Fractional-Order Predictive Functional Control of Industrial Processes with Partial Actuator Failures

Guangdong Zhicheng Champion Group Co., Ltd., Dongguan 523718, China National-Local Joint Engineering Laboratory of Digitalize Electrical Design Technology, Wenzhou University, Wenzhou 325035, China Institute of Cyber Systems and Control, Zhejiang University, Hangzhou 310018, China College of Electrical and Information Engineering, Hunan University, Changsha 410082, China College of Cyber Security, Jinan University, Guangzhou 510632, China


Introduction
In industrial processes, actuators play an important role in the industrial control system because an actuator links the controller output to the physical actions and determines the quality of products [1,2]. However, the actuator faces difficulties in executing the controller demand completely due to the physical malfunction, e.g., friction and saturation. Specially, in the control system of industrial process, an actuator fault often exists, and the control performance may be degraded caused by discrepancies between the desired actuator position and the actual position.
Generally, the actuator faults in industrial processes can be separated into three categories, i.e., the actuator outage, the actuator stuck, and the partial actuator failure [3]. As for the first two categories, it is impossible to improve any controllers' performance because the control process under these two categories is totally uncontrollable.
us, this study focuses on the third category, i.e., the partial actuator failure, which can be tackled to a certain extent by using the adequate control scheme. As a result, many related research works have been carried out. For example, Wang et al. [4] put forward an iterative learning control (ILC) scheme for batch processes under partial actuator faults according to a particular 2D Fornasini-Marchsini model. Giridhar and El-Farra [5] proposed a robust detection, isolation, and compensation of control actuator faults based on the framework of feedback robust control. Jin et al. [6] proposed an improved ILC scheme to control the nonlinear constrained system with actuator failures. Ding et al. [7] proposed a novel ILC scheme to control the uncertain multiple-input multiple-output discrete system under actuator faults. In [8], a model predictive control method was suggested for the injection molding batch process with partial actuator failures. Due to the uncertainties of actuator fault, the control system design is often mismatched [1,9]. To deal with this challenge problem, ILC [10][11][12] has been developed as one of the most popular strategies for different industrial processes. However, as discussed in [3], the ILC is largely dependent on the repetitive nature of such processes, whose performance improvement is confined by this unsuitable assumption because many real-world processes are actually time-varying and nonrepetitive and suffering from persistent disturbance.
us, the design of the ILC is not fit for industrial processes in practice.
In the past decades, model predictive control (MPC) has shown the potential ability of control design for industrial processes [13][14][15]. As a recently developed MPC, state-space predictive functional control (PFC) [3,16,17] provides a novel insight into control design for the industrial processes against partial actuator failures because it not only has theoretical basis for the control design but also has advantages in hardware implementation, computational capability, and control accuracy. For example, in [3], Tao et al. proposed a PFC method based on a linear quadratic structure for industrial process under actuator failures and highlighted that its performance is better than traditional state-space PFC. In [16], an improved version of PFC was applied into the control design for an injection modeling batch process with a partial actuator. However, this improved PFC's weighting factors on the process state and output tracking error are determined through experience. To alleviate this deficiency, a genetic algorithm-(GA-) based PFC was proposed, where GA was used to tune its weighting factors, and six cases of partial actuator failures were used to demonstrate the performance of GAbased PFC [17]. Besides, Hu et al. [18] combined PFC with GA and linear quadratic structure for industrial processes against the partial actuator failures. Although a lot of good results have been obtained by PFC strategies, the framework of PFC design for industrial control processes still needs to be further explored for achieving high-quality control performance.
Additionally, with the deep study on mathematical fractional-order calculus, the applications of fractional-order controller have been attempted by many researchers. In [19,20], the PFC based on fractional-order calculus was proposed to control the fractional model with model mismatches. Sanatizadeh and Bigdeli [21] designed the fractional-order predictive functional controller for unstable systems with time delay. In [22], the authors have successfully applied the fractional-order PFC into industrial heating furnace, and the experimental results on the temperature process showed the improvement of the fractionalorder PFC. In all aforementioned examples, fractional-order methods have shown better performance than the corresponding traditional integer-order methods. In summary, the fractional-order calculus has a potential ability to improve the performance of traditional integer-order methods.
is is one of primary motivations to incorporate fractionalorder mechanism into PFC and propose a fractional-order PFC called FOPFC for an industrial process against partial actuator faults.
Unfortunately, the fractional-order calculus also involves more tuning parameters than the integer version. In other words, tuning the related parameters in FOPFC has more difficulty than PFC. In FOPFC, there are three key parameters called fractional-order parameter c, smoothing factor λ, and prediction horizon P, which have important impacts on the performance of control system. More specifically, as discussed in [19,20], λ plays the role in smoothing reference trajectory. c and P have influence on the rapidity and stability of the system response. On the other hand, as tuning these parameters lacks analytical knowledge, the selection of these parameters in [19,20] is generally based on the trial and error method. In order to alleviate this problem, in this paper, a competitive evolutionary algorithm is employed to optimize the related parameters in FOPFC for improving the closed-loop performance. As an efficient evolutionary algorithm, population-based extremal optimization (PEO) [23] is extended from extremal optimization (EO) [24] and has shown great promising ability in a variety of fields, such as numerical optimization problems including single-objective and multiobjective problems [23,25], PID/FOPID controllers designing problems [26,27], and weighting optimization of ensemble learning [28]. To be more precise, in [25], an improved multiobjective PEO was presented for solving multiobjective problems. In [26], multivariable PID controllers were designed using real-coded PEO. In [27], multi-non-uniform mutation-based PEO was applied to designing the FOPID controllers. Abovementioned examples have revealed that PEOs have outstanding superiority over other popular evolutionary algorithms including GA and particle swarm optimization (PSO). As discussed in [23], the authors have done extensive experiments on some benchmark single-objective optimization problems. e results in [23] demonstrate the PEO performs better than or at least competitive with many reported popular single-objective evolutionary algorithms. From the design perspective of PEO algorithm, the PEO used in this paper is relatively simper than other evolutionary algorithms including GA and PSO due to its fewer adjustable parameters and only mutation operation. us, the PEO is used in this work. It is worthy to be mentioned that compared with the previous studies [17,18], the proposed control method extends the integer-order PFC to the fractional version and uses PEO algorithm to tune the main parameters in FOPFC, which is superior to the GA used in [17,18]. In addition, the existing work [20] designed the parameter FOPFC by the trial-and-error method, while this paper uses the PEO algorithm to tune the related parameters in FOPFC and applies FOPFC to solving the industrial processes under partial actuator failures. To the best of our knowledge, there exist no reported related works focusing on PEO for tuning parameters in FOPFC. erefore, in order to deal with this problem, the PEO-based FOPFC algorithm called PEO-FOPFC is proposed in this paper by adopting PEO to search for the adjustable parameters in FOPFC.

Complexity
To the best of our knowledge, this work is the first contribution to optimize the analytical parameters including the prediction horizon, fractional-order parameter, and smoothing factor in a FOPFC controller for industrial processes with partial actuator failures by means of PEO. To be more specific, the principal contributions of this paper are summarized as follows: (1) A fractional-order predictive functional control (FOPFC) strategy is firstly proposed for the industrial process with partial actuator failures. (2) Encountered the difficulties in tuning-related parameters in FOPFC strategy due to the increasing adjustable parameters and lacking analytical knowledge, and the population extremal optimization is introduced into FOPFC to search for the adjustable parameters such as the prediction horizon, fractionalorder parameter, and smoothing factor. (3) e effectiveness of the PEO-FOPFC strategy is demonstrated on two industrial processes, e.g., injection modeling batch process and process flow of coke furnace under six cases including constant faults, time-varying faults, and nonrepetitive unknown disturbance. Moreover, the simulation results show that the performance of proposed PEO-FOPFC is much better than the recently developed PFC [16]. (4) e performance of fractional-order strategy is illustrated by the comparison of FOPFC with PFC. In addition, the performance of PEO algorithm is purely verified by comparison PEO-FOPFC with FOPFC and two other popular evolutionary algorithms including GA-and PSO-based FOPFC methods on an injection modeling batch process. e remainder of this paper is given as follows. Section 2 presents preliminaries concerning fractional-order calculus, basics of EO, and problem formulation of the process. en, the proposed PEO-FOPFC is described in Section 3. e comprehensive experimental results of two industrial processes are discussed in Sections 4 and 5, respectively. Finally, Section 6 concludes the paper and gives future works.

Preliminaries
In this section, a brief overview of fractional-order calculus and canonical EO algorithm is given. en, the description of a single-input single-output industrial process is presented.

Fractional-Order Calculus.
ere are three common definitions of fractional-order calculus called Grünwald-Letnikov (GL) definition, Riemann-Liouville (RL) definition, and Caputo definition [29]. e RL form with order α is given as where f(t) means the function, [b, t] means the interval of f(t), α > 0 is the fractional-order with n − 1 < α < n, and Γ represents the gamma function.
e Caputo form with order α is given as where f(t) means the function, [b, t] means the interval of f(t), α > 0 is the fractional-order with n − 1 < α < n, and Γ represents the gamma function. e GL definition can be described as where β is the initial time, h is the calculation step, [x] means the integer part of x, c is the fractional-order parameter, and are the polynomial coefficients and can be obtained as follows: In addition, h can be substituted by the sample time T s , when considering the practical process and the characteristics of fractional order.
As suggested in [20], this paper employs GL definition to derive the discrete form of the control system. en, the discretized model of fractional-order integer operator can be described as follows [20,30]: where  [3,16], the linear deviation model can be used to describe the nonlinear industrial process. For simplicity, a single-input single-output (SISO) process is used in this study. And, the corresponding process can be obtained through linearization as follows: where k and d denote the current time and process time delay, respectively, and x(k), y(k), u(k), and ω(k) represent the process state, output, input, and unknown measurement noise, respectively. A I , B I , and C I are the system matrices with appropriate sizes. Here, the term u F (k) is the failed signal from the actuator. en, the failure model can be derived as follows: Afterwards, the version of process under actuator failures can be described as

Proposed Control Strategy
In this section, we firstly introduce the design of fractionalorder predictive functional controller (FOPFC) in Section 3.1. en, we present the control strategy of the proposed PEO-based FOPFC (PEO-FOPFC) in Section 3.2.

Design of FOPFC.
On the basis of equation (7) in Section 2.3, the state vector can be constructed as follows [17]: where the Δ denotes the difference operator. Afterwards, the new state space model can be obtained as where Note that the 0 is full vector with zero elements with the appropriate dimensions. e output tracking error is described as follows: where r(k) denotes the reference trajectory. And, r(k + i) � λ i y(k) + (1 − λ) i c(k), λ denotes the smoothing factor and c(k) is the set point. Combining equations (11) and (13), the dynamic output error can be derived as follows: where Δr(k + 1) means the differenced value of set-point at time k + 1. By adding the tracking error to the state variable, the extended state vector can be obtained as follows: en, the new model is obtained as follows: . 0 is the zero vector with appropriate sizes. As mentioned in [20], the cost function of integer PFC is chosen as follows: where P is the prediction horizon. And, the diagonal matrix Q j is often used as the weighting factor to give a specific value (1) Randomly initialize configuration S I . Calculate the global fitness of S I termed as C(S I ) (2) Set S best � SI and C(S best ) � C(SI), where S best donates the best solution found so far (3) Repeat (4) Evaluate the local fitness of each component in the current configuration S I (5) Obtain a permutation according to the assigned local fitness (6) Employ mutation operation on worst component and a new configuration is generated (7) Replace the worst component with the new configuration, unconditionally (8) Update the S best and C(S best ) (9) Until some criterion is satisfied; (10) Return S best and C(S best ) ALGORITHM 1: Canonical EO. 4 Complexity for each state variable in z(k + j). As discussed in [31,32], matrix Q j plays an important role in control performance of the predictive controller. erefore, in [31,32], GA or EO was employed to tune Q j and showed better performance than the trial and error method. In fact, equation (17) can be viewed as the continuous is the continuous integer interval. Because the fractional order is derived from the integer-order and it has been demonstrated to provide better control performance than integer-order in various domains, a natural idea is to replace integer-order with fractional order and test whether the performance can be improved. us, the cost function of FOPFC can be derived as follows [20]: l is the number of state variables in z(k), and c is the fractional-order factor. e PFC control action is based on the base functions [3,16]: where η j is the coefficient, f j (i) denotes the base function, and N is the number of base functions. (19) can be rewritten as follows: From equation (16), together with equation (20), we can obtain Complexity en, the following equation can be obtained: In addition, denote the Z � z(k + 1), z(k + 2), . . . , z(k + P)] T and ΔR � Δr(k + 1), Δr(k + 2), . . . , Δr(k+ P)] T .
We obtain where en, equation (18) can be rewritten as where Λ � block diag{Λ 1 , Λ 2 , . . ., Λ P }. Afterwards, the optimal control law can be obtained by finding the minimum value of equation (25): At last, the control signal u(k) is then derived as follows: In the realistic industrial process, there exist system uncertainties. us, a robust stability condition is needed for the closed-loop control system to ensure a stable system. In [16] and [33], the authors have given the robust stability condition of state space predictive controller. Here, extended from [16], we give a robust stability condition for the proposed FOPFC, which is described below. Theorem 1. For the industrial process considering unknown partial actuator failures, i.e., description in equation (9), if the FOPFC is designed based on the model equation (7) such that the following condition holds: where σ max (κ) is the maximum singular value ofκ, λ min (κ) and λ max (κ) are the minimum and maximum eigenvalues ofκ, respectively, and M P and W P represent the symmetric positive matrices subject to the following equation: where en, the proposed controller is the robust stability for the considered system. e proof is presented in Proof of eorem 1 in Appendix. e error tracking and constant disturbances rejecting the performance of PFC have been given in [34]. And, we can examine error estimates and sensitivity to disturbances of the proposed controller extended from [34] as follows.

Proposition 1.
If the process is treated in the form of equation (16) and the subsequent FOPFC law is designed as equation (20), then the proposed FOPFC control law tracks the constant set-point without steady error and for the constant input disturbances and output disturbances and the FOPFC can reject with no steady error. e proof is presented in Proof of Proposition 1 in Appendix.

Remark 1.
As suggested in [1,3,16], to facilitate the controller design, the process model is based on nominal state space model and the noise is not considered. In the simulation part, the noise is not ignored. To consider the ω(k) in design controller, one can use system identification technique [10].

PEO-Based FOPFC Control Strategy.
ese are two key strategies in the PEO-FOPFC method. One is that the fractional -order mechanism is applied into the cost function of PFC technique, which is presented in Section 3.1. e other is that the PEO algorithm is used to search the adjustable parameters in FOPFC. In the evolutionary algorithm, the fitness function plays an important role in searching the optimal parameters. us, we firstly define the fitness function used in the process of evolution and then describe the specific steps of the PEO-FOPFC.

Fitness Function.
In [17], the combination of overshoot and rise time was used as the fitness function to tune the weighting factors in PFC, while the integral of time weighted absolute error (ITAE) was adopted as fitness definition in [31]. As discussed in [26], a more reasonable performance index has been proposed, which considers not only integral of absolute error (IAE) but also overshoot, steady-state error, rise time, settling time, square of the input signal, and output signal. Also, its superiority to IAE and ITAE has been demonstrated on multivariable PID controllers. us, in this paper, we use the following equation (31) as fitness function for the industrial process under partial actuator faults. As seen from equation (31), this fitness function not only considers IAE, rise time, and settling time but also the square of input signal to avoid exporting a large control value and Δy to avoid getting a large overshoot: if Δy(t) ≥ 0, where t r , t s , e(t), u(t), and y(t) are rise time, settling time, system error, input signal, and output signal at the time t. w 1 , w 2 , and w 3 are weight coefficients. As suggested in [26], here w 1 , w 2 , w 3 , and w 4 are set as w 1 � 50, w 2 � 0.999, w 3 � 0.001, and w 4 � 100.
Remark 2. w 1 , w 2 , w 3 , and w 4 are the weight coefficients, which have a large influence on the control performance. From equation (31), one can see that w 1 controls the rise time and settling time of system response. e parameters w 2 and w 3 have effect on the system error and input signal, and w 4 has impact on the overshoot of system response. In real-life engineering, these weight coefficients are often determined via the experiential rules and trial-and-error method according to priority of performance indices. In general, w 2 and w 3 are subject to the equation, i.e., w 2 + w 3 � 1 and w 2 is often set much larger than w 3 . us, w 2 and w 3 are set as 0.999 and 0.001, respectively. In addition, w 4 is generally set as 50 ≤ w 4 ≤ 100 or a larger value; here, w 4 � 100 is used. After determining w 2 , w 3 , and w 4 , the parameter w 1 is determined by the trial-and-error method and set as 50. e weight coefficients, i.e., w 1 , w 2 , w 3 , and w 4 are not the optimal values in this paper. In fact, how to obtain more appropriate weight coefficients is still worth studying.

Main Description of PEO-FOPFC.
e main parameters P, c, and λ in FOPFC are optimized by the PEO algorithm. e flowchart of PEO-FOPFC strategy is shown in Figure 1. e detailed steps of optimizing three parameters in FOPFC by the PEO algorithm are described as follows.
Input: the system model, PEO's adjustable parameters including the population size N P , the maximum number of iterations I max , and mutation parameter b, number of base functions N, and upper and lower values of parameters to be optimized, i.e., P, c, and λ Complexity 7 Output: the best solution S best (i.e., the optimal parameters P, c, and λ used in FOPFC) and the corresponding fitness value C best Step 1. e parameters to be optimized are encoded into a solution S in the PEO algorithm given in Figure 2. More specifically, one initial population P I = {S 1 , S 2 , . . ., S NP } contains generated solutions with NP size, where each solution S i � [P i , c i , λ i ] denotes one group parameters used in FOPFC. en, set P � P I , S best � PI be (PI be is the best solutions in P I based on the fitness value defined in equation (27)), and C best � F(S best ).
Step 2 (for each solution S i in P).
(a) Obtain the D mutated solutions {S ik, (k � 1, 2, 3, . . ., D)} by application of multi-non-uniform mutation (MNUM) operation [35] shown in equations (32) and (33). More specially, the j-th component is mutated by MNUM operation and the other components remain unchanged, and then fitness value for each mutated solution is obtained by calculating the fitness function defined in equation (31). For example, for solution , and S 13 � [P 1 , c 1 ,λ 1 ′ ] can be obtained by using MNUM operation. And, the process of mutation operation in PEO is shown in Figure 3: where the subscript j is the j-th decision variable, x is the decision variable, U and L are the upper and lower values of decision variable, t is the current number of iteration, r and r 1 are uniformly distributed random values between 0 and 1, and b is the mutation parameter. (c) Select the best solution among S ik according to the rank index and term it as S bi . Additionally, the corresponding best fitness value is termed as C bi . en save S bi and C bi in P b and C b for the purpose of updating, respectively.
Step 3. Update the best solution and the corresponding fitness. More specially, find the best solution C nb in C b and corresponding solution S nb , if C nb is better than C best , and then set C best � C nb and S best � S nb .
Step 4. Accept P � P b unconditionally.
Step 5. Obtain the optimal parameters of FOPFC (i.e., the optimal parameters P, c, and λ used in FOPFC) and corresponding fitness value C best when the predefined I max is satisfied; otherwise, go to Step 2 with the P.

Remark 3.
To improve integer-order PFC, we have used two key operators in the PEO-FOPFC. One is the application of fractional-order calculus into PFC. is operator makes the PFC have more tuning parameters. e other operator is employing an effective PEO algorithm to optimize the related parameters in FOPFC by minimizing the constructed fitness function, i.e., equation (31). From equation (31), one can see that a comprehensive performance is considered. Additionally, a specific improvement performance can be obtained by adjusting the weighting factors. Compared with the integer-order PFC, the fractional-order PFC is more likely to obtain a better comprehensive performance by using PEO.

Remark 4.
e fractional calculus introduced into the cost function of PFC is expected to enhance the performance of PFC because of more tuning parameters. However, how to determine the adjustable parameters of FOPFC is not easy due to lacking analytical knowledge. us, this paper proposed PEO to tune the related parameters of FOPFC. In other words, the proposed PEO-FOPFC is one kind of FOPFC in essence, where the adjustable parameters of FOPFC are determined by PEO. Remark 6. In [22], the authors have successfully applied the fractional-order predictive functional control into industrial heating furnace. According to this reference, we can implement the proposed PEO-FOPFC for real-world engineering problem described as follows: after using the PEO algorithm, we can obtain the optimized FOPFC. en, Grünwald-Letnikov approximation is used to transform the fractional-order process into an integer formulation. Finally, the implementation of fractional-order mechanism can be transformed into the implementation of integer modelbased PFC.

Simulation Results on Injection Molding Process
In this section, we firstly introduce a typical industrial process, i.e., injection molding process in Section 4.1. en, in order to demonstrate the effectiveness of the proposed PEO-FOPFC, we do some experiments on control of the injection velocity under different partial faults and unknown disturbance. In this section, our experiments are divided into three parts for different purposes. e first part is to investigate the ability of fractional order. us, we compare the FOPFC with PFC in this part. e second one is to compare the proposed control method with the recently published PFC [16]. Note that the PFC in [16] is Start Input the system model and related parameters of PEO-FOPFC One initial population P I = {S 1 , S 2 , …, S NP } is obtained with N P randomly generated solutions, where each solution S i denotes one group parameters in FOPFC. Subsequently, set P = P I , S best = PI be , and C best = F(S best ) Obtain D mutated solutions {S ik , (k = 1, 2, 3, …, D)} by MNUM operation shown in equations (28) and (29)  Select the best solution among S ik based on the rank index, and term it as S bi . Additionally, the corresponding best fitness value is termed as C bi . Then save S bi and C bi in P b and C b , respectively C nb is better than C best ? C best and S best unchanged N N Y Y C best = C nb and S best = S nb Accept P = P b , unconditionally Is I max satisfied?
Obtain the optimal parameters of FOPFC and corresponding fitness value C best The i-th solution The N P -th solution      the improved version of PFC, and the performance of PFC in [16] has been proved to be better than the traditional PFC.
is part is designed to show the PEO-FOPFC's superiority to the reported PFC [16]. e third one is to validate the influence of PEO on tuning the related parameters (i.e., P, λ, and c) in PEO-FOPFC. Here, the FOPFC and two other evolutionary algorithms (i.e., GA and PSO) based FOPFC methods are considered as competitors. It should be noted that all the computer simulations are conducted in MATLAB 2016a software on a 2.5 GHz and 8 GB RAM computer under the Windows 7 operating system. [16], the process of injection molding consists of three phases: filling, packing, and cooling. Among above three phases, the packing stage is very important because it plays key role in maintaining product quality, mechanical strength, deformation, and accuracy. erefore, the injection velocity should be controlled with high-precision to get the high-quality products. Generally, it is operated through the opening of the proportional valve by regulating the flow of hydraulic oil. us, in this process, the output is injection velocity termed as y(k), and the input variable is the proportional valve opening termed as u(k), respectively. In this paper, the typical

Complexity
injection velocity control model is adopted from [16] shown as follows: where y(z) and u(z) are the z-transforms of y(k) and u(k), respectively. In addition, there exists unknown actuator failure α in the valve opening. And, the set-point r(k) form is shown in the following equation: where L is set as 80. e goal of controller design is to make the process output track the set point as closely as possible under actuator failures and unknown disturbances. In this study, six faults are considered under nonrepetitive unknown disturbance. More specially, three cases of constant faults and three cases of time-varying faults with a random white noise of standard deviation 0.2 added to the process output are used. Table 1 gives the above six conditions. Additionally, in order to quantitative analyze the control performance of different methods, we use the following performance indices: IAE, ITAE, maximum deviation M d , and steady-state error Ess. Here, IAE, ITAE, M d , and Ess are Smaller performance implies a better control performance.

Remark 7.
In the all simulations, the fault is from beginning to end, which includes the moment the set-point is changing.

Simulation 1: FOPFC and Its Comparison with PFC.
As suggested in [16], the related parameters in PFC are set as follows: prediction horizon P � 4, the weighting factors Q j � [1, 0, 1, 0], and the number of base function N � 1. To the best of the authors' knowledge, the FOPFC strategy is firstly proposed for the industrial process with partial actuator failures. us, this subsection is devoted to investigating the control performance of FOPFC against the partial actuator failures in the industrial process and the effects of fractional-order mechanism on the control performance. Here, the version of PFC equipped with fractional-order mechanism, i.e., FOPFC, is considered to be compared with PFC [16]. Note that the main parameters in FOPFC are set as the same as those of PFC, except the specific parameter, i.e., fractional-order 14 Complexity parameter c used in FOPFC. Here, c is set to 0.5 by the trail-and-error tuning approach. Figures 4-9 show the comparative results of output responses and input signals obtained by PFC and FOPFC under different six cases. In addition, Table 2 tabulates four performance indices. Clearly, the closed-loop system responses and input signals obtained by FOPFC are slightly better than those of PFC under Case 2 and Case 3. For Case 1, FOPFC has a similar performance to PFC. As another three server faults (i.e., Case 4, Case 5, and Case 6) with timevarying faults and a random white noise of standard deviation 0.2, FOPFC shows superiority to PFC, which indicates that the fractional-order mechanism have potential ability in solving more server faults.

Simulation 2: PEO-FOPFC and Its Comparison with PFC.
In PEO-FOPFC, T s is abstracted as 1 during the process of evolution because it has same influence on weighting factors, and I max , N P , and b are set as 10, 50, and 3, respectively. e parameters including P, c, and λ of FOPFC optimized by PEO are P � 2, c � 0.5659, and λ � 0.8231. And, the faults are the same with six cases used in simulation 1. e comparison results are presented in Figures 10-15 Table 3 presents the comparative results of control performance indices obtained by PEO-FOPFC and PFC [16] under six cases, where the best performance is highlighted in bold.

Complexity 15
From Figures 10(a)-15(a), the proposed PEO-FOPFC presents better output tacking performance than PFC because it shows less oscillations and smaller overshoots. Observed from Figures 10(b)-15(b), we can see that the PFC [16] presents more drastic input control signals than PEO-FOPFC, which indicates the product quality obtained by PFC [16] is worse than PEO-FOPFC. Besides, the tacking error performance in Figures 10(c)-15(c) also illustrates the superiority of the proposed PEO-FOPFC.

Simulation 3: PEO-FOPFC and Its Comparison with FOPFC, GA-FOPFC, and PSO-FOPFC.
In this section, we devote to investigating the influence of PEO algorithm on control performance of PEO-FOPFC. us, we compare PEO-FOPFC with FOPFC and two other popular evolutionary algorithms (i.e., GA [35] and PSO [36]) based FOPFC methods (termed as GA-FOPFC and PSO-FOPFC, respectively) under above six cases of faults. In order to clearly compare the three evolutionary algorithms-based FOPFC methods, we tabulate their related parameters and optimal parameters of FOPFC in Table 4. From Table 4, we can see the number of adjustable parameters used in PEO is less than that in GA and PSO, which indicates the PEO algorithm is simpler than GA and PSO from the perspective of algorithm design. Figures 16(a)-16(f) show the results of four compared methods, and the corresponding performance indices are presented in Table 5. From Figure 16 and Table 5, the main observations can be summarized as follows: (1) it can be found that the performance of  16 Complexity FOPFC for the industrial process with partial actuator failures can be improved by using evolutionary algorithms; (2) the PEO algorithm is better than GA and PSO algorithm to find the better related parameters of FOPFC in term of four indices, although M d obtained by PEO-FOPFC is worse than that by GA-FOPFC and PSO-FOPFC in Cases 4-6; (3) the proposed PEO-FOPFC can improve ensemble control performance against partial actuator failures under both constant faults and time-varying faults.

Remark 8.
In Table 1, the conditions of six actuator faults are listed and a different α is used. Smaller α value means more severe fault. From above simulations 1-3, we can find that worse control performance is obtained with smaller α values in terms of four control performance indices and output responses.

Simulation Results on Process Flow of Coke Furnace
In this section, we investigate the application of PEO-FOPFC into control of process flow of coke furnace under different partial faults and unknown disturbance. Figure 17 gives the sketch of process flow of coking furnace. e main job of this unit is coking residues oil. First, the residual oil is divided into two flows, i.e., FRC8103 and FRC8104, and the convection room of the process furnace (101/3) is preheated. en, the separated streams will be merged into the fractionating tower, i.e., T102, where the preheated oil exchanges     heat with the gas oil from the top of the coke towers. After that, the circulating oil, i.e., the heavy part of mixed oil, is pumped into two flows and sent back to the radiation room for the purpose of heating to about 495°C. Finally, the two branches join together and will be sent to the coke towers to remove coke. Each time the output temperature in the chamber has a large impact on this process. us, the output temperature needs to be accurately controlled in the coke furnace under different partial faults and unknown disturbance. As suggested in [37], a first-order flus dead time (FOPDT) can be used to test and can be described as follows: Under sampling time T s � 30 s, the discrete-time model can be obtained as follows [37]:  To T102 To T101/5,6

Complexity
Remark 9. e set-point is abstracted into 1 to 3, and the six faults considered here are the same with simulations 1-3 given in Table 1. e used adjustable parameters of PEO are also the same with simulations 1-3. e competitor is taken from [37], called predictive functional control, based on the extended state space model (denoted as EPFC here).
By employing PEO, we can obtain P � 5, c � 5.6773, and λ � 0.8335 for this simulation. Figures 18(a)-18(f ) compare the output responses of the set-point tracking obtained by PEO-FOPFC and EPFC. Table 6 presents the results of control performance obtained by PEO-FOPFC and PFC on the process flow of coke furnace. From Figure 18 and Table 6, one can see that PEO-FOPFC shows an improvement over EPFC under partial faults and unknown disturbance considered in the paper, which implies the PEO algorithm and fractional-order theory contribute a lot to obtain a better comprehensive control performance.

Conclusion
In this paper, we have proposed the PEO-FOPFC strategy for industrial processes under partial actuator failures. ere are two key operators in the PEO-FOPFC. More specifically, one is the application of fractional-order calculus into the PFC technique to improve its performance against partial actuator failures, and the other one is employing an effective PEO algorithm to search the optimal-related parameters in FOPFC. To demonstrate the strong competitiveness, the PEO-FOPFC strategy has been evaluated on two industrial processes under six cases considering constant faults, timevarying faults, and nonrepetitive unknown disturbance, compared with the recently developed PFC [16,37]. In addition, in order to purely verify the effectiveness of the fractional-order mechanism, the experiment is designed for comparing FOPFC with PFC. And, another experiment is used to purely evaluate the influence of PEO algorithm by comparing PEO-FOPFC with FOPFC, GA-FOPFC, and PSO-FOPFC. e three experimental comparison results imply that the fractional-order mechanism and PEO algorithm are largely improving the performance of PFC to maintain the control performance and against the partial actuator faults. As a consequence, the proposed PEO-FOPFC strategy can be considered as a strong competitive version of PFC strategy for industrial processes against partial actuator failures.
Overall, by introducing the fractional-order mechanism into the cost function, the FOPFC can be designed to improve the performance of PFC due to more tuning parameters. e application of PEO to tuning the adjustable of FOPFC can handle the difficulty in lacking the knowledge of analytical these parameters. However, the parameter tuning issue of FOPFC is based on the designed fitness function, which may change for different control systems. Also, the PEO algorithm costs more computation time.
In future, it will be a significant subject to extend the proposed PFO-FOPFC into more complex industrial processes and design more advanced controller for handling actuator faults [38].
Additionally, the transfer function from the output disturbance and input disturbance to the output response can be obtained, respectively, as follows: (A.21) One can see that which indicates that the constant output disturbances and input disturbances can be rejected without steady error.

Data Availability
e data used to support the findings of this study are available from the corresponding author upon request.

Conflicts of Interest
e authors declare that there are no conflicts of interest.