Scholarlink

<p>To address the challenges of parameter redundancy, limited interpretability, and the inability to effectively represent uncertainty in expert knowledge in traditional Belief Rule Base models for complex engineering system fault diagnosis, an adaptive reference value optimization framework based on reinforcement learning is proposed in this work. First, a hybrid mechanism integrating Deep Q-Network and Soft Actor–Critic is developed to jointly explore discrete structural operations and continuous reference value adjustments, achieving a dynamic balance between model structural compactness and inferential accuracy. Second, a power set identification framework is introduced. By incorporating belief variance and entropy into the parameter optimization process, this framework effectively captures the multi-dimensional uncertainty inherent in expert knowledge. Furthermore, a multi-criteria structural interpretability analysis method is presented to quantify structural changes before and after reference value optimization using rule correlation, activation entropy, and clustering separability. Finally, experimental results demonstrate that the proposed framework not only improves predictive accuracy but also achieves a more compact and interpretable model structure under uncertainty, with consistent robustness and generalization across diverse experimental scenarios.</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

Uncertainty-Aware and Interpretable Fault Diagnosis for Complex Engineering Systems Using Adaptive Rule-Based Reasoning

  • 期刊论文

  • 发表日期：

    2026年7月14日

查看全文
  • Dongyang Liu,
  • Jun Tao,
  • Tianhong Pan

摘要

To address the challenges of parameter redundancy, limited interpretability, and the inability to effectively represent uncertainty in expert knowledge in traditional Belief Rule Base models for complex engineering system fault diagnosis, an adaptive reference value optimization framework based on reinforcement learning is proposed in this work. First, a hybrid mechanism integrating Deep Q-Network and Soft Actor–Critic is developed to jointly explore discrete structural operations and continuous reference value adjustments, achieving a dynamic balance between model structural compactness and inferential accuracy. Second, a power set identification framework is introduced. By incorporating belief variance and entropy into the parameter optimization process, this framework effectively captures the multi-dimensional uncertainty inherent in expert knowledge. Furthermore, a multi-criteria structural interpretability analysis method is presented to quantify structural changes before and after reference value optimization using rule correlation, activation entropy, and clustering separability. Finally, experimental results demonstrate that the proposed framework not only improves predictive accuracy but also achieves a more compact and interpretable model structure under uncertainty, with consistent robustness and generalization across diverse experimental scenarios.