Beruflich Dokumente
Kultur Dokumente
Real time adaptive nonlinear model inversion control of a twin rotor MIMO
system using neural networks
A. Rahideh a,b, A.H. Bajodah c, M.H. Shaheed b,n
a
School of Electrical and Electronic Engineering, Shiraz University of Technology, Shiraz, Iran
b
School of Engineering and Materials Science, Queen Mary University of London, London E1 4NS, UK
c
King Abdulaziz University, Jeddah, Saudi Arabia
a r t i c l e i n f o abstract
Article history: This paper investigates the development and experimental implementation of an adaptive dynamic
Received 9 May 2011 nonlinear model inversion control law for a Twin Rotor MIMO System (TRMS) using artificial neural
Received in revised form networks. The TRMS is a highly nonlinear aerodynamic test rig with complex cross-coupled dynamics
5 October 2011
and therefore represents the control challenges of modern air vehicles. A highly nonlinear 1DOF
Accepted 29 December 2011
mathematical model of the TRMS is considered in this study and a nonlinear inverse model is
Available online 28 January 2012
developed for the pitch channel of the system. An adaptive neural network element is integrated
Keywords: thereafter with the feedback control system to compensate for model inversion errors. The proposed
Inverse-control on-line learning algorithm updates the weights and biases of the neural network using the error
Neural networks
between the set-point and the real output. The real-time response of the method shows a satisfactory
Nonlinear system
tracking performance in the presence of inversion errors caused by model uncertainty. The approach is
On-line learning
Real-time therefore deemed to be suitable to apply real-time to other nonlinear systems with necessary
modifications.
& 2012 Elsevier Ltd. All rights reserved.
0952-1976/$ - see front matter & 2012 Elsevier Ltd. All rights reserved.
doi:10.1016/j.engappai.2011.12.006
1290 A. Rahideh et al. / Engineering Applications of Artificial Intelligence 25 (2012) 1289–1297
Shital et al., 2010; Zhai and Yu, 2009). In the case of model were provided. Other extensions were made to consider struc-
inversion control, neural network training of the plant input– tural load limits (Sahani and Horn, 2006) and the output feedback
output data is being utilised to perform dynamic inversion. Both control problem (Hovakimyan et al., 1999).
on-line (Kim and Calise, 1997) and off-line (Behera et al., 1996) Linear control techniques may show satisfactory performance
training algorithms are reported in the dynamic inversion litera- for those nonlinear systems with negligible nonlinearity and/or
ture. Kim and Calise (1997) have suggested off-line training of always working around a fixed operating point. For those highly
neural networks using a nominal mathematical model, which nonlinear systems with continuously varying operating points,
provides an approximate inversion that can accommodate the linear control approaches either fail to control the system or even
total flight envelope. However in real-time situation, on-line if the control system remains stable, it may not be optimal.
training of neural networks is required to compensate for inver- Therefore to optimally control highly nonlinear unmanned aerial
sion error. Behera et al. (1996) have presented a new control vehicles in terms of, for instance, manoeuvrability and/or fuel
scheme based on inversion of a feed-forward neural model of a efficiency, nonlinear adaptive and/or optimal control techniques
robot arm. should be explored.
Another source of model inversion inaccuracy occurs due to In this investigation a nonlinear model inversion based control
plant model uncertainty. Since inversion is model based, high- methodology is developed and implemented in real-time on a
fidelity plant modelling is needed. It is well known that model TRMS, which resembles the dynamics of a helicopter. A detailed
inversion control is sensitive to the accuracy of the inversion description of the flying machine experimental test-bed is
signal, and that inexact inversion can deteriorate the performance included in Section 2. The model inversion process and the
of the feedback control system. For example, to achieve a control methodology developed and implemented are illustrated
physically realisable commanded change in aircraft attitude, the in Section 3. The results and conclusion are presented in Sections
corresponding moments must be determined very accurately 4 and 5, respectively. Section 4 consists of two subsections, one
(Ostroff and Bacon, 1999). Other difficulties encountered when for the model validation to show the accuracy of the dynamic
performing model inversion are related to the nature of plant model developed and the other presents the response of the
itself. For example, nonminimum phase plants like flexible proposed control method. This investigation is the experimental
structures cause instability of the inversion process, i.e., the version of the previously published research by the same authors
required control inputs are unbounded for bounded desired plant (Rahideh et al., 2007) in which the adaptive nonlinear model
trajectories (Devasia et al., 1996). Researchers have suggested inversion control method has only been verified in simulation.
augmenting adaptive neural network based compensators with
feedback control system to compensate for the inversion errors
effect on the closed loop system, (Yesildirek and Lewis, 1995). 2. TRMS
This type of augmentation to model inversion control has been
reported abundantly in the literature. For instance, dynamic The TRMS is a laboratory platform designed for control
inversion nonlinear control system designs were proposed for experiments by Feedback Instruments Ltd (Feedback, 1998). In
fixed (Kim and Calise, 1997) and rotary (Lee et al., 2005) wing certain aspects, its behaviour resembles the dynamics of a
aircraft models, and adaptive neural network elements were helicopter (Ahmad et al., 2003; Rahideh et al., 2008). For example,
augmented to robustify the command augmentation systems it possesses a strong cross-couplings between the collective
against model inversion errors. The methodology was also applied (main) and tail rotors. The TRMS is characterised by its complex
to enhance flying quality characteristics of tilt-rotor airplanes and highly nonlinear dynamics. Some of its states and outputs are
(Rysdyk and Calise, 1999, 2005). Other aerospace applications also inaccessible for measurements. All these typify TRMS as a
include deep-space spacecraft formation flying (Gurfil et al., 2002) challenging engineering problem. The control objective is to make
and airship vehicle flight (Park et al., 2003). An extension of the the beam of the TRMS tracks a predetermine trajectory. Fig. 1
methodology was made to consider the effects of actuator shows the TRMS considered in this investigation. The dynamic
characteristics and limitations on the adaptation process, such model as supplied by the manufacturer has been improved in this
as actuator displacement and/or rate limits, actuator dynamics study and the electric motors are modelled with respect to the
and time delay (Johnson and Calise, 2003; Johnson and Kannan, corresponding equations. The TRMS possesses two permanent
2005). Applications in reusable launch vehicles and helicopters magnet DC motors; one for the main and the other for the tail
A. Rahideh et al. / Engineering Applications of Artificial Intelligence 25 (2012) 1289–1297 1291
−v
Fig. 3. Gravity and propulsive forces in the vertical plane.
Tfric,v
Lav Rav
Ωv
iav +
+
Vv N Eav S
_ _
av
dov 1
¼ ðT ev T Lv Bmr ov Þ ð3Þ dav
dt J mr ¼ Ov ð8Þ
dt
T ev ¼ kav fv iav ð4Þ where Fv is the aerodynamic force produced by the main rotor, mt,
mtr, mts, mm, mmr, mms, mb and mcb are, respectively, the mass of
T Lv ¼ ktv 9ov 9ov ð5Þ the tail part of the beam, tail motor, tail shield, main part of the
where Vv is the voltage control input of the vertical channel, Eav, beam, main motor, main shield, counter-weight beam and coun-
Rav, Lav and iav are, respectively, the electromotive force, armature ter-weight, lt, lm and lb are, respectively, the length of the tail part
resistance, armature inductance and armature current of the main of the beam, main part of the beam and counter-weight beam, lcb
motor, kav and ktv are constants, jv is the magnetic flux, ov is the is the distance between the counterweight and the joint, av is the
rotational velocity, Tev is the electromagnetic torque, TLv is the pitch angle of the TRMS beam, Ov is the angular velocity of the
load torque, Jmr is the rotor moment of inertia and Bmr is the rotor TRMS beam in the vertical plane, Jv is the moment of inertia about
damping coefficient all of the main motor. horizontal axis, Tfric,v is the friction torque in vertical plane and
The mathematical model of the remaining parts of the system kfvp and kfvn are two positive constants.
in vertical plane is described in (6) to (8) (see Fig. 3). Fig. 4 shows the profile of the friction torque that covers
In (6) the first term denotes the torque of the propulsive force viscous, coulomb and static frictions. It is worth noting that there
due to the main rotor, the second term refers to the torque of the are differences between the input voltage levels in the MATLAB/
friction force, and the torque of gravity force is shown in the third Simulink environment and the motor terminal voltages, and that
term the relationship between these two sets of values is slightly
nonlinear. For more details on the one degree of freedom
dOv lm F v ðov ÞT f ric,v þ g ½ðABÞcos av C sin av (1DOF) TRMS model including how the parameters in Eqs. (1)–(8)
¼ ð6Þ
dt Jv have been obtained see (Rahideh et al., 2008).
1292 A. Rahideh et al. / Engineering Applications of Artificial Intelligence 25 (2012) 1289–1297
2
v,ref Model Uv Real v
1.5 Inversion TRMS
1
Fig. 7. Open loop model inversion control strategy.
0.5
Uv
Vv
−1
NN
−1.5 Compensator
PD
−2
Adaptation
−2.5 Law
−25 −20 −15 −10 −5 0 5 10 15 20 25
_
Uv +
Fig. 5. Interface circuit model (relationship between Uv and Vv). Fig. 8. Total block diagram of the adaptive nonlinear model inversion control.
TRMS
v Fv(v) Propeller v Motor Vv Interface Uv
Dynamics
Inversion Inversion Inversion Inversion
4. Results
xi ¼ av ðti þ1Þ i ¼ 1,2,3 ð14Þ
2 3 The responses of the 1DOF inverse model and the proposed
bw
6 7 nonlinear adaptive controller are presented in this section.
6 w1 7
W ¼6
6 ^ 7
7 ð15Þ
4 5 4.1. Model validation
w10
2 3 The 1DOF dynamic model of the TRMS has been tested with
bv,1 bv,2 bv,10 various input signals in order to show its accuracy. Figs. 10–12
6 7
6 v1,1 v1,2 v1,10 7 depict the responses of the model to three different types of input
V¼6
6 ^
7 ð16Þ
4 ^ & ^ 7 5 signals compared to that of the real TRMS. As seen in the figures,
v3,1 v3,2 v3,10 in all cases the model responses closely match those of the
real TRMS.
1
f j ðzj Þ ¼ , j ¼ 12, ,10 ð17Þ
1 þ eaj zj 4.2. Controller response
X
3
In this section the experimental results of the proposed non-
zj ¼ bv,j þ vi,j xi , j ¼ 1,2, ,10 ð18Þ
i¼1 linear adaptive controller are presented. More details on nonlinear
model accuracy of the TRMS can be found in (Rahideh et al., 2008).
2 3
1 The real TRMS is connected to a computer by a fast interface
6 7 circuit board-PCL-812. The PCL-812PG can be considered a high
6 f 1 ðz1 Þ 7
F ¼6
6
7
7 ð19Þ performance, high speed, multi-function data acquisition card
4 ^ 5
for IBM PC/XT/AT and compatible computers from Advantech Co.
f 10 ðz10 Þ
The responses of the control system, as experimentally obtained
After defining the neural network structure, i.e. the number of using a sampling time of 0.001 s, are shown in Figs. 13–18. The
layers and the number of neurons in each layer, the neural responses of the proposed NN-based adaptive controller and a PID
1294 A. Rahideh et al. / Engineering Applications of Artificial Intelligence 25 (2012) 1289–1297
1
Real TRMS 101
PSD of real TRMS
0.8 Model
100 PSD of the model
0.6 10−1
Pitch angle (rad)
10−2
Magnitude
0.4
10−3
0.2
10−4
0
10−5
−0.2
10−6
−0.4 10−7
0 20 40 60 80 100 0 0.5 1 1.5 2 2.5
time (sec) Frequency (Hz)
Fig. 10. Pitch angle response to a sine wave input with frequency 0.02 Hz and amplitude 1.5 V. (a) Time domain comparison; (b) frequency domain comparison.
0.6
Real TRMS 100
0.5 Model PSD of real TRMS
PSD of the model
10−1
0.4
0.3 10−2
Pitch angle (rad)
Magnitude
0.2 10−3
0.1
10−4
0
10−5
−0.1
−0.2 10−6
−0.3 10−7
0 10 20 30 40 50 0 0.5 1 1.5 2 2.5
time (sec) Frequency (Hz)
Fig. 11. Pitch angle response to a sine wave input with frequency 0.1 Hz and amplitude 1 V. (a) Time domain comparison; (b) frequency domain comparison (PSD).
0.8 100
Real TRMS PSD of real TRMS
Model PSD of the model
0.6 10−1
0.4 10−2
Pitch angle (rad)
Magnitude
0.2 10−3
0
10−4
−0.2
10−5
−0.4
10−6
0 10 20 30 40 50 0 0.5 1 1.5 2 2.5
time (sec) Frequency (Hz)
Fig. 12. Pitch angle response to a square wave input with frequency 0.1 Hz and amplitude 1 V. (a) Time domain comparison; (b) frequency domain comparison (PSD).
controller to a sine wave reference signal with an amplitude also confirmed here. Figs. 15 and 16 show the controller responses
of 0.2 are depicted in Fig. 13(a). The power spectral density to a sine wave signal with an amplitude of 0.4 but with and
results corresponding to the time domain responses shown in without an initial value of 0.2, respectively. In Fig. 15 the results of
Fig. 13(a) are depicted in Fig. 13(b). It is demonstrated from both the proposed NN-based adaptive controller are compared to those
the time domain and frequency domain point of view that the obtained from the PID controller. To validate the reliability and
controller is able to follow the reference closely. Fig. 14 also accuracy of the controller further, it is tested with a square wave
presents the response of the controller to a sine wave signal but reference signal and the responses are presented and compared
with an initial value of 0.1 rad. Satisfactory control performance is with those from the PID controller in Fig. 17. Finally, Fig. 18
A. Rahideh et al. / Engineering Applications of Artificial Intelligence 25 (2012) 1289–1297 1295
0.25 100
Reference Reference
0.2 Adaptive NN Adaptive NN
−1
PID 10 PID
0.15
0.1 10−2
Pitch angle (rad)
0.05
Magnitude
10−3
0
−0.05 10−4
−0.1 10−5
−0.15
10−6
−0.2
−0.25 10−7
0 20 40 60 80 100 0 0.5 1 1.5 2 2.5
time (sec) Frequency (Hz)
Fig. 13. Response of the NN-based adaptive controller and the PID controller to a sine wave reference signal with amplitude 0.2. (a) Time domain comparison;
(b) frequency domain comparison (PSD).
0.25 100
Reference Reference
0.2 Adaptive NN Adaptive NN
10−1
0.15
0.1 10−2
Pitch angle (rad)
0.05
Magnitude
10−3
0
10−4
−0.05
−0.1 10−5
−0.15
10−6
−0.2
−0.25 10−7
0 20 40 60 80 100 0 0.5 1 1.5 2 2.5
time (sec) Frequency (Hz)
Fig. 14. Response of the NN-based adaptive controller to a sine wave reference signal with amplitude 0.2 and initial value. (a) Time domain comparison; (b) frequency
domain comparison (PSD).
0.5 101
Reference Reference
0.4 Adaptive NN Adaptive NN
100 PID
PID
0.3
10−1
Pitch angle (rad)
0.2
10−2
Magnitude
0.1
10−3
0
−0.1 10−4
−0.2 10−5
−0.3 10−6
−0.4 10−7
0 20 40 60 80 100 0 0.5 1 1.5 2 2.5
time (sec) Frequency (Hz)
Fig. 15. Response of the NN-based adaptive controller and the PID controller to a sine wave reference signal with amplitude 0.4. (a) Time domain comparison;
(b) frequency domain comparison (PSD).
1296 A. Rahideh et al. / Engineering Applications of Artificial Intelligence 25 (2012) 1289–1297
0.5 101
Reference Reference
0.4 Adaptive NN Adaptive NN
100
0.3
0.2 10−1
Pitch angle (rad)
Magnitude
0.1
10−2
0
−0.1 10−3
−0.2 10−4
−0.3
10−5
−0.4
−0.5 10−6
0 20 40 60 80 100 0 0.5 1 1.5 2 2.5
time (sec) Frequency (Hz)
Fig. 16. Response of the NN-based adaptive controller to a sine wave reference signal with amplitude 0.4 and initial value. (a) Time domain comparison; (b) frequency
domain comparison (PSD).
101
Reference
0.6 Adaptive NN
0
10 PID
0.4
10−1
Pitch angle (rad)
Magnitude
0.2 10−2
0 10−3
−0.2
Reference 10−4
Adaptive NN
PID
10−5
−0.4
10−6
0 20 40 60 80 100 0 0.5 1 1.5 2 2.5
time (sec) Frequency (Hz)
Fig. 17. Response of the NN-based adaptive controller and the PID controller to a square wave reference signal with amplitude 0.4. (a) Time domain comparison;
(b) frequency domain comparison (PSD).
0.5 101
Reference
0.4 Adaptive NN
0
10
0.3
0.2 10−1
Pitch angle (rad)
0.1
Magnitude
10−2
0
−0.1 10−3
−0.2 10−4
−0.3 Reference
Adaptive NN 10−5
−0.4
−0.5 10−6
0 20 40 60 80 100 0 0.5 1 1.5 2 2.5
time (sec) Frequency (Hz)
Fig. 18. Response of the NN-based adaptive controller to a square wave reference signal with amplitude 0.4 and initial value. (a) Time domain comparison; (b) frequency
domain comparison (PSD).
A. Rahideh et al. / Engineering Applications of Artificial Intelligence 25 (2012) 1289–1297 1297
depicts the response of the developed method to a square wave Johnson, E.N., Kannan, S.K., 2005. Adaptive trajectory control for autonomous
reference with an initial value of 0.2. In all cases the proposed helicopters. J. Guidance Control Dyn. 28 (3), 524–538.
Johnson, E.N., Calise, A.J., 2003. Limited authority adaptive flight control for
controller shows satisfactory results and outperforms the PID reusable launch vehicles. J. Guidance Control Dyn. 26 (6), 906–913.
controller in terms of transient and steady state performance. As Kim, B.S., Calise, A.J., 1997. Nonlinear flight control using neural networks.
evident from the results, the PID controller suffers from consider- J. Guidance Control Dyn. 20 (1), 26–33.
Kumar, M.V., Suresh, S., Omkar, S.N., Ganguli, R., Sampath, P., 2009. A direct
able overshoot, which may not be acceptable in some unmanned adaptive neural command controller design for an unstable helicopter. Eng.
aerial vehicles with desired agility and manoeuvrability. Appl. Artif. Intell. 22, 181–191.
Lara, D., Romero, G., Sanchez, A., Lozano, R., Guerrero, A., 2010. Robustness margin
for a four rotor mini-rotorcraft: case of study. Mechatronics 20, 143–152.
5. Conclusion Lee, C.-T., Tsai, C.-C., 2010. Nonlinear adaptive aggressive control using recurrent
neural networks for a small scale helicopter. Mechatronics 20, 474–484.
Lee, S., Ha, C., Kim, B.S., 2005. Adaptive nonlinear control system design for
In this investigation, a real-time adaptive model inversion helicopter robust command augmentation. Aerosp. Sci. Technol. 9, 241–251.
control approach has been developed and implemented on a Marconi, L., Naldi, R., 2008. Aggressive control of helicopters in presence of
parametric and dynamical uncertinities. Mechatronics 18, 381–389.
TRMS in terms of 1DOF. An accurate dynamic model has been Ostroff, A.J., Bacon, B.J., 1999. Force and moment approach for achievable
extracted and a nonlinear inverted model has been developed dynamics using nonlinear dynamic inversion. In: Proceedings of the AIAA
accordingly. In order to eliminate the inversion error an adaptive Guidance, Navigation, and Control Conference, Portland, Oregon.
Park, C.S., Lee, H., Tahk, M.J. Bang, H., 2003. Airship control using neural network
neural network based compensator has been designed and added
augmented model inversion. In: Proceedings of the IEEE Conference on Control
to the control system. Results show that the model based Applications. vol. 1, pp. 558–563.
controller performance in tracking the trajectories of the pitch Peng, K., Cai, G., Chen, B.M., Dong, M., Lum, K.Y., Lee, T.H., 2009. Design and
implementation of an autonomous flight control law for a UAV helicopter.
angle of the beam is very satisfactory in terms of transient and
Automatica 45, 2333–2338.
steady state criteria. The developed approach can be applied to Prasad, J.V.R., Lipp, A., 1993. Synthesis of a helicopter nonlinear controller using
control any nonlinear aerodynamic system with little or no approximate inversion. Int. J. Math. Comput. Modeling 18 (3/4), 89–100.
modifications to achieve satisfactory control performance pro- Rahideh, A., Shaheed, M.H., Bajodah, A.H., 2007. Adaptive non-linear model
inversion control of a twin rotor multi-input multi-output system using
vided that an invertible dynamic model of the system is available. artificial intelligence. IMechE J. Aerosp. Eng. 221 (3), 343–351.
Rahideh, A., Shaheed, M.H., Huijberts, H.J.C., 2008. Dynamic modelling of a TRMS
References using analytical and empirical approaches. Control Eng. Pract. 16, 241–259.
Ramon, A.F., Edgar, N.S., Alexander, G.L., 2009. Neural block control for synchro-
nous generators. Eng. Appl. Artif. Intell. 22, 1159–1166.
Ahmad, S.M., Chipperfield, A.J., Tokhi, M.O., 2003. Dynamic modelling and linear Rigney, B.P., Pao, L.Y., Lawrence, D.A., 2010. Nonminimum phase adaptive inverse
quadratic Gaussian control of a twin-rotor multi-input multi-output system. control for settle performance applications. Mechatronics 20, 35–44.
Proc. IMechE Part-I: J. Syst. Control Eng. 217 (I3), 203–227. Rysdyk, R., Calise, A.J., 1999. Adaptive model inversion flight control for tilt-rotor
Ahmad, S.M., Chipperfield, A.J., Tokhi, M.O., 2002. Dynamic modelling and open- aircraft. J. Guidance Control Dyn. 22 (3), 402–407.
loop control of a twin rotor multi-input multi-output system. Proc. IMechE Rysdyk, R., Calise, A.J., 2005. Robust nonlinear adaptive flight control for consistent
Part-I: J. Syst. Control Eng. 216 (I6), 477–496. handling qualities. IEEE Trans. Control Sys. Technol. 13 (6), 896–910.
Behera, L., Gopal, M., Chudhury, S., 1996. On adaptive trajectory tracking of a robot Sahani, N.A., Horn, J.F., 2006. Adaptive model inversion control of a helicopter with
manipulator using inversion of its neural emulator. IEEE Trans. Neural Net- structural load limiting. J. Guidance Control Dyn. 29 (2), 411–420.
works 7 (6), 1401–1414. Shaheed, M.H., 2005. Feedforward neural network based nonlinear dynamic model-
Devasia, S., Chen, D.G., Paden, B., 1996. Nonlinear inversion-based output tracking. ling of a TRMS using RPROP algorithm. Aircr. Eng. Aerosp. Technol. 77 (1), 13–22.
IEEE Trans. Autom. Control 41 (7), 930–942. Shital, S., Chiddarwar, N., Babu, R., 2010. Comparison of RBF and MLP neural
Enns, D., Bugajski, D., Hendrick, R., Stein, G., 1994. Dynamic inversion: an evolving networks to solve inverse kinematic problem for 6R serial robot by a fusion
methodology for flight control design. In: AGARD Conference Proceedings, approach. Eng. Appl. Artif. Intell. 23, 1083–1092.
Active Control Technology: Applications and Lessons Learned, Turin, Italy, May Slotine, J.J.E., Li, W., 1991. Applied Nonlinear Control. Prentice Hall, New Jersy,
1994, pp. 7–12. USA.
Feedback Co., 1998. Twin Rotor Mimo System User Manual, UK, /http://www. Tee, K.P., 2008. Adaptive neural network control for helicopters in vertical flight.
feedback-group.comS. IEEE Trans. Control Sys. Technol. 16 (4), 753–762.
Gans, N.R., Dixon, W.E., Lind, R., Kurdila, A., 2009. A hardware in the loop Thampatty, K.C.S., Nandakumar, M.P., Cheriyan, E.P., 2011. Adaptive RTRL based
simulation platform for vision-based control of unmanned air vehicles. neurocontroller for damping subsynchronous oscillations using TCSC. Eng.
Mechatronics 9, 1043–1056. Appl. Artif. Intell. 24, 60–76.
Gurfil, P., Idan, M. Kasdin, N.J., 2002. Adaptive neural control of deep-space Yang, J.-H., Hsu, W.-C., 2009. Adaptive backstepping control for electricity driven
formation flying. In: Proceedings of the American Control Conference, May unmanned helicopter. Control Eng. Pract. 17, 903–913.
2002, pp. 2842–2847. Yesildirek, A., Lewis, F.L., 1995. Feedback linearization using neural networks.
Hovakimyan, N., Rysdyk, R. Calise, A.J., 1999. Dynamic neural networks for output Automatica 31 (11), 1659–1664.
feedback control. In: Proceedings of the IEEE Conference on Decision and Zhai, Y.-J., Yu, D.-L., 2009. Neural network model-based automotive engine air/fuel
Control, pp. 1685–1690. ratio control and robustness evaluation. Eng. Appl. Artif. Intell. 22, 171–180.