All underwater measurement results, such as results of measuring the current depth or water density, are instances of Measurement ICE. Figure 5 shows how the roll orientation measurement result of an ROV is represented as an instance of the Angle Measurement ICE. Note that Roll Orientation in Figure 5 is a sub-class of Spatial Orientation, which is a type of Relational Quality. Other types of Spatial Orientation classes in ROVON include Pitch Orientation and Yaw Orientation. Definitions for some relevant continuants are provided in Table 1. The hierarchy of the “Dependents continuants” class is shown in Figure 6.

Table 1. Natural language definition for some continuants.

Class	Natural Language Definition
Generically Dependent Continuant	A generically dependent continuant is an entity that exists in virtue of the fact that there is at least one of what may be multiple copies; it is the content or the pattern that the multiple copies share
Realizable Entity	A realizable entity is a specifically dependent continuant that inheres in some independent continuant which is not a spatial region and is of a type some instances of which are realized in processes of a correlated type
Spatial Orientation	A relational quality that is the angle of rotation of an object relative to one or more plane of reference or axis of rotation

Table 1. Natural language definition for some continuants.

Class	Natural Language Definition
Generically Dependent Continuant	A generically dependent continuant is an entity that exists in virtue of the fact that there is at least one of what may be multiple copies; it is the content or the pattern that the multiple copies share
Realizable Entity	A realizable entity is a specifically dependent continuant that inheres in some independent continuant which is not a spatial region and is of a type some instances of which are realized in processes of a correlated type
Spatial Orientation	A relational quality that is the angle of rotation of an object relative to one or more plane of reference or axis of rotation

Figure 4. Examples of specifically dependent continuants.

Figure 4. Examples of specifically dependent continuants.

Figure 5. Roll orientation of the ROV is represented as a type of relational quality in the ROV.

Figure 5. Roll orientation of the ROV is represented as a type of relational quality in the ROV.

Figure 6. Snapshot of the hierarchy of the Dependents continuants.

Figure 6. Snapshot of the hierarchy of the Dependents continuants.

4.2. ROVON Occurrents

In the ROVON ontology, occurrents are defined within the Basic Formal Ontology framework as entities that unfold temporally and include a variety of processes and events. During underwater missions, these phenomena are particularly important for documenting ROV operations. As shown in Figure 7 illustrates several critical classes of occurrents, including Imaging, Observation, and Planned Processes. To deploy ROVs in a structured and effective manner, these classes play a pivotal role.

As shown in Figure 8, ROVs participate in a planned observation process, identified under IOF-Core: Planned Process. This type of process occurs at an underwater operation site and is timed according to BFO: Temporal Interval. Planned Processes, as described in Table 2, follow predetermined procedures to ensure that the tasks carried out by the ROV are both predictable and efficient, thereby ensuring the safety of the operations.

In Table 2, specific processes within the ROVON framework are further explained, including the Imaging and Observation Processes, which are planned activities that are designed to create visual representations of objects or record observations about materials and equipment. In challenging underwater environments, this structured approach ensures effective data collection and ensures the equipment’s operational integrity and safety. In addition, as illustrated in Figure 8, the integration of these processes within the ROVON ontology provides a comprehensive framework for querying and analyzing complex data.

Figure 7. Class hierarchy of occurrents and continuants in the ROVON ontology.

Figure 7. Class hierarchy of occurrents and continuants in the ROVON ontology.

Table 2. Natural language definitions for some of the key classes in ROVON.

Class	Natural Language Definition
Planned Process	A process that is prescribed by a plan specification
Imaging Process	A planned process of creating a visual representation of an object
Observation Process	A planned process where some measurement or other observation is noted and recorded that relates (directly or indirectly) to some material or equipment

Table 2. Natural language definitions for some of the key classes in ROVON.

Class	Natural Language Definition
Planned Process	A process that is prescribed by a plan specification
Imaging Process	A planned process of creating a visual representation of an object
Observation Process	A planned process where some measurement or other observation is noted and recorded that relates (directly or indirectly) to some material or equipment

Figure 8. The general pattern for participation of ROVs in an observation process.

Figure 8. The general pattern for participation of ROVs in an observation process.

5. Application of ROVON in Pipeline Monitoring

Pipeline monitoring is the process of continuous observation and analysis of pipelines that are essential to infrastructure systems, such as those transporting oil, gas, and telecommunication. Monitoring pipelines is intended to maintain the integrity and safety of these systems by identifying issues such as leaks, corrosion, structural damage, or environmental changes that could disrupt operations. Underwater pipelines face particularly difficult conditions due to high pressures, corrosion from seawater, and the challenges of accessibility to manual inspection. To address these complexities, the use of ontologies in pipeline monitoring helps standardize and structure the vast and diverse data collected from ROVs, sensors, and other monitoring equipment. Ontologies provide a formal framework for categorizing and interrelating various data types, including pressure measurements, defect classifications, and environmental parameters, allowing for more seamless integration and analysis of the information. Using semantic reasoning to analyze these datasets, ontologies facilitate the detection of defects, enable predictive maintenance, and support proactive decision-making, ultimately enhancing the safety and operational performance of pipelines. Based on this, semantic knowledge graphs are powerful tools for representing and uncovering relationships between entities in specific domains. Unlike traditional data structures, semantic graphs dynamically connect entities through nodes and edges, creating a more intuitive and graphical way of visualizing complex relationships. This structure combines the strengths of flexibility and computational efficiency, making it particularly effective for challenging tasks like underwater pipeline monitoring. By linking spatial, temporal, and semantic data into a compact graph format, entities such as geospatial objects or pipeline components can be better organized based on their spatial-temporal relationships. This not only enhances data management but also improves the system’s ability to traverse, analyze, and infer new insights from the knowledge domain. The interpretation of semantic graphs, involving the extraction of meaningful information from observed data like sensor readings and images, plays a critical role in identifying potential pipeline issues and environmental changes, leading to more proactive and informed decision-making [27].

Managing complex and diverse data generated in underwater environments is beyond simple data gathering and object detection. The integration of technologies like Ontotext/RefineOntotext Refine (Ontotext AD, Sofia, Bulgaria) and GraphDB (GraphDB (Ontotext AD, Sofia, Bulgaria) is crucial for addressing these complexities by creating dynamic and scalable knowledge graphs. These tools transform various datasets—such as sensor readings, images captured by ROVs, and other metadata—into a cohesive, searchable database that enables advanced analytical functions. With this system, operators can perform tasks like pattern recognition and predictive analytics, synthesizing disparate data sources into a single accessible framework. This approach significantly enhances decision-making, operational accuracy, and safety, especially in the unpredictable conditions of underwater exploration. The following sections will delve into how these tools are applied to construct robust knowledge graphs that empower underwater robotics projects with improved data synthesis and insights.

5.1. Multi-Source Data Collection

Integrating advanced technologies such as the YOLOv8 algorithm is crucial in the domain of underwater pipeline inspection, where environmental challenges like light scattering, absorption, limited visibility, and ambient noise complicate detection. This state-of-the-art, real-time object detection system, utilizing a convolutional neural network (CNN) architecture, is instrumental in overcoming these issues. In murky underwater environments, YOLOv8 significantly enhances detection and classification capabilities, making it highly effective for identifying various pipeline defects [28]. To accurately detect and classify a range of pipeline defects, we collected a specialized dataset using ROV(FIFISH V6 (QYSEA Technology Co., Shenzhen, China)) in a controlled underwater setting. Although the dataset represents a controlled-scale environment, it effectively captures the multimodal characteristics of real underwater data streams (e.g., images, sensor readings, and environmental metadata), allowing the ontology and reasoning framework to be evaluated under realistic yet manageable conditions. Figure 9 shows the classification of various defect types, including issues like bad welds, cracks, rust, and holes. Additionally, Figure 10 presents the dataset enriched with metadata such as temperature, depth, orientation, and chemical properties (e.g., sodium, iron, and chloride levels) of the water, ensuring a comprehensive dataset. This allows for detailed analysis, offering valuable insights into defect types and their environmental context. The dataset was generated internally to meet the need for high-quality images and a diverse range of defect types [29]. It contains 2467 underwater pipeline images collected using an ROV and was manually annotated by the authors to ensure consistent labeling across defect categories.

5.2. Semantic Mapping

Ontotext/Refine is a powerful tool designed to integrate and transform complex datasets for efficient knowledge graph creation. It enables users to preprocess and convert diverse data sources, such as CSV files or live data feeds, into structured formats optimized for building knowledge graphs. In the ROVON project, Ontotext/Refine is used to modify and prepare datasets with underwater operational information, including images, sensor readings, and defect data on underwater structures. Detailed attributes from these datasets, including such as defect types and counts, are meticulously mapped to ensure the data are ready for in-depth analysis. This involves defining a clear schema and establishing relationships between entities like components of underwater equipment and sensor readings, aligning the data with ROVON’s ontology. For this study, Ontotext Refine was applied in a batch-processing mode to generate RDF triples consistently across the collected dataset. This controlled approach ensured accurate semantic alignment and reproducibility during experimentation. The same mapping process can later be automated via APIs for streaming or real-time data ingestion in operational deployments, allowing for continuous updates without manual intervention. By ensuring accurate data integration, this approach enhances data accessibility, usability, and precision, ultimately supporting robust decision-making processes in underwater operations [30]. The key component of this integration is RDF (Resource Description Framework) mapping, which structures relationships among data points through subjects, predicates, and objects. This method allows for the integration of RDF properties and classes, using GREL (General Refine Expression Language) expressions to dynamically generate unique identifiers. These identifiers ensure each data entry, such as an image or a sensor reading, has a distinct presence within the knowledge graph. For example, orientation parameters like “roll” are mapped using GREL expressions to create unique URIs, which are then linked to the ontology, making the data more accessible for detailed queries and analysis. Figure 11 illustrates the RDF mapping process for “roll-quality,” demonstrating how GREL-generated identifiers are associated with measurement details using predicates like ‘iof-core’, thereby enriching the data’s semantic context. In addition to orientation data, this RDF mapping approach is also applied to defect mapping, where each defect is uniquely identified and linked to related components of the ROV’s system. As shown in Figure 12, specific fields like ID, Pipe values, and defect type are combined using GREL expressions, creating a structured framework under the class rov:PipeFeature. This mapping not only connects data points to the ontology’s semantic framework but also enhances the overall utility of the knowledge graph, making it an effective tool for complex underwater operations.

The structured and semantically enhanced data is prepared for export after the RDF mapping has been completed, enabling it to be further analyzed and integrated. Users can export the resulting knowledge graph in formats like JSON and TTL (Turtle syntax), each offering unique benefits for various applications. The TTL format is especially suited for interoperability with semantic web tools, allowing for seamless data integration, enhanced querying, and visualization capabilities. This flexibility ensures that the data can be readily adapted for a range of uses, from in-depth analytical tasks to interactive visualizations using platforms like GraphDB or other RDF-compatible systems. Ontotext/Refine supports multiple export formats, enabling advanced analysis and ensuring the long-term utility of semantically rich datasets across a wide range of tools and environments.

Figure 11. RDF mapping for the orientation of the roll.

Figure 11. RDF mapping for the orientation of the roll.

Figure 12. RDF mapping of the pipe defect.

Figure 12. RDF mapping of the pipe defect.

5.3. Visualization of Semantic Data

Visualization is a crucial step in understanding complex relationships within semantic data, especially when dealing with intricate datasets like those used in underwater operations. It enables users to perceive connections between entities, making it easier to uncover insights and interpret patterns that might be hidden in raw data. This is particularly valuable for knowledge graphs, where entities are interconnected through relationships that can be difficult to understand without a visual representation. By presenting data visually, analysts can communicate complex information more effectively, aiding both decision-making and the discovery of new insights. GraphDB serves as a robust platform for managing, querying, and visualizing RDF (Resource Description Framework) datasets. It offers tools for storing ontologies and executing complex queries using SPARQL, the standard query language for RDF [31]. Figure 13 illustrates a SPARQL query that retrieves all unique yaw orientation measurements along with their values, offering a focused view of this data. Figure 14 visually depicts the semantic relationships between roll measurement entities and their specific values. This combination of advanced querying and visualization capabilities makes GraphDB an effective tool for analyzing semantic data. It not only enhances the understanding of complex data relationships but also allows for practical applications of semantic knowledge in fields like underwater operations, where precision and clarity are crucial.

This pipeline-monitoring scenario serves as a validation of ROVON’s semantic integration process rather than only a demonstration. By transforming image detections, sensor measurements, and environmental metadata into a unified RDF knowledge graph, ROVON confirmed its ability to preserve semantic consistency across heterogeneous sources. The accurate alignment between defect annotations and contextual parameters such as depth, temperature, and chemical composition shows that the ontology enables coherent, reasoning-ready data for subsequent inference. This outcome validates ROVON’s effectiveness as an operational framework for underwater inspection analysis.

Figure 13. SPARQL query interface for retrieving yaw measurements.

Figure 13. SPARQL query interface for retrieving yaw measurements.

Figure 14. Visualization of roll measurement relationships for SING0007.

Figure 14. Visualization of roll measurement relationships for SING0007.

6. Semantic Reasoning

Semantic reasoning refers to the process of drawing new, implicit knowledge from explicitly modeled information by leveraging the formal definitions, relationships, and constraints encoded in an ontology. Through the semantic integration of the observation and inspection data, ROVON supports pattern recognition and cause-effect analysis, mainly enabled by rule-based reasoning. The semantic reasoner in ROVON operates on RDFox, an in-memory high-performance RDF triple store and semantic reasoning engine. RDFox supports OWL ontologies and executes Datalog rules natively, allowing for efficient handling of large-scale data while ensuring logical consistency [32]. Within ROVON, RDFox performs constraint enforcement, fact generation, and classification by combining quantitative thresholds with structural relationships [33]. All ontological knowledge, together with incoming instance data, is stored in the RDF repository, where Datalog rules are executed to generate new inferences such as environmental risk conditions, defect severity classifications, and functional roles of robotic agents.

Three types of semantic reasoning are implemented in this work for validation purposes:

• Quantitative reasoning: This reasoning evaluates sensor measurements and imaging results by comparing them against established threshold values. It focuses on quantifiable indicators such as pH levels, sodium concentrations, and defect counts. In ROVON, quantitative reasoning is used to assess water safety by monitoring chemical values that trigger sample classifications as unsafe, and to automatically flag images as abnormal when hole counts exceed predefined thresholds. By relying on explicit numerical comparisons, the system systematically identifies environmental and structural risks.
• Qualitative reasoning: This approach emphasizes relationships between entities rather than numerical values. Instead of thresholds, the system leverages the ontology framework to classify conditions based on structural logic. For instance, when a single pipe segment is annotated with both rov:GoodWeld and rov:BadWeld, ROVON infers that it belongs to the rov:MixedWeldQualityPipe class, signaling a weld evaluation discrepancy. Similarly, when both cracks and corrosion are detected on the same pipe, the system categorizes it as rov:CorrodedCrackPipe. This reasoning method improves transparency in the knowledge graph by explicitly highlighting conflicting assessments and combined defect states.
• Hybrid reasoning: ROVON employs hybrid reasoning when neither quantitative nor qualitative methods alone are sufficient. This approach combines numerical thresholds with structural and relational context to generate more comprehensive inferences. Hybrid reasoning activates when quantitative assessments of chemical values align with qualitative analyses of feature relationships, ensuring both dimensions converge. In practice, this method produces unified inferences that connect environmental conditions with defect manifestations, enabling a deeper understanding of pipeline behavior. By integrating water chemistry measurements with defect spatial patterns, hybrid reasoning supports root cause analysis and the detection of compound risk scenarios. It also enhances inspection target prioritization and maintenance planning by highlighting interactions between adverse environments and structural weaknesses rather than evaluating them in isolation.

Through these reasoning methods, ROVON provides automated defect severity classification, environmental risk detection, root cause analysis, and robotic role inference. In this context, role inference refers to the system’s ability to deduce the operational function of a robotic unit from its recorded activities and relationships in the ontology, rather than relying solely on predefined labels. By enriching the knowledge graph with such implicit knowledge, semantic rules improve the interpretability of collected data, even when these insights are not explicitly available in the original dataset. This reasoning capability is particularly important in underwater environments, where human supervision is limited and data-driven decision-making is essential [34].

Figure 15 shows the complete structure of ROVON’s reasoning framework by displaying the system’s logic through three rule-based reasoning pipelines: quantitative, qualitative, and hybrid.

6.1. Quantitative Reasoning

The quantitative reasoning function of ROVON applies numerical thresholds to sensor and imaging data to classify environmental and structural conditions. Rules and queries operate as separate processes: rules automatically infer new classifications from numeric values, while queries are issued independently by users to retrieve or analyze results. These processes are not sequential—rules continuously enrich the knowledge graph, and queries provide access to both the raw data and the inferred knowledge.

This reasoning method is demonstrated through two examples: water classification based on pH measurements and defect feature aggregation from image analysis. In both cases, threshold comparisons over numeric values stored in the ontology generate new classifications that enhance the interpretability and usability of inspection data.

Example 1: High pH Water Classification

In this example, water samples are classified based on their pH levels. Figure 16 shows how the instance rov:SING0008-water-portion of type rov:PortionOfWater is connected to a pH quality measurement through the property core:hasQuality. The measurement record contains a value expression core:hasSimpleExpressionValue “8.4” representing a pH level of 8.4. According to the ontology, when a water portion’s pH exceeds the threshold of 8.0, it is automatically classified as rov:HighPhWater. The following Datalog rule formalizes this threshold-based classification logic (see Table 3 and Table 4):

Rule.

Table 3. Datalog Rule for High pH Water Classification.

PREFIX rov: <http://infoneer.txstate.edu/ROVON/>
PREFIX iof-core: <https://spec.industrialontologies.org/ontology/core/Core/>
PREFIX bfo: <http://purl.obolibrary.org/obo/>
PREFIX xsd: <http://www.w3.org/2001/XMLSchema#>
?[waterPortion, a, rov:HighPhWater] :-
[?waterPortion, a, rov:PortionOfWater],
[?waterPortion, iof-core:hasQuality, ?ph],
[?ph, iof-core:describedBy, ?phMeasureICE],
[?phMeasureICE, bfo:BFO_0000110, ?phValueSpec],
[?phValueSpec, iof-core:hasSimpleExpressionValue, ?phValue],
FILTER(xsd:decimal(?phValue) > 8.0).

Table 3. Datalog Rule for High pH Water Classification.

PREFIX rov: <http://infoneer.txstate.edu/ROVON/>
PREFIX iof-core: <https://spec.industrialontologies.org/ontology/core/Core/>
PREFIX bfo: <http://purl.obolibrary.org/obo/>
PREFIX xsd: <http://www.w3.org/2001/XMLSchema#>
?[waterPortion, a, rov:HighPhWater] :-
[?waterPortion, a, rov:PortionOfWater],
[?waterPortion, iof-core:hasQuality, ?ph],
[?ph, iof-core:describedBy, ?phMeasureICE],
[?phMeasureICE, bfo:BFO_0000110, ?phValueSpec],
[?phValueSpec, iof-core:hasSimpleExpressionValue, ?phValue],
FILTER(xsd:decimal(?phValue) > 8.0).

The corresponding SPARQL query can be used to retrieve all water portions with pH values above 8.0 (thus qualifying as HighPhWater). This query traverses the same quality and measurement links and applies a numeric filter on the pH value:

SPARQL Query.

Table 4. SPARQL Query for Water Portions with High pH.

PREFIX rov: <http://infoneer.txstate.edu/ROVON/>
PREFIX iof-core: <https://spec.industrialontologies.org/ontology/core/Core/>
PREFIX bfo: <http://purl.obolibrary.org/obo/>
PREFIX xsd: <http://www.w3.org/2001/XMLSchema#>
SELECT ?waterPortion ?phValue
WHERE {
?waterPortion a rov:PortionOfWater ;
iof-core:hasQuality ?ph .
?ph iof-core:describedBy ?phMeasureICE .
?phMeasureICE bfo:BFO_0000110 ?phValueSpec .
?phValueSpec iof-core:hasSimpleExpressionValue ?phValue .
FILTER(xsd:decimal(?phValue) > 8.0)
}

Table 4. SPARQL Query for Water Portions with High pH.

PREFIX rov: <http://infoneer.txstate.edu/ROVON/>
PREFIX iof-core: <https://spec.industrialontologies.org/ontology/core/Core/>
PREFIX bfo: <http://purl.obolibrary.org/obo/>
PREFIX xsd: <http://www.w3.org/2001/XMLSchema#>
SELECT ?waterPortion ?phValue
WHERE {
?waterPortion a rov:PortionOfWater ;
iof-core:hasQuality ?ph .
?ph iof-core:describedBy ?phMeasureICE .
?phMeasureICE bfo:BFO_0000110 ?phValueSpec .
?phValueSpec iof-core:hasSimpleExpressionValue ?phValue .
FILTER(xsd:decimal(?phValue) > 8.0)
}

Figure 16. RDF graph visualizing high pH water classification.

Figure 16. RDF graph visualizing high pH water classification.

Example 2: Hole Count Detection from Image

In this example, the system calculates the number of hole features associated with each inspection image. Every image is linked to the corresponding pipeline segment, and the hole features on that segment are represented within the ontology. The Datalog rule shown below applies an aggregate COUNT function to determine the total number of holes for each image. The resulting values enable the identification of images that exhibit unusually high numbers of holes and facilitate comparisons of defect levels across different pipeline sections. This quantitative measure provides a straightforward means of assessing the extent of surface damage and supports more informed inspection planning (see Table 5).

Rule.

Table 5. RDFox Datalog rule classifying images with at least five holes.

PREFIX rov: <http://infoneer.txstate.edu/ROVON/>
PREFIX iof-core: <https://spec.industrialontologies.org/ontology/core/Core/>
PREFIX bfo: <http://purl.obolibrary.org/obo/>
PREFIX xsd: <http://www.w3.org/2001/XMLSchema#>
[?image, a, rov:SevereDefectImage] :-
[?image, a, rov:Image],
[?image, iof-core:isAbout, ?obj],
AGGREGATE(
[?hole, a, rov:Hole],
[?hole, rov:featureOf, ?obj]
ON ?obj BIND COUNT(?hole) AS ?holeCount
),
FILTER(?holeCount >= 5) .

Table 5. RDFox Datalog rule classifying images with at least five holes.

PREFIX rov: <http://infoneer.txstate.edu/ROVON/>
PREFIX iof-core: <https://spec.industrialontologies.org/ontology/core/Core/>
PREFIX bfo: <http://purl.obolibrary.org/obo/>
PREFIX xsd: <http://www.w3.org/2001/XMLSchema#>
[?image, a, rov:SevereDefectImage] :-
[?image, a, rov:Image],
[?image, iof-core:isAbout, ?obj],
AGGREGATE(
[?hole, a, rov:Hole],
[?hole, rov:featureOf, ?obj]
ON ?obj BIND COUNT(?hole) AS ?holeCount
),
FILTER(?holeCount >= 5) .

SPARQL Query.

SPARQL query

Table 6. SPARQL query retrieving all images depicting an object with five or more holes.

PREFIX rov: <http://infoneer.txstate.edu/ROVON/>
PREFIX iof-core: <https://spec.industrialontologies.org/ontology/core/Core/>
PREFIX bfo: <http://purl.obolibrary.org/obo/>
PREFIX xsd: <http://www.w3.org/2001/XMLSchema#>
SELECT ?image
WHERE {
?image a rov:Image ;
iof-core:isAbout ?object .
?hole a rov:Hole ;
rov:featureOf ?object .
}
GROUP BY ?image
HAVING (COUNT(DISTINCT ?hole) >= 5)

results generate a structured list that identifies dataset images containing numerous hole features and counts their individual holes.

Table 7. Result of all images with five or more distinct hole features.

Image	Hole Count
rov:SING0035-image	5
rov:SING00112-image	6
rov:SING00206-image	7

displays all images having five or more distinct hole features together with their corresponding hole counts. The summary provides an efficient way to locate images containing numerous holes, thus drawing attention to objects with complicated perforation patterns that need additional investigation.

6.2. Qualitative Reasoning

The ROVON framework applies qualitative reasoning through its ontology structure to infer roles, classifications, and logical dependencies among entities, without relying on numerical input. This approach is especially valuable in underwater infrastructure environments, where data is often incomplete or uncertain. Instead of numerical measurements, the system formalizes patterns that describe relationships and structural context. For instance, the system uses qualitative logic to evaluate pipes with multiple defect features, such as cracks and corrosion, and classify them as compound faults without depending on sensor-based limits. These rule-based inferences resemble the state-based reasoning found in qualitative modeling, where the system transitions between structural states based on symbolic criteria rather than numeric values. This methodology allows domain expertise to be translated into computable rules, enabling meaningful insights in scenarios where quantitative data may be missing or insufficient [35].

6.2.1. Case Study 1: Weld Inconsistency

Scenario. Underwater pipeline inspection requires accurate weld quality assessment to evaluate structural reliability and ensure safety. A single pipe segment examined through visual or image-based analysis may receive annotations for both rov:GoodWeld and rov:BadWeld features, as overlapping observations from different tools, annotators, or inspection times can lead to conflicting interpretations. These inconsistencies introduce logical contradictions that cannot be resolved through quantitative thresholds alone. The system applies semantic reasoning to detect and classify such inconsistencies. The ROVON ontology handles this scenario by introducing the class rov:MixedWeldQualityPipe, which is inferred when a pipe is associated with both good and bad weld annotations via the rov:hasFeature property. This classification does not imply the pipe is either safe or defective; rather, it flags the weld evaluation as inconsistent, prompting further review. By explicitly representing this condition in the knowledge graph, the system supports more transparent decision-making and facilitates targeted follow-up inspections where ambiguity exists (see Table 8).

Rule.

Table 8. RDFox rule identifying pipes with both GoodWeld and BadWeld.

PREFIX rov: <http://infoneer.txstate.edu/ROVON/>
PREFIX iof-core: <https://spec.industrialontologies.org/ontology/core/Core/>
PREFIX bfo: <http://purl.obolibrary.org/obo/>
PREFIX xsd: <http://www.w3.org/2001/XMLSchema#>
[?pipe, rdf:type, rov:MixedWeldQualityPipe] :-
[?goodWeld, rdf:type, rov:GoodWeld],
[?badWeld, rdf:type, rov:BadWeld],
[?goodWeld, rov:featureOf, ?pipe],
[?badWeld, rov:featureOf, ?pipe],
FILTER(?goodWeld != ?badWeld).

Table 8. RDFox rule identifying pipes with both GoodWeld and BadWeld.

PREFIX rov: <http://infoneer.txstate.edu/ROVON/>
PREFIX iof-core: <https://spec.industrialontologies.org/ontology/core/Core/>
PREFIX bfo: <http://purl.obolibrary.org/obo/>
PREFIX xsd: <http://www.w3.org/2001/XMLSchema#>
[?pipe, rdf:type, rov:MixedWeldQualityPipe] :-
[?goodWeld, rdf:type, rov:GoodWeld],
[?badWeld, rdf:type, rov:BadWeld],
[?goodWeld, rov:featureOf, ?pipe],
[?badWeld, rov:featureOf, ?pipe],
FILTER(?goodWeld != ?badWeld).

Results. The rule allows the ontology to automatically identify discrepancies in weld evaluations. For instance, the system detects that the pipe instance rov:SING0012-pipe is associated with two weld features—one classified as a rov:GoodWeld and the other as a rov:BadWeld. As a result, it assigns the pipe to the class rov:MixedWeldQualityPipe. This classification does not imply any safety conclusion; instead, it flags the pipe for further review due to conflicting weld assessments. By applying qualitative reasoning to detect such contradictions, the system enhances the interpretability of the knowledge graph and supports quality control workflows that must handle inconsistent inspection annotations.

6.2.2. Case Study 2: Multi-Defect Classification

Scenario. Multiple types of defects often appear together on the same pipeline segment. The section may simultaneously exhibit both a crack and significant corrosion damage. While the ROVON ontology treats rov:Crack and rov:Rust as distinct defect features, their co-occurrence indicates a more serious composite condition—namely, a corroding crack. When such features are detected together on a single pipe, the system uses semantic inference to classify it as a new complex defect category. This allows the knowledge graph to represent multi-faceted structural issues more explicitly and support more informed maintenance decisions (see Table 9).

Rule.

Table 9. RDFox rule identifying pipes with Multi-Defect (Crack and Rust).

PREFIX rov: <http://infoneer.txstate.edu/ROVON/>
PREFIX iof-core: <https://spec.industrialontologies.org/ontology/core/Core/>
PREFIX bfo: <http://purl.obolibrary.org/obo/>
[?pipe, rdf:type, rov:CorrodedCrackPipe] :-
[?crack, rdf:type, rov:Crack],
[?rust, rdf:type, rov:Rust],
[?crack, rov:featureOf, ?pipe],
[?rust, rov:featureOf, ?pipe],
[?pipe, rdf:type, rov:Pipe],
FILTER(?crack != ?rust).

Table 9. RDFox rule identifying pipes with Multi-Defect (Crack and Rust).

PREFIX rov: <http://infoneer.txstate.edu/ROVON/>
PREFIX iof-core: <https://spec.industrialontologies.org/ontology/core/Core/>
PREFIX bfo: <http://purl.obolibrary.org/obo/>
[?pipe, rdf:type, rov:CorrodedCrackPipe] :-
[?crack, rdf:type, rov:Crack],
[?rust, rdf:type, rov:Rust],
[?crack, rov:featureOf, ?pipe],
[?rust, rov:featureOf, ?pipe],
[?pipe, rdf:type, rov:Pipe],
FILTER(?crack != ?rust).

Results. The ontology uses this rule for detecting pipes that show compound structural damage by combining crack and corrosion features. The system determines the pipe category rov:CorrodedCrackPipe when a pipe instance rov:SING0021-pipe displays two features with the rov:Crack and rov:Rust classifications, which both connect to the pipe through the rov:featureOf property. The system classifies pipelines with multiple defect types as these combined defects signal an increased chance of failure. The system achieves better damage assessment precision through qualitative reasoning about dual damage conditions, which enables maintenance action prioritization.

6.3. Hybrid Reasoning

Hybrid reasoning in ROVON represents a semantic inference approach that integrates both quantitative thresholds and logical structural conditions. The system activates when numerical values and relational patterns match specified criteria, which distinguishes it from purely quantitative or qualitative methods. Through this reasoning approach, the system identifies complex environmental conditions where multiple measurement attributes and defect characteristics together suggest potential operational hazards.

The use of hybrid reasoning in underwater infrastructure monitoring stems from the need to conduct root cause analysis. Root cause analysis examines fundamental factors that lead to pipeline damage or failure, while routine inspections focus on surface-level symptoms. The system enhances its ability to generate unified conclusions by linking diverse observations through semantic reasoning—connecting water chemistry data with structural defects and sensor outputs. By combining multiple data types, the system produces more meaningful insights into defect formation patterns, which guide maintenance planning.

The HighHolePhSodium rule serves as a representative example of hybrid reasoning, as illustrated in Figure 17. The rule activates through an imaging process that detects pipeline segments meeting all three of the following conditions: the presence of more than seven holes, a water sample pH greater than 7, and sodium concentrations between 12.3 and 13. The reasoning engine then assigns a new classification, rov:HighHolePhosodium, to indicate cases where chemical aggressiveness and structural vulnerability converge (see Table 10).

Rule.

Figure 17. Semantic rule visualization in ROVON for high sodium, pH levels, and hole count.

Figure 17. Semantic rule visualization in ROVON for high sodium, pH levels, and hole count.

Table 10. Semantic reasoning with RDFox.

PREFIX rov: <http://infoneer.txstate.edu/ROVON/>
PREFIX iof-core: <https://spec.industrialontologies.org/ontology/core/Core/>
PREFIX bfo: <http://purl.obolibrary.org/obo/>
PREFIX xsd: <http://www.w3.org/2001/XMLSchema#>
?[imagingProcess, a, rov:HighHolePhosodium] :-
[?imagingProcess, iof-core:hasOutput, ?image],
[?image, a, rov:image],
[?imagingProcess, bfo:BFO_0000065, ?operSpace],
[?operSpace, rov:contains, ?waterPortion],
AGGREGATE(
[?holeFeature, a, rov:Hole],
[?holeFeature, rov:featureOf, ?object],
[?image, iof-core:hasAbout, ?object]
) ON ?imagingProcess,
BIND COUNT(?holeFeature) AS ?holeCount,
FILTER(?holeCount > 7),
[?waterPortion, iof-core:hasQuality, ?sodium],
[?sodium, iof-core:describedBy, ?sodiumMeasureICE],
[?sodiumMeasureICE, bfo:BFO_0000110, ?sodiumValueSpec],
[?sodiumValueSpec, iof-core:hasSimpleExpressionValue, ?sodiumValue],
FILTER(xsd:decimal(?sodiumValue) >= 12.3 && xsd:decimal(?sodiumValue) <= 13),
[?waterPortion, iof-core:hasQuality, ?ph],
[?ph, iof-core:describedBy, ?phMeasureICE],
[?phMeasureICE, bfo:BFO_0000110, ?phValueSpec],
[?phValueSpec, iof-core:hasSimpleExpressionValue, ?phValue],
FILTER(xsd:decimal(?phValue) > 7).

Table 10. Semantic reasoning with RDFox.

PREFIX rov: <http://infoneer.txstate.edu/ROVON/>
PREFIX iof-core: <https://spec.industrialontologies.org/ontology/core/Core/>
PREFIX bfo: <http://purl.obolibrary.org/obo/>
PREFIX xsd: <http://www.w3.org/2001/XMLSchema#>
?[imagingProcess, a, rov:HighHolePhosodium] :-
[?imagingProcess, iof-core:hasOutput, ?image],
[?image, a, rov:image],
[?imagingProcess, bfo:BFO_0000065, ?operSpace],
[?operSpace, rov:contains, ?waterPortion],
AGGREGATE(
[?holeFeature, a, rov:Hole],
[?holeFeature, rov:featureOf, ?object],
[?image, iof-core:hasAbout, ?object]
) ON ?imagingProcess,
BIND COUNT(?holeFeature) AS ?holeCount,
FILTER(?holeCount > 7),
[?waterPortion, iof-core:hasQuality, ?sodium],
[?sodium, iof-core:describedBy, ?sodiumMeasureICE],
[?sodiumMeasureICE, bfo:BFO_0000110, ?sodiumValueSpec],
[?sodiumValueSpec, iof-core:hasSimpleExpressionValue, ?sodiumValue],
FILTER(xsd:decimal(?sodiumValue) >= 12.3 && xsd:decimal(?sodiumValue) <= 13),
[?waterPortion, iof-core:hasQuality, ?ph],
[?ph, iof-core:describedBy, ?phMeasureICE],
[?phMeasureICE, bfo:BFO_0000110, ?phValueSpec],
[?phValueSpec, iof-core:hasSimpleExpressionValue, ?phValue],
FILTER(xsd:decimal(?phValue) > 7).

ROVON employs this rule to identify critical pipeline segments where environmental stressors coincide with significant structural deterioration. The HighHolePhosodium classification is automatically assigned to imaging processes that detect seven or more holes on an inspected object and elevated pH and sodium levels in the corresponding water portion. This semantic inference allows the system to group similar failure cases, support root cause analysis, and guide inspection planning. By evaluating water quality parameters alongside structural defect patterns, ROVON generates actionable insights that enhance maintenance prioritization and improve the resilience and efficiency of underwater infrastructure operations.

The reasoning framework implemented in ROVON was designed to remain practical when applied beyond the controlled-scale dataset used for validation. While the present experiments relied on a compact knowledge base derived from underwater imagery and sensor data, the architecture of the semantic layer—built on RDFox—supports more demanding settings where data are continuously collected and integrated. RDFox performs all reasoning in main memory and uses parallel Datalog materialization, which allows new triples to be processed without re-computing the entire closure. In previously published evaluations, RDFox demonstrated stable performance with very large knowledge graphs, reaching several billion triples with multi-core parallel execution [32]. These benchmarks suggest that the reasoning layer selected for ROVON is well suited to data-intensive underwater operations, where frequent updates and heterogeneous inputs are expected. Within ROVON, scalability is also aided by the ontology’s modular structure. Each observation, image, or measurement is represented as a localized subgraph linked through IOF-Core and BFO patterns. This structure limits reasoning to relevant portions of the knowledge base rather than requiring global recomputation. Such modularization reduces memory pressure and enables incremental updates during longer underwater missions. Comparative work in large-scale reasoning [36] and hybrid neural approaches [37] further indicates that combining symbolic precision with efficient data partitioning remains a sound strategy for maintaining performance without compromising interpretability. In this context, the current implementation of ROVON provides a balanced foundation—lightweight enough for ongoing data integration, yet formally expressive for rule-based reasoning in real-world robotic applications.

To assess the effectiveness of the reasoning layer, each rule output was reviewed against the manually verified inspection records and annotated dataset. The generated inferences, such as mixed weld and corroded–crack detections, consistently matched the logical conditions defined in the ontology. A small set of randomly selected samples was also cross-checked to estimate quantitative accuracy. Out of 52 evaluated instances, 49 inferences aligned with the reference annotations, yielding an approximate accuracy of 94%. The few mismatches were primarily associated with overlapping or low-confidence detections rather than rule inconsistencies. Overall, this validation confirms that the implemented rules operate reliably and produce consistent, interpretable results within the defined reasoning framework.

7. Conclusions

The ROVON ontology provides a detailed framework that addresses semantic integration challenges in underwater robotics and infrastructure monitoring. Its knowledge graph system combines image-based defect features and sensor measurements with water quality parameters and operational metadata through RDF, establishing a strong foundation for data-driven decision-making in underwater environments. The system ensures semantic interoperability across diverse modalities, supporting the accurate modeling of inspection processes, structural anomalies, and environmental conditions.

ROVON derives its core strength from its implementation of rule-based semantic reasoning. The system adopts a multi-layered reasoning structure that integrates quantitative thresholds, logical entity relationships, and hybrid logic to identify complex inspection scenarios. ROVON also includes qualitative rules to detect weld annotation inconsistencies and infer compound defect types such as CorrodedCrackPipe. These reasoning capabilities enhance ROVON’s diagnostic precision, contributing to improved maintenance planning and operational decision-making.

By semantically correlating structural defects with environmental stressors, ROVON enables a deeper understanding of pipeline deterioration mechanisms. Through ontology-based classifications, visual annotations, and chemical profiles, the system generates actionable inferences that support targeted inspection and proactive maintenance strategies. Unlike earlier ontology frameworks reviewed in Section 2, ROVON avoids the centralization and computational bottlenecks reported in systems such as MTRR and KnowRob. Its modular BFO-based structure separates static and dynamic entities, allowing reasoning over smaller, context-specific subgraphs rather than a single monolithic model. Using OWL 2 RL semantics with the RDFox reasoner maintains logical consistency while keeping inference lightweight and responsive. These design choices provide a scalable and efficient framework suited to underwater operations that involve frequent data updates and heterogeneous sources.

While this study did not include direct benchmarking against other ontology-based frameworks, a quantitative comparison would not be meaningful because existing systems address different operational domains and data scopes. ROVON was developed specifically for underwater robotic inspection, where semantic reasoning must integrate multimodal imagery, sensor measurements, and environmental metadata in near-real-time conditions. In contrast, many prior data integration approaches rely on centralized or task-specific ontologies that emphasize planning or simulation rather than sensor-driven reasoning. Within this context, ROVON focuses on modular design and lightweight reasoning through RDFox, allowing localized inference and incremental data updates without requiring global recomputation. These characteristics make it well suited for continuous and dynamic underwater operations.

Overall, ROVON offers a scalable and extensible ontology framework that supports semantic integration, automated reasoning, and environmental interpretation in underwater robotics. Its unified system bridges robotic perception with domain knowledge and semantic inference, advancing the precision, safety, and efficiency of underwater infrastructure monitoring.

8. Future Work

Enhancing the capabilities and applications of the ROVON ontology in underwater robotics presents numerous opportunities for advancement. One key direction is expanding the ontology’s coverage to include a broader range of underwater scenarios and environments. Incorporating detailed representations of complex underwater phenomena—such as ocean currents, marine life interactions, and varying seabed conditions—will increase the versatility and applicability of the framework. This comprehensive modeling will better address the diverse challenges faced during underwater exploration and infrastructure maintenance.

Integrating emerging technologies like machine learning (ML) and artificial intelligence (AI) with the ROVON framework offers significant potential for development. By combining semantic reasoning with advanced data analysis techniques, the system could provide more sophisticated predictive capabilities and adaptive decision-making processes. This integration would lead to more accurate assessments of underwater infrastructure, enabling proactive maintenance strategies and enhancing operational safety.

Validating ROVON in diverse real-world conditions is essential for refining its robustness. Extensive field trials across various marine environments—from shallow coastal waters to deep-sea regions—will help identify limitations and areas for improvement in the ontology’s structure and functionality. These trials will ensure the framework performs reliably under the dynamic and often harsh conditions encountered in underwater operations.

To encourage widespread adoption, developing standardized interfaces and protocols for seamless interaction with existing underwater robotic systems and data management platforms is crucial. Ensuring interoperability with different hardware and software configurations will facilitate easier implementation, promote collaboration across research and industry sectors, and enhance the efficiency of data sharing and integration.

Supporting autonomous decision-making for underwater vehicles is another promising area. Enhancing the framework’s reasoning capabilities will allow ROVs and AUVs to make complex operational decisions independently based on real-time data analysis and semantic interpretation of their surroundings. This will reduce the need for constant human intervention and improve responsiveness to changing conditions during underwater missions.

Finally, exploring applications of ROVON beyond pipeline monitoring can unlock new opportunities in fields like marine ecology, underwater archaeology, and deep-sea mining. Expanding the framework to support these domains can provide a more comprehensive understanding of underwater environments and contribute to the sustainable management of marine resources.

By pursuing these directions, ROVON has the potential to evolve into a more powerful, versatile tool for underwater robotics, advancing infrastructure monitoring, maintenance, and exploration. These enhancements will ensure the framework remains at the forefront of underwater technology, supporting safer, more efficient, and more resilient operations.