2010 11th Int. Conf.

Control, Automation, Robotics and Vision Singapore, 7-10th December 2010

A Sligthly Improved Robust Adaptive Observer for Actuator Fault Detection

Kamel Menighed, Joseph J. Yam, Christophe Aubrun
Centre de recherche en Automatique de Nancy - CRAN - UMR 7039 Universit Henri Poincar Nancy 1 B.P. 70239, Vandoeuvre-les-Nancy F-54506, France
AbstractIn this paper, we address the problem of actuator fault detection and estimation for a plant in the presence of external disturbances. We design a robust fault estimation technique based on an adaptive observer with enhanced speed of adaptation. The construction of the observer is carried out through a transformation of the plant model into its special coordinate basis (SCB) form. This transformation allows a decoupling of the fault estimates from the disturbances. Thanks to this decoupling, the standard adaptive observer is slightly modied to achieve faster fault estimation in presence of disturbances and plant-model uncertainties. Simulation experiments are presented to illustrate the effectiveness of the proposed adaptive scheme. Index TermsFault estimation, actuator fault, robust residuals, adaptive observer, uncertain systems.

I. I NTRODUCTION With nowadays high-level automation integration in industrial systems, plants with their instrumentation - sensors, actuators- become more prone to failures. Such failures might have drastic consequences on plant operation such as closedloop performance degradation or even instability with a serious impact on human safety and the environment. It becomes therefore mandatory to increase signicantly the reliability of automated systems with devices capable of a continuous monitoring of the plant states. For economic and hardware weight reasons, an efcient realization of such devices tends to be algorithmic-based systems which can be easily implemented on existing control process computers. The impetus to this algorithmic-based approach for process monitoring begun in the early seventies [2] and is now known as the analytical approach to fault detection and isolation (FDI). The core of this approach is based on a mathematical model of the controlled plant which allows faults characterization in plant-components from a comparison of the signals generated by this plantmodel with the available actual plant-measurements. A key problem with analytical models is that the real plants are subject to unknown disturbances and models are never perfect and may be imprecisely know. These uncertainties should be taken in account at the outset for the design of FDI algorithms. Fortunately, some unmeasurable perturbations can often be incorporated in plant models by viewing them as unknown inputs. Hence, when the unknown inputs are certain causes such as faults in actuators or other plant-components, thanks to the above view, unknown input estimation may be carried out directly for the purpose of fault detection and isolation. This problem of direct fault estimation when the plant is subject c 2010 IEEE 978-1-4244-7815-6/10/$26.00

to uncertainties and disturbances is the main purpose of this paper. Amongst many methods and techniques for fault detection and estimation which have been developed over the past several years, see e.g. [3], the unknown input observer (UIO) is one of the most attractive techniques [5], [6]. For uncertain systems, FDI based on unknown input observers which are sensitive to faults but insensitive to model uncertainties and/or unknown disturbance have been proposed in ([?], [3]). In reference [4], the authors proposed an input estimator in a more general setting not directly related to the FDI problem, however their technique can be viewed as an approach for fault estimation, albeit no distinction is made between disturbances and fault signals. Although exact asymptotic estimation is not achieved with this technique, it allows asymptotic estimation of the unknown inputs to any degree of accuracy. The main drawback of the technique is that the existence of the proposed estimator depends on very strict conditions. Thanks to the advances in adaptive observer theory [8], the authors in reference [11] proposed a Luenberger-type adaptive observer to simultaneously estimate the state and the unknown fault. The fault was viewed as the adjustable (unknown) parameter of the observer and the adaptive law for updating the fault estimation was chosen judiciously in order to ensure stability and convergence of the fault estimation algorithm. However, an adaptive observer based on the nominal plant-model might have some troubles in estimating faults when the real plant is uncertain and subject to disturbances. In order to cope with this issue related to plant uncertainties/disturbances, references [13], [12] presented a nice solution with a fault isolation observer capable to produce residuals decoupled from the disturbances. In this paper, we strongly rely on the above mentioned results to construct a robust adaptive observer able to estimate actuator faults with enhanced estimation speed despite the presence of disturbances and plant uncertainties. This paper has some connections with the work from [14], however in this latter reference the robustness problem is not addressed and furthermore, we derive stronger results with generalized adaptation mechanisms for fault estimation. The paper is organized as follows. In section II, we recall some preliminary materials on the adaptive diagnostic observer. In section III, we state our problem and delineate a path to its solution. Section IV collects some of the main properties of the special coordinate basis tranformation which

1322

ICARCV2010

is used in section V to establish the algorithm used in the design of the robust adaptive observer. A numerical example is presented in section VI to illustrate the effectiveness of the robust adaptive observer algorithm. II. P RELIMINARY MATERIALS A. Plant description We consider a nominal plant subject to actuator faults and represented by the following linear time-invariant model : { () = () + () + () (1) () = () where () , () and () are the state, control input and output vector respectively and () represents the actuator faults. The matrices , and are known constant matrices of appropriate dimensions and denotes the distribution matrix for actuator faults. Furthermore, we assume that is of full column rank and the pair (, ) is observable. Note that full rank assumption on is necessary for estimation of (). The failures () are viewed as additive signals in the plant model (1) when they occur. This can be written explicitly using the indicator function as { 0, with ( ) = () = ( ) () 1, > where is the time of fault occurrence and () the fault signal. B. The adaptive diagnostic observer Using the plant model (1), the standard adaptive diagnostic observer proposed in [11] is written slightly modied as () = () + ( () () + () + ()) () = () (2)

then the adaptive observer given by () = () + ( () () + () + ()) () = () with adaptation law () = ( () ()) viz () =

(6)

(7) (8)

( ( ) ( ))

() = 0, where the () = 0 and achieves SPD matrix is used to prespecify a learning rate for (7). Note that the above standard adaptive diagnostic observer is derived on the basis of the nominal plant model (1) subject to actuator faults. This adaptive observer will get into trouble in case the plant experiences input disturbances and model uncertainties. Note also that the actuator fault estimate may be seen as the output of an integrator. Due to the low-pass ltering property of integrators, it is clear that such estimation scheme will be able only to deal with constant or very slow time-varying fault. III. P ROBLEM S TATEMENT Consider the plant subject to actuator faults and uncertainties. These uncertainties might be possibly external disturbances acting on the plant and/or dynamics uncertainties due to modelling errors. Thus, the plant is described with the following model { () = () + () + () + () (9) () = () where () is the state, () is the control input, () is the output. The matrices , , are the nominal plant matrices. As previously dened in the last section, signal () with , represents the vector of actuator faults and denotes their distribution matrix. Vector () denotes all the unknown inputs and is the distribution matrix for the unknown input signal. Note that both model uncertainties and input disturbances have been represented through the term in the above formulation. It is assumed that the matrix is perfectly known and full rank. The issue we address is how to design an adaptive observer to estimate arbitrary time-varying actuactor faults despite model-uncertainties and unknown input disturbances. In order to solve this problem, our road map will be the following. First, we need to design a disturbance decoupled residual in order that the observer fault estimation error be independent from the uncertainties and disturbances. Second, if this decoupling is feasible, the estimation updating algorithm should be devised as a fonction (or lter) of the residual signal. To proceed, we consider an adaptive observer for simultaneous residual generation and fault estimation having the following structure

() is an estimate (to be determined) of the where actuator fault (). From the observability assumption of the pair (, ), the observer gain matrix can be selected such that ( ) is stable. Set () = () () () = () (), () = () (), (3) then the error dynamics is given by () () = ( ) () + () = () (4)

We quote the following proposition from [11] for computing the fault estimate, see also [4]. Proposition 1: If there exist symmetric positive denite (SPD) matrices , and two matrices and such that the following conditions hold ( ) + ( ) = , = (5)

1323

 () = () + ( () () + () + ()) () = () () = ( ()) () = () ()

(10)

is the fault estimation of the where is the state estimation, real fault and is the convolution of the residual with the impulse response of some linear lter to be designed. In order to increase the degree of design freedom, the residual is dened as a post-ltering of the output estimation error by a constant matrix . Substracting the real dynamics (9) with that provided by the observer (10) gives the following errors dynamics () + () () = ( ) () + () = () (11)

internal structure of systems by revealing their distinctive parts which are directly associated with their zero dynamics as well as their left and right invertible dynamics. Here, we summarize the structural decomposition of a linear system in the following main theorem and gives some of its key properties [9, Chap. 3]. Proposition 2: For any given linear system described by the following equations { () = () + () () = () + () with = [0 ]

(15)

It is clearly seen that the errors dynamics is corrupted by the uncertainties and unknown disturbances (). The residual vector can be written as () = () = () (12)

there exists + 1) coordinate free nonnegative integers , , , , , and , = 1, 2, ... , = =1 and 2) nonsingular state, output and input transformation 1 , 2 and 3 which take the given into a special coordinate basis (SCB) that displays explicitly both the nite and innite zero structures of . The SCB is described by the following set of equations :
1 1 1 = = 1 , 2 , = 3 + = [( ) , ( ) ] , = [( ) , ( ) , , , ] ,

This yields the relationship between residual and faults signal and unknown inputs which we write in transfer operator form () () = ()() + () (13)

with () = ( + )1 the transfer from disturbance to residual and, () = ( + )1 the transfer from actuator fault to residual. In the light of the above adaptive observer structure, the tasks to be done are to nd 1) suitable matrices and such that is Hurwitz and, independently of the unknown inputs , the residual should be zero (or close to zero) in the absence of faults and different from zero in case of fault occurence , where 2) an adaptive law for continuously updating is seen as the output of a to-be-designed lter () which ensures convergence and speed of the fault estimation. In the two next sections, we present an algorithm based on the special coordinate basis (SCB) transformation to select and such that () = 0, and when there is no fault () = 0, > 0 when some fault occurs (14)

= [( ) , ( ) ] { () () = () + () = () , are given by where , = := 1 1 1 0 0 + 0 0 + + 0 + 0 + 0 0 0 0 0 1 0 := 1 3 0 0 0 [ ] := 1 1 0 0 0 0 2 0 0 0 0

(16)

(17)

(18)

(19)

and the adaptive law devised for updating the fault estimate. IV. SCB TRANSFORMATION FOR ROBUST RESIDUAL
GENERATION

The so-called structural decomposition basis (SCB) allows an equivalent representation of a dynamical system [10][9, Chap. 3] and it has emerged as a key tool for many aspects of system analysis and design. Furthermore, there is now available toolkits with algorithms for computing SCB [7]. Roughly speaking, this representation provides deep insight into the

Some of the key properties revealed by the system SCB form, which are of interest to us, are listed below Property 1: System is right invertible if and only if and hence do not exist, i.e., is equal to zero; it is left invertible if and only if and hence do not exist, i.e., is equal to zero. Property 2: + + = + , where , are the number of stable and unstable transmission zeros of system respectively. Property 3:

1324

 is the number of innite zeros with = =1 , where is number of innite zeros of order and is the highest order of an innite zero, also it is the row number of . Property 4: = + + + Property 5: ( , ) forms an observable pair. The following proposition which is the basis for realizing the disturbance decoupled (DDR) residual is from [13] and the proof is sketched for the reader to get some insight into the DDR algorithm. Proposition 3: [13] There exists matrices and such that the observer given in equation (10) satises the disturbance decoupling property, with being Hurwitz, if and only if the pair (, ) is detectable and system (, , ) is not right invertible. Proof : Assume (, , ) is transformed in SCB form. From the requirement of disturbance decoupled residual, the transfer from disturbance to residual () = ( + )1 (20)

with () = (1 () + 2 1()) (24) where is a matrix to be determined, the impulse response of the lter and 1 , 2 are lter tuning matrix parameters (of dimension ) which are assumed to be positive denite matrices. The symbols () and 1() denotes the Dirac impulse and the Heaviside function respectively. Let us disregard, for the moment, the unknown disturbance and assume that is computed as in proposition 1. Then, from the Kalman-Yakubovitch-Popov lemma [1], the error-dynamics system with output dened by () () = ( ) () + () = () (25) (26)

is strictly positive real (SPR). Now, consider the adjusting law (23) in connection with the error dynamics system (25)-(26) as depicted in gure.

must be equal to zero and this can be achieved if and only if has the form
= = [

0 0]

(21)

where has columns. Thanks to the SCB form of matrix , partition as = [1 2 ] where 1 and 2 have compatible dimensions with then = [0 0 2 0 1 ] (22) We claim the following result. Proposition 4: Assume that matrix is computed from conditions (5) of proposition 1 . Then, under the positive definiteness of matrices 1 and 2 , the closed-loop system (25)=0 (26)-(23)-(24) is asymptotically stable, that is, lim = 0. and lim Proof: From the positive deniteness of matrices 1 and 2 , it follows that the lter with impulse response () is a positive real system, that is, a passive system. Therefore, the closed-loop (25)-(26)-(23)-(24) is a negative feedback interconnection of a strictly passive system and a passive one. The result follows as a direct application of the passivity theorem for linear systems [1, Chapter 2]. Remark 1: A similar proportional and integral adaptation law has been used in [14]. It was only shown there that the adaptive observer achieves uniformly bounded estimation . Proposition 4 is a stronger result which shows errors and that proportional and integral adaptation laws really achieves exact asymptotic estimation of the fault. Note that under the positive deniteness of the lter matrix parameters 1 and 2 , the passive property of the lter () is independent of the magnitude of the elements of these matrices and, consequently the asymptotic stability of the closed-loop system (25)-(26)-(23)-(24) is always guaranteed. Therefore, with large magnitude elements of the proportional

Comparing (21) and (22) implies that 1 must be zero. The term 2 would be nonzero if and only if = 0. In this case, we can therefore nd a matrix such that = 0 to satisfy the requirement () = 0 and a matrix to ensure stability of ( ). V. ROBUST ADAPTIVE FAULT E STIMATION D ESIGN WITH
ENHANCED ADAPTATION SPEED

In the observer (10), the unknown parameter representing the actuator fault is replaced by its estimate. Then, by means of the updating law, the parameter should be quickly adjusted in real time such that its value converges to the real value in the plant. Since we assume that the fault might be a possible arbitrary time-varying signal, the adaptive law should be such that the fault estimate tracks closely the real value of the fault signal. Note that the adjustment law in the standard adaptive observer (6)-(7)-(8) is such that the fault estimate is the output of an integrator whose input is the ouptut estimation error . There are, surely, many other alternatives for choosing the adaptation mechanism. It might be expected that faster adaptation rule can be achieved with lter having proportional and integral terms. This means that the updating law may have the form of a ltering process () = { () ()} (23)

1325

term 1 , it can be expected that quicker adaptation can be achieved. It is clearly seen from proposition 4 that a more general convergent adaptation scheme can be envisioned as a ltering of the output-error , provided that the lter is passive. An example of such a passive lter might be a proportionalintegral-and-derivative (PID) lter leading to an updating law of the form +3 +2 ( ) +3 +2 = 1 = 1

As the pair (, ) is detectable, (, ) is also detectable. Choose a gain such that eigenvalues ( ) are assigned at desired location in ] and form the . Partition as = [ matrix as = (30) 0 = [ ] such that = Select matrix 0, and is any nonsingular matrix with compatible dimension with such that = 0 . Note that we can choose = . Finally, compute the desired and as 1 , = 1 2 1 = 2 (31) using the estimation law (23) where 2) STEP 2 - Compute () is the impulse response of a passive lter (e.g., PI or PID-type lter) a) Matrix is computed from conditions (5) b) The updating fault estimation mechanism is derived from () = { () ()} has compatible diwhere = [ ] and mension with Remark 2: Note that, apart from the existence of matrix imposed by conditions (5), the algorithm is conditioned to the non right invertibility of the triplet (, , ) which is the condition = 0. VI. N UMERICAL E XAMPLE In this example, the algorithm is applied to implement a robust adaptive fault estimation with enhanced speed using a PI-type lter for the updating estimation mechanism. Consider the following continuous linear systems with unknown inputs [14]. A. Nominal (non faulty) case {

(27) where 1 ,2 and 3 are positive denite. This general result is stated in the next proposition. Proposition 5: Assume that matrix is computed from conditions (5) of proposition 1 . Then any adaptation law obtained as a ltering of the -weighted output-estimationerror with a lter having a positive real transfer matrix () achieves exact asymptotic fault estimation. Furthermore, the closed-loop error-dynamics of the adaptive observer is asymptotically stable. Now, we are in a position to derive the robust adaptive fault estimation with enhanced adaptation speed. We simply combine the disturbance decoupling residual generation with an updating mechanism law of PI or PID type. The algorithm proceeds through the following steps. 1) STEP 1 - Compute and using "SCB transformation" Find nonsingular state, output and input transformation 1 , 2 and 3 which take the given = {, , } into the form of (SCB) given by : := 1 3 and := 1 1 := 1 1 , 1 1 2 0 0 0 = , 0 0 ] [ 0 0 = = 0 0 0 , (28) 0 0 0 0 0

() = () + () + () (32) where the row dimension of is , and row dimen () = () sion of is the maximum dimension of residual, that where matrices , , , are is, = 1-a) If = 0, stop 9.9477 0.7476 0.2632 5.0337 1-b) Otherwise continue 52.1659 2.7452 5.5532 24.4221 , = a) Since ( , ) forms an observable pair, choose a 26.0922 2.6361 4.1975 19.2774 gain such that eigenvalue ( ) are 0 0 1 0 placed at desired location in b) Dene matrices and as 0.4422 0.1761 0 1.0000 0 3.5446 7.5922 , = 0.0620 0.1320 , , = [0 0 ] = = 5.52 4.49 0 7.1890 0 0 0 . 1000 0.3000 (29)

1326

1 0 = 0 0

0 1 0 0

0 0 1 0

0 0 0 1

x 10

-12

Residual

1.5

Using SCB transformation, we can easily show that this system has = 2 > 0 and the pair (, ) is observable. Therefore, this system satises existence conditions to proceed with the algorithm. Following the design algorithm step by step for disturbance decoupling conditions, we get , as [ ] [ ] 2.3054 16.9099 9.2580 7.3713 , = , = 8.0918 36.8087 0.2822 36.6453 [ = 1 0 0 1 ] , = [ 1 0 0 1 ]

0.5

r(t)

-0.5

-1

-1.5

-2 0

200

400

600

800

1000

1200

1400

1600

1800

2000

time t

Figure 1.
5

Residual output using the robust adaptive observer for 1 .

-11

The row dimension of is = 2. Then, the maximum dimension of residual vector, that is, = = 4 2 = 2. The matrices and are designed as 3.2580 7.3713 0 0 0.2822 27.6453 0 0 = 21.5684 55.7305 2.2454 16.9099 , 81.2088 44.9979 8.0918 36.8987 [ ] 7.1890 0.0000 1.0000 0 = 0.5129 1.6129 0 1.0000 B. Fault case The objective is to estimate actuator faults despite the existence of the unknown inputs. In this situation, an actuator fault occurs in the second channel input and the actuator fault distribution matrix is 0.1761 7.5922 = 4.4900 0 We have compute and from rst stage, and we can easily verify that is Hurwitz. Now, solving (5), one obtain 1.0417 0.7267 0.1133 0.4500 0.7267 0.6072 0.0749 0.3761 = 1.0+005 0.1133 0.0749 0.0126 0.0464 0.4500 0.3761 0.0464 0.2330 = 1.0 + 005 [ 4.8255 4.1461 0.4922 2.5682 ]

x 10

Residual

r(t)

-1 0

200

400

600

800

1000 time t

1200

1400

1600

1800

2000

Figure 2.

Residual output using the robust adaptive observer for 2 .

Assume that constant actuator fault is given by: { 0 0 400 () = 30 400 < 2000 Fig.3 shows the estimation result given by the robust adaptive observer for a constant fault. Starting from the time the fault occurred, .., = 400, the estimated value of the locked position converges to the real value (.., 30) with speed and accuracy. Now,consider a time-varying actuator fault with sawtooth waveform given by ( + 400) ( + 1) T=20 () =
=0

= 1.0 + 005 [4.8255 4.1461] for () = where [1 1] . The unknown input vector [ ] is assumed to be 2(0.3) 1 () = 2(0 [ .25) ] 10 (0.3) we have (see gure 2) for 2 () = 9 (0.25)

Fig.4 illustrates the simulation result for time-varying fault estimation using the robust adaptive observer, it is shown with the solid line. It can be seen that robust fault estimator tracks the fault almost immediately without estimation error, i.e., it gives us the actual shape of the fault, even in the presence of disturbances. VII. C ONCLUSION In this paper, a robust adaptive observer for deterministic and uncertain systems has been designed using a disturbance

1327

True actuator fault and its estimation

35 estimated fault real value 30

25

20

15

10

-5 0

200

400

600

800

1000 Time

1200

1400

1600

1800

2000

Figure 3.

Estimation of constant fault using the robust adaptive observer.

True actuator fault and its estimation

5 estimated fault 4 real value

[6] P.M. Frank and J. Wunnenberg "Robust fault diagnosis using unknown input observer. In Fault Diagnosis in Dynamic Systems:Theory and Application, Patton R.J. (eds). Prentice Hall International, UK, pp. 47-98, 1989. [7] X. Liu X, B.M. Chen and Z. Lin, "Linear systems toolkit in Matlab: Structural decomposition and their applications", Journal of Control Theory and Application,Vol. 3, pp. 287-294, 2005. [8] Narendra S. K., Annaswamy A.M., Stable Adaptive Systems, Dover Publications, 2005 (republication of the work published in 1989 by Prentice-Hall Inc.) [9] Saberi A., Stoorvogel A.A., Sannuti P., Filtering Theory, With Applications to Fault Detection, Isolation, and Estimation, Birkhaser, 2007 [10] P. Sannuti and A. Saberi, "Special coordinate basis for multivariable linear systems nite and innite zero structure, squaring down and decoupling". International Journal of Control, Vol. 45(5), p.p. 16551704, 1987. [11] H. Wang and S. Delay "Actuator fault diagnosis: An adaptive observerbased technique" IEEE transactions on Automatic Control, 41, 10731078, 1996. [12] Y. Xiong and M. Saif, "Robust Fault Isolation Observer Design", Proceeding of the American Control Conference, pp. 2077-2081, 1999. [13] Y. Xiong and M. Saif, "A Novel Design for Robust Fault Diagnosis Observer" Proceeding IEEE Conference on Decision and Control, pp. 592-597, 1998. [14] K. Zhang, B. Jiang, and V. Cocquempot "Adaptive Observer-based Fast Fault Estimation". International Journal of Control, Automation and Systems, Vol 6.(3) pp. 320-326. June 2008.

Amplitude

Amplitude
3 2 1 0

-1 0

500

1000

1500

2000

2500

Time

Figure 4. Estimation of time-varying fault using the robust adaptive observer.

decoupled residual generator combined with a fairly generalized adaptation mechanism for actuator fault estimation. The adaptive observer is a Luenberger-type observer and the adaptation mechanism is an error lter which can be any system having positive real properties. It is shown that the algorithm assures a perfect disturbance decoupling residual generation, an asymtoptotic exact estimation and may improve drastically the estimation speed through particular structures of the lter, as e.g., lters with a proportional term. Numerical simulation has been performed to validate the effectiveness of the designed adaptive observer. R EFERENCES
[1] Bao J., Lee P.L. - Process Control - The Passive System Approach, Springer-Verlag, 2007 [2] R.V. Beard, "Failure Accommodation in Linear Systems Through, SelfReorganization", Dept. MTV-71-1, Man Vehicule Laboratory, Cambridge, MA. 1971. [3] J. Chen and R.J. Patton, "Robust model-based fault diagnosis for dynamic systems"Kluwer Academic Publisher, 1999. [4] M. Corless and J. Tu, State and input estimation for a class of uncertain systems. Automatica , Vol.34, pp.757-764. 1998. [5] M. Darouach, M. Zasadzinaski and S.J. Xu, "Full-order observers for linear systems with unknown inputs" IEEE Transactions on Automatic Control, 39(3), pp. 606-609, 1994.

1328