The present invention relates to generally controlling electric motors, and more particularly generating trajectories for motor controlled actuators subject to dynamics, acceleration and velocity constraints.
Motion control systems are used in number of positioning applications, e.g., single-axis positioning, and multiple-axis positioning. For example, a simple single-axis positioning motion control system generally includes sensors, controller, amplifier, and actuator motor. The actuator follows a predetermined trajectory subject to state and control constraints, i.e., dynamics, acceleration, velocity. The trajectory of the actuator can be designed to reduce vibration induced by the motor.
For two motor control cases,
Although minimal time motor controllers generate the fastest trajectory for each motion, for a complex processes, minimal time controller may not help improve the overal productivity if a bottleneck of production is due to other slower processes, such as material processing. For example, there is no advantage in rapidly moving a work piece to a next state using excessive energy, if the piece is not going to be manipulated until later.
For such systems, minimal time controllers are not only unnecessary, but also inefficient because the controllers are not energy optimal. Furthermore, the efficiency of a plant depends not only on productivity, but also on other costs, such as energy consumption. The maximum efficiency is usually generated with certain trade-off between productivity and energy consumption. Therefore, strictly minimal time controllers, although useful in certain cases, do not increase efficiency in general, and minimizing energy consumption by relaxing time constraints should be considered for optimal motor control.
Optimal Control Theory
Optimal control deals with the problem of finding a control law for a system such that a certain optimality criterion is achieved. The control problem includes a cost function of state and control variables. An optimal control has to satisfy a set of differential equations describing paths of the control variables that minimize the cost function.
Pontryagin's minimal principle for optimal control theory determines the best possible control for taking a dynamical system from one state to another, especially in the presence of constraints on the state or control inputs. The optimal control theory provides a systematic way for determining the optimal solution to the problem of minimizing certain cost functions, such as time and energy, subject to various constraints, including dynamics constraint, boundary conditions (BC), state constraint, control constraint, and path constraint. Therefore, the energy efficient motor control problem can be addressed as an optimal control problem.
The optimal control can be obtained by solving a two-point boundary value problem (TBVP), or a multi-point boundary value problem (MBVP) if the optimal solution contains multiple segments. This usually happens when control or state constraints are active. For the minimal time motor control problem, the optimal solution can be obtained analytically. Such an analytic solution forms the basis of many minimal time motor controllers.
However, for an energy saving optimal control problem, the corresponding TBVP and MBVP are difficult to solve, and no analytic solution is readily available. The existing indirect methods for solving the TBVP and MBVP, including single shooting method (SSM) and multiple shooting method (MSM), are computationally complex for real-time motion control applications. Besides, the convergence of those methods are generally not guaranteed, and rely on an initial guess of certain key parameters in the methods. Hence, due to the computation complexity issue and the reliability issue, the existing methods for solving TBVP and MBVP are difficult to be applied for real-time energy efficient trajectory generation in motor control applications.
The direct transcription method (direct method), provides an alternative way for solving optimal control problems. Similar to shooting methods, the convergence of the direct method is not guaranteed. A comprehensive evaluation of current direct methods, including a pseudo-spectral method and a mesh refinement method, shows that the direct method cannot provide motor control in real-time.
Thus, the known methods are insufficient in terms of computation efficiency and reliability for the real-time application of energy saving motor control. Due to these difficulties, there is a need for a method to generate energy efficient reference trajectories for motor control. Such a method should be computationally efficient for real-time motor control applications, and should be reliable. It is also desirable that such a method provides the capability to adjust the trade-off between execution time and energy saving for different applications.
The embodiments of the present invention provide methods for generating trajectories for motor controlled actuators subject to dynamics, acceleration and velocity constraints. The method considers the energy consumption of a motor motion control system due to resistive loss of the motor, and mechanical work. The motor motion control trajectory generation problem is formulated as an optimal control problem with different constraints comprising dynamic, acceleration, and velocity constraints.
The invention uses an analytic solution of an unconstrained case of motor control to search for the optimal solution of the constrained cases using an iterative process. Using optimal control terminologies, such an approach corresponds to solving a multi-point boundary value problem (MBVP) by iteratively solving a two-point boundary value problems (TBVP) until a termination condition is reached. For example, the termination condition is that all constraints on the solution, such as velocity and acceleration, are satisfied. Because the evaluation of analytic solutions is computationally efficient, the MBVP problem can be solved quickly. Special methods are provided to ensure that the iterative process is guaranteed to converge to the optimal solution.
The method can be performed in a processor 300 connected to a memory and input/output interfaces as known in the art. The method considers energy consumption of a motor motion control system due to resistive loss of the motor, and mechanical work. Although the example motor is rotational, other motors, such as linear motors can also be used with the invention.
Step 310 initializes all data for solving a two-point boundary value problem (TBVP), including parameters for a motor model and a positioning task. The data are input to the method 310.
Step 320 solves the TBVP using the data and an analytical solver for unconstrained motor optimal control subject to the boundary conditions, as described in greater detail below.
Step 330 identifies any violations of an acceleration constraint, and if true, then step 370 updates the boundary conditions, and iterates beginning at step 320.
Step 340 identifies any violations of a velocity constraint, and if true, then step 360 updates the boundary conditions, and iterates beginning at step 320. velocity constraint is violated to update the BC and repeat step 320.
Otherwise, step 350 sets the trajectory of the actuator 390 of the motor 38 to the solution of the TBVP if the acceleration constraint and the velocity constraint are satisfied.
Energy Efficient Motor Control Problem with an Optimal Control Formulation
The lumped inertia of a load and a motor is I, and a torque constant of the motor is Kt. We define d=
{dot over (x)}=v, (1)
{dot over (v)}=−dv−c+bu. (2)
The motion for the motor satisfies the velocity and acceleration constraints, which are
v≦v
max, (3)
A
min
≦−dv−c+bu≦A
max, (4)
where vmax is maximum allowable velocity, Amin and Amax are bounds on system acceleration with
A
min<0 and Amax>0.
The energy consumption of the motor is affected by many factors such copper loss (heat produced by electrical currents in the motor windings), iron loss (magnetic energy dissipated when the magnetic field is applied to the stator core of the motor), and mechanical work (friction in the motor). An instantaneous power of the motor consider these factors is
P(v,u)=Ru2+Kh|v∥u|γ+Ks|u|+Ktvu, (5)
where R is the resistance of the motor, Kh are hysteresis losses, γ is a constant for the hysteresis loss, Ks is a constant related to the switching loss, and Kt is a torque ratio. When P is negative, the motor becomes a generator converting mechanical work into electricity by braking. This electricity is dissipated. Hence, the total energy consumption of a motor during a time period [o, tf] is
E=∫
0
t
Q(v(t),u(t))dt, (6)
Q(v(t),u(t))=max{0,P(v,u)}.
The minimal energy motor control is given by the solution to the following optimal control problem:
The BC can be different for different cases such that x(0) and v(0) are not necessarily zero. However, they are set as zeros to simplify the notation. Because the positive rotational direction of the motor can be arbitrarily assigned, without loss of generality, it is assumed that xf>0.
Cost Function Simplification
The optimal control problem, such as Problem 1, can be solved faster with the appropriate simplified cost function. A method for cost function simplification is described below.
Problem 1 with power function of Eqn. (5) is first solved via a numerical optimization using a density function based mesh refinement process. A total of sixty-four cases with different execution time (or, final time) tf and final position xf were solved. The contribution of different terms in (5) are analyzed.
Specifically, the following quantities
are determined for all test cases and compared. The result shows that the copper loss term ru2 dominates the other terms when the average velocity of the position transit, i.e., xf/tf, is small. When xf/tf is large, the mechanical work term Ktvu dominates the others. This indicates that a simplified cost function including the copper loss and the mechanical work of the motor is a good approximation of the original cost function with the power determined by Eqn. (5).
To evaluate the optimality of the result using the simplified cost function, the following power functions are used to determine the optimal trajectory via numerical optimization:
Q
A(x,u)=Ru2Kh|x2∥u|γ+Ks|u|+Ktx2u,
Q
B(x,u)=Ru2Kh|x2∥u|+Ks|u|+Ktx2u,
Q
C(x,u)=Ru2+Ks|u|+Ktx2u,
Q
D(x,u)=Ru2+Ktx2u,
Q
E(x,u)=Ru2,
Q
F(x,u)=max{0,Ru2+Ktx2u}.
The baseline cost
J
b
=[J
b1
, . . . ,J
b64]
is used for comparing the loss of optimality when simplified power functions are used for optimization. This baseline cost is obtained by solving Problem 1 with the actual power function Q(v, u) using the numerical optimization approach for all 64 cases. The adaptive mesh scheme is applied for better accuracy. With each of the power functions QA to QF, Problem 1 is also solved via numerical optimization for all 64 cases.
The relative cost error for the ith case is estimated by
and for each cost function, the vector of relative cost error Δ
According to the numerical optimization result listed in Table 1, the numerical optimization approach using direct transcription takes 1.6 to 5.6 seconds to solve the problem, which is too slow for real time motor control applications.
Therefore, power functions used to solve Problem 1 analytically are more suitable for applications. Such power functions include QD and QE. Because the power function QD provides acceptable optimality as compared to the true cost function, and can be solved analytically, it is used to determine the energy consumption cost function as
E=∫
0
t
Q
D(v,u)dt=∫0t
Analytic Solution to the TBVP for the Unconstrained Case
In this section we present the analytical solution to Problem 1 using the simplified cost function (7) without velocity and acceleration constraints. The optimal solution for this case is given by the following problem:
Problem 2. Unconstrained Minimal Energy Motor Control with Simplified Cost Function
A description of the problem to minimize the energy can be formulated as follow:
Problem 2 is a linear system optimal control problem with quadratic cost, therefore, it can be solved analytically. The Hamiltonian for Problem 2 is given by
H=Ru
2
+K
t
vu+λ
x
v+λ
v(−dv+bu−c),
where λx and λv are the co-states for x and v dynamics, respectively. According to optimal control theory, the dynamics of the co-states are
Note that λx is constant according to Eqn. (8). The optimal control u* is determined from the first-order optimality condition ∂H/∂u=0. which yields
Bringing the expression of optimal control in Eqn. (10) into Eqns. (8-9), we have the following Two-point Boundary Value Problem (TBVP)
Two-Point Boundary Value Problem (TBVP) for Unconstrained Motor Control
The TBVP can be formulated as:
with unknown parameter λx and boundary conditions
x(0)=0, X(tf)=xf,
v(0)=0, v(tf)=0, λv(0) and λv(tf) free.
Let
and define
Then the differential equations in the TPBV can be written more compactly as
The solution to linear system (11) is given by
where M(t)=eAtε3×3, and G(t)ε3×3 is given by
G(t)=eAt∫0te−Aτdτ.
The BC of the TBVP satisfies Eqn. (12) with t=tf
from which the unknowns
λv
can be solved. After these unknowns are solved, the optimal state and co-state histories can be determined from Eqn. (12), and the optimal control is given by Eqn. (10).
Method for Solving the MBVP with Active Acceleration Constraint
Next we describe a method for determining the optimal solution to Problem 2 subject to acceleration constraints.
Problem 3. Acceleration Constrained Minimal Energy Motor Control with Simplified Cost Function
In the above description, the term Amin≦−dv−c+bu≦Amax is the acceleration constraint 320.
The analytical result for the TBVP indicate that the control solution to Problem 2 is positive at the beginning for accelerating the motor, then negative for deceleration. {dot over (v)} is larger around t=0 and t=tf. For a given final position xf when the final time tf is large enough, the acceleration constraints are not activated. As tf decreases, it requires faster acceleration at the beginning and deceleration at the end such that the motor move over the same distance within a shorter time. When tf is small enough, the acceleration constraints can be activated around t=o and t=tf.
When the acceleration constraints are activated, the optimal solution exhibits a three phase structure: maximum acceleration, unconstrained optimal solution (analytical solution), and minimal deceleration. In the first and third phases, the position and velocity of the motor are determined explicitly by
In the second phase, because the acceleration constraints are inactive, the optimal solution during this phase is given by the analytic solution to Problem 2 with BC
x
m(t1*)=xl(t1*), vm(t1*)=vl(t1*), xm(t2*)=xr(t2*), vm(t2*)=vr(t2*),
where t1* and t2* are the optimal switching times from an acceleration constrained arc ({dot over (v)}=Amax) to an unconstrained arc, and from an unconstrained arc to a deceleration constrained arc ({dot over (v)}=Amin), respectively, and xm and vm are the optimal position and velocity solution for the second phase.
The optimal control u* for Problem 2 is continuous according to the optimal control theory. This further implies that the derivative of the optimal velocity is continuous. Hence, the junction conditions at the optimal switching times
t
1* and t2*
are described by two tangent conditions
{dot over (v)}
l(t1*)={dot over (v)}m(t1*), {dot over (v)}r(t2*)={dot over (v)}m(t2*)
Therefore, the solution to the acceleration constrained energy optimal motor control problem is determined from the following system of equations which form the MBVP:
The analytic expression for the first and third phases of optimal solution has been applied in the above MBVP for simplification, hence the BC for these phases are automatically satisfied. There are a total of nine equations and nine unknowns
λx,t1*,t2*,xm(t1*),vm(t1*),λv(t1*),xm(t2*),vm(t2*),λv(t2*),
therefore the MBVP is solvable. However, the whole system is nonlinear and no analytic solution can be found for the MBVP. Besides, there is no guarantee that the current numerical method can solve this problem. Solving such a system of equations is also time-consuming.
Because velocity and reliability are crucial for the real-time application of optimal motor control, we describe a method for solving Problem 3. The optimal switching times t1* and t2* are solved by identifying the optimal velocity profiles vk for an unconstrained minimal energy control problems.
This method is presented by flow chart
The steps in
{dot over (v)}(ta
The above method above is guaranteed to generate the optimal solution for the acceleration constrained energy optimal motor control problem.
Method for Solving the MBVP with Active Velocity Constraint Next, we describe a method for solving the energy optimal motor control problem when the velocity constraint is active.
As shown in
Similar to the acceleration constrained case, the optimal control approach for solving the velocity constrained case also leads to a MBVP, which is even more complicated and difficult to solve than that of the acceleration constrained case. Hence, we provide a method for solving the velocity constrained energy optimal motor control problem.
Let (x*, v*, u*) be the optimal solution to Problem 1 with final position xf and final time tf. Suppose the state constraint v≦vmax is active on the interval [t3*,t4*], such that v≦vma if and only if tε[t3*,t4*], where t3* and t4* are the optimal switch time entering and exiting the state constraint. Let Δt*=t4*−t3*, and let ({tilde over (x)}*, {tilde over (v)}*, ũ*) be the optimal solution to Problem 3 with final position xf−Δt*vmax, and final time tf−Δt*. Then (x*, v*,u*) and ({tilde over (x)}*, {tilde over (v)}*, ũ*) are related by
Hence, if Δt* is determined, then ({tilde over (x)}*, {tilde over (v)}*, ũ*) can be solved, from which (x*, v*, u*) can be determined using Eqns. (14-16).
The value Δt* is decided from the condition that
maxt{tilde over (v)}*(t)=vmax,
or, equivalently,
{tilde over (v)}*(t)|{tilde over ({dot over (v)}(t)=0=vmax.
The steps in
Detailed Description Step
Otherwise, update δi using the Newton's method as
Optimal Solutions
The optimal solution for three representative cases given by the disclosed method are shown in
Using our method, it takes less than 40 ms to find the optimal solution for each test case. The average computation time is 7.2 ms, which is fast enough for real time energy saving motor control applications.
Although the invention has been described by way of examples of preferred embodiments, it is to be understood that various other adaptations and modifications may be made within the spirit and scope of the invention. Therefore, it is the object of the appended claims to cover all such variations and modifications as come within the true spirit and scope of the invention.