Article citation info: 1

Ouarhlent S, Terki N, Hamiane M, Dahmani H. Fault detection in robots based on discrete wavelet transformation and eigenvalue of
energy. Diagnostyka. 2023;24(4):2023407. https://doi.org/10.29354/diag/174124.

e-ISSN 2449-5220
DIAGNOSTYKA, 2023, Vol. 24, No. 4
DOI: 10.29354/diag/174124
1

FAULT DETECTION IN ROBOTS BASED ON DISCRETE WAVELET

TRANSFORMATION AND EIGENVALUE OF ENERGY
Saloua OUARHLENT 1,* , NadjibaTERKI 1 , Madina HAMIANE 2 , Habiba DAHMANI 3
1
Department of Electrical Engineering, University of Biskra,Biskra, Algeria.
2
College of Engineering, Royal University for Women, Bahrain.
3
Department of Electrical Engineering, University of Mohamed Boudiaf, M’Sila, Algeria.
*
Corresponding author, e-mail: [email protected]

Abstract
This article addresses the problem of fault detection in robot manipulator systems. In the production field,
online detection and prevention of unexpected robot stops avoids disruption to the entire manufacturing line.
A number of researchers have proposed fault diagnosis architectures for electrical systems such as induction
motor, DC motor, etc..., utilising the technique of discrete wavelet transform. The results obtained from the use
of this technique in the field of diagnosis are very encouraging. Inspired by previous work, The objective of
this paper is to present a methodology that enables accurate fault detection in the actuator of a two-degree of
freedom robot arm to avoid system performance degradation. A partial reduction in joint torque constitutes
the actuator fault, resulting in a deviation from the desired end-effector motion. The actuator fault detection is
carried out by analysing the torques signals using the wavelet transform. The stored energy at each level of the
transform contains information which can be used as a fault indicator. A Matlab/Simulink simulation of the
manipulator robot demonstrates the effectiveness of the proposed technique.

Keywords: 2 DOF robot, fault detection System. Discrete wavelet transform (DWT). Energy eigen value

List of Symbols/Acronyms fault detection [3, 4], such as fast Fourier transform
(FFT), short-time Fourier transform (STFT), and
DWT– Discrete Wavelet Transform; discrete wavelet transform (DWT). Despite its
FFT – Fast Fourier Transform; common use in industry for fault diagnosis, fast
PID-Proportional-Integral-Derivative;
Fourier transform (FFT) analysis remains unsuitable
Two DOF robot–Two degree of freedom robot;
for signals of a transient nature and rapidly reveals
its limitations [5, 6]. As a result, the discrete wavelet
1. INTRODUCTION
transform (DWT) of the torque signal has been
suggested as an alternative to overcome the
The use of robotic systems has the potential to
shortcomings of the FFT. The wavelet transform can
improve our lives by increasing efficiency, reducing
provide information simultaneously in both the time
human error, and enabling us to explore and achieve
and frequency domains of the signal under
things that would otherwise be impossible. For most
examination, while offering a time-frequency
of these uses indeed, any failure or malfunction of a
representation of the signal [5-7].
robot can have catastrophic and costly
Under normal operating conditions, each joint in
consequences. The integration of fault diagnosis
robotic systems moves at different angular velocities
(FD) capabilities into robotic systems allows for
(and accelerations), requiring varying torques, thus
early detection of the presence of a fault and
resulting in discrete, short-term varying signals[8].
localization of its source.
On the other hand, signals from a faulty state also
Recently, several fault diagnosis and detection
have non-stationary behaviour. If the Fourier
techniques have been proposed in the field of
transform is used to calculate the frequency
robotics, through the integration of qualitative and
component of nonstationary signals, the results will
quantitative information models [1, 2]. Signal
not provide information about the temporal location
processing has a significant role in the construction
of the regime shift in the signal. In contrast, time-
of any condition monitoring system. To this end,
frequency analysis provides information about the
several signal analysis techniques have been used for
composition and frequency variation of the
introduced signal at different times [9].

Received 2023-06-07; Accepted 2023-10-13; Available online 2023-10-23

© 2023 by the Authors. Licensee Polish Society of Technical Diagnostics (Warsaw. Poland). This article is an open access article distributed
under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).
2 DIAGNOSTYKA, Vol. 24, No. 4 (2023)
Ouarhlent S, Terki N, Hamiane M, Dahmani H: Fault detection in robots based on discrete wavelet …

The discrete wavelet transform (DWT) algorithm 2. PROBLEM FORMULATION AND

is well-established in various signal processing MATHEMATICAL MODELLING
research areas because it provides a time-frequency
representation for non-stationary time-varying 2.1. The dynamic model
signals[10]. The Discrete Wavelet Transform Through the use of the Lagrange-Euler
(DWT) was implemented to solve the resolution formalism, the dynamic model of a robot arm having
problem associated with the Short-Time Fourier n degrees of freedom (n_DOF) can be
mathematically represented as [14]:
Transform (STFT)'s fixed window size. Moreover,
there are faults that are described as catastrophic. An
𝜏 = 𝑀(𝜃)𝜃̈ + 𝐶(𝜃, 𝜃̇) + 𝐺(𝜃) + 𝐹(𝜃̇ ) +
example of such a fault is a seal failure [11]. Other
𝜂(𝜃, 𝜃̇, 𝜏, 𝑡) (1)
examples of a catastrophic fault in an actuator is a
short circuit, a voltage drop or power loss in an Where:
electric motor. 𝜃, 𝜃̇ , 𝜃̈ ∈ 𝑅𝑛 denote the vectors of joint positions,
velocities, and accelerations, respectively; 𝜏 ∈ 𝑅𝑛 is
In our work, we use a robot arm. In general
the vector of input torques, 𝑀(𝜃) ∈ 𝑅𝑛×𝑛 is the
robotic arms are widely used in manufacturing and
inertia matrix whose inverse exists, 𝐶(𝜃, 𝜃̇ ) ∈ 𝑅𝑛 is
industrial settings, where they can perform repetitive
the vector representing Coriolis and centripetal
and precise tasks such as welding, painting, and
forces, 𝐺(𝜃) ∈ 𝑅𝑛 is the vector of gravitational
assembly. They are also used in medical
torque, 𝐹(𝜃̇) ∈ 𝑅𝑛 is a vector containing the
applications, such as surgery and rehabilitation, as
unknown static and dynamic friction terms,
well as in research and exploration, such as in space
and 𝜇( 𝜃, 𝜃̇, 𝜏, 𝑡) ∈ 𝑅𝑛 encompasses all the terms
exploration or deep sea exploration. There are
that account for unmodeled dynamics and external
various types of robotic arms, including cartesian,
disturbances .
cylindrical, polar, and articulated arms, each with
different advantages and application domains. The 2.2. Failures of the robotic arm
choice of arm depends on the specific task and In robotic systems, the occurrence and magnitude
requirements of the application. Robotic arms are of faults are subject to time-dependent variations and
typically controlled by a computer system that sends are influenced by multiple parameters [12].
commands to the arm's actuators that move the arm's
joints to the desired position and orientation. The The fault dynamics can be generally represented as
control system can be programmed to perform follows:
specific tasks, or can be operated in real-time using 𝐹(𝜃, 𝜃,̇ 𝜏, 𝑡) = 𝛽(𝑡 − 𝑇)𝑓𝑚 (𝜃, 𝜃,̇ 𝜏) (2)
sensors such as cameras or force sensors.
In this work, we have focused only on faults that 0 𝑖𝑓 𝑡 < 𝑇
𝛽(𝑡 − 𝑇) = { (3)
are not severe and which can be handled using 1 𝑖𝑓 𝑡 ≥ 𝑇
standard control methods. [12, 13]. To control our Where:
robot arm with two degrees of freedom (2 DoF), we 𝑓𝑚 (𝜃, 𝜃,̇ 𝜏) ∈ 𝑅𝑛 is a vector that represents the fault
consider the use of the proportional integral in the robot manipulator,𝛽(𝑡 − 𝑇)represents the
derivative (PID) controller. We study the detection appearance or non-appearance of the fault and T is
the time of occurrence of the fault. The fault fm is
of actuator faults in robot manipulators, particularly
given by:
faults that affect the joint drive systems. For accurate
fault diagnosis, we adopted a discrete wavelet
𝑓𝑚 (𝜃, 𝜃,̇ 𝜏) = 𝑓𝑚𝜃 (𝜃, 𝜃̇) + 𝑓𝑚𝜏 (𝜏) (4)
transform (DWT)-based time-frequency signal
analysis to extract the most salient fault-related Where:
features. We applied the fault detection method to 𝑓𝑚𝜏 (𝜏)and𝑓𝑚𝜃 (𝜃, 𝜃̇) represent torque-dependent and
the motor torque signal. state-dependent faults respectively.
This paper is organised as follows: In Section 2, The dynamics of the manipulator with failure is
we discuss the problem of actuator fault diagnosis. defined by:
Then, the PID controller for the robot arm is ⏟
𝜏 =
discussed in Section 3. Section 4, the discrete 𝐼𝑛𝑝𝑢𝑡 𝑇𝑜𝑟𝑞𝑢𝑒

wavelet transform method is detailed. The 𝑀(𝜃)𝜃

⏟ ̈ + 𝐶(𝜃, 𝜃̇) + 𝐺(𝜃) + 𝐹(𝜃̇) + 𝜂(𝜃, 𝜃̇, 𝜏, 𝑡) +
simulation results obtained from the proposed 𝑅𝑜𝑏𝑜𝑡𝑖𝑐 𝑆𝑦𝑠𝑡𝑒𝑚𝑠 𝐷𝑦𝑛𝑎𝑚𝑖𝑐𝑠
control technique and diagnostic method are
presented and discussed in Section 5. Finally, the 𝛽(𝑡 − 𝑇)𝑓𝑚 (𝜃, 𝜃,̇ 𝜏)
⏟ (5)
conclusions and prospects for future work are 𝐹𝑎𝑢𝑙𝑡 𝐷𝑦𝑛𝑎𝑚𝑖𝑐𝑠

presented in Section 6. Actuators in robot manipulators commonly

consist of electric motors. Faults occurring in
rotating electric motors can be categorised into three
 DIAGNOSTYKA, Vol. 24, No. 4 (2023) 3
Ouarhlent S, Terki N, Hamiane M, Dahmani H: Fault detection in robots based on discrete wavelet …

types: electrical faults, rotational faults, and benefits of the DWT is its capability to analyse
vibration faults. signals that are non-stationary, which are signals that
The mathematical model of the electric faults in change over time or have varying frequency
motors is given by the following equation [15]: components.
𝑓(𝜏) = 𝛼𝜏 − 1 < 𝛼 ≤ 𝐻 < ∞ (6) Figure 1 presents a diagram that illustrates the
Where H is the maximum value that the parameter DWT process of performing multilevel signal
α can take. 𝛼 ∈ ℝ decomposition. After calculating the approximation
and detail coefficients at various levels of
2.3. The PID controller decomposition, it is possible to reconstruct the
For the robotic system (1) the control objective is approximation and detail signals at each level,
to follow a predetermined trajectory. The PID facilitating the extraction of features.
controller is the most efficient choice for robot By employing the discrete wavelet transform, it
control because it offers a combination of is possible to conduct a multiresolution analysis of
operational simplicity and ease of implementation. the torque signals in both healthy and faulty states.
The robot's mechanism is modelled as a This analysis, described in [4], involves
decoupled linear system, with each joint controlled decomposing the analysed torque signal S into
by a PID type controller. The fault detection scheme components at multiple levels. At the initial level,
developed in this paper is independent of the type of the signal is decomposed into an approximation
control applied to the robot system. component (𝑎1 ) and a detail component (𝑑1 ).
Under faultless operating conditions, the formula Mathematically, these coefficients can be expressed
used for finding the control signal τ is given as
as:
𝜏(𝑡) = 𝐾𝑃 (𝜃𝑑 − 𝜃) + 𝐾𝐷 (𝜃̇𝑑 − 𝜃̇ ) + 𝐾𝐼 ∫ (𝜃𝑑 − 𝜃)𝑑𝑡 (7)
𝑎 = ∑𝑛𝑘 𝐿(𝑘 − 2𝑛) ∗ 𝑆𝑖 (𝑘)
{ 1 (8)
Where: 𝑑1 = ∑𝑛𝑘 𝐻(𝑘 − 2𝑛) ∗ 𝑆𝑖 (𝑘)
𝜃𝑑 , 𝜃̇𝑑 , 𝜃̈𝑑 are the vectors of desired joint positions,
The subsequent decomposition level is built upon
velocities, and accelerations, respectively.
the coefficient (𝑎1 ). In this process, (𝑎1 ) is
e(t) = θd (𝑡) − θ(t) is the joint tracking error
decomposed into another approximation component
between the desired trajectory and the actual one.
(𝑎2 ) and another detail component (𝑑2 ), and this
𝐾𝑃 , 𝐾𝐷 , 𝐾𝐼 are 2× 2 constant, diagonal and positive
iterative decomposition continues further as
matrices.
illustrated in Figure 1. The coefficients at level 2 can
Moreover, we have the following assumptions:
be mathematically represented as follows:
Assumption 1: After a fault occurs, the states of the
𝑎 = ∑𝑛𝑘 𝐿(𝑘 − 2𝑛) ∗ 𝑎𝑖 (𝑘)
robotic system remain bounded; { 2 (9)
𝑑2 = ∑𝑛𝑘 𝐻(𝑘 − 2𝑛) ∗ 𝑎𝑖 (𝑘)
i.e., 𝜃(𝑡), 𝜃̇ (𝑡) ∈ 𝐿∞
Assumption 2: The vector field 𝜇(𝑡)representing the Following the decomposition process, the
unknown additive disturbances and noise is both original signal S can be reconstructed in the
bounded and small, meaning that it remains within following manner:
limited and negligible ranges, i.e., 𝑆(𝑡) = 𝑎1 + 𝑑1 = (𝑎2 + 𝑑2 ) + 𝑑1
|𝜇𝑖 (𝑡)| ≤ 𝜇̅𝑖 (𝑡), ∀ 𝑡 ≥ 0, 𝑖 = 0, . . 𝑛 = (𝑎3 + 𝑑3 ) + 𝑑2 + 𝑑1 (10)
Where 𝜇̅𝑖 (𝑡) is a small value constant. or more generally,
𝑆(𝑡) = 𝑎𝑛 + 𝑑𝑛 + 𝑑(𝑛−1) + ⋯ 𝑑2 + 𝑑1 (11)
2.4. Discrete wavelet transform
The wavelet transform (DWT) is a mathematical There are many types of wavelets, such as Haar,
tool employed for processing and analysis of signals. Daubechies, Symlet, etc. Proper selection of the
It is a way of representing a signal as a sum of mother wavelet type and the number of
decomposition levels is crucial prior to applying the
wavelets, which are small waves that are scaled and
DWT [16]. Haar wavelet has some important
translated in time. The DWT has the advantage of
properties that make it useful in signal processing
being able to capture both frequency and time- and image compression. In this work, The Haar
domain information, which makes it useful in many wavelet is chosen as a mother wavelet for the
different applications, such as image compression, discrete wavelet transform due to its excellent time
audio processing, and data analysis. The DWT localization properties and its straightforward
works by dividing a signal into smaller segments, or hardware implementation. The important property of
"sub-bands", at different scales or resolutions. Each the Haar transform is that it preserves the energy of
sub-band contains information about a different the signals. This means that the total energy of a
range of frequencies, with the highest frequencies in signal remains unchanged after undergoing the Haar
the smallest sub-bands and the lowest frequencies in transform [17].
the largest sub-bands [3, 15].
The DWT can be used for a variety of
applications, such as denoising, feature extraction,
compression, and data analysis. One of the main
4 DIAGNOSTYKA, Vol. 24, No. 4 (2023)
Ouarhlent S, Terki N, Hamiane M, Dahmani H: Fault detection in robots based on discrete wavelet …

Fig. 1. Multi-level (n) signal decomposition using DWT

The Haar wavelet is a mathematical function The eigenvalue E is calculated based on the
named after the Hungarian mathematician Alfréd DWT decomposition level (j), DWT decomposition
Haar, who first introduced it in 1909. The Haar time (n), and the magnitude of the wavelet
wavelet is a simple piecewise function that has a coefficient (H) of the signal at each discrete point in
compact support, meaning that it is zero outside a the corresponding frequency band.
finite interval. It consists of a square wave that starts The energy values at different signal
at zero and ramps up to one, followed by a negative decomposition levels contain valuable information
square wave that starts at one and ramps down to that can be utilized for diagnosing and quantifying
zero. In wavelet analysis, two main functions are the degree of torque faults.
utilized: the scaling function ∅, also known as the
father wavelet, and the wavelet 𝜓(𝑡) also referred to 3. SIMULATION RESULTS
as the mother wavelet. The Haar mother wavelet
function can be described as[18]: The simulation model is a two-link planar
1 manipulator as depicted in Figure 3 [20].
1 0≤𝑡< , The variables 𝜃1 , 𝐿1 , 𝑀1 represent the joint angle,
2
𝜓(𝑡) = 1 (12) length, and mass of the first link (𝑖 = 1),
−1 ≤𝑡 < 1, respectively. Similarly, 𝜃2 , 𝐿2 , 𝑀2 represent the joint
2
{0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 . angle, length, and mass of the second link (𝑖 = 2),
respectively. g represents the gravitational
Whose graph is shown in Figure 2. acceleration. The robot links are assumed to be
uniform rods of length 1 m and masses 3 kg and 2
kg, respectively. The simulation is implemented in
Simulink/Matlab.

Fig. 2. The Haar wavelet 𝜓(𝑡)

The Haar wavelet's lack of continuity and
differentiability can be advantageous for analyzing
signals with abrupt transitions. This is because the
Haar wavelet can effectively capture and represent
sharp changes in signals, making it suitable for
detecting faults or abrupt changes in machine
systems.[19]. Fig. 3. Two-link robot arm
The wavelet can detect the presence of a fault
when a torque fault occurs. Diagnosis is performed Where 𝜃, 𝜃̇, 𝜃̈ 𝑎𝑛𝑑 𝜏 represent position, velocity,
by observing and comparing the decomposition acceleration and control torque.
levels that provide information about the fault. By 𝑀 𝑀21
𝑀(𝜃) = [ 11 ] is the inertia matrix with the
assessing the associated eigenvalue for each level, it 𝑀21 𝑀22
is possible to construct a highly efficient diagnostic following elements:
tool capable of quantifying the severity of faults. 𝑀11 = (𝑀1 + 𝑀2 )𝐿1 2 + 𝑀2 𝐿2 2 + 2𝑀2 𝐿1 𝐿2 cos𝜃2 ,
The energy eigenvalue E corresponding to each 𝑀12 = 𝑀21 = 𝑀2 𝐿2 2 + 𝑀2 𝐿1 𝐿2 cos𝜃2 ,
frequency band is expressed as follows [4]: 𝑀22 = 𝑀2 𝐿2 2 ,
2
𝐸𝑗 = ∑𝑘=𝑛
𝑘=1 |𝐻𝑗,𝑘 (𝑛)| (13)
 DIAGNOSTYKA, Vol. 24, No. 4 (2023) 5
Ouarhlent S, Terki N, Hamiane M, Dahmani H: Fault detection in robots based on discrete wavelet …

𝐶11 3.1. First simulation: absence of faults for the

𝐶(𝜃, 𝜃̇) = [ ]
𝐶21 first and the second motors
Represents the vector of Coriolis and centrifugal forces, In this simulation, the first and second motors
where: have been selected free of faults. This healthy state
2
𝐶11 = −𝑀2 𝐿1 𝐿2 (2𝜃̇1 𝜃̇2 + 𝜃̇2 ) 𝑠𝑖𝑛𝜃2 , is illustrated in Figure 4 and Figure 5.
2
𝐶21 = 𝑀2 𝐿1 𝐿2 𝜃̇2 𝑠𝑖𝑛𝜃2 ,

𝐺11
𝐺(𝜃) = [ ]=
𝐺21
−(𝑀1 + 𝑀2 )𝑔𝐿1 𝑠𝑖𝑛𝜃1 − 𝑀2 𝑔𝐿2 sin (𝜃1 + 𝜃2 )
[ ]∶
−𝑀2 𝑔𝐿2 sin(𝜃1 + 𝜃2 )
is a vector of gravity torques.
In the field of control system design, selecting
appropriate numerical values for controller gains,
Fig. 4. Joint positions without fault
such as 𝐾𝑃 , 𝐾𝐷 , 𝑎𝑛𝑑 𝐾𝐼 is a critical step in creating an
effective control system. These gains dictate the
controller's response to system errors and have a
profound impact on achieving the desired
performance and stability of the system. There are
several conventional methods for tuning PID
controllers, each with its unique advantages and
applicability. These methods encompass the
Ziegler–Nichols (ZN) method [21], the Cohen–Coon
Technique [22], manual tuning [23], optimization
techniques [24], and the utilization of PID tuning Fig.5. Joint position error
software [25], among others [26]. Additionally, PID
controllers can be tuned using various intelligent It is seen from the figures that the actual trajectory of
methods such as fuzzy logic, artificial neural the position perfectly follows the desired trajectory,
networks (ANN), adaptive neuro-fuzzy inference The PID control objective is therefore achieved.
systems (ANFIS), genetic algorithms (GA), and
more [27, 28]. The choice of tuning method depends 3.2. Second simulation: actuator faults
on the specific system under study, and the control In this second simulation we have considered two
objectives. types of faults, which are:
In our work, the selection of numerical values for • Bias fault on 𝜏1 (first joint) for 𝑡∈ [3 ÷ 3.5] sec
the PID gains (𝐾𝑃 , 𝐾𝐷 , 𝑎𝑛𝑑 𝐾𝐼 ) is a critical step in • Partial fault on second joint for 𝑡∈ [3.4 ÷ 4.6]
achieving the desired control system performance. sec
To determine these values, we have chosen manual In this simulation, a failure modelled by equation
tuning using simulation tools, namely (6), which is a reduction in motor torque, occurs on
MATLAB/SIMULINK. This approach enables us to the first motor (1) at t=3sec. Similarly, a partial fault
fine-tune the PID controller parameters based on the on the second motor occurs at time t=3.4sec, which
system's response to ensure optimal control system results in the deviation of the actual position
performance. This decision is motivated by the trajectory (faulty state) from the desired trajectory
advantages of simulation, which allow us to explore (healthy state) starting from this time onwards, as
and test a wide range of PID parameter combinations illustrated in Figures 6 and 7. After the occurrence of
in a controlled and risk-free virtual environment. By a fault, it becomes clear that the local PID control
conducting simulations, we can observe how law at joints 1 and 2 loses its ability to maintain the
different PID gain values impact the system's joint position on its desired trajectory. Fig. 8 presents
response, thus facilitating the fine-tuning process. a comparison between the command torque (blue
This approach ensures that our chosen PID values line) and the actual torque (red line).
align precisely with our project's control objectives
and result in optimal system performance.
The controller gains in the control law (7) has been
chosen as follows:
𝐾𝑃 = 500𝐼2×2 , 𝐾𝐼 = 500 𝐼2×2 𝑎𝑛𝑑 𝐾𝐷 = 500𝐼2×2 ,
where 𝐼2×2represents a 2 × 2 identity matrix .
The desired joint trajectories to be tracked have been
selected as follows:
𝜃𝑑1 (𝑡) = 2 ∗ sin (0.3𝜋𝑡) Fig. 6. Joint positions with faults
𝜃𝑑2 (𝑡) = (−2.6) ∗ sin (0.3𝜋𝑡)
6 DIAGNOSTYKA, Vol. 24, No. 4 (2023)
Ouarhlent S, Terki N, Hamiane M, Dahmani H: Fault detection in robots based on discrete wavelet …

higher decomposition levels in effectively

identifying and capturing the fault-related
information.
Comparison between the results of the DWT
analysis of healthy and faulty states indicates that
when the fault level increases, the amplitude
coefficients 𝑎7 , 𝑑7 , 𝑑6 and 𝑑 5 also increase. The
same results are obtained regarding the oscillations.
Fig. 7. Joint position error This simulation demonstrates how wavelet
analysis can effectively detect the precise moment
when a signal undergoes a change, as well as the
nature of that change, whether it is a sudden
disruption in the signal, an abrupt shift in
oscillations, or a significant variation in amplitude.
Such information enables us to assess the operational
status of the robot.
Fig. 8. Control vs. actual torques

The DWT is then used in order to extract the

main components of a fault (Torque signal in faulty
state). Figures 9 and 10 show the obtained results
after conducting the DWT (The sub-signals from the
DWT analysis) on the motor torque signals in
healthy and faulty states of the robot. There are 5
sub-figures, the remaining sub-figures illustrate the
generated sub-signals that represent the detail
coefficients (d6, d5, and d7) as well as the
Fig.9. First joint
approximation coefficient (a7). Each sub-signal
corresponds to a distinct frequency band. In this
manner, the healthy torque signals and the faulty
torque signals are decomposed into seven levels. For
each level of decomposition, the detail component
corresponds to a high-frequency range, while the
approximation component encompasses the low
frequencies.
Figures 9.a and 10.a represent the DWT analysis
of the motor in a healthy state, showing the signal Fig. 10. Second joint
characteristics under normal operating conditions. In
contrast, Figures 9.b and 10.b present the DWT The torque energy can be calculated for each
analysis of faulty signals, which exhibit distinct frequency band using Equation (12). Tables 1
features such as discontinuity in value and sudden through 4, give the energy values at various
frequency changes. These characteristics suggest the decomposition levels in the simulation of the first
presence of abnormalities or faults in the motor and the second joints. Furthermore, It is important to
system. The observation reveals that the faults start note that the Haar transform guarantees the
to become apparent at level 5, as previous levels conservation of signal energies.
typically lack the capability to detect these
occurrences. This highlights the significance of

Table 1. Nomenclature of Frequency Band Numbers and Energy Distribution Across Various Scales.
Control torque of the first joint: No faults detected
Scale Range of Frequencies (Hz) Frequency band number Energy (Joules)
d1 1250–2500 b1 0.0620
d2 625–1250 b2 0.0615
d3 312.5–625 b3 0.0290
d4 156.25–312.5 b4 0.0243
d5 78.125 –156.25 b5 0.0288
d6 39.0625 – 78.125 b6 0.0576
d7 19.5312 – 39.0625 b7 0.4456
a7 0 – 19.5312 b8 19.2174
∑ 19.9262

Original signal 20.2986

 DIAGNOSTYKA, Vol. 24, No. 4 (2023) 7
Ouarhlent S, Terki N, Hamiane M, Dahmani H: Fault detection in robots based on discrete wavelet …

Table 2. Nomenclature of Frequency Band Numbers and Energy Distribution Across Various Scales.
Actual torque of the first joint: Occurrence of fault
Scale Range of Frequencies (Hz) Frequency band number Energy (Joules)
d1 1250–2500 b1 0.0582
d2 625–1250 b2 0.0610
d3 312.5–625 b3 0.0295
d4 156.25–312.5 b4 0.0246
d5 78.125 –156.25 b5 0.0427
d6 39.0625 – 78.125 b6 0.0740
d7 19.5312 – 39.0625 b7 0.5235
a7 0 – 19.5312 b8 24.8464
∑ 25.6599

Original signal 26.2470

Table 3. Nomenclature of Frequency Band Numbers and Energy Distribution Across Various Scales.
control torque of the second joint: No faults detected
Scale Range of Frequencies (Hz) Frequency band number Energy (Joules)
d1 1250–2500 b1 0.6492
d2 625–1250 b2 0.6946
d3 312.5–625 b3 0.2669
d4 156.25–312.5 b4 0.1513
d5 78.125 –156.25 b5 0.1023
d6 39.0625 – 78.125 b6 0.1259
d7 19.5312 – 39.0625 b7 0.2926
a7 0 – 19.5312 b8 9.0698
∑ 11.3526

Original signal 11.6735

Table 4. Nomenclature of Frequency Band Numbers and Energy Distribution Across Various Scales.
Actual torque of the second joint: Occurrence of fault
Scale Range of Frequencies (Hz) Frequency band number Energy (Joules)
d1 1250–2500 b1 0.6503
d2 625–1250 b2 0.6748
d3 312.5–625 b3 0.2625
d4 156.25–312.5 b4 0.1822
d5 78.125 –156.25 b5 0.1466
d6 39.0625 – 78.125 b6 0.2000
d7 19.5312 – 39.0625 b7 0.3910
a7 0 – 19.5312 b8 11.5076
∑ 14.0150

Original signal 14.5206

The analysis of the calculated torques energies at across different levels of the torque signal, reflecting
each frequency band with fault and without fault, the healthy and faulty states of the robot. Notably, an
can be used for fault detection. This is because the upsurge in energy is observed in both the first and
energy of the torque may be amplified in certain second joints, which can serve as an indicator of the
frequency bands due to the generated damage. fault. Despite the fault being characterized by a
After the occurrence of a fault, the relative decrease in actuator torque, it is noteworthy that an
energy of the signal tends to increase in certain increase in energy is generated, primarily attributed
frequency bands, resulting in a significant disparity to the influence of the PID regulator.
between the energy calculations before and after the The PID regulator's response to the fault, by
fault. increasing the actuator command, helps maintain the
Therefore, changes in energy levels within desired performance of the system, albeit at the cost
specific frequency bands serve as indicators of the of consuming more energy. The increase in energy
motor's condition, and the energy content in the can be seen as a consequence of the PID regulator's
decomposed frequency bands of the signal holds effort to counteract the effect of the fault and
valuable information about the severity of damage. maintain system stability.
Simulation results depicted in Figures 11a and The comparison of the obtained results with other
11b demonstrate the variation in stored energy techniques from the literature (signal processing
technique)[29], namely FFT (Fast Fourier
8 DIAGNOSTYKA, Vol. 24, No. 4 (2023)
Ouarhlent S, Terki N, Hamiane M, Dahmani H: Fault detection in robots based on discrete wavelet …

Transform) and STFT (Short-Time Fourier

Transform), sheds light on the effectiveness of the
chosen Discrete Wavelet Transform (DWT)
approach for actuator fault detection in the 2 DOF
robot arm context.
The evaluation of the diagnostic techniques is
based on an analysis of the advantages and
disadvantages of the presented.
As indicated in reference [29], it can be inferred
that each diagnostic approach has the capacity to
identify various forms of damage and faults during
the robot's operation, contingent upon the objectives
established.
The FFT and STFT techniques are suitable for
rapid and uncomplicated detection, particularly
when applied to steady signals or signals that exhibit
low levels of noise. These techniques are
characterized by their simplicity and do not
necessitate supplementary sensors and
computational power [30], they are, however, less
effective for transient signals and are better suited
for stationary signals or those with minimal noise.
Based on the results presented in this article and
previous research [4, 31], it can be concluded that
DWT (Discrete Wavelet Transform) is remarkable
for its capability to provide a specialized time-
frequency representation tailored for non-stationary,
time-varying signals. This characteristic makes
DWT particularly suitable for analysing transient
signals, enabling a thorough assessment of signal
components across both temporal (time) and spectral
(frequency) domains. Furthermore, the work
presented in this paper introduces the idea of
evaluating torque energy across different frequency Fig. 11. Histograms of the torque energy at
ranges through DWT analysis, thus emphasizing that each frequency band (for various
alterations in energy levels within these specific decomposition levels)
frequency ranges can act as indicators of the the
motor's state. extracted features can then be utilized to build a
comprehensive database comprising the most
4. CONCLUSION AND FUTURE WORK common faults encountered in the system. Hence, it
would be possible to design a state monitoring
In conclusion, the proposed method based on the system for the robot that is capable of detecting
Discrete Wavelet Transform for detecting torque- various electrical and mechanical faults using the
related faults in the actuator of a 2 DOF robot arm is proposed method. This system would enable
shown to be effective in detecting different types of continuous monitoring of the robot's performance,
faulty and normal operating conditions. The hence offering real-time feedback to enhance its
method's advantages in both the time and frequency overall reliability and efficiency.
domains make it an efficient and useful tool for
actuator fault detection. Author contributions: research concept and design,
As a potential improvement, the sub-signals S.O.; Collection and/or assembly of data, S.O.; Data
obtained from the DWT analysis can be directly used analysis and interpretation, S.O., N.T.; Writing the article,
as inputs to various classification schemes, such as S.O., N.T.; Critical revision of the article, H.D., M.H.;
the ANN classifier, for more accurate fault detection. Final approval of the article, N.T.
Moreover, it would be worthwhile to explore the
application of condition monitoring techniques using Declaration of competing interest: The authors declare
other types of signals, such as vibration, electrical, that they have no known competing financial interests or
personal relationships that could have appeared to
and sound signals, in order to enhance the overall influence the work reported in this paper.
effectiveness of robot condition monitoring.
Adequate processing these signals is of utmost
importance to extract the most relevant features
associated with specific types of faults. These
 DIAGNOSTYKA, Vol. 24, No. 4 (2023) 9
Ouarhlent S, Terki N, Hamiane M, Dahmani H: Fault detection in robots based on discrete wavelet …

REFERENCES http://dx.doi.org/10.4028/www.scientific.net/AMM.3
93.953.
1. Patton R., Uppal F, Lopez-Toribio C. Soft computing 17. Walker JS. A primer on wavelets and their scientific
approaches to fault diagnosis for dynamic systems: a applications. CRC press 2008.
survey. IFAC Proceedings Volumes 2000; 33(11): 18. Hariharan G, Kannan K. Review of wavelet methods
303-315. https://doi.org/10.1016/S1474- for the solution of reaction–diffusion problems in
6670(17)37377-9. science and engineering. Applied Mathematical
2. Alobaidy MAA, Abdul-Jabbar JM, Al-khayyt SZ. Modelling 2014; 38(3): 799-813.
Faults diagnosis in robot systems: a review. Al- https://doi.org/10.1016/j.apm.2013.08.003.
Rafidain Engineering Journal (AREJ) 2020; 25(2): 19. Lee B. Application of the discrete wavelet transform
164-175. to the monitoring of tool failure in end milling using
https://doi.org/10.33899/rengj.2020.127782.1051. the spindle motor current. The International Journal of
3. Jaber AA, Bicker R. Industrial robot backlash fault Advanced Manufacturing Technology 1999; 15(4):
diagnosis based on discrete wavelet transform and 238-243. https://doi.org/10.1007/s001700050062.
artificial neural network. American Journal of 20. Guechi E-H, et al., MPC control and LQ optimal
Mechanical Engineering 2016; 4(1): 21-31. control of a two-link robot arm: A comparative study.
http://dx.doi.org/10.12691/ajme-4-1-4. Machines 2018; 6(3): 37.
4. Ameid T, Menacer A, Talhaoui H, Azzoug Y. Discrete https://doi.org/10.3390/machines6030037.
wavelet transform and energy eigen value for rotor 21. Patel VVJR. Ziegler-Nichols Tuning Method:
bars fault detection in variable speed field-oriented Understanding the PID Controller. Resonance 2020;
control of induction motor drive. ISA transactions 25(10): 1385-1397. http://dx.doi.org/10.1007/s12045-
2018; 79: 217-231. 020-1058-z.
https://doi.org/10.1016/j.isatra.2018.04.019. 22. Utami AR, Yuniar RJ, Giyantara A, Saputra AD.
5. Yang WY. Signals and Systems with MATLAB: Cohen-Coon PID tuning method for self-balancing
Springer Science & Business Media 2009. robot. 2022 International Symposium on Electronics
6. Mallat S. A wavelet tour of signal processing. 1999: and Smart Devices (ISESD) 2022.
Elsevier. https://doi.org/10.1109/ISESD56103.2022.9980830.
7. Mallat SG. A theory for multiresolution signal 23. Sahrir NH, Mohd Basri MA. Modelling and manual
decomposition: the wavelet representation. IEEE tuning PID control of quadcopter. Control,
transactions on pattern analysis and machine instrumentation and mechatronics: theory and practice
intelligence 1989; 11(7): 674-693. 2022; 921: 346-357. https://doi.org/10.1007/978-981-
https://doi.org/10.1109/34.192463. 19-3923-5_30.
8. Jaber A, Bicker R. Industrial robot fault detection 24. Zárate-Ramos J, Rodríguez-Hernández J, Cruz-
based on wavelet transform and LabVIEW. First Domínguez J, Nieto-Gutiérrez N, Sánchez-López C.
International Conference on Systems Informatics, Arbitrary order PID controller design for an inverted
Modelling and Simulations. School of Mechanical and pendulum system. 2023 International Conference on
System Engineering, Newcastle University, UK 2014. Fractional Differentiation and Its Applications
http://dx.doi.org/10.1109/SIMS.2014.27. (ICFDA) 2023.
9. Bae H, Kim YT, Kim S, Lee S, Wang BH. Fault https://doi.org/10.1109/ICFDA58234.2023.10153271.
detection of induction motors using fourier and 25. Bakošová M, Oravec J, Čirka Ľ. Software for PID
wavelet analysis. Journal of Advanced Computational controller tuning. V Proceedings of the 17th
Intelligence and Intelligent Informatics 2004; 8(4): Interantional Conference on Porcess Control’09,
431-436. FCFT SUT in Bratislava 2009.
10. Baig S, Farrukh F, Mughal M. Discrete wavelet 26. Bucz S, Kozakova A. Advanced methods of PID
transforms-algorithms and applications. Intech 2011. controller tuning for specified performance. PID
11. Freddi A Longhi S, Monteriu A, Ortenzi D, Proietti Control for Industrial Processes: Books 2018.
Pagnotta D. Fault tolerant control scheme for robotic 27. Sathish Kumar A, Naveen S, Vijayakumar R, Suresh
manipulators affected by torque faults. IFAC- V, Asary AR, Madhu S, Palani K. An intelligent fuzzy-
PapersOnLine 2018; 51(24): 886-893. particle swarm optimization supervisory-based control
12. Kmelnitsky VM. Automated On-line Diagnosis and of robot manipulator for industrial welding
Control Configuration in Robotic Systems Using applications. Scientific Reports 2023; 13(1): 8253.
Model Based Analytical Redundancy. Worcester https://doi.org/10.1038/s41598-023-35189-2.
Polytechnic Institute 2002. 28. Dachang Z. Baolin D, Puchen Z, Shouyan C. Constant
13. Leuschen ML, Walker ID, Cavallaro JR. Investigation force PID control for robotic manipulator based on
of reliability of hydraulic-robots for hazardous fuzzy neural network algorithm. Complexity 2020;
environments using analytic redundancy. Annual 2020: 1-11. https://doi.org/10.1155/2020/3491845.
Reliability and Maintainability. Symposium 1999 29. Autsou S, Rassolkin A, Vaimann T, Kudelina K.
Proceedings (Cat. No. 99CH36283) 1999. Analysis of possible faults and diagnostic methods of
https://doi.org/10.1109/RAMS.1999.744107 the Cartesian industrial robot. Proceedings od the
14. Spong MW, Hutchinson S, Vidyasagar M. Robot Estonian Academy of Sciences 2022; 71(3): 227-240.
modeling and control. John Wiley & Sons 2020. https://doi.org/10.3176/proc.2022.3.04.
15. Jaber AA, Design of an intelligent embedded system 30. Ouamara D, Boukhnifer M, Chaibet A, Maidi A.
for condition monitoring of an industrial robot. Diagnosis of ITSC fault in the electrical vehicle
Springer 2016. powertrain system through signal processing analysis.
16. Ngui WK, Salman Leong M, Hee LM, Abdelrhman Diagnostyka 2023; 24(1): 2023113.
AM. Wavelet analysis: mother wavelet selection https://doi.org/10.29354/diag/161309.
methods. Applied mechanics and materials 2013; 393: 31. Halder S, Bhat S, Zychma D, Sowa P. Broken rotor bar
953-958. fault diagnosis techniques based on motor current
10 DIAGNOSTYKA, Vol. 24, No. 4 (2023)
Ouarhlent S, Terki N, Hamiane M, Dahmani H: Fault detection in robots based on discrete wavelet …

signature analysis for induction motor—A review. She works as a researcher in INRS in Canada at Institute
Energies 2022; 15(22): 8569. of Energy, Materials and Telecommunications at Quebec
https://doi.org/10.3390/en15228569. University in Montreal for finishing Her Ph.D.
Dissertation.Shehas a Postdoctoral position as a researcher
Saloua OUARHLENT Was at ÉTS(Écolede Technologie Supérieure), Quebec
born in Batna, Algeria, on University, Montreal, Canada in 2016. Her primary
March 1974. She received research interests include signalprocessing, healthy and
B.Sc., M.Sc. degrees in pathological speech processing,digital communications
Electronics from the and Currently, Babies’ Crying signal.
University of Batna, Algeria, E.mail:[email protected]
respectively in 1999 and 2009.
Currently, she is a Teaching
Assistant at the Electrical
Engineering Department,
Faculty of Science and
Technology. University of Biskra, Algeria. She works on
her PhD thesis on fault detection for robot manipulators at
the University of Biskra, Algeria.
E.mail : [email protected]

Nadjiba TERKI received her

Engineering diploma in
automatics, Magister diploma in
non destructive control and PhD
degree in signal processing from
Badji Mokhtar University of
Annaba, Algeria,in 1994, 2000
and 2009, respectively. In 2001,
shejoined Mohamed Khider
University of Biskra-Algeria,
where she works currently as
Professor at the Genie Electrics
department.
Her research interests include Digital signal and Image
processing, artificial intelligence, wavelet transform and
robust control. E.mail:[email protected]

Madina HAMIANE received

her DES in Electronics from
the University of Science and
Technology Houari Boume-
diene (USTHB), Algeria; her
MSc degree in Cybernetics
and Instrument Technology
from the University of Read-
ing, UK, and her PhD degree
in Control Engineering from
the University of Sheffield,
UK
She is now with the College of Engineering at Royal
University for Women in the Kingdom of Bahrain.
Dr.Hamiane’s current research interests span applications
of signal processing and pattern recognition, biomedical
signal and image processing, and intelligent control.
E.mail:[email protected]

Habiba DAHMANI associate

professor at Electrical
Engineering Department,
Faculty of technology,
Mohamed Boudiaf University
in Msila, Algeria. She received
her Bachelor of electronic
Engineering in in 1995, Master
in signal processing in 2003
and her Ph.D. in 2015 from
Badji Mokhtar University,
Algeria