This paper describes an approximated-scalar-sign-function-based anti-windup digital control design for analog nonlinear systems subject to input constraints. As input saturation occurs, the non-smooth saturation constraint is modeled with the approximated scalar sign function which is a smooth nonlinear function. The resulting nonlinear model is further linearized at any operating point with the optimal linearization technique, and Linear Quadratic Regulator (LQR) is then applied for a state-feedback controller optimal for each operating point. As input saturation is encountered, an iterative procedure is developed to adjust control gains by systematically updating LQR weighting matrices until the inputs lie within the saturation limits. Through global digital redesign, the analog LQR controller is converted to an equivalent digital one for keeping the essential control performance, and moreover, delay compensation is taken into account during digital redesign for compensating the potential time delays in a control loop. The swing-up and stabilization control of single rotary inverted pendulum system is used to illustrate and verify the proposed method.
Various types of hardware limitations always exist in practical control systems with potential effects on the final control performance. A typical one encountered in practice is actuator saturation. For instance, as common actuation devices, motors have limited speed and torque range, power sources have output bounds, control valves cannot be more than fully open or fully closed, etc. As the control command is saturated at the top or bottom limit during actuator saturation, the control loop is broken and the controller loses the ability to regulate plant’s behavior for the time being. This phenomenon, called controller windup, may lead to significant degradation in control performance, such as long settling time, high overshoot or even instability [
In order to circumvent the windup effect, there exist many anti-windup approaches in the literature in the past decades. Most of them are proposed for linear systems, like the extensively-studied two-phase approach [1-4] (a nominal linear controller is first designed with the saturation constraints ignored and then a conditioning scheme is developed for reducing the windup effects of saturation) and Linear Matrix Inequalities (LMI)-based methods [5-7]. To the authors’ knowledge, however, few anti-windup approaches have been developed for nonlinear systems. The very limited efforts include the work of converting the physical constraint problem to a statedependent constraint problem through a coordinate transformation method [8,9], and those based on input-output linearization or feedback linearization [8-11].
This paper proposes an approximated-scalar-signfunction-based anti-windup technique for analog nonlinear systems with input constraints. As a non-differentiable function, sign function or absolute-value function has the inherent capability to describe the instantaneous jumps where the system model loses smoothness, so they commonly appear in many analytical models of nonsmooth nonlinearities, such as Bouc-Wen hysteresis model [12,13] and Stribeck friction model [
In addition to input constraints, time-delayed systems are another practical concern in the proposed design. This concern arises from the fact that in a sampled-data control system which is a popular control scheme nowadays due to the advance of computer technology, some fundamental operations like controller computation, A/D and D/A conversions, sensing and actuation etc, could cause time delays in the control loop. Another example is networked control systems where components communicate with each other through a real-time network, which inevitably causes transmission delays. Ignoring the delays in a control loop may lead to the failure of designed control so it is of practical interest to extend the developed anti-windup methodology to time-delayed systems. In this paper, the authors propose an input-delay-compensating digital redesign approach: an analog state-feedback controller is first designed in the delayfree case for the desired control performance; then a digital controller is obtained from the analog one through global digital redesign with the delay compensation considered. The resulting digital controller is able to maintain the essential control performance of the analog counterpart even in a time-delayed environment.
The rest of this paper is organized as follows. Section 2 introduces preliminary techniques used in the proposed anti-windup design, including approximated scalar sign function, optimal linearization and LQR. In Section 3, a global digital redesign method is developed for input delay compensation. The proposed anti-windup methodology is described in Section 4. Section 5 gives simulation results of proposed method on the swing up and stability control of single rotary inverted pendulum. Finally, the paper is concluded in Section 6.
The scalar sign function is defined in [
where, and denotes the open lefthalf complex plane and the open right-half complex plane, respectively. It is noted that the imaginary axis is undefined in.
The scalar sign function has an alternative form in [
where and
It is also reported in [
and its j-th truncation can be written as
It can be shown that the j-th truncation (5) gives the better approximation of (3) as the value of j approaches the infinity, i.e.
Replacing in (2) with the j-th truncation (5) yields the approximated scalar sign function for a complex number as
where which denotes the positive integer set. It can be inferred from (6) that (7) has the limit
Particularly, the approximated scalar sign function for a real number is
which has the limit
If the definition of scalar sign function for real numbers is extended to include zero, i.e.
then
as.
The concern of this paper is limited to the real number case (9) and the definition (11) is considered. Differentiating the approximated scalar sign function (9) with respect to the real number σ yields
for. It is easy to find that (13) is continuous everywhere, which proves the approximated scalar sign function (9) is differentiable everywhere.
By utilizing the approximated scalar sign function, an approximate model can be obtained for a non-smooth
dynamical system with sign function constraints, whose approximation accuracy can be adjusted with the approximation order j. Due to the smoothness and nonlinearity of the approximated scalar sign function, the resulting approximate model is still nonlinear, but differentiable everywhere, which makes possible the further local linearization.
Remark 1: As shown in
Local linearization is a typical way of handling nonlinear systems and the most popular technique is Jacobian linearization [
In order to circumvent the limitation of Jacobian linearization, Teixeira and Zak formulated linearization prob-
lem as a convex constrained optimization problem and proposed an optimal linearization approach in [
Consider a general class of nonlinear system in the form
where is the state vector, is the input vector, with is a differentiable nonlinear function vector and is a function matrix. Its optimal local linear model at an arbitrary operating point is in the state-space form
where
is the Jacobian matrix of evaluated at the operating point xk. It is noted that the case for xk = 0 in (16) agrees with the aforementioned exception case of the operating point being an equilibrium at the origin.
Remark 2: When f(x) in (14) is a scalar nonlinear function, (16) is reduced to a scalar number as
.
Together with a linear output equation, the linearized state Equation (15) constitutes a complete local linear model as
where is the controlled output vector and is a constant matrix. According to LQR [
with and, is then given by
where
r(t) is the reference for the controlled output y(t) to track, and Pk is the positive definite and symmetric solution of the Ricatti equation
The weighting matrices Q and R should be tuned to make the resulting analog controller (20) give a desired control performance in the delay-free case.
For time-delayed systems where delays can be caused by network transmission, controller computation, A/D and D/A conversions etc, a delay compensating technique is proposed next based on global digital redesign. Compared with the conventional digital redesign methods that the conversion is limited in the scope of the controller, like the bilinear transformation [
For a Single-Input-Single-Output (SISO) plant, the time delays in a control loop from network transmission, controller computation, A/D and D/A conversions, sensing and actuation etc, can be combined and then allocated to either the input or the output side of the plant for control design purpose [
where τ is the combined input delay. Among global digital redesign methods, the prediction-based digital redesign in [
The analog linear plant (18) and the analog control law (20) constitute a complete analog control system as depicted in
where the analog control input is a piecewiseconstant signal generated from the digital control input through Zero Order Hold (ZOH) as for, and the digital state xd(kT) is the sample of the analog state xd(t) at the sampling instant t = kT; T is the sampling/control period. Equation (25c) is the digital control law to be designed so that the closed-loop state in the hybrid system can closely match the closed-loop state in the analog system at defined inter-sampling instant.
The solution of (18a) at for is
where is a piecewise-constant input,
andin which is an identity matrix of appropriate dimension. The solution of (25a) at is
Comparing (27) with (26) yields that with the assumption of, the condition of is, which leads to the prediction-based digital controller
By substituting (27) into (28), is solved as
(29)
where is an identity matrix of appropriate dimension. As a result, the desired digital control law (25c) is found from (29) having digital control gains
and the digital reference. The parameter v can be tuned to adjust the control performance.
Next, the conclusion in (28) will be extended for the input delay compensation. Considering the combined input delay τ, the state equation (25a) in the hybrid system is changed to
In order to compensate the input delay, the digital control input in (28) should be further predicted for the delay duration τ as
where the future state needs to be predicted based on the available signals and.
Compared with the sampling/control period, the time delays from controller computation, A/D and D/A conversions etc, are relatively small, so a reasonable assumption is made in the following reasoning for simplicity purpose that the total duration of time delays in the control loop should not be larger than a sampling period, i.e. the combined input delay. As for the case that, the proposed method can still handle but will produce a more complicated controller structure.
Let, so and. The solution of (31) at is
where , and.
By substituting (33) into (32), is solved as
As a result, the desired digital controller for input delay compensation is derived from (34) as
(35)
where
and. The resulting hybrid control system has the configuration as shown in
Consider a general nonlinear plant
where the symbols are defined as in (14) and (18b), is the true input vector whose i-th entry has the constraints
for, in which is the ideal i-th input and are the top and bottom limits for the i-th input, respectively. For the purpose of simplicity, let in the following reasoning. Through the optimal linearization in Section 2, the optimal local linear model of the nonlinear plant (37) at the k-th sampling instant is obtained as
where are given by (16) and (17), respectively. Should no input saturation occur, i.e. for, the state equation (39a) becomes as.
Whenever input saturation happens, say, the i-th ideal input is out of limits, the state Equation (39a) becomes
which is apparently non-smooth due to the scalar sign function. Substituting approximated scalar sign function (9) for yields an approximated equation as
Equation (41) is still nonlinear due to which can be further linearized by Remark 2 as
where is the k-th operating point of . In this way, the non-smooth saturation Equation (41) is fully linearized as
(42)
where
,
and. is called input scaling factor.
To prevent the i-th input from saturation at t = kT, the input scaling factor should be tuned such that
, that is
.
As mentioned in Remark 1, for j is even, so the approximation order j is always an even number in the proposed anti-windup design.
From the linear state Equation (42), the input matrix is required for designing the ideal control input, while the input scaling factor, which determines, involves the ideal control input at t=kT which has not been designed yet. To break this deadlock, a pre-designed control input is used to estimate which is then utilized to design the true control input. For a sampled-data control system, a pre-designed digital control input can be used for this purpose. As a result, the linear state Equation (42) is modified to
where the estimated input matrix , the estimated input scaling factor with
for j is even. is the pre-designed digital control input which can be the digitally redesigned control input for the i-th component of at t = kT from the prediction-based digital redesign (25c) or the proposed delay-compensating digital redesign (35). With arbitrary input elements saturated, the state Equation (43) is generalized as
with for, where or is given by (44) depending on whether the corresponding ideal input violates the constraints or not.
The general saturated state Equation (45) and the output Equation (39b) form a saturated system with modified system matrices. Applying LQR (21-23) to it yields that with weighting matrices (Q, R), the optimal control law is determined by solving the Riccati equation
which can be rewritten as
where. The resulting optimal control law is
where the feedback gain with, and the forward gain
with.
On the other hand, applying LQR to the original system in (39) with the modified weighting pair yields the optimal control law
where the feedback gain and the forward gain are shown above. The corresponding Riccati equation is the same as (47).
It can be observed by comparing the above two LQR designs that the problem of finding the optimal control law for the saturated system (45) with the weighting matrices (Q, R) can be reformulated as the same problem for the original system (39) using the same performance index (19), but with the new weighting matriceswhere. Generally, both weighting matrices are selected to be positive definite diagonal matrices. So when the input scaling factor’s i-th diagonal element, the corresponding i-th diagonal elements of R and have the relationship of. As the matrix Q is kept the same in the performance index, the resulting control law with will have a smaller i-th element than the counterpart with (Q, R). Accordingly, the corresponding i-th element of digitally redesigned control input would be smaller as well, thus possibly avoiding input saturation. It is noted that although is times greater than Ri, the newly resulting digital input is not necessarily times smaller than the original one and may be still out of input limits after a single update. Therefore, it might be necessary to recursively update on until the input lies within the saturation limits.
Remark 3: The approximation order j in the approximated scalar sign function can be tuned to achieve certain tradeoff between the control performance and the recursion efficiency. As shown in
approximation order j leads to a faster decaying approximated scalar sign function, which results in a smaller. This can be regarded as an ‘aggressive’ update scheme with the advantage that a smaller number of iterations are needed for suppressing the input to within the limits. But the disadvantage is that the input could be suppressed too much below the limit, which may bring a fairly slow system response.
As a typical nonlinear system, the rotary inverted pendulum system depicted in
The dynamics between the arm angle θ, the pendulum angle α and the motor voltage Vm is derived using Lagrangian equations as
(50a)
where
, , ,
, , ,
, , ,
, ,
, ,
, ,
.
The parameters have physical meanings defined in
For space-saving purpose, the derivation of optimal local linear model of (50) is skipped, and only salient simulation results and necessary explanations are presented here. The proposed anti-windup methodology in
Next, an input time delay is introduced between the pendulum’s motor and the digital controller to represent
the potential time delays in the real-world control loop. In order to demonstrate the delay-compensating capability, a reasonably long time delay τ = 0.02 s is used. The control period T is accordingly extended to 0.02 s as the proposed design assumes a control period no shorter than the delay duration. Other parameters and initial conditions are the same as the delay-free case. Simulation results showed that the controller from the prediction-based digital redesign cannot succeed any more, exhibiting an unstable behavior. In contrast, the proposed anti-windup plus delay-compensating controller can still survive with performances shown in
This paper describes the design and application of an
optimal anti-windup digital controller for analog nonlinear plants subject to input constraints. As a new antiwindup technique for sampled-data control systems, the proposed method has the following contributions: 1) the approximated scalar sign function is utilized to model non-smooth input saturations, which presents a new effective solution to sign-function constrained non-smooth problems; 2) through the optimal linearization, a general approach is developed for handling nonlinear systems with linear control theories in a broader region instead of conventionally being limited to around equilibriums; 3) aside from the anti-windup functionality, the proposed digital controller is capable of compensating time delays in the control loop, which would help guarantee its designed performance in real world implementations.
Research supported by NSF award #1238859, and Department of Education, HBGI Grant.