You are on page 1of 13

Neural Computing and Applications (2024) 36:1727–1739

https://doi.org/10.1007/s00521-023-09150-3(0123456789().,-volV)(0123456789().
,- volV)

ORIGINAL ARTICLE

Robot arm damage detection using vibration data and deep learning
Getachew Ambaye1 • Enkhsaikhan Boldsaikhan1 • Krishna Krishnan1

Received: 20 June 2023 / Accepted: 17 October 2023 / Published online: 15 November 2023
This is a U.S. Government work and not under copyright protection in the US; foreign copyright protection may apply 2023

Abstract
During robot operation, robot components like links and joints may experience collisions or excess loads that can lead to
structural damages or cracks. A crack in a structural component can degrade the overall performance of the structure. This
study examines the influence of cracks on the vibration characteristics of a baseline robot link. The approach uses the finite
element method to simulate the dynamics of planar robot link models with and without artificial cracks with different sizes,
locations, and orientations in the ABAQUS software. The robot link models include one intact model and five defective
models with cracks. A rectangular crack with a fixed length of 1 mm and a varying width from 0.001 to 0.1 mm is applied
to a specific location along the robot link. Finite element analysis and machine learning are used to simulate and
characterize the vibration of each robot link with one fixed end and one free end. The vibration responses are measured at
the free end. The measured vibration data are then transformed into two-dimensional (2D) image data using the Gramian
Angular Summation Field method. A convolutional neural network is then trained with the image data for crack detection
and analysis. The results indicate that the proposed method demonstrates 98.25% accuracy on the data generated by the
simulation experiments.

Keywords Vibration characterization  Convolutional neural network  Robot health monitoring  Deep learning

1 Introduction Damaged or defective load-bearing components can com-


promise the robot’s structural integrity and reduce its pre-
An industrial robot consists of links and joints as main cision and safety.
structural components that define its overall geometry as To improve the overall performance and lifespan of
well as the motion range in which the robot can operate. robot structural components, researchers have been devel-
During robot operation with external loads, these structural oping various structural health monitoring methods. Early
components experience elastic deflections or perhaps even detection of structural failures is critical to mend poor
damages by causing unexpected variations or spatiotem- performances or prevent severe damages. In the first place,
poral uncertainties in robot operation performances [1, 2]. it is essential to avoid excessive forces and collisions that
Industrial robots are designed to complement humans in cause structural damages. However, in industrial settings, it
handling repetitive and hazardous tasks in harsh production is challenging to maintain a fully controlled workspace.
environments like furnaces, grinders, and welding stations. Therefore, it is crucial to know the intactness of robot
structural components. Vibration diagnostics can be one
way of identifying the intactness of robot structural com-
& Getachew Ambaye ponents. Hence, this study focuses on the vibration char-
gaambaye@shockers.wichita.edu acterization of a baseline robot link via numerical
Enkhsaikhan Boldsaikhan simulation with the intent to recognize potential formations
enkhsaikhan.boldsaikhan@wichita.edu of structural damages in robot links.
Krishna Krishnan Robotic systems are susceptible to a range of different
krishna.krishnan@wichita.edu types of faults. Experimental data and system models are
1
the two commonly used methods for detecting these faults.
Department of Industrial, Systems and Manufacturing
Engineering, Wichita State University, 1845 Fairmount St., A model-based fault detection has been implemented in
Wichita, KS 67260, USA robotic systems using different algorithms [3]. However,

123
1728 Neural Computing and Applications (2024) 36:1727–1739

accurately detecting faults can be challenging due to the is to simulate different objectives by utilizing real-time
nonlinear parameters of the models, which arise from data received from the physical system.
geometrical, material, and loading factors. Recently, sig- Natural frequency analysis is a common method used to
nificant progress has been made in estimating faults in detect the health of mechanical structural components.
actuators and sensors by transforming the nonlinear model Mechanical links can be modeled as beams with different
of steer-by-wire (SBW) into a new coordinate system [4]. assembly arrangements in a workspace. Sha et al. [19]
Tracking the trajectory of the robot link will provide conducted research on methods for detecting damages in
information about its performance and structural health beams by considering both single- and multiple damage
conditions. The task of trajectory tracking on a fixed two- locations. Researchers have explored the free vibration of
dimensional rectangular coordinate system for a mobile beams with various cross sections [20], materials including
robot has been studied through the implementation of composites [21, 22] and steel [20], and different types of
iterative learning control (ILC) [5]. cracks, such as breathing cracks [23, 24] and open cracks
Vibration significantly affects the precision of robot [25, 26]. The vibration responses of mechanical links can
movements. To address this issue, Biglari et al. [6] pro- differ due to the material type and the geometry. For
posed a frictional damper for flexible robot links. As noted example, Sahu and Das [22] studied a laminated composite
in the research paper by Sarkhel et al. [7], deflection in beam with multiple cracks and found that the natural fre-
robot links can also significantly affect precision. Deflec- quencies of defective links decreased with the increasing
tion can occur due to heavy loads by causing links to bend intensity of cracks. Taima et al. [27] investigated the sen-
or deform and resulting in inaccurate movement and failure sitivity of natural frequencies of thin and long beams to the
to perform the intended task. Deflection can also lead to crack location and depth, which was confirmed by another
loose connections and premature wear and tear on robot study [28]. The presence of defects or cracks creates a local
structural components. Robot links are subjected to gravity variation in the flexibility of the structure due to the stress
and compensatory loads, which exacerbate the adverse concentration at the vicinity of the crack tip. Vibration
effects of deflection [8]. analysis can clearly indicate the health status of structures
Power transmission gears frequently experience failures with respect to the crack type, whether a crack is open,
in industrial robots due to fatigue loading. Studies have closed, or birthing [27]. Teyi and Singh [29] reviewed
been conducted to detect faulty gears. Recently, Frej et al. research on crack localization and identification in struc-
[9] proposed a new approach to identifying defects in tural components and concluded that data tools are essen-
single-stage, six-degree-of-freedom (6-DOF) industrial tial to the procedure. Data tools can help increase the
robots using dynamic modeling and vibration analysis. In reliability and safety of structures by precisely locating
another study, localized planet gear defects in the rotating cracks. Zeng et al. [30] investigated the impact of crack
vector (RV) reducer were analyzed using an angle com- location on a rotating compressor blade (CRCB) using the
pensation strategy [10]. ANSYS software. The study aimed to analyze crack-in-
Since robot links are the primary components that duced breathing behaviors and vibration characteristics as
transmit load and motion from one joint to another joint or well as other related variables.
the end effector, damages like cracks in such structural Mian et al. [31] employed multi-sensor fault diagnosis
members can significantly impact the dynamics of the with infrared thermography (IRT) and vibration analysis to
robot. The crack shape and location affect the operation detect misalignment and imbalance between a shaft and
performances of the overall structure. Machine learning disk using machine learning. In another study, researchers
and simulation tools yield unparalleled abilities to under- compared the accuracy of using infrared thermography
stand the complex nature of structural dynamics and use (IRT) and vibration data and concluded that fault diagnosis
the knowledge for decision making. Researchers [11–16] based on IRT yields better accuracy [32]. Tao et al. [33]
have employed simulation tools to investigate the behav- have detected faults in rolling bearing using vibration
iors of mechanical components under various loading and signals under various conditions by constructing an unsu-
boundary conditions, whereas the structural members were pervised cross-domain fault diagnosis model. Cross-do-
assumed to be either rigid or flexible. Numerical simula- main fault diagnosis is a new approach to alleviating the
tions are commonly used to determine the fatigue life of problem associated with a small number of labeled target
structural members. For instance, Guo et al. [17] imple- data.
mented the finite element method in the modal analysis to Machine learning (ML) is a widely adopted approach
determine the fatigue life of an amphibious spherical robot. that aids decision making in a wide range of problems,
Phanden et al. [18] conducted a literature review, which ranging from simple to complex. Artificial neural networks
highlighted that the digital twin (DT) is a computer-gen- are machine learning models inspired by biological neu-
erated equivalent of a physical system. Its primary purpose rons. The essential settings of neural network models

123
Neural Computing and Applications (2024) 36:1727–1739 1729

include the number of neurons, activation functions, the 2D images using the GASF method for further analysis and
number of layers, and the size of each layer. Baseline classification with a convolutional neural network (CNN).
neural networks include shallow neural networks with one
or no hidden layer and deep neural networks (DNNs) with
more than one hidden layers. Despite the DNN superiority 2 Methodology
over shallow neural networks in various applications,
shallow networks still possess a significant value in several Figure 1 illustrates the overall approach of this research.
fields. For instance, they can capture the deterministic To reduce computational costs, this study employed planar
governing law of temperature in turbo engines [34] and robot link models with a fine mesh for finite element
predict the concrete strengths in civil engineering appli- analysis (FEA) in the ABAQUS software. Two major
cations [35]. To enhance the performance of neural net- analyses were carried out in this study: The first analysis
works, neural network settings can be altered for better investigated the effects of the crack geometries on the robot
classification performances [36]. link vibration, whereas a crack with a specific geometry
Convolutional neural networks (CNNs) are supervised was located at the mid-span of the link; the second analysis
machine learning models commonly used for image data studied the machine learning performances of crack
analysis as they demonstrate outstanding classification and detection for cracks located at various locations along the
generalization capabilities [37]. In general, signal sampling length of the link.
produces discrete time-series data. To leverage feature In the first analysis, six cases were considered to
extraction, Gramian Angular Fields (GAF) can be utilized examine the effect of crack geometries. The first case
to convert one-dimensional (1D) time-series data into two- involved an intact link, while the remaining cases involved
dimensional (2D) image data as described by Alsalmi et al. defective links each with a rectangular crack with a fixed
[38]. The GAF image is created by first converting the length of 1 mm facing upward at the center of the link. The
time-series data into an angle matrix and then calculating different crack geometry cases for the center crack are
the cosine and sine values of each angle. Some lightweight tabulated in Table 1. Case 2 featured a vertical crack with a
variants of the GAF method have been developed to width of 0.001 mm. The crack width was increased to
improve its scalability. GAF is an essential tool in pattern 0.01 mm and 0.1 mm for Cases 3 and 4, respectively, to
recognition because it can help extract useful features from investigate the effect of the crack width on the link
time-series data and enhance its interpretability. vibration. Cases 5 and 6 demonstrated the effect of the
Researchers use the GAF method for various applications, crack orientations at 45 in the counterclockwise direction
such as wearable sensor-based human activity recognition as well as in the clockwise direction, respectively. Every
[39], land covering mapping using satellite images [40], crack in all cases was a rectangular crack with 1 mm in
and tourism forecasting [41]. A variant of GAF is named length, which was implanted into the robot link as an open
Gramian Angular Summation Fields (GASF). The GASF crack as depicted in Figure. 2. The analysis also examined
method normalizes the input time-series data into the [- 1, the effect of the crack width and orientation on the natural
1] range and then extracts crucial features in the form of 2D frequency of the robot link.
image data for further analysis. Alternatively, cyclosta-
tionary and kurtogram maps of sensor data can be
employed as raw image data for CNN. Amin et al. [42]
simulated the wind turbine gearbox under various condi-
tions using the SIMPACK software and transformed the
acceleration data into cyclostationary and kurtogram maps.
These maps were capable of accurately classifying the fault
locations, but not the fault magnitudes.
There is a noticeable research gap in detecting structural
damages of robot components using modal analysis,
machine learning, and feature extraction in two-dimen-
sional (2D) images. To bridge this gap, this work employs
baseline robot link models with and without artificial
cracks with different sizes, locations, and orientations to
simulate their vibration responses using finite element
analysis (FEA). The feature extraction process then con-
verted the time-series data of the vibration responses into
Fig. 1 Overall approach

123
1730 Neural Computing and Applications (2024) 36:1727–1739

   
Table 1 Cases for robot link crack geometries o ohðx; tÞ owðx; tÞ
EI þ KGA  hðx; tÞ
Sn Cases Crack width Crack orientation ox ox ox
ð2Þ
o2 hðx; tÞ
1 Case 1 Intact link –  qI ¼0
2 Case 2 0.001 mm Vertical
ot2
3 Case 3 0.01 mm Vertical where E, G, I, and q are Young’s modulus, the shear
4 Case 4 0.1 mm Vertical modulus, the moment of inertia, and the mass density,
5 Case 5 0.1 mm 45 CCW respectively. w(x, t) is the transverse deflection. hðx; tÞ is
6 Case 6 0.1 mm 45 CW the slope angle of the cross-section bending. The shear
correction coefficient is denoted by K that depends on the
These geometries are applied to the crack at the center of the robot
link; CW means clockwise; CCW means counterclockwise cross-section geometry [20]. f(x, t) is an external load. The
internal transverse shear force and the bending moment are
given in Eqs. 3 and 4, respectively [44, 45].
 
owðx; tÞ
In the second analysis, the performance of crack Qðx; tÞ ¼KGA  hðx; tÞ ð3Þ
detection was studied for a rectangular crack with a width ox
of 0.5 mm and a length of 1 mm located at ten evenly ohðx; tÞ
Mðx; tÞ ¼EI ð4Þ
spaced distinct locations from the pinned end to the free ox
end of the link with a 10-mm increment.
This work used the finite element method to simulate the
2.1 Finite element analysis (FEA)
dynamics of robot link models with and without cracks
with different sizes, locations, and orientations in the
This work created a planar robot link model in the ABA-
ABAQUS software. Explicit dynamics was implemented in
QUS software to simulate and study the effect of different
the simulation to obtain the vibration responses of robot
crack geometries and locations on the vibration response of
link models. The vibration response was the acceleration of
the robot link model. The planar robot link model has a
the free end of the link. To study the features of the
pinned pivot point at the center of its eyelet end as shown
acceleration data, the feature extraction process used the
in Fig. 3. The eyelet end has an outer diameter of 10 mm
GASF method to produce 2D images from the acceleration
and an inner diameter of 5 mm. The numerical model
data. A CNN was then employed to classify and evaluate
utilized an explicit dynamics approach that consists of two
the 2D images with respect to the intactness of the baseline
steps, which are explained in Sect. 2.1.1. A quadrilateral
robot link.
element was used for finite element simulation. The
A baseline robot link model with a length of L and a
material of the planar robot link was steel with a density of
uniform cross section of A is simulated as a cantilever
786 kg/m3, an elastic modulus of 207 GPa, and a Poisson’s
beam in a quasi-static process for each case in Table 1. The
ratio of 0.3.
following formulas are the governing differential equations
The explicit dynamics approach enables the simulation
of the bending vibration dynamics [43]:
   of transient behaviors to analyze the vibration of the
o owðx; tÞ o2 wðx; tÞ structure in response to momentary external loading. By
KGA  hðx; tÞ  qA ¼ f ðx; tÞ
ox ox ot2 utilizing this approach, the aim is to demonstrate the fea-
ð1Þ sibility of detecting early stage cracks in a baseline struc-
tural link using vibration responses and machine learning.

Fig. 2 Baseline robot link with


an open vertical crack

123
Neural Computing and Applications (2024) 36:1727–1739 1731

Fig. 3 Simulation setup

2.1.1 Boundary conditions

Robot links and joints may experience extreme loading or


collisions during robot operation. For simplicity, the
boundary conditions of the simulation experiment restric-
ted the eyelet end of the planar robot link to only one
rotational degree of freedom about the pinned pivot point
as depicted in Fig. 3. Additionally, a deformable circular
object with a 3 mm radius was modeled and positioned at
145 mm and 10 mm from the center of the link’s pivot
point. In the simulation, the deformable circular object is
used to impose momentary loading on the robot link to Fig. 5 Mesh density convergence
assimilate foreign object impact. In this regard, the circular
object remains in place and then exerts a cyclic impact load Von Mises stress starts converging to 140 MPa at a mesh
on the robot link. During the first step, the pinned end of size of 0.6 mm, which consists of 2548 grid elements.
the link is subjected to a torque of 100 Nmm, while the Therefore, the simulation of all cases implemented a mesh
circular object is subjected to a unit load in the downward size of 0.4 mm. The mesh element size around the vicinity
direction for 0.2 s. Once the link and the circular object of the robot link crack was set to 0.05 mm.
come into contact, the vertical load imposed on them The accelerations of Node 8 and Node 851 have been
becomes inactive, and the interaction between them is extracted from a 1-second simulation. Figure 6 illustrates
controlled by the gravity and the torque at the eyelet end of the acceleration of Node 8 and Node 851 located at the top
the planar robot link. free end of the link as shown in Fig. 4 of the intact link. A
total of 98 nodes have been involved in the sampling of
2.1.2 Mesh density convergence acceleration including the two mentioned nodes. The
accelerations of the remaining 96 nodes near the vicinity of
The mesh density analysis has been done to carry out a the free end were used for specific verification. Likewise,
reasonable execution cost. An explicit quad plane stress the data collection of all defective robot link cases was
element with 4 nodes, as depicted in Fig. 4, has been conducted in the same way.
implemented by varying the mesh size from 5 to 0.4 mm.
Additionally, the accelerations of Node 8 and Node 851
were used for data acquisition. As illustrated in Fig. 5, the

Fig. 4 Planar model with


quadrilateral elements

123
1732 Neural Computing and Applications (2024) 36:1727–1739

The time-series data of acceleration measured at the free


end of the planar robot link are considered raw data. As
depicted in Fig. 7a, the acceleration fluctuation increases in
a highly nonlinear pattern in the first 0.3 s due to the impact
load of the circular object. The magnitude of the acceler-
ation fluctuation decreases over time because the load of
the circular object is removed after the impact on the robot
link, so that only the gravitational force and the moment
load are applied to the link. The polar representation of the
raw data is shown in Fig. 7b. It is further transformed into
2D image data using the GASF method, which is also
Fig. 6 Accelerations of Node 8 and Node 851 in the vertical direction
shown in Fig. 7c.
2.2 Dataset preparation
2.3 Convolutional neural network (CNN)
The time-series datasets from the simulation experiments
have then been converted into 2D images using the method This study used a convolutional neural network to classify
called Gramian Angular Summation Field (GASF). In this 2D images of the acceleration data. The neural network
case, the time-series acceleration data have been normal- architecture is composed of six hidden layers in addition to
ized and transformed from the Cartesian coordinate system the input and output layers as shown in Fig. 8. The first
into a polar coordinate system. This transformation keeps convolutional layer generates a tensor with dimensions of
the temporal information of the input data point. The next 387  387  16, which is then subjected to max-pooling to
step is to identify temporal relationships between any two reduce the feature dimensions to 193  193  16. Another
data points in the polar coordinate system. The [n, n]-di- convolutional layer and max-pooling layer follow this
mensional Gramian matrix can then be produced by process. The nodes use rectified logic unit (ReLU) as an
applying trigonometric functions, whereas n is the number activation function. The final max-pooling layer output is
of samples in a certain time period. Let S ¼ ½s1 ; s2 ; s3 ; . . .sn  fully connected to a dense layer with a soft-max activation
function for classification into one of the six classes.
denote an acceleration signal with n samples, and let S ¼
In this work, the CNN training was carried out using a
½s1 ; s2 ; s3 ; . . .sn  denote the normalized version of S in the
backpropagation algorithm, which has a learning rate of
interval [- 1,1], which can be given by the form in Eq. 5,
0.0001. The CNN model contains 36,973,035 trainable
where i (=1, ..., n) is a time index. Then, the angle u is
parameters, with a batch size of 64. The neural network
computed using Eq. 6.
architecture employed in this study is suitable for analyzing
si  minðSÞ 2D input images and has potential applications in computer
si ¼ ð5Þ
maxðSÞ  minðSÞ vision and pattern recognition.
 Following the preparation of the datasets, the CNN
u ¼ arccosðSÞ ð6Þ
model is trained and tested on a cloud-based Google Collab
Temporal correlation between adjacent time-series points server to take advantage of the computational speed.
(i, j) is computed by summing the angles of the adjacent Accuracy and loss performance metrics have been used to
points (i, j), where i and j are the time indices of the data evaluate the performance of the neural network. Accuracy
points. Thus, this Gramian matrix is also referred to as is the fraction of correct predictions in the overall predic-
Gramian Angular Summation Field (GASF). These for- tion set. Formally, for multi-category classification, the
mulations can be written as in Eqs. 7–9: simple mathematical representation of accuracy is given in
Eq. 10. Utilizing accuracy as a performance evaluation
GASF ¼ ½cosðui þ uj Þ ð7Þ metric becomes imperative when the consideration of both
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
false/true–negative predictions is required. Losses also
GASF ¼ ST  S ðI  S2 Þ  ðI  S2 Þ ð8Þ
serve as performance indicators, assuming a crucial role in
2 3 detecting and mitigating prediction inaccuracies. Ulti-
cosðu1 þ u1 Þ    cosðu1 þ un Þ
6 cosðu þ u Þ    cosðu þ u Þ 7 mately, the supervised training of the CNN model aims to
6 2 1 2 n 7
GASF ¼ 6 7 ð9Þ maximize the overall accuracy of CNN predictions.
4    5
Number of correct predictions
cosðun þ u1 Þ  cosðun þ un Þ Accuracy ¼ ð10Þ
Total number of predictions
Supervised training determines good values for CNN

123
Neural Computing and Applications (2024) 36:1727–1739 1733

Fig. 7 Feature extraction: a raw


time-series data; b polar
representation of raw data; c 2D
image data

Fig. 8 Neural network


architecture. Note: ReLU is a
rectified linear unit used as an
activation function

weights and biases using labeled training datasets. The loss is a penalty score for poor predictions. In other words, the

123
1734 Neural Computing and Applications (2024) 36:1727–1739

loss is a statistic that represents how inaccurate the pre- 3.2 The effects of crack geometries
diction is for a specific example. The objective of super- on the natural frequency of the robot link
vised training is to identify a collection of CNN weights
and biases that, on average, have the lowest possible loss To understand the effects of crack geometries on the nat-
across all cases. A popular loss function is a mean square ural frequency of the robot link, finite element simulation
error (MSE) that is the average square loss for the whole has been conducted. Figure 10 illustrates the natural fre-
dataset. To calculate MSE, one needs to sum up all the quencies of the six mode shapes exhibited by the robot
squared losses for individual examples and then divide links of the cases that are tabulated in Table 1. In Mode 1,
them by the number of examples as given in Eq. 11. the maximum natural frequency is associated with the
1 intact robot link, which is the Case-1 robot link without
MSE ¼ Rða;bÞD ðb  predictionðaÞÞ2 ð11Þ cracks. Mode 2, Mode 4, and Mode 5 also exhibit com-
N
parable natural frequency patterns for the robot link cases.
where (a, b) is a labeled sample in which a is the set of In all vibration modes, Case 2, Case 3, and Case 4
features that the classification uses to make predictions, b is demonstrate how the natural frequency drops as the crack
the target, i.e., healthy, breathing crack, open crack, etc. width rises.
prediction(.) is a neural network model that makes pre- Another interesting point is that, compared to the ver-
dictions, D is the dataset of labeled samples, and N is the tical cracks involved in Case 1, Case 2, and Case 3, the
number of samples in D. robot link displays greater natural frequencies when the
crack orientation is at 45 in either the clockwise or
counterclockwise direction. In Case 6, the robot link with a
3 Results and discussion crack oriented at 45 in the clockwise direction displays a
little larger natural frequency than the one with a crack
3.1 Vibration modes oriented at 45 in the counterclockwise direction. This
trend can perhaps be attributed to the boundary conditions
Free vibration analysis has been conducted to investigate and the impact force arrangement implemented in the
the natural frequencies of the intact robot link. Six free simulation with respect to the crack orientation. Mode 3
vibration mode shapes of the intact robot link with one produced nearly equal natural frequencies in all cases,
fixed eyelet end are depicted in Fig. 9, where each mode except the Case-2 robot link with an insignificant vertical
shape is shown as a contour plot. The planar simulation crack. The maximum crack width that is considered in this
results show that the first four mode shapes and the sixth study is only 1 mm, which is small in comparison with the
mode shape exhibit pitch motion patterns. On the other size of the robot link. Based on the simulation results, it can
hand, a yaw motion is exhibited by the fifth mode shape. be assumed that the crack width has a negative effect on
the natural frequency, which means larger crack widths
decrease the natural frequency.

Fig. 9 Vibration mode shapes due to free vibration of intact robot link with one fixed eyelet end

123
Neural Computing and Applications (2024) 36:1727–1739 1735

Fig. 10 Mode shapes, natural


frequencies, and cases for robot
link crack geometries

3.3 The effects of crack locations on the robot is located far from the constrained end. Additionally, the
link vibration acceleration variation is low after 0.3 s in comparison with
the responses within the first 0.3 s.
In this experimentation, a crack that is 0.5 mm wide and The fast Fourier transform (FFT) is a discrete math
1 mm long has been applied to various locations along the function that reveals the frequency components of time-
robot link. The same crack is applied to a robot link at a series data [46, 47]. As presented by the normalized fre-
location between 10 and 100 mm from the center of the quency spectrum in Fig. 12, higher magnitudes are exhib-
pinned eyelet end. Detection and classification of these ited by the low-frequency components of the link with a
cracks are then explored with the CNN model. Figure 11 crack at 100 mm. Generally, the magnitude of acceleration
illustrates the Node 8 acceleration responses for the intact increases as the crack is located further away from the
robot link and the robot links with respective cracks at pinned end during the first few time increments.
50 mm and 100 mm from the pinned end. The vibration To evaluate whether the cracks at different locations are
responses of the robot link are not smooth when the crack detectable, ML has been used to detect and classify the

Fig. 11 Acceleration responses of Node 8 Fig. 12 Normalized frequency spectrum of Node-8 acceleration data

123
1736 Neural Computing and Applications (2024) 36:1727–1739

links with the same crack geometries, but various crack


locations. Accordingly, a convolutional neural network
(CNN) was trained with 836 images in total. The validation
set involved 231 images, which is about 20% of the total
image data. Based on the results, the accuracy and the loss
of classification are 96.17% and 0.1237, respectively, as
depicted in Fig. 13. Both the accuracy and the loss are
converging within 30 epochs.
Figure 14 illustrates the accuracy of detecting the crack
at varying locations along the robot link. Cracks closer to
the fixed end are more detectable than the cracks further
away from the fixed end. This is reasonable because a
crack located far from the fixed end has a lower effect on
the acceleration of the free end. The crack that is closest to
the fixed end of the robot link lowers the acceleration
response of the free end by damping the impact load.

3.4 Classification of planar robot links Fig. 14 Confusion matrix for robot link with cracks at different
with different crack geometries locations. Note: Cr means crack

The simulation experiment described in the methodology robot links. This outcome is especially useful for industries
section involved the six different cases of intact and that rely on the structural integrity of equipment and
cracked planar robot links to obtain their acceleration/vi- materials. Furthermore, the confusion matrix in Fig. 16
bration responses. The obtained acceleration data are then illustrates that 100% accuracy has been achieved for all the
used to evaluate the detectability of the cracks with dif- cases, except the third and the fourth cases, which exhibit
ferent geometries. Accordingly, the simulation experiments 95% accuracy. Overall, the CNN model was able to detect
produced 588 datasets, whereas 456 datasets (80%) were cracks with different geometries at a certain location along
used for training, 126 datasets were used for validation, and the robot link more accurately than the crack at different
the remaining datasets were used for testing. The training locations along the robot link.
and validation datasets were used to optimize and fine-tune The proposed method offers a new way of detecting
the CNN model, while the testing dataset was used to structural defects, such as cracks, in robot links using their
evaluate the classification performance with novel datasets. vibration responses and deep learning with feature extrac-
The results of the simulation were analyzed and are pre- tion in two-dimensional images. To demonstrate the effi-
sented in Fig. 15, which shows the final accuracy of cacy of the method, this study uses planar robot link
98.25% and the loss of 0.0606 after 150 training epochs. models with and without artificial cracks to simulate and
These results indicate that the CNN model is good enough collect their vibration responses in ABAQUS via finite
to detect structural defects in the specimen. Although the element analysis (FEA). The collected vibration responses
finite element simulation considers ideal conditions, these are then transformed into two-dimensional images for deep
results demonstrate that vibration sensing and machine learning neural networks to learn and assess the effects of
learning can be used to detect cracks in baseline planar different artificial cracks on the vibration responses of the

Fig. 13 Evolution of
classification performances for
detecting cracks at different
locations: a accuracy (Eq. 10);
b loss (Eq. 11)

123
Neural Computing and Applications (2024) 36:1727–1739 1737

Fig. 15 Evolution of
classification performances for
detecting cracks with different
geometries: a accuracy
(Eq. 10); b loss (Eq. 11)

Fig. 16 Confusion matrix for


robot link with different crack
sizes

robot link models. This study not only demonstrates a ranging from 0.001 to 0.1 mm, crack orientations including
successful use of neural computing in conjunction with 45 CCW, vertical, and 45 CW, and crack locations
modal analysis for robot health monitoring but also ranging from 10 to 100 mm from the pinned end of the
unleashes a new opportunity to recognize structural defects robot link. The robot link acceleration response, which is
in their early stages of formations using FEA, modal 1D time-series data, is collected and transformed into 2D
analysis, and deep learning. image data using the GASF method. The 2D image data
have been fed to the CNN model for evaluation and clas-
sification. The results indicate that, as the crack width
4 Conclusions decreases, the natural frequency of the robot link decreases.
Additionally, the crack orientation also affects the natural
This study demonstrated that the size and location of small frequency. The 45 CCW cracks exhibited slightly higher
cracks have an impact on the vibrations of planar robot natural frequencies than that of the 45 CW ones, when a
links. Finite element analysis (FEA) has been conducted to momentary external load is applied to the free end of the
study the vibrations of planar robot links with crack widths link in the counterclockwise direction. The crack at

123
1738 Neural Computing and Applications (2024) 36:1727–1739

different locations along the robot link was detected by the 2. Boldsaikhan E (2020) Measuring and estimating rotary joint axes
proposed ML method with 96.17% accuracy. Moreover, of an articulated robot. IEEE Trans Instrum Meas
69(10):8279–8287
cracks with different geometries at a specific location along 3. Hasan A, Tahavori M, Midtiby HS (2023) Model-based fault
the robot link were detected with 98.25% accuracy. Cor- diagnosis algorithms for robotic systems. IEEE Access
respondingly, this study draws the following conclusions: 11:2250–2258
4. Djordjević V, Stojanović V, Pršić D, Dubonjić L, Morato MM
• The intact robot link had a higher natural frequency (2022) Observer-based fault estimation in steer-by-wire vehicle.
than that of cracked robot links. Eng Today 1(1):7–17
5. Zhuang Z, Tao H, Chen Y, Stojanovic V, Paszke W (2022) An
• The crack orientation affects the natural frequency as
optimal iterative learning control approach for linear systems
the slanted cracks exhibit higher natural frequencies with nonuniform trial lengths under input constraints. IEEE
than that of the vertical cracks. Transactions on Systems, Man, and Cybernetics: Systems
• Cracks located farther away from the fixed end exhibit 6. Biglari H, Golmohammadi M, Hayati S, Hemmati S (2021)
Vibration reduction of a flexible robot link using a frictional
higher acceleration amplitudes.
damper. J Vib Control 27(9–10):985–997
• Cracks with varying geometries at a specific location 7. Sarkhel P, Dikshit MK, Pathak VK, Saxena KK, Prakash C,
along the robot link are detected more accurately than Buddhi D (2023) Robust deflection control and analysis of a
the constant crack at different locations along the robot fishing rod-type flexible robotic manipulator for collaborative
robotics. Robot Auton Syst 159:104293
link.
8. Xu P, Yao X, Liu S, Wang H, Liu K, Kumar AS, Lu WF, Bi G
• This work not only demonstrates a successful use of (2021) Stiffness modeling of an industrial robot with a gravity
neural computing in conjunction with modal analysis compensator considering link weights. Mech Mach Theory
for robot health monitoring but also unleashes a new 161:104331
9. Frej A, Chiementin X, Fakher C, Bolaers F, Haddar M (2023)
opportunity to recognize structural defects in their early
Dynamic modeling and vibration analysis for defect identification
stages of formations using FEA, modal analysis, and of single-stage gearboxes in the joints of industrial robots with six
deep learning. DOF. Proc Inst Mech Eng Part C 09544062231152347
10. Yan Y, Guo Y, Liu X (2023) Tooth root crack detection of planet
It should be noted that the robot link models used in this gear in industrial robot RV reducer. Meas Control
study are planar and the crack shapes are rectangular for 00202940231180619
simplicity. Such geometries may not provide accurate 11. Shi M, Rong B, Liang J, Zhao W, Pan H (2023) Dynamics
representations of actual robot links. Therefore, more analysis and vibration suppression of a spatial rigid-flexible link
manipulator based on transfer matrix method of multibody sys-
realistic digital 3D robot link models along with more tem. Nonlinear Dyn 111(2):1139–1159
realistic crack types must be considered in the future work 12. Ambaye GA, Lemu HG (2021) Dynamic analysis of spur gear
to further evaluate this proposed machine learning method. with backlash using ADAMS. Mater Today Proc 38:2959–2967
Ultimately, this new machine learning method should be 13. Kouritem SA, Abouheaf MI, Nahas N, Hassan M (2022) A multi-
objective optimization design of industrial robot arms. Alex Eng J
tested with actual vibration data collected from actual robot 61(12):12847–12867
manipulators with intact and defective structural compo- 14. Hu M, Wang H, Pan X, Liao L, Sun H (2022) Elastic deformation
nents. In this case, the vibration of structural components modeling of series robots with consideration of gravity. Intell
can be captured by vibration sensors to validate the simu- Serv Robot 15(3):351–362
15. Seth A, Kuruvilla JK, Sharma S, Duttagupta J, Jaiswal A (2022)
lation results. Design and simulation of 6-DOF cylindrical robotic manipulator
using finite element analysis. Mater Today Proc 62:1521–1525
16. Ambaye GA, Lemu HG (2021) Effect of backlash on transmis-
Funding The article received no funding from any private or gov- sion error and time varying mesh stiffness. In: Advanced manu-
ernment organization. facturing and automation, Springer, vol X 10, pp 18–28
17. Guo S, He Y, Shi L, Pan S, Tang K, Xiao R, Guo P (2017) Modal
Data availability The data presented in this study are available on and fatigue analysis of critical components of an amphibious
request from the corresponding author. spherical robot. Microsyst Technol 23:2233–2247
18. Phanden RK, Sharma P, Dubey A (2021) A review on simulation
in digital twin for aerospace, manufacturing and robotics. Mater
Declaration Today Proc 38:174–178
19. Sha G, Radzieński M, Cao M, Ostachowicz W (2019) A novel
Conflict of interest The authors declare no conflict of interest. method for single and multiple damage detection in beams using
relative natural frequency changes. Mech Syst Signal Process
132:335–352
20. Altunışık AC, Okur FY, Karaca S, Kahya V (2019) Vibration-
References based damage detection in beam structures with multiple cracks:
modal curvature vs modal flexibility methods. Nondestruct Test
1. Qiao G, Weiss BA (2019) Industrial robot accuracy degradation Eval 34(1):33–53
monitoring and quick health assessment. J Manuf Sci Eng 21. Song M, Gong Y, Yang J, Zhu W, Kitipornchai S (2020) Non-
141(7):071006 linear free vibration of cracked functionally graded graphene

123
Neural Computing and Applications (2024) 36:1727–1739 1739

platelet-reinforced nanocomposite beams in thermal environ- 36. Kalaiselvi T, Padmapriya S, Somasundaram K, Praveenkumar S
ments. J Sound Vib 468:115115 (2022) E-tanh: a novel activation function for image processing
22. Sahu S, Das P (2020) Experimental and numerical studies on neural network models. Neural Comput Appl
vibration of laminated composite beam with transverse multiple 34(19):16563–16575
cracks. Mech Syst Signal Process 135:106398 37. Rodriguez-Conde I, Campos C, Fdez-Riverola F (2022) Opti-
23. Wei C, Shang X (2019) Analysis on nonlinear vibration of mized convolutional neural network architectures for efficient on-
breathing cracked beam. J Sound Vib 461:114901 device vision-based object detection. Neural Comput Appl
24. Long H, Liu Y, Liu K (2019) Nonlinear vibration analysis of a 34(13):10469–10501
beam with a breathing crack. Appl Sci 9(18):3874 38. Alsalemi A, Amira A, Malekmohamadi H, Diao K (2023)
25. Tam M, Yang Z, Zhao S, Yang J (2019) Vibration and buckling Lightweight Gramian angular field classification for edge internet
characteristics of functionally graded graphene nanoplatelets of energy applications. Clust Comput 26(2):1375–1387
reinforced composite beams with open edge cracks. DMaterials 39. Qin Z, Zhang Y, Meng S, Qin Z, Choo K-KR (2020) Imaging and
12(9):1412 fusing time series for wearable sensor-based human activity
26. Chinka SSB, Putti SR, Adavi BK (2021) Modal testing and recognition. Inf Fusion 53:80–87
evaluation of cracks on cantilever beam using mode shape cur- 40. Abidi A, Ienco D, Abbes AB, Farah IR (2023) Combining 2d
vatures and natural frequencies. In: Structures, Elsevier, vol 32, encoding and convolutional neural network to enhance land cover
pp 1386–1397 mapping from satellite image time series. Eng Appl Artif Intell
27. Taima MS, El-Sayed TA, Shehab MB, Farghaly SH, Hand RJ 122:106152
(2022) Vibration analysis of cracked beam based on Reddy beam 41. Simonyan K, Zisserman A (2014) Very deep convolutional net-
theory by finite element method. J Vib Control works for large-scale image recognition. Preprint arXiv:1409.
10775463221122122 1556
28. Zhao K, Liu Y, Du J (2023) Vibration characteristics and power 42. Amin A, Bibo A, Panyam M, Tallapragada P (2023) Vibration
flow analysis of a constant cracked beam with general boundary based fault diagnostics in a wind turbine planetary gearbox using
conditions. Int J Appl Mech machine learning. Wind Eng 47(1):175–189
29. Teyi N, Singh S (2022) A review of application of data science 43. Rao SS (2019) Vibration of continuous systems. Wiley, USA
tools in crack identification and localization. Procedia Struct 44. Lee U (2009) Spectral element method in structural dynamics.
Integr 39:608–623 Wiley, Republic of Korea
30. Zeng J, Chen K, Ma H, Duan T, Wen B (2019) Vibration 45. Sarvestan V, Mirdamadi HR, Ghayour M (2017) Vibration
response analysis of a cracked rotating compressor blade during analysis of cracked Timoshenko beam under moving load with
run-up process. Mech Syst Signal Process 118:568–583 constant velocity and acceleration by spectral finite element
31. Mian T, Choudhary A, Fatima S (2023) Multi-sensor fault method. Int J Mech Sci 122:318–330
diagnosis for misalignment and unbalance detection using 46. Boldsaikhan E, Corwin E, Antonette L, Arbegast W (2011) The
machine learning. IEEE Transactions on Industry Applications use of neural network and discrete Fourier transform for real-time
32. Mian T, Choudhary A, Fatima S (2023) Vibration and infrared evaluation of friction stir welding. App Soft Comput
thermography based multiple fault diagnosis of bearing using 11(8):4839–4846
deep learning. Nondestruct Test Eval 38(2):275–296 47. Emmanuel S, Yihun Y, Nili Ahmedabadi Z, Boldsaikhan E
33. Tao H, Qiu J, Chen Y, Stojanovic V, Cheng L (2023) Unsuper- (2021) Planetary gear train microcrack detection using vibration
vised cross-domain rolling bearing fault diagnosis based on time- data and convolutional neural networks. Neural Comput Appl
frequency information fusion. J Frankl Inst 360(2):1454–1477 33:17223–17243
34. Bukovsky I (2021) Deterministic behavior of temperature field in
turboprop engine via shallow neural networks. Neural Comput Publisher’s Note Springer Nature remains neutral with regard to
Appl 33(19):13145–13161 jurisdictional claims in published maps and institutional affiliations.
35. Asteris PG, Mokos VG (2020) Concrete compressive strength
using artificial neural networks. Neural Comput Appl
32(15):11807–11826

123

You might also like