data informativity for system analysis and control part 1...
TRANSCRIPT
— 1
Data informativity for system analysis andcontrol part 1: analysis
Jaap Eising, Henk J. van Waarde, Harry L. Trentelman, and M. Kanat Camlibel
University of Groningen
39th Benelux Meeting on Systems and Control
— 2
Motivation
Model-based vs Data-driven.Control and Analysis of unknown systems:
Data Model Controller/Analysis
— 2
Motivation
Model-based vs Data-driven.Control and Analysis of unknown systems:
Model-based control
Data Model Controller/Analysis
— 2
Motivation
Model-based vs Data-driven.Control and Analysis of unknown systems:
Model-based controlSystem identification
Data Model Controller/Analysis
— 2
Motivation
Model-based vs Data-driven.Control and Analysis of unknown systems:
Model-based controlSystem identification
Data Model Controller/Analysis
Persistency of excitation
— 2
Motivation
Model-based vs Data-driven.Control and Analysis of unknown systems:
Model-based controlSystem identification
Data Model Controller/Analysis
Persistency of excitation
Data-driven control
— 3
Motivation
The main observation:
Consider the discrete-time, linear time-invariant system:
x(t+ 1) = Asx(t) +Bsu(t).
— 3
Motivation
The main observation:
Consider the discrete-time, linear time-invariant system:
x(t+ 1) = Asx(t) +Bsu(t).
Suppose that As and Bs are unknown, but that we have access to measurementson this system. We want to resolve whether the system (As, Bs) is stabilizable.
— 3
Motivation
The main observation:
Consider the discrete-time, linear time-invariant system:
x(t+ 1) = Asx(t) +Bsu(t).
Suppose that As and Bs are unknown, but that we have access to measurementson this system. We want to resolve whether the system (As, Bs) is stabilizable.We can only conclude that the system is stabilizable if all systems compatiblewith the measurements are stabilizable.
— 4
Motivation
Contents
1. Motivation
2. Informativity Framework
3. Results on informativity for analysis
4. Conclusions
Not in this talk: Data-driven control, controller design, disturbances, noise.
— 5
Informativity Framework
Informativity Framework
— 6
Informativity Framework
The informativity framework for analysis.
Three important parts.
— 6
Informativity Framework
The informativity framework for analysis.
Model Class ΣContains the ‘true’system.
— 6
Informativity Framework
The informativity framework for analysis.
Model Class ΣContains the ‘true’system.
Data DGenerated by the ‘true’system.Set of all systems thatare compatible: ΣD.
— 6
Informativity Framework
The informativity framework for analysis.
Model Class ΣContains the ‘true’system.
Data DGenerated by the ‘true’system.Set of all systems thatare compatible: ΣD.
Property PAny system-theoreticproperty.Set of all systems thathave this property: ΣP .
— 6
Informativity Framework
The informativity framework for analysis.
Model Class ΣContains the ‘true’system.
Data DGenerated by the ‘true’system.Set of all systems thatare compatible: ΣD.
Property PAny system-theoreticproperty.Set of all systems thathave this property: ΣP .
We say that the data D are informative for property P if ΣD ⊆ ΣP .
— 6
Informativity Framework
The informativity framework for analysis.
Model Class ΣContains the ‘true’system.
Data DGenerated by the ‘true’system.Set of all systems thatare compatible: ΣD.
Property PAny system-theoreticproperty.Set of all systems thathave this property: ΣP .
We say that the data D are informative for property P if ΣD ⊆ ΣP .
Informativity problemProvide necessary and sufficient conditions on D under which the data areinformative for property P .
— 7
Informativity Framework
Informativity for stabilizability.
Consider the class of discrete-time, linear time-invariant input/state systems,containing ‘true’ system: x(t+ 1) = Asx(t) +Bsu(t).
— 7
Informativity Framework
Informativity for stabilizability.
Consider the class of discrete-time, linear time-invariant input/state systems,containing ‘true’ system: x(t+ 1) = Asx(t) +Bsu(t).Suppose that we measure the state and input for time 0 up to T .
U− :=[u(0) u(1) · · · u(T − 1)
], X :=
[x(0) x(1) · · · x(T )
],
X− :=[x(0) x(1) · · · x(T − 1)
], X+ :=
[x(1) x(2) · · · x(T )
].
Then the set of all systems compatible with the data is equal to:
ΣD = {(A,B) | X+ = AX− +BU−}
— 7
Informativity Framework
Informativity for stabilizability.
Consider the class of discrete-time, linear time-invariant input/state systems,containing ‘true’ system: x(t+ 1) = Asx(t) +Bsu(t).Suppose that we measure the state and input for time 0 up to T .
U− :=[u(0) u(1) · · · u(T − 1)
], X :=
[x(0) x(1) · · · x(T )
],
X− :=[x(0) x(1) · · · x(T − 1)
], X+ :=
[x(1) x(2) · · · x(T )
].
Then the set of all systems compatible with the data is equal to:
ΣD = {(A,B) | X+ = AX− +BU−}
Consider stabilizability, and let:
ΣP = {(A,B) | (A,B) is stabilizable}
— 8
Informativity Framework
Informativity for stabilizability.
With data (U−, X) as before and sets:
ΣD = {(A,B) | X+ = AX− +BU−}ΣP = {(A,B) | (A,B) is stabilizable.}
Informativity for stabilizability
The data (U−, X) are informative for stabilizability if ΣD ⊆ ΣP .
— 9
Informativity Framework
The informativity framework for control.
Let the ‘true’ system be (As, Bs) and data (U−, X) as before.
— 9
Informativity Framework
The informativity framework for control.
Let the ‘true’ system be (As, Bs) and data (U−, X) as before.
Suppose now that we are interested in stabilization by static state feedback: FindK such that As +BsK is stable.
— 9
Informativity Framework
The informativity framework for control.
Let the ‘true’ system be (As, Bs) and data (U−, X) as before.
Suppose now that we are interested in stabilization by static state feedback: FindK such that As +BsK is stable.
We say the data D is informative for property P(K) if there exists a K such thatΣD ⊆ ΣP(K).
— 9
Informativity Framework
The informativity framework for control.
Let the ‘true’ system be (As, Bs) and data (U−, X) as before.
Suppose now that we are interested in stabilization by static state feedback: FindK such that As +BsK is stable.
We say the data D is informative for property P(K) if there exists a K such thatΣD ⊆ ΣP(K).
Informativity problem for controlProvide necessary and sufficient conditions on D under which there exists a Ksuch that the data are informative for property P(K).
Control design problem
If D are informative for P(K), find K such that ΣD ⊆ ΣP(K).
— 10
Results on informativity for analysis
Results on informativity for analysis
— 11
Results on informativity for analysis
Data-driven Hautus test.Consider the class of discrete-time, linear time-invariant input/state systems,containing ‘true’ system: x(t+ 1) = Asx(t) +Bsu(t).
— 11
Results on informativity for analysis
Data-driven Hautus test.Consider the class of discrete-time, linear time-invariant input/state systems,containing ‘true’ system: x(t+ 1) = Asx(t) +Bsu(t).With data (U−, X) as before and sets:
Σi/s = {(A,B) | X+ = AX− +BU−},Σstab = {(A,B) | (A,B) is stabilizable},Σcont = {(A,B) | (A,B) is controllable}.
— 11
Results on informativity for analysis
Data-driven Hautus test.Consider the class of discrete-time, linear time-invariant input/state systems,containing ‘true’ system: x(t+ 1) = Asx(t) +Bsu(t).
Theorem (Data-driven Hautus tests)
The data (U−, X) are informative for controllability if and only if
rank(X+ − λX−) = n ∀λ ∈ C.
Similarly, the data (U−, X) are informative for stabilizability if and only if
rank(X+ − λX−) = n ∀λ ∈ C with |λ| > 1.
— 12
Results on informativity for analysis
Example
Theorem (Data-driven Hautus test)
The data (U−, X) are informative for controllability if and only if
rank(X+ − λX−) = n ∀λ ∈ C.
— 12
Results on informativity for analysis
Example
Theorem (Data-driven Hautus test)
The data (U−, X) are informative for controllability if and only if
rank(X+ − λX−) = n ∀λ ∈ C.
Let n = 2 and
x0 =
[00
], x1 =
[10
], x2 =
[01
], u0 = 1, u1 = 0.
— 12
Results on informativity for analysis
Example
Theorem (Data-driven Hautus test)
The data (U−, X) are informative for controllability if and only if
rank(X+ − λX−) = n ∀λ ∈ C.
Let n = 2 and
x0 =
[00
], x1 =
[10
], x2 =
[01
], u0 = 1, u1 = 0.
This implies that
X+ =
[1 00 1
], X− =
[0 10 0
], =⇒ X+ − λX− =
[1 −λ0 1
]
— 12
Results on informativity for analysis
Example
Theorem (Data-driven Hautus test)
The data (U−, X) are informative for controllability if and only if
rank(X+ − λX−) = n ∀λ ∈ C.
Let n = 2 and
x0 =
[00
], x1 =
[10
], x2 =
[01
], u0 = 1, u1 = 0.
Indeed:
Σi/s =
{([0 a11 a2
],
[10
])| a1, a2 ∈ R
}
— 13
Results on informativity for analysis
Observability
Given C, consider the class of autonomous discrete-time, linear time-invariantsystems, containing ‘true’ system: x(t+ 1) = Asx(t), y(t) = Cx(t).
— 13
Results on informativity for analysis
Observability
Given C, consider the class of autonomous discrete-time, linear time-invariantsystems, containing ‘true’ system: x(t+ 1) = Asx(t), y(t) = Cx(t).With state measurements X as before we have sets:
Σs = {A | X+ = AX−},Σstab = {A | (C,A) is observable},Σcont = {A | (C,A) is detectable}.
— 13
Results on informativity for analysis
Observability
Given C, consider the class of autonomous discrete-time, linear time-invariantsystems, containing ‘true’ system: x(t+ 1) = Asx(t), y(t) = Cx(t).
Theorem (Data-driven Hautus tests)
The data X are informative for observability with C if and only if kerC ⊆ imX−and for all λ ∈ C and v ∈ Cn we have(
X+ − λX−CX−
)v =⇒ X−v = 0.
The data X are informative for detectability with C if and only if kerC ⊆ imX−and for all λ ∈ C with |λ| > 1 and v ∈ Cn we have(
X+ − λX−CX−
)v =⇒ X−v = 0.
— 14
Results on informativity for analysis
Example
Let C =(1 0 0
)and suppose that X− =
0 00 11 0
, X+ =
0 11 00 0
.
— 14
Results on informativity for analysis
Example
Let C =(1 0 0
)and suppose that X− =
0 00 11 0
, X+ =
0 11 00 0
.
Theorem (Data-driven Hautus tests)
The data X are informative for observability with C if and only if kerC ⊆ imX−
and for all λ ∈ C and v ∈ Cn we have
(X+ − λX−CX−
)v =⇒ X−v = 0.
— 14
Results on informativity for analysis
Example
Let C =(1 0 0
)and suppose that X− =
0 00 11 0
, X+ =
0 11 00 0
.
Theorem (Data-driven Hautus tests)
The data X are informative for observability with C if and only if kerC ⊆ imX−
and for all λ ∈ C and v ∈ Cn we have
(X+ − λX−CX−
)v =⇒ X−v = 0.
Indeed kerC ⊆ imX− and
(X+ − λX−CX−
)=
0 11 −λ−λ 00 0
.
— 15
Conclusions
Conclusions
— 16
Conclusions
Conclusions
> We have introduced a framework for dealing with informativity problems.
> We resolved analysis problems regarding controllability, stabilizability,observability and detectability.
> Some examples illustrated these results.
— 16
Conclusions
Conclusions
> We have introduced a framework for dealing with informativity problems.
> We resolved analysis problems regarding controllability, stabilizability,observability and detectability.
> Some examples illustrated these results.
— 16
Conclusions
Conclusions
> We have introduced a framework for dealing with informativity problems.
> We resolved analysis problems regarding controllability, stabilizability,observability and detectability.
> Some examples illustrated these results.
— 16
Conclusions
Conclusions
> We have introduced a framework for dealing with informativity problems.
> We resolved analysis problems regarding controllability, stabilizability,observability and detectability.
> Some examples illustrated these results.
Based on: H. J. Van Waarde, J. Eising, H. L. Trentelman, and M. K. Camlibel,“Data informativity: a new perspective on data-driven analysis and control,”IEEE Transactions on Automatic Control.
— 17
Conclusions
Extensions
> Data-driven control.
> Different model classes:
• Nonlinear systems.
• Disturbances.
• Parametrization (e.g. network structure).
> Different data:
• Noise.
• Input/output measurements
> Different properties.
— 18
Conclusions
Questions?