The controlled objects of actual industrial processes are mostly multivariable. One of their important characteristics is the existence of couplings among the variables. That is to say, when one of the input variables of the object changes, many output variables or even all the output variables may change. Due to the existence of couplings among the variables, when the output variable of one control loop is changed by adjusting the controller parameters in the control loop, the output variables of the other control loops are also changed. This results in a degraded control effect, or even invalidity of the whole control system. Decoupling control system is one of the five advanced control systems in multivariable process control [1], whose research is very important.
Decoupling control methods are divided into openloop decoupling control and closedloop decoupling control. The openloop decoupling control is realized in two steps. First, a decoupling compensator is designed, which transforms a multivariable system to multiple singlevariable systems. Then, a controller is designed, which makes these singlevariable control systems achieve the desired performance. By combining a decoupling compensator with a controller, the closedloop decoupling control is realized in one step. The coupling effect among the loops is reduced, and simultaneously the performance requirements are also achieved.
The earliest decoupling control method can be traced back to 1950s, where it was proposed for multivariable linear systems described by an inputoutput model. The decoupling control problem of multivariable linear systems described by a state space model was first proposed by Morgan in 1964. The sufficient and necessary conditions that a square system described by a state space model could be decoupled were given in [2]. The decoupling control problem of nonsquare systems was solved in [3]. For multivariable linear systems with unmeasurable states, the output feedback based decoupling control method was proposed in [4]. There also exist other classical linear decoupling control methods, such as the inverse Nyquist matrix method [5], the BristolShinskey method [6] and the related analysis method [7].
By the 1980s, linear multivariable adaptive decoupling control had become an important research field. Linear multivariable adaptive decoupling control methods are divided into openloop adaptive decoupling control and closedloop adaptive decoupling control. On the basis of identification, both of them use the conventional decoupling control method to decouple the system [8][10].
Since the 1990s, with the introduction of intelligent methods, nonlinear multivariable adaptive decoupling control has attracted the attention of control scholars and engineers. For a class of discretetime multivariable systems with strong coupling, strong nonlinearity and uncertainty, the sufficient and necessary conditions that a system could be decoupled in a compact set were given in [11]. In [12], the controlled system was linearized at an equilibrium point by using Taylor expansion, and then it was equivalently expressed as a combination of a linear model and a nonlinear higher order term. Based on the equivalent model, an adaptive decoupling control method was proposed by combining one stepahead optimal weighting decoupling control with neural network feedforward compensation. In [13] and [14], the system was transformed into a linear model with diagonal parameter matrices and a nonlinear term. By combining neural network feedforward compensation respectively with one stepahead optimal weighting adaptive control and generalized predictive adaptive control, two adaptive decoupling control methods were proposed. In the above literatures, the effectiveness of the proposed decoupling control methods are demonstrated only by numerical simulations. The stability and convergence of the closedloop systems are not provided. To solve this problem, in [15][18], four adaptive decoupling control methods based on multiple models and neural networks were proposed by combining multiple model switching respectively with one stepahead adaptive decoupling control, one stepahead weighting adaptive decoupling control, adaptive generalized predictive decoupling control, and adaptive proportion integration differentiation (PID) decoupling control. The stability and convergence of the closedloop systems were proved. In [19], the adaptive decoupling control method combining an openloop decoupling compensator with neural network approximate compensation was proposed, which can also ensure the stability and convergence of the closedloop system. For complex industrial processes with multivariable, strong coupling, strong nonlinearity and especially variable dynamic characteristics, two intelligent decoupling control methods were proposed by combining multiple model switching respectively with one stepahead optimal weighting decoupling control based on neural network feedforward compensation and approximate dynamic decoupling control based on neural network feedforward compensation [20], [21]. For nonlinear multivariable systems with unknown structures, an adaptive switching control method driven by virtual unmolded dynamics was proposed in [22]. With the development of intelligent control methods, in recent years, by combining intelligent control with inverse systems, the inverse decoupling control methods were proposed in [23][27].
The above methods are designed without considering the performance requirements on the systems. Although they can realize decoupling control of complex industrial processes, they cannot meet the rising performance index. Therefore, if the above decoupling control methods are used, the integrated optimization control of complex industrial processes aimed at saving energy and reducing consumption will be influenced. With the rapid development of science and technology, the performance requirements on many control systems (such as spacecraft, modern industrial equipment, and production process, etc.) become higher and higher. For a control system, it is always hoped that a certain performance index is optimal in some sense. Optimal control is a subject that studies and solves the optimal solution among all possible control schemes, which provides a feasible scheme to realize the optimal performance of a control system.
This paper proposes an optimal decoupling control method for a class of continuous time linear multivariable systems, by combining openloop feedforward decoupling control with optimal tracking control. First, by introducing one virtual input vector, the differential equation on state is converted into the differential equation on output, and then the system matrix and input matrix are decomposed into a diagonal matrix and the matrix with zero diagonal elements, such that the coupling terms are separated. Second, by introducing the other virtual input vector, the system is openloop decoupled by using feedforward and output feedback method. Finally, for the decoupled system, the conventional optimal tracking control method is adopted to realize the tracking of the system to any reference input. By choosing appropriate weighting matrices, the optimal decoupling control method is equivalent to the closedloop optimal decoupling control method. Simulation results show the effectiveness of the proposed method and the superiority as compared with conventional optimal quadratic tracking (LQT) control method.
Ⅱ. PROBLEM DESCRIPTIONSConsider a continuoustime linear multivariable timeinvariant system
$ \begin{align} \dot {x}(t)=Ax(t)+Bu(t) \end{align} $  (1a) 
$ y(t)=Cx(t) $  (1b) 
where
The conventional optimal tracking control problem is to find the optimal control policy
$ \begin{align} J(t)=\frac{1}{2}\int_{0}^{\infty}[{e(t)}^{T}{Qe(t)+u(t)^{T}Ru(t)}]{dt} \end{align} $  (2) 
where
$ \begin{align} u^*(t)=R^{1}B^{T}[Px(t)g_{ss}] \end{align} $  (3) 
where
$ \begin{align} PA+A^{T}PPBR^{1}B^{T}P+C^{T}QC=0 \end{align} $  (4) 
and the limiting function
$ \begin{align} \dot g=[ABR^{1}B^{T}P]^{T}gC^{T}Qy_r, g(T)=0. \end{align} $  (5) 
The optimal output trajectories can be then obtained by computing
$ \dot {x}(t)=[ABR^{1}B^{T}P]x(t)+BR^{1}B^{T}g $  (6a) 
$ y(t)=Cx(t). $  (6b) 
From (5) and (6), the transfer function matrix from
The purpose of this paper is to derive an optimal decoupling control law, so that the system output
To realize inputoutput decoupling control, we should first convert the differential equation on state into the equation on output.
From (1), we know
$ \begin{align} \dot {y}(t)=CAx(t)+CBu(t). \end{align} $  (7) 
Define
$ \begin{align} u(t)=K_0y(t)G_0x(t)+L_0w(t) \end{align} $  (8) 
where
$ \begin{align} \dot {y}(t)=CBK_0y(t)+CBL_0w(t)+[CACBG_0]x(t). \end{align} $  (9) 
Select
$ \begin{align} G_0=(CB)^{1}CA. \end{align} $  (10) 
(9) can then be converted into the following equation:
$ \begin{align} \dot {y}(t)=\bar{A}y(t)+\bar{B}w(t), \quad y(t_0)=y_0=Cx_0 \end{align} $  (11) 
where
Equation (11) can be viewed as a generalized system. In the sequel, it will be considered directly, and an optimal decoupling control
Since the coupling effect of the
$ \begin{align} \dot {y}(t)=\bar{A}_1y(t)+\bar{A}_2y(t)+\bar{B}_1w(t)+\bar{B}_2w(t) \end{align} $  (12) 
where
By introducing the second virtual input vector
$ \begin{align} S_1w(t)+S_2w(t)=Ky(t)+Lv(t) \end{align} $  (13) 
where
$ \begin{align} \dot{y}(t)=&\ \bar{A}_1y(t)+{S_1}^{1}(S_1\bar{A}_2\bar{B}_1K)y(t)\\ &\, +{S_1}^{1}(S_1\bar{B}_2\bar{B}_1S_2)w(t)+{S_1}^{1}\bar{B}_1Lv(t). \end{align} $  (14) 
From (14), to achieve decoupling,
$ S_1\bar{A}_2=\bar{B}_1K $  (15) 
$ S_1\bar{B}_2=\bar{B}_1S_2. $  (16) 
Then, (11) is converted to
$ \begin{align} \dot{y}(t)=\bar{A}_1y(t)+{S_1}^{1}\bar{B}_1Lv(t). \end{align} $  (17) 
From (17), since the matrices
Assumption 1:
In the following, for (17), we will design the virtual input
$ \begin{align} J(t)=\frac{1}{2}\int_{0}^{\infty}[{e(t)}^{T}{Qe(t)+v(t)^{T}Rv(t)}]{dt} \end{align} $  (18) 
where
Let
$ \begin{align} v^*(t)=R^{1}\stackrel{=}{B}^{T}Py(t)+R^{1}\stackrel{=}{B}^{T}g_{ss} \end{align} $  (19) 
where
$ \begin{align} P\stackrel{=}{A}+\stackrel{=}{A}^{T}PP\stackrel{=}{B}R^{1}\stackrel{=}{B}^{T}P+Q=0 \end{align} $  (20) 
and
$ \begin{align} \dot g=[\stackrel{=}{A}^{T}P\stackrel{=}{B}R^{1}\stackrel{=}{B}^{T}]gQy_r, g(T)=0. \end{align} $  (21) 
Then, from (13) and (19), the optimal decoupling control law is finally obtained as follows:
$ \begin{align} w^*(t)=&\, (S_1+S_2)^{1}[K +LR^{1}\stackrel{=}{B}^TP]y(t)\\ &\, +(S_1+S_2)^{1}LR^{1}\stackrel{=}{B}^{T}g_{ss}. \end{align} $  (22) 
The structure of the optimal decoupling control system is shown in Fig. 1.
Download:


Fig. 1 Structure of the optimal decoupling control system. 
According to (15) and (16), if
$ \begin{align} \begin{split} w^*(t)=\bar{B}^{1}[\bar{A}_2 +R^{1}P]y(t) +\bar{B}^{1}R^{1}g_{ss}. \end{split} \end{align} $  (23) 
where the symmetric positive definite constant matrix
$ \begin{align} P\bar{A}_1+\bar{A}_1^{T}PPR^{1}P+Q=0. \end{align} $  (24) 
and
$ \begin{align} \dot g=[\bar{A}_1^{T}PR^{1}]gQy_r, \quad g(T)=0. \end{align} $  (25) 
The optimal decoupling controller (23) is realized by first decoupling the generalized system (11) and then designing the optimal tracking controller. Therefore, it is actually an openloop decoupling controller. In the following, we will show that the openloop decoupling controller equals a closedloop optimal decoupling controller in a sense.
Theorem 1: For the generalized system (11), the optimal decoupling controller (23), with
$ \begin{align} J=&\ \frac{1}{2}\int_{0}^{T}[{e}{(t)^{T}Qe(t)}\\ &\, +(Sw(t)+Gy(t))^{T}R(Sw(t)+Gy(t))]{dt} \end{align} $  (26) 
where
Proof: The optimal control minimizing the performance index (26) must satisfy the minimum principle; thus, the Hamiltonian function is introduced.
$ \begin{align} H=&\ \frac{1}{2}[y_r(t)y(t)]^{T}Q[y_r(t)y(t)]\\ &\, +\frac{1}{2}[Sw(t)+Gy(t)]^{T}R[Sw(t)+Gy(t)]\\ &\, +\lambda(t)^{T}[(\bar{A}y(t)+\bar{B}w(t)]. \end{align} $  (27) 
From the extreme value condition
$ \begin{align} \frac{\partial{H}}{\partial{w}}=S^{T}RSw(t)+S^{T}RGy(t)+\bar{B}^{T}\lambda(t)=0 \end{align} $  (28) 
the optimal control is obtained as
$ \begin{align} w^*(t)=(S^{T}RS)^{1}[S^{T}RGy(t)+\bar{B}^{T}\lambda(t)]. \end{align} $  (29) 
Since
$ \begin{align} \dot y(t)=&\ \frac{\partial{H}}{\partial\lambda}=\bar{A}y(t)+\bar{B}w(t) \\ \dot\lambda(t)=&\ \frac{\partial{H}}{\partial y}=(Q+G^{T}RG)y(t)Qy_r(t)\\ \end{align} $  (30) 
$ +G^{T}RSw(t)+\bar{A}^{T}\lambda(t). $  (31) 
Let
$ \begin{align} \lambda(t)=Py(t)g(t) \end{align} $  (32) 
then, we have
$ \begin{align} \dot y(t)=&\ \frac{\partial{H}}{\partial\lambda}=[\bar{A}\bar{B}(S^{T}RS)^{1}S^{T}RG\\ &\, \bar{B}(S^{T}RS)^{1}\bar{B}^{T}P]y(t)\\ &\, +\bar{B}(S^{T}RS)^{1}\bar{B}^{T}g(t) \end{align} $  (33) 
$\begin{align} \dot \lambda(t)=&\ P\dot y(t)\dot g(t)\\ =&\ [P\bar{A}P\bar{B}(S^{T}RS)^{1}S^{T}RG\\ &\, P\bar{B}(S^{T}RS)^{1}\bar{B}^{T}P]y(t)\\ &\, +P\bar{B}(S^{T}RS)^{1}\bar{B}^{T}g(t)\dot g(t). \end{align} $  (34) 
From (31), we know
$ \begin{align} \dot\lambda(t)=&\ \frac{\partial{H}}{\partial y}=(Q+G^{T}RGG^{T}RS(S^{T}RS)^{1}S^{T}RG\\ &\, +\bar{A}^{T}PG^{T}RS(S^{T}RS)^{1}\bar{B}^{T}P)y(t)\\ &\, +G^{T}RS(S^{T}RS)^{1} \bar{B}^{T}g(t)\bar{A}^{T}g(t)Qy_r(t). \end{align} $  (35) 
By comparing (34) and (35), and making them to hold for any
$\begin{align} P(\bar{A}\, &\bar{B}(S^{T}RS)^{1}S^{T}RG)+(\bar{A}\bar{B}(S^{T}RS)^{1}\\ &\ S^{T}RG)^{T}PP\bar{B}(S^{T}RS)^{1}\bar{B}^{T}P+Q=0 \end{align}$  (36) 
$ \begin{align} \dot g(t)=&\, [\bar{A}\bar{B}(S^{T}RS)^{1}S^{T}RG\\ &\, \bar{B}(S^{T}RS)^{1}\bar{B}^{T}P]^{T}g(t)Qy_r(t). \end{align} $  (37) 
Substituting (32) into (29), the optimal control can be obtained as
$ \begin{align} w^*(t)=&\, (S^{T}RS)^{1}(S^{T}RG+\bar{B}^TP)y(t)\\ &\ (S^{T}RS)^{1}\bar{B}^{T}g(t). \end{align} $  (38) 
Since
$ P\bar{A}_1+\bar{A}_1^{T}PPR^{1}P+Q=0 $  (39) 
$ \dot g(t)=[\bar{A}_1^{T}PR^{1}]g(t)Qy_r(t), g(T)=0 $  (40) 
$ w^*(t)=\bar{B}^{1}(\bar{A}_2+R^{1}P)y(t)+\bar{B}_1^{1}R^{1}g_{ss}. $  (41) 
Remark 1: Since the optimal decoupling controller (23) minimizing the performance index (26) is derived by using one step, it is also a closedloop decoupling scheme.
Ⅳ. SIMULATIONSTo illustrate the effectiveness of the proposed method, in this section, we will apply it to a ball mill coalpulverizing system.
As described in [21], ball mill coalpulverizing systems are important heatpower equipment in power plants. They are used to pulverize raw coal into fine powder of desired temperature and fineness. There are plenty of coal mines in China, but the quality of raw coals varies greatly. Therefore, ball mill coalpulverizing systems are widely used to grind various raw coal. The flowchart of ball mill coalpulverizing system is shown in Fig. 2, which consists of a hopper, a feeder, a strap transmission system, a ball mill, a coarse powder separator, a fine powder separator and a blast system [21]. The raw coal is leaked from the hopper to the feeder, sent into the dryness pipeline by the strap transmission system, and then blended with dryer air. The quantity of dryer air and heat can be regulated by the flow rates of hot air and warm air in the blast system. The mixture of raw coal and dryer air are sent into the ball mill through the dryness pipeline, where it is pulverized to fine powder by knocking and grinding of iron balls when the ball mill rotates. At the same time, the coal powder is dried and brought out of the ball mill by dryer air. Then, it is transferred into the coarse separator, where the coarse powder and fine powder are separated. The unqualified coarse powder is returned into the ball mill for regrinding, while qualified fine powder is sent into the bunker and then to boiler for burning.
Download:


Fig. 2 Flowchart of the ball mill coalpulverizing system. 
According to the energy and mass equilibriums of ball mill entrances and the mass equilibrium of liquid in the blast pipes, a dynamic model of the ball mill coalpulverizing system is established as follows [21]:
$ \begin{align} \dot x_1(t)=&\ \frac{1}{C_{gq}w_{gq}+C_m\omega_m(t)}\times \bigg\{[\frac{C_{gm}t_c}{3.6}\frac{C_mx_1}{3.6}+11.5\\ &\, \Delta w(595+0.45x_1t_c)]u_1+[C_rt_rC_{tf}x_1\\ &\, +k_{lf}(C_{lk}t_{lk}C_{tf}x_1)]u_2+[C_wt_wC_{tf}x_1\\ &\, +k_{lf}(C_{lk}t_{lk}C_{tf}x_1)]u_3\bigg\} \end{align} $  (42a) 
$ \begin{align} \dot x_2(t)=&\ \frac{RT}{V_1}[\Delta wu_1+(1+k_{lf})u_2+(1+k_{lf})u_3\\ &\, \sqrt{\frac{x_2x_3+P_0}{R_1}}] \end{align} $  (42b) 
$ \begin{align} \dot x_3(t)=&\ 3(1+0.8\mu)\frac{w_{thr}^2}{2\left(V_q\frac{w_{gq}} {\rho_{gq}}\frac{\omega}{\rho_m}\right)}\\ &\, \times[\Delta wu_1+(1+k_{lf})u_2+(1+k_{lf})u_3 \\&\, \sqrt{\frac{x_2x_3+P_0}{R_1}}] \end{align} $  (42c) 
$ y(t)=\ x(t)=[x_1, x_2, x_3]^T $  (42d) 
where
From (42), the ball mill coalpulverizing system is a nonlinear process with strong couplings between each loop. Variations of each control input will cause the changes of all the system outputs. According to the observation on site, the strong coupling property is mainly reflected in following cases:
1) When the coal feeding rate,
2) When the hot air flow,
3) When the warm air flow,
Therefore, it is very important to decouple the control of the ball mill coalpulverizing system.
According to the requirements of the practical process, the outputs
For
$ \begin{align*} \begin{aligned} &A=\left[\begin{array}{ccc} 0.8548~&0~&0\\0~&0.1694~&0.1694\\3.9075~&2.0444~&1.2569 \end{array}\right]\\ &B=\left[\begin{array}{ccc} 0.1214~&0.79~&0.302\\0.2639~&4.004~&4.004\\0.6232~&17.2977~&6.4875 \end{array}\right]. \end{aligned} \end{align*} $ 
The relative gain matrix is as follows:
$ \begin{align*} \begin{aligned} \rm RGA=\left[\begin{array}{ccc} 0.6417~&0.4096~&0.0513\\0.5865~&2.0054~&1.5919\\0.2282~&1.4150~&2.6433 \end{array}\right]. \end{aligned} \end{align*} $ 
From the RGA matrix, we know that it is reasonable to select the coal feeding rate
$ \begin{align*} \begin{aligned} Q=\left[\begin{array}{ccc} 1~&0~&0\\0~&1~&0\\0~&0~&1 \end{array}\right], \quad R=\left[\begin{array}{ccc} 0.01~&0~&0\\0~&0.01~&0\\0~&0~&0.001 \end{array}\right] \end{aligned} \end{align*} $ 
and the reference input as
$ \begin{align*} y_r(t)=\left\{\begin{aligned} &y_{r1}(t)=\begin{cases} 70, & 0\leq t < 7\\84, &7\leq t < 30\end{cases} \\ &y_{r2}(t)=\begin{cases} 0.2, & 0\leq t < 14\\0.6, &14\leq t < 30\end{cases} \\ &y_{r3}(t)=\begin{cases} 0.8, & 0\leq t < 21\\1.6, &21\leq t < 30.\end{cases} \end{aligned}\right.\ \end{align*} $ 
Fig. 3 is the output tracking curves by using the proposed optimal decoupling control method, and Fig. 4 is the corresponding control input curves. From Fig. 3, the three output variables can all track their reference curves well. In order to compare with the proposed method, the conventional LQT control method is also used. Fig. 5 is the corresponding output tracking curves, and Fig. 6 is the control input curves. From Fig. 5, we can see that due to the existence of the couplings, the output
Download:


Fig. 3 Tracking results by using the proposed controller. 
Download:


Fig. 4 Input signals by using the proposed controller. 
Download:


Fig. 5 Tracking results by using the LQT controller. 
Download:


Fig. 6 Input signals by using the LQT controller. 
For
$ \begin{align*} &A=\left[\begin{array}{ccc} 1.0124~&0~&0\\0~&1.4291~&0.3426\\1.1776~&3.2199~&2.9230 \end{array}\right]\\ &B=\left[\begin{array}{ccc} 0.9339~&0.3431~&0.5119\\0.5231~&4.7891~&1.8023\\0.6633~&3.4521~&0.2324 \end{array}\right]. \end{align*} $ 
The relative gain matrix is as follows:
$ \begin{align*} \begin{aligned} \rm RGA=\left[\begin{array}{ccc} 0.6171~&0.0.574~&0.3256\\0.3439~&3.0544~&1.7105\\0.7268~&2.1118~&2.3850 \end{array}\right]. \end{aligned} \end{align*} $ 
From the RGA matrix, we know that there still exists serious coupling in the loops. Select the same weighting matrices as in the above example and the reference input as below
$ \begin{align*} y_r(t)=\begin{cases} y_{r1}(t)=\begin{cases} 71, & t\in [0, 3], [6, 9], [12, 15]\\83, &t\in [3, 6], [9, 12], [15, 18]\end{cases} \\ y_{r2}(t)=\begin{cases} 0.2, & t\in [0, 2], [5, 8], [11, 14], [17, 18]\\0.6, &t\in [2, 5], [8, 11], [14, 17]\end{cases} \\ y_{r3}(t)=\begin{cases} 0.8, & t\in [0, 1], [4, 7], [10, 13], [16, 18]\\ 1.6, &t\in [1, 4], [7, 10], [13, 16].\end{cases} \end{cases} \end{align*} $ 
Fig. 7 is the output tracking curves by using the proposed control method, Fig. 8 is the corresponding control input curves, Fig. 9 is the output tracking curves by using the conventional LQT control method, and Fig. 10 is the corresponding control input curves. From Figs. 7 and 9, we can see that the tracking effect produced by using the proposed method is better than that produced by the LQT method.
Download:


Fig. 7 Tracking results by using the proposed controller. 
Download:


Fig. 8 Input signals by using the proposed controller. 
Download:


Fig. 9 Tracking results by using the LQT controller. 
Download:


Fig. 10 Input signals by using the LQT controller. 
To verify the effectiveness of the proposed method for the ball mill coalpulverizing system, in the following, we will conduct simulations on the ball mill coalpulverizing system (42) around a small neighbourhood of the point
The weighting matrices are selected as in the above two examples, and the reference input is as
$\begin{align*} y_r(t)=\begin{cases} y_{r1}(t)=\begin{cases} 80, & t\in [0, 3], [6, 9], [12, 15]\\84, &t\in [3, 6], [9, 12], [15, 18]\end{cases} \\ y_{r2}(t)=\begin{cases} 0.4, & t\in [0, 2], [5, 8], [11, 14], [17, 18] \\0.6, &t\in [2, 5], [8, 11], [14, 17]\end{cases} \\ y_{r3}(t)=\begin{cases} 1.5, & t\in [0, 1], [4, 7], [10, 13], [16, 18] \\ 1.9, &t\in [1, 4], [7, 10], [13, 16].\end{cases} \end{cases} \end{align*} $ 
Since the nonlinearity is weak around the small neighbourhood of the equilibrium/operational point, the linear decoupling control method can be used to control the nonlinear system. Fig. 11 shows the output tracking curves by using the proposed optimal decoupling control method, from which we can see the three output variables can all track their reference curves well. In order to compare with the proposed method, the nonlinear decoupling control method [21] is also used. The sampling time
Download:


Fig. 11 Tracking results by using the optimal decoupling control in the ball mill. 
Download:


Fig. 12 Tracking results by using the nonlinear decoupling control [21] in the ball mill. 
In this paper, for a class of linear multivariable systems with strong couplings, an optimal decoupling control method is proposed by combining an openloop decoupling compensator and an optimal tracking controller. From the design method point of view, this method belongs to the openloop decoupling domain. However, it is also proved to be equivalent to the closedloop decoupling control in a sense. Therefore, the method can not only eliminate the coupling effects in the loops, but also can improve the closedloop performance index, and can realize the tracking of any reference input of the system. Two linear models are obtained by linearizing a ball mill coalpulverizing system at two equilibium/operational points. Simulations are conducted respectively for the two models. The results show the effectiveness and superiority of the proposed method as compared with the conventional LQT control method. Simulations are also conducted on the ball mill coalpulverizing system. The results show that around a small neighbourhood of the equilibium/operational point, the effect produced by the proposed linear decoupling control method is not worse than that produced by the nonlinear decoupling control method [21].
The proposed optimal decoupling control method is designed for linear systems with completely known dynamics. Although it is also effective for nonlinear systems with weak nonlinearity, it cannot be directly applied to a nonlinear system with strong nonlinearity. In the future, we will study adaptive optimal decoupling control methods for nonlinear systems with strong nonlinearity and unknown dynamics.
[1]  C. L. Smith, "Automatic process control: where are we?, " Acta Metall., vol. 28, pp. 501506, Jan. 1973. https://www.researchgate.net/publication/296612899_AUTOMATIC_PROCESS_CONTROL_WHERE_ARE_WE 
[2]  P. L. Falb and W. A. Wolovich, "Decoupling in the design and synthesis of multivariable control systems, " IEEE Trans. Autom. Control, vol. 12, no. 6, pp. 651669, Dec. 1967. http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=1098737 
[3]  W. A. Wonham, Linear Multivariable Control: A Geometric Approach. New York: Springer, 1979: 10001001. 
[4]  V. Eldem, "The solution of diagonal decoupling problem by dynamic output feedback and constant precompensator: the general case, " IEEE Trans. Autom. Control, vol. 39, no. 3, pp. 503511, Mar. 1994. http://www.researchgate.net/publication/3021773_The_solution_of_diagonal_decoupling_problem_by_dynamic_outputfeedback_and_constant_precompensator_the_general_case 
[5]  H. H. Rosenbrock, "Design of multivariable control systems using the inverse Nyquist array, " Proc. Institut. Electric. Eng., vol. 116, no. 11, pp. 19291936, Nov. 1969. http://www.researchgate.net/publication/239425919_Design_of_multivariable_control_systems_using_the_inverse_Nyquist_array 
[6]  E. H. Bristol, "On a new measure of interaction for multivariable process control, " IEEE Trans. Autom. Control, vol. 11, no. 1, pp. 133134, Jan. 1966. http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=1098266 
[7]  M. E. Salgado and A. Conley, "MIMO interaction measure and controller structure selection, " Int. J. Control, vol. 77, no. 4, pp. 367383, Jan. 2004. https://www.researchgate.net/publication/238181468_MIMO_interaction_measure_and_controller_structure_selection?ev=prf_cit 
[8]  P. Nordfeldt and T. Hagglund, "Decoupler and PID controller design of TITO systems, " J. Process Control, vol. 16, no. 9, pp. 923936, Oct. 2006. http://www.sciencedirect.com/science/article/pii/S095915240600062X 
[9]  T. Y. Chai and G. Wang, "Globally convergent multivariable adaptive decoupling controller and its application to a binary distillation column". Int. J. Control , vol.55, no.2, pp.415–429, 1992. 
[10]  T. Y. Chai, "A selftuning decoupling controller for a class of multivariable systems and global convergence analysis, " IEEE Trans. Autom. Control, vol. 33, no. 8, pp. 767771, Aug. 1988. https://www.researchgate.net/publication/265671320_A_selftuning_decoupling_controller_for_a_class_of_multivariable_systems 
[11]  L. M. Wu and T. Y. Chai, "A neural network decoupling strategy for a class of nonlinear discrete time systems, " Acta Autom. Sinica, vol. 23, no. 2, pp. 207212, Feb. 1997. http://www.aas.net.cn/EN/Y1997/V23/I02/207 
[12]  H. Yue and T. Y. Chai, "Adaptive decoupling control of multivariable nonlinear nonminimum phase systems using neural networks, " in Proc. 1998 American Control Conf., Philadelphia, PA, USA, pp. 513514. http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=694720 
[13]  K. Y. Zhu, X. F. Qin, and T. Y. Chai, "A new decoupling design of selftuning multivariable generalized predictive control, " Int. J. Adapt. Control Signal Process., vol. 13, no. 3, pp. 183196, May 1999. http://www.researchgate.net/publication/229531972_A_new_decoupling_design_of_selftuning_multivariable_generalized_predictive_control 
[14]  Q. M. Zhu, Z. Ma, and K. Warwick, "Neural network enhanced generalised minimum variance selftuning controller for nonlinear discretetime systems, " IEEE Proc. Control Theory Appl., vol. 146, no. 4, pp. 319326, Jul. 1999. http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=795864 
[15]  Y. Fu and T. Y. Chai, "Nonlinear multivariable adaptive control using multiple models and neural networks, " Automatica, vol. 43, no. 6, pp. 11011110, Jun. 2007. https://www.sciencedirect.com/science/article/pii/S0005109807000647 
[16]  Y. Fu and T. Y. Chai, "Nonlinear indirect adaptive decoupling control based on neural networks and multiple models, " in Proc. 6th World Congress on Intelligent Control and Automation, Dalian, China, 2006. http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=1657292 
[17]  Y. J. Shi and T. Y. Chai, "Nonlinear adaptive decoupling generalized predictive control using neural networks and multiple models". Control Theory Appl. , vol.25, no.4, pp.634–640, 2008. 
[18]  L. F. Zhai and T. Y. Chai, "Nonlinear decoupling PID control using neural networks and multiple models, " J. Control Theory Appl., vol. 23, no. 1, pp. 6269, Feb. 2006. http://kns.cnki.net/KCMS/detail/detail.aspx?filename=KZLL200601010&dbname=CJFD&dbcode=CJFQ 
[19]  Y. Fu and T. Y. Chai, "Neuralnetworkbased nonlinear adaptive dynamical decoupling control, " IEEE Trans. Neural Netw., vol. 18, no. 3, pp. 921925, May 2007. http://dl.acm.org/citation.cfm?id=2327864 
[20]  Y. Fu and T. Y. Chai, "Intelligent decoupling control of nonlinear multivariable systems and its application to a wind tunnel system, " IEEE Trans. Control Syst. Technol., vol. 17, no. 6, pp. 13761384, Nov. 2009. http://www.researchgate.net/publication/224402021_Intelligent_Decoupling_Control_of_Nonlinear_Multivariable_Systems_and_its_Application_to_a_Wind_Tunnel_System 
[21]  T. Y. Chai, L. F. Zhai, and H. Yue, "Multiple models and neural networks based decoupling control of ball mill coalpulverizing systems, " J. Process Control, vol. 21, no. 3, pp. 351366, Mar. 2011. http://www.sciencedirect.com/science/article/pii/S0959152410002258 
[22]  T. Y. Chai, Y. J. Zhang, H. Wang, C. Y. Su, and J. Sun, "Databased virtual unmodeled dynamics driven multivariable nonlinear adaptive switching control, " IEEE Trans. Neural Netw., vol. 22, no. 12, pp. 2154 2172, Dec. 2011. http://dl.acm.org/citation.cfm?id=2332941 
[23]  J. C. Fang and Y. Ren, "Decoupling control of magnetically suspended rotor system in control moment gyros based on an inverse system method, " IEEE/ASME Trans. Mech., vol. 17, no. 6, pp. 11331144, Dec. 2012. http://www.researchgate.net/publication/260330835_Decoupling_Control_of_Magnetically_Suspended_Rotor_System_in_Control_Moment_Gyros_Based_on_an_Inverse_System_Method 
[24]  Q. Li, G. H. Liu, H. Zhang, and Y. Jiang, "Decoupling control for fivephase faulttolerant permanentmagnet motor by using SVM inverse system method, " in Proc. 2014 Int. Joint Conf. Neural Networks, Beijing, pp. 38373840. http://www.en.cnki.com.cn/Article_en/CJFDTotalHDZJ201604038.htm 
[25]  M. Hamdy, A. Ramadan, and B. Abozalam, "A novel inverted fuzzy decoupling scheme for MIMO systems with disturbance: a case study of binary distillation column". J. Intell. Manuf. , 2016. DOI:10.1007/s108450161218x 
[26]  M. Hamdy, A. Ramadan, and B. Abozalam, "Comparative study of different decoupling schemes for TITO binary distillation column via PI controller". IEEE/CAA J. of Autom. Sinica , vol.5, no.4, pp.869–877, 2018. DOI:10.1109/JAS.2016.7510040 
[27]  M. Hamdy and A. Ramadan, "Design of smith predictor and fuzzy decoupling for MIMO chemical processes with time delays, " Asian J. Control, vol. 19, no. 1, pp. 5766, Jan. 2017. http://dl.acm.org/citation.cfm?id=3078412 
[28]  H. Modares and F. L. Lewis, "Linear quadratic tracking control of partiallyunknown continuoustime systems using reinforcement learning, " IEEE Trans. Autom. Control, vol. 59, no. 11, pp. 30513056, Nov. 2014. https://ieeexplore.ieee.org/document/6787009/ 
[29]  Y. Bai, "Application of neural decoupling control in the intermediate storage bin of ball mill, " Ph. D. dissertation, Northeastern University, 1998. 