Chapter 3

Data Acquisition and DFC

Processing

3.1 Introduction

This chapter presents the acquisition process of vibration signals and the detailed de-
scription of the dataset characteristics. Since the rotor faults are more sensitive to
vibration data, the referred datasets use different vibration sensors to collect data at
different speed, load and other working conditions. To analyze rotor faults other than
broken rotor bar fault, more than 80.0% of the works depend on rotor testbeds for
faulty data generation. The primary dataset (DS-1) is a novel one which has been
collected from a rotor kit setup (Meggitt-Mi 19003) [183], which simulates the real
plant varying speed working environment. In addition, a publically available data set
called Machinery Fault Database (MaFaulDa) [184] is employed as the second dataset
(DS-2) for general performance comparison in the proposed frameworks. The experi-
mental setup, fault simulation process, related parameters of data acquisition, etc., are
explained along with the testbed description is given in the first section of the chapter.
The remaining sections bring out the significance of DFC in SRF diagnosis in detail
to provide an insight into the effective fault diagnosis. Then the section contains the
44 3.2. Datasets Description

Data Acquisition
Variable Speed General purpose Unbalance
Machine 3 Phase Induction Disk with
Foundation Motor Drive Accelerometer Weight Proximity General
Baseplate Motor notch
Coupling Pedestal Weighing Sensor purpose
Tacho meter
Bolt Wheel Accelerometer

RTB
Rubbing Foundation
Bush Bolt

SCU

3a. Structural 3b. Pedestal

Looseness Looseness
c 4c. Rub Impact
3 Looseness
C
MAU

1 Misalignment 2 Unbalance 4 Shaft Faults

CFs cd
4b. Shaft Crack
d m
m
IA
GA bd
a1 d d1 d2
m
a2 IA
2a. Static UB 2b. Couple UB

m1 != m2
m2 4d. Shaft
Corrosion
GA
1a. Angular M.A 1b. Parallel M.A 1c. Parallel & 4a. Bent Shaft
m1
Angular M.A IA
2c. Dynamic UB

RTB: Rotor test bed; SCU: Signal-conditioning unit; MAU: Monitoring & Analysing Unit; C, c: Looseness clearances; a1, a2: Angular shifts; d: Parallel shift;
m, m1, m2: Weights; CF: Centrifugal Force; GA: Geometrical axis; IA: Inertia axis; bd: Bent depth; cd: Crack depth; d1, d2: Crack distance from shaft ends

Figure 3.1: Rotor fault implementation in a testbed

proposed DFC extraction process from varying operating conditions followed by the de-
scription of the experiment conducted to signify the dominance of DFC as the decision
parameter in SRF diagnosis.

3.2 Datasets Description

3.2.1 Meggitt dataset (DS-1)

The data acquisition setup contains the RTB, the signal-conditioning unit (SCU), and
the monitoring and analyzing unit (MAU) [16]. Fig. 3.1 presents a diagrammatic
representation of a typical rotor testbed-based data acquisition setup along with the
fault implementation and the associated frequency responses.

The main component of the testbed assembly is an electric motor controlled by a

variable frequency drive (VFD, which varies the frequency and voltage supply) con-
3.2. Datasets Description 45

Figure 3.2: Meggitt testbed (Meggitt-Mi 19003)

Table 3.1: Setup conditions for testbed of DS-1 dataset.

Features Specification / values

Accelerometer: Piezo-electric ,
Sensors
Proximity sensors: Eddy current based
RPM Range: 0 − 3600
RPM / Frequency
Sampling Frequency: 1 KHz
Motor Power 1 kW-h
750 rpm, 1500 rpm, 2250 rpm, 3000 rpm, 3250
Rotor Speeds
rpm with Runup and Rundown
Unbalance weights 13.0 g and 26.0 g (Dynamic UB)
Diameter Weighing Disc: 100.18 mm and Shaft:12.66 mm

nected to a rotating shaft via a flexible coupling. The shaft is supported by bearing
housing that is fixed by pedestal bolts to the testbed base. The bearing housing has
provisions for mounting sensors in horizontal, vertical, and axial directions. The shaft
will be mounted with discs for various purposes. Generally, a disc with one or more
notches is closely placed with a tachometer that senses the rotational speed of the shaft.
This tachometer reference, along with a sensor waveform, is used to calculate the phase
46 3.2. Datasets Description

information. Meanwhile, weighing discs, which will likely incorporate grooves, holes, or
another eccentric weight connecting provision, will also be mounted on the shafts. The
majority of the testbeds are equipped with a metal bush connecting facility to create
the rub impact [185]. Contact type general-purpose accelerometers and non-contact
type proximity sensors are commonly used for the vibration data acquisition. The ac-
quired data is then passed through an SCU, which often performs signal amplification
and analog-to-digital conversion. It is then applied to the MAU, which is generally
a computer that carries out the monitoring and analysis of the input signal and any
associated tasks with the help of specifically developed software. In SRF, the unbalance

Table 3.2: Setup conditions for testbed of DS-2 dataset.

Features Specification / values

Analog tachometer, accelerometers, and micro-
Sensors
phone.
Frequency Sampling Frequency: 51.2 kHz
RPM RPM Range: 700 − 3600 rpm
Motor Power 1/4 CV DC
Rotor Speeds 10.0 − 60.0 Hz
Unbalance weights 6.0 g, 10.0 g, 15.0 g, 20.0 g, 25.0 g, 30.0 g, 35.0 g
Horizontal: 0.50 mm, 1.0 mm, 1.50 mm, 2.0 mm
Misalignment Vertical: 0.51 mm, 0.63 mm, 1.40 mm, 1.90 mm,
1.27 mm, 1.78 mm
Diameter Rotor: 152.4 mm and Shaft: 16.0 mm

fault is created by connecting weights (bolts, nuts, washers, etc.) on the weighing disc.
Two equal weights are placed 180◦ apart to create couple unbalance, while a one-sided
weight is positioned to generate static unbalance. Much like in couple unbalance, if
unequal weights are set in place, then dynamic unbalance can be simulated [26]. The
coupling is loosened to induce a parallel shift and/or angular shift to create parallel
and/or angular misalignment [185]. In terms of loosening faults, the majority of the
works deal with pedestal loosening, which is simulated by loosening the pedestal bolt of
3.2. Datasets Description 47

the bearing housing to create clearance from the testbed base [16]. Meanwhile, struc-
tural looseness can be demonstrated by loosening the bolt connecting the testbed base
to the foundation. The stationary bush that is in contact with the shaft or disc pe-
riphery will induce rub impact [26]. The shaft bend and shaft crack provisions are also
shown in Fig. 3.1. Here, the experiment is conducted by setting different fault condi-
tions as described before, and then the testbed is set to different rotational frequencies
and load conditions to collect data via repeated trials. The SRF dataset created for
our experiments consists of six different fault classes namely healthy (HL), static UB
(S UB), couple UB (C UB), dynamic UB (D UB), misalignment (MA) and looseness
(LS). The data collection conditions and the general testbed information are provided
in Table 3.1.

3.2.2 MaFaulDa dataset (DS-2)

MaFaulDa is a publicaly available dataset that uses SpectraQuest’s machinery fault

simulator and produces multivariate TS data for SRF diagnosis. The dataset uses the
two sets of accelerometers for the collection of vibration data, which are organized in
three orthogonal directions, i.e., axial, radial, and tangential. It is equipped with a
tachometer to measure the system rotation frequency and a microphone to capture the
operating sound. The faulty conditions considered in this setup are normal (NM), unbal-
ance (UB), and horizontal misalignment (H MA) and vertical misalignment (V MA).
The unbalance conditions are simulated with seven different weights. The misalign-
ments are implemented with varying shift clearances in the data gathering process. A
total of 1951 various conditions are incorporated in the dataset. For each scenario,
the data is collected at 50 kHz over a time interval of 5 seconds, with eight prescribed
signals. The MaFaulDa data acquisition setup description is given in Table 3.2.
48 3.2. Datasets Description

Table 3.3: Description of rotor faults

FaultCategory Vibration Characteristics Assoc.
(Fault No.) Faults Cause & Effect
Plane Symptomatic Frequency Phase
Angular Axl 2x peaks are higher than 1x A phase shift of
(1a) peak. Stronger 3x axial and 180◦ in the axial Cause: Component expansion and
Misalignment (1)

torsional responses at 1/3rd of direction. cold alignment due to thermal distor-

critical speed. tions. Improper coupling alignment,
4a(C), Uneven foundation, imparted forces
Parallel Rad 2x peaks are higher than 1x A phase shift of Brg. from other components, bent shaft,
(1b) peak. Stronger 1x axial and 180◦ in the radial Flts(E) mass unbalance, etc.
torsional responses. direction. Effect: Early fatigue failure, exces-
Parallel Axl 2x peaks are higher than 1x A phase shift of sive heat and friction, and compo-
& & peaks. Presence of 3x har- 180◦ in radial and nent damage.
Strucural Rotor Faults

Angular Rad monics. axial position

(1c)
Static Rad Larger 1x and harmonics with A phase shift of 0◦ Cause: Rotor mass eccentricity due
Unbalance (2)

(2a) less than 15.0% of the 1x. in radial direction to assembly errors, manufacturing
defects, debris deposit, etc. Presence
Couple Rad Larger 1x and harmonics with A phase shift of 4a(C), of other faults: bow shaft, corrosion,
(2b) less than 15.0% of the 1x. 180◦ in radial di- 3a(C), falling of damaged rotor part, etc.
rection 4d(C) Effect: Higher dynamic load and
Dynamic Rad Larger 1x and harmonics with 0◦ to 180◦ Phase early fatigue failure, wobbling
(2c) less than 15.0% of the 1x. shift. movement.

Component Axl Dominating 2x. Presence of Unstable phase Cause: Improper fittings, tear and
(3a) & 1x to 10x harmonics, subhar- reading. wear, thermal expansion.
3 (E),
Looseness (3)

Rad monics (x/2, x/3, etc.), mul- Effect: Damage or detach of compo-
2 (E)
tiples of subharmonics (2x/3, nents, Secondary faults like misalign-
4x/3, etc.). ment and unbalance.
Structural Rad Higher 1x and/or 2x. Har- 180◦ phase shift Cause: Loosened bolts or bed-
(3b) monics frequency amplitude between base plates, deteriorated concrete foun-
above 50.0% of 1x. plate and founda- 2 (C) dation, loose or distorted machine
tion mountings.
Effect: Component looseness, unbal-
ance, and misalignment.
Bent Axl Higher 1x when bent at the A phase shift of Cause: 1) Permanent: Large un-
Shaft & middle of shaft. 2x generated 180◦ in the axial & balance force, uneven shrink fittings,
(4a) Rad when bent close to the cou- 0◦ in the radial di- residual stress or collision.
pling. rections. 2a(E) 2) Temporary: Friction and thermal
distortion, high length to width ra-
tio.
Shaft Faults (4)

Effect: Higher dynamic load and

early fatigue failure, unbalance.
Shaft Rad Dominating 2x. Presence of Directly propor- Cause: Severe thermal and mechani-
Crack 1x and certain subharmonics tional to the depth 4c(E), cal stresses or manufacturing defects.
(4b) (1/2x, 1/x, etc.). of the crack. 4a(E) Effect: Reduction in bending stiff-
ness, shaft bow.
Rub Rad Higher 1x with 2x, 3x and Inconsistent Cause: Excessive unbalance, mis-
Impact other harmonics. Presence of phase. 1 (C), alignment, self-excited instability,
Fault (4c) 1/2x, 1/3x, etc. 2 (C), resonance, thermal distortions.
4a(E) Effect: Excessive heat, Thermal bow.
Shaft Not directly evident. Cause: Electrochemical reaction due
Corrosion 4b(E) to environmental factors.
& Wear Effect: Metal lose, crack
(4d)
Broken Ro- Rad Case 1: Higher 1x with, up Inconsistent Cause: Thermal, magnetic, dy-
tor Bar (5) to 4x. Pole-pass sidebands phase. namic, mechanical, and environmen-
with 2x line frequency. Case2: 4a(E), tal stresses.
Higher 1x with 2x line fre- 2 (E) Effects: Uneven current flow. Ther-
quency sidebands. mal bow and unbalance.
*Abbreviations- Axl: Axial, Rad: Radial, Assoc.: Associated, Brg. Flts: Bearing Faults, C: Cause of, E: Effect
3.3. DFC and SRF 49

3.3 DFC and SRF

The study of symptomatic parameters of rotor faults revealed that SRF is more sensitive
to DFC of rotational frequency. Hence this section addresses the SRF-DFC correlation
and the experiments conducted to substantiate it. The impact of DFC in characterizing
the vibration of SRF can be visually verified from Fig 3.3, where DFC representation of
a healthy and unbalance vibration signal at 1500 RPM is shown. At the 1x frequency of
around 25Hz, the amplitude of unbalance signal is high, and small spikes are present at
all the other rotational frequencies. In contrast, for healthy data signal all amplitudes
of rotational frequencies are similar. This evidence that DFC can be the clear indicator
of SRF. A more detailed description of SRF is given in Table 3.3 with amplitude and
phase characteristics, along with the most affected plane, associated faults, and the
causes and effects of the faults. The fault-wise description of the impact of DFC in
SRF is given below:

(a) Healthy (TD) (b) Unbalance (TD)

(c) Healthy (DFC) (d) Unbalance (DFC)

Figure 3.3: DFC representation of signal at 1500 RPM

50 3.3. DFC and SRF

3.3.1 Structural rotor faults

The SRF, being the primary cause of vibration, strongly reflects the abnormal vibrations
by various DFC parameters. The vibrational characteristics of faults such as misalign-
ment, unbalance, and looseness are explained in this section to depict the DFC-SRF
correlation.

3.3.1.1 Misalignment

Misalignment can be analyzed by comparing the ratio between 1x (unbalance indica-

tor) and 2x (misalignment indicator) components. In normal misalignment scenarios,
2x and 1x components are present in the radial vibration spectrum, with 2x being the
predominant component with a range up to 150.0% of 1x [67]. The severe misalignment
conditions are characterized by the harmonics 3x to 8x or even a full high-frequency
harmonics series. Patel et al. [186] presented certain observations to identify the type
of misalignment as well as to distinguish between misalignment and crack faults in the
rotor. In order to uniquely identify misalignments, they suggested investigating the
presence of strong negative higher harmonic frequency components compared to that
of rotor crack. They also observed that stronger 1x axial as well as torsional responses
and weak higher harmonics represent parallel misalignment. They added that a strong
3x harmonic component in its axial and torsional vibration response at 1/3rd critical
speed indicates angular misalignments. Phase information is also used as a defining
feature of misalignments. Across the coupling or machine, a phase shift of 180◦ in the
axial position shows angular misalignment, and the same in radial position indicates
parallel misalignment, whereas that in both axial and radial position represents com-
bined misalignment [68]. The vibration waveforms follow a periodic pattern having one
or two cycles in each revolution of the shaft.
3.3. DFC and SRF 51

3.3.1.2 Unbalance

The vibration wave produced by unbalance is sinusoidal and occurs at a frequency of

‘one per revolution’ (1x), i.e., a single frequency vibration with the same amplitude in
all radial directions. Other than the severe faulty states, vibration generally contains
1x only, without any harmonics of it. The 1x with high amplitude and its harmonics
with less than 15.0% of the 1x is an indication of an unbalanced state [187]. In this
case, up to the first critical speed of the machine, the amplitude increases with speed,
and the phase from the vertical and horizontal measurements differ by 90◦ . In dynamic
unbalance, there will be 180◦ phase shift in the radial direction, while static unbalance
shows a phase shift of 0◦ . For unbalance due to the bent shaft, the phase shift of 180◦
happens in the axial direction with no phase shift in the radial direction [67].

3.3.1.3 Looseness

In the case of component looseness, the initial stage vibration signature contains mostly
1x and 2x components, but with escalated deterioration, the fractional harmonics with
increased amplitude starts to appear. Generally, looseness is characterized by several
running speed frequency harmonics (1x – 10x) with subharmonics (x/2, x/3, etc.) and
their integer multiples (2x/3, 4x/3, etc.) of magnitudes greater than 20.0% of the 1x
amplitude. Structural looseness creates 1x and/or 2x radial components with predom-
inant vertical amplitude, subject to the type of issue. For rigidly connected machines
with no belts or couplings, the radial 2x signifies looseness. The waveform generated
is periodic with one or two cycles per revolution. A phase difference of 180◦ exists
between the foundation and vibrating components in case of structural looseness [67].
52 3.3. DFC and SRF

3.3.2 Shaft faults

According to the fault categorization from the vibration perspective, the shaft faults
such as bending, cracking, and rub impacts are the secondary cause of abnormal vibra-
tion. Hence, their frequency characteristic responses are described in this section.

3.3.2.1 Bent shaft

In the case of bent shaft, if the bent is close to the middle of the shaft length, then 1x
will be dominating, and 2x will be dominant for the bents near to the couplings. The
most affected plane is axial, though vertical and horizontal planes will also give out 1x
and 2x peaks [72]. The 2x amplitude can vary from 30.0% of the 1x amplitude to 100.0%
– 200.0% of the 1x amplitude. The spectrum of bent shaft is almost similar to that of
misalignment; hence, the phase can be used as a good differentiating indicator. In bent
shafts, the radial phase measurements will be in phase, while axial measurements will
be 180◦ out of phase at opposite ends of the component [67].

3.3.2.2 Shaft crack

The first and foremost effect of shaft crack is the reduction in bending stiffness in the
direction of the crack, which results in inducing excessive 2x vibration in the shaft [71].
The second effect is the rotor bow, where the bending results in the natural axis shift cor-
responding to the direction of the crack. This effect generates 1x components, which will
progressively add to the already existing residual unbalance frequency component [72].
The cracked shaft vibration mostly affects the radial plane, and it produces increased
1x vibration along with the 2x and 3x harmonics. In certain scenarios, the presence of
subharmonics 1/2x, 1/x, etc., are also observed with this fault. Studies revealed that
the phase shift is directly proportional to the depth of the crack [71, 188].
3.4. DFC Extraction 53

3.3.2.3 Rub impact

The vibration changes in characteristic ways whenever the rub happens between sta-
tionary and non-stationary components. So other than the times there is no contact,
the waveform seems absolutely normal. The spectrum will be containing higher am-
plitude subharmonics and superharmonics of the synchronous frequency with a strong
rub impact. Chu et al. [189] observed the presence of 1/2 fractional harmonic compo-
nents and 1/3 fractional harmonic components along with the 2x, 3x harmonic com-
ponents. Meanwhile, the existence of pseudo-resonance and backward whirling compo-
nents showed by [190] as the effect of rub. The phase information shows that it can not
maintain a consistent phase in rub related vibration motion.

3.3.3 Broken rotor bar

Uneven current flow to induction motor rotor owing to crack or break results in two
kinds of vibration changes. In the first case, 1x and harmonics generally up to 4x
will accompany with pole-pass sidebands with quite low 2x line frequency [67]. In the
second case, the amplitude of rotor bar passing frequency (RBF, which is the number
of rotor bars times the running speed) will be higher, with 2x line frequency sidebands
(100Hz or 120Hz) [66]. Due to the motion of vibration, the phase will be consistent in
broken rotor bar faults.

3.4 DFC Extraction

According to the literature review, the utilization of DFC is instrumental in developing

solutions of SRF. Meanwhile, the literature study upholds the need of more realistic
industrial data with varying speed and load conditions, which in turn makes the DFC
extraction more complicatied. Hence we proposed the following procedure in DFC
extraction and data subsampled feature space creation, which suppresses the other
54 3.4. DFC Extraction

raw data related issues. Sub-sampling is performed by selecting the summary interval
points that ensure sufficient data points in each sensor segment. This stipulates proper
extraction of DFC (FD component) and TD representatives of each segment. The
process is described as follows:

Let K be the number of sensors mounted at different positions, producing univariate

TS vibration signals V = [V1 , V2 , ..., VK ] with each Vi ∈ RL where L is the length of
one sample sequence. Since the framework deals with rotational data, one complete
informative segment should consider at least the number of sampling points produced
in a single rotation. This decision is based on the given rotational frequency in rpm Rs
and the sensing frequency Sf of the data acquisition system. The rotation frequency
is converted in seconds for the convenience of calculation by Rf = Rs /60. Then, the
number of sampling points per rotation is calculated by: Sr = 60Sf /Rs or Sf /Rf .
Thus, the interval points or segment length is decided according to the condition: So ≤
Sr ≤ Sl , where So is the overlapping points in segmentation and Sl is the segment
length. The segments are extracted equally from all the sensors so that the number of
segments from an L length single sensor sequence is:

ns = (L − Sl )/(Sl − So ) + 1 (3.1)

Hence, the segments of sensor Si are denoted by Vi = [Vi,1 , Vi,1 , ..., Vi,ns ].

After the segmentation, each segment is depicted by a high-level representation in

the sub-sampled feature space. The rotational frequency and its higher-order harmonic
components are more sensitive to the SRF. Thus, the basic task is to extract these
components from the vibration spectrum. In a real plant scenario, RM working at a
fixed rpm (Rs ) undergoes speed fluctuations due to varying voltage, connected load,
the influence of other rotating components, etc. As a result, the rotational frequency
and its harmonic components can not be estimated directly by Rf and its multiples
to extract DFC. In faulty conditions, this phenomena of fluctuation of actual values
3.4. DFC Extraction 55

Figure 3.4: DFC extraction from varying RPM

around the theoretical values are more obvious. Hence, the process of extracting DFC
with FFT is assisted with proper normalization process [16] to reduce the rate difference
under varying operating conditions. This operation is shown as

fn
1X
Aamp = amp(fix ) (3.2)
fn i=1

amp(fix )
µ (fix ) = (i = 1, 2, ..., fn ) (3.3)
Aamp

Here, we consider fn frequency bands for DFC extraction and the amp(fix ) represents
the amplitude corresponding to fix frequency. The Aamp shows the mean of these
amplitudes, and finally µ(fix ) gives the normalized amplitude values. There is often
a fluctuation observed in the frequency range near the rotation frequency bands, and
hence, a multi-pass filter with a frequency range of ∆f that extracts frequency between
fix −∆f and fix +∆f is used. This has been shown in Fig. 3.4, considering the frequency
spectrum of the unbalance fault at 1500 rpm. In practical data collection experiments,
it is observed that the rpm variation is approximately less than 10.0%, which doesn’t
56 3.5. Role of DFC in SRF Diagnosis

cause significant amplitude change. Both amplitude and phase values are extracted
to generate the DFC components. Similarly, the TD representative of each segment
is generated from the data bins of length Sr . The DFC extraction is performed on
the harmonic frequencies according to the number of features required, and appended
along with the TD features to generate a subsampled vector for representing a particular
sensor segment.

3.5 Role of DFC in SRF Diagnosis

This section proves the significance of DFC in SRF diagnosis by comparing the perfor-
mance of the conventional TD and FD features against DFC features. This study is
conducted in order to substantiate DFC as the symptomatic fault feature of SRF. The
two top performing ML algorithms, SVM and ANN along with ANFIS algorithm for
evaluating fuzzy reasoning on DFC decision parameters are used for the DFC perfor-
mance comparison. The first two models are the most widely used ML technique in RM
fault diagnosis [10]. At the same time, ANFIS blends the ability of fuzzy logic to mimic
human reasoning and backpropagation learning of ANN. All three models were tested
with ten features each from TD, FD, and DFC features for performance comparison.
The TD features contain mean, standard deviation, variance, root mean square, abso-
lute maximum, coefficient of skewness, kurtosis, crest factor, margin factor, and shape
factor. The FD features include mean, variance, third moment, fourth moment, grand
mean, C factor, D factor, E factor, G factor, and H factor derived from the frequency
transformed domain. DFC is represented by amplitude and phase representatives of 1x
to 5x frequencies.

ANFIS method was developed by Jang in 1993 [111]. ANFIS model allows user
interference compared to ANN which can be utilized to incorporate domain specific
decision making criterias. It is a hybrid learning a structure that combines fuzzy logic
and ANN which consists of backpropagation learning. But in ANFIS, the fuzzy logic is
3.5. Role of DFC in SRF Diagnosis 57

Table 3.4: DFC Performance with ML Models

DS-1 DS-2
ML Models
TD FD DFC TD FD DFC
ANN 84.84 92.49 94.27 87.52 92.76 94.43
SVM 80.51 92.04 93.96 86.07 92.00 94.39
ANFIS 84.46 93.20 95.01 88.52 93.59 95.64

treated more systematically that it demands least expert knowledge. So in this imple-
mentation, eventhough it is possible to add the rules related to SRF diagnosis directly,
the system is let to learn the rules itself, providing the decision making parameters as
input. ANFIS normally has five layers of neurons of which neurons in the same layer are
of the same function family. The layers are named as the fuzzification layer (I), the rule
layer (II), the normalization layer (III), the defuzzification layer (IV), and the output
layer (V). The ANFIS algorithm uses Sugeno based fuzzy logic to deal with impreci-
sion or uncertainty and neural network for adaptability. Integrating the relatively low
learning rate neural network with fuzzy logic makes it suitable for time critical applica-
tions [191]. SVM is implemented with RBF kernel, and its regularization parameter C
is set to 1.0, and the parameter γ that decides the influence of training example is set
to 0.1 by performing grid search. The ANN has three layers, i.e., the input layer con-
tains ten nodes, the hidden layer with non-linear RBF activation function containing
20 nodes, and the output layer nodes decided depending on the number of classes.

The results are shown in Table 3.4, from which it is observed that TD features show
the least accuracy on both datasets as they are extracted directly from the raw data
segments. But FD features and DFC perform much better than TD, signifying the
importance of transformed features in identifying SRF. The FD features give almost
similar accuracy in both the datasets with the models, but an apparent discriminative
ability can be determined with DFC features in the transformed domain. This shows
that DFC contains more symptomatic fault features of SRF, with less complex extrac-
tion operation than traditional FD features. It is evident that all the three ML methods
58 3.6. Summary

provide the highest accuracy with DFC, especially the ANFIS model. The role of DFC
as decision parameters of SRF is exploited by the rules derived by ANFIS, much better
than the other models. This establishes DFC as the most dependable feature for SRF.
Among the datasets, DS-2 got higher accuracy than DS-1, especially with TD features.
But in the transformed domain, there is no significant performance enhancement for
DS-2. ANN and SVM show almost similar performance for both datasets with a slight
upper hand for ANN, while with the transformed data, ANFIS is dominating. From the
discussion, It is apparent that DFC demonstrated its role in efficacious SRF diagnosis.
The impact of DFC with sequential DL methods can be verified from section 4.5.2.
These result leads to recognizing DFC as the main component of SRF diagnosis in the
proposed framework.

3.6 Summary

In order to promote the fault-characteristics-based analysis of RM fault diagnosis, un-

derstanding the vibrational characteristics of the faults are very much essential. This
chapter serves this purpose by providing the theoretical background of rotor faults with
fault characteristics analysis. It also illustrates a general AI-based RFD framework,
which we followed in all the models described in the remaining chapters. Similarly,
data availability drives the recently developed intelligent data-driven methods. The
RM fault diagnosis suffers from various data-related issues such as data scarcity, data
imbalance, and impure data. Furthermore, it is indispensable that data must reflect
real plant working environments. Hence, this chapter is provided a description of the
datasets we used in the experiments. Similarly, this chapter emphasizes the role of DFC
in SRF with the summarization of various related findings and with the experiments
conducted in our works in this regard.

The critical difference between the DS-1 and DS-2 data sets is the variation of
speed used to capture the data. The operating range of rotational frequency is almost
3.6. Summary 59

similar in both datasets, with an upper band RPM of 3600. DS-1 has five different
RPM conditions and six faults, while DS-2 contains four fault conditions with around
50 different rotor speeds of varying other operating conditions. It summarizes that the
DS-2 dataset is more diverse as compared to the DS-1 data set. But the run-up and run-
down data provide distinguished fault diagnosis capabilities to DS-1, which is absent
in DS-2. Moreover, DS-1 is more imbalanced than DS-2, which demands additional
metrics like F1-score to ensure its performance. The performance comparison with the
conventional TD and FD features and DFC features has been done on SVM, ANN,
and ANFIS algorithms. The DFC produced apparent discriminative ability compared
to the other two features, especially over the TD features in the transformed domain.
Moreover, the role of DFC as decision parameters of SRF has been proven by the
exceptional results produced by ANFIS, which exploits the rules derived from DFC.
Thus the impact of DFC is showed in SRF diagnosis, and the result leads to recognizing
DFC as the main component of SRF diagnosis in the proposed framework