1. Title and Contents Contents:
Evolutionary Algorithm for Optimisation of Condition Monitoring and Fault Prediction Pattern Classification in Offshore Wind Turbines
J. Giebhardt
Institut fuer Solare Energieversorgungstechnik, ISET e.V, Kassel, Germany
Division Energy Conversion and Control Engineering
Contents:
Rotor faults in scope
Fuzzy classifier definition
Input and Output Pattern
Evolutionary Algorithm
First results
Conclusions and Outlook

2. Rotor Faults in Scope for Pattern Classification
Rotor mass imbalance
 Caused by loose material, penetrating water, icing, ...
 Excites transverse nacelle oscillation at rotor frequency
Aerodynamic rotor asymmetry
 Caused by pitch angel adjustment failures, pitch drive failures, ...
 Excites torsional nacelle oscillation at rotor frequency

3. Physical effects of rotor mass imbalance
Perfectly mass balanced rotor
Centrifugal forces of blades compensate when:
 No excitation of periodic nacelle oscillations
Mass imbalance
„Virtual“ mass mR and distance rR cause resulting centrifugal force:
 Excitation of periodic nacelle oscillations transverse to rotor axis with rotational („1p“) frequency

4. Physical effects of rotor aerodynamic asymmetry
Perfectly symmetric rotor
No excitation of periodic torsional nacelle oscillations (with respect to the vertical tower axis)
Aerodynamic asymmetry
 Excitation of torsional periodic nacelle Oscillations with 1p frequency caused by different thrust forces of the individual blades

5. Test Data as Input Pattern
Experimental data:
a) actual electrical power output
b) 1p amplitude of transverse nacelle oscillation (band pass filtered)
c) 1p amplitude of torsional oscillation at tower base (band pass filtered)

6. Training Data as Input and Output Pattern
Decreasing Probability
Increasing Probability

7. Fuzzy Classifier: Fuzzy Inference System (FIS)
Fuzzyfication
Rule base
Inference/Defuzzyfication
Inference:
IVy_small = min(µsmall (x1), µmedium (x2))
= 0.2
IVy_big = min(µ medium (x1), µbig (x2))
= 0.4
Rule 1 if
x1 = small
and
x2 = medium
then
y = small
Defuzzyfication:
Output value y is calculated as the “center of gravity” of the triangle shaped defuzzyfication functions
Rule 2 if
x1 = medium
and
x2 = big
then
y = big
x1 = 0.4
µsmall (x1)=0.2
µmedium (x1)=0.8
µbig (x1)=0.0
x2 = 0.7
µsmall (x2)=0.0
µmedium (x2)=0.6
µbig (x2)=0.4

8. Fuzzy Classifier: Overall Structure
Input Pattern:
Transfer Function:
y=(x, p)
Classifier
Parameter
Vector p
Fuzzy Classifier
Measured process data from a WT
Output Pattern:
Output vector
y=(y1, y2, y3, y4)
Data processing: 1p-filtering
Optimised by
Evolutionary Algorithm
Representation as probabilities for fault conditions
Data normalisation for input pattern generation:
Input vector
x=(x1, x2, x3)

9.  Rule Base Generation Rule Base Parameters Switching Parameters
OUT1 small
OUT1 medium
OUT1 big
then
Rule 1: If IN1 small and IN2 small and IN3 small
OUT1 small
OUT1 medium
OUT1 big
then
Rule 2: If IN1 medium and IN2 small and IN3 small
OUT1 small
OUT1 medium
OUT1 big
then
Rule 27: If IN1 big and IN2 big and IN3 big
 Rule Base Generation

10. Shaping Parameters Membership Functions Defuzzyfication Functions
Width (bS, bM, bB) and center abscissa values (mS, mM, mB) of triangle shaped defuzzyfication functions
Parameters:
Abscissa values of inflection points
KS1, KS2 for µsmall (x)
KM1, KM2 , KM1 for µmedium (x)
KB1, KB2 for µbig (x)

11. Evolutionary Optimisation
Evolutionary Algorithm
Flow Diagram
Random setup of 1st
parameter generation
Calculation of
individuals fitness
Evolutionary manipu- lation of individuals no
Ranking of individuals
(decreasing fitness)
winner
fitness >Thrhld
STOP
yes

12. Detection of a mass imbalance

13. Detection of a undefined condition

14. Detection of a aerodynamic asymmetry

15. Conclusions and Outlook
Principle concept (evolutionary optimised Fuzzy-Classifier) works
Rule base optimisation works in principle
Calculation time of algorithm is reasonable (some minutes)
Rule base optimisation has to be extended by shaping parameter optimisation to achieve optimum fault recognition results
Outlook / Next Steps:
Extension of the optimisation algorithm (shaping parameters)
Investigation of the algorithm’s stability
Verification of the algorithm’s parameter sensitivity (e. g. number of individuals, gene manipulation rates, …)