Chapter IV: Separable Subsystems
4.2 Separable Subsystem Control
Constructing the feedback linearizing controller in (3.6) requires the dynamics of the full-order system. However, in the case of large dimensional systems, the full dynamics may be unknown or may become computationally expensive, inhibiting feedback linearization.
Figure 4.1: Amputee-Prosthesis Separable System and Equivalent Subsystem.
(Left) Amputee-prosthesis separable system (blue), with separable prosthesis sub- system AMPRO3 (red). (Right) Equivalent prosthesis subsystem. (Middle) Control input from inverse dynamics of human-prosthesis motion capture walking data, determined with full-order system dynamics (blue) and with equivalent subsystem dynamics (red).
Control Law for Separable Subsystem
This section eliminates the need to know the full-order system dynamics for feedback linearization by constructing aseparable subsystem control law that only depends on subsystem dynamics. We begin by defining aseparable control system.
Definition 1: The affine control system(3.1)is aseparable control systemif it can be structured as
"
ยค ๐ฅ๐
ยค ๐ฅ๐
#
=
"
๐๐(๐ฅ) ๐๐ (๐ฅ)
# +
"
๐๐
1(๐ฅ) ๐๐
2(๐ฅ) 0 ๐๐ (๐ฅ)
# "
๐ข๐ ๐ข๐
# , ๐ฅ๐ โR๐๐, ๐ฅ๐ โR๐๐ , ๐ข๐ โR๐๐, ๐ข๐ โR๐๐ ,
(4.1)
where๐๐ +๐๐ =๐and๐๐ +๐๐ =๐.
Because of the structure of๐(๐ฅ) in (4.1), ๐ข๐ only acts on part of the system. This motivates defining aseparable subsystemindependent of๐ข๐.
Definition 2: For a separable control system (4.1), its separable subsystem is defined as
ยค
๐ฅ๐ = ๐๐ (๐ฅ) +๐๐ (๐ฅ)๐ข๐ , (4.2) which depends on the full-order system states๐ฅ โR๐.
Now, to construct a feedback linearizing control law for this separable subsystem, we construct output functions that solely depend on the subsystem states๐ฅ๐ โ R๐๐ and whose Lie derivatives solely depend on the subsystem (4.2).
Definition 3: For a separable subsystem (4.2) of the separable control system (4.1), a set of linearly independent output functions with vector relative degree
ยฎ
๐พ๐ =(๐พ๐
1, ๐พ๐
2, . . . , ๐พ๐
๐๐ )with respect to(4.1)areseparable subsystem outputsif they only depend on๐ฅ๐ โR๐๐ ,
๐ฆ๐ (๐ฅ๐ ) โR๐๐ , (4.3)
and meet the following cross-term cancellation conditions for ๐ =1, . . . , ๐พ๐
๐ โ1and ๐=1, . . . , ๐๐ :
๐ ๐ฟ
๐ ๐๐ ๐ฆ๐ (๐ฅ)
๐ ๐ฅ๐
๐๐(๐ฅ) =0, (D3.1)
๐ ๐ฟ
๐พ๐
๐โ1 ๐๐
๐ฆ๐ (๐ฅ)
๐ ๐ฅ๐ h
๐๐
1(๐ฅ) ๐๐
2(๐ฅ)i
= h 0 0
i
. (D3.2)
We use these outputs to introduce aseparable subsystem control lawin terms of the subsystem (4.2) alone.
Definition 4: For a separable subsystem (4.2) with separable subsystem outputs (4.3), we define a separable subsystem control law as the feedback linearizing control law
๐ขssc(๐ฅ) โ โ(๐ฟ๐๐ ๐ฟยฎ
๐พ๐ โ1 ๐๐
๐ฆ๐ (๐ฅ)
| {z }
๐ด๐ (๐ฅ)
)โ1(๐ฟยฎ
๐พ๐ ๐๐
๐ฆ๐ (๐ฅ)
| {z }
๐ฟโ
๐๐ ๐ฆ๐ (๐ฅ)
โ๐๐ )
=โ๐ดโ1
๐ (๐ฅ) (๐ฟโ
๐๐ ๐ฆ๐ (๐ฅ) โ๐๐ ).
(4.4)
This control law is independent of the rest of the system dynamics ๐๐(๐ฅ), ๐๐
1(๐ฅ), and๐๐
2(๐ฅ), but still depends on the full-order system states๐ฅ. We will address this dependence in subsequent results to develop an implementable form of this control law solely dependent on subsystem states and measurable inputs.
To compare this control law๐ขssc(๐ฅ)to๐ข๐ (๐ฅ), we constructseparable outputsfor the full-order system that include the separable subsystem outputs๐ฆ๐ (๐ฅ๐ )used for (4.4).
Definition 5: For a separable control system, a set of linearly independent output functions with vector relative degree๐พยฎareseparable outputsif they are structured
as
๐ฆ(๐ฅ) =
"
๐ฆ๐(๐ฅ) ๐ฆ๐ (๐ฅ๐ )
#
, ๐ฆ๐(๐ฅ) โR๐๐, ๐ฆ๐ (๐ฅ๐ ) โR๐๐ , (4.5) and ๐ฆ๐ (๐ฅ๐ ) are separable subsystem outputs with vector relative degree ๐พยฎ๐ . The remaining outputs ๐ฆ๐(๐ฅ) have vector relative degree๐พยฎ๐ and can depend on any of the system states๐ฅ. The number of subsystem outputs๐๐ and the number of the rest of the outputs๐๐ sums to๐, and๐พยฎ=( ยฎ๐พ๐,๐พยฎ๐ ).
For the following theorem, we define the auxilary control input ๐as divided in the following form:
๐=
"
๐๐ ๐๐
#
, ๐๐ โR๐๐, ๐๐ โR๐๐ . (4.6) Theorem 1: For a separable control system (4.1), if the control law ๐ข(๐ฅ) = (๐ข๐(๐ฅ)๐, ๐ข๐ (๐ฅ)๐)๐ (3.6) is constructed with separable outputs (4.5) and auxiliary control input(4.6), then๐ข๐ (๐ฅ)=๐ขssc(๐ฅ).
Proof: We begin by relating the 3 components (๐ดโ1(๐ฅ), ๐ฟโ
๐
๐ฆ(๐ฅ), ๐)of๐ข(๐ฅ)to the components of๐ขssc(๐ฅ), (๐ดโ1
๐ (๐ฅ), ๐ฟโ
๐๐
๐ฆ๐ (๐ฅ), ๐๐ ). We are given ๐ = โ ๐๐
by (4.6).
With condition (D3.2), we show:
๐ด(๐ฅ) = ๏ฃฎ
๏ฃฏ
๏ฃฏ
๏ฃฏ
๏ฃฏ
๏ฃฏ
๏ฃฐ
๐ ๐ฟยฎ
๐พ๐
โ1 ๐ ๐ฆ๐(๐ฅ)
๐ ๐ฅ๐
๐ ๐ฟยฎ
๐พ๐
โ1 ๐ ๐ฆ๐(๐ฅ)
๐ ๐ฅ๐
๐ ๐ฟยฎ
๐พ๐
โ1 ๐๐ ๐ฆ๐ (๐ฅ)
๐ ๐ฅ๐
๐ ๐ฟยฎ
๐พ๐
โ1 ๐๐ ๐ฆ๐ (๐ฅ)
๐ ๐ฅ๐
๏ฃน
๏ฃบ
๏ฃบ
๏ฃบ
๏ฃบ
๏ฃบ
๏ฃป
"
๐๐
1(๐ฅ) ๐๐
2(๐ฅ) 0 ๐๐ (๐ฅ)
#
=
๏ฃฎ
๏ฃฏ
๏ฃฏ
๏ฃฏ
๏ฃฏ
๏ฃฐ
โ โ
0
๐ ๐ฟยฎ
๐พ๐
โ1 ๐๐ ๐ฆ๐ (๐ฅ)
๐ ๐ฅ๐ ๐๐ (๐ฅ)
๏ฃน
๏ฃบ
๏ฃบ
๏ฃบ
๏ฃบ
๏ฃป
=
"
โ โ
0 ๐ด๐ (๐ฅ)
# ,
๐ด(๐ฅ)โ1=
"
โ โ
0 ๐ด๐ (๐ฅ)โ1
# ,
where the final step is true because of the 0 corner block and properties of block matrix inversion. Similarly, we show:
๐ฟโ
๐๐ฆ(๐ฅ)=
๏ฃฎ
๏ฃฏ
๏ฃฏ
๏ฃฏ
๏ฃฏ
๏ฃฏ
๏ฃฐ
๐ ๐ฟยฎ
๐พ๐
โ1 ๐
๐ฆ๐(๐ฅ)
๐ ๐ฅ๐
๐ ๐ฟยฎ
๐พ๐
โ1 ๐
๐ฆ๐(๐ฅ)
๐ ๐ฅ๐
๐ ๐ฟยฎ
๐พ๐
โ1 ๐๐ ๐ฆ๐ (๐ฅ)
๐ ๐ฅ๐
๐ ๐ฟยฎ
๐พ๐
โ1 ๐๐ ๐ฆ๐ (๐ฅ)
๐ ๐ฅ๐
๏ฃน
๏ฃบ
๏ฃบ
๏ฃบ
๏ฃบ
๏ฃบ
๏ฃป
"
๐๐(๐ฅ) ๐๐ (๐ฅ)
#
=
๏ฃฎ
๏ฃฏ
๏ฃฏ
๏ฃฏ
๏ฃฏ
๏ฃฐ
โ
๐ ๐ฟยฎ
๐พ๐
โ1 ๐๐ ๐ฆ๐ (๐ฅ)
๐ ๐ฅ๐
๐๐ (๐ฅ)
๏ฃน
๏ฃบ
๏ฃบ
๏ฃบ
๏ฃบ
๏ฃป
=
"
โ ๐ฟโ
๐๐
๐ฆ๐ (๐ฅ)
# .
Putting these components together to construct๐ข(๐ฅ) yields ๐ข(๐ฅ)=โ
"
โ โ
0 ๐ดโ1
๐ (๐ฅ)
# "
โ ๐ฟโ
๐๐
๐ฆ๐ (๐ฅ)
#
โ
"
โ ๐๐
#
=
"
โ
โ๐ดโ1
๐ (๐ฅ) (๐ฟโ
๐๐ ๐ฆ๐ (๐ฅ) โ๐๐ )
#
=
"
โ ๐ขssc(๐ฅ)
# , showing๐ข๐ (๐ฅ) =๐ขssc(๐ฅ) as defined in (4.4).
By Theorem 1, we can construct a stabilizing controller (4.4) for the subsystem independent of the rest of the system dynamics, identical to the portion of the controller constructed with the full-order system dynamics acting on the subsystem.
In the case of full-state feedback linearizable systems [39], this guarantees full- order system stability when (3.6) is applied to the rest of the system. This enables construction of stablemodel dependentcontrollers for separable subsystems without knowledge of the full-order system dynamics.
Equivalency of Subsystems
Although we now have a subsystem control law independent of the rest of the system dynamics, it still depends on the full-order system states๐ฅ โ R๐. Consider the case where the states ๐ฅ๐ โ R๐๐ cannot be measured. If we could construct an equivalent subsystem whose dynamics are a function of the subsystem states ๐ฅ๐ โR๐๐ , and measurable inputsF โR๐๐, we could calculate the subsystem control lawindependent of the full-order system states.
Consider another subsystem,
ยคยฏ
๐ฅ๐ = ๐ยฏ๐ (X) +๐ยฏ๐ (X)๐ขยฏ๐ X =
"
ยฏ ๐ฅ๐ F
#
=
"
๐ฅ๐ F
#
โR๐ยฏ,
(4.7)
where X is the state vector ยฏ๐ฅ๐ = ๐ฅ๐ augmented with an input F. Using the same separable subsystem outputsas (4.3), we define a control law ยฏ๐ข๐ (X)for the subsystem as:
ยฏ
๐ข๐ (X) โ โ(๐ฟ๐ยฏ๐ ๐ฟยฎ
๐พ๐ โ1
ยฏ ๐๐
๐ฆ๐ (X)
| {z }
ยฏ ๐ด๐ (X)
)โ1(๐ฟยฎ
๐พ๐
ยฏ ๐๐
๐ฆ๐ ( X)
| {z }
๐ฟโ
ยฏ ๐๐ ๐ฆ๐ (X)
โ๐๐ )
=โ๐ดยฏโ1
๐ (X) (๐ฟโ
ยฏ ๐๐
๐ฆ๐ (X) โ๐๐ ).
(4.8)
Theorem 2: For the subsystems (4.2) and(4.7), ifโ ๐ : R๐ โ R๐ยฏ s.t. ๐(๐ฅ) = X and the following conditions hold,
๐๐ (๐ฅ) = ๐ยฏ๐ (X), ๐๐ (๐ฅ) =๐ยฏ๐ (X),
(T2) then ๐ข๐ (๐ฅ) = ๐ขยฏ๐ (X). Applying these to (4.2) and (4.7), respectively, results in dynamical systems such that given the same initial condition๐ฅ๐0
๐ฅ๐ 0
= h๐ฅ๐(๐ก0)
๐ฅ๐ (๐ก0)
i yields solutions๐ฅ๐ (๐ก) =๐ฅยฏ๐ (๐ก) โ๐ก โฅ ๐ก0.
Proof. Since the subsystems have the same dynamics and outputs, the Lie derivatives comprising their control laws are also the same, hence
๐ข๐ (๐ฅ) =โ๐ดโ1
๐ (๐ฅ) (๐ฟโ
๐๐ ๐ฆ๐ (๐ฅ) โ๐๐ )
=โ๐ดยฏโ1
๐ (X) (๐ฟโ
ยฏ ๐๐
๐ฆ๐ (X) โ๐๐ ) =๐ขยฏ๐ (X).
With the same control law and dynamics, the closed-loop dynamics of the subsys- tems are the same:
ยค
๐ฅ๐ = ๐๐ (๐ฅ) +๐๐ (๐ฅ)๐ข๐ (๐ฅ)
= ๐ยฏ๐ (X) +๐ยฏ๐ (X)๐ขยฏ๐ (X)=๐ฅยคยฏ๐ . Hence, given the same initial condition๐ฅ๐0
๐ฅ๐ 0
= h
๐ฅ๐(๐ก0) ๐ฅ๐ (๐ก0)
i
, they have the same solution ๐ฅ๐ (๐ก)=๐ฅยฏ๐ (๐ก) โ๐ก โฅ ๐ก0.
By Theorems 1 and 2, we can construct a stabilizingmodel dependent controller, namely (4.8), for the subsystem (4.2) independent of the rest of the system dynamics and with measurable inputs.
Zero Dynamics. For full-state feedback linearizable systems, this control method will stabilize the full-order system. For partially feedback linearizable systems, we can apply this method to the feedback linearizable portion of the system. If the zero dynamics are stable, then we can guarantee full-order system stability, which will be proved through Lyapunov methods in the next chapter, Chapter 5.