A mechanical fault diagnosis model with semi-supervised variational autoencoder based on long short-term memory network

Condition monitoring and accurate fault diagnosis are always concerned for stable operating of mechanical equipment. The fault diagnosis based on supervised deep learning has been proved to be effective by their powerful capacities in feature extracting, but usually requiring large number of labeled...

Full description

Saved in:
Bibliographic Details
Published in:Nonlinear dynamics Vol. 113; no. 1; pp. 459 - 478
Main Authors: Qu, Yuanyuan, Li, Tao, Fu, Shichen, Wang, Zhisheng, Chen, Jian, Zhang, Yupeng
Format: Journal Article
Language:English
Published: Dordrecht Springer Netherlands 01.01.2025
Springer Nature B.V
Subjects:
Accuracy
Applications of Nonlinear Dynamics and Chaos Theory
Classical Mechanics
Cluster analysis
Clustering
Coal mining
Condition monitoring
Control
Deep learning
Dynamical Systems
Fault diagnosis
Feature extraction
Machine learning
Original Paper
Physics
Physics and Astronomy
Regularization
Semi-supervised learning
Statistical Physics and Dynamical Systems
Vibration
Vibration analysis
Fault diagnosis
Variational autoencoder
Long short-term memory network
Clustering analysis
Semi-supervised learning
ISSN:0924-090X, 1573-269X
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Condition monitoring and accurate fault diagnosis are always concerned for stable operating of mechanical equipment. The fault diagnosis based on supervised deep learning has been proved to be effective by their powerful capacities in feature extracting, but usually requiring large number of labeled data. Faced with the actual situation that labeled samples are often in short, data are imbalanced in category etc., accurate fault diagnosis based on deep learning is still challenging, so does to explore and explain the evolution of complex faults. A mechanical fault diagnosis model with Semi-Supervised Variational Autoencoder based on Long Short-Term Memory network (LSTM-SSVAE) is proposed in this paper. Through semi-supervised learning, LSTM-SSVAE uses unlabeled data to enhance the extraction of discriminant features of data, which make the model less dependent on only labeled data while giving improved fault diagnosis accuracy. The LSTM networks are applied as the encoder and decoder innovatively, and regularization constraints are added in loss function, to improve the clustering effect of the intermediate hidden variables, so that to achieve effective feature extraction and state detection. Based on open datasets, experimental results show that with the same number of labeled samples, the fault diagnosis accuracy obtained by using LSTM-SSVAE is higher than other typical semi-supervised learning models. Based on actual vibration data of working equipment in coal mining, the feasibility of clustering analysis of intermediate hidden variables also proves that the LSTM-SSVAE model is recommendable for fault evolution analysis and is potential for operating conditions prediction of mechanical equipment.
Bibliography:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:0924-090X
1573-269X
DOI:10.1007/s11071-024-10221-w