Vous êtes sur la page 1sur 7

Fault diagnosis and fault-tolerant control for a nonlinear electro-hydraulic system

Liang Chen and Steven Liu


Abstract This paper proposes a fault-tolerant control strategy for a nonlinear electro-hydraulic system. The fault-tolerant control scheme consists a fault diagnosis system and a recongurable controller. The fault diagnosis system is designed based on differential-geometric approach and is robust to system disturbance. Adaptive backstepping method is utilized for realizing the recongurable controller. The control performance of the electro-hydraulic system can be guaranteed even in the presence of considered faults.

I. INTRODUCTION Electro-hydraulic systems are widely used in todays industry. Compared to electrical drives they can generate large forces or torques very fast with simple structures. Control of electro-hydraulic systems is a challenging task due to inherent nonlinearities from complicated ow properties, friction in actuator, varying external load, etc. Advanced control strategies are necessary for highly demanding application [1]. Meanwhile, reliability and safety are most important issues of controlled electro-hydraulic systems. They often work in critical places, like in automobiles or aircrafts. A small fault, such as leakages in cylinders or valves, sensor faults (noise, offset), etc, can bring serious problems. For example, if the landing gear can not work properly due to a leakage in the actuation cylinder, the airplane may crush during the landing. To improve safety and reliability of electro-hydraulic systems much research on fault diagnosis in this area has been done: in [2] the adaptive robust approaches is applied to construct the fault detection observer; in [3] Unscented Kalman lter is used for detecting the faults in a hydraulic actuator; in [4] linear observer and adaptive threshold are employed for robust fault diagnosis of a nonlinear electro-hydraulic system. Moreover, if a controller is able to tolerate possible faults automatically, the control is known as fault-tolerant control (FTC) [6], which can broadly classied into two types: passive FTC and active FTC. The passive FTC systems usually can be designed with robust control approaches [7]. The active FTC contains a fault diagnosis system and reacts to faults actively by reconguring control actions based on the fault information. Thus the pre-specied performance of the system can be ensured. Various active FTC approaches are already available, such as pseudo inverse [8], model predictive control [9] and adaptive control [10]. This paper proposes a novel fault diagnosis integrated fault-tolerant control for a nonlinear electro-hydraulic sysLiang Chen and Steven Liu are with Control Systems Research Group, University of Kaiserslautern, Erwin-Schroedinger-Str. 12/332, 67663 Kaiserslautern, Germany. Email: sliu@eit.uni-kl.de

tem. The fault diagnosis system based on differentialgeometric approach is robust to the system disturbance and provides updated fault information for the controller. The controller is designed with adaptive backstepping technique and can be easily recongured to accommodate the faults. This scheme make the whole system more intelligent and reliable. Furthermore, it is not extreme complicated and can be implemented in practice. The experimental results are shown to illustrate the advantages. II. S YSTEM D ESCRIPTION A. System modeling
:RUNLQJ F\OLQGHU /RDG F\OLQGHU

'9

'9

'9

'9

'9 '9 '9

0
0

6HUYRSXPS 7DQN

*HDU SXPS

Fig. 1.

The hydraulic circuit of the testbed

The electro-hydraulic system used in our laboratory, whose hydraulic diagram is shown in Fig. 1, simulates the extruding machine and is a pump-controlled hydraulic system. It consists of two identical cylinders, which are named working cylinder and load cylinder, respectively. The working cylinder represents the press stem of an extrusion machine. Its uid power is supplied by a variable displacement axial piston servo-pump. The load cylinder simulates the resistance force, which is generated when the material is pressed. With a proportional pressure relief valve an arbitrary load can be realized. The load cylinder is connected with a gear pump so that load pressure can be built up quickly. The direction change of uid ow is realized through several directional valves, which are labeled with DVi in Fig. 1. The system dynamics consists mainly of two parts: pump model and cylinder model. The servo-pump itself is a closed-loop control system. The output ow of the pump Qp is linearly adjusted by a non-

rotating swashplate. The model of the pump is governed by the dynamics of swashplate, which can be approximated by the following second order system v Qp = = = v p1 p2 v + p1 kpu uin , KQ (1) (2) (3)

Combining the equations (1)-(6) the model of the system in fault free case can be derived xc = vc 1 vc = (PA A1 PB A2 Ff FL ) m EB (A2 vc QB ) PB = V20 A2 xc EA (A1 vc + KQ ) PA = V10 + A1 xc = v v = p1 p2 v + p1 kpu uin .

(9)

where and v are the angle and angular velocity of the swashplate respectively, uin is the input voltage, p1 , p2 and kpu are constant. KQ is a constant determined by the pump geometry. The cylinder piston motion can be described as: vc (4) 1 vc = (PA A1 PB A2 Ff FL ), (5) mc where xc is the piston position, vc is the piston velocity, A1 and A2 are cross-section area of piston and ring-side respectively, PA and PB are the pressures in chamber A and B respectively, mc denotes the moving mass, Ff represents the friction force and FL is the varying external load. By ignoring the static friction the mathematical description of Ff is Ff = D v v c + f c , where Dv is the coefcient of viscous friction and fc is the Coulomb friction. Dv and fc are constant and can be identied through experiments. The pressure dynamics in cylinder chambers can be expressed as follows [1] EB (A2 vc QB ) (6) V20 A2 xc EA PA = (A1 vc + QA ). (7) V10 + A1 xc Here Ei is the effective bulk modulus and dependent on the pressure, Qi denotes the volume ow, V10 and V20 are the initial volume of chamber A and B respectively. The leakage ows of cylinder are omitted in fault free case. The volume ow QA is the same as the output ow of the servo-pump due to the pump-controlled structure of the hydraulic system. The volume ow QB is dependent on the characteristic of the directional valves and can be described by if > 0, QB = 0 if 0, PB = where = kqb1 (PB Po ) + kqb2 with tank pressure Po , constant parameters kqb1 and kqb2 . The effective bulk modulus has signicant effect to the pressure dynamics and can not be simplied as a constant. It can be calculated by an empirical formula Pi /Pat + kair Ei = , kair Eo /Pi + Pi /Pat (8) xc =

All system variables can be directly measured except the vc and v , which are calculated through numerical differentiation of xc and followed by a low-pass lter respectively. The external load FL varies during the motion of cylinder piston and is viewed as an unknown disturbance of the system model. B. Considered faults The usual faults of hydraulic systems include pulsation and pressure drop of pumps, leakages in valves and cylinders, sensor noise or offsets, oil pollution, etc. The considered faults of the testbed are internal leakage of the cylinder QLin and the offsets of pressure sensors, which are denoted by PA and PB . The articial internal leakage is realized through the bypass pipeline between the chamber A and B of working cylinder, as shown in Fig. 2. The leakage ow can be controlled by the proportional direction valve. The dynamics

Fig. 2.

Articial internal leakage of cylinder

of pressures are rewritten as PB PA = = EB (A2 vc QB + QLin ) V20 A2 xc EA (A1 vc + QA QLin ). V10 + A1 xc (10) (11)

The rest of system equations remains the same. The sensor offsets are simply accomplished by addition of a constant after A/D conversion. III. M ODEL - BASED FAULT DIAGNOSIS A. Disturbance decoupling The system (9) is obviously highly nonlinear and subject to the unknown disturbance FL . Therefore, it is reasonable to decouple the disturbance rst for keeping the robustness of the residual generator. Among various robust residual

where Eo is the bulk modulus of oil, Pat is the atmospheric pressure and kair is the air/oil volume ratio.

generation methods the differential-geometric approach is appropriate for the system because it can also be applied to nonlinear systems. The principle of the differential-geometric approach is to nd a distribution, with which a system transformation can be carried out. The transformed system is only affected by the considered faults. Then fault detection observers can be designed for the transformed system [5]. As the introduced faults only affect the cylinder dynamics residual generator is constructed based only on the cylinder model governed by equations (4)-(7). Following the observability codistribution algorithm proposed in [5] the state variables of the transformed system are x = [xc Its dynamics is x1 x2 x3 = = = vc EB (A2 vc QB ) V20 A2 x1 EA (A1 vc + KQ ) V10 + A1 x1 (12) (13) (14) PB PA ]T = [x1 x2 x 3 ]T .

C. Fault isolation Let r = [rPB rPA ]T corresponding to [PA PB ]T . rPB and rPA are both affected by the fault QLin . However the fault PA only affects the residual rPA and PB affects the rPB merely. Then the three faults can be isolated from each other theoretically. In practice the leakage possibly can not make |rPA | and |rPB | larger than respective thresholds synchronously. A proper latency time tw should be set for isolating the faults. For example, if |rPA | exceeds the threshold it can not concluded immediately that the fault PA occurs. The diagnosis system will wait for tw . Then, if |rPB | also exceeds its threshold during the period tw , the fault QLin is occurred. Otherwise the fault PA is present. The same rule is also applied to isolate QLin from PB .

IV. FAULT- TOLERANT CONTROLLER DESIGN The design of the recongurable controller is based on backstepping approach, which can deal with nonlinear systems and achieve regulation and tracking properties without cancellations of stabilizing nonlinearities [11]. The backstepping controller can be easily recongured through parameter adaption or model modication. For the electro-hydraulic testbed adaptive backstepping is applied to compensate for the effect of time-variant external load in fault free case. Accommodation of the fault QLin can be accomplished by another parameter adaption. The sensor faults affect the system through the controller. The accommodation of such faults is usually realized through replacing the faulty measurement by estimation. But our system is not observable without the measurement of PA or PB . In this paper the sensor fault tolerance is achieved by model modication. The FTC structure based on backstepping approach is shown in Fig. 3. The control input un is designed based on the fault free model. If the fault occurs and is identied, recongured control input uf based on the model of faulty system is active. Thereby the control performance can be recovered.
reference values u
 43210 )('&% $   #"!

with input signals and vc . It can be observed that the system with equations (12)-(14) is only affected by the considered faults QLin , PA and PB . Furthermore, as xc is just an integration of a input signal in transformed system, it can reect none of the introduced faults. Thus, the residual generation is only based on the system equations (13)-(14). B. Residual generation Now the fault detection observer can be designed for the system (13)-(14). As all state variables are measurable and the nonlinearities can be compensated a linear observer is employed for fault detection. The structure of the observer is x = A + f (y) + g(y)u + Lo (y y ) x y = C x, (15)

where x is the state vector, u is the input and y is the output. A, f () and g() are the system matrix or vectors. C is the output matrix. Lo is the observer gain and should ensure (A Lo C) stable. x is the estimation of x. Dene the observer error eo = x x. The residual is chosen as r = y y . Then the residual generator can be obtained as eo = (A Lo C)eo r = Ceo .
y

y adap. backstepping

y Electro-hydraulic System

adap. backstepping
  

FDI System r

(16)

In fault free case the residual r satises limt r = 0. If the fault occurs, r will diverge from zero. Then the fault can be detected. It should be noted that for the presented electro-hydraulic system the step-like sensor offsets can be detected with the introduced method. A slowly changing sensor fault, like a drifting, can not be detected due to the system structure.

Fig. 3.

FTC structure of the system

A. Adaptive backstepping The nonlinear system (9) should be reformulated in a strict feedback form for the construction of backstepping controller. Assume a new state variable FHY = PA A1

PB A2 . Then the original system (9) can be rewritten as xc vc FHY = = = v = = vc 1 (FHY Ff FL ) m EA A1 (A1 vc + KQ ) V10 + A1 xc EB A2 (A2 vc QB ) V20 A2 xc v p1 p2 v + p1 kpu uin . (g(xc ) = 0),

Step 3: Dene the Lyapunov function of this step 1 V3 = V2a + 3 e2 3 2 The derivative is (3 > 0). (27)

(17)

V3

= +

1 k1 e2 2 k2 e2 + 1 2

2 e2 e3 m 3 e3 [f (xc , vc ) + g(xc )(4 + e4 )] .

(28)

Because g(xc ) = 0, 4 can be designed as 4 (18) This leads to V3 = 1 k1 e2 2 k2 e2 3 k3 e2 + 3 g(xc )e3 e4 . 1 2 3 Step 4: Dene the Lyapunov function of this step 1 V4 = V3 + 4 e2 4 2 (4 > 0). (31) (30) = 2 1 f (xc , vc ) e2 g(xc ) 3 m (k3 > 0). k 3 e3 + 3

For simplicity the dynamics of FHY is given by FHY = f (xc , vc ) + g(xc )

(29)

where EA and EB can be calculated by equation (8). The controller for fault free case can be derived step by step as follows. Step 1: Dene e1 = x c x r , e4 = 4 , e5 = v 5 , e2 = v c 2 , e3 = FHY 3

where i (i = 2, 3, 4, 5) are the stabilizing functions, xr is the reference position trajectory and is 5th order differentiable. Then the Lyapunov function in this step is 1 1 e2 (1 > 0). 1 2 Then the derivative of V1 is V1 = V1 = 1 e1 e1 = 1 e1 (2 + e2 xr ) . V1 = 1 k1 e2 + 1 e1 e2 . 1 (19)

Choose 5 as 3 5 = g(xc )e3 k4 e4 + 4 4 Then the derivative of V4 is

(k4 > 0).

(32)

V4 = 1 k1 e2 2 k2 e2 3 k3 e2 4 k4 e2 + 4 e4 e5 . (33) 1 2 3 4 (20) Step 5: Dene the Lyapunov function of this step 1 V5 = V4 + 5 e2 5 2 Chose the control input as uin = 1 (p1 + p2 v p1 kpu 4 e5 k 5 e5 + 5 ) 5 (5 > 0). (34)

Choose 2 = k1 e1 + xr , with k1 > 0. This follows that (21) Step 2: The design of stabilizing function in this step requires the information of FL . The parameter adaption should be applied. Dene 1 = FL and the estimation error e1 = 1 1 . The Lyapunov function of this step is (2 , 1 > 0). (22)

(k5 > 0).

(35)

1 1 V2 = V1 + 2 e2 + 1 e21 2 2 2 The derivative of V2 is

Then the derivative of V5 is


5

V2 = 1 k1 e2 + 1 e1 e2 + 1 e1 e1 1 + 2 e2

V5 . (23)

i ki ei 2
i=1

(ki , i > 0).

(36)

Ff 3 + e3 e + 1 + 1 2 m m m 1 me1 mk2 e2 + m2 2

Then 3 is chosen as 3 = Ff + 1 (24)

with the adaption law of 1 2 e2 . 1 = m1 This leads to 2 V2 = 1 k1 e2 2 k2 e2 + e2 e3 . 1 2 m (26) (25)

For the construction of backstepping controller the derivative of the stabilizing function is necessary. Here the numerical differentiations followed by low-pass lters replace the analytic derivatives for simplicity. Following the LaSalleYoshizawa theorem [11] the error term ei will asymptotically converge to zero with the controller (35). Meanwhile the estimation error e1 is bounded. Thus, the desired position tracking can be achieved with good transient performance. Theoretically the asymptotic stability can be ensured by positive ki and i . But every system sate has its physical limitation. Thus, the parameters ki and i should be so chosen that the stabilizing functions are limited in the physical constraints.

B. Fault accommodation Scenario 1: System fault QLin If the internal leakage occurs, the model is different from fault free case as shown in equations (10) and (11). It can be described as FHY = f (xc , vc ) + (xc )2 + g(xc ), (37)

where 2 denotes the leakage QLin . An extra parameter adaption can be utilized to compensate the effect of the leakage, which is same as handling the problem FL . Dene the estimation error e2 = 2 2 . The new Lyapunov function is 1 (38) (2 > 0). V3a = V3 + 2 e22 2 The stabilizing function is chosen as 4 = 1 f (xc , vc ) (xc )2 g(xc ) 2 (k3 > 0), e2 k 3 e3 + 3 3 m

It leads to FHY 3a FHY 3 = e3. The design of 3a is same as 3 . The error term e3 is almost unaffected by PA . The parameter EA , which is also affected by the PA , is replaced by a constant. The calculation of EA can not be accurate when the value of PA is wrong. Furthermore, if the pressure is high enough, then the bulk modulus can be viewed as a constant. Thus, the reconguration of controller with respect to PA is accomplished by setting the bulk modulus EA as constant. Scenario 3: Sensor fault PB In principle dealing with PB is similar to PA . However compared with the external load and PA the pressure PB is much smaller. PB can be viewed as a part of external load and compensated by parameter adaption of FL . Therefore the accommodation of PB is simply realized by assuming PB is zero. V. E XPERIMENTAL RESULTS

(39)

with the adaption law of 2 3 (xc ) 2 = e3 . 2 g(xc ) (40)

This leads to V3a = V3 . The rest of the design is the same as in the no fault case. The asymptotic convergence of the position tracking error can be guaranteed even with the internal leakage in system. Scenario 2: Sensor fault PA The Sensor fault PA can bring more serious consequence than QLin , especially if the system is running under high pressure. Fortunately, the reconguration can be easily accomplished thanks to the structure of adaptive backstepping. PA affects two parameters of the backstepping controller: error term e3 and the bulk modulus EA . Firstly, it can be proved that the sensor offset is estimated by the adaption law of FL (1 ). The faulty measurement can HY = FHY + FHY , where FHY = be modeled as F PA A1 . Assume there exist a stabilizing function 3a , which satises FHY 3a = FHY 3 = e3 . Then FHY = 3a + e3 FHY . Substitute (41) into equation (23). Then Ff V2 = 1 k1 e2 + 1 e1 e2 + 1 e1 e1 + 2 e2 1 m 3a + e3 1 + FHY + (42) 2 . m m FHY is a step-like signal and its derivative can be viewed as zero. So the term (1 + FHY ) can be viewed as one unknown constant and estimated by the same adaption law of 1 . Also 3a can be obtained as 1 3a = Ff + 1 + FHY me1 mk2 e2 + m2 . (43) 2 (41)

The proposed fault-tolerant control is implemented on the electro-hydraulic testbed and tested with the considered faults. For proof of the robustness of the residual generator the external load varies slowly from 1.5 Bar to 120 Bar. The sampling frequency of the experiments is set to 200 Hz. The initial errors have signicant effect to the fault detection observer. Therefore, the fault diagnosis is activated as t = 2 second. The QLin is designed as a slop-like signal and saturated with 5 L/min. The sensor offset is 20% of the measurement. The latency time tw is 0.5 second. The choice of tw and the vaules of thresholds is a compromis between the robustness and sensitivity of the diagnosis system. Here they are determined conservativly based on the test in fault free case to avoide false alarms.

0.8

0.4 |rPA | threshold |rPB | threshold 0.3

0.6

0.4

0.2

0.2

0.1

0 0

20

40

60

Time (sec.)

0 0

20 40 Time (sec.)

60

Fig. 4.

Residuals in fault free case

Velocity (mm/sec.)

300 Position (mm) x 200 100 0 0


2
r c

15 v 10 5 0 0 2
r

vc

20

40

60

20

40

60

ev (mm/sec.)
10 20 30 40 50 60

e (mm)

1 0 1 2 0

1 0 1 2 0 20 40 Time (sec.) 60

Time (sec.)

Fig. 5.

Position tracking in fault free case

1 1.6 0.8 1.2 0.6

0,4

0.6

0,3 |r | |r | 20 40 Time (sec.) 60 0.4


B

0.8 0.4

0,2

0.2
0.4 |rPA | threshold 0 0 20 40 Time (sec.) 60 0 0 20 40 Time (sec.)

0,1
0.2 |rPB | threshold 60

0 0

0 0

20 40 Time (sec.)

60

Fig. 6.

Residuals with the fault QLin

Fig. 10.

Residuals with the fault PB

Velocity (mm/sec.)

300 Position (mm) x 200 100 0 0 2 e (mm) x


r c

15 v 10 5 0 0 2 v
r c

Velocity (mm/sec.)

300 Position (mm) x 200 100 0 0 1 e (mm) 0 1 2 3 0 20 40 Time (sec.) 60 x


r c

15 v 10 5 v
r c

20

40

60

20

40

60

20

40

60

20

40

60

0 2 4 0

e (mm/sec.) 20 40 Time (sec.) 60

0 0 2

0 2 4 0

e (mm/sec.)

1 0 1 2 0 20 40 Time (sec.) 60

20 40 Time (sec.)

60

Fig. 11.

Position tracking with the fault PB

Fig. 7.

Position tracking with the fault QLin

1.5 0.4 1.2 |rPA | threshold 0.3 0.9 |rPB | threshold

0.2 0.6

0.3

0.1

0 0

20 Time (sec.)

40

60

0 0

20 Time (sec.)

40

60

Fig. 8.

Residuals with the fault PA

Velocity (mm/sec.)

300 Position (mm) x 200 100 0 0 2 e (mm) 1 0 1 2 0 20 40 Time (sec.) 60 x


r c

15 vr 10 5 0 0 2 vc

20

40

60

20

40

60

Due to the limited space only the residuals and control errors are shown here. The velocity trajectories are also displayed for illustrating the transient performance. xr and vr denote the reference trajectories of cylinder position and velocity respectively. In fault free case the residuals are robust to the system disturbance and varies under the threshold, as shown in Fig. 4. The position tracking is successfully achieved. It can be observed that the residuals deviate from zero even through there is no faults. The main reason is that the parameter EA and EB can not be accurately approximated, especially in low pressure area. If the faults are present, they can be detected and isolated correctly by the diagnosis system, which can be observed in Fig. 6, Fig. 8 and Fig. 10. As soon as the fault is detected and isolated the corresponding reconguration algorithm is activated. The position error converges to zero quickly as shown in Fig. 7, Fig. 9 and Fig. 11. It is a little difcult to evaluate the test results in the plots. The average errors of position ex and velocity ev are introduced for comparison. They are summarized in Table I. The best control performance is archived in fault free case. The performance has only a very small degradation in the presence of faults.
TABLE I T HE EXPERIMENTAL RESULTS

e (mm/sec.)

1 0 1 2 0 20 40 Time (sec.) 60

Fig. 9.

Position tracking with the fault PA

ex (mm) ev (mm/sec.)

No fault 0.38 0.25

QLin 0.41 0.27

PA 0.49 0.27

PB 0.43 0.28

VI. C ONCLUSIONS This paper introduces a fault diagnosis integrated faulttolerant control for a nonlinear electro-hydraulic system. High control performance with improved reliability and safety can be realized with the proposed control strategy. Two kinds of faults, cylinder leakage and sensor offsets, can be successfully detected, isolated and accommodated. The control performance has only a slight depredation in the presence of faults. The control scheme can be implemented for real-time application. R EFERENCES
[1] M. Jelali and A. Kroll Hydraulic servo-systems: Modelling, Identication and Control, Springer, 2004. [2] S. Gayaka and B. Yao, Fault detection, identication and accommodation for an electro-hydraulic system: An adaptive robust approach, Proceedings of the 17th IFAC World Congress, 2008, pp 13815-13820. [3] M. Sepasi and F. Sassani, On-line fault diagnosis of hydraulic systems using Unscented Kalman Filter, International Journal of Control, Automation and Systems, 2010, pp 149-156. [4] Z. Shi, F. Gu, B. Lennox, and A. D. Ball, The development of an adaptive threshold for model-based fault detection of a nonlinear electro-hydraulic system, Control Engineering Practice, 2005, pp 1357-1367. [5] C. DePersis and A. Isidori, A geometric approach to nonlinear fault detection and isolation, Proceeding of SAFEPROCESS, 2000, pp 209214. [6] M. Blanke, M. Kinnaert, J. Lunze, and M. Staroswiecki, Diagnosis and Fault-Tolerant Control, Springer, 2003. [7] H. Niemann and J. Stoustrup, Passive fault tolerant control of a double inverted pendulum: a case study example, Proceedings of SAFEPROCESS, 2003, pp 1029-1034. [8] S. Kanev and M. Verhaegen, A bank of recongurable LQG controllers for linear systems subjected to failures, Proceedings of the 39th IEEE Conference on Decision and Control, 2000. [9] J. Maciejowski and C. Jones. MPC fault-tolerant ight control case study: Flight 1862, Proceedings of SAFEPROCESS, 2003, pp 121-126. [10] E. Kececi, X. Tang, and G. Tao. Adaptive actuator failure compensation for cooperating multiple manipulator systems, Proceedings of the 5th SAFEPROCESS, 2003, pp 417-422. [11] M. Krstic, I. Kanellakopoulos, and P. Kokotovic, Nonlinear and Adaptive Control Design, John Wiley and Sons, 1995.

Vous aimerez peut-être aussi