A new hyper-parameter optimization method for machine learning in fault classification

doi:10.1007/s10489-022-04238-0

2022-10-21

PRE

Xingchen Ye

Gao Liang (Liang Gao)

Xinyu Li (Xinyu Li)

Long Wen *

DOI：10.1007/s10489-022-04238-0

原文链接

原文求助

摘要

En 中文

Accurate bearing fault classification is essential for the safe and stable operation of rotating machinery. The success of Machine Learning (ML) in fault classification is mainly dependent on efficient features and the optimal pre-defined hyper-parameters. Various hyper-parameter optimization (HPO) methods have been proposed to tune the ML algorithms' hyper-parameters in low dimensions but ignore the hyper-parameters of Feature Engineering (FE). The hyper-parameter dimension is high because both FE and the ML algorithm contain many hyper-parameters. This paper proposed a new HPO method for high dimensions based on dimension reduction and partial dependencies. Firstly, the whole hyper-parameter space is separated into two subspaces of FE and the ML algorithm to reduce time consumption. Secondly, the sensitive intervals of hyperparameters can be recognized by partial dependencies due to the nonlinearity of the relationship between the hyperparameters. Then HPO is conducted in intervals to acquire more satisfactory accuracy. The proposed method is verified on three OpenML datasets and the CWRU bearing dataset. The results show that it can automatically construct efficient domain features and outperforms traditional HPO methods and famous ML algorithms. The proposed method is also very time efficient.

Keyword:

Hyper-parameter optimization

Fault classification

Dimension reduction

Partial dependencies