Outline Introduction Statical virtual sensors Dynamical virtual sensors

Virtual Sensing for Control Ulf Holmberg

[email protected]

Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Outline 1

Introduction Towards intelligent sensors Virtual (soft) sensors

2

Statical virtual sensors Car engine virtual sensing and control Ankle gait assistance

3

Dynamical virtual sensors Kalman filters A virtual microphone headrest system A “sensor-free” power assisting wheelchair

Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Towards intelligent sensors Virtual (soft) sensors

Outline 1

Introduction Towards intelligent sensors Virtual (soft) sensors

2

Statical virtual sensors Car engine virtual sensing and control Ankle gait assistance

3

Dynamical virtual sensors Kalman filters A virtual microphone headrest system A “sensor-free” power assisting wheelchair

Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Towards intelligent sensors Virtual (soft) sensors

Primary sensors → Intelligent sensors Major defects in primary sensors: - nonlinearity - cross-sensitivity - time (or frequency) response - noise - parameter drift

Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Towards intelligent sensors Virtual (soft) sensors

Primary sensors → Intelligent sensors Major defects in primary sensors: - nonlinearity - cross-sensitivity - time (or frequency) response - noise - parameter drift Intelligent sensors compensation techniques: - structural compensation - tailored compensation - monitored compensation (‘sensor-within-a-sensor’) - deductive compensation (soft sensors) Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Towards intelligent sensors Virtual (soft) sensors

Example of “intelligent” sensors piezo sensors: vibration damping of ski, tennis rackets, etc. encoder: ‘noiseless’ → high gain feedback accelerometer: mechanic-electronic on chip laser: autonomous trucks ultra-sound: submarine navigation, fish detection computer vision: gestures replacing remote control fiber optic sensors: -

strain/displacement force/pressure moisture/temperature/chemical parameters vibration/acoustic emission distributed sensing Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Towards intelligent sensors Virtual (soft) sensors

Accelerometer: a micro-electromechanical sensor Applications: air bags, impact detection active suspension adaptive breaks inertial navigation machine control robotics tilt/attitude sensor vibration sensor aerospace, flight control

Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Towards intelligent sensors Virtual (soft) sensors

Sensor fusion

improve measurand by combination of sensors - range of operation - reliability (fault detection) - accuracy - bandwidth

Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Towards intelligent sensors Virtual (soft) sensors

Sensor fusion of thermal and visible images

Visible

Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Towards intelligent sensors Virtual (soft) sensors

Sensor fusion of thermal and visible images

Thermal

Visible

Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Towards intelligent sensors Virtual (soft) sensors

Sensor fusion of thermal and visible images Thermal

Visible

Fused

Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Towards intelligent sensors Virtual (soft) sensors

Alternative to sensor fusion Elvira, a Humanoid

Inclinometer only stationary angle sensor Dynamical angle sensing: sensor fusion with rate gyro

Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Towards intelligent sensors Virtual (soft) sensors

Alternative to sensor fusion Elvira, a Humanoid

Inclinometer only stationary angle sensor Dynamical angle sensing: sensor fusion with rate gyro include sensor dynamics → compliant control design Angle step responses 15

Open loop 10

Closed loop

5

0

-5

-10

0

1

2

3

4

5

6

7

time [s]

Ulf Holmberg

Virtual Sensing for Control

8

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Towards intelligent sensors Virtual (soft) sensors

Outline 1

Introduction Towards intelligent sensors Virtual (soft) sensors

2

Statical virtual sensors Car engine virtual sensing and control Ankle gait assistance

3

Dynamical virtual sensors Kalman filters A virtual microphone headrest system A “sensor-free” power assisting wheelchair

Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Towards intelligent sensors Virtual (soft) sensors

Virtual sensors extract information from other related signals Virtual sensors are model-based estimators static model based on related signal internal state reconstruction (Kalman filter, observer)

Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Towards intelligent sensors Virtual (soft) sensors

Virtual sensors extract information from other related signals Virtual sensors are model-based estimators static model based on related signal internal state reconstruction (Kalman filter, observer) Example Object not physically accessible: - nuclear reactor, bio reactor - human brain - cylinder chamber in an internal combustion engine Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Car engine virtual sensing and control Ankle gait assistance

Outline 1

Introduction Towards intelligent sensors Virtual (soft) sensors

2

Statical virtual sensors Car engine virtual sensing and control Ankle gait assistance

3

Dynamical virtual sensors Kalman filters A virtual microphone headrest system A “sensor-free” power assisting wheelchair

Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Car engine virtual sensing and control Ankle gait assistance

Optimal ignition control

Optimal control of ignition: Pressure peak at a constant crank angle

Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Car engine virtual sensing and control Ankle gait assistance

Today’s approach

Drawbacks with feedforward table complex suboptimal hard to calibrate

Why not use feedback from pressure sensors?

Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Car engine virtual sensing and control Ankle gait assistance

Feedback using pressure sensor

Pressure Drawbacks with pressure sensors expensive (US$ 2000) intrusive short life-time (500h)

Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Car engine virtual sensing and control Ankle gait assistance

Virtual sensor from spark plug ion current Ion current measurement system

Averaged pressure and ion current Ion current Pressure trace (Scaled)

2

Volts

1.5

1

virtual sensing 0.5

0

−40

−30

−20

−10

0

10 20 Crank Angle

30

Correlation between pressure peak position ion current last peak Ulf Holmberg

Virtual Sensing for Control

40

50

60

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Car engine virtual sensing and control Ankle gait assistance

Novel approach - using virtual sensors

Spark plugs as virtual sensors for control of ignition air/fuel ratio exhaust gas recirculation Result: up to 10% reduction in fuel consumption!

Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Car engine virtual sensing and control Ankle gait assistance

Ion current measurement

Single cycle ion current characteristic phases ignition: coil charging and ringing–disturbance flame front: flame propagation–chemical ionization post flame: temperature increase–thermal ionization Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Car engine virtual sensing and control Ankle gait assistance

Pressure peak position—ion current Dynamometer data under constant ignition, speed and load

Ion current signal shows large combustion cycle variations

Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Car engine virtual sensing and control Ankle gait assistance

EGR—ion current Average ion current with and without EGR

EGR decreases combustion rate and temperature

Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Car engine virtual sensing and control Ankle gait assistance

EGR—ion current Average ion current with and without EGR

EGR decreases combustion rate and temperature → smaller ion signal and delayed peak Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Car engine virtual sensing and control Ankle gait assistance

Decreased fuel consumption and NOx emission Three load cases in dynamometer showing EGR reducing fuel consumption NOx emission

Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Car engine virtual sensing and control Ankle gait assistance

Measure of combustion variability (EGR level) Indicated mead effective pressure (IMEP) Z 1 IMEP = p(θ)dV (θ) Vd Coefficient of variation COV (IMEP) =

σ(IMEP) µ(IMEP)

σ standard deviation µ mean Notice: pressure sensor needed (only in labs!) Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Car engine virtual sensing and control Ankle gait assistance

Virtual sensing of combustion variability

Feature based on ion current signal I (ck ) (ck crank angle) M=

n X

I (ck )

“Mass”

k=1

Virtual sensor (find relation) M ∼ COV (IMEP)

Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Car engine virtual sensing and control Ankle gait assistance

Virtual sensing of combustion variability COV(M)

COV(M) 6∼ COV(IMEP) (low EGR) Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Car engine virtual sensing and control Ankle gait assistance

Virtual sensing of combustion variability COV(M)

COV(M) 6∼ COV(IMEP) (low EGR) Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Car engine virtual sensing and control Ankle gait assistance

Virtual sensing of combustion variability COV(M)

Mean(M)

COV(M) 6∼ COV(IMEP) (low EGR)

Mean(M) ∼ COV(IMEP) (low EGR)

Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Car engine virtual sensing and control Ankle gait assistance

Virtual sensing of combustion variability COV(M)

Mean(M)

COV(M) 6∼ COV(IMEP) (low EGR)

Mean(M) ∼ COV(IMEP) (low EGR)

Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Car engine virtual sensing and control Ankle gait assistance

Virtual sensing of combustion variability COV(M)

Mean(M)

COV(M) 6∼ COV(IMEP) (low EGR)

Mean(M) ∼ COV(IMEP) (low EGR)

Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Car engine virtual sensing and control Ankle gait assistance

Virtual sensing and control of combustion variability

Variable of interest (combustion variability) v ≡ COV (IMEP) Virtual sensor vˆ = d1 · M + d2 Controller u(t) = u(t − 1) +

1 (vdes (t) − vˆ (t)) Ti

u(t): EGR valve position at cycle t

Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Car engine virtual sensing and control Ankle gait assistance

Virutal sensing of pressure peak position Ion current samples I (ck ) at crank angle ck Center of mass: T

cM = c m/M

mk = I (ck ) m = [m1 , . . . , mn ]T c = [c1 , . . . , cn ]T

Model pressure peak position θp as θˆp = a · cM + b Interpolate between non-EGR (a1 , b1 , M1 ) and EGR (a2 , b2 , M2 ) a(M) = b(M) =

a2 −a1 M2 −M1 (M b2 −b1 M2 −M1 (M

Ulf Holmberg

− M1 ) + a1 − M1 ) + b1

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Car engine virtual sensing and control Ankle gait assistance

Virutal sensing of pressure peak position

6000 cycles of each no-EGR/EGR Measured and estimated pressure peak position [cad] 20

a)

10 0

2000

4000

6000

8000

1e4

cycles

8000

1e4

cycles

8000

1e4

cycles

Measured and estimated pressure peak position [cad] 20

b)

10 0

2000

4000

6000

θˆp = a1 · cM + b1 θˆp = a2 · cM + b2 θˆp = a(M) · cM + b(M)

Measured and estimated pressure peak position [cad] 20

c)

10 0

2000

4000

6000

Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Car engine virtual sensing and control Ankle gait assistance

Closed-loop step response of combustion variability Road experiment

Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Car engine virtual sensing and control Ankle gait assistance

Car engine virtual sensing and control - summary

Virtual sensing of pressure peak position θp Virtual sensing of combustion variability v = COV (IMEP) Simultaneous control of θp and v (I-controllers) Reduction of fuel consumption (10%) and NOx Implemented on a SAAB and tested on Highway

Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Car engine virtual sensing and control Ankle gait assistance

Outline 1

Introduction Towards intelligent sensors Virtual (soft) sensors

2

Statical virtual sensors Car engine virtual sensing and control Ankle gait assistance

3

Dynamical virtual sensors Kalman filters A virtual microphone headrest system A “sensor-free” power assisting wheelchair

Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Car engine virtual sensing and control Ankle gait assistance

Prosthesis control

Ankle gait assistance Estimate inclination θ Adjust ankle angle ϕ to θ Accelerometer sensor for gait phase detection for θ estimation

Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Car engine virtual sensing and control Ankle gait assistance

Gait phase detection 2

af (k) = h1 af (k − 1) + a(k) − a(k − 1)

0

g

Band pass filter

a)

−2

Heel lift [Stance stops] |af (kHL )| > γS Toe off [Swing starts] af (kTO ) > γT

g

Foot down [Stance starts] |af (k)| < γS , k ∈W

2

2

4

b)

0 −2

ST FL SW 2

4 s

Stance: Estimate (a → θ) Swing: Control (foot unloaded)

Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Car engine virtual sensing and control Ankle gait assistance

Ground angle estimation Ground angle estimation Sample estimate ˆ θ(k) = − arcsin(a(k)/g ) Average over stance samples kX HL −1 1 ˆ θ(n) = θ(k) kHL − kFD k=kFD

Average over steps θf (n) =

Ankle reference Ankle error ea (n) = θf (n) − ϕ(n) Threshold  f (ea (n)) =

ea (n), |ea (n)| > d 0 else

Reference

N−1 1 X θ(n − i) N

ϕref (n) = ϕref (n − 1) + f (ea (n))

i=0

Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Car engine virtual sensing and control Ankle gait assistance

Ankle control

deg

5

a)

0 !5

deg

5

100

200

b)

0 !5

100

200 s

Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Car engine virtual sensing and control Ankle gait assistance

Stair sensing during gait

2-axis accelerometer + gyro

aX = ax cos φ − ay sin φ aY = ax sin φ + ay cos φ − g At Stance ax = g sin φ ay = g cos φ



ax → φˆ = arctan ay

Estimate at Stance ˆ φ¯ = mean[φ]

Ulf Holmberg

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Car engine virtual sensing and control Ankle gait assistance

Swing estimation Velocity Integrate gyro during swing

vˆX ,k+1 = vˆX ,k + aX ,k h vˆY ,k+1 = vˆY ,k + aY ,k h

φˆk+1 = φˆk + ωk h, k = 0, . . . , T φˆ0 = φ¯Before (stance)

Compensate drift

Compensate drift φk = φˆk TT−k + φ¯After Tk , k = 0, . . . , T making φT = φ¯After (next stance)

Ulf Holmberg

vX ,k = vY ,k =

T −k ˆX ,k T v T −k v T ˆY ,k

Position xk+1 = xk + vX ,k h yk+1 = yk + vY ,k h

Virtual Sensing for Control

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Car engine virtual sensing and control Ankle gait assistance

Classification of stairs

Classification variable (xT > 0.2m) 0.015

yT QT = xT 0.01

Classification Up: γU < QT Horizontal: γD < QT < γU Down: QT < γD

0.005

0 0.4

0.2

0

0.2

0.4

QT

Ulf Holmberg

Virtual Sensing for Control

0.6

0.8

1

1.2

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Car engine virtual sensing and control Ankle gait assistance

Foot orthosis strain sensing in hill walking

S1

Estimate inclination from strain from gait cycles at different walking speeds 0.5 −3 degrees 0 degrees 0

−5degrees

Volt

3 degrees

−0.5 5 degrees

0

25

50 % Gait cycle

75

100

Ulf Holmberg

Virtual Sensing for Control

S2

Outline Introduction Statical virtual sensors Dynamical virtual sensors

Car engine virtual sensing and control Ankle gait assistance

Fourier series representation Sample strain from gait cycle s = ( S(t1 ) . . . , S(tN ) )T Introduce Cij = cos(Ωij) Sij = sin(Ωij) 1 = ( 1 . . . 1 )T

C ∈