Accepted Manuscript

A novel adaptive feed-forward - PID controller of a SCARA parallel robot

using pneumatic artificial muscle actuator based on neural network and
modified differential evolution algorithm

Nguyen Ngoc Son, Cao Van Kien, Ho Pham Huy Anh

PII: S0921-8890(16)30440-7
DOI: http://dx.doi.org/10.1016/j.robot.2017.06.012
Reference: ROBOT 2871

To appear in: Robotics and Autonomous Systems

Please cite this article as: N.N. Son, C.V. Kien, H.P.H. Anh, A novel adaptive feed-forward - PID
controller of a SCARA parallel robot using pneumatic artificial muscle actuator based on neural
network and modified differential evolution algorithm, Robotics and Autonomous Systems (2017),
http://dx.doi.org/10.1016/j.robot.2017.06.012

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to
our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form.
Please note that during the production process errors may be discovered which could affect the
content, and all legal disclaimers that apply to the journal pertain.
 A Novel Adaptive Feed-Forward - PID Controller of a SCARA Parallel Robot
using Pneumatic Artificial Muscle Actuator based on Neural Network and
Modified Differential Evolution Algorithm

Nguyen Ngoc Son*, Cao Van Kien**,a and Ho Pham Huy Anh**,b

*Faculty of Electronic Technology, Industrial University of Ho Chi Minh City, Viet Nam
**Faculty of Electrical-Electronic Engineering, DCSELAB, HCM City University of Technology, VNU-HCM, Viet
Nam
Email: *[email protected] [email protected] b
[email protected]

Abstract

This paper proposes a novel control system combining adaptively feed-forward neural controller
and PID controller to control the joint-angle position of the SCARA parallel robot using the
pneumatic artificial muscle (PAM) actuator. Firstly, the proposed inverse neural NARX (INN)
model dynamically identifies all nonlinear features of the SCARA parallel PAM robot. Parameters
of the inverse neural NARX model are optimized with the modified differential evolution (MDE)
algorithm. Secondly, combining the inverse neural NARX model that provides a feed-forward
control value from the desired joint position and the conventional PID controller applied to improve
the precision and reject the steady state error in the joint position control. Finally, the new adaptive
back-propagation (aBP) algorithm, based on Sugeno fuzzy system, proposed for online updating the
weight values of the inverse neural NARX model as to adapt well to the disturbances and dynamic
variations in its operation. Experimental tests confirmed the performance and merits of the proposed
control scheme in comparison with the traditional control methods.

Keyword Modified Differential Evolution (MDE), Adaptive Back-Propagation (aBP) Algorithm,

Adaptive Neural Control, Pneumatic Artificial Muscle (PAM), Parallel PAM Robot Arm.

1. Introduction

The Pneumatic Artificial Muscle (PAM) actuator is now widely used in many various applications
such as medical robot [1]-[4], humanoid robot [5]-[7] and other industrial applications [8]-[10].
Pneumatic artificial muscles were first developed in the 1950s for using in artificial limbs [11].
PAMs are devices operated by pressurized air supplying a pneumatic bladder. The PAM actuator
has the ability of the providing the advantages such as high power/weight ratio, full of hygiene, easy
in maintenance and especially capable of human compliance which is the most requirement in

1
medical and human welfare domain. Therefore, PAM-based device is regarded as a promisingly
alternative to electric actuators.

Nowadays, there are various studies concerned about the accurate control of the PAM actuator.
Some studies designed the controller based on a mathematical model [12]-[14], a linear parametric
PAM model was identified by the experimental input-output data [15]. Authors in [16-18]
developed the sliding mode controller using a nonlinear model of PAM actuator to obtain robust
control. All of these methods implemented base on the modeling of PAM dynamic systems.
Therefore, it was hard to accurately control of PAM actuator using these methods because of the
strong nonlinear characteristics and sensitive depending on working conditions such as temperature
or pressure resource. To overcome this drawback, the controller based on the neural network
engineering is an ongoing research. Such as the controller based on neural network [19]-[20]; an
advanced PID controller with parameters tuned by neural network [21]-[22]; hybrid neural-fuzzy
controller [23]-[25] and hybrid neural network and genetic algorithm [26] were used for achieving
stable high performance control of the PAM actuator.

As we know, the performance of the neural network is dependent on the training process. In the
training process, the training algorithm is to produce the optimal weighting value of the neural
network scheme by minimizing the fitness function that is defined as an error value between neural
network predicted output and the target output. A popular training algorithm is the back-
propagation (BP). Nonetheless, the BP algorithm only realized local search around the initial values
and produce local optimal values, in the studies [27]-[28]. Under such consideration, the
evolutionary algorithms (EAs) based training processes are regarded as potential alternatives. The
EAs produced global optimum due to their potentiality of seeking the global solution space without
falling into a local optimum. For example, Leung et al. [29] introduced the modification of the
parameters of a neural set based on genetic algorithm (GA). Das et al. [30] introduced the particle
swarm optimization (PSO) for training the neural network model. Socha et al. [31] used the ant
colony optimization to train the feed-forward neural network for pattern classification. Valian et al.
[32] proposed the improved cuckoo search algorithm for training feed-forward neural network for
two benchmark classification problems. Although these proposed methods obtained good results,
two challenges need to be still in progress to improve in training neural network focus on how to
find the optimal solution and how to achieve a fast convergence speed.

Among evolutionary approaches, the differential evolution (DE) algorithm was regarded as a
promisingly globally stochastic optimization method. The first DE technique paper was proposed
with Stone R. and Price K.V. in 1995, [33]. Its advantages are as follows: simplicity in installation,
better performance, fewer parameters involved, and lower complexity. Over recent years, the DE

2
technique is to train a neural model by optimally minimizing real and constrained weighting values.
For example, Ilonen et al. [34] successfully proposed a DE based training neural model. Subudhi et
al. [35] introduced the DE algorithm in the neural training for nonlinear plant identification. Su et
al. [36] used the DE technique to train the wavelet neural model for the micro-vibration mechanism
of micro-gears fault diagnosis. These results showed that the DE technique could be efficiently
applied to learn the neural models used in different applications. Nonetheless, like all other
evolutionary algorithms, DE encounters some drawbacks as follows: DE possesses a good globally
finding the capability to get the globally optimal value, but it seems slow at the exploration around a
potential solution. DE presents a slow convergence rate for high-sized optimization tasks.
Furthermore, DE often gets a risk of stagnation situation in which the optimal searching procedure
stagnates before obtaining a global optimum value, [37].

As to overcome such drawbacks, some suggestions proposed to improve the DE performance on the
training neural network. For example, Slowik et al. [38] successfully implemented adaptive DE
algorithm with multiple trial vectors for training artificial neural networks. Curteanu et al. [39]
proposed a new neuro-evolutionary approach that is applied a variant of DE consist in using an
opposition-based learning initialization. Authors in [40], [41] proposed the combination of a global
search of DE algorithm with another local search method such like BP, Levenberg Marquardt (LM)
for building an approximation of the neural network models.

In summary, as to improve the quality control of a neural network model-based control, two
problems need to be considered that how to obtain the accurate approximation and how to online
update the weights of the neural controller. This paper proposes a novel control system with
adaptive feed-forward controller and PID controller (called the novel adaptive feed-forward - PID
controller) for position tracking control of the SCARA parallel PAM robot. The ideas of this new
control approach are as follows,

1. To initiatively introduce the novel modified differential evolution (MDE) algorithm for
training the multi-layers perceptron neural network model. Firstly, the mutation phase of the
differential evolution is modified by mean of combining two mutation strategies rand/1 and
best/1 together to create trial vectors instead of only using one mutation operator or rand/1
or best/1 as the standard DE. The modification that aims to equalize between global
exploration and local exploitation capacities finds global potential optimum solutions. Then
the local searching ability of the back-propagation algorithm is applied to converge to the
optimal solution.

3
 2. To propose the inverse neural NARX (Nonlinear Auto-Regressive with eXogenous input)
model optimized by MDE training algorithm - called MDE-INN model - to identify the
nonlinear dynamic characteristics of the SCARA parallel PAM robot.

3. To combine the feed-forward controller based on MDE-INN model and the traditional PID
controller. Where the MDE-INN model provides a feed-forward control value based on the
joint reference position, and the PID feedback controller is to improve the precision and
reject the steady state error for the joint position control.

4. To propose the new adaptive back-propagation (aBP) learning algorithm using fuzzy logic
to update online the weighting values of the MDE-INN model as to adapt well to the
disturbances and dynamic variations of the nonlinear SCARA parallel PAM robot in its
operation.

The rest of the paper is organized as follows: Section 2 presents the configuration of the SCARA
parallel PAM robot. Section 3 presents the proposed neural NARX model trained by the modified
differential evolution (MDE) algorithm for nonlinear system identification. Section 4 identifies the
inverse model of the SCARA parallel PAM robot using the neural NARX model and the MDE
training algorithm. Section 5 introduces the proposed adaptive feed-forward - PID controller for the
SCARA parallel PAM robot. Finally, section 6 contains some conclusions.

2. Configuration of the SCARA parallel PAM robot

A block diagram of the experimental SCARA parallel PAM robot architecture is shown in Fig.1,
and the photo of the SCARA parallel PAM robot system figured in Fig.2.
p1
p2

1
u2 u1
1
d l1
2 l2

2

Fig.1. Diagram of the experimental SCARA parallel PAM robot

The SCARA parallel PAM robot is a four-bar mechanism and constructed of two pneumatic
artificial muscles (PAM) actuators, two coil springs, two rotary encoders, two EV2500-008
4
proportional valves, NI PCI 6221 card, air compressor and computer. The mechanical parameters of
the SCARA parallel PAM robot are described in Table 1 and the device parameters of SCARA
parallel PAM robot are described in Table 2. The joint-angle rotation of the parallel PAM robot is
created with a single PAM in connection with a coil spring and a proportional valve. Two rotary
encoders are directly mounted to the two joints to measure the joint-angle position of parallel PAM
robot. The PCI 6221 card is used for data acquisition and controlling the robot system. All control
algorithms are implemented in the Real-Time Windows Target Toolbox of Matlab and the sampling
time is selected as Ts = 0.01 (Sec).

Fig.2. Photo of the experimental SCARA parallel PAM robot

Table 1. Mechanical parameters of SCARA parallel PAM robot

Parameter Notation Value Unit
Length link 1 l1 15 cm
Length link 2 l2 20 cm
Distance between two rotary joints d 20 cm
Diameter of pulley  3 cm
Stiffness of spring k 500 N/m

Table 2. Experimental device parameters of SCARA parallel PAM robot

No Devices Parameters
- Nominal length of PAM: 22 cm
PAM Festo MAS- - Internal diameter: 10 mm
1
10 N220 - Operating pressure: 0-8 bar
- Max. permissible contraction: 25% of nominal length
- Incremental encoder E40S6 3600 Autonics Co.
2 Rotary encoder - Resolution: 3600 (pulses/rotation)
- Power supply: 24 V ± 10%
- Two 16-bit analog outputs (833 KS/s); 24 digital I/O;
3 PCI 6221 (NI)
32-bit counters

5
 - 16 Analog Inputs, 16-Bit, 250 KS/s
- Power voltage: DC 24 V ± 10%
Proportional valve - Control pressure range: 0 to 0.49 Mpa (4.9 bar)
4
EV2500-008 - Input signal: 0-10 V DC
- Resolution: 0.5% F.S
3. The proposed neural NARX model optimized by the modified differential evolution
(MDE) algorithm for nonlinear system identification

3.1. The proposed neural NARX model structure

The proposed neural NARX model is implemented by combining the Multilayer Perceptron Neural
Networks (MLPNN) model with the nonlinear Auto-Regressive with eXogenous (NARX) structure.
The weighting values of the neural NARX model are optimally selected with the modified
differential evolution (MDE) training algorithm. Due to this combined structure, the novel neural
NARX model obtains a promisingly approximating feature of the MLPNN model and a powerful
predictive ability of the NARX model. The neural NARX model structure used for nonlinear system
identification presented in Fig.3.

u (k ) y (k )
Nonlinear System

u (k  1) n v jn
q1
v1n f j (.)
u (k  nb) 2 v j2 w1 j
q  nb
NARX v12 w11 ŷ  k 
v j1 f1 (.) F (.) - +
y (k  1)
model 1 error
q1 v11
w10
v j0 1
y (k  na ) 1 v10 Training
q  na Algorithm
MLPNN structure

Fig.3. Proposed neural NARX model for nonlinear system identification.

Where, the neural NARX model is a fully connected 3-layer neural network; u(k) represents the
input signal of the nonlinear system; y(k) represents the output signal of the nonlinear system; na
and nb are the order of output y(k) and input u(k), respectively; q is the backward time shift operator
defined as q-1u (k ) = u (k -1) ; vjn represents the weighting value of input layer, vj0 represents the

weighting value of bias input layer, w1j and w10 represent the weighting values of hidden layer and
bias hidden layer, respectively; fj(.) and F(.) represent the sigmoid function at hidden layer and the

6
linear activation function at output layer, respectively. The dimension number D is the sum of the
number of weights and thresholds of the neural NARX model.
Then the prediction output is as follows
æ j æ n ö ö
yˆ (k , q ) = F ççå w1i f i ççå viljl (k ) + vl 0 ÷÷ + w10 ÷÷÷ (1)
çè i=1 è l =1 ø ø

Where j(k ) represents the regression vector and q denotes the weighted vector. These parameters
are denoted as follows
  k   1  k  ,...,  n  k     y  k  1 ,..., y  k  na  , u  k  1 ,..., u  k  nb  
T T
(2)
T
q = éë v10 ,..., v1n , v j 0 ,..., v jn , w10 ,..., w1 j ùû = [ w1 ,...., wD ]
T
(3)

The MDE training algorithm is to produce the weighting value of the neural NARX model
optimally. The output of the neural NARX model plays a prediction function of the synaptic weight
vector  and regression vector   k  . During the learning procedure, both the input signal u(k) and

the output signal y(k) of a nonlinear system are determined, and the weighting vector  is adapted
to get the response from the input u(k) to the output y(k). The adaptation procedure realized by
minimally optimizing the neural network error function EN determined as follows
N
1 T
EN (q, Z N ) = å éêë y (k )- yˆ (k q)ùúû é y (k ) - yˆ (k q)ù
êë úû (4)
2N k =1

In which, the learning data Z N are denoted by Z N = {éë y ( k ) , u (k )ùû k = 1,..., N } . The optimized

purpose is to optimize the fitness function EN by optimally selecting the values of the weighting

vector q = [ w1 ,...., wD ] . In this paper, the proposed evolutionary training algorithm is the new MDE-
T

BP optimization algorithm.
3.2. Traditional DE algorithm
R. Storn and K.V. Price firstly introduced the differential evolution (DE) technique in 1995, [33].
Nowadays, it proves popular and plays the important role of optimization algorithms. The DE
algorithm includes five main phases as follows

Initialization. Assume that it needs to optimize a function containing D real-parameters. Firstly, the
population size NP is chosen. Then the DE technique starts by par hazard producing NP N-
dimensional vectors. Each vector is considered as an individual, which is defined as,

X i ,G = [ x1,i ,G , x2,i ,G ,..., xD,i ,G ] (5)

with G represents some generations, G  0,1,..., Gmax , i  1, 2,..., NP

7
Each parametric vector is restricted to a bounded seeking space, and they can not overrun either end

of these bounds. The lower and upper limits are represented as X min = { x1,min , x2 ,min ,..., x D ,min } and

X max = { x1,max , x2 ,max ,..., x D ,max } . Thus, the j th component of the i th vector is initially started as

x j ,i = x j ,min + rand j ,i [0,1] ( x j ,max - x j ,min ) , with 0 £ rand j ,i [0,1] £ 1 (6)

Mutation. In the DE method, a parent vector is also considered as a destination vector, then a
mutant vector emerged through the differential mutation procedure is named as donor vector, and
an offspring vector structured by recombining the donor with the destination vector is considered as
a trial vector. In order to produce the donor vector for each ith destination vector, some popular
mutation operations are usually used as Table 3. Where the indices r1i , r2i , r3i , r4i and r5i are mutually

exclusive integers randomly chosen from the range 1, NP  such that r1i  r2i  r3i  r4i  r5i  i ; the

scale factor F is randomly chosen within  0,1 .

Table 3. Some popular mutation operations of the DE algorithm

   
1 DE/best/1 vi ,G = xbest ,G + F xr ,G - xr ,G ( i
2 3
i )
   
2 DE/rand/1 vi ,G = xr ,G
1
i + F (x r2i ,G
- xr ,G
3
i )
     
3 DE/best/2 (
vi ,G = xbest ,G + F xr ,G + xr ,G - xr ,G - xr ,G
1
i i
2 3
i i
4
)
     
4 DE/rand/2 vi ,G = xr ,G
5
i + F (x r1i ,G
+ xr ,G - xr ,G - xr ,G
i
2 3
i i
4
)
Crossover. After producing the donor vector using a mutation procedure, the crossover procedure
is to improve the variety of the population tank. The donor vector interchanges its pieces with the
 
destination vector X i,G as to structure the trial vector Ui,G = [u1,i,G , u2,i,G ,..., uD,i,G ] . The DE technique
usually applies the binomial crossover approach. The binomial crossover can be described as
ì
ï if ( rand j ,i [0,1] £ C )
ïv j , i , G
u j ,i ,G = ï
í (7)
ï
î x j ,i ,G
ï
ï otherwise

With i  1, 2,..., NP , j  1, 2,..., D . Remark that rand j ,i [0,1] represents a uniquely distributed stochastic
number, C represents the crossover rate.
 
Selection. The destination vector X i,G is in comparison with the trial vector Ui,G and the one

possesses a minor function value will be survived to the next iteration. Where f ( X ) denotes the

minimally optimized fitness function. The selection procedure is expressed as

  

ì
ï
ï
ï
U i ,G
X i ,G +1 = í 
( ) (
if f U i ,G £ f X i ,G ) (8)
ï
ï
î X i ,G
ï otherwise

8
Termination. The abovementioned steps are realized in a ring of generations. The DE technique is
ended when one of the three followed requirements are responded: when the chosen amount of
generations Gmax is attained; when the best fitness of the population does not vary over consecutive
generations; and when a pre-defined fitness function value is reached.

3.3. The proposed modified MDE training algorithm

In this study, the modified differential evolution (MDE) is proposed for successfully learning the
neural model. The idea of the proposed MDE algorithm is illustrated as follows

3.3.1. The adaptive mutation phase

The success of the meta-heuristic methods is significantly determined by balancing two

contradictory aspects: global exploration and local exploitation, [42]. Commonly, a meta-heuristic
algorithm with good at global exploration abilities is often worse at local exploitation abilities. In
contrast, a meta-heuristic algorithm that seems better at local exploitation abilities is often worse at
global exploring capabilities. In the DE, the mutation operation plays a pivotal role in its seeking
ability and its convergence rate. There exist, at least four mutation procedures have been introduced
for the DE with different purposes. For example, with the mutant procedure rand/1, the DE is
strong at global finding but weak in local seeking. Then, it converges to the global optimum result.
Otherwise, with the mutant procedure best/1, the DE is excellent at local seeking, but poor at global
finding. Thus, it is easy to be falling into the local optimum results, [43]. For these reasons, the
performance by applying the standard DE with only one mutation operation for training NN is still
low. Based on the above analyses, in this work, the mutation phase of the DE is modified by mean
of combining two mutation strategies rand/1 and best/1 together to create trial vectors instead of
only using one mutation operator either rand/1 or best/1 as the standard DE. The modification aims
to equalize between global exploration and local exploitation capacities. The novel mutation
scheme is described as follows

if rand[0,1]>0.3
v i  x r1  F (x r2  x r3 ) % use “rand/1”

else
vi  xbest  F (x r1  xr2 ) % use “best/1”

End

From the above mechanism, it can be recognized that for each target vector, only one of the two
mutation operators is applied for creating the current trial vector, depending on a uniformly
distributed random value within the range [0,1]. For each target vector, if the random value is

9
bigger than 0.3, the rand/1 is performed. Otherwise, the best/1 is. Therefore, with this scheme, at
any particular generation, the proposed strategy has a chance to enhance the exploration and
exploitation abilities. It is important to note that because of training neural is a complex nonlinear
problem; it is more necessary exploration than exploitation capabilities. In this study, we realize that
the rate of 0.7/0.3 for rand/1 and best/1 is a well-balanced value for both of exploration and
exploitation capabilities.

3.3.2. Hybrid with the back-propagation algorithm

During many last decades, the back-propagation (BP) algorithm is known as a popular method for
training ANN [44]. Nonetheless, because the BP is a gradient-based optimization method, it is only
good when its searching process starts with a good initial point. Many researchers have recognized
that the combination of global search methods such as GA, PSO, DE with local search methods like
BP, Levenberg Marquardt (LM), etc. are an effective approach to building approximation models of
ANN. For the DE, there are some studies hybridizing it with gradient-based methods investigated in
[40]-[41]. Following this trend, this paper presents a new way of integration of the BP into the DE
algorithm. In which, the BP is applied after the selection phase through the searching process of the
DE, and it only uses the best individual as an initial point for searching. This technique does not
lose much time, but it may help enhance local solutions in each generation and has a high
probability to obtain a good solution after finishing search process.

3.3.3. Control parameters setup

The control parameters in DE are C, F and NP, which significantly influence the effectiveness and
performance of the algorithm. R. Storn and Price [45] suggested that F should be 0.5 as a good
choice, C should be set to 0.1 or 0.9 and NP should be between 5D and 10D. Gamperle et al. [46]
proposed that NP should be between 3D and 8D, C should be to [0.3, 0.9] and F = 0.6. Ronkkonen
et al. [47] proposed that a better choice for F = 0.9 and C should be set between 0.9 and 1 for non-
separable multimodal fitness functions and C between 0 and 0.2 for separable multimodal objective
functions.

Some researchers recently proposed new approaches to adjust control parameters in adaptive or
self-adaptive. For example, Liu et al. [48] introduced a fuzzy logic to modify crossover and mutant
rate. Brest et al. [49] presented jDE algorithm by self-adaptive control variables of encoding the
variables into individual and updating them by DE. Qin et al. [43] introduced a self-adaptive DE
(SaDE) technique, in which both of trial production strategy and control parameters are self-adapted
by training based on their previous experiences in producing good solutions.

10
Based on the above analysis, we see that it is difficult to adequately select all parameters of DE
algorithm. In this proposed MDE algorithm, the mutant F and crossover C do not fix as in the
standard DE, they are randomly generated within the range [0.4, 1] and [0.7, 1], respectively. These
parameters are referred from [50] to improve the searchability of the MDE-BP in many different
searching directions. The NP parameter is not changed during the running algorithm.

3.3.4. Flowchart of MDE training algorithm

This section uses the MDE algorithm to optimize the weighting value of the proposed neural NARX
model by minimizing the fitness function in Eq.(4). The working steps involved in employing the
MDE training algorithm with its content is as Algorithm 1.

Algorithm 1. Proposed MDE-BP training algorithm

1. Begin

2. Generate randomly initialized population i ,G   w1,i ,G , w2,i ,G , ..., wD,i ,G 
3. Evaluate the fitness of individual in the population
4. For G=1 to GEN do
5. For i =1 to NP do
6. jrand= randint(1,D)
7. F = rand[0:4; 1:0], C = rand[0:7; 1:0]
8. For j =1 to D do
9. If rand[0,1] < C or j == jrand then
10. If rand[0,1] > threshold then
11. Select randomly r1  r2  r3  i
12. u j ,i , G 1  w j , r 1,G  F ( w j , r 2,G  w j , r 3, G )
13. Else
14. Select randomly r1  r2  best  i
15. u j ,i , G 1  w j , best , G  F ( w j , r 1,G  w j , r 2, G )
16. End if
17. Else
18. u j , i , G 1  w j , i , G
19. End if
20. End for
 
21.  
If f U i ,G 1  f  i ,G then  
 
22.  i , G 1  U i , G 1
23. Else
 
24.  i ,G 1   i ,G
25. End if
26. End for

27. Use  i ,G 1 as an initial point for BP algorithm
 
28. Replace  i ,G 1 by bp obtaining from BP algorithm
29. End for
30. End

4. Identification the inverse dynamic of the SCARA parallel PAM robot based on the neural
NARX model and MDE algorithm

11
This section studies the performance of our proposed MDE algorithm trained the neural NARX
model for identifying the inverse model of the SCARA parallel PAM robot. All of the simulations
were performed by MATLAB version 2013b on an Intel Core i3 computer with a clock rate of
2.53GHz and 2.00GB of RAM. Each algorithm runs 10 times. Table 4 gives the parameters used in
identification. The identification process composes of four key stages as follows,

Table 4. Parameters used in identification process

Method Parameters Value

Number of hidden neurons S
Population size NP
General
Number of generations GEN
Upper and lower bound on weight [-5,5]
Learning rate, η 0.00001
BP
Number of iterations 700
Mutation factor, F 0.5
DE, [33]
Crossover constant, C 0.9
Mutation factor, F [0.4,1]
Crossover constant, C [0.7,1]
MDE
Learning rate, η 0.00001
Number of iterations 5

Collecting training data. This stage uses the experimental parallel PAM robot to produce a
collection of experimental data on the voltage applied to the joint angle of the SCARA parallel
PAM robot arm, as illustrated in Fig.4. Where u1 (k ) and u2 (k ) represent the applied voltage inputs.

theta1 (k ) and theta2 (k ) represent the joint angle output. This experimental input-output training
data are applied for estimation (a) and validation (b) of the neural NARX model.

12
 (a) (b)

u1v (V)
u1e (V) 5 5

0 0
0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000

theta1v (rad)
theta1e (rad)

0 0

-0.5 -0.5
0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000

u2v (V)
u2e (V)

5 5

0 0
0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000

theta2v (rad)
theta2e (rad)

0 0

-0.5 -0.5
0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000
time (samples) time (samples)

Fig.4. Collection of data for estimating and validating targets

Select model structure. Assuming that the data set acquired, the second step is to select a model
structure. This section investigates an inverse neural NARX (INN) model implemented by creating
the 3-layer MLPNN with 15 nodes of the hidden layer, which is illustrated as follows

u2 ( k ) theta2 (k  1)
u1 (k ) theta1 (k  1)

û1  k 
z 1
û2  k 

z 1

Fig.5. Proposed MDE-INN model structure for identifying the inverse dynamic of PAM system

In which the MDE training algorithm is to optimize the weights of the inverse neural NARX
model (MDE-INN model). The number of hidden layer neurons S = 15, the population size
NP = 120, the number of generation GEN = 700. All parameters are selected by the trial and
error method.
Estimate and validate the MDE-INN model. Finally, the estimation and validation process are to
identify the MDE-INN model. To assess the performance of the proposed method, we make a
comparison with the traditional DE and BP algorithm. Figure 6 shows the convergence rates of BP-
INN model, DE-INN model and MDE-INN model, respectively, based on the average values of

13
MSE on estimating process. Table 5 shows the performance comparison of BP-INN model, DE-
INN model and MDE-INN model over the training and validation process.
1
10
BP

MSE average on training process

DE
MDE

0
10

0 100 200 300 400 500 600 700

Iteration

Fig.6. Fitness convergence comparison of the estimation stage

Table 5. Performance comparisons of the BP-INN, DE-INN and MDE-INN models
Method MSE
Training Validation
Best Worst Average Average
MDE 0.3013 0.4611 0.3681 0.4674
DE 0.6643 0.6707 0.6685 0.6121
BP 0.6895 0.7977 0.7414 0.7250
In conclusion, the estimating and validating results show that the inverse dynamic model of
SCARA parallel PAM robot arm achieves very good identification performance. Furthermore, the
neural NARX model based on modified MDE technique (MDE-INN model) performs faster
convergence and better accuracy than the traditional BP approach and the lonely DE algorithm.
Figure 7 shows the performance of proposed MDE-INN model based on the average values of MSE
on validating the identification process. Table 6 tabulates the resulted identified weights of the
inverse neural NARX model optimized by the modified MDE algorithm.

u1 (solid) and u1_hat (dashed) u2 (solid) and u2_hat (dashed)

6
6
4
4
2 2

0 0
0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000
2 2
prediction error

0 0

-2 -2
0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000
time (samples) time (samples)

14
 Fig.7. Performance of MDE-neural NARX model for identification of the inverse PAM 2-DOF model

Table 6. Identified weights of input layer of the MDE-based neural NARX model
vji: weights of input layer; vj0: weights of bias input layer;
wij:weights of the hidden layer; wj0: weights of bias hidden layer
vji vj0 wij wj0
j i 1 2 3 4 0 1 2 0
1 -1.17309 -0.8122 -3.3633 -3.32308 -1.46143 1.136563 0.333504 1.721648
2 -0.69413 0.57009 0.415405 -0.4689 -0.64255 -1.14148 -3.47826 2.284609
3 -5.22287 -5.47881 0.717332 0.495934 0.251332 -3.34107 -1.15889
4 2.817097 1.372835 -2.21479 -1.09485 0.104201 -0.19803 1.984171
5 -0.07536 -0.22398 -0.46853 -0.33141 1.841336 1.631216 1.438443
6 -1.14283 1.447392 1.209587 -1.69594 -0.08121 2.661457 -2.64498
7 0.847571 0.774313 1.179246 1.011308 2.204621 -2.52673 -1.93821
8 0.724726 0.122334 -0.54428 0.41416 0.780499 -0.70586 0.509736
9 -0.20272 0.260797 0.069284 -0.03905 -0.21991 -0.46186 -0.96751
10 0.833561 0.308518 -2.77539 -2.03287 -0.75892 -0.2135 1.777313
11 -0.57361 -0.2377 -0.66476 -1.17879 0.725741 1.156856 -1.15823
12 -0.25865 -0.04098 -0.1892 -0.83768 1.71663 1.357704 0.643406
13 3.749973 2.586134 -1.00402 0.395873 0.176609 -2.77468 1.391287
14 -0.44888 0.216482 0.729229 -0.9257 0.083027 1.534043 -0.39856
15 0.492768 2.68352 1.06386 -0.70518 0.280303 0.642529 -2.60906

5. Adaptive feed-forward-PID controller of SCARA parallel PAM robot

5.1. The proposed adaptive feed-forward-PID controller structure

This section presents the new adaptive feed-forward-PID controller approach for the two-joint
position control of the SCARA parallel PAM robot. The basic idea of this proposed controller
consists of three steps. The first step aims to train offline the inverse neural NARX model using the
MDE algorithm (MDE-INN model) to learn the dynamic characteristics of the SCARA parallel
PAM robot. The second step is to combine the feed-forward controller based MDE-INN model and
the traditional PID feedback controller. Where the MDE-INN model provides a feed-forward
control value from the desired joint angle, and the PID controller is to improve the precision and
reject the steady state error in the joint trajectory control. The third step is to use an adaptive back-
propagation (aBP) learning algorithm for online update weighting values of the MDE-INN model as
to adapt well to the disturbances and dynamic variations of the nonlinear SCARA parallel PAM
robot in its operation. The block diagram of the proposed controller is illustrated in Fig.8.

15
 eu 2
eu1
uinn 2
uinn1

q1ref e y1 u pid 1 uPID-INN 1 q1

q2ref ey 2 u pid 2 q2
uPID-INN 2

Fig.8. Block scheme of the novel adaptive feed-forward - PID controller approach

The objective of the proposed controller is to minimize the error between the reference signal and
the SACRA parallel PAM robot output. Based on Fig.8, the voltage control law is expressed as,

uPID-INN 1 = u pid 1 + uinn1

(9)
uPID-INN 2 = u pid 2 + uinn 2
Where uPID-INN 1 and uPID-INN 2 represent the required voltage signal applied to the proportional Valve 1

and Valve 2; u pid 1 and u pid 2 are the voltage control value produced by a PID controller; uinn1 and

uinn 2 represent the feed-forward controller signal. The proposed controller uses the inverse dynamic
model of the SCARA parallel PAM robot as a feed-forward controller to cope with the tracking of
the reference signal. The feed-forward controller structure based on the MDE-INN model contains
the similar model scheme presented in section 4 with its initial weighting values are described in
Table 6.

5.1.1. Selecting the gains of the proposed controller

In this proposed controller, the PID controller roles to guarantee global asymptotic stability to the
whole system. The PID control signal is as follows
 t
de y1  t 
u pid 1  t   k p e y1  t   ki  e y1   d  kd
 0
dt
 a (10a)

t
de y 2  t 
u pid 2  t   k p e y 2  t   ki  e y 2   d  kd dt
 0

In which ey1 and ey 2 are defined as ey1 = q1ref - q1 and ey 2 = q2 ref - q2 , respectively, represent the

error of the system; kp, ki, and kd denote the gains of the PID controller. The PID controller carefully
implemented to guarantee the system stability. For this reason, in [51]-[52], these studies based on
the Lyapunov analysis in which the relation kd2 > k p is a necessary condition for the asymptotic

convergence applied in a class of second order single input - single output (SISO) systems. With the

16
other nonlinear systems, the selection of the PID parameters to guarantee the stable system and the
trial and error method can be used [53-54]. Furthermore, meta-heuristic optimization algorithms
seem also available for self-tuning PID controller parameters [55]. Thus, depending on nonlinear
systems, we must be carefully selected the PID parameters in the proposed controller to achieve
stable learning as well as good tracking performance.

5.1.2. The proposed adaptive back-propagation learning algorithm

The aim of the training algorithm is to adjust the network weights through the minimization of
following cost function.
1 1
E = (eu1 + eu 2 ) = ((uPID-INN 1 - uinn1 ) + (uPID-INN 2 - uinn 2 ))
2 2
(10b)
2 2
By using the adaptive back propagation (aBP) learning algorithm, the weighting of the feedforward
controller based on the MDE-INN model adaptively adjusted. The aBP learning algorithm briefly
written as follow,
E
w  w  new   w  old    (11)
w
with λ > 0 represents the learning rate. Using the chain rule, we have
E E uinn
 (12)
w uinn w
E E
 (u1  uinn1 );  (u2  uinn 2 ) (13)
uinn1 uinn 2

From equations (11), (12) and (13), it can return to the detailed calculation to online update the
weights of the MDE-INN model by using the Pseudo-code BP in [44]. Obviously, the selection of
the suitable learning rate  plays an important role for the convergence by adaptively adjusted
online the weight-update. If learning rate  value is too small, it needs much time to obtain an
acceptable solution. On the contrary, a large learning rate value will possibly lead to oscillation,
preventing the error Ec to converge to zero.

For this reason, this paper proposes a novel self-adaptive strategy for selecting the learning rate
 based on a Sugeno fuzzy model. The input of the fuzzy model is the error Ec indicated in Eq.(10),
and the derivative of error Ec , namely dEc . The corresponding output is the learning rate  . The
Sugeno fuzzy model is constructed using the fuzzy rules shown in Table 7, where
z1  0; z 2  5e 5 ; z 3  5e 4 and z 4  5e 3 . The membership functions for fuzzy input variables are
shown in Fig.9. The corresponding fuzzy surface of the adaptive learning rate  is shown in Fig.10.

Table 7. Fuzzy rules table for the self-adaptive scaling learning rate

17
 dEc LN SN Zero SP LP
Ec
LN z2 z3 z4 z4 z4
SN z4 z3 z3 z3 z4
Zero z3 z2 z1 z2 z3
SP z4 z3 z3 z3 z4
LP z4 z4 z4 z3 z2

LN SN Zero SP LP
1
0.8
0.6
0.4
0.2
Degree of membership

0
-0.05 -0.04 -0.03 -0.02 -0.01 0 0.01 0.02 0.03 0.04 0.05
Ec

LN SN Zero SP LP
1
0.8
0.6
0.4
0.2
0
-0.1 -0.08 -0.06 -0.04 -0.02 0 0.02 0.04 0.06 0.08 0.1
dEc

Fig. 9. Membership functions of the fuzzy input variables

-3
x 10

4
lamda

0.1
0.05
0
0

-0.1 -0.05
dEc Ec
Fig. 10. The corresponding fuzzy surface of the scaling learning rate - 

5.2. Experimental result

The block diagram of the new adaptive feed-forward - PID control approach applied to position
control of the SCARA parallel PAM robot is illustrated as Fig.11.

18
 Fig.11. Block diagram of proposed adaptive feed-forward - PID controller scheme

Fig.11 shows the experiment SIMULINK block of the new adaptive feed-forward - PID control scheme
applied for real-time control in the Real-time Window Target of MATLAB environment. The PID
factors are turned with the trial and error method and chosen to be K p   10, 10 ; Ki   25, 25

and Kd   0.5, 0.5 .

Case study 1. In this case, the closed loop performance is verified in changing the reference signal
to survey the position control performance of the parallel PAM robot using the proposed controller.
The SCARA parallel PAM robot is required to work stably and accurately in a range of reference
frequency from 0.1 [Hz] up to 0.5 [Hz]. Fig.12 and Fig.13 compare the control performance of the
proposed controller and the PID controller on Joint 1 and Joint 2 of the parallel PAM robot in two
cases of Sine 0.1 Hz and Sine 0.5 Hz without the payload, respectively. Continually, Fig.14 and
Fig.15 compare the tracking performance of the proposed controller and the PID controller on both
joints in two cases of triangle 0.2 Hz and triangle 0.5 Hz without the payload, respectively. Table 8
describes the performance comparison based on mean of the squared error (MSE) index in case
study 1. All these results show that the PID controller causes the big error values and cannot make it
smaller. On the contrary, the proposed controller always adaptively minimizes the error value to
converge to zero successfully.

19
 Joint 1 - Sine 0.1 Hz without Load Joint 2 - Sine 0.1 Hz without Load
0

theta (rad)
-0.2

-0.4
-0.5
0 20 40 60 80 0 20 40 60 80
0.05
0.05
error (rad)

0 0
-0.05
-0.05
0 20 40 60 80 0 20 40 60 80
6
voltage (V)

4 5

2
0
0 20 40 60 80 0 20 40 60 80
time (sec) ref PID PID-INN time (sec)

Fig.12. The performance comparison of the proposed controller with PID controller (with sine 0.1Hz)

Joint 1 - Sine 0.5 Hz without Load Joint 2 - Sine 0.5 Hz without Load
0 0
theta (rad)

-0.5 -0.5
0 10 20 30 0 10 20 30
0.1 0.2
error (rad)

0 0

-0.1 -0.2
0 10 20 30 0 10 20 30
voltage (V)

5
5

0 0
0 10 20 30 0 10 20 30
time (sec) ref PID PID-INN time (sec)

Fig.13. Performance comparison of the proposed controller (with sine 0.5Hz)

20
 Joint 1 - Triangle 0.2 Hz without Load Joint 2 - Triangle 0.2 Hz without Load
0 0
theta (rad)
-0.2

-0.4 -0.5
0 20 40 60 80 0 20 40 60 80

0.02 0.02
error (rad)

0 0
-0.02 -0.02

0 20 40 60 80 0 20 40 60 80
voltage (V)

5
5

0 0
0 20 40 60 80 0 20 40 60 80
time (sec) ref PID PID-INN time (sec)

Fig.14. Performance comparison of the proposed controller (with triangle 0.2Hz)

Joint 1 - Triangle 0.5 Hz without Load Joint 2 - Triangle 0.5 Hz without Load
0 0
theta (rad)

-0.2

-0.4 -0.5
0 20 40 60 80 0 20 40 60 80
0.05 0.05
error (rad)

0 0

-0.05 -0.05
0 20 40 60 80 0 20 40 60 80
voltage (V)

5
5

0 0
0 20 40 60 80 0 20 40 60 80
time (sec) ref PID PID-INN time (sec)

Fig.15. Performance comparison results of the proposed controller (with triangle 0.5Hz)

Table 8. Performance comparision in case study 1

Error criteria-MSE
Case study 1 – without Load
Theta1 Theta2
PID 7.0581e-04 0.0014
Sine 0.1 Hz
PID-INN 1.0271e-04 6.9645e-05
PID 0.0042 0.0087
Sine 0.5 Hz
PID-INN 3.0657e-04 6.1199e-04
PID 1.6886e-04 2.8307e-04
Triangle 0.2 Hz
PID-INN 5.0311e-05 6.7966e-05
PID 3.1073e-04 4.0936e-04
Triangle 0.5 Hz
PID-INN 2.1595e-04 2.0988e-04

21
Fig.16 and Fig.17 show the overall control signal (uPID-INN), the output of feed-forward controller
based on INN model (uinn) and the PID controller (upid). After a finite time, the feed-forward
controller based the MDE-INN model learns the inverse dynamic of the PAM robot and take the
responsibility of the control system. Simultaneously, the output of the PID controller tends to go to
zero. Fig.16 and Fig.17 also show that the weights varied automatically during control operation in
two cases of sine trajectory and triangle trajectory without payload, respectively. The adaptation
process based on the aBP algorithm makes it possible to control the parallel PAM robot without any
information about the Jacobian of the PAM robot system.
(a) Sine 0.1 Hz without Load (b) Sine 0.5 Hz without Load
8
pid inn PID-INN 6
6

Control signal - voltage (V)

Joint 1

4 4
2 2
0 0
0 20 40 60 80 0 10 20 30
8 8
6 6
Joint 2

4 4
2 2
0 0
0 -3 20 40 60 80 0 -4 10 20 30
x 10 x 10
2 5

The adaption process

dv/dt

0 0
-2 -5
20 40 60 80 5 10 15 20 25 30
0.2 1
dw/dt

0 0
-0.2 -1
20 40 60 80 5 10 15 20 25 30
time (sec) time (sec)

Fig.16. Proposed controller parameters during control process with Sine trajectory reference

(a) Triangle 0.2 Hz without Load (b) Triangle 0.5 Hz without Load
8 8
pid inn PID-INN
Control signal - voltage (V)

6 6
Joint 1

4 4
2 2
0 0
0 20 40 60 80 0 20 40 60 80
8 8
6 6
Joint 2

4 4
2 2
0 0
0 -3 20 40 60 80 0 20 40 60 80
-3
x 10 x 10
2 2
The adaptation process
dv/dt

0 0

-2 -2
20 40 60 80 20 40 60 80
0.5 2
dw/dt

0 0

-0.5 -2
20 40 60 80 20 40 60 80
time (sec) time (sec)

Fig.17. Proposed controller parameters during control process with Triangle trajectory reference

22
Case study 2. In this case, the closed loop performance is verified in changing the weight of an
object mounted on the end-effecter of the parallel PAM robot. Fig.18 and Fig.19 show the
experimental comparison results between the traditional PID controller and the proposed controller
on Joint 1 and Joint 2 of the parallel PAM robot in two cases of Sine 0.1 Hz and Sine 0.5 Hz with
the payload, respectively. Continually, Fig.20 and Fig.21 compare the tracking performance of the
proposed controller on both Joints of the parallel PAM robot in two cases of triangle 0.2 Hz and 0.5
Hz with the payload, respectively. Table 9 describes the performance comparison based on mean of
the squared error (MSE) index in case study 2. All these results show that the PID controller makes
the large error and then cannot make it smaller. Moreover, the PID controller also causes the error
signal larger than in the case of without payload. On the contrary, the proposed controller adapt well
to payload variation of nonlinear parallel PAM robot in its operation, the error between reference
signal trajectory and joint trajectory tracking output is adaptively and continually minimized.
Joint 1 - Sine 0.1 Hz with Load [200g] Joint 2 - Sine 0.1 Hz with Load [200g]
0
theta (rad)

-0.2

-0.4
-0.5
0 20 40 60 80 0 20 40 60 80
0.05 0.1
error (rad)

0 0

-0.05 -0.1
0 20 40 60 80 0 20 40 60 80
6
voltage [V]

4 5

2
0
0 20 40 60 80 0 20 40 60 80
time (sec) ref PID PID-INN time (sec)

Fig.18. Tracking performance of the proposed controller in case of sinusoidal 0.1[Hz] with Payload
Joint 1 - Sine 0.5 Hz with Load [200g] Joint 2 - Sine 0.5 Hz with Load [200g]
0 0
theta (rad)

-0.5 -0.5
0 10 20 30 40 50 0 10 20 30 40 50
0.2
0.2
error (rad)

0 0
-0.2
-0.2
0 10 20 30 40 50 0 10 20 30 40 50
10
voltage (V)

5
5

0 0
0 10 20 30 40 50 0 10 20 30 40 50
time (sec) time (sec)
ref PID PID-INN

Fig.19. Tracking performance of the proposed controller in case of sinusoidal 0.5[Hz] with Payload

23
 Joint 1 - Triangle 0.2 Hz with Load (200g) Joint 2 - Triangle 0.2 Hz with Load (200g)
0 0
theta (rad)
-0.2

-0.4 -0.5
0 20 40 60 80 0 20 40 60 80
0.05
0.02
error (rad)

0 0
-0.02
-0.05
0 20 40 60 80 0 20 40 60 80
voltage (V)

5
5

0 0
0 20 40 60 80 0 20 40 60 80
time (sec) ref PID PID-INN time (sec)

Fig.20. Tracking performance of the proposed controller in case of triangle 0.2[Hz] with Payload

Joint 1 - Triangle 0.5 Hz with Load (200g) Joint 2 - Triangle 0.5 Hz with Load (200g)
0 0
theta (rad)

-0.2

-0.4 -0.5
0 20 40 60 80 0 20 40 60 80
0.05
0.05
error (rad)

0 0
-0.05
-0.05
0 20 40 60 80 0 20 40 60 80
voltage (V)

5
5

0 0
0 20 40 60 80 0 20 40 60 80
time (sec) ref PID PID-INN time (sec)

Fig.21. Tracking performance of the proposed controller in case of triangle 0.5[Hz] with Payload

Table 9. Performance comparision in case study 2

Error criteria-MSE
Case study 2 – with Load
Theta1 Theta2
PID 0.0012 0.0025
Sine 0.1 Hz
PID-INN 2.1674e-04 2.1242e-04
PID 0.0073 0.0153
Sine 0.5 Hz
PID-INN 0.0011 0.0045
PID 2.2468e-04 4.0609e-04
Triangle 0.2 Hz
PID-INN 7.3457e-05 1.2134e-04
PID 3.7189e-04 5.7445e-04
Triangle 0.5 Hz
PID-INN 2.7663e-04 4.3211e-04

24
Fig.22 and Fig.23 show the overall control signal (uPID-INN), the output of feed-forward controller
based INN model (uinn) and the PID controller (upid) in two cases of Sine trajectory and Triangle
trajectory tracking with Payload, respectively. Fig.22 and Fig.23 also show how the adaptation
process of the weights changes over time in these two cases.

(a) Sine 0.1 Hz with Load (200g) (b) Sine 0.5 Hz with Load (200g)
8 pid inn PID-INN 8

Control signal - voltage (V)

6
6
Joint 1

4
4 2
2 0
0 -2
0 20 40 60 80 0 10 20 30 40 50
10 10
Joint 2

5 5

0 0
0 -3 20 40 60 80 0 -3 10 20 30 40 50
x 10 x 10
2 5

The adaption process

dv/dt

0 0
-2 -5
20 40 60 80 10 20 30 40 50
1 5
dw/dt

0 0

-1 -5
20 40 60 80 10 20 30 40 50
time (sec) time (sec)

Fig.22. Proposed controller parameters adaptively adjusted during control process

(Sinusoidal trajectory tracking with Payload)
(a) Triangle 0.2 Hz with Load (200g) (a) Triangle 0.5 Hz with Load (200g)
8 pid inn PID-INN
8

Control signal - voltage (V)

6 6
Joint 1

4 4
2 2
0 0
0 20 40 60 80 0 20 40 60 80
8 8
6 6
Joint 2

4 4
2 2
0 0
0 -3 20 40 60 80 0 -3 20 40 60 80
x 10 x 10
2 2
The adaption process
dv/dt

0 0
-2 -2
20 40 60 80 20 40 60 80
0.5 1
dw/dt

0 0

-0.5 -1
20 40 60 80 20 40 60 80
time (sec) time (sec)

Fig.23. Proposed controller parameters adaptively adjusted during control process

(Triangle trajectory tracking with Payload)

25
In Fig.24, the comparison between conventional PID controller and the proposed controller is
performed on the end-effecter trajectory tracking in case of without and with the payload. From
these experimental results, the response of the proposed controller is always in good agreement with
that of reference trajectory in comparison with the response of PID controller. The error between
reference signal trajectory and end-effecter trajectory tracking output proves adaptively and
continually minimized during control process.
Trajectory without Load Trajectory with Load (200g)
11.5 11.5

11 11

10.5 10.5

10 10
y (cm)

y (cm)
9.5 9.5

9 9
ref
8.5 8.5
PID
PID-INN
8 8
32 34 36 38 32 34 36 38
x (cm) x (cm)

Fig. 24. Performance comparison of end-effecter trajectory tracking

In summary, based on the above results, we see that the proposed adaptive feed-forward-PID
controller possesses an adaptive auto-tuning ability for accurately and robustly position control the
parallel PAM robot. The proposed control method maintains a very good performance despite the
impact of uncertain factors such as load changing and other external disturbances. The adaptation of
the proposed controller reflected in the fact that control error adaptively minimized and the weights
of the proposed neural MDE-INN model has been online adjusted during the control process.
Therefore, it is verified that the proposed controller ensures the adaptive online control ability with
a simple structure and obtains a very good dynamic property and a perfect performance as well in
comparison with the traditional PID controller. Furthermore, the proposed controller does not need
any information about the Jacobian of the nonlinear dynamic system. Thus, the proposed controller
is quite suitable for the robust control of various plants, including linear and nonlinear dynamic
systems.

6. Conclusions
This paper proposes the novel adaptive feed-forward - PID controller to control the end-effecter
position of the SCARA parallel PAM robot. The main idea of this study uses the inverse neural
NARX model optimized the new MDE algorithm in the feed-forward leg. The proposed controller
parameters are online auto-tuned based on the adaptive back-propagation algorithm during control

26
process. Due to this combination, the proposed controller obtains an excellent control quality,
strongly adaptive and robust capability regardless of external disturbances. Experimental results
prove that the novel control method possesses the capability to learn and update automatically along
with the inverse dynamics of the nonlinear parallel PAM robot as to reduce the tracking error to
nearly zero.
Acknowledgements
This paper is funded with the Vietnam National Foundation for Science and Technology
Development (NAFOSTED) under grant number 107.01-2015.23 and with the Vietnam National
University Ho Chi Minh City (VNU-HCM) under grant number B2016-20-03.
References
[1]. Andrikopoulos, G. Nikolakopoulos, and S. Manesis, “Design and development of an
exoskeletal wrist prototype via pneumatic artificial muscles,” Meccanica, vol. 50, no. 11, pp.
2709–2730, 2015.
[2]. I. Koo, B. B. Kang, and K.-J. Cho, “Development of hand exoskeleton using pneumatic
artificial muscle combined with linkage,” J. Korean Soc. Precis. Eng., vol. 30, no. 11, pp.
1217–1224, 2013.
[3]. T. Noritsugu, “Wearable power assist robot driven with pneumatic rubber artificial muscles,”
Virtual Real. Enhanc. Robot. Syst. Disabil. Rehabil., p. 235, 2016.
[4]. E. Ball and E. Garcia, “Effects of Bladder Geometry in Pneumatic Artificial Muscles,” J.
Med. Device., 2016.
[5]. M. Tian, Y. Xiao, X. Wang, J. Chen, and W. Zhao, “Design and Experimental Research of
Pneumatic Soft Humanoid Robot Hand,” in Robot Intelligence Technology and Applications
4, Springer, 2017, pp. 469–478.
[6]. B. Tondu, S. Ippolito, J. Guiochet, and A. Daidie, “A seven-degrees-of-freedom robot-arm
driven by pneumatic artificial muscles for humanoid robots,” Int. J. Rob. Res., vol. 24, no. 4,
pp. 257–274, 2005.
[7]. F. Daerden and D. Lefeber, “Pneumatic artificial muscles: actuators for robotics and
automation,” Eur. J. Mech. Environ. Eng., vol. 47, no. 1, pp. 11–21, 2002.
[8]. B. K. S. Woods, C. S. Kothera, and N. M. Wereley, “Wind tunnel testing of a helicopter rotor
trailing edge flap actuated via pneumatic artificial muscles,” J. Intell. Mater. Syst. Struct., vol.
22, no. 13, pp. 1513–1528, 2011.
[9]. R. Ban, Y. Kimura, and T. Nakamura, “Development of insertion-type peristalsis pump using
pneumatic artificial muscles,” Adv. Robot., vol. 29, no. 13, pp. 877–888, 2015.
[10]. B. K. S. Woods, M. F. Gentry, C. S. Kothera, and N. M. Wereley, “Fatigue life testing of
swaged pneumatic artificial muscles as actuators for aerospace applications,” J. Intell. Mater.
Syst. Struct., vol. 23, no. 3, pp. 327–343, 2012.
[11]. H. F. Schulte, “The characteristics of the McKibben artificial muscle,” Appl. Extern. power
prosthetics Orthot., vol. 874, pp. 94–115, 1961.
[12]. G. Andrikopoulos, G. Nikolakopoulos, and S. Manesis, “Non-Linear Control of Pneumatic
Artificial Muscles,” in Control & Automation (MED), 2013 21st Mediterranean Conference
on, 2013, pp. 729–734.
[13]. X. Li, F. He, X. Hong, and T. Guan, “Implicit Generalized Predictive Control of Hip-Joint
Rehabilitation Training Device Driven by Pneumatic Muscle Actuator,” in Applied Mechanics
and Materials, 2012, vol. 138, pp. 273–278.
[14]. A. Pujana-Arrese, A. Mendizabal, J. Arenas, R. Prestamero, and J. Landaluze, “Modelling in
Modelica and position control of a 1-DoF set-up powered by pneumatic muscles,”
Mechatronics, vol. 20, no. 5, pp. 535–552, 2010.

27
[15]. S. Ganguly, A. Garg, A. Pasricha, and S. K. Dwivedy, “Control of pneumatic artificial muscle
system through experimental modelling,” Mechatronics, vol. 22, no. 8, pp. 1135–1147, 2012.
[16]. K. Xing, J. Huang, Y. Wang, J. Wu, Q. Xu, and J. He, “Tracking control of pneumatic
artificial muscle actuators based on sliding mode and non-linear disturbance observer,” IET
Control theory Appl., vol. 4, no. 10, pp. 2058–2070, 2010.
[17]. X. Shen, “Nonlinear model-based control of pneumatic artificial muscle servo systems,”
Control Eng. Pract., vol. 18, no. 3, pp. 311–317, 2010.
[18]. B. Tondu and P. Lopez, “Modeling and control of McKibben artificial muscle robot
actuators,” IEEE Control Syst., vol. 20, no. 2, pp. 15–38, 2000.
[19]. K. Xing, Y. Wang, Q. Zhu, and H. Zhou, “Modeling and control of McKibben artificial
muscle enhanced with echo state networks,” Control Eng. Pract., vol. 20, pp. 477–488, 2012.
[20]. K. K. Ahn and H. P. H. Anh, “Design and implementation of an adaptive recurrent neural
networks (ARNN) controller of the pneumatic artificial muscle (PAM) manipulator,”
Mechatronics, vol. 19, no. 6, pp. 816–828, 2009.
[21]. J. Fan, J. Zhong, J. Zhao, and Y. Zhu, “BP neural network tuned PID controller for position
tracking of a pneumatic artificial muscle,” Technol. Heal. Care, vol. 23, no. s2, pp. S231–
S238, 2015.
[22]. H. P. H. Anh, “Online tuning gain scheduling MIMO neural PID control of the 2-axes
pneumatic artificial muscle (PAM) robot arm,” Expert Syst. Appl., vol. 37, no. 9, pp. 6547–
6560, 2010.
[23]. D. Q. Truong and K. K. Ahn, “Synchronization controller for a 3-R planar parallel pneumatic
artificial muscle (PAM) robot using modified ANFIS algorithm,” Mechatronics, vol. 23, no. 4,
pp. 462–479, 2013.
[24]. X. Jiang and et. al, “Fuzzy neural network control of the rehabilitation robotic arm driven by
pneumatic muscles,” Ind. Robot An Int. J., vol. 42, no. 1, pp. 36–43, 2015.
[25]. D. T. Liem, D. Q. Truong, H. G. Park, and K. K. Ahn, “A feedforward neural network fuzzy
grey predictor-based controller for force control of an electro-hydraulic actuator,” Int. J.
Precis. Eng. Manuf., vol. 17, no. 3, pp. 309–321, 2016.
[26]. C. Song, S. Xie, Z. Zhou, and Y. Hu, “Modeling of pneumatic artificial muscle using a hybrid
artificial neural network approach,” Mechatronics, vol. 31, pp. 124–131, 2015.
[27]. M. Gori and A. Tesi, “On the problem of local minima in backpropagation,” IEEE Trans.
Pattern Anal. Mach. Intell., vol. 14, no. 1, pp. 76–86, 1992.
[28]. X. Wang, Z. Tang, H. Tamura, M. Ishii, and W. D. Sun, “An improved backpropagation
algorithm to avoid the local minima problem,” Neurocomputing, vol. 56, pp. 455–460, 2004.
[29]. F. H. F. Leung, H.-K. Lam, S.-H. Ling, and P. K. S. Tam, “Tuning of the structure and
parameters of a neural network using an improved genetic algorithm,” Neural Networks, IEEE
Trans., vol. 14, no. 1, pp. 79–88, 2003.
[30]. G. Das, P. K. Pattnaik, and S. K. Padhy, “Artificial neural network trained by particle swarm
optimization for non-linear channel equalization,” Expert Syst. Appl., vol. 41, no. 7, pp. 3491–
3496, 2014.
[31]. K. Socha and C. Blum, “An ant colony optimization algorithm for continuous optimization:
application to feed-forward neural network training,” Neural Comput. Appl., vol. 16, no. 3, pp.
235–247, 2007.
[32]. E. Valian, S. Mohanna, and S. Tavakoli, “Improved cuckoo search algorithm for feedforward
neural network training,” Int. J. Artif. Intell. Appl., vol. 2, no. 3, pp. 36–43, 2011.
[33]. R. Storn and K. Price, Differential evolution-a simple and efficient adaptive scheme for global
optimization over continuous spaces, vol. 3. ICSI Berkeley, 1995.
[34]. J. Ilonen, J.-K. Kamarainen, and J. Lampinen, “Differential evolution training algorithm for
feed-forward neural networks,” Neural Process. Lett., vol. 17, no. 1, pp. 93–105, 2003.
[35]. B. Subudhi and D. Jena, “Evolutionary Computing Approaches to System Identification,” in
Handbook of Research on Computational Intelligence Applications in Bioinformatics, IGI
Global, 2016, pp. 329–368.

28
[36]. L. Y. Su, “Micro-Vibration Mechanism of Micro-Gears Fault Diagnosis Based on Fault
Characteristics and Differential Evolution Wavelet Neural Networks,” in Applied Mechanics
and Materials, 2014, vol. 508, pp. 219–222.
[37]. A. P. Piotrowski, “Differential Evolution algorithms applied to Neural Network training suffer
from stagnation,” Appl. Soft Comput., vol. 21, pp. 382–406, 2014.
[38]. A. Slowik, “Application of an adaptive differential evolution algorithm with multiple trial
vectors to artificial neural network training,” IEEE Trans. Ind. Electron., vol. 58, no. 8, pp.
3160–3167, 2011.
[39]. S. Curteanu, G. D. Suditu, A. M. Buburuzan, and E. N. Dragoi, “Neural networks and
differential evolution algorithm applied for modelling the depollution process of some gaseous
streams,” Environ. Sci. Pollut. Res., vol. 21, no. 22, pp. 12856–12867, 2014.
[40]. P. P. Sarangi, A. Sahu, and M. Panda, “A hybrid differential evolution and back-propagation
algorithm for feedforward neural network training,” Int. J. Comput.Appl., vol.84, no.14, 2013.
[41]. B. Subudhi and D. Jena, “Nonlinear system identification using memetic differential evolution
trained neural networks,” Neurocomputing, vol. 74, no. 10, pp. 1696–1709, 2011.
[42]. S. Das, A. Abraham, U. K. Chakraborty, and A. Konar, “Differential evolution using a
neighborhood-based mutation operator,” Evol. Comput. IEEE Trans., vol. 13, no. 3, pp. 526–
553, 2009.
[43]. A. K. Qin, V. L. Huang, and P. N. Suganthan, “Differential evolution algorithm with strategy
adaptation for global numerical optimization,” Evol. Comput. IEEE Trans., vol. 13, no. 2, pp.
398–417, 2009.
[44]. M. T. Hagan and M. B. Menhaj, “Training feedforward networks with the Marquardt
algorithm,” IEEE Trans. Neural Networks, vol. 5, no. 6, pp. 989–993, 1994.
[45]. R. Storn and K. Price, “Differential evolution–a simple and efficient heuristic for global
optimization over continuous spaces,” J. Glob. Optim., vol. 11, no. 4, pp. 341–359, 1997.
[46]. R. Gämperle, S. D. Müller, and P. Koumoutsakos, “A parameter study for differential
evolution,” Adv. Intell. Syst. fuzzy Syst. Evol. Comput., vol. 10, pp. 293–298, 2002.
[47]. J. Ronkkonen, S. Kukkonen, and K. V Price, “Real-parameter optimization with differential
evolution,” in Proc. IEEE CEC, 2005, vol. 1, pp. 506–513.
[48]. J. Liu and J. Lampinen, “A fuzzy adaptive differential evolution algorithm,” Soft Comput.,
vol. 9, no. 6, pp. 448–462, 2005.
[49]. J. Brest, B. Bošković, S. Greiner, V. Žumer, and M. S. Maučec, “Performance comparison of
self-adaptive and adaptive differential evolution algorithms,” Soft Comput., vol. 11, no. 7, pp.
617–629, 2007.
[50]. Y. Wang, Z. Cai, and Q. Zhang, “Differential evolution with composite trial vector generation
strategies and control parameters,” Evol. Comput. IEEE Trans., vol.15, no.1, pp. 55–66, 2011.
[51]. A. M. Ideta, “Stability of feedback error learning method with time delay,” Neurocomputing,
vol. 69, no. 13, pp. 1645–1654, 2006.
[52]. J. Nakanishi and S. Schaal, “Feedback error learning and nonlinear adaptive control,” Neural
Networks, vol. 17, no. 10, pp. 1453–1465, 2004.
[53]. K. Sabahi, S. Ghaemi, and S. Pezeshki, “Application of type-2 fuzzy logic system for load
frequency control using feedback error learning approaches,” Appl. Soft Comput., vol. 21, pp.
1–11, 2014.
[54]. K. Kurosawa, R. Futami, T. Watanabe, and N. Hoshimiya, “Joint angle control by FES using a
feedback error learning controller,” IEEE Trans. Neural Syst. Rehabil. Eng., vol. 13, no. 3, pp.
359–371, 2005.
[55]. K. Sabahi and M. Teshnehlab, “Recurrent fuzzy neural network by using feedback error
learning approaches for LFC in interconnected power system,” Energy Convers. Manag., vol.
50, no. 4, pp. 938–946, 2009.

29
 Nguyen Ngoc Son received the B.S. and the M.Sc. degrees in the Faculty of
Electrical and Electronics Engineering (FEEE) from HCM City University of
Technology in 2010 and in 2012, respectively. He is now PhD candidate with
the Faculty of Electrical and Electronic Engineering, HCM City University of
Technology, VNU-HCM, Vietnam. He is currently a Lecturer in the Faculty of
Electronics, Industrial University of HCM.

Cao Van Kien received the M.Sc. degrees in the Faculty of Electrical and
Electronics Engineering (FEEE) from HCM City University of Technology in
2016. He is now PhD candidate with the Faculty of Electrical and Electronic
Engineering, HCM City University of Technology, VNU-HCM, Vietnam. He
is currently a Researcher in Power Enginerring Research Laboratory
(PERLAB) in HCM City University of Technology, VNU-HCM, Vietnam.

Ho Pham HuyAnh received the B.S. and the M.Sc. degrees in the Department of
Electrical and Electronics Engineering from HCM City University of Technology in
1987 and in 1993, respectively. He received the Ph.D. degree from University of
Ulsan, Korea in 2008. He is currently a Senior Lecturer in the Faculty of Electrical
and Electronics (FEEE), HCM City University of Technology,VNU-HCM, HCM
City, Viet Nam. Up to now, he co-authored 4 books and published over 60 papers
on national and international journals and conference proceedings. His current
research interests include intelligent control, robotics, novel energy applications,
modeling and identification of nonlinear dynamic system, soft-computing.

30