Content uploaded by Duraivelu .K
Author content
All content in this area was uploaded by Duraivelu .K on Sep 25, 2022
Content may be subject to copyright.
Vol.:(0123456789)
1 3
Journal of Vibration Engineering & Technologies
https://doi.org/10.1007/s42417-022-00566-0
ORIGINAL PAPER
Vibration‑Based Fault Diagnosis ofBroken Impeller andMechanical
Seal Failure inIndustrial Mono‑Block Centrifugal Pumps Using Deep
Convolutional Neural Network
S.Manikandan1 · K.Duraivelu1
Received: 14 September 2021 / Revised: 8 May 2022 / Accepted: 9 May 2022
© Krishtel eMaging Solutions Private Limited 2022
Abstract
Purpose Hydraulic pump failure results in a high rate of energy loss, performance degradation, high vibration levels, and
continuous noise emission. An unexpected pump failure might result in a sudden collapse of the hydraulics, resulting in sig-
nificant financial losses and the shutdown of the whole factory. Fault diagnosis plays a critical function in diagnosing flaws
before they occur. Early detection is crucial for identifying problems and may save money, time, and potentially dangerous
circumstances.
Methods In recent years, many studies in intelligent fault diagnosis utilizing various machine learning approaches have
been conducted. A vibration-based fault diagnosis in industrial mono-block centrifugal pumps is presented in this study.
An experimental configuration for structuring databases, required for developing algorithms for running machine learning
programs, is designed. Standard condition vibration signals are collected from the setup when the pump is healthy and free
of defects. This study considers the two major defective conditions of broken impeller (B.I.) and seal failure (S.F.). The
faults are introduced in the pump one after the other, and the vibration signals are obtained. The image processing approach
converts these analog signals to 2D images.
Results Later, the images are trained and tested using a deep convolution neural network (DCNN) classifier, and the fault
accuracy is verified. The results show an accuracy of 99.07% after training and testing the image dataset.
Conclusion The suggested DCNN architecture exhibits high and accurate fault diagnosis accuracy for the industrial mono-
block centrifugal pump.
Keywords Vibration signals· Fault diagnosis· Broken impeller· Seal failure· Deep convolution neural network
Introduction
In most industries, mono-block centrifugal pumps are being
utilized in various essential engineering applications, such
as heating units, mining operations, construction sites,
mineral processing, and high-pressure applications. Their
failures cause adverse effects, resulting in complete machin-
ery failure [1]. The impeller and mechanical seal are the
two primary components of a hydraulic centrifugal pump
that directly influence the pump’s performance properties
[2]. It is necessary to improve the pump dependability and
safety [3]. Vibration data processing is a hotbed of signal
processing research, and it is critical to diagnose the excel-
lent and faulty vibration signals obtained from the equip-
ment [4]. Vibration signal-based approaches, model-based
techniques, composite techniques, and knowledge-based
techniques are the four types of defect diagnostic methods
[5]. Defect detection methods mainly focus on mechanisms,
specific frequency, and feature extraction to solve an issue
[6]. Currently, hydraulic system fault diagnosis concentrates
primarily on three factors (i) the primary signals, which typi-
cally combine with the additional noise, (ii) there seems to
be a lack of controller design that is capable of generaliza-
tion, and (iii) the absence of a unified solution for system
integration [7].
* S. Manikandan
ms1016@srmist.edu.in
K. Duraivelu
duraivek1@srmist.edu.in
1 SRM Institute ofScience andTechnology, Kattankulathur
campus, Chennai, India
Journal of Vibration Engineering & Technologies
1 3
Data-driven signal processing method evaluates the
gathered signals and extract relevant fault features. These
techniques can adapt the data acquired and extract sensi-
tive information to evaluate machine health [8]. Recently,
the use of deep learning-based approaches in fault detec-
tion of rotary machines, particularly motors and pumps, has
been thoroughly examined and addressed. The diagnostic
performance of such newly developed procedures is empha-
sized. They provide concepts and advice for investigating
and implementing fault diagnostics in rotational machinery
[9]. A deep neural networkis the crucial machine learn-
ing technique that uses both supervised and unsupervised
learning to learn several levels of computational models at
various stages [10]. Deep learning techniques may adap-
tively learn the data structure from the basic information
by numerous transformations which are not linear and com-
plicated nonlinearity functions, rather than extracting fault
features manually.
The major applications of deep neural networks (DNN)
are image recognition and language processing domains.
DNN-based fault detection process has gained prominence
in recent times [11]. The study of unsupervised learning
features is widely used to tackle proper classifications and
image processing issues with the precision of learned traits
demonstrated [12]. DNN has been shown to benefit by iden-
tifying more characteristics and displaying enormous data in
several studies. They are considered to be the critical factor
in the detection of multiple machine faults [13].
In DCNNmodels, the empirical mode segmentation pro-
cess is mainly used to diagnose the faults in rotating equip-
ment [14]. Usually, the bearing failure diagnosis is examined
using the SVM algorithm (support vector machine) with the
Hilbert–Huang transform approach [15]. The K-singular
value decomposition (K-SVD) and the ML representation
techniques are used to diagnose the damage in the wind tur-
bine bearings [16]. Based on multi-kernel learning oncen-
trifugal pumps, the autonomous fault diagnosisincreases
the production efficiency [17]. The combination of an
extreme learning machine (ELM) and the Mel-frequency
cepstrum coefficients (MFCC) isused in an axial piston
pump fault diagnostic technique [18]. Various machine
learning approaches detect multi-faults on the high-speed
gearbox [19]. For improved efficiency, a proper flaw detec-
tion method in rotating machinery is required.
In recent days, the impact of signal decomposition on
various fault diagnostic techniques has been investigated
[20]. The vibration processing signals and mixed multiple
object deep learning [21] process focus more on the fault
diagnosis of rotary equipment. The other prominent method
is the acoustic emission method used in the fault detection
on roller bearings [22]. Another strategy with a squared
inverted feature representation is utilized to increase effi-
ciency using the k-Nearest Neighbor (kNN) classifier [23,
24]. The Teager Kaiser Energy approach is used with vibra-
tional mode decomposition to identify failure mechanisms
in the bearings [25]. Considering the varying motor loads
and speeds, vibration analysis’s time domain characteristics
determine the gear cracking levels [26]. Methods are studied
based on K-nearest neighbors on identifying the various gear
fracture levels at various motor loads and speeds [27]. The
identification of electrical defects using DNN for additional
problems related to the operation of water pumps is inves-
tigated [28].
Vibration signal analysisof functionally graded compos-
ite structure components having crack formation in centrifu-
gal pumps are the main focus considered for the study [29].
An Entropy-Based feature in the FAWT (flexible analytic
wavelet transform) is used to detect a fault in an automatic
gearbox [30]. Of late, time–frequency approaches are used
to diagnose and classify the spot in diesel engines [31].
The rotating machinery fault diagnosis is studied [32]. The
dynamics of the fiber-reinforced plastics (FRP) cracked
beams are investigated by processing the vibration signals
through DNN [33]. From various studies, the neural network
model and fuzzy logic technique were the most efficient
way to identify the faults in rotary machine [34]. There are
various applications in feature extraction process. Out of
which image processing is one of the most powerful meth-
ods. ResNet is used to extract the defined features using the
predefined network model. The ResNet model feature extrac-
tion is the fastest and easiest way to identify the features of
sound and image characteristics [35] when compared with
the other type of extraction technique. They remove the main
drawback of vanishing gradient and outperform complicated
networks. The fault diagnosis in wind turbine gearbox shows
a maximum accuracy 98.79% using the hybrid ResNet [36].
In the transfer, CNN for the fault diagnosis based on the
ResNet carried out on motor and pump shows an accuracy of
98.95% which is considered to be high when compared with
other neural networks [37]. In the thermal-based fault diag-
nosis, ResNet model with CNN shows a maximum accuracy
of 95.7% [38], whereas Alexnet and VGG net shows 93.57%
and 93.75%, respectively, on comparison.
The deep convolutional neural network (DCNN) plays
an important role in the traditional image classification
algorithm. The convolution kernels of the DCNN model
effectively identifies the features in this study [39]. The tra-
ditional ML techniques are found to be time-consuming and
more complex to calculate the results. DCNN shows excel-
lent performance than any other ML algorithms [40]. They
are more effective in image processing. The important aspect
of this study is to develop an algorithm to train and test the
faulty images and to get more precise fault accuracy results
through DCNN algorithm. Here, an experimental setup is
arranged, where the signals of normal conditions are col-
lected from the form when the pump is healthy and free of
Journal of Vibration Engineering & Technologies
1 3
defects. In this study, defective conditions such as broken
impeller (B.I.) and seal failure (S.F.) are introduced, and
the corresponding signals are also obtained. The limitation
of the study is restricted to only two critical faults consid-
ered in this paper. The image processing approach converts
these signals to two-dimensional (2D) images. The datasets
are processed by the proposed deep convolutional neural
network (DCNN) classifier [41]. The images are separated
into two sections for training and testing, with a ratio of
70:30 [42]. The classification is carried out using the DCNN
algorithm, which predicts the enhanced system with various
layers such as convolution, max pooling, and an activation
function. In this study, rectifier linear unit (ReLU) is used
as an activation function. Finally, the validated results are
collected to determine the fault diagnosis efficiency in the
mono-block centrifugal pump.Figure1 shows the methodol-
ogy flowchart carried out in this study.
Acquisition ofVibration Signals
In this study, the two significant faults considered are B.I.
and S.F. that occurs in a mono-block centrifugal pump. The
other significant faults may be an internal bearing failure,
outer bearing failure, cavitation, bearing seal failure, and
overheating. The vibration signals introduced in the mono-
block centrifugal pump determine the fault diagnosis accu-
racy using the vibrational signals. In most rotary equipment,
vibration signals are used to extract the data and perfor-
mance of the system. Signals are analyzed using the accel-
erometer sensor that is placed near the motor. The major-
ity of conventional general-purpose accelerometer sensor
includes a powerful sufficient signal to be monitored without
considering the signal contamination. The signal is sent to
the computer through the power amplifier module and the
interface card from the sensor linked to the vibrating system.
Sensors take measurements through analog format, and the
results are then converted to digital outputs for continuous
usage, including analog to digital conversion. This sensor
transfers the vibrational information towards the data acqui-
sition device (DAQ), after that the signals are amplified and
filtered. The DAQ system connected to the laptop is shown
in Fig.2. The different sensor outputs classify the charac-
teristics of the mono-block centrifugal pump. The classified
features are the data of good pump, mechanical seal failure,
and broken impeller. These signals are analyzed using Lab
VIEW software, and then the signals in the form of raw data
are saved. Figure3 shows the data acquisition’s flow chart.
Later, the one-dimensional (1d) raw data were converted to
2D images.
Experimental Setup
A single-phase 0.5 HP industrial mono-block centrifugal
pump having a flow of 120 LPM with a rated head of 15m
and a rating speed of 2900 RPM is stated as a study object
as shown in Fig.4. The major faults such as mechanical seal
failure and broken impeller are analyzed in this research. A
NI-9234 data acquisition system and a uniaxial accelerometer
Fig. 1 Methodology flowchart Fig. 2 DAQ system
Journal of Vibration Engineering & Technologies
1 3
sensor with an operating frequency of 0.5–10,000Hz are used.
The DAQ module has 51.2 kS/s/channel, 4-Channel, ± 5V,
and 24-bit IEPE signal conditioning with A.C. coupling. The
accelerometer used in this study is uniaxial and has a sensi-
tivity of 100 mv/g. The sensor is installed near the motor to
organize the signal waveforms and parameters.
Image Processing Through Normalization
Technique
The time domain features of the vibration signals within are
transformed into 2D dimensions utilizing energy valuesduring
thismethod. The signals areseparatedinto equal subparts as a
frame.The amountof structure is computed by multiplying the
signal’s frame length and harmonic rate. This study obtaineda
selectednumber of frames after sub-dividing the signals to
establish the depth and breadth of the matrices. The matrix’s
height is equivalent to the frame size, and the matrix’s width is
identical to the frame quantity. For example, if the frame size
is M and the amount of the frame is N, then the ideal matrix
size is MxN. The energy levels of definedstructures are then
entered into the matrix’s cells. Each frame value is vertically
put into the matrix. The first frame energy values are entered
into the matrix’s first column, with the first value going into
a first-row cell and the second value going into the second
row first cell. As the matrix height is equivalent to the frame
quantity, the final value of the first frame is placed in the first
cell of the last column, and every subsequent value fits into
the column. Consequently, the additional frame energy levels
are entered one by one into the matrix. All the frames are pro-
vided inside the matrix since the matrix’s breadth is equal to
the number of frames. The normalization is carried out with
values of the frames in the range of 0–120 after the values are
entered into the matrix. The matrix’s lowest numerical value
is 0, and its greatest numerical value is 120. All numerical
numbers are derived based on their ratios using this range. The
previous values inside the matrix are then replaced by the new
values. The normalizing process aids in lowering the original
signal’s resolution and supports the reduction of noise using
the identified signals. Signal characteristics in the time domain
are preserved in a 2D representation, and texture features of
the signal in 2D may be retrieved to identify the signal. This
study also utilizes the signal frequency domain that represents
the signal as a combination of magnitude and phase values for
each element frequency. The signal levels are then normalized
in the range of 0–120. Virtually every signal has an efficient
pattern that converges in the low-frequency region, so the fre-
quency signals do not generate any unique textures.
Normalization is the process of transforming an n-dimen-
sional gray scale picture:
(1)
I∶{X∈Rn}n{Min, …., Max}
Fig. 3 Flow chart for data acquisition
Fig. 4 Experimental setup
Journal of Vibration Engineering & Technologies
1 3
into a range of intensity values (Min, Max). The intensity
values of the new range (new Min, new Max) are denoted by
the new image conversion of
The formula for linear normalizing of a grayscale digi-
tized image is as follows:
CNN Approach
The input layer, a pooling layer, a convolutional layer, a fully
connected layer, as well as an output layer, are all standard
components of a DCNN, as shown in Fig.5. It is the most
powerful machine learning techniques that enhances learn-
ing by sharing convolutional kernels and down-sampling to
increase resilience. The convolutional layer extracts feature
from a particular region on the map, with multiple convolu-
tional kernels corresponding to different feature extractors.
It transforms the input image layer to bring out the features
present. The convolution layer input represents the N*N
square neuron layer, and the m*m filter means the output of
(2)
IN∶{X∈Rn}−{new Min, …., new Max}
(3)
I_N =(I−Min)(
new Max −new Min)∕(Max−Min)+new Min
the convolution layer. The following equation calculates the
convolutional layer size:
The weights of the pre-nonlinearity of the xl
ij layer are
calculated by the following equations:
The following equations calculate the convolution output
layers for the multiple inputs. ω denotes the function bias, xl
ij
denotes the weight function, a and b are the processing param-
eters, and i, j are the weight function parameters:
Xl
cm is the m-th feature map acquired by activating lst con-
volutional layer; f1 is the convolutional layer activation func-
tion; bl
cm is biased with m-th feature map of l-th convolutional
layer; Mn is the collection of specified feature maps; kl
nm is the
convolution kernel linking the n-th feature map layer l-1 and
(4)
(N−m+1)x(N−m+1)
(5)
x
1
ij =
m
∑
a=0
−1
m−1
∑
b=0
𝜔abyi−1(i+a)(j+b
)
(6)
X
l
cm =f1
⎛
⎜
⎜
⎝
�
xl−1
n∈Mn
Xl−1
n∗kl
nm +bl
cm
⎞
⎟
⎟
⎠
Fig. 5 DCNN architecture
Journal of Vibration Engineering & Technologies
1 3
the m-th feature map layer l; Mn is the collection of specified
input feature mappings.
The pooling layer is a type of non-linear down-sampling
that decreases computation time by lowering network param-
eters, and it controls overfitting to a degree. The following
equation calculates the pooling layer:
Here, Xl
pm is a feature map derived from the m-th pool-
inglayerand f2 is the pooling layer’s activation function. wl
n
is the weighting of the m-th predicted output in layer l; down
is the pooling function; bl
pm is the biases of the convolution
layer derived from the m-th pooling layer.
A conventional feed-forward neural network serves as a
fully connected layer. The ReLU activation function is then
used as activation function to solve the multi-classification
problem. The prime advantage of using the ReLU function
is that it is computationally effective compared with other
activation functions such as sigmoid and tan h. The char-
acteristic features are collected from the transformed 2D
images using MATLAB to build the convolutional neural
networking method to categorize the input signals.
Figure5 shows the complete DCNN architecture. To train
the classifier, the feature vectors are used as the training
data. The entire amount of samples for assessment is 16384
samples, and the total dataset considered per fault is 120.
The sampling rate is 8.192kHz, and the number of epochs
is 25, with a learning rate of 0.0001. The study used 70%
of the dataset for classifier training and 30% for testing to
distinguish the training dataset from the test dataset. Based
on that confusion matrix is obtained to measure the experi-
ment’s correctness and confirm that the approaches men-
tioned above are adequate. In thedataprocessing problems,
DCNN has very high accuracy. DCNN recognizes key char-
acteristics without human intervention, and the computation
weight distributed is organized. It is necessary to organize
large amount of training data.
(7)
X
l
pm =f2
wl
ndown
xl
cm
+bl
pm
Results andAnalysis
In this study, the principal critical faults of B.I. and S.F. are
considered as they constitute the significant share of defects
[43] occurring in a mono-block centrifugal pump. Figure6a
shows the broken impeller fault at the corner of the Impeller,
and Fig.6b shows a damaged mechanical seal. Both the fault
conditions are fitted. The raw vibrational data are acquired
at regular intervals into the pump, one after the other. The
initial process is extracting features from the excellent pump,
followed by the faulty critical pump. After the extraction
process, they are converted to 2D images through normali-
zation and finally fed into the DCNN algorithm to check the
fault accuracy.
Extraction ofFeatures
The concept of extracting features for evaluating pump deg-
radations is an essential step in developing pump vibration
analysis. Different unique characteristics that are derived
from centrifugal pump vibration signals are explored.
Mechanical seal failure and broken impeller are the two typi-
cal problems that are physically introduced into the pumps.
An acceleration sensor on the pump collects the vibration
data. The vibration signals are collected from a single-phase
0.5 HP industrial centrifugal mono-block pump with the
standard operating condition. 16,384 samples are collected
from each fault and are in good condition. It is worth noting
that the electrical outlets of various health problems split
from the pump’s good need to fault-based conditions. The
vibration signals are analyzed over amplitude vs. time as
shown in Fig.7a good pump signal, Fig.7b broken impeller
fault signal, and Fig.7c mechanical seal failure signal as
stated as follows.
Fig. 6 a Broken impeller, b seal
failure
Journal of Vibration Engineering & Technologies
1 3
Image Processing Analysis
First, various phases of the motor current’s signals are moni-
tored simultaneously. Each of the detected signals from each
stage of the motor current is kept as a distinct dataset that
is analyzed separately. The initial signals are measured in
a 1D format. By quickly reorganizing the array of signal
amplitudes into a square matrix form, these signals are
transformed into 2D form using MATLAB software. In a
square matrix n*n, each part of the present signal of the
length is reorganized as n2. The transformation converts
a one-dimensional signal into a two-dimensional signal.
This two-dimensional representation of current signals is
identified as a gray-level image. The analysis is done for
all the datasets, and a sample of 2D gray images are shown
in Fig.8a good healthy pump, Fig.8b broken impeller, and
Fig.8c mechanical seal failure.
Fault Diagnosis Through Deep Convolutional
Neural Network (DCNN)
A DCNN starts with alternating convolutional and various
pooling layers and then connects the fully connected layer,
activation function before the output layer. The features’
absolute precision of its generated system is assessed and
(a)
(b)
(c)
-0.01
-0.005
0
0.005
0.01
00.511.5 2
Series1
-0.02
-0.01
0
0.01
0.02
00.5 11.5 2
Series1
-0.02
-0.01
0
0.01
0.02
00.5 11.5 2
Series1
AMPLITUDE
(mm/s2)
AMPLITUDE
(mm/s2)
AMPLITUDE
(mm/s2)
Time (s)
Time(s)
Time (s)
Fig. 7 1D Raw signal: a good pump, b broken impeller, and c seal failure
Journal of Vibration Engineering & Technologies
1 3
retrieved based on the networks in various layers. Finally,
the graphic patterns of the pump’s response are used to
determine the efficiency of the generated model. The main
advantage of using DCNN is to extract the best features from
the 2D gray-level images. This approach is more power-
ful and efficient in classifying the elements when compared
with other algorithms. The one main disadvantage of using
DCNN is overfitting. However, as more data are added, over-
fitting is reduced. The DCNN parameters considered are
shown in Table1. Initially, the CNN is trained using 11,468
signal samples from the training set. The trained CNN is
then tested using 4916 samples from the testing set from a
total sample of 16,384, and these samples are analyzed at
an 8.192k(Hz) sampling rate. For these training data, the
parameter consideration is 70%, while it is about 30% for
testing. 100 filters with a convolution kernel size of 5 × 1
are defined in convolutional layers 1 and 2. The pooling size
is set to two by the maximum pooling layer. The activation
function is ReLU, assigning zero weights to the network’s
neurons at random. Table2 shows the total number of data-
sets considered for each fault and a healthy pump. A total of
15 datasets for each spot are taken at different pressures from
0.1kg/cm2 to 1.5kg/cm2 at eight time intervals. Thus, a total
of 120 datasets are obtained for classification of accuracy.
To prepare the classifier, the training data features are
used. Furthermore, a similar approach is used for feature
extraction from the testing dataset, where the components
are categorized. The DCNN algorithm results focus on the
confusion matrix outcomes for the defined faults occurring
in the pump’s mechanical parts. The confusion matrix results
are validated by the convolution layers, normalization layers,
pooling layers, and activation functions. The proper class
and predicted class in the confusion matrix were identified
based on the true positive (T.P.), true negative (T.N.), false
positive (F.P.), and false negative (F.N.) values. The catego-
rization precision is accessed using the confusion matrix
that is shown in Fig.9, and their predicted class and accurate
class accuracy are depicted in Table3. As seen in Table3,
the CNN classifier’s prediction classification accuracy for a
good pump, B.I., and S.F. classes is 100%, 100%, and 97.2%.
This study primarily gathered vibration signals indicative of
the pump’s faulty state. The pump dataset ischosen for the
test verification, having a sampling frequency of 8.192kHz,
and the pumps are used to simulate faults such as a broken
Fig. 8. 2D gray-level images: a good pump, b broken impeller, and c seal failure
Table 1 DCNN parameters
Parameter name Value
Number of convolutional layers 2
Number of convolutional kernels (256, 512)
Convolutional kernel size (5 × 1, 5 × 1)
Number of maximum pooling layers 2
Pooling size for each layer (3 × 1, 3 × 1)
Activation function ReLU
Learning rate 0.0001
Maximum iterations 30
Mini batch size 32
Epochs considered 30
Iterations per epoch 1
Table 2 Dataset classification
Fault condition Dataset
Good pump 120
Broken impeller 120
Mechanical seal failure 120
Journal of Vibration Engineering & Technologies
1 3
impeller and mechanical seal failure in this study. The over-
all training validation of the dataset is shown in Fig.10. The
validation accuracy of the DCNN algorithm data is 99.07%
for both learning and testing. The overall process of the
CNN technique took 34s, with 30 epochs and 30 iterations,
with a learning rate of 0.0001 and the loss function of this
approach is approximately near zero. The research discovers
that the DCNN algorithm is a quicker and better classifier,
and the results prove to be more efficient with an overall
fault accuracy of 99.07%
Conclusion
This research uses DCNN architecture to analyze the
two critical faults B.I. and S.F. in mono-block centrifu-
gal pump with the vibrational signals. The research
demonstrates that the pump fault diagnosis is accurate,
and the suggested technique analyses the non-stationary
and non-linear vibration signals. The activation function
ReLU is considered for this DCNN architecture. Finally,
based on the training data, the confusion matrix for the
predicted class accuracy is 97.3%, 100%, and 100% for a
good pump, broken impeller, and seal failure, respectively,
and the actual class accuracy is 100%, 100%, and 97.2%.
The fault diagnosis accuracy using this DCNN architec-
ture is identified as 99.07%. The suggested diagnostics
architecture may also be used to consider other mechanical
faults such as cavitation, outer bearing failure, inner bear-
ing failure, and overheating. In the future scope, the sever-
ity of each faults can also be determined. Even the combi-
nation of various faults with the suggested architecture can
be done by comparing other traditional machine learning
algorithms such as SVM, kNN, and random forest.
Fig. 9 Confusion matrix
Journal of Vibration Engineering & Technologies
1 3
Funding The authors did not receive support from any organization
for the submitted work.
Data Availability The datasets generated during and/or analyzed during
the current study are not publicly available as the research work with
few more faults are yet to be carried out in future but are available with
the corresponding author on reasonable request.
Declarations
Conflict of interest The author(s) declared no potential conflicts of in-
terest concerning the research, authorship, and/or publication of this
article.
Financial interests The authors have no relevant financial or non-finan-
cial interests to disclose.
References
1. Azadeh A, Saberi M, Kazem A, Ebrahimipour V, Nourmoham-
madzadeh A, Saberi Z (2013) A flexible algorithm for fault diag-
nosis in a centrifugal pump with corrupted data and noise based
on ANN and support vector machine with hyper-parameters opti-
mization. Appl Soft Comput 13(3):1478–1485. https:// doi. org/ 10.
1016/j. asoc. 2012. 06. 020
2. Muralidharan V, Sugumaran V, Indira V (2014) Fault diagnosis of
mono-block centrifugal pump using SVM. Engineering Science
and Technology, an International Journal 17(3):152–157. https://
doi. org/ 10. 1016/j. jestch. 2014. 04. 005
Fig. 10 Overall training validation of dataset
Table 3 Accuracy based on the confusion matrix
Fault condition Predicted class accuracy True
class
accuracy
Good pump 97.3% 100%
Broken impeller 100% 100%
Seal failure 100% 97.2%
Journal of Vibration Engineering & Technologies
1 3
3. Muralidharan V, Sugumaran V (2013) Feature extraction using
wavelets and classification through decision tree algorithm for
fault diagnosis of mono-block centrifugalpump. Measurement
46(1):353–359. https:// doi. org/ 10. 1016/j. measu rement. 2012. 07.
007
4. Goyal D, Mongia C, Sehgal S (2021) Applications of digital signal
processing in monitoring machining processes and rotary compo-
nents: a review. IEEE Sens J 21(7):8780–8804. https:// doi. org/ 10.
1109/ JSEN. 2021. 30507 18
5. Gao Z, Cecati C, Ding SX (2015) A survey of fault diagnosis
and fault-tolerant techniques—Part I: Fault diagnosis with model-
based and signal-based approaches. IEEE Trans Ind Electron
62(6):3757–3767. https:// doi. org/ 10. 1109/ TIE. 2015. 24175 01
6. Tian Y, Lu C, Wang ZL (2015) Approach for hydraulic pump fault
diagnosis based on wpt-svd and svm. In Applied Mechanics and
Materials. Trans Tech Publications Ltd 764:191–197. https:// doi.
org/ 10. 4028/ www. scien tific. net/ AMM. 764- 765. 191
7. Jiang W, Li Z, Zhang S, Wang T, Zhang S (2021) Hydraulic pump
fault diagnosis method based on EWT decomposition denoising
and deep learning on cloud platform. Shock Vib. https:// doi. org/
10. 1155/ 2021/ 66743 51
8. Bin GF, Gao JJ, Li XJ, Dhillon BS (2012) Early fault diagnosis
of rotating machinery based on wavelet packets—empirical mode
decomposition feature extraction and neural network. Mech Syst
Signal Process 27:696–711. https:// doi. org/ 10. 1016/j. ymssp. 2011.
08. 002
9. Tang S, Yuan S, Zhu Y (2019) Deep learning-based intelligent
fault diagnosis methods toward rotating machinery. IEEE Access
8:9335–9346. https:// doi. org/ 10. 1109/ ACCESS. 2019. 29630 92
10. LeCun Y, Bengio Y, Hinton G (2015) Deep learning nature
521(7553):436–444. https:// doi. org/ 10. 1038/ natur e14539
11. Lu C, Wang ZY, Qin WL, Ma J (2017) Fault diagnosis of rotary
machinery components using a stacked denoising autoencoder-
based health state identification. Signal Process 130:377–388.
https:// doi. org/ 10. 1016/j. sigpro. 2016. 07. 028
12. Ali JB, Fnaiech N, Saidi L, Chebel-Morello B, Fnaiech F (2015)
Application of empirical mode decomposition and artificial neural
network for automatic bearing fault diagnosis based on vibration
signals. Appl Acoust 89:16–27. https:// doi. org/ 10. 1016/j. apaco ust.
2014. 08. 016
13. Hinton GE, Salakhutdinov RR (2006) Reducing the dimensionality
of data with neural networks. Science 313(5786):504–507. https://
doi. org/ 10. 1126/ scien ce. 11276 47
14. Xie Y, Zhang T (2017) Fault diagnosis for rotating machinery based
on convolutional neural network and empirical mode decomposition,
2017 Shock and Vibration. https:// doi. org/ 10. 1155/ 2017/ 30841 97
15. Kabla A, Mokrani K (2016) Bearing fault diagnosis using Hilbert-
Huang transform (HHT) and support vector machine (SVM). Mech
Ind 17(3):308. https:// doi. org/ 10. 1051/ meca/ 20150 67
16. Yang B, Liu R, Chen X (2017) Fault diagnosis for a wind turbine
generator bearing via sparse representation and shift-invariant
K-SVD. IEEE Trans Ind Inf 13(3):1321–1331. https:// doi. org/ 10.
1109/ TII. 2017. 26622 15
17. Feng H, Wang Y, Zhu J, Li D (2018) Multi-kernel learning based
autonomous fault diagnosis for centrifugal pumps. 2018 Interna-
tional Conference on Control, Automation and Information Sciences
(ICCAIS), 548–553. https:// doi. org/ 10. 1109/ ICCAIS. 2018. 85703 22
18. Jiang W, Li Z, Li J, Zhu Y, Zhang P (2019) Study on a fault identi-
fication method of the hydraulic pump based on a combination of
voiceprint characteristics and extreme learning machine. Processes
7(12):894. https:// doi. org/ 10. 3390/ pr712 0894
19. Su X, Tomovic MM, Zhu D (2019) Diagnosis of gradual faults in
high-speed gear pairs using machine learning. J Br Soc Mech Sci
Eng 41(4):1–11. https:// doi. org/ 10. 1007/ S40430- 019- 1701-3
20. Yadav, P., Chauhan, S., Tiwari, P., Upadhyay, S.H. and Rakesh, P.K.,
(2020). Mode selection in variational mode decomposition and its
application in fault diagnosis of rolling element bearing. Reliab
Safety Hazard Assess Risk-Based Technol 663–670. https:// doi.
org/ 10. 1007/ 978- 981- 13- 9008-1_ 56
21. Xin Y, Li S, Wang J, An Z, Zhang W (2020) Intelligent fault diag-
nosis method for rotating machinery based on vibration signal
analysis and hybrid multi-object deep CNN. IET Sci Meas Technol
14(4):407–415. https:// doi. org/ 10. 1049/ iet- smt. 2018. 5672
22. Chacon JLF, Kappatos V, Balachandran W, Gan TH (2015) A novel
approach for incipient defect detection in rolling bearings using
acoustic emission technique. Appl Acoust 89:88–100. https:// doi.
org/ 10. 1016/j. apaco ust. 2014. 09. 002
23. Sharma A, Jigyasu R, Mathew L, Chatterji S (2018) Bearing fault
diagnosis using weighted K-nearest neighbour. In: 2nd International
Conference on Trends in Electronics and Informatics (ICOEI),
1132–1137. https:// doi. org/ 10. 1109/ ICOEI. 2018. 85538 00
24. Wang H, Yu Z, Guo L (2020) April. Real-time online fault diagnosis
of rolling bearings based on KNN algorithm Journal of Physics:
Conference Series 1486:032019. https:// doi. org/ 10. 1088/ 1742- 6596/
1486/3/ 032019
25. Verma V, Ratol JS, Singh A (2019) Bearing fault detection using
ensemble empirical mode decomposition integrated with Teager
Kaiser energy method. Int J Technolol Comput 4:167–171
26. Wang D (2016) K-nearest neighbors based methods for identification
of different gear crack levels under different motor speeds and loads:
Revisited. Mech Syst Signal Process 70:201–208. https:// doi. or g/ 10.
1016/j. ymssp. 2015. 10. 007
27. Sanchez RV, Lucero P, Macancela JC, Cerrada M, Cabrera D,
Vasquez R (2019) Gear crack level classification by using KNN and
time-domain features from acoustic emission signals under differ-
ent motor speeds and loads. SDPC Proceedings-2018, International
conference on sensing, diagnostics, prognostics, and control. https://
doi. org/ 10. 1109/ SDPC. 2018. 86649 79
28. Gundewar SK, Kane PV (2021) Condition monitoring and fault
diagnosis of induction motor. J Vib Eng Technol 9:643–674. https://
doi. org/ 10. 1007/ s42417- 020- 00253-y
29. H.S., R.C., Bharadwaj, S.C., Umashankar, S., Padmanaban, S.,
Dutta, N. and Holm-Nielsen, J.B., (2019), June. Electrical fault
detection using machine learning algorithm for centrifugal water
pumps. IEEE International Conference on Environment and Electri-
cal Engineering and IEEE Industrial and Commercial Power Sys-
tems Europe (EEEIC/I&CPS Europe), 1–6. https:// doi. org/ 10. 1109/
EEEIC. 2019. 87838 41
30. Sinha GP, Kumar B (2021) Review on vibration analysis of function-
ally graded material structural components with cracks. J Vib Eng
Technol 9(1):23–49. https:// doi. org/ 10. 1007/ s42417- 020- 00208-3
31. Ramteke DS, Pachori RB, Parey A (2021) Automated Gearbox Fault
Diagnosis Using Entropy-Based Features in Flexible Analytic Wave-
let Transform (FAWT) Domain. J Vib Eng Technol 9(7):1703–1713.
https:// doi. org/ 10. 1007/ s42417- 021- 00322-w
32. Ramteke SM, Chelladurai H, Amarnath M (2021) Diagnosis and
Classification of Diesel Engine Components Faults Using Time-
Frequency and Machine Learning Approach. J Vib Eng Technol.
https:// doi. org/ 10. 1007/ s42417- 021- 00370-2
33. Lobato TH, da Silva RR, da Costa ES, Mesquita AL (2020) An inte-
grated approach to rotating machinery fault diagnosis using, EEMD,
SVM, and augmented data. J Vib Eng Technol 8(3):403–408. https://
doi. org/ 10. 1007/ s42417- 019- 00167-4
34. Jena PC, Parhi DR, Pohit G (2019) Dynamic Investigation of FRP
Cracked Beam Using Neural Network Technique. J Vib Eng Technol
7(6):647–661. https:// doi. org/ 10. 1007/ s42417- 019- 00158-5
35. Mahajan, A., & Chaudhary, S. (2019). Categorical Image Classifi-
cation Based On Representational Deep Network (RESNET).3rd
International conference on Electronics, Communication and Aer-
ospace Technology (ICECA), 327–330. https:// doi. org/ 10. 1109/
ICECA. 2019. 88221 33
Journal of Vibration Engineering & Technologies
1 3
36. Kai Zhang;Baoping Tang;Lei Deng;Xiaoli Liu; (2021).A hybrid
attention improved ResNet based fault diagnosis method of wind
turbines gearbox. Measurement, Vol:179. https:// doi. org/ 10. 1016/j.
measu rement. 2021. 109491.
37. Wen L, Li X, Gao L (2020) A transfer convolutional neural net-
work for fault diagnosis based on ResNet-50. Neural Comput Appl
32:6111–6124. https:// doi. org/ 10. 1007/ s00521- 019- 04097-w
38. Pinedo-Sánchez LA, Mercado-Ravell DA, Carballo-Monsivais CA
(2020) Vibration analysis in bearings for failure prevention using
CNN. J Br Soc Mech Sci Eng 42:628. https:// doi. org/ 10. 1007/
s40430- 020- 02711-w
39. Lei X, Pan H, Huang X (2019) A dilated CNN model for image
classification. IEEE Access7 124087–124095. https:// doi. org/ 10.
1109/ ACCESS. 2019. 29271 69
40. Hao Wu, Zhao J (2018) Deep convolutional neural network
model based chemical process fault diagnosis. Comput Chem Eng
115:185–197. https:// doi. org/ 10. 1016/j. compc hemeng. 2018. 04. 009
41. Malla C, Panigrahi I (2019) Review of condition monitoring of
rolling element bearing using vibration analysis and other tech-
niques. J Vib Eng Technol 7(4):407–414. https:// doi. org/ 10. 1007/
s42417- 019- 00119-y
42. Manikandan S, Duraivelu K (2021) Fault diagnosis of various rotat-
ing equipment using machine learning approaches–A review. Proc
Inst Mech Eng Part E: J Process Mech Eng 235(2):629–642. https://
doi. org/ 10. 1177/ 09544 08920 971976
43. Hasan MJ, Rai A, Ahmad Z, Kim J-M (2021) A fault diagnosis
framework for centrifugal pumps by scalogram-based imaging and
deep learning. IEEE Access 9:58052–58066. https:// doi. org/ 10.
1109/ ACCESS. 2021. 30728 54
Publisher's Note Springer Nature remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.