Global-Local Continual Transfer Network for Intelligent Fault
Diagnosis of Rotating Machinery
Jipu Li
1
, Ke Yue
2
, Jingxiao Liao
1
, Tao Wang
1
and Xiaoge Zhang
1
1
The Hong Kong Polytechnic University, Hung Hom, Hong Kong, 999077, China
jipu1994.li@polyu.edu.hk
jingxiao.liao@connect.polyu.hk
tao3wang@polyu.edu.hk
xiaoge.zhang@polyu.edu.hk
2
South China University of Technology, Guangzhou, Guangdong, 511442, China
202210190718@scut.mail.edu.cn
ABSTRACT
Existing fault diagnosis methods face three fundamental chal-
lenges when deployed under dynamic environments: limited
continuous diagnostic capability, poor generalization, and in-
adequate protection on data privacy. To address these prob-
lems, we develop a novel continual fault diagnosis framework
named Global-Local Continual Transfer Network (GLCTN)
for classifying unlabeled target samples under varying work-
ing conditions without accessing source samples. To this end,
the proposed GLCTN incorporates a consistency loss and a
mutual information loss to facilitate the transfer of learned
diagnostic knowledge from one domain to another domain.
Moreover, a dual-speed optimization strategy is employed to
retain the acquired diagnostic knowledge while empowering
the model to acquire new information. Experiments con-
ducted on an automobile transmission dataset demonstrate
that the proposed GLCTN achieves robust diagnostic perfor-
mance across multiple continuous transfer diagnostic tasks.
1. INTRODUCTION
Data driven-based intelligent fault diagnosis (IFD) methods
have been implemented in various high-end equipment, in-
cluding wind turbines, airplanes, high-speed trains, to name
a few. Precision IFD models are essential for ensuring the
reliable operations of mechanical equipment, reducing ma-
chine breakdowns, and minimizing economic losses (Li et al.,
2024). Consequently, the development of precision IFD mod-
els has become a key research focus in the field of machinery
fault diagnosis.
Jipu Li et al. This is an open-access article distributed under the terms of
the Creative Commons Attribution 3.0 United States License, which permits
unrestricted use, distribution, and reproduction in any medium, provided the
original author and source are credited.
Deep learning (DL), a crucial branch of machine learning,
has garnered widespread attention in recent years and it
has achieved significant success in machinery fault diagno-
sis (Yan et al., 2023). For example, Zhou et al. (2022) pro-
posed a probabilistic Bayesian DL framework to improve the
reliability of diagnostic results for rotating machinery. Zhang
et al. (2023) proposed an end-to-end DL framework for the
IFD of wind turbine gearboxes under non-stationary condi-
tions.
Despite the rapid progress, DL models experience significant
performance degradation when applied to testing conditions
with distributions different from the training data. To over-
come the poor generalization of DL models, transfer learning
(TL) has been leveraged along the development of IFD mod-
els. The key idea of TL is to apply learned knowledge to ad-
dress a similar but distinct task. In machinery fault diagnosis,
researchers have combined DL with TL to leverage DL’s fea-
ture extraction capabilities and TL’s knowledge transfer abil-
ity simultaneously as a means of enhancing the generalization
performance of IFD models. For example, Chen et al. (2023)
proposed a deep parameter-free reconstruction classification
network to solve the fault classification problem of bearing
under different working conditions. Li et al. (2020) pro-
posed a two-stage transfer adversarial network for detecting
the multiple unknown faults in the unlabeled target samples.
Zhao et al. (2022) utilized extreme learning machine and TL
techniques to achieve the IFD of the aero engine. Meanwhile,
continual learning (CL) is introduced to improve the continu-
ous diagnostic ability of IFD models. Inspired by the continu-
ous learning ability of humans, CL enables the model to learn
continuously from new data without the need of retraining the
entire model. This capability is well-suited for the diagnostic
requirements of mechanical equipment in continuous opera-
tion. Li et al. (2023) proposed a deep continual TL method to
1
