可解释性
特征选择
马尔可夫毯
特征(语言学)
数据挖掘
模式识别(心理学)
计算机科学
中心性
人工智能
断层(地质)
排名(信息检索)
特征提取
机器学习
数学
马尔可夫链
马尔可夫模型
统计
语言学
哲学
地震学
马尔可夫性质
地质学
作者
Chongchong Yu,Mengxiong Li,Zongning Wu,K. Gao,Fei Wang
标识
DOI:10.1088/1361-6501/ad3c5f
摘要
Abstract Feature selection is a crucial step in fault diagnosis. When rolling bearings are susceptible to compound faults, causal relationships are hidden within the signal features. Complex network analysis methods provide a tool for causal relationship modeling and feature importance assessment. Existing studies mainly focus on unweighted networks, overlooking the impact of the strength of causal relationships on feature selection. To address this issue, we propose a compound fault feature selection method based on the causal feature weighted network. First, we construct a weighted network using the incremental association Markov blanket discovery algorithm and Pearson correlation coefficient. Then, we quantify the importance of features by treating node strength as a centrality index and rank them to partition the feature subset. Finally, the optimal feature subset is obtained through a neural network with the accuracy of compound fault diagnosis as the threshold. Analysis of public datasets and comparative experiments demonstrate the advantages of our method. Compared to existing research, our method not only effectively reduces the number of optimal feature subsets to 11 but also improves the accuracy of compound fault diagnosis to 95.2%. Furthermore, we employ the SHapley Additive exPlanations to interpret the contribution of each feature in the optimal subset to the accuracy of compound fault diagnosis. This provides reference from both physical and network perspectives to feature selection and compound fault diagnosis in rolling bearings in practical working conditions.
科研通智能强力驱动
Strongly Powered by AbleSci AI