USE OF BARKHAUSEN NOISE IN INSPECTION OF THE
SURFACE CONDITION OF STEEL COMPONENTS
1
Aki Sorsa
16.1.2014
16.1.2014
2
CONTENTS
• Background
• Barkhausen noise ◦ Origin
◦ Literature
◦ Applications
• BN Studies ◦ Research problem
◦ Approach
◦ Results
◦ Conclusions
• Summary
16.1.2014
3
BACKGROUND
• Material properties can be measured destructively or
non-destructively
• Destructive methods are not applicable to quality
control ◦ non-destructive methods are preferred
• Non-destructive testing methods: visual inspection,
ultrasonics, acoustic emission, magnetic methods etc.
• Barkhausen noise (BN) is an intriguing technique for
ferromagnetic materials
◦ Fast, low costs, simple equipment
16.1.2014
4
BARKHAUSEN NOISE ORIGIN
• The specimen is placed in an external, varying magnetic
field ◦ Magnetic domain wall movements
◦ The walls get trapped behind pinning sites
◦ Rapid and stochastic movements caused by walls
breaking out of the pinning sites
• Rapid movements of the domain
walls cause a noise-like signal
• Wall movements are influenced
by material properties
16.1.2014
5
BARKHAUSEN NOISE SIGNAL
-300
-200
-100
0
100
200
300
-1500
-1000
-500
0
500
1000
1500
0 0.2 0.4 0.6 0.8 1
Ap
pli
ed m
ag
net
ic f
ield
Ba
rkh
au
sen
no
ise
Time (relative)
Barkhausen noise Applied magnetic field
16.1.2014
6
BARKHAUSEN NOISE LITERATURE
• BN has been shown to be very sensitive to
microstructure and material properties: residual stress,
hardness, etc.
• Some feature is calculated from the BN signal and
compared to the material property studied ◦ RMS, peak height, width and position
• Results are usually only
qualitative
Figure from Sorsa (2013)
16.1.2014
7
BARKHAUSEN NOISE APPLICATIONS
• Material characterisation ◦ Quality control
• Case depth evaluation ◦ Remaining layer thickess
• Grinding burn detection ◦ Soft spot detection
16.1.2014
8
BN STUDIES RESEARCH PROBLEM
• Changes in material properties cumulate to the BN
signal. ◦ How to distinguish the influence of different factors?
• Interactions between material properties and BN are
complex.
• Stochastic phenomenon ◦ Only averaged properties are reproducible.
• Indirect measurement ◦ Models are needed
◦ Significance of calculated features?
16.1.2014
9
BN STUDIES APPROACH
• Mathematical models used for describing the
interactions between material properties and BN ◦ Residual stress, hardness
• Identification of the model divided into 4 steps ◦ Feature generation
◦ Feature selection
◦ Model identification
◦ Model validation
16.1.2014
10
BN STUDIES APPROACH: STEP 1
• Feature generation ◦ BN signal is useless by itself
◦ Information needs to be converted into useful form
◦ Calculation of features with different mathamatical
operations - Statistical
- RMS, BN energy and entropy
- Factors
- Features from BN profile
◦ Produces a big set of features (about 150)
16.1.2014
11
BN STUDIES APPROACH: STEP 2
• Feature selection ◦ The most significant features are case-dependent
◦ Automatic procedures are needed
◦ Deterministic methods - Stepwise selection (forward-selection, backward-
elimination and their combinations)
- May not lead to optimal solution
◦ Stochastic methods - Simulated annealing, genetic algorithms
- Reported to give better results
16.1.2014
12
BN STUDIES APPROACH: STEP 3
• Model identification ◦ Usually carried out simultaneously with feature selection
• Different model structures can be used ◦ MLR, PLSR, PCR, ANN …
• Model is identified with the training data set
• In BN studies, the number of data points is limited ◦ Cross-validation methods are used (efficient data usage)
16.1.2014
13
BN STUDIES APPROACH: STEP 4
• Model validation ◦ Prediction models are useless if they are valid only for
the data they were trained with.
◦ Models must be validated with an independent testing
data set
◦ Validation should also include validation of the selected
features with expert knowledge
16.1.2014
14
BN STUDIES RESULTS
Study Feature selection Modelling technique Reference
1 Manual MLR Sorsa et al. (2012a)
2 Forward-selection MLR Sorsa et al. (2012b)
3 GA MLR Sorsa et al. (2013a)
4 Preselection + GA Nonlinear regression Sorsa et al. (2014)
5 Preselection +
exhaustive
ANN Sorsa et al. (2013b)
• Material characterisation: prediction of residual stress
• All studies carried out in cooperation with the
Department of Materials Science, Tampere University of
technology
16.1.2014
15
BN STUDIES RESULTS: STUDY 1
-600
-400
-200
0
200
400
600
800
-600 -400 -200 0 200 400 600 800
Pre
dic
ted r
esi
dual
stre
ss
Measured residual stress
Data set 1 Data set 2 Perfect fit
Training Validation
R 0.85 0.91
RMSEP 57.68 MPa 139.37 MPa
16.1.2014
16
BN STUDIES RESULTS: STUDY 2
Training Validation
R 0.87 0.94
RMSEP 53.11 MPa 111.82 MPa
-600
-400
-200
0
200
400
600
800
-600 -400 -200 0 200 400 600 800
Pre
dic
ted r
esi
dual st
ress
[M
Pa]
Measured residual stress [MPa]
Data set 1 Data set 2 Perfect fit
16.1.2014
17
BN STUDIES RESULTS: STUDY 3
Training Validation
R 0.95 0.96
RMSEP 75.62 MPa 93.80 MPa
-1000
-800
-600
-400
-200
0
200
-1000 -800 -600 -400 -200 0 200
Pre
dic
ted
re
sid
ua
l s
tre
ss
[M
Pa
]
Measured residual stress [MPa]
Training External validation Perfect fitb)
16.1.2014
18
BN STUDIES RESULTS: STUDY 4
Training Validation
R 0.96 0.92
RMSEP 51.95 MPa 91.16 MPa
16.1.2014
19
BN STUDIES RESULTS: STUDY 5
Training Validation
R 0.88 0.93
RMSEP 51.1 MPa 45.8 MPa
-1000
-800
-600
-400
-200
0
-1000 -800 -600 -400 -200 0
Pre
dic
ted r
esi
dual st
ress
(M
Pa)
Measured residual stress (MPa)
training testingb) RBF
16.1.2014
20
BN STUDIES RESULTS: CONCLUSIONS 1/2
• Feature selection ◦ Manual selection
- Reasonable results but not as good as with other methods
- Features based on expert knowledge
◦ Deterministic methods - Efficient
- Good results but not as good as with stochastic methods
◦ Stochastic methods - Best results
- Computationally expensive
16.1.2014
21
BN STUDIES RESULTS: CONCLUSIONS 2/2
• Model structures ◦ Linear / MLR
- Not as good results as with nonlinear
- Robust
- Main interactions
- Computationally efficient
◦ Nonlinear regression
- Not as good results as with ANN
- Computationally expensive
- More robust than ANN
◦ ANN
- Best results
- Computationally expensive
- Risk of overfitting
16.1.2014
22
SUMMARY
• BN is a non-destructive testing method suitable for
ferromagnetic materials ◦ Sensitive to many material properties
◦ Indirect measurement models are needed
• Evaluation of material properties with mathematical
models ◦ 4 steps: feature generation, feature selection, model
identification, model validation
• Results illustrated with residual stress predictions ◦ Reasonable results
16.1.2014
23
REFERENCES
• Sorsa A, Ruusunen M, Leiviskä K, Santa-aho S, Vippola M and Lepistö T (2014) An attempt
to find an empirical model between Barkhausen noise and stress. Materials Science
Forum 768-769: 209-216.
• Sorsa A (2013) Prediction of material properties based on non-destructive Barkhausen
noise measurement. Doctoral thesis, University of Oulu Graduate School, Acta
Universitatis Ouluensis, 122p.
• Sorsa A, Leiviskä K, Santa-aho S, Vippola M and Lepistö T (2013a) An efficient procedure
for identifying the prediction model between residual stress and Barkhausen noise.
Journal of Nondestructive Evaluation 32(4): 341-349.
• Sorsa A, Santa-aho S, Vippola M, Lepistö T and Leiviskä K (2013b) A case study of using
radial basis function neural networks for predicting material properties from Barkhausen
noise signal. Proceedings of 18th Nordic Process Control Workshop, 6p.
• Sorsa A, Leiviskä K, Santa-aho S and Lepistö T (2012a) A data-based modelling scheme
for estimating residual stress from Barkhausen noise measurements. Insight - Non-
Destructive Testing and Condition Monitoring 54(5): 278-283.
• Sorsa A, Leiviskä K, Santa-aho S and Lepistö T (2012b) Quantitative prediction of
residual stress and hardness in case-hardened steel based on the Barkhausen noise
measurement. NDT&E International 46: 100-106.
16.1.2014
24
THANK YOU FOR
YOUR ATTENTION !!