Sunteți pe pagina 1din 6

HSI 2008

Krakow, Poland, May 25-27, 2008

Fault Detection and Diagnosis (FDD) in Heating Ventilation Air Conditioning Systems (HVAC) Using an Interactive Multiple Model Augmented Unscented Kalman Filter (IMMAUKF)
Nicolae Tudoroiu, Mohamed Zaheeruddin, Elena-Roxana Tudoroiu , and Victor Jeflea Building Civil and Environmental Engineering, Concordia University, Montreal, Canada West University Politehnica Timisoara, Timisoara, Romania Spiru Haret University Constanta, Timisoara, Romania
tnicolae@excite.com, zaheer@encs.concordia.ca, tudelena@excite.com, jefleavf@yahoo.com

Abstract-Nowadays monitoring and controlling the modern and sophisticated Heating Ventilation Air Conditioning (HVAC) building systems under a wide variety of occupancy and load related operating conditions is becoming a difficult and challenging task. Their complexity drastically increases and the control becomes more difficult task due to the several control loops that interact between them. Among these control loops the discharge air temperature (DAT) loops, the static pressure loop (SP), and the Variable Air Volume (VAV) terminal unit loop are the candidate loops requiring frequent re-tuning. Equipment failures and loss of control leading to less than acceptable indoor environment conditions is a common problem reported in these systems. In our paper we consider the degradation in the DAT loop performance caused by a gradual increase in backlash of the valve actuator and several disturbances that occur in the process. The main objective of this paper is to describe the application of an interactive multiple model (IMM) [9] based on an augmented Unscented Kalman Filter (UKF) estimation algorithm (IMMAUKF) [2] to the problem of fault detection diagnosis and isolation (FDDI) of the valve actuator failures in DAT loop of the HVAC systems [4]. The proposed algorithm is an alternative to the interactive multiple model (IMM) developed in the literature based on the Extended Kalman Filter standard technique [2], [3][5], [7]-[11], the most popular estimation technique used in the last 40 years. The main advantage of the proposed algorithm is the less computation, consequently faster, high accuracy, robustness and eliminates completely the linearization of the system dynamics.

severe equipment operating problems. To increase the energy efficiency and equipment reliability, there is a need to develop robust fault detection and diagnosis tools capable of detecting and isolating any sensor, actuator or system component faults so that remedial actions could be taken as soon as possible. The proposed algorithm deals directly with the nonlinear dynamics of the system, the unanticipated valve actuator failures and statistic properties of the process noise and measurements such as the mean and covariance matrices in the most general formulation based on the Unscented Kalman filter estimation theory. The FDDI based on the IMM algorithm is implemented in a simulation environment, and the fault diagnosis results are presented for several fault scenarios in terms of mode probabilities and mode probability indices. From the preliminary simulations [3]-[5] for different scenarios we found that the IMM algorithm is robust to the choice of the matrix probability and to the small changes in process and measurement noise level, a result that is also confirmed by [6]. As tools we use MATLAB with SIMULINK toolboxes from control systems, system identification and signal processing.
II. THE DISCHARGE AIR TEMPERATURE SYSTEM

Keywords - Fault Diagnosis and Detection (FDD), HVAC Systems, Interactive Multiple Model (IMM) , System Estimation Unscented Kalman Filter (UKF).
I. INTRODUCTION

A. Physical System Figure 1 shows a schematic diagram of a DAT loop of a HVAC system. A central air supply provides air at a controlled temperature and flow rate for use in heating coil, Ta 0 and Q respectively. A heating coil interposed in the central air supply space heats the discharged air using hot water. The temperature of the air leaving the heating coil Ta is controlled by regulating the rate at which the hot water flows through the heating coil as shown by the feedback control loop. The discharged air flow rate is regulated to maintain a

In the most HVAC systems several failures caused by malfunctions in components, actuators, and sensors can be as result of unexpected interference, or gradual aging of system components. Consequently these failures have a great impact on the energy consumption, thermal comfort and generate

1-4244-1543-8/08/$25.00 2008 IEEE

predetermined static air pressure within the temperature controlled space. In Figure 1, u the hot water flow rate, is the DAT loop subsystem input and temperature

Ta , is the DAT output. The entering air

Ta 0 is considered as a disturbance on the system.

leaving the heating coil, Ta [oC], y(k). Hot water supply temperature Tw[oC] and the average tube temperature Tt [oC] act as disturbances on the DAT loop of HVAC system. Also such as input disturbances could be considered & ma , Ta ,in , and Tw,in . In practice the hot water flow rate is regulated such that the temperature of air leaving the heating coil is maintained close to a chosen set-point. In this representation the heating coil output temperature set point is more directly related to the overall DAT subsystem performance, and the model uncertainties are due to variable heat and mass transfer rates. C. Discrete time DAT System Model The discrete time space system representation version of the DAT subsystem (1)-(7) has the following form: x1 (k + 1) = x1 (k ) + c1ha (k )ov (k )Ts ( x1 (k ) x3 (k )) + (8) c2Ts ( x1 (k ) p1 (k )) p2 (k ) x2 (k + 1) = x2 (k ) + c3 hw (k )Ts ( x3 (k ) x2 (k )) + c4Ts ( x2 (k ) p3 (k ))u (k ) x3 (k + 1) = x3 (k ) + 1 s Ts [c6u( x1 (k ) p2 (k )) p1 (k ) + s + c5 10) (9)

B. DAT System Control Oriented Structure The DAT system was modeled as a SISO system as shown by the block diagram in Fig.2. The model is nonlinear and described by he following equations: dTa & = c1haov (Ta Tt ) + c2 ma (Ta Ta ,in ) (1) dt dTw & = c3 hw (Tt Tw ) + c4 mw (Tw Tw,in ) (2) dt dTt 1 s c & = [c6 ma (Ta Ta ,in ) + 7 hw (Tt Tw ) + dt s + c5 1 s (3) c9 + (c8 + )haov (Ta Tt )] 1 s For the simulations we use the following values for the coefficients: c1 = 1.2046, c2 = 38.9034, c3 = 4.5714 105 , c4 = 0.2023,

c5 = 1.1192, c6 = 38.9034, c7 = 1.4926 10 , c8 = 1.2046, c9 = 4.8995 104 , and the air and water heat transfer coefficients ha , hw , the sensible efficiency quadratic functions s ,ov are given by: & & ha = 17.58ma + 70.562ma + 8.1796 ,
2

c7 hw (k )( x3 (k ) x2 (k )) p2 (k ) + 1 s (c8 + c9 )ha (k )ov (k )( x1 (k ) x3 (k ))] 1 s

(4) (5) (6) (7)

& hw = 1403.2(mw ) &2 0.16375m a

0.8

& s = 0.39483ma + 0.92805 , ov = 1 0.9(1 s ) .


& and the control input mw , are given by:

y (k ) = x1 (k ) (11) where x(k ) = [ x1 (k ) x2 (k ) x3 (k )]T = [Ta (k ) Tw (k ) Tt (k )]T represent & s the state space vector, u (k ) = mw (k ) is the DAT subsystem input, y (k ) = [Ta (k )] is the DAT subsystem output, and & p(k ) = [ p1 (k ) p2 (k ) p3 (k )]T = [Ta ,in (k ) ma (k ) Tw,in (k ) ]T represents the disturbance input vector. In our simulations we consider the sampling time Ts =4 seconds. The sensible efficiency quadratic functions s ,ov and the (12) (13) (14) (15) time variable water heat transfer coefficients ha , hw , become:

& The variation ranges for the disturbance inputs ma , Ta ,in , Tw,in ,

kg & ma [0.1, 0.8] , Ta ,in [16, 22][0 C ], Tw,in [35,50][0 C ] s The variables that appear in this DAT model have the following significance: Ta [0 C ] -the air temperature leaving the heating coil, Tw [0 C ] the hot water supply temperature, with their average values Ta + Ta ,in Tw + Tw,in Ta = , Tw = , Tt [0 C ] - the given by 2 2 average tube temperature, Ta ,in [0 C ] -the inlet air temperature leaving the heating coil, T w,in [0 C ] -the inlet water supply kg kg & & temperature, ma -the input air flow rate, and mw s s the water flow rate. The input of the simplified SISO model is hot water & flow rate, mw , u(k), and the output is the temperature of air

s (k ) = 0.16375 p2 2 (k ) 0.39483 p2 (k ) + 0.92805 ov (k ) = 1 0.9(1 s (k ))


ha (k ) = 17.58 p2 2 (k ) + 70.562 p2 (k ) + 8.1796 hw (k ) = 1403.2(u (k )) x1 (k ) =
0.8

The average values of the states x1 (k ) and x2 (k ) are given by x2 (k ) + x2,in (k ) , x2 ( k ) = 2 2 where x1,in (k ) = Ta ,in (k ), and x2,in (k ) = Tw,in (k ) . x1 (k ) + x1,in (k ) (16)

The dynamics of the valve actuator is integrated in the plant dynamics, and will be represented by a nonlinear block with backlash characteristic of dead zone width =2r. The input-

335

output behavior of the backlash nonlinearity can be described by two modes of operation, as either tracking or in the dead zone:
Tracking described by the following differential equation:
& & w= y with the solution given by:

(17)

dy >0, and (18) dx dy w = ( y + r ) for <0 (19) dx Dead zone described by the following differential equation: w = ( y r ) for
& (20) w = 0 | w y | r where r is the dead zone width and is the slope of tracking region such as depicted in the Fig.3, with y = ( x) , .and

through the system dynamics. Using the principle that a minimal set of carefully chosen sample points can be used to parameterize mean and covariance, the UKF estimator yields superior performance compare to EKF estimator, specially for nonlinear systems. These sample points completely capture the true mean and covariance of the Gaussian random state variable x k R n , and are propagated through the true nonlinear system dynamics. The main advantage of the UKF estimator is that does not require the calculation of the Jacobean matrices that could lead to implementation difficulties. B. UKF Architecture In our research we developed an augmented UKF architecture [2], [9]. The algorithm is developed to compute a collection of sigma points corresponding to the state vector x k or noise signals w k and v k stored in the column of the L (2 L + 1) sigma point matrix X k 1 , where L = dim( X ) + dim(W ) + dim(V ) . C. UKF Algorithm procedures/steps [ 1]-[3],{6] Step 1. Initialization The Gaussian random variable x is augmented with the process noise w and measurement noise v , according to

w(t ) = (0 , x([0, t )) = [ x, 0 ](t ) , t [0, T ] . Unlike memory less nonlinearities, hysteresis output at any given time is function of the entire history of the input, and the initial condition of the output, 0 .
III. UNSCENTED KALMAN FILTER APPROACH

A. Unscented Kalman Filter (UKF) Methodology The Unscented Kalman Filter (UKF) technique is based on the unscented transformation (UT) [1]-[2],[7], [9] and addresses the general problem of state estimation x k R n of a discrete-time controlled process that is governed by the nonlinear stochastic difference state-space equation with a measurement y k R m that is given by the output equation

xa = [xT w T v T ]T . The new augmented Gaussian random


state variable xa is initialized with the mean T a = E[xa ] = [x0 0 0]T and the covariance matrix as follows: x0

a P0

E[(xa 0

0 0 - xa )(xa

0 - x a )T ] =

x k +1 = f (x k , u k ) + w k y k = g (xk , u k ) + v k

(21) (22)

P0 0 0

0 Qw 0

0 0 Rv

(26)

The random variables w k and v k representing the process and measurement noise injected to drive the system dynamics are generated by the additive noise sources. They are assumed to be independent, zero mean, white, and with normal probability distributions: p(w ) ~ N (0, Q w ) (23) p( v) ~ N (0, R v ) (24) The covariance matrices Q w (process noise covariance) and

where x0 = E[x0 ] P0 = E[(x0 - x0 )(x0 - x0 ) ]


T

(27) (28)

Step 2. Sigma Points Matrix Computation The set of sigma points are created for the augmented state vector using the following relationships:
+ k X a -1 = [xa -1 Z k -1 Z- -1 ] for k {1,...., } k k

R v (measurement noise covariance) might change with each time step or measurement, but in our approach we assume they are constant. Due to the process noise injected in the state space equation the state vector x k R n becomes random variable with its distribution approximated by a Gaussian distribution function: p(x) ~ G (x, Px ) (25) The vital operation performed in the Kalman Filter is the propagation of a Gaussian random state variable x k R n

(29)

a k Z+-1 =xa-1[1,1,1......1]+ L+ Pk-1 k 123 44 dim(X) a k Z - -1 = x a -1[1,1,1.....1] L + P -1 k 1 24 4 3 dim( X )

(30)

Step 3. Prediction Step (Time Update)

336

Once the augmented sigma points matrix X a 1 is computed, k the prediction step is performed by first propagating each column of this matrix through the nonlinear function f (., uk ,.) and obtain the transformed sigma points matrix for the state variable x according to: Xx k k 1
x X k k 1

Step 4. Correction Step (Measurement Update) Once the new measurement state estimate

y k is available the a posteriori

x k and covariance Pk are computed as follows


k x k = x + K G ( yk yk ) Pk = Pk K G Pyk yk K T G (40) (41)

= f

x w ( X k 1 , u k , X k 1 )

(31)

where

X w1 k

represents the set of sigma points corresponding


where the Kalman gain KG and covariance and cross-

k to the process noise vector w k .The predicted priori mean x and the predicted a priori covariance matrix as weighted averages given by
xk =

Pk

are computed

covariance matrices Pxk yk , Pyk yk , respectively, are defined by K G = Px y (Py y )1 k k k k (c) 2 = (W0 + 1 )(Y0,k k 1 yk )(Y0, k k 1 yk )T + (42)

W
i =0

2n

(m)

Xix,k k 1

(32)

Pyk yk

Pk

= (W0
2n

(c)

+1

x )( X0, k k 1

x yk )( X0, k k 1

yk )T

+
(33)

W
i =1

2n

(c )

[(Yi ,k k 1 yk )(Yi ,k k 1 yk )T ] + Rv

(43)

i =1

Wi(c ) [( Xix, k k 1 xk )( Xix, k k 1 x k )T ] + Q w

) x Pxk yk = (W0( c ) + 1 2 )( X0,k k 1 x )(Y0, k k 1 yk )T + k

where the weights are defined deterministically by


( W0 m) =

W
i =1

2n

(c)

k [( Xix,k | k 1 x )(Yi ,k k 1 yk )T ]

(44)

( L + )
2

+1

(34)

Step 5. Parameter selection/settings


= 0.001 ,
R v = 1 10
2

k=0

= 2,

P0 = 1.5 10

I3 ,

Q w = 2 10 I 3 ,

Wi

( m)

0.5 ( L + ) 2

.
IV. FAULT DETECTION AND DIAGNOSIS METHODOLOGY

, i = 1,.....2n

(35)

( W0 c ) = W0( m) + (1 2 + )

(36)

Wi( c ) =

0.5 ( L + ) 2

, i = 1,.....2n

(37)

Furthermore, the columns of the transformed sigma point matrix X a k 1 will be propagated through the nonlinear k function

g (., uk ,.) to yield

In this paper we assume a DAT subsystem having one single actuator valve and three potential disturbances that could generate faults. Consequently there are only five modes, first mode will be normal mode corresponding to the healthy system, labeled by the index j = 1, and the last four modes will be the faults modes, labeled by the indexes j = 2, 3, 4, 5. The root cause of these fault modes is the valve backlash described by the relations (17)-(20), and the disturbances p1 = Ta ,in , p2 = ma , p3 = Tw,in that appear explicitly in the model equations (8)-(10). Theoretical aspects and computation steps in this algorithm are well developed and documented in [1]-[3], [7]. According to the different scenarios only three possible approaches could be developed: soft, hard and random decisions. The interactive estimation algorithm runs each of the four parallel filters bank corresponding to the normal and the fault modes (j =1, 2, 3, 4) only once in each cycle. Each of these filters at time tk+1 = k+1 has its own input, the estimate state at the time t , x0 (k | k ) , and its covariance matrix,
k

Y = g (X x , u , Xv ) (38) k k 1 k k 1 k k 1 where Xv 1 represents the sigma points matrix corresponding k


to the measurement noise v k . The a priori output mean now given by

y k is

y = Wi Yi ,k k 1
k (m) i =0

2n

(39)

P0 (k | k ) , which form a good quasi-sufficient statistic of all old information and the assumption that the model of each filter matches the system mode. FDDI assumes that the actual

337

system at any time can be modeled sufficiently accurately by the following jump Markov hybrid linear system:
x(k + 1) = F (k , m(k + 1)) x(k ) + G(k , m(k + 1))u(k ) + T (k , m(k + 1)) (k, m(k + 1)),

(45)
(46) (47)

x(0)

) N ( x0 , P0 )

z (k ) = H (k , m(k )) x(k ) + (k , m(k ))

The mode (normal or fault mode) of the system at time k is selected by a discrete process m j and modeled as a discrete time, s-state, first-order Markov chain with transition probabilities ij (k ) given by:
ij ( k ) = P{m j ( k + 1) | mi ( k )}, mi , m j S

Prediction step corresponding to the steps 1-3, relations (26)-(39), of the augmented UKF algorithm, developed in Section 3. (ii) Correction step, corresponding to the step 4, relations (40)-(44) of the augmented UKF algorithm described in Section 3. Step3. Update the mode probability at k+1, (k + 1) , [1], [2], [7] 3.1 Compute the likelihood function at k+1, that is L j (k + 1) :

(i)

(48) (49)

) Lj (k +1 =

1 exp[ jT(k +1 P,yy(k+1| k))1j (k+1 )( j %% )], 2 |(2 )P,yy(k +1| k)| j %% (55)

and
0 ij ( k ) 1, i = 1, N , j = 1, N ,
ij ( k ) = 1, i = 1, N

, j =1 2,........., N 3.2 Update the mode probability at k+1, that is j (k + 1)


j ( k + 1 | k ) L j ( k + 1) j ( k + 1) = N , j ( k + 1 | k ) L j ( k + 1)
1

The initial state distribution of the Markov chain is (0) = [ 1 2 3 ..... N ] , where
N 0 j 1, j = 1, N , j = 1 j =1

(56)

(50)

where x(k ) is the state vector, z (k ) is the mode-dependent measurement vector, u (k ) is control input vector. The process noise and noise measurement vectors (k ) and (k ) are mutually independent, additive, white Gaussian noises of zero mean and covariance matrices Q(k ) and, R (k ) , and are independent of the initial state x(0) . In the relation (48) P{} denotes the probability operator. The event that m j is in effect at time
k

j = 1, 2 , .........., N Step 4. Fault detection 4.1 Compute the mode probability vector at r is ( k + 1) :
r ( k + 1) = 1 ( k + 1) 2 ( k + 1) 3 ( k + 1)..... N ( k + 1)

k+1, that (57)

4.2 Obtain the maximum value of the mode probability vector components: = (58) max{ j ( k + 1)) , for j =1, 2, ..., N FDD max j 4.3. Determine the index of the maximum value of the mode probability vector component r r j = find ( == max( )) (59) 4.4 Fault decision FDD logic [1]-[3],[6]: if FDD max > Threshold then the fault occurs otherwise no fault occurs
V. SIMULATION RESULTS

is denoted as

m j (k ) = {m(k ) = m j } , and

S = {m1 , m2 ,..., mN } represents the set of all possible system modes. The system modes may jump from one to another at a random time. The interactive multi-model algorithm is a recursive estimator that consists in each cycle of four steps [1][3],[7],[9]:

Step1. Model-conditional reinitialization (interacting or mixing the estimates): 1.1 Compute the predicted mode probability from k to k+1, that is (k + 1| k ) :
j ( k + 1 | k ) = 1 (i , j ) i ( k ), j = 1, 2, ..., N N

(60)

(51)

1.2 Compute the mixing probability at k,


ij ( k ) = (i , j ) i ( k )
j (k + 1 | k ) , i , j = 1, 2, ...., N

ij (k ) :
(52)

1.2 Compute the mixing estimates at k, that is x(k | k ) :


x j 0 ( k | k ) = xi ( k | k ) ij ( k ), i , j = 1, 2,...., N

(53)

1.4 Compute the mixing covariance from at k:


T Pj0(k | k) = [P (k | k) +(xj0(k | k) xi (k))(xj0(k | k) xi (k | k)) ]ij (k), i j =1, 2,...., N

(54)

Step2. Model conditional filtering [1]-[3], [6]

We consider in our paper the following scenario: z = [z1 z3 z2 z5 z2 z5], u = [u1 u3 u2 u5 u2 u5 ] , when the first fault mode occurs inside the time horizon windows [70, 150], [200, 250], the second fault mode inside the time horizon windows [30, 70], the fourth fault mode inside the time horizon windows [150, 200], [250, 300]. The results obtained in this paper are very similar to those obtained for the linear case in [4] using the standard Extended Kalman Filter (EKF). The simulation results are presented in Fig.4-5. Due to the linear approximation of the DAT system the results obtained could degrade in accuracy. However the gain of the developed algorithm is for nonlinear cases, but in many cases it is very difficult to obtain the analytical model of the nonlinear system. However the superiority of the IMM and

338

UKF algorithms is well documented in literature [2], [6], and in our research in [3]-[5]. VI. CONCLUSIONS
entering air temperature u

controller

In this paper, we develop an interactive multi-model approach first to detect the fault valve actuator and the fault modes generated by the possible perturbations in the HVAC systems. The approach is probabilistic and gives results more accurate than the spectral analysis very well spread in the literature. The results obtained are encouraging and the performance of the IMMAUKF algorithm will be tested. The recovery mode development to complete the IMM algorithm in order to avoid the degrading in performance of the closed-loop DAT subsystem during the failures periods will be explored in the future work.
REFERENCES

Ta0 Heating Coil

Ta=y discharge air temperature

Fig.1. Schematic diagram of DAT system u= mw DAT System y=Ta

Fig.2.The block scheme of the DAT system

[1].

[2]. [3].

[4].

[5].

[6].

[7].

[8].

[9].

[10].

[11].

S.J. Julier, J.K. Uhlman, A New Extension of the Kalman Filter to Nonlinear Systems, Proceedings of Aero Sense, The 11th International Symposium on Aerospace/Defense Sensing, Simulation and Controls, 1997, pp.182-193. S.Haykin, Adaptive filter theory - Fourth edition, Prentice Hall, 2003. N. Tudoroiu, M. Zaheerudin , Fault Detection and Diagnosis of the Valve Actuators in HVAC Systems Using Interactive Parallel Kalman Filters Bank, The 29th Annual Congress of the American Romanian Academy of Arts and Sciences (ARA), Proceedings, Bochum, Germany, , 2004, pp 395-398, September 7-12. N.Tudoroiu, M.Zaheeruddin, Fault Detection and Diagnosis of Valve Actuators in HVAC Systems, Proceeding of the 2005 IEEE International Conference on Control Applications, August 29-31, 2005, Toronto, Canada, CD-ROM , ISBN:07803-9354-6. N.Tudoroiu, M.Zaheeruddin , Fault Detection and Diagnosis of Valve Actuators in Discharge Air Temperature (DAT) Systems, using Interactive Unscented Kalman Filter Estimation, International Symposium on Industrial Electronics , ISIE2006, Montreal., Canada, 9-13 July 2006. Y. Zhang, Rong-Li Xiao , Detection and Diagnosis of Sensor and Actuator Failures using IMM Estimator, IEEE Transaction on Aerospace and Electronic Systems, Vol.34, No.4, pp. 1293-1311, October 1998. J. Schein, S. T. Bushby, A Hierarchical Rule-Based Fault Detection and Diagnostic Method for HVAC Systems, ASHRAE,HVAC& RESEARCH, Vol.12, No.1, January 2006. Y. Chetovani, Fault Detection in a Chemical Reactor by Using the Standardized Innovation, IChemE, vol. 84, Issue B1, pp.27-32, January 2006. C.H.Lo, P.T.Chan, Y.K.Wong, A.B.Rad,K.L.Cheug, Fuzzygenetic Algorithm for Automatic Fault Detection in HVAC systems, Applied Soft Computing, Vol 7, Issue 2, pp.554-560, March 2007. Zhirmin Du, Xin Qiao Jin, Detection and Diagnosis for Sensor Fault in HVAC Systems, Elsevier, Energy Conversion and Management, Vol.48, No.3, pp. 693-702, 2007. D. Nikovski, A. Divakaran, Radhakrishnan Regunathan, Peker, Kadir , Detecting and Diagnosing Faults in HVAC equipment, USA Patent 20080033674, Mitsubishi Electric Research Laboratories, Inc., February 2008.

Fig.3.The backlash nonlinearity (the dead zone width =2r)


The mode probability 0.3 0.28 0.26 0.24 0.22 u i m miu3 0.2 miu4 0.18 0.16 0.14 0.12

miu1

miu2

miu5

50

100

150 samples

200

250

300

Fig.4. The evolution of the mode probability for the sequence


z = [ z1 z3 z2 z5 z5 z 2 z5 ], u = [u1 u3 u 2 u3 u5 u 2 u5 ]
(d)-Robustness to PImatrix 5 4.5 4 3.5 x e d n i y t i l i b a b o r p e d o m 3 2.5 2 1.5 1 0.5 0

50

100

150 samples

200

250

300

Fig.5.The index of the mode probability for the sequence z = [ z1 z3 z2 z5 z2 z5], u = [u1 u3 u2 u5 u2 u5 ]

339

S-ar putea să vă placă și