linear dynamical systems with inputs & outputsee263.stanford.edu/lectures/linsys.pdf ·...
TRANSCRIPT
EE263 Autumn 2015 S. Boyd and S. Lall
Linear dynamical systems with inputs & outputs
I inputs & outputs: interpretations
I impulse and step responses
I examples
1
Inputs & outputs
recall continuous-time time-invariant LDS has form
x = Ax+Bu, y = Cx+Du
I Ax is called the drift term (of x)
I Bu is called the input term (of x)
picture, with B ∈ R2×1:
x(t)Ax(t)
x(t) (with u(t) = 1)
x(t) (with u(t) = −1.5)B
2
Interpretations
write x = Ax+ b1u1 + · · ·+ bmum, where B =[b1 · · · bm
]I state derivative is sum of autonomous term (Ax) and one term per input
(biui)
I each input ui gives another degree of freedom for x (assuming columns of Bindependent)
write x = Ax+Bu as xi = aTi x+ bTi u, where aTi , bTi are the rows of A, B
I ith state derivative is linear function of state x and input u
3
Response to input
I the solution to x = Ax+Bu is
x(t) = etAx(0) +
∫ t
0
e(t−τ)ABu(τ) dτ
I etAx(0) is the unforced or autonomous response
I etAB is called the input-to-state impulse response or impulse matrix
4
Impulse response
impulse response h(t) = CetAB +Dδ(t)
with x(0) = 0, y = h ∗ u, i.e.,
yi(t) =
m∑j=1
∫ t
0
hij(t− τ)uj(τ) dτ
interpretations:
I hij(t) is impulse response from jth input to ith output
I hij(t) gives yi(t) when u(t) = ejδ(t)
I hij(τ) shows how dependent output i is, on what input j was, τ seconds ago
I i indexes output; j indexes input; τ indexes time lag
5
Step response
the step response or step matrix is given by
s(t) =
∫ t
0
h(τ) dτ
interpretations:
I sij(t) is step response from jth input to ith output
I sij(t) gives yi when u = ej for t ≥ 0
for invertible A, we have
s(t) = CA−1(etA − I
)B +D
6
Example 1
u1 u2u1 u2
I unit masses, springs, dampers
I u1 is tension between 1st & 2nd masses
I u2 is tension between 2nd & 3rd masses
I y ∈ R3 is displacement of masses 1,2,3
I x =
[yy
]
7
Example 1
system is:
x =
0 0 0 1 0 00 0 0 0 1 00 0 0 0 0 1−2 1 0 −2 1 01 −2 1 1 −2 10 1 −2 0 1 −2
x+
0 00 00 01 0−1 10 −1
[u1
u2
]
eigenvalues of A are
−1.71± i0.71, −1.00± i1.00, −0.29± i0.71
8
Example 1: Impulse response
0 5 10 15
−0.2
−0.1
0
0.1
0.2
0 5 10 15
−0.2
−0.1
0
0.1
0.2
t
t
h11
h21
h31
h12
h22
h32
impulse response from u1
impulse response from u2
roughly speaking:
I impulse at u1 affects third mass less than other two
I impulse at u2 affects first mass later than other two
9
Example 2: Circuit
u
C3 C4
C1 C2
I u(t) ∈ R is input (drive) voltage
I xi is voltage across Ci
I output is state: y = x
I unit resistors, unit capacitors
I step response matrix shows delay to each node
10
Example 2: Circuit
system is
x =
−3 1 1 01 −1 0 01 0 −2 10 0 1 −1
x+
1000
u, y = x
eigenvalues of A are
−0.17, −0.66, −2.21, −3.96
11
Example 2: Circuit
step response matrix s(t) ∈ R4×1:
0 5 10 150
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
s1
s2
s3s4
I shortest delay to x1; longest delay to x4
I delays consistent with slowest (i.e., dominant) eigenvalue −0.17
12
DC or static gain matrix
I DC gain describes system under static conditions, i.e., x, u, y constant:
0 = x = Ax+Bu, y = Cx+Du
eliminate x to get y = H0u where
H0 = −CA−1B +D
I if system is stable,
H0 =
∫ ∞0
h(t) dt = limt→∞
s(t)
if u(t)→ u∞ ∈ Rm, then y(t)→ y∞ ∈ Rp where y∞ = H0u∞
13
DC gain matrix
DC gain matrix for example 1 (springs):
H(0) =
1/4 1/4−1/2 1/2−1/4 −1/4
DC gain matrix for example 2 (RC circuit):
H(0) =
1111
(do these make sense?)
14
Discretization with piecewise constant inputs
linear system x = Ax+Bu, y = Cx+Du
suppose ud : Z+ → Rm is a sequence, and
u(t) = ud(k) for kh ≤ t < (k + 1)h, k = 0, 1, . . .
define sequences
xd(k) = x(kh), yd(k) = y(kh), k = 0, 1, . . .
I h > 0 is called the sample interval (for x and y) or update interval (for u)
I u is piecewise constant (called zero-order-hold)
I xd, yd are sampled versions of x, y
15
Discretization with piecewise constant inputs
xd(k + 1) = x((k + 1)h)
= ehAx(kh) +
∫ h
0
eτABu((k + 1)h− τ) dτ
= ehAxd(k) +
(∫ h
0
eτA dτ
)B ud(k)
xd, ud, and yd satisfy discrete-time LDS equations
xd(k + 1) = Adxd(k) +Bdud(k), yd(k) = Cdxd(k) +Ddud(k)
where
Ad = ehA, Bd =
(∫ h
0
eτA dτ
)B, Cd = C, Dd = D
called discretized system. If A is invertible, we can express integral as∫ h
0
eτA dτ = A−1(ehA − I
)16
Example: Force on mass
Newton’s law gives continuous-time LDS
x(t) =
[0 10 0
]x(t) +
[01
]u(t)
y(t) =[1 0
]x(t)
let’s compute the discretization
Ad = eAh
= I +Ah+1
2A2h2 + · · ·
= I +Ah
=
[1 h0 1
]17
Example: Force on mass
Bd =
∫ h
0
eAsB ds
=
∫ h
0
[1 s0 1
] [01
]ds
=
∫ h
0
[s1
]ds =
[12h2
h
]
so the discretization is
xd(k + 1) =
[1 h0 1
]xd(k) +
[12h2
h
]ud(k)
yd(k) =[1 0
]xd(k)
18
Stability of discretization
stability: if eigenvalues ofA are λ1, . . . , λn, then eigenvalues ofAd are ehλ1 , . . . , ehλn
discretization preserves stability properties since
<λi < 0 ⇔ |ehλi | < 1
for h > 0
19
Extensions and variations
I offsets: updates for u and sampling of x, y are offset in time
I multirate: ui updated, yi sampled at different intervals
(usually integer multiples of a common interval h)
both very common in practice
20
Causality
interpretation of
x(t) = etAx(0) +
∫ t
0
e(t−τ)ABu(τ) dτ
y(t) = CetAx(0) +
∫ t
0
Ce(t−τ)ABu(τ) dτ +Du(t)
for t ≥ 0:
current state (x(t)) and output (y(t)) depend on past input (u(τ) for τ ≤ t)
i.e., mapping from input to state and output is causal (with fixed initial state)
21
Idea of state
x(t) is called state of system at time t since:
I future output depends only on current state and future input
I future output depends on past input only through current state
I state summarizes effect of past inputs on future output
I state is bridge between past inputs and future outputs
22
Change of coordinates
start with LDS x = Ax+Bu, y = Cx+Du
change coordinates in Rn to x, with x = T x
then˙x = T−1x = T−1(Ax+Bu) = T−1ATx+ T−1Bu
hence LDS can be expressed as
˙x = Ax+ Bu, y = Cx+ Du
whereA = T−1AT, B = T−1B, C = CT, D = D
TF is same (since u, y aren’t affected):
C(sI − A)−1B + D = C(sI −A)−1B +D
23
Standard forms for LDS
can change coordinates to put A in various forms (diagonal, real modal, Jordan. . . )
e.g., to put LDS in diagonal form, find T s.t.
T−1AT = diag(λ1, . . . , λn)
write
T−1B =
bT1
...
bTn
, CT =[c1 · · · cn
]
24
Discrete-time systems
discrete-time LDS:
x(t+ 1) = Ax(t) +Bu(t), y(t) = Cx(t) +Du(t)
we have:x(1) = Ax(0) +Bu(0),
x(2) = Ax(1) +Bu(1)
= A2x(0) +ABu(0) +Bu(1),
and in general, for t ∈ Z+,
x(t) = Atx(0) +
t−1∑τ=0
A(t−1−τ)Bu(τ)
25
Discrete-time systems
Solution is
x(t) = Atx(0) +
t−1∑τ=0
A(t−1−τ)Bu(τ)
write this asy(t) = CAtx(0) +H ∗ u
where ∗ is discrete-time convolution
y(t) = CAtx(0) +t∑
τ=0
H(t− τ)u(τ)
and
H(t) =
{D, t = 0CAt−1B, t > 0
is the impulse response
26
Block Toeplitz matrices
we have
y(0)y(1)y(2)
...y(t)
=
DCB DCAB CB D
.... . .
CAt−1B CAt−2B . . . CB D
u(0)u(1)u(2)
...u(t)
+
CCACA2
...CAt
x(0)
I this matrix gives the output sequence y(0), y(1), . . . in terms of the inputsequence u(0), u(1), . . . and the initial state x(0)
I block Toeplitz means blocks are constant along diagonals from top-left tobottom right
I we can use this to find controllers and estimators
27
Example: Point mass
unit point mass, with actuators applying force in directions
v1 =
[10
]v2 =
[−0.51
]v3 =
[11
]has dynamics
x(k + 1) =
1 h 0 00 1 0 00 0 1 h0 0 0 1
x(k) +
12h2 0h 00 1
2h2
0 h
[ v1 v2 v2] u1(k)
u2(k)u3(k)
y(k) =
[1 0 0 00 0 1 0
]x(k)
here
I x1, x2 = position, velocity in x-directionx3, x4 = position, velocity in y-direction
I h = sample time; we’ll use h = 1.
I ui(k) force applied by actuator i at time k28
Example: Point mass
we would like to drive it through the positions
y(20) =
[53
]y(40) =
[10−1
]y(70) =
[41
]at the above times
we have
y(t) =
t−1∑τ=0
CAt−1−τBu(τ) +Du(t)
this gives the rows of y(20)y(40)y(70)
= Aact
u(0)...
u(70)
here Aact is 6× 213.
29
Example: Point mass
let’s find the minimum norm sequence of forces that meets the specifications
u(0)...
u(70)
= A†act
5310−141
trajectory is
0 2 4 6 8 10−2
−1
0
1
2
3
4
30
Example: Point mass
sequence of force inputs is
0 10 20 30 40 50 60 70 80−0.03
−0.02
−0.01
0
0.01
0.02
0.03
31