1. Unsupervised adversarial and cycle consistent feature extraction network for intelligent fault diagnosis.
- Author
-
Yi-Die, Wang, Pei-Pei, Chao, Rui-Yuan, Zhang, Tang, Hong, Yu-Cheng, Wei, and Hong-Liang, Dai
- Subjects
MECHANICAL engineering ,FAULT diagnosis ,CONVOLUTIONAL neural networks ,INTELLIGENT networks ,NOISE - Abstract
Machine failures often arise from cumulative aging and anomalies, with early-warning indicators hidden in sensor-collected data. When high-quality labeled data are scarce for determining machine conditions, such failures can lead to performance decline and unforeseen incidents. To address these challenges, an unsupervised adversarial and cycle-consistent feature extraction network is presented, named CycleFeature. Specifically engineered for mechanical fault diagnosis, CycleFeature excels in scenarios characterized by limited labeled data and progressive machine aging. It integrates four crucial components: a sample feature extractor, a noise feature extractor, a generator, and a discriminator. This architecture employs multi-layer perceptions and convolutional neural networks. The training protocol for CycleFeature comprises two discrete phases: pre-training and integrated training. In the pre-training phase, the network collaboration ensures that the extracted features maintain cycle consistency and accurately represent the sample data. In the integrated training phase, the noise feature extractor captures features from noise and sample data. Through adversarial learning involving the sample feature extractor and the discriminator, the model fine-tunes the generated noise features to closely mirror the sample features. Tests are conducted based on the early health and pre-retirement high-risk data from 20 normally retired and 4 accident-retired vehicles. CycleFeature shows excellent prediction accuracy in the early warnings of these cases. It garnered higher frequency and value scores on the pre-retirement high-risk data from 24 vehicles compared to early health data, robustly confirming its accuracy and effectiveness. [Display omitted] • The unlabeled battery dataset is processed using MLP and CNN techniques. • A novel neural network architecture is specifically engineered for fault diagnosis. • Empirical validation robustly confirms the efficacy of the proposed approach. [ABSTRACT FROM AUTHOR]
- Published
- 2024
- Full Text
- View/download PDF