1. Introduction
Bearings are important to any processing or manufacturing plant that uses rotating equipment. Unfortunately, however, many bearings fail prematurely in service because of contamination, poor lubrication, misalignment, temperature extremes, poor fitting/fits, shaft unbalance, and misalignment. All of these factors lead to an increase in bearing vibration and may show typical patterns in spectrum. It is essential to have a spectrum of vibration of bearings to diagnose the source of evolving fault(s) for preventive maintenance, as bearings are the most common elements in rotating machinery and with a large number in any factory. Therefore, their failure often leads to unacceptably long production downtimes [
1,
2,
3], and it is important to monitor any potential source of evolving bearing fault. Due to the reasons above, bearing fault diagnosis is receiving more and more attention.
The vibration of rolling bearings results from external forces of machine system, i.e., an unbalance of forces, or bearings act as the excitation sources producing time varying forces that cause system vibration. As for the latter, the components of rolling bearings, i.e., rolling element, inner raceway, outer raceway, and cage interact to generate complex vibration. The sources of vibration include variable compliance [
4] and geometrical imperfections (e.g., surface roughness [
5] and waviness [
6,
7]), and some of these show typical effects in a spectrum [
8]. For example, the increased amplitude at ball or roller pass frequency indicates emerging clearances in bearings or loss of loading, and that of ball spin or at cage rotational frequency gives a signature of cage slipping or collisions. Thus, as stated above, it is essential to have a spectrum of vibration of bearings to diagnose the source of evolving fault(s). However, the vibration signals are usually affected by the variable operating conditions and background noise, which lead to the diversity and complexity of the vibration signal characteristics, and it is a challenge to effectively identify the bearing faults from such vibration signals without further fault information. Thus, it is of significance to extract the fault-related features embedded in the raw signals. In the past few decades, many methods have been developed for feature extraction. Generally, features can be extracted from signals in time domain [
9,
10], frequency domain [
11,
12], and time-frequency domain [
13,
14,
15]. In the time domain, the dimensional statistical indices (e.g., root mean square value and central moment) and dimensionless statistical indices (e.g., skewness and kurtosis) are commonly used that achieve good results. In the view of frequency and time-frequency domains, the vibration signals of the rolling element bearing usually contain a rich spectrum. Different defect positions, such as outer race, inner race, and the ball, have a corresponding fault characteristic frequency. Based on this, some useful methods including the Fast Fourier Transform (FFT), Envelope Spectrum, Short-time Fourier Transform (STFT), Wavelet Transform (WT) [
13], and Empirical Mode Decomposition (EMD) [
14] have been developed for feature selection and are proved effective in determining existing faults and classifying failure modes of mechanical systems. Rai et al. [
16] carried out bearing fault diagnosis by combining FFT and Hilbert–Huang transform. In this study, the transformation results of HHT were further processed by FFT. It was verified that the proposed method could clearly indicate the effectiveness of fault-related frequencies and improve the robustness of diagnosis. Gao et al. [
17] employed STFT and non-negative matrix factorization to extract features for rolling bearing fault recognition. STFT was used to acquire the time-frequency distribution of raw signals. The recognition results yielded a higher accuracy than traditional ANN methods. Ali et al. [
18] made an effort to characterize and classify seven different bearing classes, where EMD is applied to perform primarily processing on the non-stationary vibration signals. The energy entropy is then calculated on each of the obtained intrinsic mode functions (IMFs) and used as features for classification. Guo et al. [
19] proposed a novel signal compression method for bearing vibration signals where the ensemble empirical mode decomposition method is employed for separating fault-related signal components from irrelevant ones. The results proved that the proposed method provided a higher compression ratio, while retaining the bearing defect characteristics. Vafaei et al. [
20] developed a new signal processing method by combining indicated runout with wavelet decomposition, which was employed to effectively window around regions of interest with the indicated runout method. An example in the paper showed that the method could pinpoint with remarkable accuracy the physical cause of the vibration response. Huo et al. [
21] used the impulse modeling continuous wavelet transform (IMCWT) model to diagnose bearing faults under multi-speed, where quasi-Newton optimization algorithms were applied to optimize IMCWT model so that fault information from vibration signals could be extracted more effectively. However, on one hand, the above studies need careful engineering and substantial domain expertise in selecting the useful features related to faults, and on the other hand, information of single domain, such as time-frequency domain, is mostly adopted for fault diagnosis and the time domain and frequency domain information are not considered. These have limited the further development of fault diagnosis especially with the growth of complexity of mechanical systems and computation requirements.
In recent years, with the achievement in deep learning and computation ability, the data-driven methods accompanied with the 2D visualization technique [
22,
23,
24] for feature selection are widely developed and show excellent performance in fault diagnosis. Deep Convolutional Neural Network is one of the most used deep learning structures that has been gaining more and more attention, particularly in the field of image processing [
25], pattern analysis [
26], and fault diagnosis [
27,
28,
29]. For the DCNN-based fault diagnosis, data are primarily processed and transformed into 2D samples as the input of classifier, and the final features for decision-making are obtained by iterations of self-study and not dependent on manual selection. One effective method for primary data processing is to make transformations on the data in time domain, frequency domain, and time-frequency domain to obtain related features which are further used for sample construction. However, as stated above, features of single domain are adopted for fault diagnosis in many earlier studies, which only describe the specific aspect of vibration signals, and with the rich features of the raw signals it is not easy to figure out the domain(s) where features give the most contribution to the fault diagnosis. On this account, some works combine features from more than one domain for a better fault diagnosis performance. For example, Roozbeh et al. [
30] carried out diagnosis on gear faults, where features are first extracted from the sensor measurements in frequency domain and time-frequency domain by corresponding tools such as FFT and WPT, and statistics are then calculated on the obtained results, after which all the features are fused in sequence as the input of diagnosis scheme where DCNN is used. Zhao et al. [
31] extracted features in the time domain (statistical induces), frequency domain (spectral skewness and kurtosis), and time-frequency domain (wavelet energy) for reducing the dimension of raw signals. The obtained results are then used for machine health monitoring.
For most of studies on DCNN-based fault diagnosis, a single network is used, which means there is only one input. This leads to the problem of features integration across different domains, in other words, a careful construction of sample is needed to contain useful information as much as possible. However, the analysis results of raw signals may have different dimensions in different domains, due to which features in different domains cannot be fused directly. For example, the transformation results of raw signal by FFT and WT are [frequency × amplitude] and [time × frequency × amplitude], corresponding to frequency domain and time-frequency domain, respectively. Beside those mentioned above, in single DCNN-based fault diagnosis, plenty of work must be done with respect to structure designing and parameter tuning to obtain a satisfying performance (e.g., diagnosis accuracy). This is generally done by running multiple trials where each trial is a complete execution of training application with values for the chosen structure and hyperparameters and is time-consuming. There is not a common approach applicable for all cases to provide a theoretical guide on this process, so situations may arise in which the performance of the deep learning network is hard to be improved by personal expertise.
Considering the problems above, we propose a scheme that combines diverse CNN learners and AdaBoost tree-based ensemble classifier. The information of raw signals in the time domain, frequency domain, and time-frequency domain are employed in the form of 2D images for diagnosis. Differently, samples are constructed separately and contribute to fault diagnosis simultaneously, which eliminates the situation that only one domain is focused or an elaborate feature integration across different domains must be determined in traditional vibration-based fault diagnosis, thus providing more freedom for the sample construction where each type of sample can concentrate on one aspect of the signal. To our best knowledge, although there were some studies assembling CNNs and the AdaBoost algorithm-based classifier [
32,
33,
34] for classification, proving its effect in improving classification accuracy, few researches have paid attention to the change of this improvement and the robustness of such an ensemble method. Herein, we present a universal framework to improve the diagnosis performance, and give instruction for the practical application of the framework where the proposed methodologies for 2D sample construction in this paper can also be applied in other vibration-based analysis.
6. Conclusions
In this paper, we propose a scheme that combined diverse CNN learners and AdaBoost tree-based ensemble classifier. The information of raw signals in the time domain, frequency domain, and time-frequency domain are employed in form of 2D images for diagnosis. Differently, samples are constructed separately and contribute to fault diagnosis simultaneously, which eliminates the situation that only one domain is focused or an elaborate feature integration across different domains must be determined in traditional vibration-based fault diagnosis, thus providing more freedom for the sample construction where each type of sample can concentrate on one aspect of the signal.
An example based on the CWRU datasets is illustrated to prove the superiority of the proposed scheme by two steps: (1) comparing the performance between DCNN and Ada-DCNN (single DCNN embedded), and (2) comparing the performance between Ada-DCNN and hybrid Ada-DCNN (multi-DCNNs embedded). Both comparisons above are carried out under accuracy levels growing from S1 to S4. The results show that the diagnosis accuracy is averagely improved by an absolute value of 15.37%, 12.29%, 10.40%, and 5.14% from S1 to S4 by using Ada-DCNN instead of DCNN, and 29.26%, 24.59%, 18.68%, and 12.95% by using hybrid Ada-DCNN. Besides, robustness of the proposed hybrid Ada-DCNN is also investigated by dropping the performance of one or more DCNNs embedded. It presents a reduction of only 1.83%, 4.31%, and 12.29% of hybrid Ada-DCNN versus the average drop of 10.11%, 20.21%, and 26.75% of one, two, and three DCNNs, respectively.
According to the results above, a hybrid Ada-DCNN is recommended for use in practical vibration-based fault diagnosis instead of traditional single DCNN method. The example in this paper has provided a specific way for using the proposed scheme to diagnose where the signal preprocessing and image sample construction methods can also be applied in other vibration-based analysis.