Wind Generator Tower Vibration Fault Diagnosis and Monitoring Based on PCA

Ning Fang, Peng Guo

School Of Control and Computer Engineering, North China Electric Power University, Changping, Beijing 102206
Email: fangningqq@163.com

ABSTRACT: Vibration signal is one kind of important variables for wind turbine in Supervisory Control and Data
Acquisition (SCADA) System. With modeling and analysis of vibration signals, incipient failure of key components
such as tower, drive train and rotor could be detected. In this paper, Principal Component Analysis (PCA) was applied
to model the tower vibration providing a good understanding of the tower vibration dynamic characteristics and the
main factors influencing these. The merits and physical interpretation of PCA are also discussed and analyzed. With
SCADA of one wind turbine during April to June in 2006, PCA model for tower vibration in normal working condition
was established and validated. This method monitoring the operation of the wind generator by calculating the
monitoring statistic Hotelling T2˄T2 for short˅and SPE. If T2 and SPE statistics over confidence limit, the decision
system is abnormal. We can calculate contribution rate of each variable on the main componentˈthen get the cause of
the fault . Wind generator fault data set simulation verify the effectiveness of the proposed method.

KEY WORDS: Condition monitoring; Tower vibration; Principal Component Analysis ; Modeling

1 Introduction
to solve the problem. Second, there is strong
Wind turbine condition monitoring technology has aerodynamic and vibrating coupling between different
become a research focus in the field of wind power at turbine components and thus many interconnected
home and abroad in recent years. In wind turbine factors may influence the vibration signatures [3]. With
condition monitoringˈvibration signal is to reflect the flexible tower as an example, it as a support unit, not
important parameters of the unit rotating parts (such as only need to support weighed dozens of tons of
the transmission chain, impeller) and supporting parts impeller and the cabin, and also need to inherit and
(tower) in the running state [1]. However, there are two resist wind acting on the axial load of the impeller [4]
difficulties in the application of vibration analysis to has simplified tower as cantilever beam with the
wind turbines. First, large-scale wind turbines operate method of the finite element numerical analysis to
in a variable speed mode to optimize performance in study the influence on the impeller and engine weight
relation to time changing wind speed and so the to tower natural frequency. [5] takes the engineering
rotational speed of the rotor, gearbox and generator is measurement method, they has installed eight vibration
changing significantly in time. For example, because measuring point on different height and position in the
the rotating speed of the gearbox is changing, the width tower wall. It gets tower drum modal parameters and
of the vibration sidebands is not fixed, and this creates resonance frequency through the vibration signal and
difficulties in locating the exact locations of gear or the power spectrum curve, and then puts forward the
bearing faults. It is conventional as in [2] to use order measures to avoid tower resonance. But these analysis
analysis to deal with this problem, or equivalently of the tower vibration are obtained by simplifing the
azimuthal data sampling (rather than fixed time actual tower complex state. At the same time large
interval sampling) in which the rotor vibration is amounts of calculation, analysis of complex is suitable
analyzed based on samples recorded at equidistant for wind power unit online condition monitoring or not
rotational angles instead of time equidistant samples, still to be determined. In wind power unit SCADA
and then use the FFT frequency domain analysis method system, there are also records of different component
vibration sensor measurement of vibration data usually.
Project Supported by the Fundamental Research Funds for the Chinese
Taking a wind farm of SCADA operating data as the
Central University(12MS58)

978-1-4673-5534-6/13/$31.00 2013
c IEEE 1924
foundation, on the basis of qualitative analysis, we use and rated ones, the wind turbine runs in a Maximum
principal component analysis (PCA) to compress many Power Point Tracking (MPPT) regime. In this regime,
variables into a small number of independent variables. the blade angle is usually fixed (at around two degrees
The variable retained the original features of the data depending on the blade design) and the rotational speed
information and abandoned the redundant information. for rotor is controlled to be proportional to wind speed
We can calculate the monitoring statistic HotellingT2 in order to maintain operation at C pmax and thus
and SPE calculate to realize production data mining maximize energy capture. Figure 1 shows trends of
and analysis in a large quantity of data, high data tower vibration and related variables from 25/03/2006
dimension, variables with a correlation. Then we can to 30/03/2006. Fig 2 shows trends of tower vibration
calculate contribution rate of each variable on the main and related variables from 30/03/2006 to 03/04/2006.
component to get the causes of the failure of may.
The physical units used for tower vibration and related
Then we can more easily find out the root cause of
variables are shown in Table 1.
the failure from a few independent variables. This kind
of vibration modeling method based on actual data as 1500

the foundation, we can call it operation data driven

Power
modeling method. The vibration model reveals the
0
main influence factors of the vibration mode, at the 15

Wind speed
same time can be used as scale real-time monitoring
the running state of relevant parts, such as impeller,
variable propeller system, etc. In order to lay the 0
100
foundation for the follow-up to be carried out based on
Blade angel

the state of vibration monitoring and early fault

diagnosis.
0
200
2 Wind Turbine SCADA Data Preparations tower drive train
Vibration

and Tower Vibration Analysis

The wind turbine studied in this paper is a GE 0
0 100 200 300 400
1.5MW (model 1.5SLE) variable pitch, variable speed Vector sequence
wind turbine, located in Zhangjiakou, northwest from
Fig. 1 Trends for tower vibration and related variables from
Beijing. The cut-in and rated wind speeds for wind
2006/03/25 to 2006/03/30
turbine are respectively 3m/s and 12m/s, and the cut
Tab.1 physical unit for SCADA variables
out wind speed is 25m/s. The SCADA system records
Variable name Physical unit Notice
all wind turbine parameters every 10 minutes. Each 2
Tower vibration mm/s üü
record includes a time stamp, power, wind speed, blade
Drive train
angle, tower and drive train vibration amongst many mm/s2 üü
vibration
others, about 46 parameters in all. The accelerometer
Power MW Rated: 1.5MW
for measuring tower vibration is mounted at the top of
Wind speed m/s Rated: 12m/s
tower, where it meets the nacelle. The accelerometer
Torque % Rated: 880KNm
for drive train vibration (DTV) is mounted on the
Blade angle degree Below rated˖2
high-speed shaft bearing. 10-minute SCADA data from
the wind turbine from March to April was used, and Below rated, the pitch angle of this GE turbine is
there were about 5500 10-minute records. fixed at 2 degrees. From Fig 1, it can find that the
following variables have a great influence on tower
For a large-scale wind turbine, there are several
vibration magnitude.
different operational regimes reflecting different wind
speed ranges. When the wind speed is between cut-in A. Wind speed. Wind speed is stochastic and produces

2013 25th Chinese Control and Decision Conference (CCDC) 1925

time varying forces and loads on the rotor. [6] Most
relevant to this analysis are the torque and thrust, both
3 PCA tower vibration modeling
approximately proportional to wind speed squared
below rated. Even below rated, the higher the wind 3.1 PCA principle
speed, the larger magnitude tower vibration, as shown The object of the principal component analysis
in Fig.1. This is because the amplitude of thrust (PCA) is process variable sample data matrix.
variation increases with wind speed as would be It produce a compressed primary element model
expected since wind speed standard deviation increases by spectral decomposition process data matrix. The
with wind speed, assuming roughly constant turbulence model gives the main trends of the linear combination
intensity. of variables to describe the data changes.
For a given data matrix X, it has m line sampling
B. Torque and Power. At regime, torque will increase
value and n column variable. X’s covariance is:
approximately with wind speed squared as indicated
above, output power with wind speed cubed. Torque X 'X (1)
COV ( X ) =
m −1
and power reflect how hard the wind turbine works.
If X has been normalized, covariance becomes
C. Drive train vibration. Drive train for wind turbine
correlation. Matrix X will be decomposed into k (k
includes main bearing, gearbox, and generator bearing.
<min {m, n}) pairs of vector product of vectors and a
Because the drive train is located in the nacelle,
residual matrix E:
vibration of the drive train will be transmitted to the
supporting structure, in this case through the yaw X = t1 p1T + t2 p2T +……tk pkT = TPT + E (2)
bearing to the tower and will directly influence tower
vibration. ti is the score vector, and pi is the feature vector of
covariance. In order to determine the number of
2000
principal component. We usually use variance
Power

cumulative percentage contribution principle. We

0
usually choose the number when the percentage points
15
more than 85%. In order to describe the process of the
Wind speed

normal operation we generally choose no mare than 3

0
principal component based on the process of principal
100
component analysis in fault diagnosis. [8]
Blade angel

3.2 PCA fault detection

0

200 tower
When use PCA method for the performance
monitoring, we use Hotelling T2 statistic and residual
Vibration

drive train

space SPE statistics of the monitoring primary space to

0 50 100 150 200 250 300 350 400
vector sequence obtain the whole production process the operation
Fig. 2 Trends for tower vibration and related variables from condition of real-time information. Among them, the
30/03/2006 to 03/04/2006 SPE statistics reflect the measurement value at one
point of model deviation degree; And T2 statistic
During such an event the aerodynamic forces on
reflects the change of variables though the PCA model
the rotor reverse over a very short period of time
internal main vectors die fluctuations. [7] It reflects
(typically less than 10 seconds) as it moves from
distance between the mapping point of sampling data
turbine mode to propeller mode. This results in a large
on the main element planar and multivariate data mean.
impulse force on the tower.
The T2 statistics are defined as follow:

T 2 = [t1, t2 ,…䯸 t p ]Λ-1[t1䯸 t 2䯸 …䯸 tp ]T (3)

1926 2013 25th Chinese Control and Decision Conference (CCDC)

 Score vector tk can be obtained by formula (2). ȁ-1 Figure 3 depicts a model based on the realization
is the inverse matrix of the score vector corresponding of the process monitoring and fault diagnosis
eigenvalue matrix. T2 statistic be obtained as follows˖ algorithm.

P( N − 1) 4 Tower Vibration Modelling using

T 2 P , N ,α = FP , N − P ,α (4)
N −P Reasonably determine the tower vibration related
Where N is the number of samples in the model variables set and calculate the PCA model is the key to
and p is the number of pivot element referring to the using PCA to tower vibration modeling.
level of detection. F pˈN-pˈ¢is F distribution threshold As the first quarter mentioned, 10-minute SCADA
corresponding to the level of detection is Į, the degree data from the wind turbine from March to April 2006
of freedom is p and under the conditions of Np. The was used, and there were 5500 10-minute records.
statistics of SPE in the model of PCA reflect the degree According to the alarms and stop records of the units,
of deviation of measured value relative to the model at the SCADA record when fault outage or cut-in speed
a certain moment, which is stated as the distance from happens will be centralized and removed, then the
the sample point to the model space. During the running trends of variables under normal conditions
performance monitor of PCA, the statistic of SPE is can be obtained and the principle components can be
defined as following: calculated. So we can detect and diagnose the data in
Figure 2. [8]
SPE = ei eiT = X i ( I − Pk Pk T ) (5)

Where ei is in the i-th row of matrix E, I is the

unit of matrix. Pk=[p1,p2,Ă,pk], and the statistic of SPE
represents the data changes, which is not explained by
the model of pivot element. There has been abnormal
in the process when the SPE is too large, it is no longer
applicable of the established model .The calculation of
the control limits are established on the basis of a
certain of assumptions. When the level of detection is Į,
control limits of SPE statistic can be calculated as
follows:
1
ª C 2θ h 2 θ h䯴 h -1䯵 º o
h

Q a =θ i « a 2 o
+1+ 2 o 2 o » (6)
«¬ θ1 θ1 »¼ Fig.3 Process monitoring and fault diagnosis algorithm
realization model
Simulate the dataˈthe SPE statistic and Hotelling
n
θi = ¦ λ j ( i = 1, 2, 3)
i
(7) T2 are displayed as figure 4 and figure 5˖
j = P +1
4
SPE st at ist ics
SPE cont rol limit s

2 θ 1θ 3
h o = 1- (8)
3θ 2 2
SPEstatistics

Ȝi is the characteristic value of the covariance

matrix, Ca is the critical value for the next test level of
the normal distribution, p is to retain the number of
0
principal of the model, n is the main number of the 0 100 200
Vector sequence
300 400

principal. Fig .4 PCA statistics monitoring diagram of SPE with Untreated

2013 25th Chinese Control and Decision Conference (CCDC) 1927

 80
eliminated. There are a few points, we consider is the
T square statist ics
T square cont rol limit s
result of process data correlation caused by.
As shown in fig 8 and 9.they are the contribution
of the first and the second principal component among
squarestatistics

40
every variable at point 316. We can see that the fourth
T

variables which named tower vibration make the main

influence (1. power; 2. wind speed; 3. torque; 4 .tower
0
0 100 200 300 400
vibration; 5.drive chain vibration).
vect or sequence

Fig .5 PCA statistics monitoring diagram of T2 with Untreated 4

SPE Stat istics
SPE cont rol limits
2
At point 316ˈthe SPE,T statistic are more greater 3

than the control limit value and the value of the other

SPEStatistics
sampling points. From the above we can judge at this 2

point appeared fault. Diagnostic result matches the

experimental conditions, so we get the correct 1

verification. For 270 points after the large fluctuations,

0
as well as 280 points or even over the control limit of 0 100 200
vec tor sequenc e
300 400

the condition. We can understand that the appropriate

alarm thresholds or use a sliding window before the Fig .6 PCA statistics monitoring diagram of SPE with no data
failure. processing

T2 statistic describes the degree of fluctuation in 80

T square st at ist ics

the data explained by the first K principal component; T square cont rol limit s

The SPE statistics describe the data change which is

squarestatistics

not explained by the main element model. In the case

40

of T2 statistic diagnostic accuracy, Figure 4 shows the

T

SPE statistics which still have a number of points

outside the control limits before 270 points. SPE
statistics for small perturbations will act more sensitive, 0
0 100 200 300 400
Vect or sequence

it requires us to do some processing of data.
2
Fig .7 PCA statistics monitoring diagram of T with no data
By equation (9) and (10), the wind speed is
processing
proportional relationship to the cubic of the power and
the square of the torque respectively [9]. So before the
6
simulation is performed, we need for data processing.
Through arithmetic operation we can get the cubic root
of the power value and the square root of the torque 4

value. In this way we can eliminate the errors which

are not in the same comparison relations.
2
1
P= ρ SC p v 3 (9)
2
1 1 (10)
dT = ρ rlω 2Cl ⋅ (1 − ⋅ cot I )dr 0

2 ξ
-1
After the simulation of the processed data, we can 1 2 3 4 5


get SPE statistics and T2 statisticsˈas shown in figure5
Fig .8 The contribution of the first principal component among
and figure6. In Figure 5 and Figure 7 compares we can
every variables
clearly see that most of the interference has been

1928 2013 25th Chinese Control and Decision Conference (CCDC)

 follow-up condition monitoring.

6
Reference
[1] Hameed Z, Hong Y.S, Cho Y.M, et al. Condition
4 monitoring and fault detection of wind turbines and related
algorithms: a review [J]. Renewable and Sustainable
Energy Reviews, 2009, 13(1): 1-39.
2
[2] Peter Caselitz, Jochen Giebhardt. Rotor condition
monitoring for improved operational safety of offshore
0
wind energy converters [J]. Journal of Solar Energy
Engineering, 2005, 127: 253-261.
-1
1 2 3 4 5 [3] Yao Xinjia, Liu Yingming, Liu Guangde, et al. Vibration
analysis and online condition monitoirng technolog for
Fig .9 The contribution of the second principal component large wind turbine [J]. Journal of Shenyang University of
among every variables
Technology, 2007, 29(6) :627-632. (in Chinese)
5 Conclusion [4] Li Deyuan, Liu Shengxiang, Huang Xiaohua. Numerical
analysis of vortex-induced vibration of the large scale
Due to the variable-speed operation of the wind wind turbine cylindrical tower [J]. ACTA ENERGIAE
turbine and random variation of the operating condition, SOLARIS SINICA, 2008, 29(11) : 1432-1437.
there is great difficulty and challenge in the component [5] Ma Renle, Ma Yueqiang, Liu Huiqun, et al. Ambient
vibration test and numerical simulation for modes of wind
condition monitoring to traditional time domain and
turbine towers [J]. Journal of Vibration and Shcok, 2011,
frequency domain vibration analysis. With actual
30(5) : 152-155.
operation data, this dissertation analyses the important
[6] He Dexin. Wind engineering and industrial aerodynamics
influencing factors of tower vibration. The author [M]. Beijing : Defense Press, 2006.
constitutes a set of related variables with the factors Gross K CˈSinger R MˈWegerich S Wˈet alˊApplication
found from the SCADA run data, also establishes a [7] LiQiang,WangXingqun,ZhangXi. Process Monitoring and
tower vibration model under below-rated wind speed in Fault Identification of Turbine Generator Based on
principal component analysis method, whose algorithm Principal Com-ponent Analysis and Similar
is simple and physical meaning is clear. The accuracy DegreeAnalysis[J].microcomputer imformation
of the tower vibration model is verified through a 2010,26(3-1) :138-140
validation example. Tower vibration model built based [8] LiXiang.Industrial Process Fault Diagnosis Algorithm
on operation data is basis of condition monitoring for Research Based on PCA [D].Shenyang University.
key components in unit. Data is analyzed before 2010,03,03:1-94
modeling to reduce errors brought by different [9] XuDaping,LiuYibing,LvYuegang. The principle of wind
comparative. However, there are still several problems energy generation.[M]. China Machine PRESS.
in overcoming the auto-correlation of the process data. 2011:44-47
This is an important research direction for the

2013 25th Chinese Control and Decision Conference (CCDC) 1929