Search

Intelligent fault diagnosis has been a promising way for condition-based maintenance. However, the small sample problem has limited the application of intelligent fault diagnosis into real industrial manufacturing. Recently, the generative adversarial network (GAN) is considered as a promising way to solve the problem of small sample. For this purpose, this paper reviews the related research results on small-sample-focused fault diagnosis methods using the GAN. First, a systematic description of the GAN, and its variants, including structure-focused and loss-focused improvements, are introduced in the paper. Second, the paper reviews the related GAN-based intelligent fault diagnosis methods and classifies these studies into three main categories, deep generative adversarial networks for data augmentation, adversarial training for transfer learning, and other application scenarios (including GAN for anomaly detection and semi-supervised adversarial learning). Finally, the paper discusses several limitations of existing studies and points out future perspectives of GAN-based applications. • Generative adversarial network for intelligent fault diagnosis under small sample is discussed. • A systematic description of the generative adversarial network, and its variants, is provided. • Existing studies based on generative adversarial network for mechanical fault diagnosis are systematically reviewed and classified in this paper. • Limitations of existing studies, as well as future perspectives, are provided in this paper. [ABSTRACT FROM AUTHOR]

Intelligent fault diagnosis of machines has yielded fruitful achievements; however, the application in engineering scenarios is still not satisfactory. There are two reasons: 1) lacking fault signals of machines to support the training of diagnosis models and 2) difficulty in sensitive fault feature extraction from raw signals. In view of these, an attention-weighted multidepth feature fusion scheme with signals augmentation is proposed. In the proposed scheme, generative adversarial networks (GANs) are applied to expand the fault training samples set, which helps improve the generalization ability of the diagnosis model. Then, we design a multidepth feature fusion net based on a deep convolutional neural network for sensitive feature learning from raw signals. Since features of different depths may differ in their ability to represent faults, the attention mechanism is utilized to adaptively assign weights to features of different depths based on the real-time training results of the model, thus achieving weighted fusion of multidepth features. The experimental results indicate that the proposed scheme can achieve high diagnostic accuracy with only a small number of fault training samples. Moreover, the attention module in the scheme is proved to be effective in sensitive fault feature learning in two-bearing diagnosis cases. [ABSTRACT FROM AUTHOR]

Intelligent fault diagnosis of machines has long been a research hotspot and has achieved fruitful results. However, intelligent fault diagnosis is a difficult issue in the case of a small sample due to the lack of fault signals. To solve this problem, a small sample focused intelligent fault diagnosis method via multimodules gradient penalized generative adversarial networks is proposed. The proposed method consists of three network modules: generator, discriminator, and classifier. By adversarial training, the generator can generate mechanical signals in different health conditions. Because of the high similarity to the signals obtained in practice, the generated signals can also be used as training data so that the limited training dataset of the proposed method is expanded. The classifier has a strong ability to extract fault features from raw mechanical signals and then classify different health conditions. The experimental results on two bearing vibration datasets indicate that the proposed method can not only generate bearing vibration signals but also obtain fairly high fault classificati on accuracy under the small sample condition. [ABSTRACT FROM AUTHOR]

Intelligent fault diagnosis models based on transfer learning achieve cross-domain fault identification under small training samples. Existing cross-domain diagnosis models assume the source and the target data share the health condition label space. However, new fault types will occur in the application stage, and it is usually defined as an open-set fault diagnosis problem. In this study, a one-stage self-supervised momentum contrastive learning model (OSSMCL) is proposed for open-set cross-domain fault diagnosis. We propose a self-supervised contrastive learning based momentum encoder to capture the distinguishable features between sample pairs. Different from common self-supervised pipelines, the proposed method is a one-stage framework fusing a meta-learning paradigm, through which OSSMLC can learn to identify new faults with a few labeled samples in the target domain. The one-stage framework can achieve cross-domain fault diagnosis directly without network fine-tuning, and it will reduce the risk of overfitting under limited target data. The effectiveness of the proposed model is evaluated by three open-set cross-domain fault diagnosis experiments. Compared with the state-of-the-art methods, the proposed model can obtain higher diagnostic accuracy in the open-set scenario with small training samples. • A one-stage self-supervised momentum contrastive learning is proposed for open-set fault diagnosis with a small sample. • We propose an open-set fault diagnosis method OSSMCL to tackle problems that common transfer learning methods are unable to identify new fault types in the target domain. • The performance of OSSMCL is proved by three open-set cross-domain small sample experiments. [ABSTRACT FROM AUTHOR]

The advances of intelligent fault diagnosis in recent years show that deep learning has strong capability of automatic feature extraction and accurate identification for fault signals. Nevertheless, data scarcity and varying working conditions can degrade the performance of the model. More recently, a tool has been proposed to address the above challenges simultaneously. Meta-learning, also known as learning to learn, uses a small sample to quickly adapt to a new task. It has great application potential in few-shot and cross-domain fault diagnosis, and thus has become a promising tool. However, there is a lack of a survey to conclude existing work and look into the future. This paper comprehensively investigates deep meta-learning in fault diagnosis from three views: (i) what to use, (ii) how to use, and (iii) how to develop, i.e. algorithms, applications, and prospects. Algorithms are illustrated by optimization-, metric-, and model-based methods, the applications are concluded in few-shot cross-domain fault diagnosis, and open challenges, as well as prospects, are given to motivate the future work. Additionally, we demonstrate the performance of three approaches on two few-shot cross-domain tasks. Typical meta-learning methods are implemented and available at https://github.com/fyancy/MetaFD. • Review the advances of meta-learning in fault diagnosis for the first time. • Demonstrate deep meta-learning in fault diagnosis via algorithms and applications. • Illuminate meta-learning algorithms by mathematical optimization. • Stimulate future work with open challenges and prospects. [ABSTRACT FROM AUTHOR]