AI-driven similarity assessment between P&ID nodes in support of HAZOP report generation

Piping and Instrumentation Diagrams (P&IDs) are the primary input for Hazard and Operability (HAZOP) studies and the generation of related reports. Each of these diagrams is divided into small sections, referred to as P&ID nodes, for systematic analysis to identify potential hazards and malfunctions. Currently, node division and report generation rely heavily on the team’s experience, making the process time-consuming and limiting the reuse of HAZOP knowledge from past similar projects. In response, a new framework is developed to reuse such knowledge for efficient HAZOP report generation by identifying similarity between current and past P&ID nodes. A set of similarity methods are considered, and a new one is developed based on object, location, and functional similarities, using object detection and large multimodal models (LMMs) as direct inputs. You Only Look Once (YOLO-based) models are used to recognize and localize objects within P&ID nodes, while Gemini is primarily used to generate node descriptions and is benchmarked against three open-source LMMs—DeepSeek-VL, Qwen2-VL, and a LLaMA-based vision model. The framework was validated on 348 nodes collected from diverse sources. The results show that YOLOv8x achieved the best detection performance, with 91.1% precision, 78.3% recall, and 88.7% mAP@50. For description generation, LLaMA and Gemini achieved the highest F1-scores 82.89% and 82.30%, respectively. The similarity method achieved an average similarity score of 92.2%, demonstrating its potential to support more efficient HAZOP report generation.