Professional Documents
Culture Documents
discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/283989773
READS
3 AUTHORS:
Paolo Castaldi
University of Bologna
79 PUBLICATIONS 327 CITATIONS
SEE PROFILE
All in-text references underlined in blue are linked to publications on ResearchGate, Available from: Silvio Simani
letting you access and read them immediately. Retrieved on: 02 February 2016
9th IFAC Symposium on Fault Detection, Supervision and
9th IFAC Symposium on Fault Detection, Supervision and
Safety
9th
9th of Symposium
IFAC
IFAC Technical Processes
on
on Fault
Fault Detection,
Detection, Supervision
Supervision and
Safety of Symposium
Technical Processes and
September
Safety
Safety of 2-4, 2015.
of Technical
Technical Processes
Processes
Available online
Arts et Mtiers ParisTech, at www.sciencedirect.com
Paris, France
September 2-4, 2015. Arts et Mtiers ParisTech, Paris, France
September
September 2-4,
2-4, 2015.
2015. Arts
Arts et
et Mtiers
Mtiers ParisTech,
ParisTech, Paris,
Paris, France
France
ScienceDirect
IFAC-PapersOnLine 48-21 (2015) 13451350
FaultTolerant
FaultTolerant Control of an Offshore
FaultTolerant Control
Control of
of an
an Offshore
Offshore
Wind
Wind Farm via Fuzzy Modelling and
Wind Farm via Fuzzy Modelling and
Farm via Fuzzy Modelling and
Identification
Identification
Identification
S.
S. Simani
,1
Simani ,1 S.
S. Farsoni
Farsoni P.
P. Castaldi
Castaldi
S. Simani
S. Simani ,1 S. Farsoni
,1
P. Castaldi
S. Farsoni P. Castaldi
fuzzy models estimated via the data acquired from the xc (t) = fc (xc (t), u(t))
process under diagnosis. The controller accommodation (1)
exploiting further control loops depends on the online y(t) = xc (t)
estimate of the fault signals themselves. The suggested
nonlinear fault estimation procedure resembles the ap- T
where u(t) = [vw (t), i (t)] and y(t) = xc (t) =
proach described in (Simani and Castaldi (2014)), even T
if in this paper the nonlinear filters are described as fuzzy [i g (t), Pi g (t)] are the input and the monitored output
prototypes and identified directly from the wind farm data. measurements, respectively. The subscript i indicates the
Both the adaptive filters and the FTC strategy are anal- measurement from the ith wind turbine of the wind
ysed with respect to the wind park simulator described in farm (i = 1, . . . , 9). fc () represents the continuoustime
(Odgaard and Stoustrup (2013)), in the presence of faults, nonlinear function that will be approximated with the
disturbances, measurement noise, and modelling errors. discretetime fuzzy prototype from N sampled data of u(t)
The design of the overall FTC for the wind park simulator and y(t), and using the procedure presented in Section 3.
and based on fuzzy FDD modules represents the novel
contribution of this paper. The proposed solution will 2.1 Simulated Fault Conditions
be also compared with respect to two different strategies
presented in (Simani and Castaldi (2012); Simani et al. In this benchmark three faults are considered that in-
(2014)). fluence the measured variables from the wind turbine,
i.e. i (t), i g (t), and Pi g (t). It is also assumed that the
considered faults can be detected at a wind farm level by
comparing the performance from other wind turbines in
2. OFFSHORE WIND PARK SIMULATOR the wind farm, but they are difficult to detect at a wind
turbine level. Moreover, these three faults affect different
In this benchmark model a simple wind farm with 9 wind wind turbines at different times, as described in more
turbines is considered, arranged in a square grid layout detail in (Odgaard and Stoustrup (2013)).
(Odgaard and Stoustrup (2013)). The distance between The remainder of this section describes the relations
the wind turbines in both directions are 7 times the rotor among the fault cases considered above, and the monitored
diameter, L. Two measuring masts are located in front measurements acquired from the wind park benchmark,
of the wind turbines, one in each of the wind directions in the presence of uncertainty and measurement errors. In
considered in this benchmark model, e.g. 0o and 45o . The this way, it will be shown that the fault isolation task can
wind speed is measured by these measuring masts and be easily solved by using the fuzzy scheme proposed in this
they are located in a distance of 10 times L in front of work, thus representing one of the main motivations of the
the wind farm. The wind turbines of the farm are defined suggested approach. In particular, Table 1 shows the fault
by their row and column indices in the coordinate system effect distribution in the case of single fault occurrence,
illustrated in (Odgaard and Stoustrup (2013)). The farm with respect to the acquired inputs and outputs of the
uses generic 4.8MW wind turbines, which are threebladed wind park simulator.
horizontal axis, pitch controlled variable speed wind tur-
bine. Each of the wind turbines are described by simplified Table 1. The FMEA results for the wind park
models including control logics, variable parameters and 3 benchmark.
states. The ith wind turbine model generates the electri-
Fault affecting Selected measurements Fault
cal power, Pi g (t), the collective pitch angle, i (t), and the wind turbine nr. after FMEA case
generator speed, i g (t). Note that only one measured pitch i = 2, i = 7 {vw (t), 9 (t), P4 g } Fault 1
angle is provided since it is assumed that the wind turbine i = 1, i = 5 {vw (t), 2 (t), P6 g } Fault 2
controller regulates the pitch angles in the same way. The i = 6, i = 8 {vw (t), 3 (t), P7 g } Fault 3
two scenarios with different wind directions but driven
both by the same wind speed sequence vw (t) (possibly with Table 1 was obtained by performing the socalled fault
a time shift) are considered. The wind sequence contains sensitivity analysis, i.e. the Failure Mode & Effect Analysis
a wind mean speed increasing from 5 m/s. to 15 m/s, and (FMEA). In practice, Table 1 is thus built by selecting
with a peak value of about 23 m/s. In this benchmark the most sensitive measurement (ui or yj ) with respect to
model a very simple wind farm controller is used, which the simulated fault conditions. Obviously, when different
provides the wind turbine controllers with a power refer- fault conditions have been considered with respect to the
ence Pi ref (t). If the wind farm is requested to generate scenario of this work, different measurements will probably
a power lower than the available one, the references are be taken into account.
evenly distributed among the wind turbine controllers.
More details on wind farm model considered in this paper 3. FUZZY MODELLING AND IDENTIFICATION
can be found in (Odgaard and Stoustrup (2013)). It is FOR PASSIVE FTC DESIGN
worth noting that the wind farm considered here could
be seen as simplistic model. However, the work (Odgaard
and Stoustrup (2013)) describes how the simulator can fit The proposed FTC method consists of two phases. The
actual wind farm. first stage requires the identification of the nonlinear
dynamic filters, which are required for fault estimation.
With these assumptions, the complete continuoustime From this FDD module, the fault reconstruction is thus
description of the wind farm under diagnosis has the exploited by the control scheme for compensating the
following form: alterations of both the measured and control signals.
1346
SAFEPROCESS 2015
September 2-4, 2015. Paris, France S. Simani et al. / IFAC-PapersOnLine 48-21 (2015) 13451350 1347
The nonlinear identification approach suggested in this residual generator function for fault estimation can be
work employs fuzzy clustering techniques to partition the directly identified again by exploiting a fuzzy identification
available data into subsets characterised by linear be- scheme, inspired by the unknown input reconstruction.
haviours. Relationships between clusters and linear regres- The nonlinear estimator based on a fuzzy inverse model
sion are exploited, thus allowing for the combination of takes into account the process nonlinearities, including the
fuzzy logic techniques with system identification tools. In inherent saturation (level) constraints of the control inputs
addition, an implementation in the Matlab r Toolbox of and other process variables, and is capable of controlling
the Fuzzy Modelling and IDentification (FMID) technique a nonlinear system in the entire operating range.
recalled in the following is available (Babuska (1998)). In
this study, TS fuzzy models are exploited as they are able First, it is assumed that the rulebased model (2) has been
to provide the nonlinear function between measurements identified for describing the continuoustime behaviour of
and faults. The switching and the scheduling between the the complete wind park system in the discretetime form
submodels is achieved through a smooth function of the (3):
system state, the behaviour of which is defined using fuzzy y(k + 1) = F (x(k), f (k)) (3)
set theory.
where x(k) represents the generic system state and f (k)
3.1 Fuzzy Modelling and Identification the (unknown) fault signal. Its TS prototype has the form
(4):
In more detail, the fuzzy modelling and identification al- M (m) (m) (m) (m) (m)
x (k) a x (k) + b
gorithm is based on a twostep procedure, in which at y(k + 1) =
i=1 i i i
(4)
M (m) (m)
first, the operating regions are determined using the data x (k)
i=1 i
clustering technique, and in particular, the Gustafson
Kessel (GK) fuzzy clustering, since already available in The input of the model is the current state x(m) (k) that
(Babuska (1998)). Then, in the second stage, the estima- collects the lagged inputs u(k) and outputs y(k), as well
tion of the controller parameters is achieved using the as the unknown input f (k). The output is a prediction of
identification algorithm already proposed by one of the the systems output at the next sample y(k + 1). In (4)
same authors in (Simani et al. (1999)), which can be seen (m)
as a generalisation of classical leastsquares. The TS fuzzy the estimated membership functions i , the state x(m) ,
(m) (m)
models have the form of: and the parameters ai , bi of the monitored system are
M denoted by the superscript (m).
i=1 i (x(k)) yi
y(k + 1) = M
(2)
i=1 i (x(k)) 3.2 Fault Estimation for FTC Design
where yi = ai x + bi , with ai the parameter vector The objective of the estimator is to compute the unknown
(regressand), and bi is the scalar offset. M is the number
of clusters. x = x(k) represents the regressor vector, input f(k), such that the system output at the next
which can contain delayed samples of u(k) and y(k). sampling instant is equal to the desired output y(k + 1).
The antecedent fuzzy sets i are extracted from the In principle, this can be achieved by inverting the model
fuzzy partition matrix (Babuska (1998)). The consequent of the process. Given the generic current state x(k) and
parameters ai and bi are estimated from the data using the output y(k), the unknown input f(k) is given by:
the procedure presented in (Simani et al. (1999)). f(k + 1) = F 1 (x(k), y(k)) (5)
This section represents the main contributions of the work
and addresses the identification of the fault estimator Generally, it is difficult to find the analytical inverse func-
structure. Once a reasonably accurate fuzzy description of tion F 1 . Therefore, the method exploited here uses the
the considered benchmark has been available, it is used to identified fuzzy TS models of the process under investiga-
directly estimate the nonlinear behaviour between mea- tion (4) for providing the particular state x(m) (k) at each
surements and fault signals. In particular for this study, time step k. From this mapping, the inverse mapping f(k+
the proposed residual generator design methodology relies 1) = F 1 (x(k), y(k)) is easily identified as a prototype in
on the unknown input reconstruction principle, solved the form (2), if the whole system is stable, and in particular
within the fuzzy identification framework. It is wellknown in the form (6):
that for stable fuzzy systems, whose inverted dynamics M (r) (r) (r) (r) (r)
x (k) a x (k) + b
are stable, a nonlinear estimator for the unknown input i=1 i i i
f(k + 1) = M (r) (r) (6)
can be simply designed by inverting the fuzzy model. x (k)
i=1 i
Moreover, in the ideal situation of no modelling errors and
disturbances, this estimator provides perfect reconstruc- where the inputs of the identified controller model are
tion of the unknown input signal with zero steadystate the state x(r) (k) and the current output y(k). In (6) the
errors. However, in practice, one has to deal with both (r)
estimated membership functions i and the parameters
modelling errors and disturbances, which can be tackled (r) (r)
with an arbitrary degree of accuracy by exploiting the ai , bi of the identified controller model are denoted now
fuzzy modelling approach. by the superscript (r). Therefore, the series connection of
the fault estimator and the identified inverse model, should
It is worth noting that fuzzy identification can been yield to an identity mapping,
used to accurately approximate any nonlinear function. when f (k) exists such that
In the same way, this section tries to explain how the y(k + 1) = F x(m) (k), f(k) . However, due to modelling
1347
SAFEPROCESS 2015
1348
September 2-4, 2015. Paris, France S. Simani et al. / IFAC-PapersOnLine 48-21 (2015) 13451350
errors, noise, and disturbance, by means of the fuzzy iden- controller can be easily designed considering the faultfree
tification procedure
described
in Section 3, the difference plant.
|y(k + 1) F x(k), f (k) | is made arbitrarily small by an Finally, regarding the stability analysis of the overall FTC
appropriate choice of the parameters of the fuzzy residual system, the simulation results shown in Section 4 highlight
(r)
generator, i.e. the fuzzy membership functions i , the that the model variables remain bounded in a set, which
(r) (r) assures control performance, even in the presence of faults.
number of clusters M , and the regressand ai , bi . The Moreover, the assumed fault conditions do not modify the
process fuzzy model is used for the recursive prediction system structure, thus guaranteeing the global stability.
of the state vector x(m) (k). Therefore, the state of the However, a few more issues can be considered here. It
fuzzy fault estimator x(r) (k) is updated using the process should be clear that in steadystate conditions, when the
model state x(m) (k) and the output y(k). Apart from the fault effect is completely eliminated, the performances of
(r)
computation of the membership degrees i , both the pro- the FTC method are the same of the faultfree situation.
cess model and the controller are estimated using standard Therefore, the performances of the complete system are
matrix operations and linear interpolations, which makes the same of the faultfree nominal controller. The stability
the algorithm suitable for realtime implementation. properties of the FTC scheme should be considered only in
transient conditions, when the fault is not compensated.
Note that the nonlinear filter estimation and the fault However, it is possible to show that the fault estimation
reconstruction are performed online, thus providing non- error is limited and convergent to zero, thus the stability
linear adaptive filters. The fault estimations are therefore of the complete system is maintained.
exploited for the compensation of both the actuated and
measured signals used by the wind park controller. In
particular, in order to compute the simulation results 4. SIMULATED RESULTS
described in Section 4, the FTC scheme has been com-
pleted by means of the wind park controller described In order to show the capabilities of the proposed FTC
in (Odgaard and Stoustrup (2013)). In this way, the sig- strategy, the system has been simulated as described in
nals from the wind farm controller are compensated by (Odgaard and Stoustrup (2013)). Extended simulations
the FDD module providing the fa control additive fault have been also performed by the authors according to a
signals. On the other hand, also the measured outputs more realistic testrig developed in (Odgaard and Johnson
(2013)). The designed nonlinear filters provide the esti-
are compensated via the fy sensor additive fault signals
mate of the magnitude of the different faults acting on the
reconstructed by the FDD module. After these corrections,
the wind park model, as shown in Section 2.
the wind park controller provides the nominal tracking of
the reference signals. Note that these identified nonlinear
filters, organised into a bank structure, are also able to per- 4.1 PFTCS Performances
form the fault isolation task, as described e.g. in (Simani
(2013)). The overall FTC strategy based on the fuzzy FDD As an example, the identified nonlinear fuzzy filter pro-
module is depicted in Fig. 1. vides the reconstruction fa for the fault case 1 of Table 1,
which is decoupled from the effect of both the wind speed
v(t) and the Cp map uncertainty. In order to compute the
simulation results described below, the FTC scheme has
been completed by means of the standard wind park con-
troller implemented in (Odgaard and Stoustrup (2013)).
The following results refer to the simulation of the accom-
modated controller with fault case 1 of Table 1. Hence,
after the proper identification of the fuzzy residual genera-
tor parameters, the nonlinear fuzzy estimator provides the
reconstruction of the fault mode, with minimal detection
delay. The tests refer to the simulation of the actuator
fault modelled as a sequence of rectangular pulses with
Fig. 1. Diagram of the integrated FTC strategy. variable amplitude and length. Fig. 2 shows the estimate
of the intermittent actuator fault fa (dotted grey line),
Fig. 1 describes the structure of the FTC system where compared with the actual fault (dashed black line). It is
u is the controller output, y and Ref are the measured shown that the fault estimation module provides a quite
and the reference outputs, respectively. The signal f is the accurate reconstruction of the fault signal.
generic estimated fault, i.e. the fa or fy additive signals. Under this condition, Fig. 3 shows the power reference
Therefore, Fig. 1 shows that the FTC system is obtained
signal (continuous black line) compared with its desired
by integrating the fuzzy FDD module with the baseline
value (grey dotted line), with fault accommodation.
control system. The FDD module, consisting of a bank
of fuzzy estimators, provides the correct estimates of the In particular, Fig. 3 depicts the power reference to the
fault signals. Each fuzzy estimator of the bank provides wind farm, Pref , which is constant and equal to 43.6MW
the correct detection (isolation) and the estimation of the until t = 2000 s., when it changes to about 30 MW. It
corresponding faults. These estimated signals are injected is worth observing that the suggested fuzzy estimators
into the control loop in order to compensate the effect of provide not only the fault detection, but also the fault
the faults. Thanks to this fault estimation feedback, the estimate. Moreover, a fault modelled as a sequence of
1348
SAFEPROCESS 2015
September 2-4, 2015. Paris, France S. Simani et al. / IFAC-PapersOnLine 48-21 (2015) 13451350 1349
1349
SAFEPROCESS 2015
1350
September 2-4, 2015. Paris, France S. Simani et al. / IFAC-PapersOnLine 48-21 (2015) 13451350
designed control algorithms considering more realistic con- Mahmoud, M., Jiang, J., and Zhang, Y. (2003). Active
ditions that the wind park may experiment with during Fault Tolerant Control Systems: Stochastic Analysis and
its working situations. Therefore, for this purpose, a HIL Synthesis. Lecture Notes in Control and Information
testbed was exploited in this paper and applied to the Sciences. SpringerVerlag, Berlin, Germany. ISBN:
wind park, in order to analyse the capabilities of the 3540003185.
developed FTC approach also in realtime conditions. The Noura, H., Theilliol, D., Ponsart, J.C., and Chamseddine,
results achieved for a single test are summarised in Table A. (2009). Faulttolerant Control Systems: Design and
4 for the proposed FTC system. Moreover, the approach Practical Applications. Advances in Industrial Control.
described in this study has been compared with respect Springer, London, 1 edition.
to the PFTCS developed in (Simani et al. (2014)) and the Odgaard, P.F., Stoustrup, J., and Kinnaert, M. (2013).
adaptive FTC system in (Simani and Castaldi (2012)). On FaultTolerant Control of Wind Turbines: A Bench-
the other hand, the FTC strategy relying on the analytic mark Model. IEEE Transactions on Control Systems
disturbance decoupling approach addressed in (Simani and Technology, 21(4), 11681182. ISSN: 10636536. DOI:
Castaldi (2014)) is also considered. 10.1109/TCST.2013.2259235.
Odgaard, P.F. and Johnson, K. (2013). Wind Turbine
Table 4. HIL results with the considered FTC Fault Diagnosis and Fault Tolerant Control an En-
solutions: N SSE% values. hanced Benchmark Challenge. In Proc. of the 2013
FTC Scheme Fault 1 Fault 2 Fault 3 American Control Conference ACC, 44474452. IEEE
N SSE% Control Systems Society & American Automatic Con-
Fuzzy identified FDD trol Council, Washington DC, USA. ISSN: 07431619.
module + FTC 13.74% 14.37% 15.01% ISBN: 9781479901777.
Odgaard, P.F. and Stoustrup, J. (2013). Fault Tolerant
Analytic Wind Farm Control a Benchmark Model. In Pro-
disturbance 14.07% 15.06% 15.34%
decoupled FTC
ceedings of the IEEE Multiconference on Systems and
(Simani and Castaldi (2014)) Control MSC2013, 16. Hyderabad, India.
Simani, S. (2013). Residual Generator Fuzzy Identification
PFTCS 32.64% 33.53% 34.98% for Automotive Diesel Engine Fault Diagnosis. Interna-
(Simani et al. (2014)) tional Journal of Applied Mathematics and Computer
Science AMCS, 23(2), 419438. Invited Contribution
Adaptive FTC 31.53% 31.94% 32.07% to the AMCS Quarterly. Organisers: Koscielny, M. J.
(Simani and Castaldi (2012)) and Syfert, M. ISSN: 1641876X. DOI: 10.2478/amcs
20130032.
Regarding the FTC system proposed in this study, Table Simani, S. and Castaldi, P. (2012). Adaptive faulttolerant
4 illustrates that there are some deviations between the control design approach for a wind turbine bench-
achieved results, but consistent with the ones from the mark. In C.M. Astorga-Zaragoza and A. Molina (eds.),
MonteCarlo analysis. Although there are some deviations 8th IFAC Symposium on Fault Detection, Supervision
between the simulation and the experimental results, these and Safety of Technical Processes SAFEPROCESS
deviations are not critical. Moreover, the results obtained 2012, volume 8, 319324. Instituto de Ingeniera, Cir-
here seem reliable enough for future wind farm real ap- cuito escolar, Ciudad Universitaria, CP 04510, Mexico
plications. Moreover, the comparison of Table 4 highlights D.F., IFAC, Mexico City, Mexico. Invited session pa-
that this FTC scheme can achieve better performances in per. ISBN: 9783902823090. ISSN: 14746670. DOI:
terms of tracking error. 10.3182/20120829-3-MX-2028.00066.
Simani, S., Fantuzzi, C., Rovatti, R., and Beghelli, S.
5. CONCLUSION (1999). Parameter identification for piecewise linear
fuzzy models in noisy environment. International Jour-
The paper addressed the development of a passive fault nal of Approximate Reasoning, 1(22), 149167. Pub-
tolerant control scheme applied to an offshore wind farm lisher: Elsevier.
simulator. In particular, a datadriven approach relying on Simani, S., Farsoni, S., and Castaldi, P. (2014). Fault
fuzzy model identification was exploited to build the fault tolerant control of an offshore wind turbine model via
accommodation solution. This simple approach relying on identified fuzzy prototypes. In J.F. Whidborne (ed.),
fuzzy rules can represent an alternative to possible more Proceedings of the 2014 UKACC International Con-
complex design solutions that would have been inevitably ference on Control (CONTROL), 494499. UKACC
derived from purely nonlinear modelbased methods. This (United Kingdom Automatic Control Council), IEEE,
point can represent also a very important aspect when on Loughborough University, Loughborough, UK. ISBN:
line implementations are required for a viable application 9781467306874. Special session invited paper. DOI:
of the proposed strategy. A highfidelity offshore wind 10.1109/CONTROL.2014.6915188.
farm simulator realised via the hardwarein-the-loop tool Simani, S. and Castaldi, P. (2014). Active Actuator Fault
was finally exploited to validate the achieved performances Tolerant Control of a Wind Turbine Benchmark Model.
of the suggested scheme. International Journal of Robust and Nonlinear Control,
24(8), 12831303. John Wiley. DOI: 10.1002/rnc.2993.
REFERENCES
Babuska, R. (1998). Fuzzy Modeling for Control. Kluwer
Academic Publishers, Boston, USA.
1350