1. Introduction
Electrical machines are used ubiquitously in industrial applications nowadays. With the development and advancement in science and technology, modern industries are also developing rapidly. As a result, machinery equipment functions on a day-to-day basis and for almost every application, which means that these types of machinery work under unfavorable circumstances, excessive loads, and humidity. This results in motor failures, leading to massive maintenance expenditures, declines in production levels, severe financial losses, and a possible risk of loss of lives. The rotating machinery and induction engines play a vital role in the manufacturing systems. These rotating machines consist of numerous elements, such as a stator, rotor, shaft, and bearings. Rolling element bearings (REBs) are generally termed bearings and are the most vital and vulnerable components in the machine, whose fitness state affects the effectiveness and performance, stability, and lifespan of the machinery [
1,
2]. The four components of REBs are the ball (B), cage, inner-race (IR), and outer-race (OR). The experimental test rig of the Case Western Reserve University (CWRU) ball bearing system and the bearing components are shown in
Figure 1. The bearing fault, one of the most common faults in machinery, accounts for 30% of the total faults, causing the machine to break down and eventually resulting in a severe loss of safety, property, and even the loss of lives in some cases. Hence, bearing fault detection and diagnosis have attracted researchers and scientists and have become essential for scientific advancement [
3,
4]. With the growing concept of Industry 4.0 and smart manufacturing, intelligent methods for detecting and classifying machinery faults have been a key part of scientific research and interest. The fault detection system and the bearing health-state monitoring system are anticipated to provide information regarding the actual working state of the machinery equipment continuously without hampering the production line. Again, the mechanical vibration signals are considered rich sources of information for the appropriate analysis of processes related to bearing faults. These vibration signals can provide enough information about the location and type of the fault, which is helpful for fault diagnosis [
1,
5].
The working process of the ball bearing system consists of sensors placed in different locations in the equipment, via which the signals are transferred to the data acquisition system for additional processing.
Figure 2 shows the vibration data collection process of the CWRU ball bearing system. The performance of fault detection methods depends both on the quality of the vibration signals collected and on the efficiency of the applied signal processing and feature extraction methods [
5,
6]. Traditionally, the maintenance of these REBs used to be a posterior task, usually taking place after the occurrence of the fault. Moreover, this kind of posterior maintenance procedure leads the machine to break down, resulting in financial loss and other casualties [
3]. Hence, it is of great significance to surveil the bearing condition during the working state of the engine. Many signal processing, machine learning (ML), and deep learning (DL)-based methods have been suggested and implemented in bearing condition monitoring and bearing fault detection and diagnosis.
Data-driven methods use signal processing techniques in the time domain, frequency domain, and time-frequency domain to analyze vibration signals. With the use of these signal processing approaches, the appropriate height of fault detection and diagnosis accuracies were stated [
7,
8]. Nevertheless, these conventional signal processing methods carry some limitations. The time-domain method uses the natural properties of the vibration signals in the time domain, such as root mean square, crest factor, quadratic mean, and skewness. These characteristics can be used in dynamic system monitoring applications to effectively reflect transient machine conditions assuming a stationary signal. In actual industrial settings, condition monitoring signals are often complicated by time-varying environmental conditions such as temperature and lubricants. In addition to background noise and interference, spectral changes and nonlinear behavior are also complicated. In other words, the actual vibration signal captured is not stationary, which limits the validity of time-domain statistics. Again, due to the weak amplitude and short duration of structural changes in the vibration signal in the initial stage, the frequency-domain approaches may be unreliable for evaluating non-stationary machine conditions. Another limitation of these methods is that they are inadequate to deal with non-stationary signals [
9].
To be more specific, temporal analysis is not capable of finding the faulty component of the machine. The frequency peak of the bearing fault is not easily distinguishable through FFT analysis. Correspondingly, cepstrum analysis is computationally expensive, and it generates many undesired large peaks near the zero point, making the output complex to interpret. The prerequisite of some experience and knowledge regarding the resonance frequency and filtering band makes envelope analysis challenging to use. Moreover, wavelet transform remains weak in the selection of an appropriate mother wavelet, decomposition level, and respective frequency band, which is necessary information for fault analysis and detection [
6,
10]. Again, the maxima of these methods are troublesome as they require features such as the mean, median, minimum, maximum, peak-to-peak, kurtosis, skewness, standard deviation, absolute mean, and root-mean-square (RMS) value for describing the actual bearing condition [
11]. One finds it difficult to choose the exact features for analyzing the particular signal used in the classification [
12]. Therefore, many ML/DL-based algorithms have been used ubiquitously for the ease of selecting the exceptional patterns present in the data, which are challenging for a human being to identify.
Machine learning, which is a subfield of artificial intelligence, can generate insights in data, even if they are not specifically instructed regarding what to search for in the data [
13]. Many ML-based methods have been proposed and implemented to develop a knowledge-based architecture for the prior diagnosis of bearing faults to prevent catastrophic failure and reduce operating costs. ML-based algorithms such as artificial neural networks, principal component analysis, support vector machines, k-Nearest Neighbors, and singular value decomposition are broadly used in bearing fault detection and diagnosis. A comprehensive review of such approaches can be found in [
14]. However, the problem-solving approach of ML is not satisfactory. ML algorithms first divide the problem statement into different parts and then combine the result. Alternatively, deep learning algorithms are widely accepted and implemented. Deep learning is a subfield of machine learning that defines both higher- and lower-level categories with greater accuracy. Deep learning techniques provide better efficiency and accuracy [
15]. The efficient working ability of these algorithms with a huge amount of data, their end-to-end problem-solving approach, and the pleasing result has attracted many scholars [
16]. DL-based algorithms are ubiquitously used in almost every field.
1.1. Methodology
The methodology used in this study is shown in
Figure 3. First, raw bearing vibration data were collected from the CWRU bearing database, originally stored as MATLAB files in a one-dimensional format. They were then pre-processed for 1-D and 2-D CNN inputs; for the 1-D CNN input, each sample was created from 1600 data points; for the 2-D CNN input, a two-dimensional image representation of 40 × 40 pixels was created. We then used each model to perform feature extraction and classification. Finally, we performed a performance analysis and drew conclusions.
1.2. Contribution and Organization
Inspired by the widespread use of convolutional neural networks, a typical deep learning model in computer vision, we used a 1-D CNN-based model to detect and classify bearing faults on CWRU time-series data. The proposed model appropriately utilizes the feature extraction and classification properties of CNNs. Thus, it is simple to apply for time-sequence data and efficient in terms of computational complexity. Furthermore, we also compared the proposed model’s performance with a 2-D CNN using a two-dimensional image illustration of raw data as an input. Four different datasets were used in this research. With the use of a smaller amount of training data also, we achieved promising results. We also performed a sensitivity analysis of the proposed 1-D CNN model. Precision, recall, and f-measure were calculated, along with the accuracy, which are helpful in demonstrating the efficiency of the proposed architecture. Simplicity and computational feasibility are the main advantages of this model.
The remainder of the paper is organized as follows. A fundamental introduction to bearings and faults in bearings, conventional signal processing, and ML/DL approaches is presented in
Section 1.
Section 2 contains the related theory and related work—a brief explanation of the published works employing CNN and other DL-based architectures.
Section 3 describes the experimental analysis using the proposed method. This section is the core section of this research. The analysis of the results and the sensitivity analysis are presented in this section.
Section 4 presents the comparison between the proposed method and two-dimensional CNN using 2-D illustrated images as input. We also compares the proposed method with some of the published works as well in this section. The paper is summarized in
Section 5, with a discussion and conclusions.
5. Discussion and Conclusions
In the context of the growing concepts of Industry 4.0 and smart manufacturing, intelligent methods for detecting and classifying machine faults are the subject of increasing scientific research and interest. Different signal processing and ML-based approaches have been used in detecting and classifying bearing faults. Signal processing techniques in the time domain, frequency domain, and time-frequency domain have been used to analyze vibration signals. However, due to the various limitations of the typical signal processing and ML-based approaches, DL-based methods are preferred over them. An intelligent method for bearing fault detection and classification in one of the most used benchmark datasets, the CWRU dataset, is presented in this paper. The 1D CNN-based deep learning approach is implemented for the time-sequence bearing data. The raw vibration data from four datasets are divided into N samples, with each sample containing 1600 data points, and then fed into a 1-D CNN for feature extraction and classification.
Moreover, a comparison of the proposed method with the 2-D CNN using 2-D image representation of the raw bearing signal as input is carried out. The result shows that the 1-D CNN performs efficiently for time-series data. In addition, sensitivity analysis of the proposed model is performed, in which metrics such as precision, recall, and f1-score are determined. We also compare the proposed method with some of the published works as well. We use Keras [
50] for training the model, using TensorFlow [
51] at the backend. The model is trained through the GeForce RTX 2080 Ti GPU of NVIDIA [
52]. The results analysis also shows that the time for training the 1-D CNN model is lower than that for the 2-D CNN. The Results section thus shows that implementing a 1-D CNN is more efficient in terms of computational complexity for time-series data. With only 276,522 parameters, the proposed method achieves state-of-the-art accuracy. Simplicity and computational feasibility are the main advantages of this model.