Bearings are considered as indispensable and critical components of mechanical equipment, which support the basic forces and dynamic loads. Across different condition monitoring (CM) techniques, infrared thermography (IRT) has gained the limelight due to its noncontact nature, high accuracy, and reliability. This article presents the use of IRT for the bearing fault diagnosis. A two-dimensional discrete wavelet transform (2D-DWT) has been applied for the decomposition of the thermal image. Principal component analysis (PCA) has been used for the reduction of dimensionality of extracted features, and thereafter the most relevant features are accomplished. Furthermore, support vector machine (SVM), linear discriminant analysis (LDA), and k-nearest neighbor (KNN) as the classifiers were considered for classification of faults and performance assessment. The results reveal that the SVM outperformed LDA as well as KNN. Noncontact condition monitoring shows a great potential to be implemented in determining the health of machine. The utilization of noncontact thermal imaging-based instruments has enormous potential in anticipating the maintenance and increased machine availability.
Condition-based maintenance and condition monitoring are associated with maintenance of equipment based on the real-time condition of subsystem(s) of the machine. Every year, industries around the world spend billions of dollars on plant maintenance processes; it has been documented that maintenance expenses can account for up to a third of production expenses [
IRT has been considered as one of the most emanating CM techniques having numerous applications. IRT has been used in civil construction [
The experimental methodology utilized during the fault diagnosis is shown in Figure
Framework of proposed methodology adopted for bearing fault diagnosis.
The experimentation was carried out on a bearing test rig having a single phase 2 HP, 2 poles, and 220 V induction motor with diverse bearing conditions, and the FLIR E60 thermal camera was utilized to capture the thermal images for fault diagnosis of bearings in rotating machines. The setup used for the experimentation work is shown in Figure
Experimental setup.
Specifications of the FLIR E60 thermal camera.
Parameter | Description |
---|---|
IR resolution (array size) | 76,800 (320 × 240) |
Thermal sensitivity | <0.05°C (50 mK) |
Field of view | 25° × 19 °; optional lenses available |
Temperature range | −20 to 650°C (−4 to 1202°F) |
Accuracy | ±2% |
Image frequency | 60 Hz |
Focus | Manual |
Display screen | 3.5″ landscape touchscreen (widescreen) |
Visual camera | 3.1 MP |
Digital zoom | 4X continuous |
Image fusion/picture-in-picture | Yes/scalable |
Specifications of the induction motor.
Parameter | Description |
---|---|
Brand/model | Bonvario/BM 90S-2 |
Phase/pole | 1/2 |
Frequency/rated voltage | 50 Hz/220 V |
Protecting grade | IP-55 |
Ambient temperature | 45°C |
The type of bearing used in the experimentation work along with the specifications is presented in Table
Specifications of 1205 EKTN9 SKF self-aligning ball bearing.
Type of bearing | Outer diameter (mm) | Inner diameter (mm) | Magnitude of balls | Pitch diameter (mm) | Width (mm) |
---|---|---|---|---|---|
1205 EKTN9 SKF self-aligning ball bearing | 52 | 25 | 26 | 44.60 | 15 |
Different bearing states, namely, healthy, inner race, and outer race.
The real-time thermal images of the three bearing states captured from the thermal camera are presented in Figure
Raw thermal images of different bearing states: (a) healthy; (b) inner race fault; (c) outer race fault.
Representation of thermal images at different load and shaft speed of three bearing conditions.
Bearing conditions | No. of dataset | Total dataset | |||||||||
---|---|---|---|---|---|---|---|---|---|---|---|
At no load | At 2 kg load | At 4 kg load | |||||||||
500 rpm | 750 rpm | 1000 rpm | 500 rpm | 750 rpm | 1000 rpm | 500 rpm | 750 rpm | 1000 rpm | |||
HB | 6 | 6 | 6 | 6 | 6 | 6 | 6 | 6 | 6 | 162 images | 0.65 to 0.77 |
IRF | 6 | 6 | 6 | 6 | 6 | 6 | 6 | 6 | 6 | ||
ORF | 6 | 6 | 6 | 6 | 6 | 6 | 6 | 6 | 6 |
Image processing is a crucial step in fault diagnosis as it converts the images acquired from the thermal camera into digital form to extract some useful information from it. In this article, three different bearing conditions were taken into consideration and the acquired thermal images of these bearings contain immense noise and higher dimensionality data; therefore, their decomposition or denoising becomes a crucial step. For this purpose, discrete wavelet transform (DWT) has been applied in this article. The real-time decomposition of the original image into the denoised image using DWT is shown in Figure
Histogram of raw and decomposed thermal images using DWT.
DWT has been accepted as one of the outstanding approaches for the decomposition of thermal images [
Sub-bands block diagram of 2D-DWT.
Here, ↓ and
Extracting the features from the procured thermal images is an essential step of fault diagnosis. The various attributes such as texture, pixel, and region of interest can be obtained through the features of the input image. In this article total, eleven statistical features have been extracted for further processing. These features include mean, standard deviation, variance, skewness, kurtosis, median, energy, correlation, entropy, contrast, and homogeneity. These features have been extracted from thermal images for different healthy and faulty bearing conditions. These sets of features will serve as an input for the feature selection stage which is a very crucial stage in the fault diagnosis of bearings.
Selecting the most appropriate features from a complete set of features reduces the computation as well as enhances the classification accuracy. In the present work, PCA has been utilized as the feature selection technique for the selection of appropriate features. From the eleven statistical features, only six most relevant features have been selected by using PCA, namely, mean, standard deviation, entropy, kurtosis, skewness, and energy. These relevant features will serve as an input for the classification output.
For the fault classification, the set of appropriate features were trained to three different classifiers, namely, LDA, KNN, and SVM, and the accuracy of all these classifiers was compared for the performance evaluation. LDA was not only used as the classifier but it was a well-known method used for feature dimension reduction. With the assistance of a linear transformation matrix, LDA projects the features from parametric to feature space. LDA can even be computed for a set of large data. In the present work, KNN was also utilized for fault classification. KNN is a supervised learning algorithm that works on the principle of storing the data during training, and based on the similarity feature, it classifies any new data to that category whose features are quite similar to the new data. Further, another supervised technique named SVM was used for solving the classification problems. In contrast to other machine learning techniques, SVM proves to be more accurate and reliable especially for the classification problems associated with IRT [
This section presents the experimental results obtained from the thermal analysis of bearings for fault diagnosis. The current section presents the measurement of the response parameter which is the temperature of the region of interest for different bearing states. In addition to that this section also presents the classification and performance evaluation for different bearing conditions utilizing LDA, KNN, and SVM.
During experimentation, the temperature of the region of interest (bearing mounted at the free end of the shaft of the bearing test rig) was measured. The recorded temperature values for each bearing state considering different shaft speeds and load are shown in Table
Experimental design.
Run | Input parameters | Response parameter | ||
---|---|---|---|---|
Bearing fault | Rotational speed (rpm) | Load (kg) | Temperature (K) | |
1 | H | 500 | 0 | 291.2 |
2 | IRF | 500 | 0 | 294.7 |
3 | ORF | 500 | 0 | 297.7 |
4 | H | 750 | 0 | 293.5 |
5 | IRF | 750 | 0 | 296.4 |
6 | ORF | 750 | 0 | 299.6 |
7 | H | 1000 | 0 | 295.8 |
8 | IRF | 1000 | 0 | 298.2 |
9 | ORF | 1000 | 0 | 301.3 |
10 | H | 500 | 2 | 293.1 |
11 | IRF | 500 | 2 | 296.1 |
12 | ORF | 500 | 2 | 299.4 |
13 | H | 750 | 2 | 295.2 |
14 | IRF | 750 | 2 | 298.8 |
15 | ORF | 750 | 2 | 301.2 |
16 | H | 1000 | 2 | 298 |
17 | IRF | 1000 | 2 | 301 |
18 | ORF | 1000 | 2 | 304 |
19 | H | 500 | 4 | 297 |
20 | IRF | 500 | 4 | 300.1 |
21 | ORF | 500 | 4 | 303.4 |
22 | H | 750 | 4 | 299.2 |
23 | IRF | 750 | 4 | 302.3 |
24 | ORF | 750 | 4 | 305.2 |
25 | H | 1000 | 4 | 302 |
26 | IRF | 1000 | 4 | 305 |
27 | ORF | 1000 | 4 | 308 |
The thermal performance curves for each bearing state at different shaft speeds and different loads are shown in Figures
Thermal performance curves of different bearing states at different load and shaft speed. (a) Thermal performance curve at no load and 500 rpm. (b) Thermal performance curve at no load and 750 rpm. (c) Thermal performance curve at no load and 1000 rpm. (d) Thermal performance curve at 2 kg load and 500 rpm. (e) Thermal performance curve at 2 kg load and 750 rpm. (f) Thermal performance curve at 2 kg load and 1000 rpm. (g) Thermal performance curve at 4 kg load and 500 rpm. (h) Thermal performance curve at 4 kg load and 750 rpm. (i) Thermal performance curve at 4 kg load and 1000 rpm.
At no load and 500 rpm, there is a slight increment in the temperature for the healthy bearing state which means the rise in temperature at 500 rpm in comparison to the room temperature is less. For both the faulty bearings states, the difference in the temperature is more as compared with healthy bearing state which signifies that fault in the bearing leads to more heat generation and in turn the temperature rise. As the load increases to 2 and 4 kg, the rise in the temperature is more for faulty bearings because as the shaft load increases the load on the motor also increases which in turn increases the temperature of the bearings which was recorded by the thermal camera.
In image processing, entropy and energy provide a measure of how the pixel values are distributed along with the gray level range. Usually, the image with few gray levels will have higher energy than the others with many gray levels. Contrast is the difference in luminance or color that makes an object distinguishable. In statistics, homogeneity and, its opposite, heterogeneity, arise in describing the properties of a dataset or several datasets. They relate to the validity of the often convenient assumption that the statistical properties of any one part of an overall dataset are the same as any other part whereas standard deviation gives important information about the contrast of the image. 2D feature space of different selected features such as contrast, homogeneity, mean, kurtosis, entropy, and the standard deviation is shown in Figure
Scatter representation of extracted features at different operating conditions. Scatter plot of (a) contrast vs. STD; (b) homogeneity vs. STD; (c) mean vs. STD; (d) Kurtosis vs. STD; (e) energy vs. STD.
Standard deviation is taken as the reference among all the extracted features because it is a measure of variability or diversity used in statistics. In terms of image processing, it shows how much variation exists from the average value. A low standard deviation indicates that the data points tend to be very close to the mean whereas a high standard deviation indicates the data points are spread out over a very large range of values.
The current section presents the results for diverse bearing conditions utilizing LDA, KNN, and SVM. The results obtained through these classifiers are presented in the form of a matrix known as confusion or error matrix. The confusion matrix describes the performance of each classifier for each bearing condition in the form of rows and columns. Rows in the matrix represent the predicted class whereas columns refer to the true class. The confusion matrix of the LDA classifier for diverse bearing states at diverse load and shaft speed is presented in Table
Confusion matrix of three bearing states at different rpm using LDA.
Decision | Predicted class | ||||||||
---|---|---|---|---|---|---|---|---|---|
500 rpm | 750 rpm | 1000 rpm | |||||||
HB | IRF | ORF | HB | IRF | ORF | HB | IRF | ORF | |
True class | |||||||||
No load | |||||||||
HB | 6 | 0 | 0 | 6 | 0 | 0 | 6 | 0 | 0 |
IRF | 0 | 6 | 0 | 0 | 6 | 0 | 1 | 5 | 0 |
ORF | 0 | 1 | 5 | 0 | 1 | 5 | 0 | 1 | 5 |
2 kg load | |||||||||
HB | 6 | 0 | 0 | 6 | 0 | 0 | 6 | 0 | 0 |
IRF | 1 | 5 | 0 | 1 | 5 | 0 | 0 | 4 | 2 |
ORF | 0 | 1 | 5 | 0 | 1 | 5 | 0 | 1 | 5 |
4 kg load | |||||||||
HB | 6 | 0 | 0 | 6 | 0 | 0 | 6 | 0 | 0 |
IRF | 1 | 5 | 0 | 0 | 4 | 2 | 0 | 6 | 0 |
ORF | 0 | 1 | 5 | 0 | 1 | 5 | 0 | 3 | 3 |
The confusion matrix obtained by using KNN as a classifier for diverse bearing states at diverse load and shaft speed is presented in Table
Confusion matrix of three bearing states at different rpm using KNN.
Decision | Predicted class | ||||||||
---|---|---|---|---|---|---|---|---|---|
500 rpm | 750 rpm | 1000 rpm | |||||||
HB | IRF | ORF | HB | IRF | ORF | HB | IRF | ORF | |
True class | |||||||||
No load | |||||||||
HB | 6 | 0 | 0 | 6 | 0 | 0 | 6 | 0 | 0 |
IRF | 0 | 5 | 1 | 0 | 5 | 1 | 0 | 6 | 0 |
ORF | 0 | 0 | 6 | 0 | 0 | 6 | 0 | 1 | 5 |
2 kg load | |||||||||
HB | 5 | 1 | 0 | 6 | 0 | 0 | 4 | 1 | 1 |
IRF | 1 | 5 | 0 | 0 | 4 | 2 | 0 | 6 | 0 |
ORF | 0 | 0 | 6 | 0 | 1 | 5 | 0 | 2 | 4 |
4 kg load | |||||||||
HB | 6 | 0 | 0 | 6 | 0 | 0 | 6 | 0 | 0 |
IRF | 0 | 4 | 2 | 0 | 4 | 2 | 0 | 4 | 2 |
ORF | 0 | 1 | 5 | 0 | 2 | 4 | 0 | 2 | 4 |
Confusion matrix of three bearing states at different rpm using SVM.
Decision | Predicted class | ||||||||
---|---|---|---|---|---|---|---|---|---|
500 rpm | 750 rpm | 1000 rpm | |||||||
HB | IRF | ORF | HB | IRF | ORF | HB | IRF | ORF | |
True class | |||||||||
No load | |||||||||
HB | 6 | 0 | 0 | 6 | 0 | 0 | 6 | 0 | 0 |
IRF | 0 | 6 | 0 | 0 | 6 | 0 | 0 | 5 | 1 |
ORF | 0 | 0 | 6 | 0 | 0 | 6 | 0 | 0 | 6 |
2 kg load | |||||||||
HB | 6 | 0 | 0 | 6 | 0 | 0 | 6 | 0 | 0 |
IRF | 0 | 6 | 0 | 0 | 5 | 1 | 1 | 5 | 0 |
ORF | 0 | 0 | 6 | 0 | 0 | 6 | 0 | 1 | 5 |
4 kg load | |||||||||
HB | 6 | 0 | 0 | 6 | 0 | 0 | 5 | 1 | 0 |
IRF | 0 | 6 | 0 | 1 | 5 | 0 | 1 | 5 | 0 |
ORF | 0 | 1 | 5 | 0 | 1 | 5 | 0 | 0 | 6 |
It has been cleared from the results that SVM outperformed LDA and KNN in every aspect for bearing fault classification. An overall summary of fault diagnosis of various faults using IRT is presented in Table
Comparative analysis of present work with similar research work using IRT.
References | Camera used | Fault examined | Image preprocessing | Classifiers | Accuracy (%) |
---|---|---|---|---|---|
Li et al. [ | Fluke-Ti32 | Unbalance, outer race fault, ball fault, inner race fault | Region selection | CNN | 98.59 |
Huo et al. [ | FLIR-A35 | Outer race fault, inner race fault | 2D-DWT | Naive Bayes, SVM | 91.67 |
90.67 | |||||
Glowacz et al. [ | FLIR-E4 | Broken rotor bar | MoASoID | NN, K-mean | 100 |
100 | |||||
Janssens et al. [ | FLIR-SC655 | Rotor imbalance, bearing fault, lubrication | Windowing and subsampling | SVM | 88.25 |
Lim et al. [ | FLIR-SC5000 | Unbalance, ball bearing fault, shaft misalignment | 2D-DWT | SVM | 96.25 |
Eftekhari et al. [ | FLIR-I60 | Stator winding inner turn fault | Single Gaussian model | — | — |
Younus et al. [ | FLIR-A40 | Shaft misalignment, bearing fault, unbalance | 2D-DWT | LDA, SVM | — |
Younus et al. [ | FLIR-A40 | Shaft misalignment, bearing fault | Image segmentation | SVM | — |
Zhiyi et al. [ | FLIR-A35 | Outer race fault, ball fault, inner race fault | — | CNN | 98.26 |
Glowacz et al. [ | FLIR-E4 | Faulty fan, damaged gear-train | BCAoID | NN, BNN | 100 |
97.91 | |||||
Present work | FLIR-E60 | Inner race fault, outer race fault | 2D-DWT | LDA, KNN, SVM | 94.4 |
88.9 | |||||
100 |
MoASoID: method of area selection of image difference; BCAoID: binarized common areas of image differences; SR: softmax regression; CNN: convolutional neural network; NN: nearest neighbor; BNN: back propagation neural network.
The present work proposed an intelligent IRT-based system for fault classification of distinct bearing states. The acquired thermal images of distinct bearing conditions were initially preprocessed utilizing 2D-DWT accompanied by selecting the most appropriate features through PCA which further helps in classification and performance evaluation done through LDA, KNN, and SVM wherein SVM outperformed both LDA and KNN. The main outcomes obtained from the present work are as follows: DWT gives diverse resolutions at diverse frequencies while analyzing the signal which makes DWT a better approach for the decomposition of the thermal images PCA has been applied for the selection of the most relevant features among the set of features which in turn reduces the computation and enhances the classification accuracy Fault classification was done by using three classifiers, namely, LDA, KNN, and SVM among which SVM outperformed both LDA and KNN The present research work using the IRT approach for fault diagnosis is well compared with that of the various approaches utilized by the researchers
The classification outcomes proclaim that the present scheme could be utilized for detecting and inspecting the rotating machine faults and their condition. A multisensor-based approach combining a thermal camera and accelerometer or acoustic emission sensor can be considered for studying different behavior of rotating components. ANSYS tools can also be used to investigate the thermal behavior of bearings in greater depth.
The data used to support the findings of this study are available from the corresponding author upon request.
The authors declare that they have no conflicts of interest.
Condition monitoring
Nondestructive testing
Acoustic emission
Infrared thermography
2-dimensional discrete wavelet transform
Artificial neural network
Resistance temperature detector
Method of areas selection of image differences
Binarized common area of image differences
Nearest neighbor classifier
Continuous wavelet transform
Support vector machine
Independent component analysis
Principal component analysis
Mahalanobis distance
Standard deviation
Region of interest
Linear discriminant analysis
Short-term Fourier transform
Fast Fourier transform
Motor current signature analysis
Back propagation neural network
Speeded up robust features
Gaussian mixture model
Bag-of-visual word
Scale invariant feature transform
Convolutional neural network.