1. Introduction
Bearings are crucial parts of rotating machinery and are widely used in many industries; it was recorded that 44% of machine faults experienced in manufacturing industries are related to bearing failures [
1]. Fault detection is a critical part of system design and maintenance because it helps to improve production efficiency resulting in reduced costs and accidents. Fault detection has gained interest in academia and industry, and has been a hot topic of research because of its significance [
2]. Fault detection and diagnosis methods are generally classified into two groups: model-based methods and data-driven methods. In model-based methods, the model’s output and the actual system’s signals are used to generate several symptoms differentiating normal from faulty machine states; based on these symptoms, faults are determined through classification or inference methods. However, data-driven methods rely on the sensor data collected from the plant and usually use artificial intelligence (AI) to learn and classify characteristic fault features from data. AI plays an inevitable role in industry and manufacturing systems [
3].
Deep learning methods, which are known for their capability to process massive amounts of data and are relatively robust against noise, are the best methods for intelligent fault detection [
4]. Convolutional neural networks (CNNs) [
5,
6], stacked auto Encoders [
7,
8], and deep belief networks (DBNs) [
9] are among the most studied algorithms that could reach very high accuracy. Industrial machines mainly operate in normal conditions, so there are more normal data than fault data, making the available data imbalanced. Even though some methods (such as one-class classification and novelty detection) can detect faults in such conditions, identifying the type of faults is not possible [
10]. To solve this problem, generative algorithms can be employed to generate fault data. Generative algorithms are unsupervised learning paradigms that automatically discover patterns in input data so the model can produce new examples. Variational autoencoders (VAE) and generative adversarial networks (GANs) are the most famous generative models and have been wildly used for bearing fault detection [
11,
12], data augmentation [
13], and predicting remaining useful life. By using generative algorithms, the problem of lack of samples and patterns in industrial data can be solved [
14]. CGAN is a variation of GAN, which can generate conditional new data [
15]. In [
16], where datasets are limited and imbalanced, a conditional deep convolutional generative adversarial network is used for machine fault diagnosis. Yin et al. [
17] also applied a data generation method based on the Wasserstein generative adversarial network and the convolutional neural network for bearing fault detection. These methods aim to extract the input data probability distribution and hidden information so that they can be sampled and used to generate new data. Moreover, the distribution of any condition is unique, so the distributions of unknown conditions can be found with the information of known ones and used to generate data for new conditions where fault data are not available.
In this paper, a novel method inspired by image-to-image translation [
18] was introduced and tested on vibration signals to generate fault data from normal data. Pairs of normal and fault data were fed as inputs to the network at given conditions. After the training phase, the network can generate new fault data under different conditions. It has been assumed that no fault sample is available in other conditions, and the data generation is only conducted using normal data. The efficiency of the proposed method and the quality of generated data were evaluated using different classifiers and visualization methods. The paper’s organization is as follows: A literature review was conducted in
Section 2. Next,
Section 3 introduces some background theories and
Section 4 elaborates on the proposed method and the normal to fault GAN (N2FGAN). The N2FGAN was tested on the Case Western Reserve University (CWRU) dataset and verified in
Section 5, and finally, conclusions are available in
Section 6. In
Figure 1, a flow diagram using N2FGAN to generate synthetic fault data for new conditions in different systems is shown.
,
, …,
refers to different working conditions.
2. Literature Review
Numerous AI techniques, such as traditional machine learning methods and deep learning approaches, have been used for fault recognition and diagnosis in roller ball bearings or rotating parts of machinery. Lei et al. [
19] systematically reviewed the development of intelligent fault diagnosis (IFD) since the adoption of machine learning approaches, presenting the past, present, and future artificial intelligent approaches. Schwendemann et al. [
20] surveyed machine learning in predictive maintenance and condition monitoring of bearings and studied different approaches to classify bearing faults and the severity detection.
Liu et al. [
4] also presented a review of the literature on the applications of artificial intelligence algorithms for the fault detection of rotating machinery with a focus on traditional machine learning methods, such as the naive Bayes classifier, K-nearest neighbor (k-NN), support vector machine (SVM), and artificial neural network algorithms. In earlier years of intelligent fault diagnosis, traditional machine learning approaches involved collecting raw sensor data of various fault types, extracting features from the collected data, and developing diagnosis models from the features to automatically recognize the machine’s health status. Although traditional machine learning methods can automate the fault detection processes, these approaches cannot handle increasingly large data due to their low generalization performances, thereby reducing their accuracy in fault diagnosis. For instance, support vector machine classifiers can be applied to classification and regression problems. However, they do not perform well when applied to multi-class classifications or pairwise classification problems. Some approaches, such as SVMs, are computationally expensive and cannot deal with massive industrial data efficiently [
21].
In recent times, deep learning paradigms for intelligent fault diagnoses have become prominent because they can automatically learn fault characteristics from the data without direct feature extraction. Moreover, they can handle large amounts of industrial data, which is one of the drawbacks of traditional machine learning methods; this has helped to reform intelligent fault diagnoses since the 2010s. Li et al. [
22] reviewed the literature on the applications of deep learning methods for fault diagnoses, analyzing the deep learning approaches in relevant publications to point out the advantages, disadvantages, areas of imperfections, and directions for future research. Although the adoption of deep learning methods has led to many successes, these approaches assume that labeled data are sufficient for training diagnosis models [
19]. However, this assumption is impractical, given the working conditions in most industries. The collected data are inadequate as machines seldom develop faults, and condition data healthier than faults are collected. So, even with deep learning approaches, the collected data are unbalanced and insufficient to train reliable fault diagnosis models. This poses some limitations in using intelligent fault diagnoses in industries. As mentioned earlier, the lack of fault data during the network training process is generally termed a small sample problem [
23]. Researchers have come up with three significant ways to solve the small sample problem data augmentation-based, transfer learning/domain adaptation-based, and model-based strategies. The data augmentation and transfer learning-based methods attempt to increase the amount of data by generating similar data from the existing fault data; slightly modified copies of existing fault data are used to create synthetic data for training the neural network, transfer learning-based approaches use pre-trained networks from similar domains to train the new models in a bid to minimize the amount of data required for training.
GANs have unveiled promising capabilities in intelligent fault diagnoses for data argumentation and adversarial training purposes. They can be considered as potential solutions to the small sample problem because GANs can be used to generate additional data with the same distributions as the original data. The generative adversarial network was first introduced by Goodfellow in 2014 [
24]. Generally, a standard GAN comprises two modules, the generator and the discriminator. The generator learns the distribution of the training data, and a discriminator’s goal is to distinguish the samples of the original training set from the generated ones. This capability exhibited by the generative adversarial network has made its application in intelligent fault diagnosis. Pan et al. [
25] reviewed the related literature on small sample-focused fault diagnosis methods using GANs. Their paper describes the GAN approaches and reviews GAN-based intelligent fault diagnosis applications in the literature while discussing the limitations and future road maps of GAN-based fault diagnosis applications. Li et al. [
26] also presented research on GANs with a focus on the theoretical development and achievements of GANs while introducing and discussing the improved GAN methods and their variants.
Liu et al. [
26] presented a rotating machinery fault diagnostics framework that is based on GANs and multisensor data fusion to generate synthetic data from the original data. Zhang et al. [
27], Wang et al. [
28], and Lv et al. [
29] all made use of one-dimensional time-domain signals to generate synthetic data using GANs for classification and diagnosis of rotating machinery. Similarly, Li et al. [
30], Wang et al. [
31], Zheng et al. [
32], and Wang et al. [
33] used one-dimensional frequency domain signals, and Huang et al. [
34] and Shi et al. [
35] used two-dimensional images while Pan et al. [
36], and Zhou et al. [
37] used one-dimensional feature sets to generate synthetic data.
The original GAN has been extended into various forms, such as the Wasserstein GAN (WGAN), convolutional-based GANs, semi-supervised GANs, and condition-based GANs to enhance the quality of data synthesis and improve the training process. For instance, the complexities of controlling the adversarial process between generator and discriminator cause a mode collapse/gradient disappearance phenomenon leading to unsatisfactory data generation performance of the GAN models. To overcome this challenge, Arjovsky et al. [
38] introduced the Wasserstein GAN to deal with the mode collapse phenomena. It provided a solution to the instability problem of GAN but had the challenge of weight clipping, which was addressed by Gao et al. [
39] through the combination of WGAN with a gradient penalty. Zhang et al. [
40] also attempted to solve the small sample problem, focusing on intelligent fault diagnosis via the multi-module gradient penalized GAN. The proposed method comprises three network modules: generator, discriminator, and classifier. The mechanical signals were generated by adversarial training and were then used as training data. References [
41,
42] also used GANs for the fault diagnosis problem of rotating machinery.
These improved variants of GANs have been extensively applied to roller-bearing fault diagnoses. There have also been many combinations of GANs with other generative models for fault diagnoses, namely encoder, autoencoder (AE), and variational autoencoder. Wang et al. [
31] combined GAN and the conditional variational autoencoder to enhance the quality of generated samples for fault pattern recognition in planetary gearboxes. Reference [
43] proposed an improved fault diagnosis approach to learn the deep features of the data by combining an encoder with GANs, integrating the discriminator with the deep regret analysis method to avoid the mode collapse by imposing the gradient penalty on it. Reference [
43] also proposed a novel method called upgraded GAN, which is a combination of energy-based GANs, auxiliary-classifier, and conditional variational autoencoders. Some other applications of GANs for data argumentation in the literature for fault diagnoses were demonstrated by Liu et al. [
44], who proposed a data synthesis approach using deep feature-enhanced GANs for roller bearing fault diagnoses; [
45] used wavelength transform to extract image features from time-domain signals with GANs to generate more training samples and CNN for fault detection. Generative algorithms have proven to be beneficial for solving the small sample problem encountered when using data-driven approaches for intelligent fault diagnoses. This method is widely accepted and more improvements and modifications to the standard GAN have been embraced in the literature to develop highly effective models capable of detecting and classifying industrial fault data and other applications in intelligent fault diagnosis; it was also adopted in this research work to develop new fault samples.
6. Conclusions
In industrial environments, fault data are scarce, and in many cases, normal data are abundant. Machines work in different conditions (i.e., numerous motor loads and speeds), for which fault samples are rarely available. This makes the utility of any machine learning-based method limited since the developed model will be greatly biased to normal conditions. By augmenting normal data with sufficient fault data in a certain condition, the proposed framework enables machine learning-based models that are more robust for fault diagnoses, even in unforeseen fault conditions.
This paper introduces a novel data augmentation algorithm to synthesize fault data. In this algorithm, a variation of CGAN is proposed that can be trained on normal and fault data of one condition. The trained generator of the network was used to generate fault data from the normal samples for each motor speed for which there were no fault data available. The generated data were compared with the actual data and the normal input data using t-SNE. The results illustrate that the generated fault data have the same characteristics as the real fault data.
Moreover, three different classifiers were employed to validate the quality of the synthesized data. The classifiers were trained on various normal and actual fault samples. For the test phase, a new dataset was extracted from the primary dataset with the actual faults from the target class replaced by generated faults and fed into the trained classifiers for testing. In our experiments, three different conditions were tested with respect to different motor speeds. The results demonstrate that the generated faults are correctly classified with high accuracy (more than 97% in all cases). This proves that the generated fault data are very similar to the actual fault data. On the other hand, three frameworks were provided (including CGAN and WGAN) to evaluate the effectiveness of the proposed model in an imbalanced condition. N2FGAN, compared to the others, has demonstrated a higher similarity to the real data and improved the classification performance significantly.
Future extensions of the present work will focus on exploring the effectiveness of generating the signal features instead of the raw vibration samples. In addition, the work should explore an efficient hyperparameter tuning framework to train the generator faster without compromising its performance. Furthermore, reducing the complexity of the network to reduce the training time can be another avenue for future work.