TY - JOUR
T1 - Rolling bearing fault diagnosis using optimal ensemble deep transfer network
AU - Li, Xingqiu
AU - Jiang, Hongkai
AU - Wang, Ruixin
AU - Niu, Maogui
N1 - Publisher Copyright:
© 2020
PY - 2021/2/15
Y1 - 2021/2/15
N2 - Rolling bearing fault diagnosis with unlabeled data is a meaningful yet challenging task. Recently, deep transfer learning methods with maximum mean discrepancy (MMD) have achieved great attention. To further enhance the performance of individual models, this paper proposes an optimal ensemble deep transfer network (OEDTN). The proposed method takes advantage of parameter transfer learning, domain adaptation and ensemble learning. Firstly, different kernel MMDs are used to construct multiple diverse deep transfer networks (DTNs) for feature adaptation. Secondly, parameter transfer learning is applied to initialize these DTNs with a good start point. Finally, ensemble learning is used to combine these DTNs to acquire the final results. Considering no labeled information available for ensemble, a novel comprehensive metric is designed to guide the particle swarm optimization to assign suitable voting weights for each DTN. By this way, the ensemble strategy of OEDTN can be adaptively constructed. Experiments on three bearing test rigs are carried out, and the results show that the proposed method is more effective than the existing methods.
AB - Rolling bearing fault diagnosis with unlabeled data is a meaningful yet challenging task. Recently, deep transfer learning methods with maximum mean discrepancy (MMD) have achieved great attention. To further enhance the performance of individual models, this paper proposes an optimal ensemble deep transfer network (OEDTN). The proposed method takes advantage of parameter transfer learning, domain adaptation and ensemble learning. Firstly, different kernel MMDs are used to construct multiple diverse deep transfer networks (DTNs) for feature adaptation. Secondly, parameter transfer learning is applied to initialize these DTNs with a good start point. Finally, ensemble learning is used to combine these DTNs to acquire the final results. Considering no labeled information available for ensemble, a novel comprehensive metric is designed to guide the particle swarm optimization to assign suitable voting weights for each DTN. By this way, the ensemble strategy of OEDTN can be adaptively constructed. Experiments on three bearing test rigs are carried out, and the results show that the proposed method is more effective than the existing methods.
KW - Domain adaptation
KW - Fault diagnosis
KW - Kernel maximum mean discrepancy
KW - Optimal ensemble deep transfer network
KW - Rolling bearing
UR - http://www.scopus.com/inward/record.url?scp=85098094030&partnerID=8YFLogxK
U2 - 10.1016/j.knosys.2020.106695
DO - 10.1016/j.knosys.2020.106695
M3 - 文章
AN - SCOPUS:85098094030
SN - 0950-7051
VL - 213
JO - Knowledge-Based Systems
JF - Knowledge-Based Systems
M1 - 106695
ER -