Fault Diagnosis of Rolling Bearings under Small Sample

Principal Investigator:

Professor Yining DONG
Assistant Professor, School of Data Science

Co-Principal Investigator:

Professor Lishuai LI
Associate Professor, School of Data Science

 Project Period:  1 October 2022 - 30 September 2024

Rolling bearings are essential components of rotating machinery, and their failures are one of the most frequent causes of mechanical failures. Therefore, it is of great importance to monitor and diagnose rolling bearings in a timely and effective manner to ensure safe and efficient operations and prevent unplanned production interruptions. With the advancement of machine learning and artificial intelligence technology, many machine learning based approaches have been developed for fault diagnosis of rolling bearings, such as support vector machines, random forests and convolutional neural networks.

However, many machine learning methods suffer from two limitations in practice. First, the amount of data under failure states is limited. As machines are typically operated under healthy conditions and shut down immediately when a failure occurs to reduce risks, the amount of data collected under unhealthy states are very limited. In addition, the limitation of transmission channel bandwidth and transmission rate can also lead to the limited samples collected in a fault state. Second, it is difficult and time-consuming to label the data collected during operations. Since many fault diagnosis methods require a large amount of labeled data to build the model, these two limitations make it challenging to develop reliably and accurately identify the health condition of the rolling bearings. In this project, we aim to

(1) Develop data augmentation approaches to generate new samples with high diversity and high fidelity

Data augmentation is one of the most effective methods to solve the problem of fault diagnosis under limited data. Given a small amount of data samples, the major challenge of data augmentation is to generate new data with broad diversity while retaining the key feature information in the original data samples. Traditional oversampling approaches, such as nearest neighbor interpolation (NNI), linear and quadratic interpolation, and synthetic minority oversampling technique (SMOTE) are easy to implement. However, the generated samples are not representative of the original feature information, which significantly affect the fault diagnosis performance. With the rapid development of machine learning and artificial intelligence, researchers have proposed many deep learning methods for data augmentation. Shao et al. proposed a framework based on the auxiliary classifier GAN (ACGAN), which learns from sensor signals and generates realistic one-dimensional data^[1]. Gao et al. proposed a data augmentation method based on Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP), which improves the accuracy of fault diagnosis^[2]. However, these data augmentation approaches based on GAN networks suffer from the drawbacks of networks training difficulties and limited data generation capacities. Zhao et al. used Variational Auto-Encoder (VAE) for vibration data augmentation^[3]. However, the generated new samples are often less realistic than GAN due to differences in the objective functions. We will take the fault diagnosis of rolling bearings as the research object, target the weak enhancement of small samples to develop an effective and efficient method of data augmentation for fault diagnosis of rolling bearings. By studying the diversity of the generated signals in the time domain and the representativeness of feature information in the frequency domain, and by exploring the sparseness of the vibration signals and the reconstruction ability of compressed sensing, a data augmentation method based on compressed sensing will be developed for vibration signal data augmentation to enable accurate fault diagnosis of rolling bearings.

(2) Develop transfer learning methods for fault diagnosis using limited unlabeled data

Transfer learning based intelligent fault diagnosis has received increasing attention in recent years due to the limitation of the labeled fault state data and the discrepancies in the training and testing data distributions. However, in practice, the complexity of operating conditions and the differences in failure degrees can lead to the large discrepancies in the data distribution between domains, which in turn reduces the amount of the transferrable knowledge and the accuracy of fault diagnosis. To deal with this issue, multisource domain transfer learning has been proposed where multiple source domains are used to provide more transferable knowledge. Zheng et al. proposed a multisource domain generalization for fault diagnosis where a Grassmann manifold is used to describe data from multiple source domains^[4]. Huang et al. proposed a multisource dense domain adaptation adversarial network for fault diagnosis that use multi-sensor data information to establish an unsupervised learning framework^[5].
The existing multisource domain transfer learning approaches for fault diagnosis are not able to solve the applicability issue under limited unlabeled data and are not adaptive to the target online data after training. In real applications, although it is convenient to collect some unlabeled operating data of the target machine, it is often difficult and time-consuming to obtain various types of fault data. What strategies can be used to perform cross-machine fault diagnosis under small sample and build an adaptive transfer learning model? To this end, we take the fault diagnosis of rolling bearings as the research object, and we aim to develop a time adaptive multi-source online transfer learning framework for fault diagnosis of rolling bearings. An offline model is first built using limited unlabeled target machine data, and the model is then updated dynamically. The framework is not only able to solve problems with a limited number of samples for each fault type, but also problems with a limited number of available sample types. The integration of the offline training and online updating of the framework will effectively solve the limited unlabeled sample issue in the rolling bearing fault diagnosis, providing guidance and support for the intelligent predictive health maintenance.

References

1. Shao, S., Wang, P., & Yan, R. (2019). Generative adversarial networks for data augmentation in machine fault diagnosis. Computers in Industry, 106, 85-93.
2. Gao, X., Deng, F., & Yue, X. (2020). Data augmentation in fault diagnosis based on the Wasserstein generative adversarial network with gradient penalty. Neurocomputing, 396, 487- 494.
3. Zhao, B., & Yuan, Q. (2021). Improved generative adversarial network for vibration-based fault diagnosis with imbalanced data. Measurement, 169, 108522.
4. Zheng, H., Wang, R., Yang, Y., Li, Y., & Xu, M. (2019). Intelligent fault identification based on multisource domain generalization towards actual diagnosis scenario. IEEE Transactions on Industrial Electronics, 67(2), 1293-1304.
5. Huang, Z., Lei, Z., Wen, G., Huang, X., Zhou, H., Yan, R., & Chen, X. (2021). A multi-source dense adaptation adversarial network for fault diagnosis of machinery. IEEE Transactions on Industrial Electronics.