7MW LDT Pitch System Health monitoring with Machine Learning

2020 VirtualWind II ORE Catapult & Fraunhofer IWES

08, July, 2020 Hyunjoo Lee

GLASGOWORE Catapult

Agenda

• Study approach and conditions/assumptions• Target turbine• Operational characteristics and damages• Investigation concept• Pitch cycle and load estimation method• Bearing life estimation method• Data sets and conditions for study

• Study result and observations• Measured wind speeds & power production• Bearing life result comparison• Loads & Pitch cycle comparison• Change history – IPC control algorithm• Wear risk indication

• Closing remarks / Next works

Target Turbine

• 7MW Levenmouth Demonstration Turbine (LDT)• Developed by SHI (25 years design life)• Collective and individual pitch controlled by

internal algorithm developed by DNV/GL-GH• Currently owned by OREC for research purpose

R&D / Testing assets

Head office

Operational Characteristics and Damages

• Wind turbine pitch bearings are subject to very hostile operating conditions:

• Huge bending moments whilst low speed oscillating or standing still

• Large structural deformation due to limited stiffness• Transient events, i.e. start / stop• IPC is becoming common in wind turbine (frequent

small amplitude oscillation)

• Associated possible damages:• Ring fracture• Contact ellipse truncation• Raceway fatigue• False brinelling / Fretting corrosion

Photos reference: RBB engineering

Pitch system failures account for most downtime Failures are even more critical on offshore 10 MW+ turbines No effective detection method for whole pitch system

Pitch System failure & Approach

SCADA based Pitch system health monitoring with Machine Learning

Project aim : Pitch system health monitoring with existing SCADA- No extra hardware- Automatic Warning/Alarm in advance- Reduce the unplanned maintenance

Source: windeurope.org

Investigation Concept – Cause of issue

Time-series simulation

Oscillation cycle counting algorithm

CONTROL STRATEGY

Aeroelastic simulation(Bladed model)

SCADA signals from actual turbine

Pitch bearing life with IPC

Control strategy feed back/Validation

Bearing performance & life simulation

-1

-0.5

0

0.5

1

1.5

2

2.5

3

3.5

4

4.5

300 310 320 330 340 350 360 370 380 390 400

Pitc

h an

gle [

deg]

Time [s]

Pitch angle - SCADA

SubPtchPosition1 [deg]

SubPtchPosition2 [deg] - 1 deg offset

SubPtchPosition3 [deg]

Pitch bearing condition

• Fatigue damage• Wear damage

Comparison/correlation

Pitch Cycle and Load Estimation Method

• Recent reference from IWES:• Cycle counting of roller bearing oscillations – case study of wind turbine individual pitching system,

2018• Oscillation cycle counting achieved by range-pair counting method • Load estimation supplemented by bin counting method

• Cycle counting MatLab code developed by OREC project team with the suggested core algorithm• Handling the time series input from both Bladed simulation and SCADA data signal

Cycle counting concept (range-pair)

Bearing Life Estimation Method

• Pitch bearing is subject to oscillation movement, so life calculation method for oscillating bearing is required for this investigation

• Calculation method suggested by NREL (Harris method) was used for this study:- Wind Turbine Design Guideline DG03: Yaw and Pitch Rolling Bearing Life, 2009

Amplitude angle of one cycle oscillation

Critical amplitude angle,

in million cycles

when

when

Conditions and Dataset for Study

• Following conditions and dataset were considered just for comparison purpose, not for the suitability assessment toward 25years design life:

• Bearing reference life with the modification factors of 1.0• Mx and My only (major contributory load components)• One year SCADA data (2017) under normal power production• 1m/s step wind speed bin, 3 SCADA datasets randomly chosen over each wind speed bin (total 36

datasets to be analysed)• Wind field to be re-generated in simulation based on the wind speed measured at the hub centre• Very small variation less than the amplitude of 0.05 deg to be considered as stationary condition

for bearing• Double-row 4 point contact ball bearing, approximately 4.4m size

Measured Wind Speed Distribution

• Occurrence of 10 min average wind speed was recorded from LDT operation during 2017• Total 36 datasets that were analysed representing normal power production, cover around 73%• Probability of each mean wind speed to be used for one-year cycle counting and bearing life calculation

0.73

4%

3.79

4%

7.06

8%

7.09

1%

8.03

3% 9.90

6%

13.1

70%

11.4

08%

10.3

02%

9.01

0%

6.25

6%

4.45

4%

3.06

0%

2.17

9%

1.43

8%

0.92

1%

0.53

3%

0.27

5%

0.15

3%

0.09

9%

0.04

6%

0.02

0%

0.02

1%

0.01

3%

0.00

4%

0.00

3%

0.00

4%

0.00

3%

0.00

2%

0.00

1%

B 1 B 2 B 3 B 4 B 5 B 6 B 7 B 8 B 9 B 1 0 B 1 1 B 1 2 B 1 3 B 1 4 B 1 5 B 1 6 B 1 7 B 1 8 B 1 9 B 2 0 B 2 1 B 2 2 B 2 3 B 2 4 B 2 5 B 2 6 B 2 7 B 2 8 B 2 9 B 3 0

PROB

ABILI

TY [%

]

MEAN WIND SPEED

SCA DA WIN D SPEED DISTR IB UTIO N (2 0 1 7 O N E YEA R )

Power Production Comparison

• Wind field was re-generated in simulation based on the wind speed measured at the hub centre• Consistent wind speed between SCADA and simulation• Good correlation in power production with small localized deviation

0

5

10

15

20

25

0 100 200 300 400 500 600

Win

d Sp

eed

[m/s

]

Time [s]

Wind Speed

WindSpeed_mps (SCADA)HubWindSpeedMagnitude_m/s (Bladed)

0

1000

2000

3000

4000

5000

6000

7000

8000

0 100 200 300 400 500 600

Pow

er [k

W]

Time [s]

Power

Power_kW (SCADA)

ElectricalPower_kW (Bladed)

Example comparison result; 05/30/2017 16:20

Bearing Life with Measured Dataset (SCADA)

• Accumulated pitch bearing fatigue damage from one year operation was estimated to 1.2% based on the measured SCADA datasets

Index, i

Amplitude range, ϴ

[deg]No. of half cycles, ni

Operational time, ti

[%]

Mean amplitude, ϴi

[deg]

Mean frequency, fi

[Hz]

Equv. Loads, Pea_i[kN]

Capacity, Ca,osc_i

[kN]L10h,osc_i

[h]Damage, D

[%]1 0.05 - 0.08 75661 5.40% 0.064 0.098 6510.1 35165 446718 0.024%2 0.08 - 0.2 270271 18.27% 0.135 0.104 6534.9 28118 213756 0.170%3 0.2 - 0.5 531292 27.00% 0.342 0.138 6394.7 21260 74152 0.723%4 0.5 - 1 496469 25.86% 0.707 0.134 6160.8 17094 44177 1.162%5 1 - 1.5 199939 11.05% 1.200 0.127 5833.6 14586 34283 0.640%6 1.5 - 2 87926 6.01% 1.713 0.102 5344.3 13110 40049 0.298%7 2 - 3 51263 4.11% 2.264 0.087 5076.9 12057 42624 0.191%8 3 - 4 5753 0.56% 3.364 0.071 6397.6 10707 18246 0.061%9 4 - 5 665 0.07% 4.259 0.063 5457.5 9521 23601 0.006%

10 5 - 10 2276 0.29% 5.304 0.054 4515.1 8850 38594 0.015%11 10 - 15 1579 0.62% 11.370 0.018 2274.5 6864 427444 0.003%12 15 - 20 1579 0.75% 16.997 0.015 2750.3 6003 196370 0.008%13 20 - 25 0 0.00%14 25 - 30 0 0.00%15 30 - 90 0 0.00%

sum 1724672 100% L10h,osc [h]: 60142 3.301%

Considering the pitch activation ratio of 34.92%, the damage of 1.15% is estimated for one year (L10,osc,act [h]: 172,209)

Bearing Life with Simulated Dataset (Bladed)

• Accumulated pitch bearing fatigue damage from one year operation was estimated to 11.4% based on the corresponding simulated datasets

• Approximately 10 times more damage than that from SCADA dataset – WHY?

Index, i

Amplitude range, ϴ

[deg]No. of half cycles, ni

Operational time, ti

[%]

Mean amplitude, ϴi

[deg]

Mean frequency, fi

[Hz]

Equv. Loads, Pea_i[kN]

Capacity, Ca,osc_i

[kN]L10h,osc_i

[h]Damage, D

[%]1 0.05 - 0.08 141420 4.57% 0.065 0.166 8181.3 34981 130911 0.090%2 0.08 - 0.2 251421 12.04% 0.132 0.112 8140.1 28296 104320 0.299%3 0.2 - 0.5 228530 18.84% 0.337 0.065 8291.8 21345 72983 0.670%4 0.5 - 1 180950 14.55% 0.730 0.067 8285.2 16935 35637 1.059%5 1 - 1.5 148474 4.50% 1.268 0.176 8315.3 14347 8084 1.445%6 1.5 - 2 152520 4.13% 1.736 0.198 8111.8 13057 5855 1.829%7 2 - 3 408466 12.11% 2.525 0.181 7835.0 11670 5082 6.181%8 3 - 4 473408 14.46% 3.477 0.175 7604.2 10601 4294 8.737%9 4 - 5 308828 9.63% 4.456 0.172 7329.5 9379 3389 7.369%

10 5 - 10 162188 5.17% 5.607 0.168 7352.4 8688 2726 4.916%11 10 - 15 0 0.00%12 15 - 20 0 0.00%13 20 - 25 0 0.00%14 25 - 30 0 0.00%15 30 - 90 0 0.00%

sum 2456205 100% L10h,osc [h]: 7958 32.596%

Considering the pitch activation ratio of 34.92%, the damage of 11.38% is estimated for one year (L10,osc,act [h]: 22,788)

Loads Comparison

• Simulated loads were approximately 1.5 times higher based on average value

Example time-series comparison result; 05/30/2017 16:20

6510

.1

6534

.9

6394

.7

6160

.8

5833

.6

5344

.3

5076

.9 6397

.6

5457

.5

4515

.1

2274

.5

2750

.3

8181

.3

8140

.1

8291

.8

8285

.2

8315

.3

8111

.8

7835

.0

7604

.2

7329

.5

7352

.4

0.05

-0.

08

0.08

-0.

2

0.2

-0.5

0.5

-1

1 -1

.5

1.5

-2

2 -3

3 -4

4 -5

5 -1

0

10 -

15

15 -

20

20 -

25

25 -

30

30 -

90

Pea_

i [kN

]

Amplitude range, [deg]

Equivalent Loads Distribution

Equv. Loads, SCADA [kN] Equv. Loads, Bladed [kN]

-15000

-10000

-5000

0

5000

10000

15000

0 100 200 300 400 500 600Mx

[kN

m]

Time [s]

Edgewise MomentMx [kNm], SCADA

Mx [kNm], Bladed

0

5000

10000

15000

20000

25000

30000

0 100 200 300 400 500 600

My

[kN

m]

Time [s]

Flapwise MomentMy [kNm], SCADA

My [kNm], Bladed

Pitch Cycle and Amplitude

• Reduced total cycles for SCADA, most frequent amplitude shifted toward far smaller range – WHY?• Consistent collective pitch behaviour but large difference in amplitude under IPC

Example time-series comparison result; 05/30/2017 16:20

7.6E

+04

2.7E

+05

5.3E

+05

5.0E

+05

2.0E

+05

8.8E

+04

5.1E

+04

5.8E

+03

6.7E

+02

2.3E

+03

1.6E

+03

1.6E

+03

0.0E

+00

0.0E

+00

0.0E

+00

1.4E

+05

2.5E

+05

2.3E

+05

1.8E

+05

1.5E

+05

1.5E

+05 4.

1E+0

5

4.7E

+05

3.1E

+05

1.6E

+05

0.0E

+00

0.0E

+00

0.0E

+00

0.0E

+00

0.0E

+00

0.05

-0.

08

0.08

-0.

2

0.2

-0.5

0.5

-1

1 -1

.5

1.5

-2

2 -3

3 -4

4 -5

5 -1

0

10 -

15

15 -

20

20 -

25

25 -

30

30 -

90

ni [c

ycle

s]

Amplitude range, [deg]

Half Cycle Counting Distribution

No. of half cycles, SCADA No. of half cycles, Bladed

-4-202468

1012141618

0 100 200 300 400 500 600

Pitc

h po

sitio

n [d

eg]

Time [s]

Pitch Position SubPtchPosition1_Deg (SCADA)SubPtchPosition2_Deg (SCADA)SubPtchPosition3_Deg (SCADA)Blade1PitchAngle_deg (Simulation)

SCADA: 1.7E06 cyclesSimulation: 2.4E06 cycles

Evolution – IPC Control Algorithm

• Controller tuning from prototype commissioning (2012)• IPC operation > increase of actuator motor current > excessive current over the rated current > current

cut-off > fault event• Improvement action: IPC amplitude factor introduction to IPC control algorithm – Continuous

adjustment of IPC amplitude against actuator motor current signal• The introduced IPC amplitude factor is not applied to Bladed simulation model, leading to the major

difference in IPC behaviour that have been observed

IPC amplitude factor

Utilising the Power of A.I. (Machine Learning)

There can be 1,000s of non-linear turbine operating modes Analysis is too difficult using conventional method Currently most collected SCADA data is unused

=> We apply advanced machine learning to detect abnormal condition

0 500 1000 1500 2000 2500 3000 3500 4000 45000

0.2

0.4

0.6

0.8

1

Ale

rt:

Hig

h S

ide-S

ide V

ibra

tion

0 500 1000 1500 2000 2500 3000 3500 4000 4500-10

-5

0

5x 10-3

Gearb

ox V

ibra

tion (

Mean)

0 500 1000 1500 2000 2500 3000 3500 4000 45000

0.010.020.03

0.040.05

0.06

Gearb

ox V

ibra

tion (

Std

ev)

Sample Number

(a)

(b)

(c)

10 20 30 40 50 60 70 80

Sensory Signal sample

10 20 30 40 50 60 70 80

10 20 30 40 50 60 70 80

Continuous Alert/Error Status Signal

ON OFF ON OFF

Sampling Time

SCADA Alert Status

Ale

rt/E

rror

St

atus

ON

OFF

10 20 30 40 50 60 70 80

Ale

rt/E

rror

St

atus

ON

OFF

Synchronised Alert/Error Status

SCA

DA

Log

Sens

ory

valu

es (a)

(b)

(c)

(d)

Sample time (minutes)

Closing Remarks

• Process to evaluate the influence of turbine controller on pitch bearing performance defined

• Case study showed the evaluation result with the simulated dataset largely deviated from that with the measured dataset (SCADA):

• Observed difference in pitch behaviour under IPC was majorly caused by previous control algorithm improvement action with IPC amplitude factor introduction (not applied to simulation controller), leading to less total cycles having smaller amplitude

• With lower measured moment loads, resulting in much less fatigue damage for actual turbine. However, increasing the portion of possible wear regime

• Advanced machine learning algorithm is applied to detect abnormal condition

• Note that the case can differ over turbines due to different controller

• Recommend to monitor pitch system to evaluate the design life

Thank You For Your Attention!

Contact us

ore.catapult.org.uk@orecatapult

GLASGOW BLYTH LEVENMOUTH HULL ABERDEEN CORNWALL PEMBROKESHIRE CHINA

Email us: info@ore.catapult.org.uk Visit us: ore.catapult.org.uk

Engage with us: