Automatic Anti-swing Gantry Crane based on

Gradient Descent optimization Proportional

Integrated Derivative (PID): A Comparison
Simulation and Real Experiment
Munzir Qadri Danang Karuna Harjanto Jonathan Miharja
Automotive and Robotics Program, Automotive and Robotics Program, Automotive and Robotics Program,
Computer Engineering Department, BINUS Computer Engineering Department, BINUS Computer Engineering Department, BINUS
ASO School of Engineering, Bina Nusantara ASO School of Engineering, Bina Nusantara ASO School of Engineering, Bina Nusantara
University, University, University,
Jakarta, Indonesia 11480 Jakarta, Indonesia 11480 Jakarta, Indonesia 11480

Winda Astuti Yosica Mariana Sofyan Tan

Automotive and Robotics Program, Product Design program, Industrial Automotive and Robotics Program,
Computer Engineering Department, BINUS Engineering Department, BINUS ASO Computer Engineering Department, BINUS
ASO School of Engineering, Bina Nusantara School of Engineering, Bina Nusantara ASO School of Engineering, Bina Nusantara
University, University, University,
Jakarta, Indonesia 11480 Jakarta, Indonesia 11480 Jakarta, Indonesia 11480

wastuti@binus.edu Mahmud Iwan Solihin

Departement of Mechanical &
Mechatronics Engineering,
UCSI University, Kuala Lumpur, Malaysia

Abstract— Transporting payload with precise swing angles control the crane and precisely halt the swing, further
and transfer times poses a challenging task in building impeding the unloading process. Failure to effectively
construction, which is why gantry crane systems are regulate the crane can result in accidents and jeopardize the
extensively utilized. Various control strategies, including open- well-being of individuals involved.
loop and closed-loop systems, have been proposed to tackle this
issue. However, most of these controllers rely on a model and
Several attempts have been undertaken to govern gantry
parameters specific to the crane system, making the process of crane operations through the utilization of open and closed-
modeling and parameter identification complex and time- loop control systems. Manson (1992) and Auernig et.al
consuming. To address this challenge, this work presents a (1981) have employed open-loop time optimal strategies in
practical and simulation control method for automatic gantry crane control [1]–[3]. However, these studies yielded
cranes, which is evaluated simulation and experimentally. The unsatisfactory results as open-loop approaches are
lab-scale Gantry crane is developed in this work, which susceptible to variations in system parameters like rope
consists of mechanical and electrical parts. Proportional length and do not consider the influence of wind
Integral and Derivative (PID) and Proportional Derivative (PD) disturbances. Another similar technique known as input
controllers which applied optimization method are
implemented in the electrical part of the system. The PID and
shaping has been suggested by Teo et.al (1998) and
PD controller applied in controlled the position and swing Singhose et.al (1997), but it remains an open-loop
angle, respectively. The Gradient Decent optimization method method[3].
is the best optimization method in determining the parameter Efforts have been made to regulate the gantry crane
of the controller, since it has the best simulation performance system by employing feedback control, which is more
in the output of the system. The experimental result is resilient to disturbances and parameter fluctuations
developed and compared between simulation and lab-scale compared to open-loop control. Omar has suggested PD
Gantry crane hardware implementation part. The controllers for position and anti-swing control, but they are
performance of experimental results is shown appropriately not highly efficient in eliminating steady state error. PID
approaching the performance of the simulation result, with the
approximately settling time.
controllers have also been proposed, but their effectiveness
diminishes when the actuator reaches saturation. Yang and
Keywords—Crane, PID controller, PD controller, Gradient Yang have introduced an adaptive control strategy, but it
Decent optimization method, Lab-scale Gantry Crane. necessitates nonlinear control theory and intricate
mathematical analysis[4][5].
I. INTRODUCTION In general, the control strategies mentioned above
exhibit commendable performance when the precise model
In construction projects, gantry cranes are frequently and its parameters are utilized in controller design, a task
utilized for the efficient transportation of heavy loads and typically handled by a skilled control engineer. However,
hazardous materials. The primary goal is to swiftly relocate the modeling and parameter identification processes
the load to its destination while minimizing unnecessary involved are intricate and time-consuming, while advanced
movement. However, conventional gantry cranes often controllers often prove too complex for real-time
exhibit swinging when the payload abruptly stops after rapid implementation in crane systems. To tackle these challenges,
motion. Mitigating this swinging is a time-consuming this paper introduces a practical and intelligent control
process that negatively impacts operational productivity. approach for automatic gantry cranes that eliminates the
Additionally, it requires a skilled operator to manually

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE

need for modeling or system identification. The proposed
method is based on a straightforward open-loop experiment
and is specifically designed for easy implementation by non-
expert engineers. Its effectiveness and ability to handle
parameter variations are assessed through experimental
evaluations, comparing it to classical PID, with implement
the various optimization method to get the best PID
parameter. The experimental results vividly demonstrate the
superior performance and robustness of the proposed control
method over the other controllers. The remaining sections of Figure 2. Trolley crane model
this paper are organized to provide a detailed presentation of
the proposed control method and a comprehensive The position of the swinging load with respect to the
evaluation of its performance. horizontal plane can be represented as xm and vertically as
The structure of the remaining sections in this paper is as ym, and can be formulated as:
follows. Section 2 provides a summary of the proposed �� = � + � sin � 1
system, lab scale gantry crane, modelling and controller �� =− � cos � . 2
applied while Section 3 outlines the discuss the result of the With the Lagrangian method of the kinetic and potential
experiment. The PID controller for anti-swing control is energy of the system, the result of the equation:
discussed in Section 4. Section 5 the paper concludes with � = �1 + �2 � + �1 �� 3
final remarks in the last section. � + �� ++ �� = 0. 4
The derivation of the dynamic model for a 2-D gantry crane
II. PROPOSED SYSTEM
prototype is based on the application of Lagrange's equation.
2.1 Lab Scale Gantry Crane The diagram of the gantry crane mechanism is depicted in
The proposed system is applied and develop in the lab- Figure 1, with the variables m1, m2, l, x, θ, B, g, and F
scaled gantry crane system, as shown in Figure 1. system representing the payload mass, trolley mass, cable length,
system consists of mechanical part and electrical part. The horizontal position of the trolley, swing angle, damping
mechanical part involved, an actuation mechanism for coefficient, gravitational constant, and driving force,
payload transfer, position and swing angle sensors, and real- respectively. By applying the Lagrange equation, the
time control. The electrical part consists of DC motor and its equation of motion for the trolley crane system is obtained.
driver are employed, with the payload connected to it. The The translational motion of the gantry crane
lab-scale gantry crane system solely focuses on the planar prototype is driven by a motor. By modeling the dynamic
movement of the trolley, assuming a fixed load and string motor according to the equivalent motor circuit, the overall
length. Galanized hollow iron, Stepper motor Nema 17, transfer function model for the gantry crane can be achieved.
wheel, A4988 motor driver, gyro sensor (MPU-6050), It is known that the torque of the motor (T) is proportional
Arduino UNO, copper wire, and Polivinil Clorida (PVC) to the armature current and the back emf voltage is also
applied on the system. The length, weight and hight of the proportional to the rotational speed (ω). And from the motor
lab-scaled gantry crane are 1m, 0.5, and 0,7 meters, circuit it is also known that there are armature resistance (R)
respectfully. The load weight light is 0.5 kg applied to the and induction (L). Using Kirchoff's law, the relationship
system. between current and voltage can be written as in formula
(18).
� = �� . � 5
�� = �� � = �� �� 6
��
� = �. � + � + �� . 7
��
The gantry crane system also has a pulley mechanism to
convert the rotational torque from the motor shaft into
translational motion of the trolley. With r being the radius of
the pulley, J being the total moment of inertia, and b being
the friction force constant, the total torque can be formulated
Figure 1. Lab-Scale Gantry Crane as follows:
�� . � − ��� − �� = ��� . 8
2.2 Dynamic Model of Gantry Crane The final linear transfer function result, as shown in
The derivation of the dynamic model for a 2-D gantry crane equation below.
prototype is based on the application of Lagrange's equation. � � �0
The diagram of the gantry crane mechanism is depicted in = 2
9
� � � �2 � + �1 � + 1
Figure 2, with the variables m1, m2, l, x, θ, B, g, and F Θ � − �2
representing the payload mass, trolley mass, cable length, = 2 10
� � �� + �
horizontal position of the trolley, swing angle, damping
coefficient, gravitational constant, and driving force,
respectively. By applying the Lagrange equation, the The mathematical model of the lab-scale gantry crane was
equation of motion for the trolley crane system is obtained. developed, and its parameters are identified [10]. The
developed gantry crane model is only used to design
classical PID controller, which is used as a comparator, and neural network learning models. This algorithm works by
to make simulation. The developed model of the crane is [10] finding the minimum point of the function used. Within the
starting points of the function, the algorithm shifts the
�(�) 5.10
=�(0.0788�2+0.9804�+1) (11) solution in the negative direction of the gradient to reach a
�(�)
minimum. The model will continue to adjust the parameters
until it reaches the smallest error until the function is close
�(�) −�2
= (12) to 0 or equal to 0. With this algorithm, the cost function can
�(�) 50�2+981
be minimized, and can predict the error between the
predicted and actual results. The iteration process of the
where U(s), X(s) and Θ(s) are input voltage, trolley gradient descent algorithm is as follows [9].
displacement and load swing angle respectively. ��+1 = �� − �� ∇� �� = �� − �� � �� . (15)
where �� > 0:
2.3 Control Design � �� − �� � �� = min� �� − �� � �� . (16)
�� >0
Figure 3 depicts how the gantry crane's anti-swing control is
where λ is step size, ∇ is gradient operator of the
set up. Gradient Decent (GD) Optimization is used to
function f(x), and g(xi) is the gradient at the current point.
perform limited optimization for the anti-swing gantry crane
By moving to the point where the function f reaches its
system tuning of resilient PID and PD controllers. The goal
minimum value, the derivative equation of the direction can
(fitness) function and constraint are the two components that
be written as follows:
are most crucial for a constrained optimization to succeed. � �
Due to the popularity and simplicity of this strategy, the �(��+1 ) = ∇� ��+1 � ∙ �
optimization's goal is to reduce the error function of the ��� ��� �+1
system response. The system's closed-loop poles must be in = ∇� ��+1 � � �� . (17)
a position that complies with Kharitonov's interval If equation (17) equals 0, and by applying the step size
polynomials theorem [6]. To provide strong stability across (λ) of the functions ∇f(X_(i+1) ) and g(X_i ), then the
the system's spectrum of parameter fluctuations, this is direction derivative of the function will be orthogonal in
proposed. shape or have a zig-zag pattern. The correct value for λ is
very important to obtain. If the value of λ is too small, it can
increase the convergence time, while a value that is too
large can cause divergence. The exact value of λ to produce
a stable condition is as follows;
� ��+1 ≤ � �� . (18)

2.4.2 Simplex Search Optimization

The Simplex search optimization method, also known
Figure 3. PID diagram of the system
as the Nelder-Mead method, is a numerical optimization
method for finding minimizing functions. This method is
2.4 Proportional Integral Derivative Controller
very useful when the function used is non-linear and its
The Proportional Integral Derivative (PID) Controller
gradient or derivative cannot be found. This method works
is a combination of the three control systems, namely
by preserving the simplex which is a geometric shape
proportional, integral, and derivative controls. The PID
consisting of N+1 points in an N dimension, where n is the
controller functions to get fast response results, eliminate
variable of the function. In N dimensions, a simplex is a
delays, and eliminate offsets according to Ogata et al. In this
polytope that has N+1 vertices[10].
PID controller, it has a feedback component that aims to get
Each vertex of the simplex is denoted as xk, where k =
appropriate and precise response results. The equation of the
1, 2, …, N+1. Starting from the initial simplex, this
PID controller is as follows[7]:
algorithm determines the worst corner (xw), second worst
�� �
� � = �� � � + �� � � �� + �� 13 (xs), and best corner (xb) of the current simplex. And then
�� the algorithm will move the simplex in the direction away
By applying the Laplace transform, the transfer from the worst vertex. This algorithm iteratively moves the
function of the PID controller is obtained as follows: simplex towards the minimum of the function, by exploring
� � �� the design with 3 operations, namely reflection, expansion
= �� + + �� � 14
� � � and contraction, and also replaces the worst vertex with a
To make the gantry crane stable and not have excessive new point. This operation modifies the simplex and its
swing angles, a controller with optimized parameters is angles until it reaches the maximum iteration level, the
made to obtain more optimal results. In the PID controller, tolerance level of the function value, or the size of the
the Kp, Ki, and Kd parameters can be optimized so that the simplex.
resulting output can have better performance and as desired. All centroids (xc) of all points other than the worst
To achieve the appropriate parameters for the control system, points are calculated, so that new points can be searched as
there are several optimization methods that can be used. The shown in the following equation:
following is an optimization method that can be used[8]. 1
2.4.1 Gradient Descent Optimization �� = �� . (19)
�
The gradient descent method is an algorithm �≠�
optimization that is often used for machine learning and
 and point xy is reflected as xr=2xc-xw. This reflected
point is then taken to become a new point. However, it can
be revised depending on the following conditions[10]:
1. If f(xr) < f(xb), then xr potentially provides a good
direction to move. Therefore, an expansion of the
simplex is carried out to get a new point xnew as
���� = �� + �(�� + �� ), α>1. If f(xnew) < f(xr),then xnew
is taken as the new point.
2. Jika f(xr) ≥ f(xw), maka bergerak ke arah yang
disarankan tidak baik. Oleh karena itu, dilakukan Figure 4. schematic of the Gantry crane simulation system
kontraksi negatif untuk mendapatkan titik baru xnew,
sebagai ���� = �� − �(�� − �� ), 0<β<1. In the hardware implementation, Arduino UNO applied
3. If f(xs) < f(xr) < f(xw), then xr looks like a move in the as controllers with a 1-ms sampling time. the electrical part
right direction, but too far. Therefore, a positive as shown in Figure 5. Two motor applied to balance the
contraction is performed to get the new point xnew,, as trolley position and payload swing angle.
���� = �� + �(�� − �� ), 0<β<1.
The simplex search method is relatively easy to implement,
does not require calculation of derivatives or gradients, and
can handle poor and noisy functions. However, this method
may converge slowly or get stuck at a local optimum,
especially in high-dimensional spaces. Therefore, it is often
combined with other optimization methods or used as a
starting point for more sophisticated algorithms.
2.4.3 Sequential Quadratic Programming Optimization
Sequential Quadratic Programming (SQP) is the most
successful numerical method for solving constrained non-
linear optimization problems. This method uses a gradient-
based approach that iteratively solves quadratic subproblems Figure 5. schematic of the electrical circuit of the system
at each step by using a quadratic approximation of functions
and constraints. SQP is not a single algorithm, but a IV. PERFORMANCE EVALUATION
conceptual method of several algorithms that have evolved. An automated gantry crane controlled by traditional PID
The SQP algorithm combines ideas from Newton's method controllers with optimization is tested against the
and linear programming approaches[9]. construction of an intelligent gantry crane system. Using an
2.4.4 Pattern Search Optimization optimization technique, the PID controllers were designed
Pattern search optimization, optimizing the search and improved. Three types of optimization methods:
space by providing a spatial location for the algorithm that is Gradient Descent, Pattern Search, and Simplex Search,
expected to end when the termination criteria are met, is a respectively, and the result as shown in Table 1 [8].
type of optimization algorithm without derivatives that
works by re-searching the search space based on different Table 1. Result of three optimization methods for position.
search directions and step sizes rather than finding the
minimum function without information about gradients or
high-order derivatives that require complex calculations.
In pattern search optimization, the search direction is
chosen based on a user-defined pattern that determines the
direction and step size to be taken at each iteration, the step
size is often modified based on the search goals found and
can be different for each search method[9].
Table 2. Result of three optimization methods for swing
III. EXPERIMENT RESULT angle.
The proposed system is then experimented by simulation
and experimental for controlling the lab-scale gantry crane
system. Figure 4 shows the simulation, which consisting of
two controller, namely PID for the position control and PD
for swing control, respectively .
The simulation results for position control and anti-
swing control, respectively, are shown in Tables 1 and 2. It
was assessed how well the position control system worked.
The maximum swing amplitude and settling time, on the
other hand, were the foundation for anti-swing control. As
shown in Table 1 PID controller parameters Controller Kp,
Ki, Kd Position Control. As shown in Table 1, the best result
from the system applied Sequential Quadratic Programming,
Kp, Ki, Kd is 1.8495, -0.0165, and 0.8747, respectfully. Figure 8. experiment result from the hardware with PID-GD
Table 2 shows the three optimization methods applied for optimization implementation.
anti-swing control, as shown from result are 0.7899 and
3.0327 for peak time (sec) and settling time (sec) result Table 3. The comparation result of simulation and real
respectfully. experiment for swing angle.
The results show that the Gradient Descent optimization
controller for position control have the smaller overshoot
value, shorter settling time and smaller steady-state
error than the PID controller, as shown in Figure 6.
Therefore, it can be concluded that the proposed PID with As shown experiment result Table 3 above, simulation
Gradient Descent optimization method position control performance and the real experiment have a little difference
system is better than the other optimization method for PID
of approximately 0.2 sec.
controller. Figure 7 show the PID optimization with the
method of GD controller for anti-swing control gave faster V. CONCLUSION
settling time than other two optimization method
The implementation of the PID and PD with
implemented in PD controller.
implementation of optimization method in determining the
coefficient parameters PID and PD for position and PD for
swing angle, respectfully. the performance of simulation and
experimental are applied and compared. the simulation result
of implementation PID, PD with three optimization method
are compared. the result show that Sequential Quadratic
Programming (SQP) with the result of Kp, Ki, Kd is 1.8495,
-0.0165 , and 0.8747, respectfully, for the PID coefficient
parameter. On the other hand, anti-swing control was based
on maximum swing amplitude and settling time, with the
best simulation result of result are 0.7899 and 3.0327 for
peak time (sec) and settling time (sec) result respectfully.
Figure 6. simulation positioning result using three types of The results show that the Gradient Descent optimization
optimization method for PID controller. controller for position control gave smaller overshoot,
shorter settling time and smaller steady-state error
than the PID controller. the SQP optimization method for
PD controller gave the shorter settling time (Ts) among other
optimisation method. The performance of the best parameter
above from the simulation is then applied to the arduino in
order to applied in the real system of the lab-Scale gantry
crane, and the result show still have different in the hardware,
where the setelling time of the swing angle is slower in lab-
scale ganrty crane compare to the result in simulation
REFERENCES
Figure 7. simulation swing angle result using three types of [1] G. A. Manson, “Time‐optimal control of an
optimization method for PID controller. overhead crane model,” Optim. Control Appl.
Methods, vol. 3, no. 2, pp. 115–120, 1982, doi:
The best simulation results from position and swing angle as 10.1002/oca.4660030202.
shown in Figure 8, which the sampling time around 2.8 sec. [2] J. W. Auernig and H. Troger, “Time optimal control
the comparative result between simulation and real of overhead cranes with hoisting of the load,”
Automatica, vol. 23, no. 4, pp. 437–447, 1987, doi:
implementation to the hardware as shown in Table 3, with
10.1016/0005-1098(87)90073-2.
the result is approximation between simulation and
[3] Wahyudi and J. Jalani, “Design and implementation
hardware implementation.
of fuzzy logic controller for intelligent gantry crane
system,” Icom’05, no. May, pp. 345–351, 2005,
[Online]. Available: http://eprints.uthm.edu.my/179/
[4] M. I. Solihin, C. Y. Chuan, and W. Astuti,
“Optimization of fuzzy logic controller parameters
using modern meta-heuristic algorithm for gantry
crane system (GCS),” in Materials Today:
Proceedings, 2019, vol. 29. doi:
10.1016/j.matpr.2020.05.641.
[5] M. Gopal, “Digital Control and State Variable
Methods.” p. 998, 2003.
[6] M. Gopal, Control system principles and design, vol.
 24, no. 1. 1988. doi: 10.1016/0005-1098(88)90017- of the PID controller for stable and unstable systems
9. using nonlinear optimization technique,” Lect. Notes
[7] N. S. Nise, Control Systems Engineering. John Eng. Comput. Sci., vol. 2209, no. January, pp. 292–
Wiley & Sons, Ltd, 2004. 296, 2014.
[8] O. E. Engineering, “Control System | Closed Loop [10] The MathWorks Inc, “User ’ s Guide R 2016 b,”
Open Loop Control System,” Electrical4u.com, 2016.
2015.
[9] F. Alariqi and A. AbdulRahman, “Optimum tuning