You are on page 1of 10

Authors responses to Viva committee comments

Responses to the examiners comes are typed in red front.


The details (i.e chapter, section, page and lines) of the modifications in the revised
submission are also provided at the end of every response.
Modifications in the revised submission are also shown in highlighted by red font.

1. Please insert end of chapters summaries too.


Authors response:
The summary section of each chapter is inserted at the end of the corresponding chapter in the
revised submission
Please see the revised submission

Section 1.8, page 16.


Section 2.6, page 28.
Section 3.5, page 41.
Section 4.6, page 69.
Section 5.4, page 112.
Section 7.5, page 161.

2. Novelty- identify precisely what this is:


2.1. Healthy signal application of SSA-discuss.
Authors response: Modifications are made to the thesis in order to clarify novelty in subjecting only healthy signal to
the SSA. The author would like to say the following in brief:

The baseline space in all the levels of the fault diagnosis is made from subjecting the training
sample (corresponding to healthy category) to decomposition stage of the SSA. The features
are obtained from the norms of the PCs corresponding to this training sample. A threshold is
made from the features corresponding to this training sample.
Till the moment, none of damage classes are involved in building the baseline space or
determining the threshold and that is what we meant by using only healthy category signals.
Any new signal (healthy or faulty) is projected onto this baseline space and from the
projection; the corresponding feature vector is obtained. Then the Mahalanobis distance of
new feature vectors are measured to the baseline category. This Mahalanobis distance is
then compared to the predetermined threshold and a decision of assigning of the new signal
to baseline or non-baseline category is made.
Compared to previous studies discussed in section 3.4.1, none is used only signals from
healthy bearing category to build a baseline space. They usually subjecting all categories (i.e
Healthy and faulty) to SSA. In addition, healthy bearings signals are subjected only to the
decomposition stage which offers minimum transformation of signals.
The conclusions which we obtained in our work are:
o The Mahalanobis distances corresponding to the features from faulty category are
significantly deviated from those corresponding to baseline category.
1

The Mahalanobis distance levels (range) changes with the change of fault severities
and fault locations.
o Based on the two points mentioned, we claim that our methodology is not able to
distinguish not only among baseline and none-baseline categories but among the
none-baseline themselves. In other word, the methodology is able to detect faults
corresponding to different locations (such as inner race and outer race) and of
different severities such as small and large.
The author also makes modifications in the revised submission to clarify the points mentioned
above.
Please see in the revised submission, chapter 1, lines (2-17) in page 11 and, lines (2-9) page12
2.2. New way to use LTIVAR model discuss this too.
Authors response
The author apologizes for any confusion caused in this point of novelty. The author did not claim
that a new model was used because the LTIVAR model is already existed but it is not popular for
bearing fault diagnosis. The LTIVAR is suitable for representing stationary signals while bearing
vibration signals are almost always non-stationary. The author claims that a fresh direction of
autoregression based fault diagnosis technique is suggested. This includes a combination of an
advanced signal pretretment with the simplest form of autoregression (i.e LTIVAR). The advance
signal pretreatment includes combination of SSA-based densoing (which is used for the first time in
our current work) and differencing-based stationaristion (which also used for the first time in the
current thesis). This signal pretreatment make the signals amenable to be analysed by the simplest
form of autoregression (i.e LTIVAR). The combination of this advanced signal pretreatment and
LTIVAR model provides a high accuracy fault diagnosis methodology as shown in chapter 6 and 7.
Please see in the revised submission, chapter 1, starts from line 18 in page 12 till the 8th line in
page 13.

Compared to other autoregression-based signal, the current methodology is much simpler. This is
because the other forms of autoregression that researchers usually use in the literature are complex
and requires several assumptions. For example, using time-varying autoregressive (TVAR) models
has some challenges which can be affecting on the modelling preciseness. One of these challenges is
the need to describe a relationship for the models coefficients evolution over time. And we need
also to propose an initial set of models coefficients. If these assumptions were not precise enough,
the modelling process is incorrect. When we use LTIVAR model, neither the shape of models
coefficients evolution over the time nor the initial set of models coefficients are to be assumed.
Only what we need is to simplify the signal by de-nosing it and transforming it to stationary signal.
Please see in the revised submission, chapter 3, lines (4-7) and lines (12-20) in page 39.

2.3. Validation very wide range of test data expand.


Authors response
The author agrees that the description of the experimental work should be expanded and the variety
of the data sets should be clarified and that is what he modifies in the revision copy of the thesis.
The author has included a new section 4.5 to stress on the wide range of data used and its variety.
To clarify the variety of the data sets in brief, please see the below:

The three bearing test rigs cover a variety of bearing rotational speed range. For test bearing at
Case Western Reserve University the speed ranges from 1730 RPM to 1797 RPM. For test
bearing at Strathclyde university the speed ranges from 250 RPM (which is relatively low speed)
to 1250 RPM. For test bearing at and Politecnico di Torino university the speeds are very high
and range from 18000 RPM to 30000 RPM. For all of these speeds the methodology where
performing very well and shows a 100% correct classification for most of the cases.
The test bearings cover various fault severities which are detect by the methods. Some of these
fault severities were very small such as in the around 0.005 inch (around one tenth of a
millimetre).
The structure of the test bearings was not the same and the bearings used are different. For
example, the bearing used in the test rig at Politecnico di Torino is a special made rolling
element of a special dimension. While the other tests rigs used different ball bearing models.
Sampling rate of Politecnico di Torino is different to what was used in the other test bearing rigs.
The loads applied to the bearing are different in the entire test bearing rigs.

We believe that this variety of data sets specifications enables and support the generality of the
methodology. For the modifications done in the revised submission
Please see in the revised submission, chapter 4, a new section 4.5 in page 68.
3. Need to acknowledge you are using supervised learning.
Authors response
Modifications are made to the thesis in order clarify the confusion that arose. To reply to this
concern we would like to say the following: 1. The baseline space in all the levels of the fault diagnosis is made by subjecting the training
sample which is only made from healthy data to a decomposition stage of the SSA. The features
are obtained from the norms of the PCs corresponding to this training sample. A threshold is
made from the features corresponding to this training sample.
2. Till the moment, none of damage classes are involved in building the baseline space or
determining the threshold and that is what we meant by using only healthy category signals.
3. Any new signal (healthy or faulty) is projected onto this baseline space and from the projection,
features are obtained and the Mahalanobis distance is measured between the feature (i.e
feature vector) and the baseline category. This Mahalanobis distance is then compared to the
predetermined threshold and a decision of assigning of the new signal to baseline or nonbaseline category is made.
4. From point three we believe that the case is unsupervised learning as we did not involve the
information from fault category in assignment decision.
5. :From the above points one can deduct
3

5.1. The Mahalanobis distances corresponding to the features from faulty category are
significantly deviated from those corresponding to baseline category.
5.2. The Mahalanobis distance levels (range) change with the change of fault severities and fault
locations.
5.3. Based on points 5.1 and 5.2 above, we claim that our methodology is able to distinguish not
only between baseline and none-baseline categories (point 5.1) but also among the nonebaseline categories themselves (point 5.2). In other word, the methodology is able to detect
faults corresponding to different locations (such as inner race and outer race) and of
different severities such as small and large.
6. So, we believe that the work done, according to the description in point (5), is unsupervised
learning method because so far we did not use the information about the fault categories.
7. When we move to the next phase/ level of diagnosis we build a number of feature matrices
which belong to different fault categories. More specifically when we need to know the
category of a specific feature set (i.e to which fault location category or severity it belongs) we
build a number of signal categories from training samples. In this case it can be considered as a
supervised identification process.
This was clarified in the revised submitted by adding new paragraphs
Please see in the revised submission, chapter 1, lines (2-17) in page 11 and, lines (10-15) page12

4. Main motivation for using vibration analysis needs to be stated.


Authors response
In brief, vibration-based methodology is able to detect 90% of all machinery defects. In addition, the
machines vibration signal contains information about the global behaviour of the machine and any
changes within the machine result in changes in the vibration signal. Furthermore, it is an online and
non-destructive methodology as vibration signals can be acquired several times from the machine
without switching the machine off.
Please see in the revised submission, chapter 1, modified section 1.4.5 in page 6.
5. Please provide a list of aims and objectives.
Authors response
The author adds a new section which explains the aims and objectives of the research.
Please see the revised submission, chapter 1, a new section 1.5 in page 7.
6. Simplicity- how it was achieved relative to other methods.
Authors response
The author would like to clarify the simplicity mentioned in the thesis. The simplicity in
methodology achieved by:

For the first approach, our methodology is simpler when compared to other SSA- based
methods because we use only signals from healthy bearing category to build the
baseline space by subjecting them to decomposition stage only which makes it practical.
4

When compared to other SSA based fault diagnosis techniques, which uses signals from
both healthy and fault bearing categories and subject each of them to SSA, the present
approach is simpler.
Please see the revised submission, chapter 1, lines 12-18, in page 10.

For the second approach, the suggestion of the new signal pretreatment facilitates the
efficient use of the simplest form of the autoregression. By using the LTIVAR, neither
initial set of model coefficients and shape of coefficients evolution over time need to be
assumed.

Please see in the revised submission, chapter 1, line 16 - to the end of page 12.
Please see in the revised submission, chapter 3, lines (4-7) and lines (13-21) in page 39.
Please see in the revised submission, chapter 3, last two lines in page 40.

7.

reduced dimensionality space- please extend this in relation to the literature.

Authors response
The author would like to clarify that the current SSA-based methodology (i.e chapter 4) was able to
achieve 100% of correct classification by using only the first PC. Thus we notice that in comparison to
other SSA-based techniques, discussed in section 3.4.1, the way that we apply the SSA reduces the
necessary number of PCs required for achieving maximum correct classification.
Please see in the revised submission, line 12-16, page 112.
Page 25- please insert [79] in each caption.
Authors response:
Sources for all images and figures have been provided and the proper captions with sources are
given in the revised thesis.
8. Consider putting chapter 3 earlier on in the thesis.
Authors response:
Chapters 2 and 3 are swapped in the revised submission.
Please see the revised submission, chapter 2 starts at page 17 and chapter 3 starts at page 29.
9. Re-think how you express the removal of PCs in order to clarify the methodological precision
of doing this.
Authors response
From the decomposition stage, a number of principal components (i.e PCs) are produced. Each of
these PCs has a portion from the original signal variance. The portions of the variance are arranged
ascendingly with the number of PCS as described in the so called scree plot (see for example Figure
5.2. This means portion of the original variance decreases when we move to the last PCs.

There are several criteria for removing of PCs explained in the literature. In our work, we have
checked the methodologys correct classification rate at different number of PCS (i.e from 1 to 10)
and eventually select the number of PCs which gives maximum correct classifications. In most cases
the number of PCs preserves at least 75% of the original variance which meet with one of the criteria
mentioned in the source [82].
Please see the revised submission, lines 8-14 in page 73.
10. Clean up the analysis to explain how you put these existing equations together.
Authors response
For the methodology presented in chapter 4, we have explained in section 4.3.1 how the equations
corresponding to decomposition stage are utilised in obtaining the eigenvector and eventually used
in building the baseline space.
Please see the revised submission, chapter 4, section 4.3.1, page 46.

For the methodology presented in chapter 6, we have explained in section 6.2.2 how we incorporate
the decomposition and reconstruction stage in cleaning the signal noise. It is important to mention
that SSA has not used before for signal denosing for fault diagnosis of REBs.
Please see the revised submission, chapter 6, section 6.2.2, page 115

11. Lognormal following- why is this? (Please explain why is this best choice?. Refer to page 169
and extend.
Authors response
We have mentioned that we select a probability density function which fits the distribution of
Mahalanobis data precisely. Since Mahalanobis distances are positive values and their distribution is
skewed to the right so we select the lognormal one. The similarity of the lognormal to the histogram
of the MD is checked visually and it is shown they are very similar.
Please see the revised submission, chapter 4, lines (8-11) in page 51.
A figure 5-B.1 was inserted in the appendix B which shows the similarity between the lognormal
distribution and the histogram of the Mahalanobis distance.
Please see also the revised submission, page 192, figure 5-B.1.

12. Please show you have questioned the assumptions you have made.
Authors response
Please see the revised submission, chapter 4, lines (4-13) in page 51.

13. Page 47- why did you choose cumulative probability to be 0.99? a low threshold.
Authors response
In classification problem, misclassifications are divided into two main groups: 1) False-positive
which indicates the presence of a fault when it is not and 2) False-negative which indicates no fault
when it is present. In the current study, a statistical-based threshold is determined such that these
false alarms are kept to minimum during the fault detection process.
The author has checked the misclassification rates at different cumulative probability values (i.e 0.97
and 0.99) and eventually 0.99 was selected as it gives minimum misclassification rates.
Please see the revised submission, line 8-end of page 52, and the first line of page 53.

14. Page 52 & 53- figure 4.5 and 4.6 explain source of data please.
Authors response
Data used in this example belongs to case study 1 -1772 RPM for H and IRF case where 30 feature
vectors from each category are used.
Please see also the revised submission, the captions of Figure 4.5 and Figure 4.6 on pages 57 and
58 respectively.
15. Page 60. Explain the test sequence procedure clearly.
Authors response
A healthy bearing is mounted on the shaft and inside the bearing housing and then data are
acquired for acceleration vibration signals are acquired at a specific speed such as 250 RPM. Then
the motor speed is increased and another five signals are obtained at the new speed and the same
done for the speed considered in the analysis. When the measurement from the healthy signals end,
another faulty bearing, such as a bearing with a fault in inner raceway, is mounted in the test rig and
the author follows the same measurement procedure. The measurements are repeated for the
other fault bearings.
Please see the revised submission, a new paragraph, lines (14-21) in page 65.
16. Page 60 emphasise that the 0.05 fault diameter was deliberately small.
Authors response
This fault size was deliberately made small to challenge the performance of the methodology in
detecting small faults. Due to the limitations of the EDM tool size at the time of conducting the
experiments, smaller fault size was not made. However, cases of smaller fault sizes were covered in
the other case studies (i.e CS1 and CS3).
7

Please see the revised submission, lines 7-11, page 65.


17. Page 62. It would be helpful to conclude the difference between and with of experimental
data.
Kindly see the authors response to comment 2.3 above.
18. Insert a comment about length of FV data when applying machine learning methods.
Authors response
In machine learning it is preferable and will be easier to work with few features (i.e short FV
lengths). In the current study, it is noticed generally that Increasing of FV length affect positively on
the correct classification rates it was illustrated in the results presented in chapter 5. This can be
interpreted as increasing FV length means incorporating more PCS and then more information about
the fault presents is included.
Please see the revised submission, lines (3-9) in page 110.

19. Page 73. Please state that there is a general phenomenon in the in the FV dimensionality
comment with respect to PCs. Refer to page 84.(please relate this to information in Figure 5.22
on page 95)
Authors response
A general phenomenon is seen in the relationship between FV dimensionality (i.e number of PC
involved in building the baseline space) for all the levels of fault diagnosis. The genral
phenomenon can be described as increasing FV dimensionality improves the correct
classification rates. This can be interpreted as the increase of FV dimensions (i.e more PC
components are used) means involving more information about the presence of the fault
presence.
Please see in the revised submission, lines 1-6 in page 81.
Figure 5.22 shows the variation of Mahalanobis distance level at different fault severities. In this
Figure, the level of Mahalanobis distance to the baseline is shown on the basis of using minimum
baseline space dimension that provides maximum correct classification rates. Although, all the
FV corresponding to fault categories are correctly detected, their Mahalanobis distance levels do
not follow a monotonic function. The exploration of the reasons behind this non-monotonic
behaviour can be an interesting future study.
Please see in the revised submission, chapter 5, section 5.3, lines (16-21) in page 110 and first
two lines in page 111.

20. Page 82. Re-consider Figure 5.13 to more for what you have explained the phenomenology, as
clearly as you can, regarding the various nonlinear processes that lie behind it. Please include a
discussion of this.
Authors response
Figure 5.13 shows the clustering of the 3D baseline space for FVs corresponding to baseline and
different IRF severities. Although all the FV are clustered clearly, they locate in a non-monotonic
path. It is clear in Figure 5.13, that the small IRF cluster locates the farthest from the baseline cluster.
This kind of behaviour is only shown for CS1 while for the CS3 (i.e Figure 5.23) the Mahalanobis
distance follows a monotonic function. In fact, the reasons behind the non-monotonic path followed
by the data of CS1, is not investigated in the current study. This might explained as the feature
selected are not able always to behave linearly. A further future study might help in the further
understanding the physical meaning of the features.
Please see in the revised submission, chapter 5, section 5.3, lines (16-21) in page 110 and first two
lines in page 111.
21. Section 5.3. Please consider depth and extend as you think necessary.
Authors response
In the revised submission, the section 5.3 has been extended and modified to have more deep.
Please see in the revised submission, chapter 5, section 5.3, page 108
22. Page 109-110 models of different orders- please explain this all more clearly.
Authors response

All the criteria of model order estimation suggest that the optimum model order
corresponds to the minimum value of a loss function. However, it is not always easy to find a
clear minimum value of the loss function. Therefore, a criterion is used for guidance only and
the model order is an estimated one. In some cases, different optimum models orders
might be estimated for different signals/ signal classes. In the current study, different
optimum models orders means different FV lengths. The equalising of FV lengths is
necessary when NN ruled-based classifier (i.e more specifically based on Mahalanobis
distance) is used. This is because both the training and testing FVs must have the same
length when such classifier is used. So, the equalising of FV lengths is applied if there is a
difference in the optimum models order.
Four different possibilities for equalising FVs lengths were investigated to see whether the
way of equalising may affect the performance of the methodology. From the results
obtained, we found that both the zero padding and trimming procedures perform well.

Please see the revised submission, section 6.2.6, last two lines in page 120 and lines (1-11) in page
121.

23. Please insert a statement on classifier and regressor around page 112.
Authors responses

The diagnosis of the severity condition can be considered as a classification problem or a regression
problem. It might be desirable to use a regression model, however this might complicate the
methodology as it requires determination the coefficients of regressor. In the current study, we
dealt with the quantification of the bearing fault severity as a classification problem assuming that a
faulty bearing signal can be assigned to one of fault severity categories (i.e small, medium and large).
Please see in the revised submission, section 6.3.2, last 5 lines in page 123 and first two lines in
page 124.
24. End of chapter 7 and chapter 8. Consider extending conclusions and discussion of assumption
etc. chapter 8 needs to be a little deeper.
Authors response
In the revised submission, the conclusions and discussion of chapter 7 are modified and extended for
more depth.
Please see in the revised submission, chapter 7, section 7.4, page 159.
Modifications including extending the conclusion in more depth, the limitations and the future
recommendation to cover these limitations and/ or extending the research are made in the revised
submission
Please see in the revised submission, see the modified version of chapter 8, from page 162-172.

10

You might also like