Online transition matrix identication of the state evolution model for the extended Kalman lter in electrical impedance

tomography
Fernando S Moura, Julio C C Aya, Agenor T Fleury and Raul G Lima
Escola Politcnica of University of So Paulo, Department of Mechanical Engineering, Brazil E-mail: fernandosmoura@gmail.com, jccaya@gmail.com, lima.raul@gmail.com

Abstract. One of the electrical impedance tomography objectives is to estimate the electrical
resistivity distribution in a domain based only on contour electrical potential measurements caused by an imposed electrical current distribution into the boundary. In biomedical applications, the random walk model is frequently used as evolution model and, under this conditions, poor tracking ability of the Extended Kalman Filter is observed. An analytically developed evolution model is not feasible at this moment. The present work investigates the possibility of identifying the evolution model in parallel to the EKF and updating the evolution model with certain periodicity. The evolution model is identied using the history of resistivity distribution obtained by a sensitivity matrix based algorithm. To numerically identify the linear evolution model, the Ibrahim Time Domain Method is used, which is normally employed to identify the transition matrix on structural dynamics. The investigation was performed by numerical simulations of a domain with time varying resistivity and addition of noise. Numerical diculties to compute the transition matrix were solved using a Tikhonov regularization. The EKF numerical simulations suggest that the tracking ability is signicantly improved.

1. Introduction

Electrical Impedance Tomography (EIT) is an imaging method of the resistivity distribution within a domain. The images are estimated from a set of electrical potential measurements on the boundary of a domain. This set of electrical potentials are obtained by applying current through electrodes and measuring the resulting electrical potentials on electrodes. Electrical Impedance Tomography have a wide range of applications [1]. It may be used to monitor cardiac function, detect internal hemorrhage and breast cancer, for instance. Examples of non-clinical applications are the visualization of multiphase ow, detection of minerals in the soil, soil pollution monitoring and crack detection on mechanical components. The application in mind of the present work is lung monitoring. Lung condition may change dramatically within few respiratory cycles [2]. Furthermore, inadequate lung ventilation in Intensive Care Units may cause barothrauma or hypoxia. The EIT problem can be seen as a state observation problem within the control theory framework. Among several absolute EIT algorithms, the Extended Kalman Filters (EKF) have been investigated due to its ability to track the state of time varying non-linear systems. The EKF is a probabilistic estimation algorithm in the sense that it minimizes the variance of the estimation error. Good references on the Kalman Filters are [3] and [4].

The rst work that solves the EIT problem considering temporal variations of resistivity is [5]. In this work, the authors implemented the linearized Kalman Filter to estimate the state, the resistivity distribution of a circular domain. They argue that any attempt to use static algorithms to solve dynamic problems ends up using fewer measurements, which decreases spatial resolution, or ends up using a faster rate of measurements, which raises noise levels in the measurements. In [6], the smoothing extension of the Kalman lter is implemented for estimating time-varying impedance distributions. They also discuss how to introduce prior spatial information in the state space formulation. The authors argue that without the need of online estimations, the lter can use past and future measurements. The resistivity resolution decreases if the choice of the reference state for the linearization, in the linearized Kalman Filter, is far from the real state. To minimize this problem, in [7] the Extended Kalman Filter (EKF) was implemented. The advantage of EKF is that the linearization is performed at every step, using the last state estimation as reference. One disadvantage of the EKF is the increase of computational time. The Tikhonov regularization was used. They show that EKF has better spatial resolution and smaller position errors of objects than the linearized version. In [8], prior information about subregions with known resistivity is included. The results show that the prior information improved tracking ability of EKF. In [2] and [9], the electrode contact resistivities and the domain resistivity distribution were estimated using the EKF in two phases. In the rst phase skin/electrode contact impedance is estimated while the domain resistivity is not updated. In the second phase the domain resistivity is estimated while the skin/electrode contact impedances are not updated. The tracking ability was improved since each EKF was adjusted independently. The EKF requires a state evolution model. All of the above cited works adopted the Random Walk for the state evolution model. However poor tracking ability of the Kalman lters resulted, for lung monitoring applications. In [10], a procedure, called Interacting Multiple Model, with three EKFs running in parallel were implemented. Each EKF was running with dierent evolution models, namely, Random Walk, First Order and Second Order. The three resulting images are wheighted at the end of the iteration according to a probabilistic criterion. The present work investigates whether the tracking ability of EKF is improved if the evolution model is frequently estimated using images obtained by a sensitivity matrix algorithm. Once a set of images obtained from the sensitivity matrix algorithm is available, the transition matrix of the evolution model can be estimated by the Ibrahim Time Domain Method (ITD) [11], [12]. The Extended Kalman Filter was developed in the early 60's. It is a predictor-corrector estimator. It minimizes the trace of the covariance matrix of the estimation error. The EKF takes into account modelling errors and measurement errors [13]. The state in the present work is a vector of the resistivities of each nite element of the model that represents the domain. The discrete time EKF is described by equation (1) to equation (5),

2. The Extended Kalman Filter

k Pk k Pk

()

= k1 = P k1 T + Qk1
() () () () () (+)

(+)

(1) (2) (3) (4) (5)

()

()

Gk = P k H T (k )[H k (k )P k H T (k ) + Rk ]1 k k
(+) (+)

= k

()

+ Gk [v k H k k ]
() ()

= [I Gk H k k ]P k

where (1) and (2) represent the propagation phase and (3) to (5) represent the update phase. The vector is the state vector, v is the measurements vector, is the transition matrix of the

evolution model, P is the covariance matrix of the estimation error, Q is the covariance matrix of the state noise, H is the linear observation matrix1 , R is the covariance of the measurement noise and G is the Kalman gain matrix. In order to use this set of equations, the number of lines of the observation matrix H must be the number of electrodes e times the number of current patterns p, ep [9]. Eectively, there are ep dierent sensors, in the sense that, the observation model depends, also, on the current pattern. When this fact is not taken into account, there must be p dierent P matrices, one for each current pattern.

3. The Evolution Model

The EKF requires a transition matrix , which is part of the state evolution model. The dynamics of ventilation and perfusion will be represented by a second order homogeneous set of ordinary dierential equations (ODEs). The evolution model required in Kalman lters is a set of rst order ODEs. To accommodate the second order evolution model into a rst order Kalman evolution model, the size of the state must be duplicated [14]. Which means that the state vector for the remaining of this work have dimension 2n, where n is the number of nite elements in the mesh, and will be denoted by . The transition matrix is estimated in this paper through the Ibrahim Time Domain Method (ITD) [15], [11], [12].The ITD requires two matrices X and X + lled with state vectors in its columns. Each column of these matrices are formed by the augmented state (k) . The state vectors that compose X are a time increasing sequence of state vectors, from discrete time k to discrete time k + m 1 and the state vectors that compose X + are a time increasing sequence of state vectors, from discrete time k + 1 to discrete time k + m.

X= X+ = (k) =

(k) (k+1) . . . (k+m1) (k+1) (k+2) . . . (k+m) (k1) (k)

(6) (7) (8)

It is good to have m > 2n augmented state vectors to estimate to form a least square solution. For this solution, a pseudo-inverse of X is required and a Tikhonov regularization is also necessary [15], [16], [14]. The transition matrix is estimated according to

= (X + X T )(XX T + I)1 ,

(9)

where is the Tikhonov regularization parameter and I is the identity matrix. For the sequence of absolute images (k) , a sensitivity matrix algorithm was used [17] and the absolute images are approximated through the addition of the resistivity distribution used for the Taylor series expansion. The sensitivity matrix images are under estimated, but the method is robust, gives good localization of the objects and, therefore, are used to estimate the transition matrix. The images resulting from EKF take a longer time to become informative. With a certain periodicity, a new transition matrix can be estimated in parallel with the EKF estimation. The periodicity is problem dependent. In the present work, the transition matrix is estimated only once.
1

The observation matrix is obtained by assembling the rst order matrices from Taylor series expansions of the nonlinear equation of the nite element model, one for each current pattern cj , v j = Y 1 cj , where v j is the electrical potentials at the electrodes, cj is the current pattern injected and Y is the global conductivity matrix [14].

4. Numerical Simulations

A nite element numerical phantom comprising a cylindrical domain of 300mm of diameter, n = 514 linear triangular elements, e = 32 electrodes, p = 32 current patterns and an object of 40mm of diameter was developed. The object, which has time varying resistivity, was placed in two positions: in the center of the domain and 126mm distant from the center of the domain. Two dierent functions were used to represent the time varying resistivity of the object. The rst function is used to generate the sequence of voltages that will result in a sequences of images obtained by the sensitivity matrix algorithm, which will comprise the input data for ITD.

object (k) =

35.0 , k0 10 k + 35.0 , k > 0 50

m basal = 35.0m

(10) (11)

The second function is a step function in the object resistivity at k = 0.

object (k) =

35.0 , k < 0 100.0 , k 0

(12)

basal = 35.0m.
The second function is used to generate the sequence of voltages that EKF will use as measurements, so it has to estimate a step function using a transition matrix of a ramp function. Two dierent functions were used to avoid inverse crime. The complete electrode model was used [18] and Gaussian noise was added to the electrical potential measurements.

5. Numerical Results

The results will be presented in two parts, the rst part shows the estimation of the evolution model, the second part shows the state estimation using the EKF.

5.1. Part 1 The numerical phantom was used to generate electrical potential data when the object is in the center or o the center. The initial uniform resistivity was 0 = 35m and the current intensity was I = 2mA peak-to-peak2 . Figure (1) shows the time history of two representative elements taken from the basal region and from the object region, using the estimated transition matrix. It can be noted that the object's resistivity estimation is smaller than the real resistivity. It can also be noted that there is an asymptotic behavior in the estimation. It's partially caused by the nite precision measurements that can't register electrical potential changes caused by the resistivity changes in the object when there is a big resistivity dierence between object and basal and also caused by sensitivity matrix algorithm, which is a linearized map from observations to image of a non-linear map. This asymptotic limit depends on the position of the object relative to the electrodes and the dierence between the medium and the object resistivities.

In electronics, peak-to-peak refers to twice the amplitude of a sinusoidal signal.

350 estimated resistivity estimated basal object real resistivity basal real resistivity

350 estimated resistivity estimated basal object real resistivity basal real resistivity

300

300

250 resistivity ( m) resistivity ( m) 500 iteration (k) 1000 1500

250

200

200

150

150

100

100

50

50

0 0

0 0

500 iteration (k)

1000

1500

(a)

(b)

Figure 1. Time history. (a) Object at the center; (b) Object o the center.
These eects are not caused by the ITD method. To show this, the gure (2) presents the mean dierence between the evolution state used as input data for the ITD method and the state evolution using the transition matrix computed. The graph shows the mean dierence in a continuous line and the mean plus or minus one standard deviation in dashed lines 3 . It can be noted that the error in g. 2 is much smaller than the dierence between the real resistivity and the estimated resistivity using the sensitivity matrix algorithm.

This error index is given by

ek = k , k

(13)

where k is the iteration, is the input image for the ITD and is the image obtained from the transition matrix calculated by ITD. The mean error and it's standard deviation are given by
ek = 1X ei k n i=1 v u n u 1 X =t (ei k ek )2 , n 1 i=1
n

(14) (15)

ek

where n is the size of .

1 0.8 0.6 0.4 error ( m) error ( m) 500 iteration (k) 1000 1500 0.2 0 0.2 0.4 0.6 0.8 1 0

1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1 0 500 iteration (k) 1000 1500

(a)

(b)

Figure 2. Mean dierence between the state evolution using the input data for the ITD and

the state evolution using the transition matrix computed with the ITD method. (a) Object at the center; (b) Object o the center. Figure (3) shows the state at k = 1495 using the estimated transition matrix. The circle marks the correct object location in two dierent positions.
0.15 75
0.15 200 180 0.1 160 140 120

0.1

70

65 0.05 60
0.05

55
0 100 80 0.05 60 40 0.1 20 0

50 0.05

45

40 0.1 35 0.15 0.15 0.1 0.05 0 0.05 0.1 0.15

0.15 0.15 0.1 0.05 0 0.05 0.1 0.15

30

(a)

(b)

Figure 3. Dierence image at iteration k = 1495. (a) Object at the center; (b) Object o the
center. The transition matrix was computed using the regularization parameter = 1.0 to avoid instabilities in the state evolution. Smaller values of generate instabilities.

5.2. Inuence of the regularization parameter The model used in this work is a second order with n degrees of freedom system, so each pair eigenvalue-eigenvector represents one mode of the evolution. Each mode has a damping factor that depends on the magnitude of its eigenvalue. If the magnitude is smaller than 1.0, the mode is stable, if the magnitude is 1.0, the mode is critically stable, if the magnitude is greater than 1.0 the mode is unstable. Figure (4) shows the Nyquist plot of the transition matrix's eigenvalues for two dierent values of when the object is o the center of the domain. Increasing , the most of the eigenvalues have their magnitudes reduced, which increases the damping. The eigenvalues close

to the unitary circle are not heavily inuenced by . Similar results are observed when the object is located in the center of the domain.

0.8

0.8

0.6

0.6

0.4

0.4

0.2 imaginary imaginary 1 0.5 0 real 0.5 1

0.2

0.2

0.2

0.4

0.4

0.6

0.6

0.8

0.8

0.5

0 real

0.5

(a)

(b)

Figure 4. Nyquist plot of the eigenvalues of matrix , object o the center. (a) = 1.0; (b)
= 0.01.
The dominant modes, the modes with larger eigenvalues, have eigenvalues close to the point (1.0, 0.0) in the unit circle and their eigenvectors represent the position of the object according to gure (5) and gure (6). The correlation of the eigenvectors and the most recent image was computed using the Modal Assurance Criterion (MAC) [15] revealing that only the rst three eigenvectors have high correlation, M AC > 0.5, with the image.
0.1

0.15

z=1 0

0.15

z=0.999 0.05

0.1

0.02

0.1

0.04

0.05
0.06

0.05

0.05

0.1

0.08

0
0.15

0.1

0.05
0.12

0.05

0.2

0.25

0.1

0.14

0.1
0.3

0.16

0.15 0.15 0.1 0.05 0 0.05 0.1 0.15

0.15
0.35

0.15

0.1

0.05

0.05

0.1

0.15

(a)

(b)

Figure 5. Dominant eigenvectors of matrix , ramp resistivity variation with = 1.0. (a)
Eigenvector 1; (b) Eigenvector 2.

0.15

z=0.993 0.05

0.15

z=0.772

0.1 0.08

0.1

0.1

0.06 0.04

0.05

0.05

0.05
0.02

0.1

0 0.02

0.05

0.15

0.05

0.04 0.06

0.2

0.1
0.25

0.1
0.08 0.1

0.15 0.15 0.1 0.05 0 0.05 0.1 0.15

0.15 0.15 0.1 0.05 0 0.05 0.1 0.15

(a)

(b)

Figure 6. Dominant eigenvectors of matrix , ramp resistivity variation with = 1.0. (a)
Eigenvector 3; (b) Eigenvector 4.

5.3. Part 2 In the second part is shown the state estimation using the EKF with the estimated evolution model. The results are grouped according to the object localization. In EIT, one can nd information about the adjustments in [5], [2], [9]. 5.3.1. Object o the center The procedure to adjust the EKF parameters was done adjusting one parameter at a time. The initial uniform resistivity 0 = 60m was chosen to overestimate the real resistivity. The initial estimation error covariance matrix was chosen such that P 0 = 101 I . Matrix Qk was updated following
Qk = P k1 , k 3 , Qk1 , k>3
(16)

to keep the lter open to new information[19]. The Kalman gain matrix is inuenced by errors in H due to wrong linearization states. Therefore, a relaxation factor k was used in eq. 3 according to

Gk = k (P k H T [H k P k H T + Rk ]1 ), k k
where k is a relaxation factor dened by 0.1 , k < 5 0.5 , k < 8 , k = 1.0 , k 8 to avoid large oscillations of the initial state updates.

()

()

(17)

(18)

Matrix R adjustment
Although Rk can be estimated from electronic noise, modeling errors in the observation equation eectively increases Rk [5]. Therefore, Rk was adjusted taking into consideration the

observation residue history from a few EKF tests, keeping the observation residue inside the limit of three measurement noise standard deviation [9]. Table (1) shows the parameters used4 .

Table 1. Parameters used in the tests to adjust R

Test Rk (constant) P0 Q0 0 [m] A 102 I 101 I 103 P 0 60 103 B 103 I 101 I 103 P 0 60 103 C 104 I 101 I 103 P 0 60 103 D 105 I 101 I 103 P 0 60 103

Figure (7) shows the history of the observation residue for dierent values of Rk . An adequate value is Rk = 104 I (test C), since the observation residue keeps inside the limit.
10
0

10 residue superior limit

residue superior limit

observation residue (V)

10

observation residue (V) 50 100 150 iteration 200 250 300

10

10

10

50

100

150 iteration

200

250

300

(a)
10
0

(b)
10 residue superior limit
0

residue superior limit

observation residue (V)

observation residue (V) 50 100 150 iteration 200 250 300

10

10

10

10

10

50

100

150 iteration

200

250

300

(c)

(d)

Figure 7. History of the state estimation error. (a) Test A; (b) Test B; (c) Test C; (d) Test D;
4

All tests were ordered alphabetically and the parameters of each test are given in tables of each subsection.

Matrix Q adjustment
The Qk was adjusted taking in account the estimated state evolution and the quantity of artifacts on the images. The table (2) shows the parameter used in these tests.

Table 2. Parameters used in the tests to adjust Q

Test Rk (constant) P0 Q0 0 [m] E 104 I 101 I 103 P 0 60 103 F 104 I 101 I 105 P 0 60 105

Figure (8) shows the state history and the image on iteration k = 300. It suggests that a reduction of Qk causes a reduction of the tracking ability of EKF and increases the spatial resolution. Test F was adopted as the best set of parameters for EKF.
0.15
100 90 80
80

0.1

70

0.05
70 resistivity [ m] 60 50 40 30 20 10 0

60

50

0.05

40

0.1

30

0.15
50 100 150 iteration 200 250 300

20

0.15

0.1

0.05

0.05

0.1

0.15

(a)
0.15
90

(b)
70

65

80 70

0.1
60

0.05
resistivity [ m] 60

55

50

50 40 30

0
45

0.05

40

35

0.1
20 10 0
30

0.15
50 100 150 iteration 200 250 300

25

0.15

0.1

0.05

0.05

0.1

0.15

(c)

(d)

Figure 8. State history for two values of Qk . (a) state history, test E; (b) state on test E at
k = 300; (c) state history, test F; (b) state on test F at k = 300;

5.3.2. When the object is at the center of the domain The set of parameters from test F were used in test G. Figure (9) shows the state history, the image at iteration k = 300, and the observation residue. Table (3) shows the parameters used.
110 100 90 observation residue (V) 80 resistivity [ m] 70 60 50 40 30 20 10 0 50 100 150 iteration 200 250 300 10
3

10

residue superior limit

10

10

50

100

150 iteration

200

250

300

(a)
0.15

(b)

80

0.1
70

0.05
60

0
50

0.05
40

0.1
30

0.15 0.15 0.1 0.05 0 0.05 0.1 0.15

(c)

Figure 9. Test G. (a) State history; (b) Observation residue; (c) State at k = 300.
It can be observed that the observation residue is lower in test G than in test F. When the object is in the center, errors in the state estimation have low inuence in the predicted measurements. Therefore, a direct comparison between observation residues is not fair. Since the tracking ability can be improved, another set of parameters was proposed, based on the parameters of test G. A smaller value of matrix Rk was attempted. Table (3) shows the parameters used.

Table 3. Parameters used in the tests to adjust R

Test Rk (constant) P0 Q0 0 [m] G 104 I 101 I 105 P 0 60 105 H 106 I 101 I 105 P 0 60 105

Figure (10) shows the state history, the image at iteration k = 300 and the observation residue.
120 10
0

residue superior limit 100

observation residue (V) 50 100 150 iteration 200 250 300

80 resistivity [ m]

10

60

40

10

20

0
3

20 0

10

50

100

150 iteration

200

250

300

(a)
0.15

(b)

80

0.1
70

0.05
60

0
50

0.05
40

0.1
30

0.15 0.15 0.1 0.05 0 0.05 0.1 0.15

(c)

Figure 10. Test H. (a) State history; (b) Observation residue; (c) State at k = 300.
In about 25 iterations test H converged. Comparing test G and test H, clearly the tracking ability improved by reducing Rk .

6. Performance of EKF with the Random Walk Evolution Model

In this subsection the performance of EKF using the Random Walk Model is presented. Figure (11) shows the state history, the observation residue and the state at iteration k = 800 for the two dierent object positions with the Random Walk model. The object was not visible in the center although the observation residue lies within three standard deviation range. It can be observed that there are many artifacts in both images.

70

120

60
100

50 resistivity [ m]
80 resistivity [ m]

40

60

30

40

20
20

10

0 0

100

200

300

400 iteration

500

600

700

800

0 0

100

200

300

400 iteration

500

600

700

800

(a)
10
0

(b)
10 residue superior limit
0

residue superior limit 10 observation residue (V)

1

10 observation residue (V)

10

10

10

10

10

10

10

100

200

300

400 iteration

500

600

700

800

10

100

200

300

400 iteration

500

600

700

800

(c)
0.15
60

(d)
0.15
110 100

0.1
50

0.1

90 80

0.05
40

0.05
70

0
30

60 50

0.05
20

0.05

40 30

0.1
10

0.1
20 10

0.15 0.15 0.1 0.05 0 0.05 0.1 0.15

0.15 0.15 0.1 0.05 0 0.05 0.1 0.15

(e)

(f)

Figure 11. Results using the Random Walk model. (a) State history, object at the center; (b)

State history, object o the center; (c) observation residue, object at the center; (d) observation residue, object o the center; (e) State at k = 800, object at the center; (f) State at k = 800, object o the center. The convergence of EKF with the ITD estimated evolution model when the object was in the center of the domain happened after 25 iterations and when the object was o the center happened after 200 iterations. The convergence of EKF with the Random Walk evolution model

when the object was in the center of the domain ocurred after 800 iterations and when the object was o the center did not happened within 800 iterations. The tracking ability was accelerated 32 times by the use of the ITD estimated evolution model when the object was in the center of the domain and mores than 4 times when the object was o the center of the domain.

7. Final Comments

The results show that the Ibrahim Time Domain method can be used to estimate the evolution model based on a sequence of images obtained by a sensitivity matrix algorithm. Although the resistivity distribution obtained by the sensitivity matrix algorithm is underestimated, the tracking ability of the EKF was improved when compared to the use of the Random Walk model for the evolution model. The computational eort to implement the direct identication of the transition matrix seems to be justied since the tracking ability improved approximately thirty two times when the object was in the center and more than four times when the object was near the boundary. The use of observation vector of dimension 1024 instead of 32 caused each iteration to be longer but, on the other hand, the P update can be performed according to the equation (5) and an oscillation with periodicity 32 was avoided.

Acknowledgments

The authors want to express their gratitude to Marko Vauhkonen, for suggesting the use of an evolution model other than random walk model. The authors are also grateful for the nancial aid from The State of So Paulo Research Foundation, processes 01/05303-4, 05/01088-2 and 07/04069-4.

8. References
[1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19]

Cheney M, Isaacson D and Newell J C 1999 Society for Industrial and Applied Mathematics 41 85101 Trigo F C, Lima R G and Amato M B P 2004 IEEE Transactions on Biomedical Engineering 51 7281 Maybeck P S 1979 Stochastic models, estimation and control 1st ed vol 1 (New York: Academic Press Inc.) Jazwinski A H 1970 Stochastic Processes and Filtering Theory (New York: Academic Press Inc.) Vauhkonen M, Karjalainen P A and Kaipio J P 1998 IEEE Transactions on Biomedical Engineering 45 486493 Kaipio J P, Karjalainen P A and Somersalo E 1999 Annals-New York Academy Of Sciences 430439 Kim K Y, Kim B S, Kim M C, Lee Y and Vauhkonen M 2001 Measurement Science and Technology 10321039 Kim K Y, Kang S I, Kim M C, Kim S, Lee Y and Vauhkonen M 2002 IEEE Transactions on Magnetics 38 13011304 Trigo F C 2005 Estimao no linear de parmetros atravs dos ltros de Kalman na tomograa por impedncia eltrica Ph.D. thesis Escola Politcnica da Universidade de So Paulo So Paulo Kim B S, Kim M C, Kim S and Kim K Y 2004 Measurement Science and Technology 15 21132123 Ibrahim S R and Mikulcik E C 1973 Shock and vibration 43 2137 Ibrahim S R and Mikulcik E C 1977 Shock and vibration 47 183198 Trigo F C 2001 Filtro Estendido de Kalman Aplicado Tomograa por Impedncia Eltrica Master's thesis Escola Politcnica da Universidade de So Paulo So Paulo Moura F S 2006 Identicao da Matriz de Transio do Modelo de Evoluo do Filtro Estendido de Kalman na Tomograa por Impedncia Eltrica (So Paulo) Ewins D J 1984 Modal Testing: Theory and Pratice 1st ed (New York: Research Studies Press LTD.) Neto L S F 2004 Um mtodo para anlise modal de estruturas submetidas a excitaes ambientes Master's thesis Escola Politcnica da Universidade de So Paulo So Paulo Moura F S, Aya J C C, Mirandola L A S, Nan P C, Schweder R K and Lima R G 2005 International Congress of Mechanical Engineering Hua P, Woo E J, Webster J G and Tompkins W J 1993 IEEE Transactions on Biomedical Engineering 40 335343 Brogan W L 1991 Modern Control Theory 3rd ed (New Jersey: Prentice Hall)