Calibrating method and system, recording medium for this method

Abstract

A method for calibrating coefficients of an observer of a variable state of a physical system from measurements of physical quantities of the system, at different instants includes measuring a variable of the physical system at several instants, the variable being a function of a system state variable, and determining a vector of coefficients that minimizes a sum of the number of measurements of a square of a norm of a vector that is the difference between the measured variable and a function of the system state variable and the vector of coefficients. The minimization is subject to a constraint that the trajectory of the measured variable be within a corridor of uncertainty on either side of a trajectory of its estimate, at least for the measurement instants.

Description

CROSS REFERENCE TO RELATED APPLICATIONS

This application claims the benefit of the Mar. 29, 2010 priority date of French application FR1052261. The contents of the foregoing application are incorporated herein by reference.

The invention pertains to a method and system for calibrating coefficients of an observer of a variable state x_k of a physical system. The invention also pertains to a recording medium to implement this method.

An observer estimates the state variable x_k from measurements y_k of physical quantities or magnitudes of this system where the index k identifies the instant of measurement.

There are known ways of calibrating the coefficients of an observer by successive experiments with different sets of coefficients until one set of coefficients gives the desired operation by this observer. Typically, the behavior desired for the observer is:

• • fast convergence towards the most exact possible estimation {circumflex over (x)}_k of the variable x_k, and
  • high stability.

This calibration by experiment is part of a heuristic approach. The calibration is often lengthy, and it is difficult to be certain that the best possible calibration of the observer has been obtained.

To resolve this problem, other alternative methods have been proposed. For example, the patent application CA 2 226 282 describes a method for calibrating coefficients of a Kalman filter by means of a neural network. However, these alternative methods are often as complex to implement as the method that uses experimentation.

The invention is therefore aimed at proposing a method of calibration that is simpler to implement.

Thus, an object of the invention is a method for calibrating coefficients of an observer comprising:

• • the measurement of a variable z_k of the physical system at N different instants, this variable z_k being a function of the state variable x_k;
  • the determining of a vector p of coefficients which minimizes the following criterion in complying with a predetermined set Δ of constraints:

$\left\{f\left(p\right)=\sum _{k=1}^N{z_k-\varphi \left({\hat{x}}_k,p\right)}^2\right\}$ where:

• • {circumflex over (x)}_k is the estimation of the variable x_k built by the observer calibrated with the coefficients of the vector p at the instant k;
  • φ is a known function which links the estimation {circumflex over (x)}_k to an estimation {circumflex over (z)}_k of the variable z_k;
  • ∥ . . . ∥² is a norm of the difference between z_k and φ({circumflex over (x)}_k,p), and
  • the constraint or the constraints of the set Δ dictate that the trajectory of the variable z_k should be included in a corridor of uncertainty situated on either side of the trajectory of the estimation {circumflex over (z)}_k at least for the majority of the instants k.

The above method is used to determine the coefficients enabling an efficient functioning of the observer in a simple way and without recourse to different experiments. Furthermore, the coefficients thus determined make it possible to ensure that the criterion f(p) has been minimized.

Finally, the use of the set Δ takes account of the fact that there is an uncertainty on the estimation {circumflex over (z)}_k. Taking account of this uncertainty when determining the vector p that minimizes the criterion f(p) gives better calibration than if this uncertainty were not taken into account. In particular, aberrant approaches which could minimize the criterion f(p) are eliminated.

The embodiments of this calibration method may include one or more of the following characteristics:

• • the set Δ includes a constraint which dictates that the variable z_k should be included between {circumflex over (z)}_k−a and {circumflex over (z)}_k+b, where a and b are predetermined positive constants;
  • the constants a and b are proportional to a mean square deviation σ_{{grave over (z)}k} representing the uncertainty in the estimation {circumflex over (z)}_k;
  • the set Δ of constraints includes a constraint which dictates that the mean of the deviations between the variable z_k and its estimation {circumflex over (z)}_k should be below a predetermined threshold S₁;
  • the threshold S₁ is a function of the uncertainty on the estimation {circumflex over (z)}_k;
  • the observer is a Kalman filter, an extended Kalman filter or an Unscented Kalman Filter and the coefficients to be calibrated are coefficients of the covariance matrices R and Q of this filter;
  • the R and Q matrices are chosen to be diagonal matrices.

These embodiments of the calibration method furthermore have the following advantages:

• • dictating that the variable z_k should be included between {circumflex over (z)}_k−a and {circumflex over (z)}_k+b reduces the space of search in which the vector p must be sought, thus accelerating the execution of this method;
  • using constants a and b proportional to the mean square deviation σ_{{circumflex over (z)}k} improves the calibration,
  • using a constraint which dictates that the mean of the deviations in terms of absolute value between the variable z_k and its estimation {circumflex over (z)}_k should be below the threshold S₁ relaxes the dictated constraints thus facilitating the determination of a vector p;
  • using a threshold S₁ which is a function of the uncertainty of the estimation of {circumflex over (z)}_k improves the calibration;
  • choosing diagonal R and Q matrices decreases the number of coefficients to be determined, thus facilitating the calibration.

An object of the invention is also an information-recording medium comprising instructions to implement the above method when the instructions are executed by a programmable electronic computer.

An object of the invention is also a system for calibrating coefficients of an observer of a state variable x_k of a physical system from measurements y_k of physical quantities of this system where the index k identifies the instant of measurement, the system comprising:

• • at least one sensor capable of measuring a variable z_k of the physical system at N different instants, this variable z_k being a function of the state variable x_k, and
  • a computer capable of determining a vector p of coefficients which minimizes the following criterion in complying with a predetermined set Δ of constraints:

$\left\{f\left(p\right)=\sum _{k=1}^N{z_k-\varphi \left({\hat{x}}_k,p\right)}^2\right\}$ where:

• • {circumflex over (x)}_k is the estimation of the variable x_k built by the observer calibrated with the coefficients of the vector p at the instant k,
  • φ is a known function which links the estimation {circumflex over (x)}_k to an estimation {circumflex over (z)}_k of the variable z_k,
  • ∥ . . . ∥² is a norm of the difference between z_k and φ({circumflex over (x)}_k,p), and
  • the constraint or the constraints of the set Δ dictate that the trajectory of the variable z_k should be included in a corridor of uncertainty situated on either side of the trajectory of the estimation {circumflex over (z)}_k at least for the majority of the instants k.

The invention shall be understood more clearly from the following description, given purely by way of a non-exhaustive example and made with reference to the appended drawings of which:

FIG. 1 is a schematic illustration of the architecture of a localizing system including an observer and a system for calibrating this observer,

FIG. 2 is a flowchart of a method of calibration of the observer of the system of FIG. 1, and

FIG. 3 is a graph illustrating the trajectory of a variable z_k and its estimation {circumflex over (z)}_k.

Here below in this description, the characteristics and functions well known to those skilled in the art shall not be described in greater detail.

Here, a trajectory designates the variations in time of a variable. The variable that varies in time may be any unspecified variable. In particular, this variable is not necessarily a position in space. For example, it may a temperature or any other physical quantity which varies over time.

FIG. 1 represents a physical system, i.e. in this case an object 4 and a system 8 for localizing this object 4 in the reference system 8.

For example, the object 4 is a probe or a catheter introduced into a human body. The object 4 is mobile in the referential system 8.

The referential system 8 is a fixed referential system having three orthonormal axes X, Y and Z.

The localization of the object 4 in the referential system 8 consists in finding its x_p, y_p, z_p position and its θ_x, θ_y and θ_z orientation. The angles θ_x, θ_y and θ_z represent the orientation of the object 4 respectively in relation to the axes X, Y and Z.

To localize the object 4, the system 6 herein comprises magnetic field sources and magnetic field sensors some of which are linked to the referential system 8 while others are fixed without any degree of freedom to the object 4 to be localized. Here, the magnetic field sensor 10 is linked without any degree of freedom to the object 4. The sensor 10 is for example a triaxial sensor, i.e. a sensor capable of measuring the projection of the magnetic field on three non-colinear measurement axes. Such a sensor measures the direction of the magnetic field. More generally, this sensor also measures the amplitude of the magnetic field.

For example, the measurement axes are mutually orthogonal. These axes are fixedly linked to the object 4.

This sensor 10 is connected by means of a flexible wire link to a processing unit 12.

The unit 12 is also connected to three sources 14 to 16 of the magnetic field. These sources are for example magnetic field triaxial sources. A magnetic field triaxial source emits a magnetic field, along three mutually non-colinear emission axes. For example, such a source is formed by several aligned magnetic field uniaxial emitters, respectively on each of the emission axes of the source. The uniaxial emitter mainly emits the magnetic field along a single axis. For example, it is a coil whose turns are wound about a same emission axis. In this case, the emission axis coincides with the winding axis of the turns.

Here, the sources 14 to 16 are identical to each other. These sources are fixed in the referential system 8.

The processing unit 12 powers the sources 14 to 16 with alternating current to generate the magnetic field measured by the sensor 10. The unit 12 also acquires the measurements of the magnetic field made by the sensor 10.

The unit 12 is equipped with an observer 18 which localizes the object 4 in the referential system 8 from the measurements acquired. Here, the observer 18 determines the position and orientation of the object 4 by resolving a system of equations. This system of equations is obtained by modeling the magnetic interactions between the sources 14 to 16 and the sensor 10. In this system of equations, the position x_p, y_p and z_p and the orientations θ_x, θ_y and θ_z of the object 4 are the unknown quantities while the values of the other parameters are obtained from measurements made by the sensor 10. Further information on such systems of equations may for example be found in the patent application EP 1 502 544.

Here, the observer 18 is a Kalman filter.

This observer 18 builds an estimation {circumflex over (x)}_k of a state variable x_k from a measurement y_k. The variable x_k is a vector containing the position x_p, y_p, z_p of the object 4 and its orientation θ_x, θ_y and θ_z in the orthonormal referential system 8, the measurement y_k is a measurement vector comprising the measurements made along the three measurement axes of the sensor 10 at an instant k.

The equations of the Kalman filter of the observer 18 are the following:

$\begin{array}{cc}\left\{\begin{array}{c}{\hat{x}}_{k+1/k}=F_k{\hat{x}}_k\\ P_{k+1/k}=F_k{P}_{k/k}{F}_k^T+Q_k\end{array}\right\}& \left(1\right)\\ \left\{\begin{array}{c}{\hat{x}}_{k+1/k+1}={\hat{x}}_{k+1/k}+K_{k+1}\left(y_{k+1}-C_{k+1}{\hat{x}}_{k+1/k}\right)\\ P_{k+1/k+1}=P_{k+1/k}-K_{k+1}{C}_{k+1}{P}_{k+1/k}\end{array}\right\}& \left(2\right)\end{array}$

where:

• • {circumflex over (x)}_k+1/k is the estimation of the variable x_k obtained solely from data available at the instant k,
  • F_k is the matrix linking the estimation {circumflex over (x)}_k to the estimation {circumflex over (x)}_k+1/k,
  • P_k/k is the matrix for estimating the covariance, of the errors on the estimation {circumflex over (x)}_k,
  • Q_k is a matrix of covariance of the noise of the equation of the model containing inter alia the modeling errors,
  • {circumflex over (x)}_k+1/k+1 is the estimation {circumflex over (x)}_k+1 corrected at the instant k+1 from the measurements made at the instant k+1,
  • y_k is the vector of the measurements made at the instant k,
  • C_k is the matrix that links the variable x_k to the measurement y_k,
  • P_k+1/k+1 is the matrix of covariance of the error of estimation {circumflex over (x)}_k corrected on the basis of the measurements y_k,
  • K_k+1 is the gain of the Kalman filter.

The matrix F_k is for example obtained by modelizing the magnetic interactions between the sources 14 to 16 and the sensor 10 on the basis of the laws of electromagnetism.

The gain K_k+1 of the Kalman filter is given by the following relationship:K_k+1=P_k+1/kC^T_k+1[C_k+1P_k+1/kC^T_k+1+R_k+1]⁻¹  (3)where R_k+1 is the matrix of covariance of the noise in the measurement y_k.

The given equations of the Kalman filter are classic and are therefore not described in greater detail herein.

The use of a Kalman filter gives a precise estimation of the position and orientation of the object 4. However, before this, the matrices R_k and Q_k have to be calibrated, i.e. the values of their coefficients have to be determined. To simplify the description, it is assumed here that the values of these coefficients are constant. The matrices R_k and Q_k are therefore also denoted as R and Q here below.

Here p_R and p_Q denote the set of coefficients respectively of the matrices R and Q to be determined.

Once the Kalman filter has been calibrated, these sets P_R and P_Q are registered in a memory 20 connected to the unit 12 so that they can be used by the observer 18 to estimate the position of the object 4.

FIG. 1 also represents a system 30 for calibrating the observer 18, this system automatically determines the sets p_R and p_Q of coefficients of the matrices R and Q of the observer 18.

To this end, the system 30 is equipped with one or more sensors 32 which measure a variable z_k. This variable z_k is a variable whose value is a function of the variable x_k. More specifically, the variable z_k is connected to the variable x_k by a known function φ. This function φ may be a linear function or on the contrary a non-linear function. For example, the sensor 32 measures the acceleration along the X, Y and Z axes of the object 4. In this case, the sensor 32 is formed by several accelerometers.

In another embodiment, the sensor 32 can directly measure the position of the object 4 with sensors other than those used to carry out the measurement y_k. For example, the sensor 32 may, to this end, have several cameras or other sensors and other magnetic field sources to measure the position of the object 4. In this particular case, the function φ is the identity function.

The system 30 also has a computer 34 capable of acquiring the measurements of the sensor 32. This computer 34 is also capable of calibrating the observer automatically by determining the set of coefficients P_R and P_Q from the measurements made by the sensor 32.

In the particular case shown in FIG. 1, the computer 34 is connected to the memory 20 so that it can directly record the set of coefficients p_R and p_Q in this memory 20.

The computer 34 is made with a programmable electronic computer capable of executing the instructions recorded in an information-recording medium. Here, the computer 34 is connected to a memory 36 and this memory 36 contains the instructions needed to execute the method of FIG. 2.

The working of the calibration system 30 shall now be described in greater detail with reference to the method of FIG. 2 and the graph of FIG. 3.

Initially, at a step 40, the relationship φ which connects the variable z_k to the variable x_k is set up. For example, to this end, the laws of physics are used.

Then, at a step 42, the number of coefficients to be used to calibrate the observer 18 is determined. To this end, knowledge on the physical working of the system 6 is used. Furthermore, this step applies the principle of parsimony according to which the number of coefficients to be calibrated is limited to the greatest possible extent.

For example, in the particular case described herein, it is assumed that the measurements made on each of the measurement axes are independent of one another. Thus, the matrix R is taken to be a diagonal matrix. Similarly, it is assumed that the errors in the state equation are independent. Thus, the matrix Q is taken to be a diagonal matrix. They therefore have the following form:R=diag(p¹_R, . . . ,p^l_R),pⁱ_R≧0,i=1:l  (7)Q=diag(p¹_Q, . . . ,pⁿ_Q),pⁱ_Q≧0,i=1:n  (8)where:

• • the index i represents the index of the coefficient in the set of coefficients p_R or p_Q,
  • l and n are respectively the dimensions of the matrices R and Q.

Furthermore, the matrices R and Q are symmetrical matrices defined as being positive, i.e. all matrices of which all the Eigen values are positive. To ensure this condition, the matrices R and Q can be written as noted here below:

R=X^TX and Q=Y^TY, where the matrices X and Y are diagonal matrices. It can be noted that when the matrices R and Q are any unspecified matrices, the matrices X and Y are higher triangular matrices.

At the end of the step 42, with the assumptions made, the vector p of coefficients to be determined is defined by the following relationship:p=[p^T_R,p^T_Q]^T  (9)

At a step 44, the criterion to be optimized to determine the vector p is chosen. Here, this vector is given by the following relationship:

$\begin{array}{cc}\underset{p\in \Delta }{\min }\left\{f\left(p\right)=\sum _{k=1}^N{z_k-\varphi \left({\hat{x}}_k,p\right)}_2^2\right\}& \left(10\right)\end{array}$ where:

• • ∥ . . . ∥₂ represents the L₂ norm operation corresponding to the Euclidian norm,
  • <<min>> represents the operation of finding the vector p which minimizes the function f(p),
  • Δ is the set of constraints that must be verified by the vector p which minimizes the function f(p).

The function φ which appears in the criterion (10) is the same as the one introduced here above. This function φ is used to obtain the estimation {circumflex over (z)}_k of the variable z_k from the estimation {circumflex over (x)}_k as indicated in the following relationship:{circumflex over (z)}_k=φ({circumflex over (x)}_k,p)  (11)

It will be noted that the function φ also depends on the vector p since the estimation {circumflex over (x)}_k which constitutes the argument of the function φ is itself a function of the vector p. For this reason, in the above relationship, the estimation {circumflex over (x)}_k and the vector p are indicated as the argument of the function φ.

At a step 46, the set Δ of constraints to be met by the vector p is defined. This set Δ comprises at least one constraint which dictates that estimation {circumflex over (z)}_k and the variable z_k should be proximate to each other. More specifically, this constraint herein dictates that the trajectory of the variable z_k should be included, at least on an average, in a corridor 48 of uncertainty (FIG. 3) situated on either side of the trajectory of estimation {circumflex over (z)}_k.

This corridor 48 is demarcated by two boundaries L_min and L_max respectively situated on each side of the trajectory of estimation {circumflex over (z)}_k. The corridor of uncertainty represents the fact that the estimation {circumflex over (z)}_k is only known with an uncertainty attached to it. This uncertainty comes from the fact that there is uncertainty in the estimation {circumflex over (x)}_k used to build the estimation {circumflex over (z)}_k. Thus, to calibrate the observer 18 properly, it is necessary that, for the majority of the instants k, the variable z_k should be included in the corridor 48.

In the particular case shown in FIG. 3, the boundaries L_max and L_min which demarcate the corridor 48 are separated from the trajectory of {circumflex over (z)}_k, and respectively by distances a and b. In the case of FIG. 3, these distances a and b are constant in the course of time.

Preferably, the distances a and b are a function of the uncertainty in the estimation {circumflex over (z)}_k. For example, the distances a and b are all the greater as the mean square deviation σ_{{circumflex over (z)}k} is great. σ_{{circumflex over (z)}k} is the mean square deviation of the error in the estimation {circumflex over (z)}_k. For example, here, the set Δ includes a constraint which dictates that the mean of the divergences between the variable z_k and the estimation {circumflex over (z)}_k should be below a predetermined threshold S₁. This constraint can be expressed by means of the following relationship:

$\begin{array}{cc}\sqrt{\sum _{k=1}^N{z_k-{\hat{z}}_k}_2}\le S_1& \left(11\right)\end{array}$

The threshold S1 is for example defined by the following relationship:S₁=√{square root over (N)}ασ_{{circumflex over (z)}k}  (12)where:

• • N is the number of instants k, and
  • α is a constant, for example equal to three.

For each vector p potentially minimizing the criterion (10), it is possible to determine the mean square deviation σ_{{circumflex over (z)}k}. Indeed, when a vector p is known, the matrices R and Q are also known. From the matrices R and Q, it is possible to compute the matrix P_k/k of the covariance of the error in the estimation of {circumflex over (x)}_k. From this matrix P_k/k it is possible to determine the uncertainty affecting the estimation of {circumflex over (x)}_k. With this uncertainty being known, it is also possible to determine the uncertainty in the estimation {circumflex over (z)}_k since the function φ connecting these two estimations is known. The uncertainty in the estimation {circumflex over (z)}_k is then expressed in the form of the mean square deviation σ_{{circumflex over (z)}k}.

At the step 46, other constraints of the set Δ may be defined. For example, on the basis of the physical sense of some of the coefficients of the sets p_R and p_Q, it is possible to set limits on these coefficients. For example, if pieces of information on the measurement noise of the sensor 10 are known, it is possible on the basis on these pieces of information, to limit the values that can be taken by certain coefficients of the set p_R and p_Q.

Once the set Δ has been defined, in a step 50, the sensor 32 measures the variable z_k at N different instants. The number N of instants is typically greater than the number of coefficients to be determined contained in the vector p.

Once the variables z_k have been measured, they are acquired by the computer 34. Then, a step 52, the different variables and coefficients needed to minimize the criterion (10) are initialized. For example, the initial matrix P_0/0 of estimation of the covariance in the error of the estimation {circumflex over (x)}₀ is defined by the following relationship:P_0/0=λI,λ>>1  (13)where:

• • I is the identity matrix, and
  • λis a constant.

The constant λ is adjusted according to the confidence placed in the initialization of {circumflex over (x)}₀.

The choice of a constant λ far greater than one means that the uncertainty in the initial estimation {circumflex over (x)}₀ is very great.

Here, the initial estimation {circumflex over (x)}₀ and the initial value of the coefficients of the vector p are drawn randomly.

Then, a step 54 is performed for determining the vector p which minimizes the criterion (10). In this step, the values of the vector p are made to vary until a vector p is found that minimizes the criterion (10) and at the same time meets the constraints of the set α. This step is performed by means of a tool of optimization under constraints. Such tools of optimization under constraints are known. For example such a tool especially is the “Fmincon” function available in the “MATLAB®” development environment, inter alia for digital computation.

Once the vector p which minimizes the criterion (10) has been determined, at the step 56 the values of the coefficients of this vector p are used to calibrate the coefficients of the matrices R and Q of the observer 18. For example, during this step, the values of these coefficients are copied and recorded in the memory 20.

Then, at a step 58, the localizing system 6 uses the coefficients determined by the system 30 to determine the position and orientation of the object 4 from the measurements made by the sensor 10.

Many other embodiments are possible. For example, for a same physical system and for a same observer, it is possible to make different choices on the number of coefficients of this observer that have to be determined in order to calibrate it. For example, the number of coefficients may be even further reduced by requiring that the different coefficients of the diagonal matrix R or Q should all be proportional to a same coefficient. Conversely, the method described here above can also be applied to the case where the matrices Q and R are not diagonal matrices but only positively defined symmetrical matrices. This assumption on the contrary increases the number of coefficients to be determined by the system 30.

In the criterion (10), the Euclidian norm can be replaced by another norm.

Constraints other than those described here above can be used. For example, it is possible to require that the variable z_k should be systematically or routinely included within the corridor of uncertainty 48 at each instant k. This constraint is stronger than the one described with reference to FIG. 3 which requires only that the variable z_k should on an average be included in the corridor of uncertainty 48.

Different sets of constraints can be used at different instants. For example if, at certain instants belonging to a set L, the uncertainty in the measurement of the variable z_k is very low, it is then possible to use a greater constraint at these instants. For example if the instant k belongs to the set L, then it is possible to require that the variable z_k should be routinely included in a more limited corridor of uncertainty. For the instants that do not belong to the set L, the constraint used will be weaker. For example, it will require solely that, on an average, the variable z_k should be included in the corridor of uncertainty 48. The corridor of uncertainty used to define the constraint when the instant k belongs to the set L is not necessarily the same as the one used to define the constraint when the instant k does not belong to this set L.

Additional constraints other than the ones requiring that the variable z_k should be in the corridor of uncertainty of the trajectory of the estimation {circumflex over (z)}_k may be used. For example, it is possible to define constraints that dictate a low variation of estimation {circumflex over (z)}_k between two successive instants.

Choices other than the ones described above are possible for initializing the optimization of the criterion (10). For example, if the initial value of the variable x₀ is known, this value is used to initialize the variable, x₀ and the coefficients of the matrix P₀ are set as a function of the trust in this value.

The uncertainty on the estimation {circumflex over (z)}_k may be obtained by means other than the observations of the observer. For example, this uncertainty may be determined experimentally or on the basis of systems of equations other than those defining the observer 18.

The measured variable z_k and the measurement y_k may coincide. Thus it is not necessary to resort to an additional sensor such as the sensor 32.

The calibration method described here above also applies to the Extended Kalman Filter as well as to the Unscented Kalman Filter (UKF) or any other recursive filter having parameters to be set. More generally, the calibration method described here above can be used to calibrate any observer of a state variable whose coefficients have to be determined beforehand. Thus, the coefficients liable to be determined by means of the above method are not limited to those of a covariance matrix.

The invention has been described in the context of a given application. However, it is not limited to this type of application. Other applications may be taken into account, such as for example applications for obtaining measurements in the probing of seabeds for which the physical system makes it possible to have measurements representing surface/seabed distances or the depths of seas.

Claims

1. A method for calibrating coefficients of an observer of a variable state xk of a physical system from measurements yk of physical quantities of the system, wherein index k identifies an instant of measurement, the method comprising causing a sensor to measure a variable zk of the physical system at N different instants, the variable zk being a function of a state variable xk, and causing a computer to determine a vector p of coefficients that minimizes the following criterion in complying with a predetermined set Δ of one or more constraints:

2. The method of claim 1, wherein the predetermined set Δ includes a constraint that dictates that the variable zk be included between {circumflex over (z)}k−a and {circumflex over (z)}k+b , wherein a and b are predetermined positive constants.

3. The method of claim 2, wherein the constants a and b are proportional to a mean square deviation σ{circumflex over (z)}k representing uncertainty in the estimation {circumflex over (z)}k.

4. The method of claim 1, wherein the predetermined set Δ of constraints includes a constraint that dictates that a mean of deviations between the variable zk and its estimation {circumflex over (z)}k be below a predetermined threshold S1.

5. The method of claim 4, wherein the threshold S1 is a function of the uncertainty in the estimation {circumflex over (z)}k.

6. The method of claim 1, wherein the observer is a filter selected from a group consisting of a Kalman filter, an extended Kalman filter, and an unscented Kalman Filter, and wherein the coefficients to be calibrated are coefficients of the covariance matrices R and Q of the filter.

7. The method of claim 6, wherein the R and Q matrices are chosen to be diagonal matrices.

8. A manufacture comprising a non-transitory computer-readable medium having encoded thereon software for calibrating coefficients of an observer of a variable state xk of a physical system from measurements yk of physical quantities of the system, wherein index k identifies an instant of measurement, the software comprising instructions for causing a sensor to measure a variable zk of the physical system at N different instants, the variable zk being a function of a state variable xk, and causing a computer to determine a vector p of coefficients that minimizes the following criterion in complying with a predetermined set Δ of one or more constraints:

9. An apparatus for calibrating coefficients of an observer of a state variable xk of a physical system from measurements yk of physical quantities of the system, wherein index k identifies an instant of measurement, the apparatus comprising: at least one sensor capable of measuring a variable zk of the physical system at N different instants, the variable zk being a function of a state variable xk, and a computer capable of determining a vector p of coefficients that minimizes the following criterion in complying with a predetermined set Δ of one or more constraints: