Elsevier

Neural Networks

Volume 141, September 2021, Pages 133-144
Neural Networks

Residual wide-kernel deep convolutional auto-encoder for intelligent rotating machinery fault diagnosis with limited samples

https://doi.org/10.1016/j.neunet.2021.04.003Get rights and content

Abstract

This paper deals with the development of a novel deep learning framework to achieve highly accurate rotating machinery fault diagnosis using residual wide-kernel deep convolutional auto-encoder. Unlike most existing methods, in which the input data is processed by fast Fourier transform (FFT) and wavelet transform, this paper aims to learn important features from limited raw vibration signals. Firstly, the wide-kernel convolutional layer is introduced in the convolutional auto-encoder that can ensure the model can learn effective features from the data without any signal processing. Secondly, the residual learning block is introduced in convolutional auto-encoder that can ensure the model with sufficient depth without gradient vanishing and overfitting problems. Thirdly, convolutional auto-encoder can learn constructive features without massive data. To evaluate the performance of the proposed model, Case Western Reserve University (CWRU) bearing dataset and Southeast University (SEU) gearbox dataset are used to test. The experiment results and comparisons verify the denoising and feature extraction ability of the proposed model in the case of very few training samples.

Introduction

Rotating machinery plays an important role in modern industry. Rolling bearings and gears are the most important ingredient in the rotating machinery, which would bring huge economic losses and serious accidents to the industry once they malfunction. Therefore, it is necessary to monitor and control the operating conditions of rotating machinery. Moreover, how a fast and effective method can be developed for detecting faults in bearings and gears is still an essential issue in the modern industry (Klausen et al., 2017, Li et al., 2018, Li et al., 2020, Yin et al., 2014, Zarei et al., 2014).

Hence, many kinds of artificial intelligent algorithms were applied in the problem of rotating machinery fault detection (Zhang et al., 2020, Zhao and Lai, 2019, Zhao, Lai et al., 2019), including machine learning algorithms (e.g. artificial neural networks Gangsar and Tiwari, 2020, Yu et al., 2006, support vector machine (SVM) Liu et al., 2015, Wu et al., 2014, k-nearest neighbors Wang, 2016 and logic regression Gangsar & Tiwari, 2020) and deep learning algorithms (e.g. convolutional neural networks (CNN) Jiang, He, Yan, & Xie, 2018, recurrent neural network (RNN) Li et al., 2018, Miao et al., 2019, Zhao et al., 2017, deep auto-encoders (DAE) Haidong et al., 2018, Sun et al., 2018, and deep belief network (DBN) Yu & Liu, 2020) depending on their superior properties. These artificial intelligent methods can get a good performance on rotating machinery fault diagnosis without any manually selected features. However, there are still some issues related to these methods because of the structure of the rotating machinery vibration signal, such as data processing by FFT (Razavi-Far, Hallaji, Farajzadeh-Zanjani and Saif, 2018), wavelet transform (Abdelgayed, Morsi, & Sidhu, 2017) or other feature transformation techniques (Razavi-Far et al., 2018), the dimensions of input data and the number of training samples. For example, the authors in Dubey and Agrawal (2015) combined artificial neural network with empirical mode decomposition and footprint analysis of Hilbert transform for ball bearing fault analysis. The authors in Soualhi, Medjaher, and Zerhouni (2014) used the Hilbert–Huang transform to extract the health indicators from vibration signals and then used SVM and support vector regression to detect the degradation state and fault. The authors in Shao, McAleer, Yan, and Baldi (2018) combined transfer learning, Visual Geometry Group Network (VGG-16) and wavelet transform, which can transform the vibration signals into two-dimension pictures, to build a high-precision fault diagnosis model. The authors in Zhao et al. (2017) first used the local feature module to extract the features and then used an enhanced gated recurrent unit network to learn the traits among the local features. While these state-of-the-art models have achieved successes, they use some signal processing technologies which can reduce the difficulty of the model to extract signal features. Even if these appropriate methods can extract valid features from the original signal, they still rely on adequate labeled training samples.

Based on the above-mentioned results, because of difficulty in obtaining a large amount of labeled data without any noise, the performance of the model is not completely addressed in the following three aspects, (i) the performance of using testing and training samples that are not from the same distribution, (ii) training and testing samples that contain a lot of noise, and (iii) very few training samples. Nevertheless, CNN as a supervised learning model, the more layers it has, the more labels it needs to train (Guo et al., 2018, Sun et al., 2018). For instance the authors in Shao et al. (2018) used more than 1000 samples per category to train a deep CNN, and the authors in Zhang, Peng, Li, Chen, and Zhang (2017) applied more than 2000 samples to train a wide-kernel CNN model.

AE as a popular unsupervised learning method could learn effective features with unlabeled data, which would improve the disadvantage of CNN that require a large amount of data. Therefore, many researchers try to combine AE and CNN, as convolutional auto-encoder (CAE) (Ince et al., 2016, Yu and Zhou, 2020), for image classification and fault diagnosis based on their powerful performance with limited image pictures (Chen et al., 2020, Chen et al., 2017, Dos Santos et al., 2020, Luo et al., 2017). However, Deep CAE is susceptible to the risk of overfitting and gradient disappearance when the number of training samples is small.

In this paper, we propose a new DNN model, residual wide-kernel deep convolutional auto-encoder (RWKDCAE), to address the above-mentioned problems for rotating machinery fault diagnosis. The main contributions of this paper are summarized as follows.

(1) A new DNN model is proposed for rotating machinery fault diagnosis. In RWKDCAE, one-dimensional wide-kernel convolutional neural network is used to extract important features from raw limited time-domain vibration signals. Besides, the residual learning part is used to make complex networks applicable to learning fault features effectively.

(2) The proposed model is based on a wide-kernel CNN and AE, it inherits the advantages of the wide-kernel CNN and AE that can extract features and recognize fault raw time-series signals without the help of signal processing techniques and feature extraction techniques.

(3) The proposed fault diagnosis model is evaluated through a bearing dataset and a gearbox dataset. The results are compared with the existing ones under different situations.

The remainder of this paper is organized as follows. The preliminary methods related to residual learning methods are given in Section 2. The details of the structure of the proposed model are illustrated in Section 3. In Section 4, the performance of the proposed model is evaluated on two rotating machinery fault datasets. Finally, Section 5 concludes this paper.

Section snippets

Preliminaries

This section describes related methods in residual learning, auto-encoder and convolutional neural networks.

Framework of the proposed method

The AE is proved to have a good performance when the data is processed by FFT for feature extraction before training, while many kinds of AEs cannot learn features well from the raw rotating machine fault data. Therefore, the proposed intelligent fault diagnosis frame with high accuracy is combined with one-dimensional wide-kernel convolutional neural networks, residual learning and auto-encoder. In order to keep the deep learning model with the ability to learn features from the original fault

Experiments

To test the performance of the proposed RWKDCAE, the training samples and testing samples are raw rotating machine time-domain vibration signals without any duplicate points, there are 1024 data points in each sample. The raw signal processing process is shown in Fig. 5. Since the samples do not contain similar data to each other, the training sample set and the test sample set are limited. And two real-life case studies are bearing fault diagnosis and gearbox fault diagnosis. Besides, in order

Conclusion

This paper proposed a new deep learning model, RWKDCAE, for rotating machinery fault diagnosis with limited raw time-domain vibration signal. Firstly, the one-dimensional wide-kernel convolutional layer was introduced into CAE to avoid feature extraction before training and could extract features from raw limited vibration signals. Secondly, the residual learning block was introduced to avoid overfitting and gradient vanishing in the deep learning model. Thirdly, the feature extraction effects

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgments

This research is supported in part by the scholarship from the China Scholarship Council (CSC), China under Grant CSC N201906050158 and in part by the Italian Ministry of Education, University and Research, Italy for the support provided through the Project “Department of Excellence LIS4.0 - Lightweight and Smart Structures for Industry 4.0”.

References (39)

  • YinS. et al.

    Data-driven design of robust fault detection system for wind turbines

    Mechatronics

    (2014)
  • YuY. et al.

    A roller bearing fault diagnosis method based on EMD energy entropy and ANN

    Journal of Sound and Vibration

    (2006)
  • YuJ. et al.

    Knowledge extraction and insertion to deep belief network for gearbox fault diagnosis

    Knowledge-Based Systems

    (2020)
  • ZareiJ. et al.

    Vibration analysis for bearing fault detection and classification using an intelligent filter

    Mechatronics

    (2014)
  • ZhaoH. et al.

    Neighborhood preserving neural network for fault detection

    Neural Networks

    (2019)
  • ZhaoH. et al.

    Global-and-local-structure-based neural network for fault detection

    Neural Networks

    (2019)
  • AbdelgayedT.S. et al.

    Fault detection and classification based on co-training of semisupervised machine learning

    IEEE Transactions on Industrial Electronics

    (2017)
  • ChenH. et al.

    Low-dose CT with a residual encoder-decoder convolutional neural network

    IEEE Transactions on Medical Imaging

    (2017)
  • DubeyR. et al.

    Bearing fault classification using ANN-based Hilbert footprint analysis

    IET Sci Meas Technol

    (2015)
  • Cited by (76)

    View all citing articles on Scopus
    View full text