Fault Diagnosis of Rolling Element Bearing using Acoustic Condition Monitoring and Artificial Neural Network Technique

1. Lt Cdr Jaskaran Singh
19RE61D02
Subir Chowdhury School of Quality and Reliability
IIT Kharagpur
Fault Diagnosis of Rolling Element Bearings using
Acoustic Condition Monitoring and Artificial Neural
Network Technique

2. Conclusions and Future Scope
Results and Discussions
Combined Acoustic Emission with Neural Network Model
Feature Selection
Experimental setup
Introduction
Contents
09-02-2021
2

3. • Rolling element bearings - critical components of rotating machines.
• Defective bearings source of vibration signals - same utilised to assess faulty bearings.
• Acoustic Emission (AE) - complementary method for bearing condition monitoring - very
sensitive to incipient defects.
• AE - stress wave emission (transient elastic wave) in materials - can be detected by
transducers placed on it.
• AE monitoring can detect :-
 Growth of subsurface cracks.
 Signals between 100 kHz to 1 MHz in frequency.
 Low frequency problems generated by fatigue cracks, incipient damage.
• Limitations of AE technique - Difficulty in processing, interpreting and classifying the
acquired data.
Introduction
09-02-2021
3

4. • Artificial neural networks (ANN) - Interconnected network of models based on
biological learning processes of human brain.
• ANN - self-possessed large number of artificial neurons working simultaneously to
solve a specific problem.
• Neural network – adaptive system – changes network architecture based on
information flowing through it.
• Two important factors in ANN - Training-and-learning.
• Current problem employs Multilayer Feed Forward Back Propagation (MLP)
architecture.
• MLP most commonly used and successfully applied architecture.
Introduction
09-02-2021
4

5. Experimental Setup
Figure1. Experimental test set-up 09-02-2021
5

6. Experimental Setup
09-02-2021
6
Figure 2. NTN N312 Cylindrical
roller bearing and specs
Figure 3. Defect cutting on wire EDM

7. Experimental Setup
09-02-2021
7
Figure 4. Outer race seeded line
defect (0.5 mm width)
Figure 5. Test bearing

8. Experimental Setup
09-02-2021
8
Figures 6 and 7. Bearing Test Rig without and with Probes

9. • Statistical methods employed for physically characterizing time and frequency domain data.
• Different descriptive statistical parameters selected for the study:-
 Root Mean Square (RMS)
 Peak value (Pv)
 Crest Factor (CrF)
 Skewness
 Kurtosis
 Clearance Factor (ClF)
Feature Selection
09-02-2021
9

10. • Total 144 test runs conducted – bearing damage in incipient stage detected.
• Defect size couldn’t be predicted.
• ANN learnt behavior of a specific fault in bearing to correlate obtained AE values
with given parameters to defect size.
• ANN model constructed with three layers - input layer, output layer and one hidden
layer.
• Learning of neural network done with feed forward back propagation algorithm.
• Neural network trained with 53 samples and validated with 6 samples.
• Learning stopped after 25000 cycles with average training error less than 0.01.
• Post training the network, fault size predicted at required features.
Combined AE with Neural Network Model
09-02-2021
10

11. Combined AE with Neural Network Model
09-02-2021
11
Figure 8. ANN Topology
(9-10-1)

12. Results – AE Time Wave and Frequency Spectrum
09-02-2021
12
• Figure 9. (a) Time wave
• Figure 9. (b) Enlarged time
wave
• Figure 9. (c) Frequency
spectrum

13. Results – Defect Frequencies and Test Program
09-02-2021
13
RPM FTF (Hz) BSF (Hz) IRF (Hz) ORF (Hz)
500 3.38 21.43 59.35 40.61
700 4.74 30.03 83.15 56.89
900 6.09 38.59 106.88 73.13
1100 7.45 47.16 130.60 89.36
1300 8.80 55.76 154.40 105.64
1500 10.16 64.32 178.13 121.98
Load (kN) Defect Size Width (mm) Speed (RPM)
L1 = 2 D1 = 0.3 N1 = 500
L2 = 4 D2 = 0.5 N2 = 700
D3 = 0.7 N3 = 900
D4 = 0.9 N4 = 1100
D5 = 1.1 N5 = 1300
N6 = 1500

14. Results – Comparison of ORF at D2 (0.5mm) and L2 (4kN)
09-02-2021
14
RPM ORF (Hz)
Theoretical
dB Level ORF (Hz)
Actual
Difference in
ORF
500 40.61 21.43 40.404 0.206
700 56.89 30.03 56.608 0.282
900 73.13 38.59 72.224 0.906
1100 89.36 47.16 88.328 1.032
1300 105.64 55.76 104.342 1.298
1500 121.98 64.32 120.536 1.344

15. Experimental and Predicted values of defect size in testing (AE)
09-02-2021
15
S.
No.
Loa
d
(kN)
RPM Amplit
ude
(dB
level)
RMS Peak
Value
Crest
Facto
r
Skewn
ess
Kurtos
is
Cleara
nce
Factor
Seeded
defect
size
(mm)
Predict
ed
defect
size
(mm)
%
error
1 2 500 1.8 0.89 0.27 0.33 0.0004 0.0001 0.05 0.5 0.5343 6.42
2 2 900 3.75 0.99 0.27 0.28 0.0007 0.0002 0.08 0.7 0.6967 0.47
3 4 900 4.32 0.95 0.23 0.25 0.0005 0.0002 0.07 0.5 0.6846 26.96
4 4 500 3.04 0.84 0.3 0.36 0.0004 0.0002 0.07 0.9 0.8942 0.64
5 4 1500 11.8 1.26 0.31 0.25 0.0009 0.0003 0.13 1.1 1.0999 0.01
Average of % error 6.90

16. • Small defect width size, negligible change in ORF as:-
 Rolling element easily rolls over.
 Force exerted over defect edge very less.
 Little disturbance/ stress in defect area of outer race.
• As defect width increases, peaks at ORF in AE frequency spectra
observed as:-
 Fault edge obstructs rolling motion.
 Greater change in momentum leads to greater impact.
 Increased stress over defect area.
 AE probe captures change in energy as stress waves.
Discussions
09-02-2021
16

17. Discussions
09-02-2021
17
 Calculated % error of 6.90% proves closeness of predicted values to experimental values.
 Shows reliability of proposed network model in predicting bearing defect size in given conditions.
Figure10. Defect size
comparison with actual vs.
ANN predicted values

18. Conclusions and Future Scope
09-02-2021
18
 Combination of ANN and AE can be effectively utilized in predicting defect size.
 Increase in rise of peaks of ORF frequency spectra observed at increased defect width size.
 Apart from spectral analysis, statistical features of AE proposed to classify and quantify severity
levels.
 This work is extendable to bearings of different shapes and sizes such as spherical, tapered
bearings, etc.
 Bearing performance at different speeds and load conditions can be monitored.
 Showcases reliability of proposed network model in predicting bearing defect size in given
conditions.

19. References
09-02-2021
19
 Mary Jasmin Nerella, Ch. Ratnam, V. Vital Rao, “Fault Diagnosis of a Rolling Element
Bearings using Acoustic Condition Monitoring And Artificial Neural Network Technique” International
Research Journal of Engineering and Technology, Vol. 5, Issue 03, March 2018.
 C.Senthilraja, L.Vinoth, “A Review on Fault Diagnosis of Ball Bearing Using Sound and Vibration
Signals” International Journal of Innovative Research in Science, Engineering and Technology, Vol. 4,
Special Issue 13, December 2015.
 V.V. Rao , Ch. Ratnam , “A Comparative Experimental Study on Identification of Defect
Severity in Rolling Element Bearings using Acoustic Emission and Vibration Analysis”, Vol. 37, No. 2 ,
176-185, (2015).
 Vana Vital Rao , Chanamala Ratnam, Estimation of Defect Severity in Rolling Element
Bearings using Vibration Signals with Artificial Neural Network, Jordan Journal of Mechanical and
Industrial Engineering, Volume 9 Number 2, Pages 113 – 120 ,April.2015.

20. 09-02-2021
20