Filtering and Identification

Lecture 5:
Subspace Identification
The Deterministic case

Michel Verhaegen and Jan-Willem van Wingerden

1/31

Delft Center for Systems and Control

Delft University of Technology
Overview
• Recap course outline
• Recap “classical” system identification
• A different approach: Subspace Identification
• A deterministic Identification problem
• Subspace Identification of an autonomous
system
• Subspace Identification with inputs

2/31

Delft Center for Systems and Control

Recap “goal” of the course
• Estimate linear static models from data:

min ky − F xk2W
x

Relevant and to introduce the numerical fundaments of this course!

• Estimate the state of an LTI state space model:

x(k + 1) = Ax(k) + Bu(k) + Ke(k) e(k) ∼ (0, Re )

y(k) = Cx(k) + Du(k) + e(k)

given the model (A, B, C, D) and K and input-output data.

• “Identify” state space model matrices (A, B, C, D, K) from
input-output data {u(k), y(k)}N
k=1 .

3/31

Delft Center for Systems and Control

Recap “goal” of the course
• Estimate linear static models from data:

min ky − F xk2W
x

Relevant and to introduce the numerical fundaments of this course!

• Estimate the state of an LTI state space model:

x(k + 1) = Ax(k) + Bu(k) + Ke(k) e(k) ∼ (0, Re )

y(k) = Cx(k) + Du(k) + e(k)

given the model (A, B, C, D) and K and input-output data.

• “Identify” state space model matrices (A, B, C, D, K) from
input-output data {u(k), y(k)}N
k=1 .

4/31

Delft Center for Systems and Control

Overview
• Recap course outline
• Recap “classical” system identification
• A different approach: Subspace Identification
• A deterministic Identification problem
• Subspace Identification of an autonomous
system
• Subspace Identification with inputs

5/31

Delft Center for Systems and Control

Recap “Classical” (PEM) Identification methods
w(k)
SGM v(k)
u(k) y(k)

+ ǫ(k, θ)
−
A(θ) B(θ) K(θ)
C(θ)D(θ) ŷ(k, θ)

\
1. Assume SGM(θ)
2. Derive the “optimal” predictor for
ŷ(k, θ).
3. Find the “best” estimate
θ̂N = argminJN (θ) with
1 X
N
with JN (θ) = ky(k) − ŷ(k, θ)k22
N k=1

6/31

Delft Center for Systems and Control

Recap “Classical” (PEM) Identification methods
w(k)
SGM v(k)
Challenges:
u(k) y(k) 1. You need to fix the order of the SS
model a priori!
+ ǫ(k, θ)
− 2. (MIMO) parametrization is a diffi-
A(θ) B(θ) K(θ) cult?
C(θ)D(θ) ŷ(k, θ)
3. non-convex optimization
100
\
1. Assume SGM(θ) 90 53.1 %

2. Derive the “optimal” predictor for 80

70
ŷ(k, θ). 60

3. Find the “best” estimate 50

40
θ̂N = argminJN (θ) with
30

1 X
N 20

with JN (θ) = ky(k) − ŷ(k, θ)k22 10

N k=1 0
0 20 40 60 80 100

6/31

Delft Center for Systems and Control

Example: Spatial Discretization of vibrating membrame

Approximationvia “multiple” mass-spring-damper

  systems:

0 I 0
ẋ(t) =   x(t) +   u(t)
−M −1 B −M −1 K −M −1

7/31

Delft Center for Systems and Control

Example: Spatial Discretization of vibrating membrame

Approximationvia “multiple” mass-spring-damper

  systems:

0 I 0
ẋ(t) =   x(t) +   u(t)
−M −1 B −M −1 K −M −1

Multiple key problems:

1. Number of mass-spring-damper segments?
2. Parametrization of the “mass”, “stiffness” and “damping”
matrices. Following “The Timoshenko”, “The
Bernoulli-Euler”, etc. beam theories?

7/31

Delft Center for Systems and Control

Overview
• Recap course outline
• Recap “classical” system identification
• A different approach: Subspace Identification
• A deterministic Identification problem
• Subspace Identification of an autonomous
system
• Subspace Identification with inputs

8/31

Delft Center for Systems and Control

Deterministic Identification problem

SGM:
(
x(k + 1) = Ax(k) + Bu(k)
x(k) ∈ Rn
y(k) = Cx(k) + Du(k)

Determine n AND the system matrices A, B, C, and D from a

finite number of measurements of u(k) and y(k) without
parametrizing the system matrices, etc.

9/31

Delft Center for Systems and Control

Deterministic Identification problem

SGM:
(
x(k + 1) = Ax(k) + Bu(k)
x(k) ∈ Rn
y(k) = Cx(k) + Du(k)

Determine n AND the system matrices A, B, C, and D from a

finite number of measurements of u(k) and y(k) without
parametrizing the system matrices, etc.

In general no unique solution!

9/31

Delft Center for Systems and Control

Overview
• Recap “classical” system identification
• A different approach: Subspace Identification
• A deterministic Identification problem
• Subspace Identification of an
autonomous system
• Subspace Identification with inputs

10/31

Delft Center for Systems and Control

Identification of an autonomous system

SGM: Autonomous system

(
x(k + 1) = Ax(k)
x(k) ∈ Rn
y(k) = Cx(k)

Output: y(j) = CAj x(0) j =0:s−1

11/31

Delft Center for Systems and Control

SIDemo.m
For a second order system (n = 2), we display:
     
y(0) y(1) y(2)
     
y(1) y(2) y(3)
y(2) y(3) y(4)

in R3 . What do we see?

12/31

Delft Center for Systems and Control

Identification of an autonomous system
SGM: Autonomous system
(
x(k + 1) = Ax(k)
x(k) ∈ Rn
y(k) = Cx(k)

Output: y(k + j) = CAj x(k) j =0:s−1

   
y(k) C
   
   
 y(k + 1)   CA 
   

y(k + 2)
 
= CA2 x(k) = O x(k)
    s
   
 ..   .. 
 .   . 
   
y(k + s − 1) CAs−1

13/31

Delft Center for Systems and Control

SubidPlot.m
For the 2nd order system:
" #
0.9 0
x(k + 1) = x(k)
−0.5 0.7
h i
y(k) = 1 −1 x(k)
 
y(k)
 
we plot the vector y(k + 1) for k = 1 : 10 in R3 .
y(k + 2)
What do we see?
14/31

Delft Center for Systems and Control

Revealing information
We take y(k) ∈ R, n = 2, s = 3.
 
y(k)
 
y(k + 1) = Os x(k)
y(k + 2)

Plot for k = 0, 1, 2, . . .

15/31

Delft Center for Systems and Control

Revealing information
We take y(k) ∈ R, n = 2, s = 3.
 
y(k)
 
y(k + 1) = Os x(k)
y(k + 2)

Plot for k = 0, 1, 2, . . .
   
y(0) y(1) · · · y(N − 2) C
   h i
y(1) y(2) · · ·
 y(N − 1) =
 
 CA  x(0) x(1) · · · x(N − 2)

y(2) y(3) · · · y(N ) CA2
15/31

Delft Center for Systems and Control

Data equation: 2nd-order example
   
y(0) y(1) · · · y(N − 2) C
   h i
y(1) y(2) · · · y(N − 1) =  CA  x(0) x(1) · · · x(N − 2)
   
2
| {z }
y(2) y(3) · · · y(N ) CA X0,N −1
| {z } | {z }
Y0,3,N −1 O3

Use Sylvester’s inequality

 
rank(O3 )+rank(X0,N −1 )−2 ≤ rank(Y0,3,N −1 ) ≤ min rank(O3 ), rank(X0,N −1 )

16/31

Delft Center for Systems and Control

Data equation: 2nd-order example
   
y(0) y(1) · · · y(N − 2) C
   h i
y(1) y(2) · · · y(N − 1) =  CA  x(0) x(1) · · · x(N − 2)
   
2
| {z }
y(2) y(3) · · · y(N ) CA X0,N −1
| {z } | {z }
Y0,3,N −1 O3

Use Sylvester’s inequality

 
rank(O3 )+rank(X0,N −1 )−2 ≤ rank(Y0,3,N −1 ) ≤ min rank(O3 ), rank(X0,N −1 )

If (A, C) is observable and (A, x(0)) is controllable:

rank(Y0,3,N −1 ) = 2 = n
16/31

Delft Center for Systems and Control

SubidPlot.m
Let the SVD of the matrix Y0,3,N −1 be given as:
" #
h i Σ 0 h i
⊥ 2
Y0,3,N −1 = U2 U2 (V2 )T (V2⊥ )T
0 0

with Σ2 ∈ R2×2 and rank(Σ2 ) = 2.

Illustration of the column space equivalence
between U2 and O3 .
Open File: SubidPlot.m and inspect the final 2
lines! minT kO3 − U2 T k2F

17/31

Delft Center for Systems and Control

Compute the matrices A and C
   
CT CT
   

 CT T −1 AT   
 CT AT 

Un = Os T =  ..  =  . 
 .   .. 
   
s−1
−1
CT (T AT ) CT ATs−1

18/31

Delft Center for Systems and Control

Compute the matrices A and C
   
CT CT
   

 CT T −1 AT   
 CT AT 

Un = Os T =  ..  =  . 
 .   .. 
   
s−1
−1
CT (T AT ) CT ATs−1

We can determine A and C up to an unknown

similarity transformation T .

18/31

Delft Center for Systems and Control

Compute the matrices A and C
   
CT CT
   

 CT T −1 AT   
 CT AT 

Un = Os T =  ..  =  . 
 .   .. 
   
s−1
−1
CT (T AT ) CT ATs−1

We can determine A and C up to an unknown

similarity transformation T .

CT equals the first ℓ rows of Un : CT = Un (1 : ℓ, :)

AT is computed by solving:
Un (1 : (s − 1)ℓ, :)AT = Un (ℓ + 1 : sℓ, :)
18/31

Delft Center for Systems and Control

Zoom in: Computation of the A matrix
 
As Un is equal to, CT
 
 CT AT 
 
Un =  . 
 .. 
 
CT ATs−1

Therefore, we can find AT by,

   
CT CT AT
   
 CT AT   CT A2 
   T 
Un (1 : (s−1)ℓ, :)AT =  .  AT =  .  = Un (ℓ+1 : sℓ, :)
 ..   .. 
   
CT ATs−2 CT ATs−1
19/31

Delft Center for Systems and Control

A systematic Approach
• Step
 1:Data stacking into Hankel matrices
Y0,s,N .

20/31

Delft Center for Systems and Control

A systematic Approach
• Step
 1:Data stacking into Hankel matrices
Y0,s,N .
• Step 2: The Data Equation relating these
Hankel
 matrices 
to the state space matrices.
Y0,s,N = Os X0,N .

20/31

Delft Center for Systems and Control

A systematic Approach
• Step
 1:Data stacking into Hankel matrices
Y0,s,N .
• Step 2: The Data Equation relating these
Hankel
 matrices 
to the state space matrices.
Y0,s,N = Os X0,N .
• Step 3: Retrieving the column space of Os .

20/31

Delft Center for Systems and Control

A systematic Approach
• Step
 1:Data stacking into Hankel matrices
Y0,s,N .
• Step 2: The Data Equation relating these
Hankel
 matrices 
to the state space matrices.
Y0,s,N = Os X0,N .
• Step 3: Retrieving the column space of Os .
• Step 4: Calculating the system matrices
(AT , CT )

20/31

Delft Center for Systems and Control

Overview
• Recap “classical” system identification
• A different approach: Subspace Identification
• A deterministic Identification problem
• Subspace Identification of an autonomous
system
• Subspace Identification with inputs

21/31

Delft Center for Systems and Control

Identification with inputs
(
x(k + 1) = Ax(k) + Bu(k)
SGM
y(k) = Cx(k) + Du(k)

22/31

Delft Center for Systems and Control

Identification with inputs
(
x(k + 1) = Ax(k) + Bu(k)
SGM
y(k) = Cx(k) + Du(k)

x(1) = Ax(0) + Bu(0)

x(2) = Ax(1) + Bu(1)
= A2 x(0) + ABu(0) + Bu(1)
x(3) = A3 x(0) + A2 Bu(0) + ABu(1) + Bu(2)

22/31

Delft Center for Systems and Control

Identification with inputs
(
x(k + 1) = Ax(k) + Bu(k)
SGM
y(k) = Cx(k) + Du(k)

x(1) = Ax(0) + Bu(0)

x(2) = Ax(1) + Bu(1)
= A2 x(0) + ABu(0) + Bu(1)
x(3) = A3 x(0) + A2 Bu(0) + ABu(1) + Bu(2)
k−1
X
State at time k: x(k) = Ak x(0) + Ak−i−1 Bu(i)
i=0
22/31

Delft Center for Systems and Control

 Data equation with inputs
Output at time k:
k−1
X
y(k) = CAk x(0) + CAk−i−1 Bu(i) + Du(k)
i=0
      
y(0) C D 0 0 ··· 0 u(0)
      
      
 y(1)   CA   CB D 0 · · · 0  u(1) 
      
   2    
 y(2)  =  CA x(0)+ CAB CB D 0  u(2) 
      
 ..   ..   .. .. .. .
 . 
 .   .   . . .  . 
      
y(s − 1) CAs−1 CAs−2 B CAs−3 B · · · CB D u(s − 1)
| {z } | {z }
Os Ts
23/31

Delft Center for Systems and Control

 Data equation with inputs
Output at time k:
k−1
X
y(k) = CAk−j x(j) + CAk−i−1 Bu(i) + Du(k)
i=j
      
y(1) C D 0 0 · · · 0 u(1)
      
      
y(2)  CA   CB D 0 · · · 0 u(2)
      
   2    
y(3) =  CA x(1) +  CAB CB D 0 u(3)
      
 ..   ..   .. .. .. .
 . 
 .   .   . . .  . 
      
y(s) CAs−1 CAs−2 B CAs−3 B · · · CB D u(s)
| {z } | {z }
Os Ts
24/31

Delft Center for Systems and Control

 Data equation with inputs
Output at time k:
k−1
X
y(k) = CAk−j x(j) + CAk−i−1 Bu(i) + Du(k)
i=j
     
y(k) C D 0 0 ··· 0 u(k
     
     
 y(k + 1)   CA   CB D 0 · · · 0  u(k +
     
   2   
 y(k + 2)  =  CA x(k)+ CAB CB D 0  u(k +
     
 ..   ..   .. .. ..  ..
 .   .   . . .  .
     
y(k + s − 1) CAs−1 CAs−2 B CAs−3 B · · · CB D u(k + s
| {z } | {z }
Os Ts
25/31

Delft Center for Systems and Control

Data equation with inputs
Data equation Y0,s,N = Os X0,N + Ts U0,s,N

26/31

Delft Center for Systems and Control

Data equation with inputs
Data equation Y0,s,N = Os X0,N + Ts U0,s,N

Step 3 of the Strategy:

How to estimate the column space of Os ?

26/31

Delft Center for Systems and Control

Data equation with inputs
Data equation Y0,s,N = Os X0,N + Ts U0,s,N

Step 3 of the Strategy:

How to estimate the column space of Os ?
Step 4 of the Strategy: Estimate the system
matrices?
• Determine A and C (up to a similarity
transformation) from the column space of Os .
• Given A and C determine B and D from the
input and output measurements.
26/31

Delft Center for Systems and Control

Removing U0,s,N from Y0,s,N = Os X0,N + Ts U0,s,N

27/31

Delft Center for Systems and Control

Removing U0,s,N from Y0,s,N = Os X0,N + Ts U0,s,N

Can we transform the data equation, i.e. can we

find a matrix (of appropriate dimensions) Π such
that:
Y0,s,N Π = Os X0,N Π + Ts U0,s,N Π
is “transformed” into
Y0,s,N Π = Os X0,N Π

27/31

Delft Center for Systems and Control

Removing U0,s,N from Y0,s,N = Os X0,N + Ts U0,s,N

Can we transform the data equation, i.e. can we

find a matrix (of appropriate dimensions) Π such
that:
Y0,s,N Π = Os X0,N Π + Ts U0,s,N Π
is “transformed” into
Y0,s,N Π = Os X0,N Π
Example Matlab: U = randn(2,5);
Pu = (eye(5) - U’*inv(U*U’)*U);
U Pu? What is Pu - Pu’?

27/31

Delft Center for Systems and Control

Removing U0,s,N from Y0,s,N = Os X0,N + Ts U0,s,N

28/31

Delft Center for Systems and Control

Removing U0,s,N from Y0,s,N = Os X0,N + Ts U0,s,N
Let Π⊥ U0,s,N be the Projection matrix onto the
orthogonal complement of the row space of
U0,s,N . This matrix is given as:
 
Π⊥
U0,s,N = I N −U T
0,s,N (U U T
0,s,N 0,s,N ) −1
U 0,s,N ∈ R N ×N

Then: U0,s,N Π⊥
U0,s,N = 0.

28/31

Delft Center for Systems and Control

Removing U0,s,N from Y0,s,N = Os X0,N + Ts U0,s,N
Let Π⊥ U0,s,N be the Projection matrix onto the
orthogonal complement of the row space of
U0,s,N . This matrix is given as:
 
Π⊥
U0,s,N = I N −U T
0,s,N (U U T
0,s,N 0,s,N ) −1
U 0,s,N ∈ R N ×N

Then: U0,s,N Π⊥
U0,s,N = 0.

⊥ ⊥
Y0,s,N ΠU0,s,N = (Os X0,N + Ts U0,s,N ) ΠU0,s,N
= Os X0,N Π⊥
U0,s,N

28/31

Delft Center for Systems and Control

Dealing with inputs
Y0,s,N Π⊥
U0,s,N = O X Π ⊥
s 0,N U0,s,N

range(Y0,s,N Π⊥
U0,s,N ) ⊆ range(Os )

29/31

Delft Center for Systems and Control

Dealing with inputs
Y0,s,N Π⊥
U0,s,N = O X Π ⊥
s 0,N U0,s,N

range(Y0,s,N Π⊥
U0,s,N ) ⊆ range(Os )

Lemma:  
X0,N
If u(k) is such that rank   = n + sm
U0,s,N
 
and (A, C) observable ⇒ rank Y0,s,N Π⊥
U0,s,N = n

range(Y0,s,N Π⊥
U0,s,N ) = range(Os ) ⇒ AT , CT

29/31

Delft Center for Systems and Control

Determine B and D
k−1
X
y(k) = CT AkT x(0) + CT Ak−i−1
T BT u(i) + DT u(k)
i=0

Given AT and CT , the output depends affinely on BT , DT and x(0).

30/31

Delft Center for Systems and Control

Determine B and D
k−1
X
y(k) = CT AkT x(0) + CT Ak−i−1
T BT u(i) + DT u(k)
i=0

Given AT and CT , the output depends affinely on BT , DT and x(0).

Using property (§2.3) vec(XY Z) = (Z T ⊗ X)vec(Y ) ,

k−1
X
y(k) = CT AkT x(0) + (u(i)T ⊗ CT Ak−i−1
T ) vec(B T ) + (u(k) T
⊗ I)vec(D
i=0
h iT
= F (k, AT , CT )θ θ = x(0)T vec(BT )T vec(DT )T

Find θ by solving a Linear least squares problem!

N −1
1 X
min ||y(k) − F (k, AT , CT )T θ||2
θ N
k=0
30/31

Delft Center for Systems and Control

Preparation and next lecture
Preparation:
Study Chapter 9 (9.1-9.2)
Download Homework 4 and Matlab Session 1

Next lecture:
Lecture 6: Subspace Identification Step Inputs and Ambient
Excitation
Keep your eyes focussed on “Guide-
lines/Rules/Schedule sc4040: 2016-2017” for the
correct deadlines!

31/31

Delft Center for Systems and Control