Scholarlink

<p>Explainability is essential for artificial intelligence (AI) systems, especially in high-safety areas like civil aviation. The Local Interpretable Model-agnostic Explanations (LIME) algorithm gives different results for the same input when run multiple times. The different results reduce the reliability of explainable AI in aviation. Therefore, we propose a Bayesian-Enhanced (BE) mechanism. It combines a Bayesian inference module and a prior knowledge module to accomplish robustness to kernel settings. The result can improve the efficiency of LIME and thus can solve the current problems of LIME. By integrating BE mechanism with LIME, we develop a new explainable method named BE-LIME. BE-LIME produces more stable and consistent explanations. We apply BE-LIME to the Landing Approach Runway Detection (LARD) dataset. Experimental results show that BE-LIME resolves the inconsistency issue in LIME. Using F1 scores, AUC, and LIME scores demonstrates that BE-LIME consistently achieves higher stability, greater attribution focus, and improved robustness over LIME, SHAP, and GradCAM. Additionally, BE-LIME aligns with the explainability goals outlined in the European Union Aviation Safety Agency (EASA) concept paper on machine learning guidelines and supports the data-learning assurance process.</p>

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

Explainable object detection for aircraft visual landing system based on BE-LIME method

  • 期刊论文

  • 发表日期：

    2025年7月28日

查看全文
  • Xi Chen,
  • Ruchao Zhang,
  • Lei Dong,
  • Jiachen Liu

摘要

Explainability is essential for artificial intelligence (AI) systems, especially in high-safety areas like civil aviation. The Local Interpretable Model-agnostic Explanations (LIME) algorithm gives different results for the same input when run multiple times. The different results reduce the reliability of explainable AI in aviation. Therefore, we propose a Bayesian-Enhanced (BE) mechanism. It combines a Bayesian inference module and a prior knowledge module to accomplish robustness to kernel settings. The result can improve the efficiency of LIME and thus can solve the current problems of LIME. By integrating BE mechanism with LIME, we develop a new explainable method named BE-LIME. BE-LIME produces more stable and consistent explanations. We apply BE-LIME to the Landing Approach Runway Detection (LARD) dataset. Experimental results show that BE-LIME resolves the inconsistency issue in LIME. Using F1 scores, AUC, and LIME scores demonstrates that BE-LIME consistently achieves higher stability, greater attribution focus, and improved robustness over LIME, SHAP, and GradCAM. Additionally, BE-LIME aligns with the explainability goals outlined in the European Union Aviation Safety Agency (EASA) concept paper on machine learning guidelines and supports the data-learning assurance process.