Smart Wind Turbine: Artificial Intelligence based Condition

Monitoring System

Afshin Tafazzoli

and Alvaro Novoa Mayo

Global Services, Siemens Gamesa Renewable Energy, Calle Ramirez Arellano, 37, Madrid 28043, Spain

Energy Consultant, KPMG, Torre de Cristal, Paseo de la Castellana, 259C, Madrid 28046, Spain

Keywords: Wind Turbine Generator (WTG), Artificial Intelligence (AI), Condition Monitoring System (CMS).

Abstract: This project is motivated by the importance of wind energy and reducing the financial and operational impact

of faults in wind turbine generator using artificial intelligence based condition monitoring system. It is to

classify the fault alarms and diagnose smart solutions at level zero to resolve the faults without service expert’s

intervention. Big data analysis of the large historical data pool results in the intelligent algorithms that can

power the diagnostic models. For maximum efficiency, wind turbines tend to be located in remote locations

such as on offshore platforms. However, this remoteness leads to high maintenance costs and high downtime

when faults occur. These factors highlight the importance of early fault detection and fast resolution in great

extent. The aim of the project has been to have smart wind turbines integrated with artificial intelligence. The

condition monitoring system should have the capability to detect, identify, and locate a fault in a wind turbine

and remotely reset the turbines whenever possible.

1 INTRODUCTION

Wind turbine generator (WTG) condition monitoring

systems are an example of predictive diagnostic tools

using big data and Artificial Intelligence (AI) that

allow automatic fault detection in the rotor and stator

of WTGs with substantial time before a critical fault

occurs.

The installation of such systems is developing at

a fast pace in industry with the current increasing

popularity of offshore wind projects and the problems

that are derived from their O&M such as high

downtime, often a result of complex repair

procedures in remote locations.

The main motivation is to avoid reduced

efficiency and production of WTGs as wells as an

increase in the overall costs as, when a fault occurs

and there is not predictive maintenance in place,

WTG supervisors must first send the maintenance

crew to the turbine to identify the fault’s location.

This would imply repairs which may involve using

specialist equipment (cranes and support vessels)

increasing the risk of potential delays caused by

unfavourable weather or wave conditions (Crabtree,

2010). During these steps it is likely that no energy

will be produced as without knowledge of the fault

type the risk to operate cannot be taken.

Condition Monitoring systems perform an early

fault detection, location and identification which,

under some circumstances, would not be possible

otherwise. Electrical faults are a clear example, as

protection relays cannot be attached to all the parallel

paths of the windings individually and the generator

could keep operating if a small fault happening in a

path of one of the windings goes undetected due to

the small unbalance. The fault will eventually erode

the parallel path winding and cause a catastrophic

fault. Small electrical faults can also create pulsating

torque in the machines and, with time, this can lead to

machine failures. This is also the reason why these

type of systems aim for very fast operation; which

implies strong design requirements and high tech

equipment selection.

These are situations where condition monitoring

(CM) systems come into effect to detect and locate

the electrical fault using big data analysis and AI to

prevent severe damages in wind turbines.

194

Tafazzoli, A. and Mayo, A.

Smart Wind Turbine: Artiﬁcial Intelligence based Condition Monitoring System.

DOI: 10.5220/0007767701940198

In Proceedings of the 8th International Conference on Smart Cities and Green ICT Systems (SMARTGREENS 2019), pages 194-198

ISBN: 978-989-758-373-5

2 METHODS

2.1 Fault Detection

Electrical and mechanical (bearing or gearbox) faults

are the most frequent damaging and expensive type

of faults in wind turbine generators. Vibration

analysis of the machine has been used in industry for

many years as a way of identifying both types of fault;

but lately industrial companies and research

organizations have started to look and analyze the

current output of the generator in the search for fault

indicators. In this way, current is used to detect

electrical and mechanical faults, and vibration used as

a second way to look for mechanical faults.

CM systems use different techniques for fault

detection and analysis, being spectral (frequency)

analysis one of the most popular mainly due to the

accuracy and speed of its predictive technology.

This method for fault detection relies on the

principle that the spectral magnitude of specific fault

frequencies of the electrical or mechanical signal’s

spectrum, increase when the particular fault related to

such frequency occurs.

The working sequence of CM systems using

spectral analysis involves:

1. obtaining information from the sensors, as well as

from the power measurement equipment present

on WTG,

2. processing this data to transform it to its frequency

domain,

3. check for particular fault frequencies present on

them.

Lastly, fault detection algorithms will check the

spectral magnitudes of the fault frequencies, within

an error margin. A threshold magnitude, set on

automatic tests performed on the WTG, will

determine if specific fault frequencies’ magnitudes

are high enough to be taken into consideration and

reported to the control center. The latter will be

performed using big data to maintain active

inspection on particular fault frequencies for long

periods (over a year) and AI to send fault severity

warnings and details to remote control centers.

2.1.1 Electrical Faults

Research at The University of Manchester (Djurovic,

2009) has defined a set of frequency characteristic

equations (Figure 1) for each possible type of

electrical fault in both Doubly-Fed Induction

Generators (DFIGs) and Squirrel Cage Generators.

These equations are the core of the CM fault detection

software.

Figure 1: Fault Frequencies Characteristic Equations.

k = 0  Stator excitation frequency; k > 0  Speed-

dependent high frequency components

p = number of pole pairs; f = supply frequency; s =

generator slip

(Synchronous speed assumed to be 1500rpm for the

generators used in the project)

2.1.2 Mechanical Bearing Faults

Bearing faults are the most common faults in a

generator. They can be categorised in single-point

defects and distributed faults. This report focuses on

single-point defects.

Figure 2: Diagram of bearing (Blodt, 2008).

The following equations have been identified

(Stack, 2003) as the fault frequency characteristic

equations for different bearing defects:

Cage Fault









1



cosβ

(1)

Outer Raceway Fault











1



cosβ

(2)

Inner Raceway Fault











1



cosβ

(3)

Ball Fault















1









cos





(4)

Smart Wind Turbine: Artiﬁcial Intelligence based Condition Monitoring System

195

=speed of rotor, N

=number of balls,

=diameter of ball,P

=ball pitch diameter,

β=ball contact angle

Mechanical faults can also be identified by

analysis of the stator current as vibrations cause air-

gap eccentricities that cause disturbances in the air-

gap flux density of the generator which changes

induction affecting, therefore, stator current.

2.2 Spectral Analysis

To account for the non-periocity of signals such as the

current and vibration signals obtained from a

generator, whose output and condition changes with

the wind and the electrical supply, the Short Term

Fourier Transform (STFT) is one of the methods

identified for signal processing and fault detection

software (Dudek, 2011).

The STFT analyzes real time data in predefined

periods of time, called windows, where the speed of

the rotor and the supply frequencies are assumed

constant and applies Fast Fourier analysis for such

window length. Other methods of fault detection

(Empirical Mode Decomposition or Wavelets) have

been researched to be commercially applied but

Fourier analysis was one of the methods that provided

higher frequency resolution.

The STFT relies on the assumption that during a

single time window, the signals that it processes are

periodic.

STFT







τ,ω



≡X



τ,ω



 xtωtτe







(5)

w(t) = Window function; x(t) = Signal to be

transformed; X(τ,ω) = Fourier Transform;

x(t)w(t-τ) = Function representing the magnitude and

phase of the signal; ω = Frequency; τ = time

The STFT truncates the given window with

specific functions to obtain better resolutions when

processing the signal. These functions are defined

window types mathematical defined and used for

many applications. Some examples are: Hann,

Rectangular, Hamming, Blackman window types

(LDS, 2003).

2.2.1 Frequency and Time Resolutions

Heisenberg’s uncertainty principle applies here: the

more it is known about the frequency resolution of a

signal (width of the window), the less precisely it is

known about the time resolution of the signal (narrow

window). A balance is necessary in the system for

good quality results.

Figure 2: (A) Good frequency resolution but high spectral

leakage; (B) Good time resolution but low frequency

resolution.

2.2.2 Envelope Analysis - Hilbert Transform

In vibration spectral analysis, envelope analysis acts

like a filter (Wavemetrics, 2012). It eliminates the

signal originated from the initial vibration of the

machine, facilitating the fault detection after the

spectral analysis has taken place.

Figure 3: Overall Process Diagram.

2.3 CM System’s Fault Alarms

CM systems monitor each of the fault frequencies’

spectral magnitude and their harmonics in order to

check for an increase through time.

(B) (A)

SMARTGREENS 2019 - 8th International Conference on Smart Cities and Green ICT Systems

196

A fault in the system inevitably increases the

energy of the fault frequencies given by the

characteristic equations by a value between 7 to 10

times the energy of the selected frequencies under

normal, healthy, operation in the case of electrical

faults. This value would depend on the severity and

type of fault. When such a change in energy occurs,

the fault detection algorithm gives notice of the fault

to the operator, including the location and the severity

of the fault.

In this respect, AI algorithms are also able to

identify the severity of each fault according to the

evolution of the spectral energy content of the

monitored fault frequencies in comparison to energy

threshold levels tuned up previously. This allows CM

systems to send different alarms, and required

corrective actions, depending on the severity of the

fault to the remote control centers of the system

operators.

Alarms could range from level zero, indicating

that a particular fault is starting to develop and that it

might be required to reset the WTG or send an O&M

operator in the mid-term (weeks) while in the

meantime the WTG can be kept running, to level one

which requires expert judgment to decide if the WTG

can be kept running while O&M operators would

need to be sent in the short term (days), or to level two

were the WTG is required to be stopped for safety

reasons and O&M operators need to be sent as soon

as possible.

3 RESULTS

The following examples show electrical and

mechanical faults that CM systems are able to detect

thanks to big data processing and energy threshold

algorithms.

3.1 Electrical Fault Detection by

Current Spectral Analysis

The following example shows the current spectrum of

a turn to turn fault in the stator windings of a squirrel

cage generator which happens around the sixth

second and how the spectral energy content of such

frequency band increases with the fault.

Figure 4: (A) MarelliMotori current spectrum; (B) Spectral

energy increase with fault.

3.2 Bearing Fault Detection by

Vibration Spectral Analysis

The following example shows an inner race bearing

fault simulated in a squirrel cage generator where the

energy increase in the whole spectrum is clearly

visible (Figure 5(A) below) while the energy

summation of all the calculated fault frequencies in

the first 8 time harmonics is shown in Figure 5(B).

This case differentiates itself from the electrical faults

because looking at the 8 checked individual

harmonics is no longer necessary. The energy

increase is present in all of them.

4 CONCLUSIONS

The use of big data and data processing algorithms in

systems becomes key for the detection of

Time

(

secs

)

Frequency (Hz)

7 7.5 8 8.5 9 9.5

1000

1010

1020

1030

1040

1050

1060

1070

1080

1090

1100

Smart Wind Turbine: Artiﬁcial Intelligence based Condition Monitoring System

197

Figure 5: (A) Vibration spectrum with inner race bearing

fault; (B) Energy variation in calculated fault frequencies.

electrical and mechanical faults as many faults slowly

develop, increasing their fault’s frequencies spectral

energy, through time.

Big data is required to store real and processed

operation data from electrical and vibration sensors in

the WTG for at least one year. AI would require

algorithms to tune up the fault thresholds for each

machine automatically, as well as to identify which

should be the normal, healthy, energy levels taking

into consideration variables such as rotor speed.

The solutions provided in this paper will provide

a competitive advantage to the organization since

many tasks done previously by the operators can now

be executed by the computers. This will leave more

time for the operators to review complex alarm

diagnostics and have them focus on more delicate

services. Another advantage is the speed and more

diagnostics that were not possible without computer

interventions.

The idea is to implement the collective

intelligence where the CMS system can do the level

zero diagnostics and filter the difficult tasks that is

hard for computers to be done by operator experts.

This collaboration will also lead to more accurate

and precise predictions and save a lot of time and

money in the offshore business.

ACKNOWLEDGEMENTS

We would like to thank Professor A.C. Smith and Dr.

Sinisa Durovic for their invaluable advice, guidance

and support during this project. We are also grateful

for further help from Dr. Damian Vilchis Rodriguez

and Professor Patrick Gaydecki for advising on the

technical content of the project.

REFERENCES

Crabtree, C. J., 2010. Survey of commercially available

condition monitoring systems for wind turbines,

Durham university.

Djurovic, S. S., 2009. Origins of stator current spectra in

DFIGs with winding faults and excitation asymmetries,

IEEE international.

Blodt, M., 2008. Models for bearing damage detection in

induction motors using stator current monitoring, IEEE

transactions on industrial electronics.

Stack, J. R., 2003. Fault classification and fault signal

production for rolling element bearings in electric

machines, IEEE international symposium on

diagnostics for electric machines.

Dudek, P., 2011. Digital signal processing course,

University of Manchester.

LDS, G., 2003. Understanding FFT windows, SPX

company.

Wavemetrics, I. P., 2012. Hilbert transform, http://

www.wavemetrics.com/products/igorpro/dataaalysis/si

gnalprocessing/hilberttransform.htm.

STFT

Frequency (Hz)

100

200

300

400

500

600

700

800

900

1000

0.5

1.5

2.5

x 10

Energy variation with fault

SMARTGREENS 2019 - 8th International Conference on Smart Cities and Green ICT Systems

198