Scholarlink

<p>As aircraft engines undergo frequent operational changes across different flight phases, traditional transfer learning methods—heavily reliant on stationary operating conditions—struggle to maintain diagnostic accuracy. To address the challenges posed by degraded pseudo-label quality and class distribution shift under varying conditions, a hierarchical confidence-adversarial conditional alignment (HCACA) method is proposed. Unlike typical dual-classifier adversarial methods, which primarily guide feature adaptation by maximizing classifier disagreement, HCACA places greater emphasis on leveraging the confidence of target-domain samples and information near class boundaries. First, a memory module is constructed to store source domain feature representations, and initial pseudo-labels for target domain samples are generated using a K-means clustering algorithm. Then, a hierarchical confidence-adversarial mechanism is designed to intelligently partition target domain features into core samples (high confidence) and peripheral samples (low confidence). Adversarial objective maximization is applied to core samples, while adversarial objective minimization is applied to peripheral samples, thereby enhancing the discriminability and reliability of pseudo-labels. Finally, a conditional alignment strategy is introduced to perform fine-grained feature alignment across source and target domains at the category level, ensuring effective correction of class-wise distribution discrepancies. Experiments conducted on datasets encompassing multiple real-world engine operating phases demonstrate that the proposed HCACA method achieves significantly improved diagnostic performance compared to state-of-the-art approaches, confirming its superior effectiveness in variable-condition fault diagnosis.</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

A hierarchical confidence-driven adversarial alignment method for aircraft engine fault diagnosis under variable operating conditions

  • 期刊论文

  • 发表日期：

    2026年5月8日

查看全文
  • Hang Ge,
  • Suiyuan Jia,
  • Feitong Xu,
  • Guodong Liu,
  • Yingchen Dai

摘要

As aircraft engines undergo frequent operational changes across different flight phases, traditional transfer learning methods—heavily reliant on stationary operating conditions—struggle to maintain diagnostic accuracy. To address the challenges posed by degraded pseudo-label quality and class distribution shift under varying conditions, a hierarchical confidence-adversarial conditional alignment (HCACA) method is proposed. Unlike typical dual-classifier adversarial methods, which primarily guide feature adaptation by maximizing classifier disagreement, HCACA places greater emphasis on leveraging the confidence of target-domain samples and information near class boundaries. First, a memory module is constructed to store source domain feature representations, and initial pseudo-labels for target domain samples are generated using a K-means clustering algorithm. Then, a hierarchical confidence-adversarial mechanism is designed to intelligently partition target domain features into core samples (high confidence) and peripheral samples (low confidence). Adversarial objective maximization is applied to core samples, while adversarial objective minimization is applied to peripheral samples, thereby enhancing the discriminability and reliability of pseudo-labels. Finally, a conditional alignment strategy is introduced to perform fine-grained feature alignment across source and target domains at the category level, ensuring effective correction of class-wise distribution discrepancies. Experiments conducted on datasets encompassing multiple real-world engine operating phases demonstrate that the proposed HCACA method achieves significantly improved diagnostic performance compared to state-of-the-art approaches, confirming its superior effectiveness in variable-condition fault diagnosis.