Scholarlink

<p>As marine diesel engines serve as the primary power source for ships. Traditional intelligent fault diagnosis methods fundamentally struggle to extract weak drift trends masked by strong noise, lacking adaptive parameter optimisation prevents the accurate delineation of fuzzy health-state boundaries in high-dimensional space, as well as poor explainability due to the black-box dynamic fusion of multi-source parameters during inference. This paper develops a novel Co-evolutionary Decoupled Attention Framework (CDAF). Driven by the Cooperative Rime Information Migration Evolutionary&#xa0;(CRIME) optimisation algorithm, to overcome the local minima inherent in manual tuning, the framework seamlessly integrates physically-informed dual-path feature decoupling (DPFD) to isolate weak drifts in high-noise environments, and employs a joint channel-spatial attention mechanism to resolve the fuzzy boundaries of highly similar fault states in high-dimensional space and accurate multi-source feature fusion. Fusion-CAM is further introduced to establish ensure interpretability throughout the inference process, which quantifies branch contributions of the fusion layer and maps thermal parameter responses to engine health status. Validation using real-ship and simulation data shows that the proposed method achieves 99.71% and 95% accuracy under standard and high-noise conditions, respectively. DPFD and CRIME modules enhance accuracy by over 2.2%, demonstrating the value of feature decoupling and global optimisation. Furthermore, Fusion Class Activation Mapping (Fusion-CAM) outperforms Grad-CAM + + in visualization accuracy, as confirmed by feature occlusion tests. Fusion-CAM quantitative analysis reveals that the weight of the trend-guided path exceeds that of the average path. This indicates that parameters such as maximum pressure during combustion, maximum temperature during combustion, brake mean effective pressure, exhaust gas pressure after the turbocharger, and exhaust manifold temperature are critical indicators for diagnosing subtle intake valve leakage.</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

CDAF: a co-evolutionary decoupled attention framework for explainable weak thermal fault diagnosis of marine diesel engines

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月14日

查看全文
  • Zaimi Xie,
  • Chunmei Mo,
  • Baozhu Jia

摘要

As marine diesel engines serve as the primary power source for ships. Traditional intelligent fault diagnosis methods fundamentally struggle to extract weak drift trends masked by strong noise, lacking adaptive parameter optimisation prevents the accurate delineation of fuzzy health-state boundaries in high-dimensional space, as well as poor explainability due to the black-box dynamic fusion of multi-source parameters during inference. This paper develops a novel Co-evolutionary Decoupled Attention Framework (CDAF). Driven by the Cooperative Rime Information Migration Evolutionary (CRIME) optimisation algorithm, to overcome the local minima inherent in manual tuning, the framework seamlessly integrates physically-informed dual-path feature decoupling (DPFD) to isolate weak drifts in high-noise environments, and employs a joint channel-spatial attention mechanism to resolve the fuzzy boundaries of highly similar fault states in high-dimensional space and accurate multi-source feature fusion. Fusion-CAM is further introduced to establish ensure interpretability throughout the inference process, which quantifies branch contributions of the fusion layer and maps thermal parameter responses to engine health status. Validation using real-ship and simulation data shows that the proposed method achieves 99.71% and 95% accuracy under standard and high-noise conditions, respectively. DPFD and CRIME modules enhance accuracy by over 2.2%, demonstrating the value of feature decoupling and global optimisation. Furthermore, Fusion Class Activation Mapping (Fusion-CAM) outperforms Grad-CAM + + in visualization accuracy, as confirmed by feature occlusion tests. Fusion-CAM quantitative analysis reveals that the weight of the trend-guided path exceeds that of the average path. This indicates that parameters such as maximum pressure during combustion, maximum temperature during combustion, brake mean effective pressure, exhaust gas pressure after the turbocharger, and exhaust manifold temperature are critical indicators for diagnosing subtle intake valve leakage.