Академический Документы
Профессиональный Документы
Культура Документы
Research
Article
Key words:
wind turbine;
Online Fault
Detection;
SCADA Analysis;
Multi-Agent System;
Anomaly Detection;
Normal Behaviour
Modelling
This paper describes a set of anomaly-detection techniques and their applicability to wind
turbine fault identification. It explains how the anomaly-detection techniques have been
adapted to analyse supervisory control and data acquisition data acquired from a wind
farm, automating and simplifying the operators analysis task by interpreting the volume
of data available. The techniques are brought together into one system to collate their
output and provide a single decision support environment for an operator. The framework
used is a novel multi-agent system architecture that offers the opportunity to corroborate
the output of the various interpretation techniques in order to improve the accuracy of fault
detection. The results presented demonstrate that the interpretation techniques can provide
performance assessment and early fault identification, thereby giving the operators sufficient time to make more informed decisions regarding the maintenance of their machines.
Copyright 2009 John Wiley & Sons, Ltd.
Received 7 October 2008; Revised 12 December 2008; Accepted 12 December 2008
Introduction
Wind energy is currently the fastest growing source of renewable energy used for electrical generation around
the world. Wind farms are adding a significant amount of electrical generating capacity, with farms consisting
of hundreds of units being built around the country. Over 10 GW of wind capacity is currently in planning in
the UK alone. Once these projects are completed, the overall installed UK capacity will comprise of a total of
24 GW.1 This increase in the number of wind farms has led to the need for more effective operations and
maintenance procedures. Improving the maintenance policies used helps to ensure that the operation of the
wind farms is optimized by maximizing the availability of the individual turbines.
Condition monitoring (CM) systems can be used to aid plant owners in achieving these goals. They aim to
provide operators with information regarding the health of their machines, which, in turn, can help them
improve operational efficiency by allowing more informed decisions regarding maintenance. This is a step
towards condition-based maintenance, and a move away from typical scheduled maintenance policies that are
in place in most industries today.2 In this way, any degradation or incipient faults can be detected and dealt
with before they result in more serious and costly failures. In addition, healthy turbines can be left to operate,
thereby reducing outages that result because of unnecessary scheduled maintenance operations.
While the benefits of CM provide a strong case for their use in power plants, operators are weary of blindly
adopting the technology without a reasonable economic justification. Most modern turbines are manufactured
* Correspondence to: A. Zaher, Electronic & Electrical Engineering Department, University of Strathclyde, 204 George St, Royal
College Building, Glasgow G1 1XW, UK.
E-mail: azaher@eee.strath.ac.uk
Online Wind Turbine Fault Detection through Automated SCADA Data Analysis
575
with some form of integrated system that can monitor the main components. These systems typically monitor
parameters such as temperatures of bearings, lubricating oil and windings. Some of the more recent systems
monitor vibration levels of the drive-train.3 This monitored data is collated and stored via a supervisory control
and data acquisition (SCADA) system that archives the information in a convenient manner, usually for all of
the turbines in the wind farm. This data quickly accumulates to create large and unmanageable volumes that
can hinder attempts to deduce the health of a turbines components. It would prove beneficial, from the perspective of utility companies, if the data could be analysed and interpreted automatically to support the
operators in identifying defects.
This paper describes a set of techniques that can be used to identify incipient faults in the main components
of a turbine through the analysis of this SCADA data. The SCADA data sets are already generated by the
integrated monitoring system, and therefore, no new installation of specific sensors or diagnostic equipment
is required. The techniques developed revolve around the principle of anomaly detection in order to detect
instances of abnormal behaviour indicative of incipient component failure. In this way, only interesting information is highlighted to the operator, therefore significantly reducing the volume of data they are faced with.
It is believed that anomaly detection can provide early fault detection while minimizing the dependence on
knowledge of the exact correlation between fault modes and the SCADA data generated. The final section of
the paper describes a framework that can be used to gather and present the output of these detection modules
in a comprehensible manner to wind farm operators.
576
A. Zaher et al.
Fault record information (class 1 above) is viewed by wind farm operators as commercially sensitive, and
gaining access to this kind of data was not possible. Access to historical data, however, is not so problematic,
and permission was obtained to access almost 2 years worth of SCADA from a wind farm sited in Scotland
comprising 26 Bonus 600 kW stall-regulated turbines commissioned in November 1995.
The lack of access to defect records and expertise in how faults manifest themselves in the data is essentially
the main challenge that has to be overcome. This highlights the main reason for the use of anomaly detection,
since it places minimal dependence on the requirement of having access to this experience and knowledge
whilst still achieving the objective of identifying faults as early as possible. Anomaly detection uses a model
of the normal behaviour of sensor data recorded from an item of plant or machinery. This model captures
how the data evolves and changes with respect to factors that may influence it under normal circumstances.
This model of normality allows for the detection of anomalous behaviour, even when this type of behaviour
has not been seen previously. The model is used to provide an estimate of the sensor output based on the inputs
that can affect the values recorded by the sensor. This estimation can then be compared to the real value
recorded by the sensor, where a significant deviation from the estimated value would be viewed as an abnormality. In this way, incipient faults can be highlighted and presented to the operator, dramatically reducing
the complexity of their task, since only significant information of relevance to the health of the turbine is
presented to them.
It can be argued that anomaly detection cannot be used to classify faults since no knowledge of the different types of faults is included. Its use here, however, is to provide the initial stage of the fault identification
process. Once the anomalies are detected, the opportunity of labelling them according to specific fault classes
can be achieved through the help of an experienced operator. The operators knowledge could be captured as
rules and then used to classify various instances of failure. This, however, is out with the scope of this paper,
and left for future research.
Online Wind Turbine Fault Detection through Automated SCADA Data Analysis
577
check. Each of the inputs X1 to Xn is fed to each neuron and multiplied by the associated weight (W1 to Wn).
This shows the third characteristic, which is the degree of connectivity the network exhibits, determined by
the value of the weights. Experiential knowledge is gained through a supervised training process and is stored
in these weights, determining the connectivity of the networks neurons.
The supervised training process is achieved through splitting the data into training and testing data sets.
The term supervised refers to the process of the training data consisting of inputs and known desired outputs,
which the network attempts to relate through a nonlinear function by exposure to a sufficient number
of examples. The weights in each neuron are initially set at random values. The input values are then propagated through the network, yielding an output with an error between the initial value and the desired
output. This error is then fed back through the network and used to adjust the weights accordingly with
the aim of minimizing the magnitude of the error term. This process of adjusting the weights is determined
by the algorithm used to train the network. The back propagation algorithm was used for the networks
trained in this research because of its suitability for pattern matching and prediction purposes.20 This process
is then iterated until the weights no longer change or the set number of training cycles is complete. The network
can then be tested on the second data set, which deliberately comprises data that the network has not been
trained on or previously exposed to. Its success at estimating values that are as close as possible to the desired
output values determines how well the network has learned or captured the relation between the inputs
and outputs.
NNs have been used in this research in order to develop normal behaviour models of key SCADA parameters that indicate the condition of the gearbox and the generator, and can estimate the power output of the
turbine according to the specific wind profile it experiences. The power estimation model, however, is not
detailed in this paper because of a lack of sufficient data to develop an accurate model, and as yet still requires
further work.
578
A. Zaher et al.
Active power output (10 min average and standard deviation (SD) over 10 min interval)
Anemometer-measured wind speed (10 min average and SD over 10 min interval)
Nacelle temperature (1 h average)
Gearbox bearing temperature (10 min average)
Gearbox lubricant oil temperature (10 min average)
Generator winding (10 min average)
Power factor (10 min average)
Reactive power (10 min average)
Phase currents (10 min averages)
The techniques presented utilize only the data mentioned above. The parameters listed above are typical of
data collected by commercial wind turbine SCADA systems. This means the approach developed here can be
widely applied by wind farm operators.
Online Wind Turbine Fault Detection through Automated SCADA Data Analysis
579
from experimentation and testing have proved that an accurate estimation of the gearbox and bearing temperature can be achieved without this additional information.
The other distinction between the models is how their output is used. The model built by Sanz-Bobi et al.
uses the output of the cooling oil model as a means to recognize faults with the cooling mechanism and not
the gearbox itself or its bearings. The results from the models built here show that the output of this model
could be used to identify overheating in the gearbox at a very early stage, and so provide insight into the
current status of the mechanical health of the gearbox itself.
Training the Model
The models are trained using the parameters discussed in the previous section. When training the models, in
order to yield an accurate representation for the parameter under study, the range of each of the inputs fed
through for the training process should be as varied as possible while still ensuring they are within the bounds
of normal operation. This was achieved through trending the complete data set and looking for patterns in the
data. It was found that the majority of the turbines exhibit similar trends in their evolution of gearbox temperature. Data showing these similar trends were selected and their values were averaged to provide 3 months
worth of training data for each of the inputs. This amounts to roughly 13,000 data points for each input. The
range of values for each input parameter used to train the cooling oil model is depicted in Figure 3(a). What
this shows is the typical spread of values under normal healthy turbine operation sorted in ascending order.
The training process then attempts to capture the nonlinear relationship between these parameters, i.e. the
associated gearbox oil and nacelle temperatures for the corresponding power output. The number of training
cycles used, also known as epochs, was 1000. Determining the architecture for the network is an iterative
process and depends solely on the structure that yields the best accuracy when tested. The final architecture
used for both the gearbox models was 5-3-1, as shown in Figure 1.
Copyright 2009 John Wiley & Sons, Ltd.
580
A. Zaher et al.
The trained model was tested on new data from a healthy turbine that had not been used in training.
Figure 3(b) shows the input signals used from turbine 1 during the month of February 2006. Figure 3(c)
shows that the model provides an accurate estimate of the expected temperature of the gearbox cooling
oil. The difference signal shows the error between the estimated and actual values. The averaged error
between the two trends is 0.049C, and the Root Mean Square (RMS) error was 1.18, which was deemed an
Online Wind Turbine Fault Detection through Automated SCADA Data Analysis
581
Figure 3. (b) Cooling oil model testing input data (turbine 1, February 2006);
acceptable level of error for the diagnostics required by the industrial partner for successful fault detection.
This means that the output of the model can be used directly as a comparison with the actual temperature trend
to assess whether an anomaly is present. A straightforward difference of the two signals can be used. If
the difference between the estimated value produced by the model and the actual value increases for a
continuous number of instances, i.e. a prolonged period of time and not a minor fluctuation, then this
would flag as an anomaly. This is to exclude false identification of anomalies as a result of erroneous
but transient data.
Copyright 2009 John Wiley & Sons, Ltd.
582
A. Zaher et al.
Figure 3. (c) Cooling oil model output (estimating turbine 1, February 2006 gearbox oil (t) parameter)
Online Wind Turbine Fault Detection through Automated SCADA Data Analysis
583
584
A. Zaher et al.
a
Figure 5. (a) Gearbox failure: gearbox cooling oil model output, anomalies detected from August 2005January 2006
and (b) Gearbox failure: gearbox cooking oil anomalies detected from August 2005January 2006
Online Wind Turbine Fault Detection through Automated SCADA Data Analysis
585
Figure 6. (a) Gearbox failure: corresponding gearbox bearing model output detected during August 2005January
2006 and (b) Gearbox failure: corresponding gearbox bearing model error signal detected during August 2005January
2006 showing no significant or prolonged periods of deviation
Copyright 2009 John Wiley & Sons, Ltd.
586
A. Zaher et al.
deviations occurred at the same time the cooling oil deviations were spotted, which depicts the expected heat
transfer between the gearbox components. According to these results, overheating was detected in the gearbox
cooling oil while the gearbox bearing temperature conformed to the normal behaviour model. Corroborating
the output from both models suggests that the failure was in fact a problem internal to the gearbox and not the
gearbox bearing or cooling mechanism. This failure has been confirmed with the industrial contacts who
supplied the data.
Figure 7(a),(b) shows another example of gearbox problems detected by abnormally high temperatures
in turbine 17s gearbox oil parameter. In this example, the turbine experienced problems towards the end
of April 2005, where it was taken offline as depicted by the trends. Data from before April 2005 was not
available, however, to identify when the first model deviations occurred. Nevertheless, deviations were
detected at the start of the month. These results highlight the importance of model accuracy, which, in turn,
is achieved through good model selection and the effective training used to capture the normal behaviour
of each parameter.
Figure 7. (a) Gearbox problems detected in turbine 17, April 2005 and (b) Gearbox problems detected in turbine
17, April 2005
Online Wind Turbine Fault Detection through Automated SCADA Data Analysis
587
Figure 8. (a) Turbine 8 generator failure or brought offline (November 2006) detected in generator winding
temperature parameter and (b) Turbine 8 generator failure: generator winding model error signal
588
A. Zaher et al.
temperature. The repetitive spikes seen in Figure 8(a) are the effects on the temperature seen when the safety
mechanism activates, interrupting normal power generation.
Results Conclusion
The results produced by the models developed for this research are very positive. They provide an early warning
of problems developing in the gearbox and generator that become apparent through abnormal temperatures.
The anomalies detected by the models reduce the volume of data that must be analysed considerably, making
the task of the operator much more practical. The models also supply this information in good time, allowing
for a more informed decision regarding the appropriate action to be made.
It can be seen that all of the factors involved in developing and training the models can have a significant
impact on the accuracy of the relationship captured. Factors such as the training algorithm, the architecture of
the network, as well as the data selected for training the model, all affect the output. The research carried out
in Sanz-Bobi et al.6 was the most advanced research to date found in the literature in this area of study.
However, the results of the models developed and described above provide earlier fault identification on the
data set they were applied to, while the framework detailed in the next section also takes the research a step
further. It allows techniques to be combined so that the analysis of many parameters and components of the
turbine can be assessed simultaneously for corroboration purposes.
Online Wind Turbine Fault Detection through Automated SCADA Data Analysis
589
interested in. This will give the user the control of further reducing the volume of data presented to them by
selecting only what is important to them.
590
A. Zaher et al.
parser has the responsibility of reading the data files, parsing them for the appropriate data and then sending
this data in an understandable format to the processing agents. This functionality is split into the following
three behaviours:
Parse data behaviour;
Parse waker behaviour; and
Subscription responder behaviour
The parse data behaviour is responsible for accessing the data file, finding the required data and storing
it temporarily in a format that can be sent to the appropriate processing agent. The parse waker behaviour
is a timed behaviour that controls the time interval between readings of the data file to parse the next
available measured value. The data parser agents provide data to the processing agents and so they
register with the DF that they can provide a specific type of data. In order to do so, they must exhibit
a subscription manager interface that allows them to handle any number of agents that are interested in
receiving this data that they can provide. The subscription responder behaviour allows it to do this by
handling any subscriptions sent to it. It then takes care of all the subscribed agents (the processing agents
that require this data to carry out their processing) by sending out the data to them as soon as it becomes
available.
Agent Communication
Agent Communication is handled by the Foundation for Intelligent Physical Agents (FIPA) Agent Communication language (ACL). This language underpins the communication between agents, allowing them to coordinate their behaviour and achieve the subscription functionality described in the previous section. There are
two main levels of protocol that an agent designer has to take into consideration when developing communication between agents. The first is the ACL message itself, and the second is the content language used for
the syntax of the message content. At the ACL message protocol level, the only parameter actually required
is the performative. This parameter indicates the type of communication such as inform, query ref, request and
subscribe to name a few examples.25 Some of the other parameters of an ACL message include participant
information (sending and receiving agents), conversation control headers (conversation identifier), content
parameter (where the real meaning of the message is stored) and content descriptors, which can all be set to
make the message more meaningful.
Figure 10 shows an example of the communication process taking place between a data parser agent, one
processing agent, the interface and the DF from system start up to an anomaly being identified and passed on
to the user. The first three requestinform dialogues taking place between each of the agents and the DF correspond to the process of each agent registering itself with the DF to provide information of the services it can
provide. The data parser registers that it can provide a specific type of data, the gearbox oil agent registers that
it can provide alarms for excessive oil temperatures and the interface agent registers that it provides information to the user. The subscribeinform dialogues in 4 and 5 correspond to the associated agents subscribing to
the services that they require, namely the gearbox agent-requiring data and the interface agent-requiring alarms.
Messages 6n represent the data that is being sent to the gearbox agent. This data is processed by the normal
behaviour model to see if it corresponds to the model output. Finally message n + 1 signifies information being
sent by the gearbox agent to the interface in relation to the data that is anomalous and doesnt correspond to
the model.
The second level of protocol, the content language, defines the format of a messages content. In order for
agents to understand what is being sent to them, it is necessary that they use a common vocabulary that both
the sender and receiver can understand. This language is known as an ontology. An ontology defines the concepts and ideas that the agents communicate between them. To design an ontology suited to any particular
domain, thorough knowledge of the sort of messages and information that will be exchanged between the
agents is necessary. Good design of this language ensures that the agents within a system have a common
ground regarding the information they can send request or query.
Copyright 2009 John Wiley & Sons, Ltd.
Online Wind Turbine Fault Detection through Automated SCADA Data Analysis
591
592
A. Zaher et al.
generator overheats, it cuts out resulting in zero power output for a short period of time. When this happens
repeatedly, as in the example shown earlier, this has a bearing on the efficiency of the turbine for that month
since it will not have performed as well as expected. When this happens, we know that the generators condition is deteriorating to the point of affecting turbine output. This leads to greater downtime and, therefore, will
have an effect on revenues. Initial overheating was detected by the generator winding model 8 months before
the safety mechanism began to cut in and affect power performance. The anomalies then continued in tandem
with an increase in the frequency of generator cut offs, until the generator eventually failed in November 2006.
These 8 months give the operator time to decide when to repair the generator rather than having to replace it
altogether, which may be more costly and lead to a greater period of downtime.
Techniques introduced in the future can also be easily incorporated into the MAS framework, which will
also allow more opportunities for corroboration, and hence a more robust fault detection system. This corroboration can be built in automatically and carried out by the interface agent or another agent dedicated to
correlating the various sources of information.
Conclusions
This paper has described a set of techniques that can be used for early fault identification for the main components of a wind turbine. The results have shown that they can automatically interpret the large volumes of
SCADA data presented to an operator and highlight only the important aspects that would be of interest to
them. In this way, they dramatically reduce the information that is presented to the operator, therefore allowing them to make more informed decisions regarding the maintenance of the turbine. The proposed multi-agent
platform allows the techniques to be brought together to corroborate their output for more robust fault detection. It also allows the development of a system that can be used to apply the techniques across a complete
wind farm, therefore offering only one point of contact for an operator that provides all of this information in
a clear and concise manner.
The next steps of the research will focus on developing this agent architecture into a fully functioning automated fault detection system.
References
1. BWEA. UK Wind farm statistics. [Online]. Available at: http://www.bwea.com/statistics/, (Accessed 25 June
2007)
2. Schneider J, Gaul AJ, Neumann C, Hografer J, Wellssow W, Schwan M, Schnettler A. Asset management techniques.
Electrical Power Energy Systems 2006; 28(9): 643654.
3. Becker E, Poste P. Keeping the blades turning: condition monitoring of wind turbine gears, Refocus 2006; 7: 2632.
4. Hameed Z, Hong YS, Cho YM, Ahn SH, Song CK. Condition monitoring and fault detection of wind turbines
and related algorithms: a review. Renewable Sustainable Energy Reviews, 2009; 13(1): 139. DOI: 10.1016/
j.rser.2007.05.008.
5. Caselitz P, Giebhardt J. Rotor condition monitoring for improved operational safety of offshore wind energy converters.
Journal of Solar Energy EngineeringTransactions of the ASME 2005; 127: 253261.
6. Sanz-Bobi MA, Garcia MC, Del Pico J. SIMAP: intelligent system for predictive maintenance application to the health
condition monitoring of a wind turbine gearbox. Computers in Industry 2006; 57: 552568.
7. Leany V, Sharpe DJ, Infield DG. Condition monitoring techniques for optimisation of wind farm performance.
International Journal of COMADEM 1999; 2: 513.
8. Leany V, Sharpe DJ, Infield DG. The applicability of spatial analysis techniques to monitoring wind farms. In
Proceedings of the European Wind Energy Conference, 15 March 1999. 216219.
9. Singh S, Bhatti TS, Kothari DP. Wind power estimation using artificial neural network. Journal of Energy Engineering
2007; 133(1): 4652.
10. Baines N, Riddle AF, Hart DJ. Intelligent Data Management System for Wind Farm Monitoring and Operation
Optimisation. Harwell Laboratory, Energy Technology Support Unit: Harwell. 1998.
11. Fayyad UM. Advances in Data Mining and Knowledge Discovery. MIT Press: Cambridge, Massachusetts, 1996.
Copyright 2009 John Wiley & Sons, Ltd.
Online Wind Turbine Fault Detection through Automated SCADA Data Analysis
593
12. McArthur SDJ, Booth CD, McDonald JR, McFadyen IT. An agent-based anomaly detection architecture for condition
monitoring. IEEE Transactions on Power Systems 2005; 20; 16751682.
13. Stephen B, Strachan SM, McArthur SDJ, McDonald JR, Hamilton K. Design of trip current monitoring system for
circuit breaker condition assessment. IET Generation, Transmission & Distribution 2007; 1(1): 8995.
14. Tarassenko L, Nairac A, Townsend N, Buxton I, Cowley P. Novelty detection for the identification of abnormalities.
International Journal of Systems Science 2000; 31: 14271439.
15. Devendra T, Murat Y, Asok R, Vigor Y. Anomaly detection in aircraft gas turbine engines. Journal of Aerospace
Computing, Information and Communication 2006; 3: 4451.
16. McArthur SDJ, Strachan SM, Jahn G. The design of a multi-agent transformer condition monitoring system. IEEE
Transactions on Power Systems 2004; 19: 18451852.
17. McArthur SDJ, Catterson VM, McDonald JR. A multi-agent condition monitoring architecture to support transmission
and distribution asset management. Proceedings of the 3rd IEE International Conference on Reliability of Transmission
and Distribution Networks, London, 2005.
18. Stack JR, Habetler TG, Harley RG. Fault classification and fault signature production for rolling element bearings in
electric machines. IEEE Transactions on Industry Applications 2004; 40: 735739.
19. Yang J, Montakhab M, Pipe AG, Davies TS. Application of multi-agent technology to fault diagnosis of power
distribution systems. Proceedings of Engineering of Intelligent Systems, 2004.
20. Haykin S. Neural Networks: a Comprehensive Foundation. Prentice-Hall International Editions: Englewood Cliffs,
NJ.
21. Hansen MOL. Aerodynamics of Wind Turbines (2nd edn). Earthscan: London, UK.
22. McArthur SDJ, Davidson EM, Catterson VM, Dimeas AL, Hatziargyriou ND, Pomci F, Funabashi T. Multi-agent
systems for power engineering applicationsPart I: concepts, approaches, and technical challenges. IEEE Transactions
on Power Systems, 2007; 22: 17431752.
23. Wooldridge MJ, Jennings NR. Intelligent agents: theory and practice. Knowledge Engineering Revision; 1995; 10:
115152.
24. Java Agent Development Framework (JADE). [Online]. Available: http://jade.cselt.it/. (Accessed 15 July 2008)
25. Foundation for Intelligent Physical Agents. FIPA Communication Act library specification, [Online]. Available: http://
fipa.org/specs/fipa00037. (Accessed 22 July 2008).