PhD Oral Exam - Ehab Gaballah Hussein Ahmed Elhosary, Building Engineering

When studying for a doctoral degree (PhD), candidates submit a thesis that provides a critical review of the current state of knowledge of the thesis subject as well as the student’s own contributions to the subject. The distinguishing criterion of doctoral graduate research is a significant and original contribution to knowledge.

Once accepted, the candidate presents the thesis orally. This oral exam is open to the public.

Abstract

The Hazard and Operability (HAZOP) study is a widely used method for hazard identification and analysis in process industries. It systematically identifies potential hazards and related operability malfunctions by dividing piping and instrumentation diagrams (P&IDs) into small sections referred to as P&ID nodes. A node consists of one or more equipment and related instrumentation, with its boundaries defined using different criteria such as main equipment, process flow, material type, and/or operating variables. This is followed by a detailed analysis of each node to generate the study report, referred to as HAZOP notes. These notes typically consist of deviations, possible causes, consequences, safeguards, and recommendations. The process of generating these notes is often time-consuming, costly, and heavily reliant on expert judgment.

A wide range of intelligent systems reported in the literature, as well as software and AI-based tools developed in the market, aim to facilitate HAZOP studies and support efficient generation of related reports. These systems, however, only automate specific aspects of HAZOP notes, lacking full automation and integration with P&ID nodes to ensure correct applicability. Software tools primarily facilitate the initial setup of HAZOP studies and serve as knowledge bases for storing generated reports, while AI-based tools aim to speed up the generation of HAZOP reports. These AI-based tools, however, require manual data entry and significant time and resources to build multilevel functional models (MFMs) for each equipment within a node rather than for the entire nodes. They may also generate inaccurate notes, commonly known as hallucinations, and do not incorporate graphical representations of P&ID nodes.

This research presents an AI-assisted framework to shorten the time required to conduct HAZOP studies and generate related reports by leveraging knowledge from previously analyzed projects, including their P&ID nodes and associated HAZOP notes. The framework consists of four newly developed modules: (i) P&ID node similarity assessment; (ii) Note categorization and hazard prioritization; (iii) Domain augmented note generation; and (iv) A HAZOP-based user interface.

Each module is implemented using a set of artificial intelligence (AI) models. Four You Only Look Once (YOLO) object detection models (YOLOv8n, YOLOv8l, YOLOv8x, and YOLOv11x) were used to detect equipment and instrumentation classes within P&ID nodes and extract their bounding box coordinates, which are subsequently used for similarity computation. Six multimodal large language models (LLMs)—GPT-4o, GPT-4o-mini, Gemini, LlaMa, DeepSeek, and Qwen—were employed to generate node descriptions and assist in generating HAZOP notes. Topic modeling techniques, including BERTopic, Latent Dirichlet Allocation (LDA), and Latent Semantic Analysis (LSA), were applied to cluster HAZOP notes into meaningful topics. Retrieval-augmented generation (RAG) models were used to improve the accuracy and reliability of automated HAZOP note generation.

The First module is applied to identify the degree of similarity (expressed in 3 levels: high, medium, or low) between P&ID nodes of projects at hands and those previously executed considering three metrics: object, location, and functional similarities. Object and location similarity metrics are computed using the detected classes and bounding boxes coordinates obtained from YOLO models. Functional similarity is computed using embeddings of node descriptions generated by multimodal LLMs. These metrics are then combined to compute the P&ID Node Similarity Index (PNSI), which determines the similarity level to guide subsequent analysis. For the high level, the associated HAZOP notes are directly retrieved and used in the preparation of related HAZOP reports. For the medium level, the second module is used to cluster the associated HAZOP notes into key topics and prioritize them based on likelihood, severity, and risk rating, using topic modeling techniques. For the third and last, low level, the third module is applied to automatically generate HAZOP notes. It employs a domain-adaptive model and RAG-based models to enhance the reliability of LLMs outputs. The three modules, referred to above, are integrated into a user-friendly interface to facilitate their respective applications to automatically generate the report of the HAZOP study. The interface is coded using Streamlit, an open-source Python tool for developing interactive web applications.

The developed framework is tested and validated primarily using a dataset of 510 P&ID node images, 6,120 HAZOP notes and 1,140 incident investigation reports collected from diverse sources, including actual HAZOP reports, technical guides, literature, and relevant material on websites. The findings demonstrate good performance with significant potential to support HAZOP teams in generating HAZOP reports. The first module achieved the best detection performance using YOLOv8x, with 91.1% precision, 78.3% recall, and 88.7% mAP@50, and an average PNSI of 92.2% for highly similar nodes. The second module, based on BERTopic, achieved a coherence score of 84.6% and a topic diversity score of 90.7% for classifying safeguards and recommendations into 15 safety systems. In addition, it achieved coherence and diversity scores of 80.0% and 92.4% for grouping causes into 13 main risk factors, outperforming LDA and LSA models. The third module shows that RAG-based models achieved higher semantic similarity (F1-score > 92%) than non-grounded LLMs (F1-score ≈ 87%). The use of RAG-based models improved the valid scenario ratio by approximately 10% and enhanced deviation identification, reaching a maximum valid ratio of 81.25%. In addition, retrieval grounding reduced the over-generation of administrative safeguards and increased the proportion of active safeguards, producing distributions closer to the reference safeguards. The developed interface was demonstrated using three representative use cases, illustrating how the modules interact to enable users to upload a P&ID node and generate the final HAZOP report. The framework is expected to enable HAZOP teams to benefit from measures taken to address hazards encountered in previous projects, thereby capturing the experience gained and reducing the time required to perform HAZOP studies and prepare related reports.