Research Article | Open Access
Neural Adaptive PID and Neural Indirect Adaptive Control Switch Controller for Nonlinear MIMO Systems
This paper proposes an adaptive switch controller (ASC) design for the nonlinear multi-input multi-output system (MIMO). In fact, the proposed method is an online switch between the neural network adaptive PID (APID) controller and the neural network indirect adaptive controller (IAC). According to the design of the neural network IAC scheme, the adaptation law has been developed by the gradient descent (GD) method. However, the adaptive PID controller is built based on the neural network combining the PID control and explicit neural structure. The strategy of training consists of online tuning of the neural controller weights using the backpropagation algorithm to select the suitable combination of PID gains such that the error between the reference signal and the actual system output converges to zero. The stability and tracking performance of the neural network ASC, the neural network APID, and the neural network IAC are analyzed and evaluated by the Lyapunov function. Then, the controller results are compared between APID, IAC, and ASC, in this paper, applying to a nonlinear system. From simulations, the proposed adaptive switch controller has better effects both on response time and on tracking performance with smallest MSE.
Tracking performance and fast response time are a major concern in industrial control theory. This interest comes in order to facilitate the use of the systems, to satisfy stricter design criteria, and to avoid the lack of some knowledge of the system. In this context, a great effort is being made within the area of system control. A lot of control system approaches have been studied, and many researchers have developed new techniques.
For instance, the proportional integrate derivative (PID) controller has been widely spread in various applications until these days due to its simple structure and easy implementation [1, 2]. A challenging problem in designing a PID controller is to find its appropriate gain values . For nonlinear systems, PID controller performance decreases under a great deal of tuning efforts of constant gains to get local stability. Then, there remains to explore autotuning schemes for the nonlinear system, which clearly leads to consider some sort of system estimation . A new strategy called adaptive PID is needed which automatically detects the changes in the controlled system and readjusts its gains by improving the response of the PID controller.
In [4–7], a PID controller-based neural network is proposed with self-learning and adaptive ability for nonlinear systems. In , the author proposed an adaptive PID controller using the recursive least square algorithm for single-input single-output stable and unstable systems. However, in  and in , respectively, a PID controller based, firstly, on the wavelet neural network is applied to the AC motor and, secondly, on radial basis function is applied to quadrator. In , online tuning of the neural PID controller is proposed for plant hybrid modelling. However, in , a PID-type fuzzy logic controller based on particle swarm optimization is proposed.
The indirect adaptive control (IAC) is well used in [11–13]. This method is based on two parts: identification and control of the system. For the identification phase, the aim is to obtain a structure with the capacity of representing the behavior of the unknown system. This process may be performed offline or online using measurements of input and output of the system. Neural networks, under certain condition, can approximate any nonlinear function with arbitrary precision. It also has a strong ability of adaptive, self-learning, and self-organization. For the control phase, the goal is to design a controller with the capacity of generating the appropriate signal law in order to carry out the output of the system to the desired target. The IAC is applied for nonlinear dynamic systems using self-recurrent wavelet neural networks via adaptive learning rates , using fuzzy neural networks via kernel smoothing , and using radial basis function . However, in , a performance comparison of neural network training approaches in indirect adaptive control is proposed. In , an indirect hierarchical FCMAC control is proposed for the ball and plate system.
However, hybrid control has received considerable attention during the past two decades, and the class of switching systems is specifically employed in many industrial applications such as in [17–24] and . In fact, in  and , respectively, the PID is mixed with the model reference adaptive control and the model predictive controller. In [19, 20], a mixed method between the PID controller and model reference control is proposed for the nonlinear hydraulic actuator and for an aerial inertially stabilized platform. These methods are proposed to suppress the effect of nonlinear and time-varying parameters and to get faster response time and better tracking performance. Motivated by the above discussion, we aim, in this paper, to combine the neural network adaptive PID (APID) controller with the neural network indirect adaptive control (IAC) and establish a unified framework to get fast better tracking performance for the nonlinear control system. By constructing the Lyapunov function, a sufficient condition is presented to ensure the asymptotic stability of the state tracking control. The proposed switch controller is applied to a nonlinear multi-input multi-output system.
The rest of this paper is organized as follows. Section 2 briefly introduces the nonlinear system problem of the control system. In Section 3, the proposed method is detailed and the stability properties of the adaptive designs are studied. In Section 4, an example of a nonlinear system is presented to illustrate the proposed efficiency of the methods. At last, concluding remarks are detailed in Section 5.
2. Problem Statement
Consider the above discrete nonlinear multivariable system with inputs and outputs expressed in terms of its difference equation in the following form :where , (), and , (), are, respectively, the input and output vectors; and are the number of past process input and output, respectively; is the nonlinear function mapping specified by the model; and is the discrete time index.
The problem is to find a controller to ensure that the system output tracks asymptotically the desired reference , .
In this paper, the control object is to combine the neural network PID controller and the neural network indirect adaptive control as a switch controller for the nonlinear system.
3. Adaptive Switch Controller Structure
In this section, an adaptive switch controller is proposed for the nonlinear MIMO system (1). The controller consists of the neural network APID and neural network IAC approach. We assume that the switching point is decided by the norm of the error between the system output and the desired value during the neural network APID and the neural network IAC stages.
3.1. Neural Network Indirect Adaptive Control
In this section, an indirect adaptive control scheme for a class of MIMO nonlinear systems, shown in Figure 1, is proposed. The multilayer perceptron is used in the model block and in the controller block. Each block consists of three layers with the sigmoid activation function for all neurons. The tracking error and the identification error are used, respectively, to train the neural network controller and the neural network model. Since the learning rate is an essential factor for determining the performance of the neuroidentifier and the neurocontroller trained via the GD method, it is important to find the optimal learning rate.
The ith control law is , , and is given as follows:or in the compact vector-valued form aswith , , , and , where is the neuron number of the input layer, is the neuron number of the hidden layer, and is the Jacobian matrix of .
To update the neural network controller parameters, let us consider the cost function:where the tracking error is
The gradient descent method is applied to train the weights of the neural network controller. The output synaptic weights and the hidden weights are given by the following expressions:where the incremental change of the output weights and the hidden weights of the neural network controller are given as follows :
The used partial derivatives of the cost function with respect to ( can take or ) iswhere is the Jacobian of the multivariable by applying the system sensitivity approximation system and is as follows:
An important advantage of the presented method is that it allows the learning rate to change adaptively during the training process giving better performance than with the fixed one.
On the contrary, the neural network model outputs are given as follows:and can be rewritten in the following compact form:with , , , and , where is the neuron number of the input layer, is the neuron number of the hidden layer, and is the Jacobian matrix of .
The update of the model synaptic weights is given bywhere the incremental change of the output weights and the hidden layer weights of the neural network model are given as follows [13, 16–24]:where , is the identification error vector between the system outputs’ vector and the model network outputs’ vector and the model network parameters are updated by the minimization of the cost function, :
For the stability of the neural network model, the Lyapunov function is detailed. Indeed, let us define a discrete Lyapunov function aswhere , , is the identification error. The change in the Lyapunov function is obtained by
The identification error difference can be represented as follows :where is the synaptic weights of the neural network identifier (, ). Using equation (19), the identification error is going to bewith
The online algorithm suitable for real-time applications is the variable learning that occurs when the learning rate is .
Similar to the neural network model, for the stability of the neural network controller, the Lyapunov function is presented. Indeed, let us define a discrete Lyapunov function, , as follows:where , , is the control error. The change in the Lyapunov function is obtained by
The online algorithm suitable for real-time applications is the variable learning that occurs when the learning rate is .
This controller will be applied in Section 4 to a nonlinear MIMO system.
3.2. Neural Network Adaptive PID Controller
In this section, a control structure based on PID for nonlinear MIMO systems is presented in Figure 2 where the bold arrows indicate the vector-valued form. The PID structure includes the controlled system, the PID controller, and the neural network tuner that is used to adjust adaptively the controller parameters. For such system with outputs, PID controllers and neural network PID tuners have to be stacked in parallel. Each PID tuner gets the current system input , the reference signal , and the control error . The control system itself can be represented with a neural network model which represents this system.
The adaptive controller output is given as follows:where , and are the proportional, integral, and derivative gains of the PID controllers, where the proportional error is given asthe derivative error asand the integral error as
Then, the ith adaptive controller output can be represented in the updating algorithm as follows:
The PID-MLP structure is presented in Figure 3, where the network parameters are given as follows:
The PID-MLP tuner structure is presented in Figure 3, where the output is given by the following expression:withand in the following compact vector-valued form:with as the scaling coefficients, , , and .
The weights update of the PID tuner is obtained by minimizing the cost function as follows:where can take or .
Finally, the parameters’ update of the PID-MLP tuner is obtained as follows:withwhere is the learning rate corresponding to the vector component .
The PID controller parameters are dynamically tuned in accordance with the identification information and optimization index .
For the stability of the neural network PID tuner, the Lyapunov function is presented. Indeed, let us define a discrete Lyapunov function, , as follows:where , , is the control error. The change in the Lyapunov function is obtained by
The control error difference can be represented as follows :where .
Using (24), the control error is going to bewith
The online algorithm suitable for real-time applications is the variable learning that occurs when the learning rate is .
This controller will be applied also in Section 4 to a nonlinear MIMO system.
3.3. Neural Network Adaptive Switch Controller Design
In this section, the adaptive switch controller (ASC) is proposed for the nonlinear system (1) to achieve closed-loop stability and tracking. In fact, in this structure, we design the switching controller consisting of 2 controllers: a neural network adaptive PID and a neural network indirect adaptive control, as shown in Figure 4.
In this section, a summary of the proposed algorithm of the ASC for the nonlinear MIMO system is presented.
3.3.1. Offline Phase
Initialization of parameters , , , , , , , , and of the neural network model, neural network indirect adaptive controller, adaptive neural network PID tuner, and PID controller, respectively, is based on expressions (3), (12), (31), and (34) using a reduced number of observations.
3.3.2. Online Phase
(i)At time instant (), we have a new data , using the obtained input vector (ii)If the condition , where is a given small constant, is satisfied, then the neural network model, given by equation (3), approaches sufficiently the behavior of the system(iii)If the condition , where is a given small constant, is satisfied, then the neural network adaptive PID controller provides sufficiently the control law using expression (31)(iv)If the condition and , where is a given small constant, is satisfied, then the neural network indirect adaptive controller provides sufficiently the control law using expression (2)(v)If is not satisfied, the update of the synaptic weights of the neural network indirect adaptive controller is necessary, using equations (6) and (7)(vi)If is not satisfied, the update of the synaptic weights of the neural network model is necessary, using equations (14) and (15)(vii)If is not satisfied, the update of the synaptic weights of the neural network adaptive PID controller is necessary, using equations (36) and (37)(viii)End
4. Simulation Results
In this section, we examine the effectiveness of the proposed control system for the nonlinear system. Indeed, three different cases concerning the controller structure have been tested for comparison. First, the neural network adaptive PID controller is used. Second, the neural network indirect adaptive control is applied. Third, the new design of the switch structure is applied. Two simulation experiments were carried out for this study. The first one did not consider any disturbance or parameter perturbation. The second experiment was to demonstrate the control performance in presence of added disturbances.
Consider the nonlinear MIMO system taken for this purpose described by the following equation :where and are, respectively, the output and the input of the studied system, .
The model performance is evaluated using the mean square error (MSE)  between the system output and the model one:where is the number of observations.
The tracking control objective for the system is to follow the reference signal, . In this simulation, the reference signal is defined as
4.1. Neural Network Indirect Adaptive Control Results
The first case is to apply the neural network indirect adaptive controller.
Indeed, the mechanism of this approach is based on the online identification which ran firstly and gave good results. The found model parameters are used secondly to synthesize the neural network controller.
Figure 5 presents both the reference signal and the output system , , using neural network indirect adaptive control.
To demonstrate the system stability and tracking performance, in the presence of noise, a white noise was applied to the system output.
The control results after 150 observations are compared in Table 1 where the MSE and time simulation are used as performance indexes in two cases: without noise and in presence of noise.
4.2. Neural Network Adaptive PID Controller
In this case, the neural network adaptive proportional integrate derivate controller is applied. However, the parameters , , and are obtained and adjusted using the neural network.
Figure 9 shows the response of the system toward the reference inputs . The evolution of the tracking errors is presented in Figure 10, and the evolution of the control laws is presented in Figure 11.
The results show that the tracking of the neural network adaptive PID has some oscillations (overshot) at the beginning that might be caused by the randomly set neural network weights.
As time goes on, the system outputs track the reference inputs satisfactorily.
The PID parameters , and are adjusted by the self-learning neural network until the tracking errors approach zero asymptotically. In this simulation, the three learning rates , , and we used were set to 0.46, 0.46, and 0.35, respectively. Figure 12 corresponds to the behavior presented by the neural network adaptive PID controller gains during 150 iterations.
In Table 2, the performance of the online adaptive tuning based neural network PID algorithm in terms of MSE is summarized.
Analysis shows that the response speed, stability, and small system error of the neural network PID controller have been guaranteed.
In the next section, we focus on the proposed switch controller.
4.3. Neural Network Adaptive Switch Controller
The proposed method, presented in Figure 4 and detailed by the proposed algorithm, is applied in this case.
Indeed, the minimum of the tracking error between the used method is used to adapt the correspondent controller, as detailed in the proposed algorithm.
Table 3 Presents the performance of the adaptive switch controller in the two cases: with and without added noise compared to the other used controllers.
According to these Figures 13 and 14 and Table 3, we can get that the switch controller has the best performance for the nonlinear system compared to other controllers (APID and IAC). And the mean squared errors are the smallest, controlled by the switch controller, as shown in Table 3. The output system tracks correctly the reference signals. This method gives good results because it uses in the same time an adaptive indirect controller based on a neural network with a variable learning rate and an adaptive self-tuning proportional integrate derivate.
In Figures 5 and 9, we can observe many oscillations caused by disturbances or the failure of the controller to reject them, especially in Figure 9. However, using the proposed neural network ASC, all these disturbances are rejected, as shown in Figure 13.
Compared to others mentioned here or not, this scheme is based on the neural network which uses a variable learning rate in the neural network indirect adaptive control and in the neural network adaptive PID. This technique (neural network) is a universal approximator and well used in modeling and control. The use of the variable learning rate ameliorates the learning speed, the performance tracking, and the control system stability.
In this paper, is proposed an adaptive switch controller consisting of indirect adaptive control and adaptive proportional integrate derivate controller for multi-input multi-output nonlinear system. The switch algorithm is based on the less tracking error between these two controllers and makes the designed system have faster response time and better tracking performance, simultaneously. This proposed controller is applied for a nonlinear MIMO system. According to the simulation, this method is the best in performance tracking.
The data used to support the findings of this study are included within the article.
Conflicts of Interest
The authors declare that they have no conflicts of interest.
- R. A. Fahmy, R. I. Badr, and F. A. Rahman, “Adaptive PID controller using RLS for SISO stable and unstable systems,” Advances in Power Electronics, vol. 64, no. 1, pp. 27–35, 2018.
- C. Cheng and M.-S. Chiu, “Robust PID controller design for nonlinear processes using JITL technique,” Chemical Engineering Science, vol. 63, no. 21, pp. 5141–5148, 2008.
- O. A. Dominguez-Ramirez, L. E. Ramos-Velasco, L. E. Ramos-Velasco et al., “Online identification for auto-tuning PID based on wavelet neural networks: an experimental validation on an AC motor,” in Proceedings of the European Control Conference, Zürich, Switzerland, July 2013.
- S. Slama, A. Errachdi, and M. Benrejeb, “Adaptive PID controller based on neural networks for MIMO nonlinear systems,” Journal of Theoretical and Applied Information Technology, vol. 97, no. 2, pp. 361–371, 2019.
- X. Yan, L. Xu, and Y. Wang, “The loading control strategy of the mobile dynamometer vehicle based on neural network PID,” Mathematical Problems in Engineering, vol. 2017, Article ID 5658983, 7 pages, 2017.
- W. Lu, J. H. Yang, and X. D. Liu, “The PID controller based on the artificial neural network and the differential evolution algorithm,” Journal of Computers, vol. 7, no. 10, pp. 2368–2375, 2012.
- R. Kumar, S. Srivastava, and J. R. P. Gupta, “Artificial neural network based PID controller for online control of dynamical systems,” in Proceedings of the 1st IEEE International Conference on Power Electronics. Intelligent Control and Energy Systems, Delhi, India, July 2016.
- H. Hasanpour, M. H. Beni, and M. Askari, “Adaptive PID control based on RBF NN for quadrotor,” International Research Journal of Applied and Basic Sciences, vol. 11, no. 2, pp. 177–186, 2017.
- A. Andrášik, A. Mészáros, and S. F. de Azevedo, “On-line tuning of a neural PID controller based on plant hybrid modeling,” Computers and Chemical Engineering, vol. 28, pp. 1499–1509, 2004.
- S. Bouallègue, J. Haggège, M. Ayadi, and M. Benrejeb, “PID-type fuzzy logic controller tuning based on particle swarm optimization,” Engineering Applications of Artificial Intelligence, vol. 25, no. 3, pp. 484–493, 2012.
- C.-F. Hsu, C.-J. Chiu, and J.-Z. Tsai, “Indirect adaptive self-organizing RBF neural controller design with a dynamical training approach,” Expert Systems with Applications, vol. 39, no. 1, pp. 564–573, 2012.
- K. Itamiya and M. Sawada, “Indirect model reference adaptive control system based on dynamic certainty equivalence principle and recursive identifier scheme,” International Journal of Engineering Practical Research, vol. 2, no. 2, 2013.
- A. Errachdi and M. Benrejeb, “Performance comparison of neural network training approaches in indirect adaptive control,” International Journal of Control, Automation and Systems, vol. 16, no. 3, pp. 1448–1458, 2018.
- S. Yoo, J. Park, and Y. Choi, “Indirect adaptive control of nonlinear dynamic systems using self recurrent wavelet neural networks via adaptive learning rates,” Information Sciences, vol. 177, no. 15, pp. 3074–3098, 2007.
- I. C. Vega, L. Moreno-Ahedo, and W. Y. Liu, “Indirect adaptive control with fuzzy neural networks via kernel smoothing,” in Proceedings of the Mexican International Conference on Artificial Intelligence: Advances in Computational Intelligence, pp. 386–398, San Luis Potosí, Mexico, October 2012.
- M. A. Moreno-Armendáriz, C. A. Pérez-Olvera, F. O. Rodríguez, and E. Rubio, “Indirect hierarchical FCMAC control for the ball and plate system,” Neurocomputing, vol. 73, no. 13–15, pp. 2454–2463, 2010.
- M. Cetin and S. Iplikci, “A novel auto-tuning PID control mechanism for nonlinear systems,” ISA Transactions, vol. 58, pp. 292–308, 2015.
- J. Habibi, B. Moshiri, and A. K. Sedigh, “Performance benefits of hybrid control design for switched linear systems,” in Proceedings of the SICE-ICASE International Joint Conference, Bexco, Busan, Korea, October 2006.
- X. Zuo, J.-W. Liu, X. Wang, and H.-Q. Liang, “Adaptive PID and model reference adaptive control switch controller for nonlinear hydraulic actuator,” Mathematical Problems in Engineering, vol. 2017, Article ID 6970146, 15 pages, 2017.
- X. Y. Zhou, C. Yang, and T. Cai, “A model reference adaptive control/PID compound scheme on disturbance rejection for an aerial inertially stabilized platform,” Journal of Sensors, vol. 2016, Article ID 7964727, 11 pages, 2016.
- S. Armaghan, A. Moridi, and A. K. Sedigh, “Design of a switching PID controller for a magnetically actuated mass spring damper,” in Proceedings of the World Congress on Engineering Vol III, London, UK, July 2011.
- K. S. Narendra and K. Parthasarathy, “Identification and control of dynamical systems using neural networks,” IEEE Transactions on Neural Networks, vol. 1, no. 1, pp. 4–27, 1990.
- S. Slama, A. Errachdi, and M. Benrejeb, “Online MRAC method using neural networks based on variable learning rate for nonlinear systems,” International Journal of Computer Applications in Technology, vol. 60, no. 3, pp. 215–224, 2019.
- A. Errachdi, I. Saad, and M. Benrejeb, “Neural modeling of multivariable nonlinear stochastic system. variable learning rate case time-varying system,” Intelligent Control and Automation, vol. 2, no. 2, pp. 167–175, 2011.
- F. Lajmi, A. J. Talmoudi, and H. Dhouibi, “Fault diagnosis of uncertain systems based on interval fuzzy PETRI net,” Studies in Informatics and Control, vol. 26, no. 2, pp. 239–248, 2017.
- Y. Feng, Y. Peng, N. Cui, D. Gong, and K. Zhang, “Modeling reference evapotranspiration using extreme learning machine and generalized regression neural network only with temperature data,” Computers and Electronics in Agriculture, vol. 136, pp. 71–78, 2017.
Copyright © 2019 Sabrine Slama et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.