AI-Assisted HAZOP Study

Description:

HAZOP (Hazard and Operability) studies [1,2] are a critical component of the process safety of chemical plants. HAZOP studies are traditionally conducted in expert groups and require detailed manual exploration of process deviations which consumes a lot of time.

In this master thesis, large language models (LLMs) are used to assist the generation of HAZOP studies. The tasks include the fine-tuning or prompt-engineering of LLMs [3] for HAZOP-specific tasks, the integration piping and instrumentation diagrams (P&IDs) and textual information, and the evaluation of LLMs in real-world HAZOP case studies with expert feedback.

Approach:

Literature search and familiarization with the topic of HAZOP and LLM

Data Collection

Evaluate (existing) HAZOP reports and process documents
Annotate deviations, causes, consequences, and actions

Model Development

Fine-tune or adapt multimodal LLM-based models with safety engineering language
Use Parameter-Efficient Fine-Tuning (PEFT) approaches to minimize resource requirements

System Integration

Create a prototype that accepts (P&IDs) and textual information
Generate guidewords, deviations, and proposed consequences

Phase 4: Evaluation

Use expert-led sessions to compare AI-assisted vs manual HAZOP tables
Evaluate based on accuracy, completeness, and time efficiency

Supervisors:

Prof. dr. ing. M. B. Franke - m.b.franke@utwente.nl

Literature:

[1] Risk Assessment - ROSS - NTNU

[2] Kletz, T. A.: “Hazop – past and future”. Reliability Engineering and System Safety, 55:263-266, 1997

[3] How are large language models trained? Pre-training | LinkedIn Learning