Experimental Study of Condition Monitoring of Conical Journal Bearing using Machine Learning Techniques

1

VEERMATA JIJABAI TECHNOLOGICAL INSTITUTE

Presented by- Harshal Bhat Sagar Nikam Rahul Yadav Sarang Bhudhar

Under the Guidance of:

Dr. Vikas M. Phalle

181020011

181020054

191020901

191020907

​

CONTENTS

• Introduction
• Literature Review
• Conical Journal Bearing & Different types of Defects
• The Test Rig
• Testing & Vibration Signals Recording
• Signal Processing & Fast Fourier Transform
• Introduction of Machine Learning
• Support Vector Machine
• Convolutional Neural Network (CNN)
• Comparison between HB & SB
• SVM & CNN Results
• Conclusion

2

Experimental Study of Condition Monitoring of Conical Journal Bearing using Machine Learning Techniques

Introduction

3

Fig. pressure distribution

• A journal bearing is the most common type of hydrodynamic bearing. The construction of the journal bearing has a circular shaft called the journal. The journal is made to rotate inside a fixed sleeve which is called the bearing. It supports high loads due the wedging action between journal and bearing because of considerable relative motion

• The bearing performance is highly dependent on the type of lubrication occurring and the viscosity.
• The type of lubrication can be classified as boundary lubrication, mixed film lubrication and full lubrication.

• The CJBis proposed to sustain radial and adjust the effect of axial load on rotating elements.
• Due to its low cost, compact size and application in axial and radial load condition, the journal and thrust hydrodynamic bearing have been replaced by conical bearing in the precision spindle of rotating machines

Experimental Study of Condition Monitoring of Conical Journal Bearing using Machine Learning Techniques

4

Need of Condition Monitoring

Fig. Condition Monitoring

• Early detection of the symptoms of faults that lead to serious journal bearing damages
• To allow remedial action to be taken to prevent catastrophic severe failures.
• To decrease possible loss of production due to breakdown of the machine caused due to major problems not detected earlier.

​

• Lubricating oil analysis (fluid property, fluid contamination and wear debris analysis)
• Vibro-acoustic analysis (vibration, vibro-acoustic and acoustic emission analysis)

​

• Bearing performance analysis (temperature, pressure, and thermography analysis)

Various Monitoring Techniques

Experimental Study of Condition Monitoring of Conical Journal Bearing using Machine Learning Techniques

Literature Review

5

Experimental Study of Condition Monitoring of Conical Journal Bearing using Machine Learning Techniques

6

Experimental Study of Condition Monitoring of Conical Journal Bearing using Machine Learning Techniques

Different types of Defects

7

Experimental Study of Condition Monitoring of Conical Journal Bearing using Machine Learning Techniques

System Layout

8

Fig. Entire System Layout

Experimental Study of Condition Monitoring of Conical Journal Bearing using Machine Learning Techniques

The Test Setup

9

Experimental Study of Condition Monitoring of Conical Journal Bearing using Machine Learning Techniques

Fig. Technical Details of the CJB

Fig. Test Rig

Testing & Vibration Signals Recording

10

Fig. Accelerometer Orientation & Location

Table: Speed & Loading Conditions

In doing this, the following are specified:

1. The time period that can be viewed
2. The highest frequency that can be seen

Fig. Mounting of Accelerometer

1) mounted onto a surface that is free from oil and grease as close as possible to the source of vibration

​

2) sensors should be located on the motor and pump drive end bearings

Experimental Study of Condition Monitoring of Conical Journal Bearing using Machine Learning Techniques

Signal Processing & Fast Fourier Transform

11

Fig. Representative image of Vibration analysis

Fig. Frequency components & Amplitudes

Fig. Complex Time Waveform Components

Fig. Complex Time Waveform

Fig. Frequency Bands

Experimental Study of Condition Monitoring of Conical Journal Bearing using Machine Learning Techniques

Spectral Kurtosis and Kurtogram

12

Fig. Spectral Kurtosis & Kurtogram (example)

Spectral kurtosis (SK) is a statistical tool that can indicate and pinpoint nonstationary or non-Gaussian behavior in the frequency domain, by taking:

• Small values at frequencies where stationary Gaussian noise only is present
• High positive values at frequencies where transients occur

Signal

Window Function

Short Time Fourier Transform of the signal, S(t, f)

The kurtogram shows kurtosis results for a range of window lengths and frequencies.

Experimental Study of Condition Monitoring of Conical Journal Bearing using Machine Learning Techniques

13

Fig:Healthy Bearing at 500 rpm & 750 N Radial Load

Fig:Faulty Bearing at 500 rpm & 750 N Radial Load

Experimental Study of Condition Monitoring of Conical Journal Bearing using Machine Learning Techniques

Functions used in MATLAB:

X = Track3(1:2560)- mean(Track3(1:2560))

kurtosis(X)

kurtogram (X, Fs)

​

​

Introduction to Machine Learning in fault diagnosis

14

• Various machine learning techniques have been used for fault classification in anti-friction bearings, electric motors, pumps, gearboxes, agricultural machinery, turbines etc.
• Machine Learning models like ANN, KNN, SVM and CNN are widely used in fault classification for machineries
• Eg:Fault Diagnosis of Induction motor using SVM for detecting broken rotor bars

CNN

SVM

KNN

ANN

Figs:Architecture of Machine Learning Algorithms

Experimental Study of Condition Monitoring of Conical Journal Bearing using Machine Learning Techniques

Machine Learning Proposed Framework

15

Experimental Study of Condition Monitoring of Conical Journal Bearing using Machine Learning Techniques

Support Vector Machine (SVM)

16

Fig: SVM Process Flow Diagram

• Fault classification in small sample cases such as fault diagnosis
• Train Test Split 101 & 21 out of total 122 dataset
• 9 Input features and 2 classes are used in the model
• Cost Minimization function is:

J

Fig : Decision Boundary of an SVM

Experimental Study of Condition Monitoring of Conical Journal Bearing using Machine Learning Techniques

Tuning SVM Parameters

17

Experimental Study of Condition Monitoring of Conical Journal Bearing using Machine Learning Techniques

Fig: Underfitting

Fig: Overfitting

GridSearchCV - tuneSVM

1)function used for hyperparameter tuning

2) Iterates through all hyperparameters & provides those with best average score

3) C = 1, γ = 0.1, best average score = 81.09%

Fig:Good Fit

SVM Results

18

Confusion Matrix

Predicted Class

Healthy Bearing

Faulty Bearing

Healthy Bearing

Faulty Bearing

Experimental Study of Condition Monitoring of Conical Journal Bearing using Machine Learning Techniques

Actual Class

False Positive(FP): Type I Error 8 & 1

False Negative(FN): Type II Error 8 & 2

Comparison between Healthy & Scratched Bearing (Results)

19

Fig:Healthy Bearing at 1100 RPM &350N Radial load

Fig:Faulty Bearing at 1100 RPM &350N Radial load

Experimental Study of Condition Monitoring of Conical Journal Bearing using Machine Learning Techniques

Operating Frequency = (2*pi*N)/(60*2*pi)=18.33Hz

Healthy Bearing: Peaks are observed only at multiples of Operating Frequency, otherwise almost flat line

​

Faulty Bearing: For same operating frequency, other than multiples, noise portion is clearly visible

Comparison between Healthy & Scratched Bearing (Results)

20

Experimental Study of Condition Monitoring of Conical Journal Bearing using Machine Learning Techniques

Convolutional Neural Network (CNN)

​

21

Fig. CNN architecture

Experimental Study of Condition Monitoring of Conical Journal Bearing using Machine Learning Techniques

Fig. Convolutional Operation

Fig. Graphical Representation of CNN

• CNN was chosen as a classifier because of its superior capacity to learn 2D data. We choose a shallow CNN with 155,681 trainable parameters. The network was built to take 32✕32 px RGB pictures as input and forecast bearing conditions.
• Three convolutional (Conv) and three Max-Pooling layers made up the network design. The first layer's kernel size was set at 3✕3 and remained constant throughout the model with a stride of 1 has been considered for experimental purposes. The output of the 3rd Max-Pooling layer was flattened and fed through the dense, allowing the state of the conical journal bearing to be predicted.

​

CNN Parameters

22

Experimental Study of Condition Monitoring of Conical Journal Bearing using Machine Learning Techniques

Fig. ReLU activation

Fig. Flatten Layer

Fig. Backpropagation

• In order to use stochastic gradient descent with backpropagation of errors to train deep neural networks, an activation function is needed that looks and acts like a linear function, but is, in fact, a nonlinear function allowing complex relationships in the data to be learned. ReLU activation function is simple to compute and has predictable gradient .
• Adam optimizer is a combination of Momentum and Root mean squared propagation(RMSprop) optimizer which gives much higher performance than the previously used and outperforms them by a big margin into giving an optimized gradient descent.

​

Fig.Fully Connected Layers

CNN Results

​

​

​

​

​

​

​

23

Experimental Study of Condition Monitoring of Conical Journal Bearing using Machine Learning Techniques

Summary of Results

24

Experimental Study of Condition Monitoring of Conical Journal Bearing using Machine Learning Techniques

Fig. Precision Histogram

Table: Results summary

Fig. Recall Histogram

Fig. Training Accuracy Histogram

Fig. F1 Histogram

Conclusion

25

1. Significant differences of acceleration amplitudes between healthy and scratched conical bearing in case of variable speed and constant load conditions
2. Higher amplitudes are observed for scratched bearing over the frequency ranges in the amplitude spectrum, these are mainly due to following reasons: 
1. Variation in the pressure distribution caused by the presence of the scratches in the bearing consequently altering the excitation forces thereby increasing the amplitudes.
2. Lubrication mixing caused due to scratches in the bearing.
3. Operating acceleration amplitude ranges from 0.5 to 1.8 m/s^2. Maximum value observed for the scratched bearing is 1.8 m/s^2. & maximum acceleration amplitudes are observed in the radial direction of the bearing observed from the recorded signals.
4. For 500 RPM and 1050N conditions, the first amplitude spike occurs at 8.48 Hz i.e. operating frequency of the CJB and other spikes are observed at 2nd, 3rd and consequent multiples of the operating frequency.
5. Support Vector Machine (SVM) method is relatively more reliable than CNN method with efficient classification for a small dataset and less execution time i.e. 76.76 sec including training and prediction times.
6. Convolutional Neural Network(CNN) deep learning approach can be applied when the dataset is larger, contains noise and outlier data.

Experimental Study of Condition Monitoring of Conical Journal Bearing using Machine Learning Techniques

Future Scope

26

1. Higher sampling rates vibration signals can be obtained from OROS software to capture frequency characteristic with more precision and fed to CNN for more accuracy.
2. Condition monitoring methods are wide and not limited to vibration analysis. They can be coupled with acoustic emission, lubricating oil analysis,etc. can be studied and implemented on the CJB using the test rig.
3. Vibration behaviour of the bearing can be studied using different types of lubricating oils with different properties.
4. The analytic tools i.e. machine learning algorithms and artificial intelligence techniques will be shifted to cloud and the machine will be condition monitored real time using the cloud with data visualization and alarming features.
5. A IOT based application can be developed to transmit the data from sensors to app via the Internet.
6. To increase validation and test accuracy we can employ bigger ML models as show in figure in combination with K-Fold Cross Optimization.
7. Use CutMix and MixUp techniques to augment kurtogram images and enhance accuracy combining with Ensemble learning.

Experimental Study of Condition Monitoring of Conical Journal Bearing using Machine Learning Techniques

References

Analysis Of Hydrodynamic Journal Bearing: A Review. Priyanka Tiwari and Veerendra Kumar

[2] Childs, P. R. N. (2019). Journal bearings-Mechanical Design Engineering Handbook

[3] Performance analysis of conical hydrodynamic journal bearing. Ajay Kumar Gangrade, Vikas M. Phalle, S. S. Mantha.

[4] Wear In Hydrodynamic Journal Bearings: A Review, Harbansh Singh , Dr. S. S. Sen

[5] Harnoy A. Bearing design in machinery: engineering tribology and lubrication. CRC press; 2002

[6] Wear of hydrodynamic journal bearings. L. Rozeanu and F.E. Kennedy

[7]Scott R 2008 Basic wear modes in lubricated systems. Machinery Lubrication Noria Publication No. 1375

[8] B. Samanta*, K.R. Al-Balushi, S.A. Al-Araimi, Artificial neural networks and support vector machines with genetic algorithm for bearing fault detection

[9] Poley J 2007 The History of Oil Analysis, Machinery Lubrication Noria Publication No. 1113

[10] Fault diagnosis of ball bearings using machine learning methods. P.K. Kankar, Satish C. Sharma, S.P. Harsha

​

27

[11] Roylance B J 2003 Machine failure and its avoidance-what is tribology’s contribution to effective maintenance of critical machinery? Proc IMechE Volume 217 Part J Engineering Tribology J05302

[12] Cerrada M, Sánchez RV, Li C, Pacheco F, Cabrera D, de Oliveira JV, Vásquez RE (2018) A review on data-driven fault severity assessment in rolling bearings. Mech Syst Signal Process 99:169–196

[13] El-Thalji I, Jantunen E (2015) A summary of fault modeling and predictive health monitoring of rolling element bearings. Mech Syst Signal Process 60:252–272

[14] Kankar P, Sharma SC, Harsha S (2011) Fault diagnosis of ball bearings using machine learning methods. Expert Syst Appl 38(3):1876– 1886

[15] Bellini A, Filippetti F, Tassoni C, Capolino GA (2008b) Advances in diagnostic techniques for induction machines. IEEE Trans Ind Electron 55(12):4109–4126

[16] Dai X, Gao Z (2013) From model, signal to knowledge: a data-driven perspective of fault detection and diagnosis. IEEE Trans Ind Inform 9(4):2226–2238

[17] Journal Bearing Fault Detection Based on Daubechies Wavelet Narendiranath Babu T.(1), Himamshu H.S.(1), Prabin Kumar N.(1) Rama Prabha D.(2), Nishant C.(1)

[18]An Improved GA-SVM Algorithm WEI Chen, Yuan Hui-mei

[19] Artificial neural networks and support vector machines with genetic algorithm for bearing fault detection B. Samanta*, K.R. Al-Balushi, S.A. Al-Araimi

​

Experimental Study of Condition Monitoring of Conical Journal Bearing using Machine Learning Techniques

THANK YOU

28