In this thought leadership piece, Professor Eyad Elyan and Dr Thanh Nguyen of Robert Gordon University and Martin Longmuir of AquaTerra Ltd explore a new Artificial Intelligence (AI)-powered approach to solving complex inspection challenges. The article focuses on leveraging multi-AI agent systems and advanced large language models (LLMs) to automate inspection processes and reporting effectively.
Inspection tasks often involve analysing diverse data sources, including numerical and imaging data. Multi-AI-agent systems in which specialised agents focus on distinct tasks are efficiently suitable in such scenarios. For example, one agent based on Computer Vision techniques is used to process imaging data for anomaly detection, while another focuses on numerical data analysis for potential failure prediction. These agents exchange information and refine outputs to obtain more comprehensive and accurate inspection results.
Machine Learning (ML) and Deep Learning (DL) techniques are the cores of these specialised agents. Imaging agents, for instance, use DL models trained on large datasets to identify patterns or anomalies, while numerical data agents employ ML algorithms to uncover trends and predict future issues. Besides, to ensure the quality and consistency of incoming data, a pre-processing and validation stage will be introduced. It detects and handles missing, noisy, or inconsistent data, enabling downstream agents to work with reliable inputs.
An important aspect of a multi-agent system is agent orchestration - the coordination of multiple agents to work together towards shared goals. An orchestrator AI oversees the process, assigning tasks to agents based on their expertise and combining their outputs to ensure cohesive insights. For instance, in subsea monitoring, the orchestrator can synchronise imaging agents that identify structural damage with numerical agents analysing environmental data, ensuring timely and accurate risk assessments. This dynamic and adaptive coordination enables efficient problem-solving even in highly complex environments.
Developing a Reporting Agent Based on LLMs
A significant innovation in AI-driven inspection systems is the deployment of a specialised reporting agent powered by LLMs. This agent is designed to autonomously generate, customise, and refine inspection reports by interpreting outputs from multi-agent systems.
The reporting agent combines all findings from imaging and numerical data agents, ensuring that the insights are presented in a coherent and structured manner. It can translate technical results into recommendations, adapting the tone and detail level to be suitable for various stakeholders, such as engineers, management teams, or regulators. The agent also integrates contextual information, such as industry-specific standards or past inspection records, to provide reports that are both insightful and compliant with regulations.
Moreover, the reporting agent supports interactive features, allowing users to request clarifications or adjustments in real-time. For example, engineers can query the agent for more detailed explanations of detected anomalies or request alternative formats for the report. This flexibility ensures that the output aligns closely with user requirements, enhancing communication and decision-making.
Continuous Learning for System Improvement
An essential feature of the proposed AI system is its ability for continuous learning and improvement. Agents within the system are equipped with mechanisms to automatically update their models using new data. Besides, the system should have the ability to receive feedback from experts and users. Inspection engineers can provide annotations or corrections to the system's outputs, enabling the agents to learn from their feedback. This continuous learning approach ensures that the system remains adaptive to evolving conditions, improving its performance and reliability over time. By combining automated retraining with human expertise, the system achieves a robust balance between adaptability and precision.
Further Challenges
Despite their advantages, integrating multi-agent systems and LLM-based reporting agents into inspection workflows requires overcoming challenges, such as data privacy, model interpretability, and ensuring compliance with regulatory standards. Collaborative research and development, alongside continuous conversation between AI developers and industry professionals, is essential for addressing these issues.
To conclude, multi-AI agent systems and LLM-powered reporting agents, are paving the way for smarter, more efficient inspection processes. By combining the precision of data analysis with the adaptability of generative models, these technologies offer unprecedented opportunities to enhance operational safety and efficiency while reducing costs.
Case Study
The Net Zero Operations team present a case study from the KTP project with AquaTerra Ltd., focusing on their AquaCLAM equipment, which uses an Ultrasonic Testing (UT) system to collect data and identify issues within tubular structures such as Caissons, Conductors and Marine Piles. In simple terms, UT works by transmitting sound waves through a material. These waves travel through the material until they reach the back wall, where they reflect to the receiver. The time taken for the waves to return is measured and converted into a distance, providing an accurate measurement of the material's thickness. The device exports this data, which can then be analysed to detect anomalies within the material, such as corrosion in conductors, without causing any internal damage. AquaCLAM is also equipped with a camera system that captures the condition of conductor surfaces and monitors surface-related issues.
In this project, the team aims to:
1) Pre-process UT data by removing errors, outliers, and noise, detect and segment anomaly zones, and create accurate 3D visualisations
2) Use video data collected by AquaCLAM to automatically detect and segment surface defects, such as erosion and cracks, on conductors
3) Develop a fully automated system for generating comprehensive inspection reports.
Multi-AI-agent systems excel in this scenario as the project team develops an AI-powered Multi-Agent System where specialised agents process numerical data from the UT system and surface information from video analysis, coordinated by an orchestration agent to ensure seamless collaboration for information exchange and output refinement. Meanwhile, a reporting agent is developed to generate comprehensive inspection reports with annotated visuals, statistical summaries, and maintenance recommendations to support efficient decision-making.
