An adversarial model for electromechanical actuator fault diagnosis under nonideal data conditions

Electromechanical actuators (EMAs) are safety-critical components that work under various conditions and loads. Realizing robust and precise fault diagnosis for an EMA increases the overall availability/safety of the whole system. However, the monitoring data of an EMA are collected under different...

Full description

Saved in:
Bibliographic Details
Published in:Neural computing & applications Vol. 34; no. 8; pp. 5883 - 5904
Main Authors: Wang, Chao, Tao, Laifa, Ding, Yu, Lu, Chen, Ma, Jian
Format: Journal Article
Language:English
Published: London Springer London 01.04.2022
Springer Nature B.V
Subjects:
Actuators
Artificial Intelligence
Artificial neural networks
Computational Biology/Bioinformatics
Computational Science and Engineering
Computer Science
Critical components
Data Mining and Knowledge Discovery
Fault diagnosis
Feature extraction
Image Processing and Computer Vision
Original Article
Probability and Statistics in Computer Science
Robustness
Safety critical
Signal processing
Training
Working conditions
Fault diagnosis
Few-shot learning
Unbalanced samples
Electromechanical actuator
ISSN:0941-0643, 1433-3058
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Electromechanical actuators (EMAs) are safety-critical components that work under various conditions and loads. Realizing robust and precise fault diagnosis for an EMA increases the overall availability/safety of the whole system. However, the monitoring data of an EMA are collected under different working conditions and consist of numerous unlabeled samples and few unbalanced labeled samples; this severely limits the applications of intelligent data-driven diagnosis approaches. Therefore, this paper provides an extended convolutional adversarial autoencoder (ECAAE) as an adversarial model to achieve end-to-end fault diagnosis for EMAs based only on vibration signals. This approach combines the feature extraction ability of convolutional neural networks (CNNs) with the semisupervised learning and data generation capabilities of adversarial autoencoders (AAEs) by activating different submodels during different training phases and is thus able to utilize both unlabeled and unbalanced labeled samples. With the help of a hyperparameter-free signal conversion method and an imbalance-compensation loss function, the adversarial training process of the model results in a feature extractor that is robust to varying working conditions as well as a sample generator capable of generating samples belonging to a given class. Consequently, after fine-tuning on samples rebalanced by the generator, the classifier of the ECAAE is able to perform robust and precise fault diagnosis even under various working conditions, unbalanced samples and few-shot situations. The proposed method is validated under 12 multicondition data scenarios and achieves a diagnostic accuracy above 90%, even in cases worse than 5-way 5-shot scenarios, thereby revealing its superiority over 3 state-of-the-art models.
Bibliography:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:0941-0643
1433-3058
DOI:10.1007/s00521-021-06732-x