Scholarlink

The oil and gas industry faces significant safety challenges, making behavioral-based safety (BBS) a critical approach to mitigating risks by focusing on human behavior. Traditionally, safety observations are analyzed manually, a process prone to errors and inefficiencies, particularly when handling large datasets with unstructured narratives. This research introduces an automated pipeline leveraging open-source pretrained language models (LLMs) to streamline BBS narrative report analysis and categorize at-risk observations. Three 4-bit quantized LLMs (Llama 3.1: 8B, Gemma 2: 9B, Mistral-Nemo: 12B) were evaluated for accuracy and inference speed on 10 sample BBS reports, with Gemma 2 selected for its superior performance. To improve the performance of the auto BBS analyzer, techniques such as in-context learning (ICL) and chain-of-thought (CoT) prompting were employed. The model's results were then manually evaluated on 400 sample BBS reports, yielded accuracies of 98% for task identification, 94.5% for positive observations, 97% for at-risk observations, 95.3% for follow-up actions, and 97% for risk categorization. However, due to the absence of a ground truth dataset, further validation by safety experts is necessary to refine the model’s understanding of industry-specific terms and scenarios. The analysis results were visualized in Power BI dashboards, highlighting frequent terms in positive and at-risk observations, and supporting proactive safety interventions. This study demonstrates the potential of open-source LLMs to enhance safety management in high-risk industries by automating BBS report analysis on corporate servers, ensuring secure handling of proprietary data and offering insights to improve safety outcomes.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Leveraging Quantized Language Models for Automating Behavioral-Based Safety Observations Analysis

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Hossein Zavari,
  • Temitayo Matthew Fagbola

摘要

The oil and gas industry faces significant safety challenges, making behavioral-based safety (BBS) a critical approach to mitigating risks by focusing on human behavior. Traditionally, safety observations are analyzed manually, a process prone to errors and inefficiencies, particularly when handling large datasets with unstructured narratives. This research introduces an automated pipeline leveraging open-source pretrained language models (LLMs) to streamline BBS narrative report analysis and categorize at-risk observations. Three 4-bit quantized LLMs (Llama 3.1: 8B, Gemma 2: 9B, Mistral-Nemo: 12B) were evaluated for accuracy and inference speed on 10 sample BBS reports, with Gemma 2 selected for its superior performance. To improve the performance of the auto BBS analyzer, techniques such as in-context learning (ICL) and chain-of-thought (CoT) prompting were employed. The model's results were then manually evaluated on 400 sample BBS reports, yielded accuracies of 98% for task identification, 94.5% for positive observations, 97% for at-risk observations, 95.3% for follow-up actions, and 97% for risk categorization. However, due to the absence of a ground truth dataset, further validation by safety experts is necessary to refine the model’s understanding of industry-specific terms and scenarios. The analysis results were visualized in Power BI dashboards, highlighting frequent terms in positive and at-risk observations, and supporting proactive safety interventions. This study demonstrates the potential of open-source LLMs to enhance safety management in high-risk industries by automating BBS report analysis on corporate servers, ensuring secure handling of proprietary data and offering insights to improve safety outcomes.