Neural Network Lec10

Outline An Actuator Additive FDI [1] An Actuator Gain FDI [1] A Sensor and Actuator FDI [1]
Neural Networks
Lecture 10: Fault Detection and Isolation
(FDI) Using Neural Networks
H. A. Talebi, Farzaneh Abdollahi
H.A. Talebi
Farzaneh Abdollahi
Department of Electrical Engineering

Amirkabir University of Technology
Winter
2011
Neural
Networks
Lecture 10
1/37
An Actuator Additive FDI [1]

Example:FDI on Robotic Manipulation
An Actuator Gain FDI [1]

Example Flexible Joint Manipulator
A Sensor and Actuator FDI [1]

Example of Satellite Attitude Control
Neural Networks
Lecture 10
2/37
Additive FDI
I
Consider the nonlinear dynamics:

x
I
I
= f (x, u) + N1 (x, u) + TF (x, u, t)
f : known nonlinear fcn.

N1 (x, u): Unmodeled dynamics and uncertainties of nominal model (fault
free), estimated by NN1
TF (x, u, t): vector valued fcn. of unknown actuator fault
Objective: Diagnose and Isolate Actuator Fault: (TF )
By defining a Hurwitz matrix A, (1) can be rewritten

x
I
(1)
= Ax + g (x, u) + N1 (x, u) + TF (x, u, t)
g (x, u) = f (x, u) Ax
M(s) = (sI A)1
Neural Networks
Lecture 10
3/37
Additive FDI
Neural Networks
Lecture 10
4/37
NN2
I
f (x, W
,V
, t) = W
(V
x)
Tf is estimated by NN2 : T
I
I
A three layer MLP

Series Parallel identifier
The estimated model

(V
x)
x = A
x + g (x, u) + N1 (x, u) + W
It has been shown that by updating the weights of NN2 :
= 1 ( J ) 1 k
x kW
W
W
= 2 ( J ) 2 k
V
x kV
V
I
I
x = x x
x T x): objective fcn to be minimized
J = 21 (
=W W
,V
=V V
are ultimately bounded
then x, W
The main advantage of the given approach: fault detection and Isolation
is provided in a unified fashion
Neural Networks
Lecture 10
5/37
Example:FDI on Robotic Manipulation
Dynamic of a robot manipulator M(q)

q + C (q, q)
+ G (q) = + F
I
I
I
I
q: manipulator joints position

q:
joint velocity
: applied torque
F : unknown fault
Basedon the defined

model:

0
q
Tf =
; f (x, u) =
M 1 F
M 1 (q)( C (q, q)
+ G (q) + F )
Neural Networks
Lecture 10
6/37
Experimental Results on Puma 560
Neural Networks
Lecture 10
7/37
Example Puma 560

I
It has 6 joints
The unmodelled dynamics (for Ex. friction in joint which are not modeled
precisely) are estimated by N1
The modified fault free description of Puma:

M(q)
q + C (q, q)
+ G (q) + N1 (q, q,
) =
Define = q
At this stage, a PD controller is considered to control the robots
+G
behavior: C = Kp (ref ) + Kv (ref )
I
I
Desired position: ref = H + 0.3sin(t)

H = [0 /2 /2 0 0 0]
Neural Networks
Lecture 10
8/37
Fault Free Experimental Results
(a)-(c) joint
positions;
(d)-(f) Joint
velocities.
Solid lines:
actual
measurements;
Dashed lines:
their estimates.
Neural Networks
Lecture 10
9/37
Faulty Experimental Results
Consider a constant actuator fault 5Nm in the first three joints:

F = [5 5 5 0 0 0]
Estimate F by NN2: a three layer NN
I
I
I
input layers: 12 neurons (, )

hidden layer 10 neurons
output layer 6 neurons F
Neural Networks
Lecture 10
10/37
The outputs of the neural network estimating the faults
Neural Networks
Lecture 10
11/37
Experimental Results with 5 Nm fault
(a)-(c) Joint
positions
(d)-(f)Joint
velocities.
Solid lines:
actual
measurements
Dashed
lines:their
estimates
Neural Networks
Lecture 10
12/37
An Actuator Gain FDI [1]

I
Consider the nonlinear fault free dynamics:

x
= f (x, u) + x (x, u, t)
(2)
y (t) = Cx(t) + y (x, u, t)
I
I
I
I
u Rm : input; x Rn : state;y Rm : output

x (x, u, t) Rn Rm R Rn : Plant unmodeled dynamics and noise
y (x, u, t) Rn Rm R Rm : Sensor unmodeled dynamics and
disturbances
C : an m n conts. matrix
Assume a multiplicative actuator fault occur:

u(t)uf (t) = TA (x, u, t)u(t)
I
TA (x, u, t) = diag {TAi (x, u, t)}, i = 1, ..., m unknown matrix
Neural Networks
Lecture 10
13/37
An Actuator Gain FDI

I
The faulty dynamics:

x
= f (x, uf ) + x (x, u, t)
(3)
y (t) = Cx(t) + y (x, u, t)

I
Objective: Diagnose and Isolate Actuator Fault: (TF )

x
= Ax + g (x, uf ) + x (x, u, t)
y (t) = Cx(t) + y (x, u, t)

I
g (x, u) = f (x, u) Ax
M(s) = (sI A)1
Neural Networks
Lecture 10
14/37
Neural Networks
Lecture 10
15/37

I
The following assumptions should be made to facilitate the FDI design

I
I
I
I
f (x, W
,V
, t) = W
(V
x)
Tf is estimated by NN1 : T
I
The nonlinear sys (3) is observable

The nominal closed loop sys is asym. stable (x = f (x, k(Cx)))
x , y are uniformly bounded kx k x , ky k y

x x
g is Lipshtiz in x and u: kg (x, u) g (
x , u)k lg k
k
u u
A three layer MLP
The estimated model
= A
x + g (
x , uf )
y (t) = C x
(V
x)u
uf = W
Neural Networks
Lecture 10
16/37

I
It has been shown that by updating the weights of NN1 :
I
I
I
I
I
) 1 k
y kW
W
J
) 2 k
y kV
= 2 (
V
= 1 (
x = x x
y = y y
=W W
W
=V V
V
J = 12 (
y T y ): objective fcn to be minimized
,V
are ultimately bounded
then x, y , W
This FDI can diagnose and isloate simultaneous fault
Neural Networks
Lecture 10
17/37
Example of Flexible Joint Manipulator [?]

I
The dynamical model:

Dl (q1 )
q1 + C1 (q1 , q 1 ) + g (q1 ) + B1 q 1 = s
J q2 + s + B2 q 2 =
I
I
I
I
I
I
I
I
I
q1 R : link position
q2 Rn : motor shaft position
g (q1 ) Rn : gravity force
C1 (q1 , q 1 ) Rn : centrifugal and Coriolis force
B1 Rnn , B2 Rnn : viscus damping of input and output shaft
D1 (q1 ) Rnn , J Rnn robot and actuator inertia matrix
input torque
s = K (q2 q1 ) + (q1 , q2 , q 1 , q 2 ): reaction torque from rotational spring
K : stiffness matrix of rotational spring represent the flexibility between the
input and output shaft
(q1 , q2 , q 1 , q 2 ): a combination of nonlinear spring and friction at output
shafts of the manipulator
The output vector y = [q21 q22 q 21 q 22 ]T
I
Neural Networks
Lecture 10
18/37
Numerical Data of Two Link Flexible Joint Manipulator

I
J = diag {1.16, 1.16}, m = diag {1, 1}
l1 = l2 = 1m, K = diag {100, 100}, = 0.3, 1 = 2 = 0.001
A = 2I8
A PD controller is used to stabilize the open-loop system
is uncertainty in dynamics and unknown
Sensory measurement is corrupted with a 10% Gaussian noise

The NN FDI:
I
I
A Three layer NN
10 neurons in input layer; 10 neurons in hidden layer; 2 neurons in output
layer ( faults in actuator 1 and 2
The actuator
fault in both joints:

diag {0.7 0.4} t [200 300]secs
TA =
diag {1 1}
otherwise
Neural Networks
Lecture 10
19/37
Example Flexible Joint Manipulator

I
(a) angular velocities of the first link
(b) angular velocity of the second link
Red lines: real signals; Blue lines: the estimated signals by NN
Neural Networks
Lecture 10
20/37
(a) Estimated
fault in the
first actuator
(b) Estimated
fault in the
second
actuator
(c) Estimated
fault in the
first actuator;
a closer look
(d) Estimated
fault in the
second
actuator; a
closer look
Neural Networks
Lecture 10
21/37
A Sensor and Actuator FDI [1]

I
Consider the nonlinear dynamics:

x
= f (x, u) + x (x, u, t) + TA (x, u)
(4)
y (t) = Cx(t) + y (x, u, t) + Ts (x, u)
I
I
I
I
I
I
I
u Rm : input; x Rn : state;y Rm : output

x (x, u, t) Rn Rm R Rn : Plant unmodeled dynamics and noise
y (x, u, t) Rn Rm R Rm : Sensor unmodeled dynamics and
disturbances
C : an m n conts. matrix
Ts Rn Rm Rm : unknown sensor fault
TA Rn Rm Rn : unknown actuator fault
Neural Networks
Lecture 10
22/37
A Sensor and Actuator FDI
Objective: Diagnose and Isolate Actuator and Sensor Faults: (TA , TS )

x
= Ax + g (x, u) + x (x, u, t) + TA (x, u)
y (t) = Cx(t) + y (x, u, t) + Ts (x, u)

I
g (x, u) = f (x, u) Ax
M(s) = (sI A)1
Neural Networks
Lecture 10
23/37
Neural Networks
Lecture 10
24/37

I
The following assumptions should be made

I
I
I
I
I
TA is estimated by NN1 and TS is estimated by NN2 :

A (
1, V
1 , t) = W
1 (V
1 x
T
x, W
)
S (
2, V
2 , t) = W
2 (V
2 x
)
T
x, W
I
Sensor and actuator fault will not occur simultaneously

The nonlinear sys (4) is observable
The nominal closed loop sys is asym. stable (x = f (x, k(Cx)))
x , y are uniformly bounded kx k x , ky k y
g is Lipshtiz in x: kg (x, u) g (
x , u)k lg kx xk
Each is a three layer MLP
The estimated model

1 (V
1 x)
x = A
x + g (
x , u) + W
2 (V
2 x
y (t) = C x + W
)
Neural Networks
Lecture 10
25/37

I
It has been shown that by updating the weights of NN1 and NN2 :
1
1 = 1 ( J ) 1 k
y kW
W
1
W
= ( J ) k
V
y kV
1
2 V
1
3 ( J
2 )
W
2 =
2
W
3 k
y kW
4
3 = 4 ( J ) 4 k
V
y kV
V
3
I
I
I
I
I
x = x x
y = y y
i = Wi W
i , i = 1, 2
W
Vi = Vi Vi , i = 1, 2
y T y ): objective fcn to be minimized
J = 12 (
i, V
i , i = 1, 2 are ultimately bounded
then x, y , W
Neural Networks
Lecture 10
26/37

I
Dynamical equation of satellite attitude model with magnetorquer

actuator H = S()H + BM
I
= [12 3 ]T : angular velocity of satellite
0
Bz (t) By (t)
0
Bx (t) : magnetic field matrix
B(t) = Bz (t)
By (t) Bx (t)
0
H: Symmetric
P.D. inertia matrix
0
3 2
0
1 : cross product matrix
S() = 3
2 1
0
M: magnetic dipole of the coil
Neural Networks
Lecture 10
27/37

I
To define it in standard state space representation:

u = M; f (x, u) = H 1 (B(t)u S()H); x =
the unmodeled dynamics: x = H 1 (B(t) B)u; B =
Magnetorquer fault (actuator fault) is considered as an additive term a :

x
1
T
RT
0
B(t)dt
= f (x, u) + x +TA ; TA = H 1 a
Magnetometer fault (sensor fault): Bm = B + EB + B +

I
I
I
I
Bm : measured magnetic field

B: true magnetic field
B: magnetic bias
measurement
noise
0
Ez
Ey
0
Ex : magnetormeter misalignment matrix
E = Ez
Ey Ex
0
any fault or misalignment can be considered as uncertainty vector y or
additive fault TS
Neural Networks
Lecture 10
28/37
Numerical Data of rsted Satellite
181.78
0
0
0
181.25
0 kg .m2
H=
0
0
1.28
All sensory measurements are assumed to be corrupted with 10%
Gaussian noise
The actuator fault is estimated by NN1 and sensor fault by NN2 . They
have:
I
I
I
I
I
I
3 layers
6 neurons in input layer; 5 neurons in hidden layer with sigmoidal activation
fcn; 3 neurons in output layer with linear activation fcn.
1 = 2 = 20; 3 = 4 = 0.1; i = 1e 6, i = 1, 2, 3, 4
A = 2I3
A = [T
A1 T
A2 T
A3 ]T : output of NN1
T
S = [T
S1 T
S2 ]T : output of NN2
T
Neural Networks
Lecture 10
29/37

I
Note that: the output of the networks do not approach to zero due to
presence of y and x
Therefore a threshold should be considered for fault detection
The threshold can be chosen by by adaptive , fuzzy or ...approached
In this simulation it has been chosen by trial and error
A fault is detected if:

A | > thai , i = 1, ..., 3 (actuator fault)
|T
i
|TSi | > thsi , i = 1, ..., 2 (sensor fault)

t [tf , tf + tr ]
I
I
I
thai = [0.02 0.02] threshold of ith actuator

thsi = [0.02 0.02] threshold of ith sensor
tr = 30sec.
Neural Networks
Lecture 10
30/37
Fault Free angular velocities, trajectory: 0.1sin(0.001t)
Solid lines:
actual
velocities
Dashed
lines:
Estimated
velocities
Neural Networks
Lecture 10
31/37
NN1 and NN2 responses to fault free operation

I
Solid lines:
estimated
faults;
Dashed
lines:
thresholds
(a)-(c):
Estimated
actuator
faults
(d)-(e):
Estimate
sensor
faults
Neural Networks
Lecture 10
32/37
Angular velocities with bias fault: TA = [0.1 0 0]T
Solid lines:
actual
velocities
Dashed
lines:
Estimated
velocities
Neural Networks
Lecture 10
33/37
NN1 and NN2 responses to bias fault: TA = [0.1 0 0]T

I
Solid lines:
estimated
faults;
Dashed
lines:
thresholds
(a)-(c):
Estimated
actuator
faults
(d)-(e):
Estimate
sensor
faults
Neural Networks
Lecture 10
34/37
NN1 and NN2 responses;

I
Solid lines:
estimated
faults;
Dashed
lines:
thresholds
(a)-(c):
Estimated
actuator
faults
(d)-(e):
Estimate
sensor
faults
traj: 0.1rad/s, fault: TS = [0.03 0.05]T
Neural Networks
Lecture 10
35/37
NN1 and NN2 responses; traj: 0.1sin(0.001t), fault free, with sensor and state
uncertainties
I
Solid lines:
estimated
faults;
Dashed
lines:
thresholds
(a)-(c):
Estimated
actuator
faults
(d)-(e):
Estimate
sensor
faults
I
Neural Networks
Lecture 10
36/37
References
H.A.Talebi, F. Abdollahi, R.V.Patel, and K.Khorasani, Neural Network-Based State

Estimation of Nonlinear Systems: Application to Fault Detection and Isolation.
Springer, 2009.
Neural Networks
Lecture 10
37/37

Neural Network Lec10

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Neural Network Lec10

Uploaded by

Copyright:

Available Formats

Outline An Actuator Additive FDI [1] An Actuator Gain FDI [1] A Sensor and Actuator FDI [1]

H. A. Talebi, Farzaneh Abdollahi

Department of Electrical Engineering

An Actuator Additive FDI [1]

An Actuator Gain FDI [1]

A Sensor and Actuator FDI [1]

H. A. Talebi, Farzaneh Abdollahi

Consider the nonlinear dynamics:

= f (x, u) + N1 (x, u) + TF (x, u, t)

f : known nonlinear fcn.

Objective: Diagnose and Isolate Actuator Fault: (TF )

By defining a Hurwitz matrix A, (1) can be rewritten

= Ax + g (x, u) + N1 (x, u) + TF (x, u, t)

M(s) = (sI A)1

H. A. Talebi, Farzaneh Abdollahi

H. A. Talebi, Farzaneh Abdollahi

A three layer MLP

The estimated model

It has been shown that by updating the weights of NN2 :

H. A. Talebi, Farzaneh Abdollahi

Example:FDI on Robotic Manipulation

Dynamic of a robot manipulator M(q)

q: manipulator joints position

Basedon the defined

H. A. Talebi, Farzaneh Abdollahi

Experimental Results on Puma 560

H. A. Talebi, Farzaneh Abdollahi

Example Puma 560

The modified fault free description of Puma:

Desired position: ref = H + 0.3sin(t)

H. A. Talebi, Farzaneh Abdollahi

Fault Free Experimental Results

H. A. Talebi, Farzaneh Abdollahi

Faulty Experimental Results

Consider a constant actuator fault 5Nm in the first three joints:

input layers: 12 neurons (, )

H. A. Talebi, Farzaneh Abdollahi

The outputs of the neural network estimating the faults

H. A. Talebi, Farzaneh Abdollahi

Experimental Results with 5 Nm fault

H. A. Talebi, Farzaneh Abdollahi

An Actuator Gain FDI [1]

Consider the nonlinear fault free dynamics:

y (t) = Cx(t) + y (x, u, t)

u Rm : input; x Rn : state;y Rm : output

Assume a multiplicative actuator fault occur:

TA (x, u, t) = diag {TAi (x, u, t)}, i = 1, ..., m unknown matrix

H. A. Talebi, Farzaneh Abdollahi

An Actuator Gain FDI

The faulty dynamics:

y (t) = Cx(t) + y (x, u, t)

Objective: Diagnose and Isolate Actuator Fault: (TF )

By defining a Hurwitz matrix A, (3) can be rewritten

y (t) = Cx(t) + y (x, u, t)

M(s) = (sI A)1

H. A. Talebi, Farzaneh Abdollahi

An Actuator Gain FDI

H. A. Talebi, Farzaneh Abdollahi

An Actuator Gain FDI

The following assumptions should be made to facilitate the FDI design

The nonlinear sys (3) is observable

The estimated model

H. A. Talebi, Farzaneh Abdollahi

Basedon the defined