Mixed phenomenological and neural approach to induction motor speed - - PowerPoint PPT Presentation

Problem statement Feedforward neural network without preprocessing Feedforward neural network with preprocessing Some results and conclusion(s)

Mixed phenomenological and neural approach to induction motor speed estimation

• B. Beliczynski, L. Grzesiak and B. Ufnalski

Institute of Control and Industrial Electronics Warsaw University of Technology, POLAND

Bartlomiej Beliczynski, Lech Grzesiak and Bartlomiej Ufnalski Speed neuroestimation in induction motor drives

Problem statement Feedforward neural network without preprocessing Feedforward neural network with preprocessing Some results and conclusion(s)

Outline

1

Problem statement

2

Feedforward neural network without preprocessing

3

Feedforward neural network with preprocessing

4

Some results and conclusion(s)

Bartlomiej Beliczynski, Lech Grzesiak and Bartlomiej Ufnalski Speed neuroestimation in induction motor drives

Problem statement Feedforward neural network without preprocessing Feedforward neural network with preprocessing Some results and conclusion(s)

Motivation There are applications of induction motor drives that require only a moderate control precision. In such cases we tend to go into speed-sensorless operation, i.e. encoders are replaced with soft speed sensors.

Bartlomiej Beliczynski, Lech Grzesiak and Bartlomiej Ufnalski Speed neuroestimation in induction motor drives

Problem statement Feedforward neural network without preprocessing Feedforward neural network with preprocessing Some results and conclusion(s)

Motivation There are applications of induction motor drives that require only a moderate control precision. In such cases we tend to go into speed-sensorless operation, i.e. encoders are replaced with soft speed sensors. Why soft sensors?

Bartlomiej Beliczynski, Lech Grzesiak and Bartlomiej Ufnalski Speed neuroestimation in induction motor drives

Problem statement Feedforward neural network without preprocessing Feedforward neural network with preprocessing Some results and conclusion(s)

Motivation There are applications of induction motor drives that require only a moderate control precision. In such cases we tend to go into speed-sensorless operation, i.e. encoders are replaced with soft speed sensors. Why soft sensors? Hint #1: To reduce cost.

Bartlomiej Beliczynski, Lech Grzesiak and Bartlomiej Ufnalski Speed neuroestimation in induction motor drives

Problem statement Feedforward neural network without preprocessing Feedforward neural network with preprocessing Some results and conclusion(s)

Motivation There are applications of induction motor drives that require only a moderate control precision. In such cases we tend to go into speed-sensorless operation, i.e. encoders are replaced with soft speed sensors. Why soft sensors? Hint #1: To reduce cost. Hint #2: To increase reliability – it is not uncommon for electric drives to operate in harsh environments.

Bartlomiej Beliczynski, Lech Grzesiak and Bartlomiej Ufnalski Speed neuroestimation in induction motor drives

Problem statement Feedforward neural network without preprocessing Feedforward neural network with preprocessing Some results and conclusion(s)

Motivation There are applications of induction motor drives that require only a moderate control precision. In such cases we tend to go into speed-sensorless operation, i.e. encoders are replaced with soft speed sensors. Why soft sensors? Hint #1: To reduce cost. Hint #2: To increase reliability – it is not uncommon for electric drives to operate in harsh environments. Why neuroestimators as soft sensors?

Bartlomiej Beliczynski, Lech Grzesiak and Bartlomiej Ufnalski Speed neuroestimation in induction motor drives

Problem statement Feedforward neural network without preprocessing Feedforward neural network with preprocessing Some results and conclusion(s)

Motivation There are applications of induction motor drives that require only a moderate control precision. In such cases we tend to go into speed-sensorless operation, i.e. encoders are replaced with soft speed sensors. Why soft sensors? Hint #1: To reduce cost. Hint #2: To increase reliability – it is not uncommon for electric drives to operate in harsh environments. Why neuroestimators as soft sensors? Hint #1: They are innately able to cope with plant nonlinearities.

Bartlomiej Beliczynski, Lech Grzesiak and Bartlomiej Ufnalski Speed neuroestimation in induction motor drives

Problem statement Feedforward neural network without preprocessing Feedforward neural network with preprocessing Some results and conclusion(s)

Motivation There are applications of induction motor drives that require only a moderate control precision. In such cases we tend to go into speed-sensorless operation, i.e. encoders are replaced with soft speed sensors. Why soft sensors? Hint #1: To reduce cost. Hint #2: To increase reliability – it is not uncommon for electric drives to operate in harsh environments. Why neuroestimators as soft sensors? Hint #1: They are innately able to cope with plant nonlinearities. Hint #2: They can be trained using real plant data – no mathematical model of the plant is needed.

Bartlomiej Beliczynski, Lech Grzesiak and Bartlomiej Ufnalski Speed neuroestimation in induction motor drives

Problem statement Feedforward neural network without preprocessing Feedforward neural network with preprocessing Some results and conclusion(s)

Problem statement Induction motor ωm mo

?

ωm ˆ us is us

?

Bartlomiej Beliczynski, Lech Grzesiak and Bartlomiej Ufnalski Speed neuroestimation in induction motor drives

Problem statement Feedforward neural network without preprocessing Feedforward neural network with preprocessing Some results and conclusion(s)

Model of the induction machine – just to make you familiar with the plant we are talking here (not needed to synthesize the soft speed neurosensor) us = Rsis + d dt ψs + jωψsψs (1) 0 = Rrir + d dt ψr + j

• ωψs − pbωm
• ψr

(2) ψs = Lsis + Lmir (3) ψr = Lrir + Lmis (4) J d dt ωm = me − ctωm − mo (5) me = 3 2pb

• ψs × is
• (6)

Bartlomiej Beliczynski, Lech Grzesiak and Bartlomiej Ufnalski Speed neuroestimation in induction motor drives

Problem statement Feedforward neural network without preprocessing Feedforward neural network with preprocessing Some results and conclusion(s)

Straightforward approach ωm(k) = f(ωm(k − 1), mo(k), ψs(k − 1), is(k − 1), us(k)) (7)

Bartlomiej Beliczynski, Lech Grzesiak and Bartlomiej Ufnalski Speed neuroestimation in induction motor drives

Problem statement Feedforward neural network without preprocessing Feedforward neural network with preprocessing Some results and conclusion(s)

Recurrent neural network

RNN (dynamic NN) usα usβ ωm z-1 z-1 mo ˆ ˆ RNN (dynamic NN) usα usβ ωm z-1 z-1 ˆ isα isβ

Bad idea! Practically impossible to get any robustness.

Bartlomiej Beliczynski, Lech Grzesiak and Bartlomiej Ufnalski Speed neuroestimation in induction motor drives

Problem statement Feedforward neural network without preprocessing Feedforward neural network with preprocessing Some results and conclusion(s)

Feedforward neural network

FFNN (static NN) usa usb usc isa isb isc ωm ˆ FFNN (static NN) usa usb isa isb ωm ˆ FFNN (static NN) usα usβ isα isβ ωm ˆ

Still bad idea! Practically impossible to get any robustness.

Bartlomiej Beliczynski, Lech Grzesiak and Bartlomiej Ufnalski Speed neuroestimation in induction motor drives

Problem statement Feedforward neural network without preprocessing Feedforward neural network with preprocessing Some results and conclusion(s)

Rotating reference frame

FFNN usd usq isd isq ωm usα usβ isα isβ αβ/dq Flux estimator γ ˆ

Much better but still sensitive to flux estimation errors.

Bartlomiej Beliczynski, Lech Grzesiak and Bartlomiej Ufnalski Speed neuroestimation in induction motor drives

Problem statement Feedforward neural network without preprocessing Feedforward neural network with preprocessing Some results and conclusion(s)

Stationary reference frame

FFNN u1 u2 u3 u4 ωm usα usβ isα isβ Nonlinear transformations, e.g. the Akagi power components and/or the instantaneous impedance components u5 u6 ˆ

Bartlomiej Beliczynski, Lech Grzesiak and Bartlomiej Ufnalski Speed neuroestimation in induction motor drives

Problem statement Feedforward neural network without preprocessing Feedforward neural network with preprocessing Some results and conclusion(s)

Nonlinear preprocessing u1 = |us| =

• u2

sα + u2 sβ

(8) u2 = |is| =

• i2

sα + i2 sβ

(9) u3 = ℜ (u∗

sis) = usαisα + usβisβ

(10) u4 = ℑ (u∗

sis) = usαisβ − usβisα

(11) u5 = ℜ us is

• = usαisα + usβisβ

i2

sα + i2 sβ

= u3 u2

2

(12) u6 = ℑ us is

• = usβisα − usαisβ

i2

sα + i2 sβ

= −u4 u2

2

. (13) Almost the best we can do but still heuristic i.e. rather subjective choice.

Bartlomiej Beliczynski, Lech Grzesiak and Bartlomiej Ufnalski Speed neuroestimation in induction motor drives

Problem statement Feedforward neural network without preprocessing Feedforward neural network with preprocessing Some results and conclusion(s)

Nonlinear preprocessing proposed in the EANN paper v1 = isα d dt isβ − isβ d dt isα = ℑ(i∗

s

d dt is) (14) v2 = isα d dt isα − isβ d dt isβ = ℜ(i∗

s

d dt is) = 1 2 d dt |is|2 (15) v3 = i2

sα + i2 sβ = |is|2

(16) v4 = isαusβ − isβusα = ℑ(usi∗

s)

(17) v5 = isαusα + isβusβ = ℜ(usi∗

s)

(18) For the analytical justification please consult the paper. This is probably still not the best we can do (mind the derivative of current) but the analytical approach instead of the heuristic one makes the choice much less subjective and much more systematic.

Bartlomiej Beliczynski, Lech Grzesiak and Bartlomiej Ufnalski Speed neuroestimation in induction motor drives

Problem statement Feedforward neural network without preprocessing Feedforward neural network with preprocessing Some results and conclusion(s)

Schematic diagram of the system A

DSFOC Power supply Induction motor ωm mo usa usb usc

ref

PWM

ref ref

ωm

ref

• FFNN

u1 u2 u3 u4 ωm Nonlinear transformations (heuristic or derived analytically) u5 u6 ˆ isb isc

Bartlomiej Beliczynski, Lech Grzesiak and Bartlomiej Ufnalski Speed neuroestimation in induction motor drives

Problem statement Feedforward neural network without preprocessing Feedforward neural network with preprocessing Some results and conclusion(s)

Schematic diagram of the system B

DSFOC Power supply Induction motor ωm mo usa usb usc

ref

PWM

ref ref

ωm

ref

• FFNN

ωm ˆ isb isc usd usq isd isq

ref ref

Bartlomiej Beliczynski, Lech Grzesiak and Bartlomiej Ufnalski Speed neuroestimation in induction motor drives

Problem statement Feedforward neural network without preprocessing Feedforward neural network with preprocessing Some results and conclusion(s)

Results A

5 10 15 20 25 ωref

m [pu]

• 1
• 0.5

0.5 1 5 10 15 20 25 ωm [pu]

• 1
• 0.5

0.5 1 t [s] 5 10 15 20 25 mo [pu]

• 1
• 0.5

0.5 1

Bartlomiej Beliczynski, Lech Grzesiak and Bartlomiej Ufnalski Speed neuroestimation in induction motor drives

Problem statement Feedforward neural network without preprocessing Feedforward neural network with preprocessing Some results and conclusion(s)

Results A (cont.)

t [s]

5 10 15 20 25

eω

• 5%

5%

Bartlomiej Beliczynski, Lech Grzesiak and Bartlomiej Ufnalski Speed neuroestimation in induction motor drives

Problem statement Feedforward neural network without preprocessing Feedforward neural network with preprocessing Some results and conclusion(s)

Results B

5 10 15 20 25 ωref

m [pu]

• 1
• 0.5

0.5 1 5 10 15 20 25 ωm [pu]

• 1
• 0.5

0.5 1 t [s] 5 10 15 20 25 mo [pu]

• 1
• 0.5

0.5 1

Bartlomiej Beliczynski, Lech Grzesiak and Bartlomiej Ufnalski Speed neuroestimation in induction motor drives

Problem statement Feedforward neural network without preprocessing Feedforward neural network with preprocessing Some results and conclusion(s)

Results B (cont.)

t [s]

5 10 15 20 25

eω

• 5%

5%

Bartlomiej Beliczynski, Lech Grzesiak and Bartlomiej Ufnalski Speed neuroestimation in induction motor drives

Problem statement Feedforward neural network without preprocessing Feedforward neural network with preprocessing Some results and conclusion(s)

Conclusions A novel analytically derived set of input signals for the speed neuroestimator has been proposed.

Bartlomiej Beliczynski, Lech Grzesiak and Bartlomiej Ufnalski Speed neuroestimation in induction motor drives

Problem statement Feedforward neural network without preprocessing Feedforward neural network with preprocessing Some results and conclusion(s)

Conclusions A novel analytically derived set of input signals for the speed neuroestimator has been proposed. It is feasible to use FFNN to estimate the induction motor angular velocity if a phenomenological-based signal preprocessing is introduced.

Bartlomiej Beliczynski, Lech Grzesiak and Bartlomiej Ufnalski Speed neuroestimation in induction motor drives

Problem statement Feedforward neural network without preprocessing Feedforward neural network with preprocessing Some results and conclusion(s)

Conclusions A novel analytically derived set of input signals for the speed neuroestimator has been proposed. It is feasible to use FFNN to estimate the induction motor angular velocity if a phenomenological-based signal preprocessing is introduced. Hint #1: If possible avoid recurrent neural networks in speed estimators.

Bartlomiej Beliczynski, Lech Grzesiak and Bartlomiej Ufnalski Speed neuroestimation in induction motor drives

Problem statement Feedforward neural network without preprocessing Feedforward neural network with preprocessing Some results and conclusion(s)

Conclusions A novel analytically derived set of input signals for the speed neuroestimator has been proposed. It is feasible to use FFNN to estimate the induction motor angular velocity if a phenomenological-based signal preprocessing is introduced. Hint #1: If possible avoid recurrent neural networks in speed estimators. Hint #2: Try not to approach plant state estimation tasks as the black-boxed ones.

Bartlomiej Beliczynski, Lech Grzesiak and Bartlomiej Ufnalski Speed neuroestimation in induction motor drives

Problem statement Feedforward neural network without preprocessing Feedforward neural network with preprocessing Some results and conclusion(s)

Conclusions A novel analytically derived set of input signals for the speed neuroestimator has been proposed. It is feasible to use FFNN to estimate the induction motor angular velocity if a phenomenological-based signal preprocessing is introduced. Hint #1: If possible avoid recurrent neural networks in speed estimators. Hint #2: Try not to approach plant state estimation tasks as the black-boxed ones. Hint #3: Do not force your FFNN to act as a low-pass filter. She’s just a universal function approximator.

Bartlomiej Beliczynski, Lech Grzesiak and Bartlomiej Ufnalski Speed neuroestimation in induction motor drives

Problem statement Feedforward neural network without preprocessing Feedforward neural network with preprocessing Some results and conclusion(s)

Conclusions A novel analytically derived set of input signals for the speed neuroestimator has been proposed. It is feasible to use FFNN to estimate the induction motor angular velocity if a phenomenological-based signal preprocessing is introduced. Hint #1: If possible avoid recurrent neural networks in speed estimators. Hint #2: Try not to approach plant state estimation tasks as the black-boxed ones. Hint #3: Do not force your FFNN to act as a low-pass filter. She’s just a universal function approximator. Hint #4: Use your knowledge (even if incomplete) of the plant while synthesizing a neuroestimator – turn the task at least into a gray-boxed one.

Bartlomiej Beliczynski, Lech Grzesiak and Bartlomiej Ufnalski Speed neuroestimation in induction motor drives

Problem statement Feedforward neural network without preprocessing Feedforward neural network with preprocessing Some results and conclusion(s)

Questions?

Thank you for your kind attention!

And please do not hesitate to contact us at

bartlomiej.ufnalski@ee.pw.edu.pl .

This presentation is already available at

www.ufnalski.edu.pl

along with the relevant models/codes published at

www.mathworks.com/matlabcentral/profile /authors/2128309-bartlomiej-ufnalski .

Please cite the accompanying paper using the following BIBT EX entry

@INPROCEEDINGS{EANN_Ufnalski_2015, author={Beliczynski, B. and Grzesiak, L. M. and Ufnalski, B.}, booktitle={Proc. of the 16^{th} International Conference on Engineering Applications

• f Neural Networks (EANN)}, title={Mixed phenomenological and neural approach

to induction motor speed estimation}, pages={1--10}, month={Sept.}, year ={2015}, address={Island of Rhodes}} . Bartlomiej Beliczynski, Lech Grzesiak and Bartlomiej Ufnalski Speed neuroestimation in induction motor drives

Problem statement Feedforward neural network without preprocessing Feedforward neural network with preprocessing Some results and conclusion(s)

MATLAB Central

Bartlomiej Beliczynski, Lech Grzesiak and Bartlomiej Ufnalski Speed neuroestimation in induction motor drives

Problem statement Feedforward neural network without preprocessing Feedforward neural network with preprocessing Some results and conclusion(s)

It is my pleasure to announce that the 19th Conference on Power Electronics and Applications, EPE’17-ECCE Europe, will take place in Warsaw, Poland! Do not miss it! One thousand papers expected! Included in the Web of Science (WoS). Computational intelligence in control as one of the flagship topics. Guess why :-)

Bartlomiej Beliczynski, Lech Grzesiak and Bartlomiej Ufnalski Speed neuroestimation in induction motor drives