期刊
RELIABILITY ENGINEERING & SYSTEM SAFETY
卷 242, 期 -, 页码 -出版社
ELSEVIER SCI LTD
DOI: 10.1016/j.ress.2023.109704
关键词
Health monitoring; Fault diagnosis; Permutation entropy; Zero-shot learning; Semantic space
Compound fault diagnosis is crucial for the reliability and security of manufacturing equipment. Existing methods mainly focus on untrained compound faults, neglecting more common single faults in the test set. To address these issues, a novel broad zero-shot diagnosis method (BZSD) is proposed, which can identify both single faults and untrained compound faults, improving diagnostic accuracy.
Compound fault diagnosis of rotating machinery is of great significance for the operational reliability and se-curity of manufacturing equipment. Since the possible compound fault types increase exponentially, the com-pound faults appear in the test phase may not be covered during training, posing great challenge for machine health monitoring. Recently, several methods attempt to construct the semantic space for untrained compound faults. However, the semantic space suffers from the low-fidelity to identify the untrained compound faults. Besides, most of these methods only focus on untrained compound faults, ignoring more common single faults in the test set. To address these issues, a novel broad zero-shot diagnosis method (BZSD) is proposed to identify both single faults and untrained compound faults. Firstly, the multiresolution permutation entropy is presented to identify single faults and preliminarily screen out the untrained compound faults, which can prevent the com-pound fault from being biased towards the trained single fault. Then, a high-fidelity semantic space is con-structed to classify the pre-screened compound faults. The proposed fault semantics are close to the ground truth semantics, which is conducive to improving diagnostic accuracy. The experiments demonstrate the effectiveness and superiority of the BZSD for rotating machinery with untrained compound faults.
作者
我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。
推荐
暂无数据