Online Prognostic Condition

Monitoring for Wind Turbines Using

Current Signals

Wei Qiao, Ph.D.

Professor, Department of Electrical and Computer Engineering
Director, Power and Energy Systems Laboratory
University of Nebraska–Lincoln
Lincoln, NE 68588-0511 USA
Tel: +1 (402) 472-9619
E-mail: wqiao3@unl.edu
UNL: http://pesl.unl.edu
April 25, 2019
 Wind Turbine Failure

Failure frequencies of major wind turbine

subsystems and downtime caused by
failures of these subsystems.

A typical Type 3 wind turbine with main subsystems

shown (“Condition diagnostic for wind parks,”
Rovsing Dynamics, Herlev, Denmark, Nov. 2008)

 P. Tavner, “How Are We Going to Make Offshore Wind Farms More Reliable?” presented at the 2011 SUPERGEN Wind
General Assembly, Durham University, United Kingdom, March 20, 2011.

Power & Energy Systems Laboratory 2

 Wind Turbine Failure (2)
Wind turbines: situated on high towers, installed in remote areas,
distributed over large geographic regions, subject to harsh environment
and relatively high failure rates

Gear teeth: macropitting (Source: Blade damage: lighting strike (Source: Blade: erosion on the leading edge
North American Windpower) Sandia National Laboratories) (Source: Sandia National Laboratories)

Damage to the generator windings due to

over-speed and subsequent overheating of
the windings (Source: Metropolitan
Main shaft bearing: wear
Engineering Consulting and Forensics)
tracks on raceway Source: National Association of Fire Investigators

Power & Energy Systems Laboratory 3

 Major Failure Modes in Wind Turbines

Our work has been focused on fault diagnosis and prognosis for bearings,
blades, rotors/shafts, and generators of various wind turbines, gearboxes
of indirect-drive wind turbines, and power electronic converters
W. Qiao and D. Lu, “A survey on wind turbine condition monitoring and fault diagnosis−Part I: Components and
subsystems,” IEEE Trans. Industrial Electronics, vol. 62, no. 10, pp. 6536-6545, Oct. 2015.

Power & Energy Systems Laboratory 4

 Background and Motivation
 High O&M costs: 10-15% LCOE of onshore and 20-35%
LCOE of offshore wind turbines
 Online condition monitoring, fault diagnosis and prognosis
Condition monitoring: a process of monitoring operating parameters of
wind turbines
Fault diagnosis: detect, locate, and identify occurring faults and monitor
the development of the faults from defects (incipient faults)
Fault prognosis: predict the development of a defect into a failure
Remaining useful life (RUL) prediction: predict when a failure occurs
making the wind turbine inoperable and RUL of the faulty component
 An anticipated condition monitoring system (CMS): online
diagnose faults/defects at an early stage and predict fault
development and RUL of the faulty component to enable
condition-based preventive/predictive maintenance
Improve wind turbine reliability, availability, and lifespan
Reduce wind turbine failure rate, downtime, and O&M costs
Power & Energy Systems Laboratory 5
 Background and Motivation (2)
 Current wind turbine CMSs and techniques: require additional sensors
(vibration, temperature, acoustic, etc.) and data acquisition devices
Sensors are mounted on the surface or buried in the body of wind
turbine components, difficult to access during wind turbine operation
Increased cost and hardware complexity of wind turbine systems
Condition monitoring capability depends on sensor locations
Sensors and devices are inevitably subject to failure, causing additional
problems with system
reliability and additional
O&M costs
Only work when wind
turbines operate in nearly
constant speed conditions
No RUL prediction capability W. Teng, et. al, Renewable Energy, 93: 591-598, 2016.

It is desired to develop nonintrusive, lower-cost, more reliable, and more

capable prognostic condition monitoring techniques for wind turbines
W. Qiao and D. Lu, “A survey on wind turbine condition monitoring and fault diagnosis−Part II: Signals and signal
processing methods,” IEEE Trans. Industrial Electronics, vol. 62, no. 10, pp. 6546-6557, Oct. 2015.

Power & Energy Systems Laboratory 6

 Our Techniques for Online Fault Diagnosis and Prognosis
We are the first to reveal the physical and mathematical principles of using
generator current signals for fault diagnosis and prognosis and RUL
prediction of wind turbines under nonstationary operating conditions.
We are the first to have developed a suite of techniques using current
signals acquired from generator terminals or in the generator/power
converter control systems and
combining current and vibration
signals for online fault
diagnosis and prognosis
and RUL prediction of
wind turbines in all
operating conditions
Our techniques were
verified on wind turbines
of different types and
different sizes from kW
to MW scale

Power & Energy Systems Laboratory 7 7

 Advantages of Our Techniques
 No or lower cost: no additional sensors or data acquisition devices needed
 Higher flexibility and nonintrusive: flexible sensor locations and configurable
to meet different requirements
 Work in all operating conditions
 Fault prognosis and RUL prediction: enable condition-based predictive
maintenance with optimal logistics and minimum cost
 Higher accuracy
and reliability

Power & Energy Systems Laboratory 8 8

Physical/Theoretical Principle: Frequency/Amplitude Modulation
 Wind turbine generator current signal: Cs (t ) = I s (t ) ⋅ sin   2π ⋅ f1 (t ) ⋅ dt 
 A failure in wind turbine causes shaft torque vibration at a certain frequency
ffault (fault characteristic frequency), which can be detected by vibration
sensors
T (t ) = T0 (t ) + Av ⋅ cos   2π ⋅ f fault (t ) ⋅ dt 

 The shaft torque vibration at the fault characteristic frequency will modulate
frequency and amplitude of generator current signals: due to mechanical
couplings between generator and failed wind turbine component(s) as well
as electromagnetic coupling between generator rotor and stator
 Current frequency modulation:
f1 (t ) = f1, w (t ) + A1,v (t ) ⋅ sin   2π ⋅ f fault (t ) ⋅ dt + ϕ f (t ) 
 Current amplitude modulation:
I s (t ) = I s , w (t ) + As ,v (t ) ⋅ sin   2π ⋅ f fault (t ) ⋅ dt +ψ f (t ) 

X. Gong and W. Qiao, “Bearing fault diagnosis for direct-drive wind turbines via current
demodulated signals,” IEEE Trans. Industrial Electronics, vol. 60, no. 8, pp. 3419-3428, Aug. 2013.
Power & Energy Systems Laboratory 9
Challenges in Current-Based Fault Diagnosis & Prognosis
 A single fault characteristic frequency in vibration becomes multiple fault
characteristic frequencies in current due to frequency/amplitude
modulations
 Excitations at fault characteristic frequencies in current could be masked by
subbands of the dominant components that are irrelevant to fault in the frequency
spectrum of current
 Low signal-to-noise ratio (SNR): the total energy of excitations related to a fault is
dispersed into multiple fault characteristic frequencies -7
x 10
4

 Fault characteristic frequencies: nonstationary 3.5

 challenging to extract fault features 3

PSD of the stator current

 Depending on shaft rotating frequency (SRF) 2.5

(i.e., 1P frequency) 2

 Wind turbines: variable speed operation 1.5

 nonstationary fault characteristic frequencies 1

 cannot be effectively extracted by using 0.5

standard spectrum analysis Frequency (Hz)

0
28 30 32 34 36 38 40 42

 Current signals may contain both SRF-related and SRF-unrelated components

 Automated fault feature extraction and diagnosis
 Reliable fault prognosis & RUL prediction for condition-based maintenance
Power & Energy Systems Laboratory 10
 Accomplishments
 Signal conditioning: improve SNR
Current frequency and amplitude demodulation: Demodulated signals
explicitly contain fault characteristic frequency components
Synchronous (angular) resampling: Convert nonstationary fault features
to stationary
Multiscale filtering spectrum: Eliminate SRF-unrelated noise
 Fault feature extraction from conditioned signals
Frequency spectrum analysis: impulse detection
 Time-frequency domain analysis: Wavelet filter, Hilbert-Huang transform
 Statistical analysis
 Anomaly detection (outlier detection)
 Fault diagnosis: faulty type/location/severity identification
Probabilistic multiclass support vector machine-based pattern recognition
Deep classifier: deep network and multiclass support vector machine
 Statistical prediction and artificial intelligence based fault
prognosis and RUL prediction
Power & Energy Systems Laboratory 11
 Signal Conditioning: Adaptive Synchronous Resampling
Amplitude

0 S1 S2 S3 S4 S5 t

Phase (Degree)

300
240
180 C fb
120 = sin(θ )
60 I1
0 S1 S2 S3 S4 S5 t

 Synchronously resampled nonstationary sinusoidal signal and its phase

 Nonstationary: amplitude, frequency, and/or phase of the signal are variable
 The phases, instead of the time intervals, of the time-domain sampling points are
evenly distributed
 In each cycle of the synchronously resampled nonstationary sinusoidal signal, the
number of the sampling points is a constant
Power & Energy Systems Laboratory 12
Signal Conditioning: Adaptive Synchronous Resampling (2)

Calculate time locations of synchronous Calculate synchronous samples via

sampling points via linear interpolation linear interpolation

S (n + 1) − S (n) C (n + 1) − C (n)
S s (ls ) = S (n) + ⋅ [θ s (ls ) − θ (n)] Cs (ls ) = C (n) + ⋅ [ S s (ls ) − S (n)]
θ (n + 1) − θ (n) S (n + 1) − S (n)

X. Gong and W. Qiao, “Current-based mechanical fault detection for direct-drive wind turbine via synchronous sampling
and impulse detection,” IEEE Trans. Industrial Electronics, vol. 62, no. 3, pp. 1693-1720, Mar. 2015.

Power & Energy Systems Laboratory 13

Evaluation: Test Setup
Test bearing

(1) A Southwest Windpower Air Breeze wind turbine was

used in the experiment
(2) The testing bearing is located between the rotors of the
turbine and the generator
(3) The bearing was pretreated by removing the lubricant
to accelerate the failure process
(4) The wind turbine was operated in variable speed
condition (500-700 rpm) in the wind tunnel for 25 hours
Power & Energy Systems Laboratory 14
 Signal Conditioning: Bearing Fault
Test bearing

Broken cage
The test bearing before and after experiment

Healthy
bearing

PSD of the original nonstationary current signal

 Theoretical fault characteristic frequencies
in stator current: ~60±4n Hz (n = 1, 2, ∙∙∙)
 The wind turbine was operated with varying
rotor speed for 25 hours
Power & Energy Systems Laboratory 15
 Signal Conditioning: Blade Imbalance
Current
4.5
Baseline
4 Blade Imbalance 1%

PSD of the estimated shaft rotating frequency

Blade Imbalance 2%
3.5 Blade Imbalance 3%
Blade Imbalance 4%
3

2.5

1.5

0.5

 Experiments were performed in four 0

scenarios with the mass density of one 10
1

blade increased by 1.25%, 2.5%, 3.75%, Frequency (Hz)

and 5%; while the mass densities of the Blade imbalance

other two blades are held constant
 The variable 1P frequency is converted to a constant value of 10 Hz
 Health case: no excitation observed at 1P in the PSD curve
 Blade imbalance cases: magnitude of 1P excitation increases with the increase of
degree of imbalance
Power & Energy Systems Laboratory 16
 DFIG Wind Turbine Drivetrain Test Rig

DFIG: doubly-fed inductor generator

Power & Energy Systems Laboratory 17

 Signal Conditioning: Gear Crack
-7
x 10
1000 100% 4
810 Estimated speed
Estimated load 3.5

Estimated shaft rotating speed (RPM)

805
800 80% 3

PSD of the stator current

25 25.5

Estimated load (%)

2.5

600 60% 2

1.5

400 40% 1

0.5

200 20% 0
0 20 40 60 80 100 28 30 32 34 36 38 40 42
Time (s)
Frequency (Hz)

Test gear with a crack Estimated shaft speed and load Current frequency spectrum
obtained from the classic FFT
50
50
fs
fs
PSD of resampled stator current

40

PSD of resampled stator current

40
f s-f 1 f s+f 1

30
30

f s-f 3 f s-f 2
f s+f 2 f s+f 3
20 f s-f 1 f s+f 1 20
fs-f 3
f s-f 2
f s+f 2 f s+f 3
10 10

0 0
28 30 32 34 36 38 40 42 28 30 32 34 36 38 40 42
Frequency (Hz) Frequency (Hz)

Current PSD spectrum of healthy gearbox Current PSD spectrum of faulted gearbox

Power & Energy Systems Laboratory 18

 Field Test for 1.6-MW DFIG Wind Turbines

Turbine # Gearbox condition

1 May contain defect(s)
2 May contain defect(s)
3 Healthy (baseline)

 Data
 8 channels: 2 stator voltages, 3 stator currents, and 3 rotor currents
 Data length: 20 seconds
 Sampling rate: 10 kHz
 Time interval between two consecutive datasets: 15 minutes
Power & Energy Systems Laboratory 19
 Without Synchronous Resampling
 PSD spectrum of the
original one-phase stator
current in one data
record of No. 1 turbine
around the fundamental
frequency
 A constant fundamental
frequency of 60Hz is
clearly observed in the
spectrum
 Only smeared frequency ranges instead of specific characteristic
frequencies related to fault can be observed due to the varying DFIG
shaft rotating frequency
 Without using the proposed method, it is hard to extract fault features for
the wind turbines operating in varying speed conditions

F. Cheng, L. Qu, W. Qiao, C. Wei, and L. Hao, “Fault diagnosis of wind turbine gearboxes based on DFIG stator current
envelope analysis,” IEEE Trans. Sustainable Energy, in press.

Power & Energy Systems Laboratory 20

 Result of Steady State Check Method
 Steady state check: record signals during a nearly steady-state condition
 Steady state check criteria
 Variation of generator speed is less than 20 rpm.
 Variation of stator current is less than 20%.
 PSD spectra of isA(t) of the three wind turbines under a nearly steady state
condition during which the DFIG shaft rotating speed is in 1485-1490 rpm
 The HSS upwind bearing rolling element fault characteristic frequency in
stator current, fs + fB, is around 180.4 Hz, which is close to the 3rd harmonic.
 Although the amplitudes of the PSD spectra of No. 1 and No. 2 turbines at
180 Hz are higher than that of No. 3 turbine, it is hard to tell whether the
difference is caused by a rolling element fault or the 3rd harmonic

Power & Energy Systems Laboratory 21

 Result of Synchronous Resampling-Based Method
 The ball and cage fault characteristic frequencies of the bearing located at
the upwind side of the HSS, fB and fC, are 2.43fHSS and 0.4fHSS, respectively.
 No. 1 turbine
 Has an HSS upwind bearing rolling element fault since impulses appear at fB and
its multiples in its PSD spectrum
 Has a gear fault or shaft misalignment in the HSS since an impulse appears at
fHSS in its PSD spectrum
 No. 2 turbine
 Has an imbalance fault on the generator shaft since impulses appear at 3mfHSS
(m = 1, 2, 3, ⋯) in both low and high frequency ranges of the PSD spectrum
 Has an HSS upwind bearing cage fault since impulses appear at 3mfHSS ± fC in
the PSD spectrum and are much higher than the other two turbines
-10
No.1 Turbine
No.2 Turbine
-20 No.3 Turbine

-30

-40

-50

-60
0 30 60 90 120 150 180 210 240 270 300
Frequency (Hz)

Power & Energy Systems Laboratory 22

 Signal Conditioning: Angular Resampling
Time-frequency SRF-unrelated: fn(t)
distribution (TFD)
Short-time Fourier 7fr(t)
Simulated transform
signal Frequency-domain spectrum
4.5fr(t)
3fr(t)
SRF: fr(t)

Angular
𝑓 𝑡 cos 0.2𝜋𝑡 𝜋 ⁄3 10
resampling

Resampled The order 7Or is

signal submerged in the SRF-
Order-domain spectrum
unrelated component On

 Advantage: the SRF-related components can be revealed in the resultant order-

domain spectrum.
 Disadvantage: the spectrum smearing problem remains in the SRF-unrelated
components and noise, which smears the SRF-related components.
Power & Energy Systems Laboratory 23
 Signal Conditioning: Multiscale Filtering Spectrum
 Multiscale filtering spectrum (MFS): decompose the measured signal into SRF-
related mono-component signals within a local order range via Vold-Kalman filter
(VKF)-based multiscale filtering and then calculate the weighted energy spectrum.
 Multiscale filter bank construction
 Determine a series of center frequencies
corresponding to the SRF at different scales.
 VKF-based multiscale filtering
 Extract the mono-component signals whose
frequencies are continuous multipliers of the
SRF via the VKF.
 Energy spectrum calculation
 𝐸𝑛 𝑠 𝑥 𝑛 .
 Weighting factor calculation
| |
 𝑤 𝑠 .
| | | |

 MFS construction
 𝐸𝑛 𝑠 𝐸𝑛 𝑠 ⋅ 𝑤 𝑠 .

J. Wang, Y. Peng, W. Qiao, and J. L. Hudgins, “Bearing fault diagnosis of direct-drive wind turbines using multiscale
filtering spectrum,” IEEE Trans. Industry Applications, vol. 53, no. 3, pp. 3029-3038, May-Jun. 2017.

Power & Energy Systems Laboratory 24

 Signal Conditioning Result: MFS
Simulated signal TFD SRF-unrelated: fn(t)
Short-time Fourier 7fr(t)
transform
4.5fr(t)
Frequency-domain spectrum
3fr(t)
MFS SRF: fr(t)

𝑓 𝑡 cos 0.2𝜋𝑡 𝜋 ⁄3 10

Intermediate results
Energy spectrum
The order 7Or is
interfered by On 3Or 4.5Or MFS MFS
Or
7Or On reveal 7Or
suppress On
Order
Weighting factor
7Or has large weighting Order-domain spectrum
3Or 4.5Or On
value Or
Angular resampling
On have small weighting 7Or submerged in On
values Order

 Advantages: the SRF-related components are enhanced; the SRF-unrelated

components and noise are suppressed.
Power & Energy Systems Laboratory 25
 Field Test Setup
 Experimental Setup
 Wind turbine: 2.4-kW Skystream 3.7 has run for more than 5 years
 Analyzed signal: stator current signal
 Data acquisition: based on wireless sensor network
 Sampling frequency: 1000 Hz
 Shaft bearing type: 6209
 Characteristic orders: OBPFI = 8.368 (inner race), OBPFO = 5.632 (outer race), OBSF =
4.923 (ball), OFTFI = 0.598 (cage inner ring), and OFTFO = 0.402 (cage outer ring)

Power & Energy Systems Laboratory 26

 Field Test Verification Result
Current signal
Shaft rotating frequency (SRF)

Frequency (Hz)
SRF-unrelated
Fundamental frequency
SRF-unrelated Result of angular resampling

Hilbert
Transform

Current envelope signal Compare

Energy spectrum Weighted energy spectrum

0.15
Multiscale filtering spectrum
0.1
OFTFI
0.05

Spectrum smearing problem 0

0 5 10 15 20
Order

 Is there any characteristic orders  Interfering orders are removed.

buried in this area?  No characteristic orders here.

Power & Energy Systems Laboratory 27

 Fault Prognosis and RUL Prediction
Developed a particle filtering algorithm (a recursive Bayesian algorithm) to
predict propagation and progression of the diagnosed faults and the RUL
F. Cheng, L. Qu, and W. Qiao, “Fault prognosis and remaining
useful life prediction of wind turbine drivetrain gearboxes using
current signal analysis,” IEEE Trans. Sustainable Energy, vol. 9,
no. 1, pp 157-167, Jan. 2018.

Particle filtering algorithm

Initialize particles

Adaptive neuro-fuzzy inference system Propagate particles

Resampling i= 1 2 N xˆki → xˆki +1

N
Fault Prognosis xˆk +1 =  wk xˆk +1
i i
Yes
i =1

No
Weights
Yes No
degenerated? New zk+1 RUL
available? Prediction

Normalize Update weights

weights i= 1 2 N
wki → wki +1

Power & Energy Systems Laboratory 28

Particle Filtering for Fault Prognosis and RUL Prediction
Step 1: Propagate
Initialize particles
xˆki +1 = f k ( xki , xki −1 , xki − 2 , xki −3 )+uk

Propagate particles f k : fault degradation model

Resampling i= 1 2 N xˆki → xˆki +1 N
xˆk +1 =  wki xˆki +1
i =1
N

Yes Fault Prognosis xˆk +1 =  wk xˆk +1

i i
Step 2: Update
i =1

No
wki +1 ∝ wki p( zk +1 | xˆki +1 ), i = 1, 2,, N
Weights − ( zk +1 − xˆki +1 )2
Yes No
degenerated? New zk+1 RUL 1 2 μ0
available? Prediction p( zk +1 | xˆki +1 ) = e
2πμ0
N
p( xk +1 | z1:k +1 ) ≈  wki +1δ ( xk +1 − xˆki +1 )
Normalize Update weights i =1
weights i= 1 2 N
w →w Step 3: Resample
i i
k k +1

Eliminate particles with lowest

F. Cheng, L. Qu, W. Qiao, and L. Hao, “Enhanced particle filtering weights and generate a set of
for remaining useful life prediction of a wind turbine drivetrain
gearbox,” IEEE Trans. Industrial Electronics, vol. 66, no. 6, pp. new particles by duplicating
4738-4748, Jun. 2019. the remaining particles
Power & Energy Systems Laboratory 29
 Adaptive Neuro-Fuzzy Inference System (ANFIS)
 Problem: Lack of accuracy physical fault degradation model
 Solution: Use ANFIS to learn the fault degradation model

xk +1 = O4j
j

1 O 2j
uA j ( xk −i ) = 2b j
, O = ∏ uA j ( xk −i ) O =
2
3
O 4j = O 3j ( c1j x k − 3 + c2j x k − 2 + c3j x k −1 + c4j x k + c5j )
 x − mj 
O
4−i j
4−i j 2
1 +  k −i  i
 a
j
 j  j

Power & Energy Systems Laboratory 30

 RUL Prediction
RUL: Time between now and the moment when the predicted
fault indicator reaches the predefined threshold.
 State of each particle is propagated recursively with a fixed model:

xˆ ki + m = f k ( xˆ ki + m − 4 , xˆ ki + m − 3 , xˆ ki + m − 2 , xˆ ki + m −1 ) + uk , m = 1, 2, , T p

 Update and normalize the weights:

− ( zˆk +m − xˆki +m )2 wki + m
1 2 μ0 w i
= , i = 1, 2,, N
wki + m = wki +m−1
2πμ0
e k +m
 wk +m
i
i

 Thus, the state at k+m can be predicted by:

N
xˆk +m =  wki + m xˆki + m
i =1

 When the predicted state reaches the threshold at a future time instant,
the predicted RUL is determined as the time between now and that
future time instant.
Power & Energy Systems Laboratory 31
RUL Prediction: Laboratory Run-to-Failure Test

Power & Energy Systems Laboratory 32

 RUL Prediction Result: Laboratory Test
NSR
Actual failure time
Predict at 1500 min
RUL PDF
Predict at 1200 min
RUL PDF
Predict at 900 min
RUL PDF

 The gearbox failed at 1710 min

 PDFs of RUL predicted at 900, 1200 and 1500 min
 RUL prediction has much larger errors at beginning
and exceeds the ±20% confidence bounds of the
true RUL frequently from 900-1150 min
 From 1150 min onwards almost every predicted
RUL value is located within the ±20% confidence
bounds of the true RUL
 Proposed method can accurately predict the RUL at 1150 min, which is 560 min
before the actual failure time or 32.7% of the overall lifetime of the gearbox
Power & Energy Systems Laboratory 33
 Fault Prognostic Result: Field Data
 Predict the fault Mature defect
Defect
propagation in a 2.5MW propagation

DFIG wind turbine

 Data: Vibration data from Incipient

defect
sensor mounted on HSS of occurred

drivetrain gearbox. Maintenance

 Data selection: 85%-100%
rated power.
 Defect type: High speed stage
(HSS) bearing inner raceway
ball pass.
 Prognostic Indicator: Peak
value of synchronized waveform
of vibration sensor located on
gearbox HSS.
performed
 Critical timeline
 Record period: 10/1/2014 to 11/1/2015
 Incipient defect occurred: 3/8/2015
 Defect propagation: 3/8/2015 to 4/21/2015
 Mature defect: 4/21/2015 to 9/28/2015
 Component replacement: 9/28/2015

Power & Energy Systems Laboratory 34

 Fault Prognostic Result: Field Data (2)

Defect propagation

Time index Actual time

1600 3/8/2015 10:14:13 PM
1750 3/17/2015 9:49:54 AM
Incipient 2050 3/21/2015 1:09:40 AM
defect
2350 3/24/2015 7:53:44 PM
3700 4/21/2015 10:44:38 PM
5000 5/19/2015 3:15:21 AM

 Black line: Actual prognostic indicator

 Blue line: Prognosis @ 1750 to 5000 time index
 Green line: Prognosis @ 2050 to 5000 time index
 Red line: Prognosis @ 2350 to 5000 time index
 The incipient defect occurred @1600 and propagated from 1600 to 3700 time index
 Prognosis made @2350 can effectively capture the trend of the defect propagation,
which is around 27 days before defect became mature.

Power & Energy Systems Laboratory 35

 RUL Prediction Result: Field Data (3)
 Threshold: An alarm will be given when the indicator increase to 8.
 The actual indicator overpass 8 at time index 2680.

Defect Mature defect

propagation

Incipient
defect
occurred Threshold

Maintenance
performed

 The RUL can be effectively predict after Time index Actual time
time index 2100.
1750 3/17/2015 9:49:54 AM
 The corresponding time is 3/19/2015
2100 3/21/2015 9:29:40 AM
10:41:06 PM, which is around 12 days
before the alarm should be triggered 2300 3/23/2015 9:49:28 AM
(3/31/2015 4:26:45 PM). 2680 3/31/2015 4:26:45 PM

Power & Energy Systems Laboratory 36

 Summary
 Proposed methods are sensorless and nonintrusive: only using
generator current measurements, which are already used in wind
turbine control and protection systems
 Proposed methods can be easily integrated into existing wind
turbine condition monitoring, control, and protection systems
 Fault diagnosis and prognosis can be implemented remotely from
the wind turbines being monitored
 Proposed methods provide an alternative to sensor-based
condition monitoring: reduce cost, size, and hardware complexity
 Proposed methods can be combined with sensor-based methods
to improve the accuracy and reliability of condition monitoring
 Proposed methods offer an effective means to achieve condition-
based predictive maintenance for wind turbines to significantly
reduce O&M cost (e.g., expect 50% reduction) and improve
availability and lifespan of wind turbines
Power & Energy Systems Laboratory 37
 Sponsors/Partners
 This work was supported in part by the U.S.
Department of Energy, U.S. National Science
Foundation, Nebraska Public Power District through
the Nebraska Center for Energy Sciences Research,
and American Public Power Association.
 Part of the work was completed in collaboration with
GE Global Research.

Power & Energy Systems Laboratory 38

Thank you!

Power & Energy Systems Laboratory 39