Beyond LTI SI

LTI SI
Problem
The 3-step strategy
Lecture 9b: Estimating the Kalman Filter from
data

Michel Verhaegen 1

Delft Center for Systems and Control

Delft University of Technology
the Netherlands

14th December 2016

1
Based on the work: [1] A. Chiuso. The Role of Vector Autoregressive
Modeling in Predictor Based Subspace Identification. Automatica, vol.
43pp. 1034–1048, 2007
Michel Verhaegen 14th December 2016 1/8
 Contents

Beyond LTI SI

LTI SI
Problem
The 3-step strategy

1 Predictor based Subspace Identification

Problem Definition
The 3-step strategy

Michel Verhaegen 14th December 2016 2/8

Predictor Based Subspace
Identification [1]
 Problem Definition

Beyond LTI SI (
x(k + 1) = Ax(k) + Bu(k) + Ke(k)
LTI SI Σ:
Problem y(k) = Cx(k) + e(k)
The 3-step strategy
e(k)

u(k) Σ y(k)

Michel Verhaegen 14th December 2016 4/8

 Problem Definition

Beyond LTI SI (
x(k + 1) = Ax(k) + Bu(k) + Ke(k)
LTI SI Σ:
Problem y(k) = Cx(k) + e(k)
The 3-step strategy
e(k)

u(k) Σ y(k)

Identification Problem
Given i/o data sequences {u(k), y(k)}N k=1 from an unknown finite
dimensional, linear time-invariant system, determine the system matrices:
 
AT , BT , CT , D, KT

and the system order n such that they are a consistent estimate of the
orginal system matrices up to a similarity transformation.

Michel Verhaegen 14th December 2016 4/8

 The “Predictor” Model

Beyond LTI SI (  
b : x̂(k + 1) = Ax̂(k) + Bu(k) + K y(k) − C x̂(k)
LTI SI Σ
Problem
The 3-step strategy
y(k) = C x̂(k) + e(k)
y(k)

u(k) b
Σ ŷ(k) = C x̂(k)

Identification Problem
Given i/o data sequences {u(k), y(k)}N k=1 from an unknown finite
dimensional, linear time-invariant system, determine the system matrices:
 
AT , BT , CT , D, KT

and the system order n such that they are a consistent estimate of the
orginal system matrices up to a similarity transformation.

Michel Verhaegen 14th December 2016 5/8

 What is needed to Estimate the state sequence?

Beyond LTI SI

Lemma 1
LTI SI
Problem Let the LTI system be given as
The 3-step strategy

 x̂(k + 1) = (A − KC) x̂(k) + (B − KD) u(k) + Ky(k)
| {z } | {z }
Φ
 B̃
y(k) = C x̂(k) + Du(k) + e(k)
 T
and define the joint input-output data vector z(t) = u(t)T y(t)T ,
then
 
z(t − p)
  z(t − p + 1)
Φ p−1
[B̃, K] Φ p−2
[B̃, K] ··· [B̃, K]  
x̂(t) = Φp x̂(t−p)+ | {z }  .. 
Bz
 .

z(t − 1)
| {z }
Zt−p,p,1

Michel Verhaegen 14th December 2016 6/8

 What is needed to Estimate the state sequence?

Beyond LTI SI

LTI SI
Corollary 1
Problem
The 3-step strategy
The (A) matrix with the required state sequence as its row space is given
by:
 
C
 CΦ 
 .  X̂t,N = Of Φp X̂t−p,N +
 .. 
CΦf −1
| {z }
Of
 
CΦp−1 Bz CΦp−2 Bz ··· CBz
 
 CΦp Bz CΦp−1 Bz ··· CΦBz 
  Zt−p,p,N
 .. .. .. 
 . . . 
CΦp+f −1 Bz ··· CΦ f −1
Bz

Michel Verhaegen 14th December 2016 6/8

 What is needed to Estimate the state sequence?

Beyond LTI SI

Subspace “trick” 1
LTI SI
Problem
The 3-step strategy The matrix Φ is A.S. and therefore: Φℓ ≈ 0 for ℓ ≥ p ⇒
 
C
 CΦ 
 .  X̂t,N ≈ Of Φp X̂t−p,N +
 .. 
CΦf −1
| {z }
Of

 
CΦp−1 Bz CΦp−2 Bz ··· CBz
 
 CΦp Bz CΦp−1 Bz ··· CΦBz 
  Zt−p,p,N
 .. .. .. 
 . . . 
CΦp+f −1 Bz ··· CΦ f −1
Bz

Michel Verhaegen 14th December 2016 6/8

 What is needed to Estimate the state sequence?

Beyond LTI SI
Subspace “trick” 2: From Lemma 1
LTI SI Let the LTI system be given as
Problem
The 3-step strategy 
 x̂(k + 1) = (A − KC) x̂(k) + (B − KD) u(k) + Ky(k)
| {z } | {z }
Φ
 B̃
y(k) = C x̂(k) + Du(k) + e(k)

then
h i
y(t) = CΦp x̂(t−p)+ CΦp−1 Bz CΦp−2 Bz ··· CBz Zt−p,p,1 +Du(t)

With “trick” 1:
h i
y(t) ≈ CΦp x̂(t − p)+ CΦp−1 Bz CΦp−2 Bz ··· CBz Zt−p,p,1 +Du(t)

The system parameters CΦj Bz for j = 0, · · · , p − 1 can be (accurately)

approximated via solving a linear least squares problem (step 1!)
Michel Verhaegen 14th December 2016 6/8
 What is needed to Estimate the state sequence?

Beyond LTI SI

LTI SI Subspace “trick” 1 and 2 combined

Problem
The 3-step strategy
The matrix Φ is A.S. and therefore: Φℓ ≈ 0 for ℓ ≥ p ⇒
 
C
 CΦ 
 .  X̂t,N ≈ Of Φp X̂t−p,N +
 .. 
CΦf −1
| {z }
Of

 
CΦp−1 Bz CΦp−2 Bz ··· CBz
 CΦp Bz CΦp−1 Bz ··· CΦBz 
 .. .. ..  Zt−p,p,N
 . . .

CΦp+f −1 Bz ··· CΦf −1 Bz

Michel Verhaegen 14th December 2016 6/8

 Summary

Beyond LTI SI

LTI SI
" #
Problem u(t)
The 3-step strategy With the given data define z(t) =
{u(t), y(t)}···
t=0
y(t)
construct the “compound” Hankel matrix:
 
z(0) z(1) · · · z(N − 1)
 
 z(1) z(2) · · · z(N ) 
Z0,s,N =
 .. .. .. 

 . . . 
z(s − 1) ··· z(N + s − 2)
h i
Ys,1,N = y(s) y(s + 1) · · · y(N + s − 1)

Michel Verhaegen 14th December 2016 7/8

 Summary

Beyond LTI SI Step 1: Solve linear least squares problem:

 
LTI SI min kYs,1,N − ΘZ0,s,N k2F \
Θ̂ = CΦ p−1 B
z
\
CΦ p−2 B
z ··· [z
CB
Problem
Θ
The 3-step strategy
Step 2: Estimate KF state sequence via:
 
\
CΦ p−1 B
z
\
CΦ p−2 B
z ··· [z
CB
  CΦp Bz \
CΦ p−1 B ··· \z 
CΦB 
 z  T
svd  .. .. ..  Zt−p,p,N ≈ Un Σn Vn
 . . . 
CΦp+f −1 Bz ··· \
CΦ f −1 B
z

h i
x̂T (s) x̂T (s + 1) · · · x̂T (s + N − 1) = VnT
Step 3: Estimate the system matrices:
    "x̂T (s) · · · x̂T (s + N −
x̂T (s + 1) · · · x̂T (s + N − 1) Φ B K
min k − u(s) · · · u(s + N − 2
Φ,B,K,C,D y(s) · · · y(s + N − 2) C D 0
y(s) · · · y(s + N − 2

Michel Verhaegen 14th December 2016 8/8