Presentation Acc 2012

download Presentation Acc 2012

of 19

Transcript of Presentation Acc 2012

  • 8/12/2019 Presentation Acc 2012

    1/19

    INSTITUTEFOR

    SYSTEMSND

    ROBOTICS

    INSTITUTEFOR

    SYSTEMSND

    ROBOTICS

    Set-Valued Observers for Attitude

    and Rate Gyro Bias Estimation

    2012 American Control Conference

    Montral, Canada

    S. Brs P. Rosa C. Silvestre P. Oliveira

    Institute for Systems and Robotics

    Instituto Superior Tcnico

    Lisbon, Portugal

  • 8/12/2019 Presentation Acc 2012

    2/19

    2

    Introduction

    Some estimators rely on stochastic descriptions of the exogenousdisturbances and measurement noise to provide accurate estimates.

    However, in many situations no probability information about

    disturbances and noise is available and only magnitude bounds are

    known.

    In such circumstances, an estimator that computes the set of possible

    states given the sensor information is more suitable.

    Objective: Simultaneously estimate the smallest set containing the attitude

    (rotation matrix) and rate gyro bias.

  • 8/12/2019 Presentation Acc 2012

    3/19

    3

    Introduction

    Discrete attitude kinematic model

    That is an exact description of the physical

    quantities involved. No model identification isrequired.

    Angular velocity measurements

  • 8/12/2019 Presentation Acc 2012

    4/19

    4

    Introduction

    Assumption: there are at least three non-coplanar vector

    observations (noisy) Magnetometer

    Accelerometer

    Star-tracker

    Sun sensor

    The manufactures often provide the maximum noise levels in each

    axis of the sensor.

  • 8/12/2019 Presentation Acc 2012

    5/19

    5

    Set-Valued Observers

    There is uncertainty on the state and on the

    measurements.

    Uncertainty is described by means of polytopes. The dynamic model may be uncertain.

  • 8/12/2019 Presentation Acc 2012

    6/19

    6

    Set-Valued Observers

    Prediction Update

    (model) (measurements)

  • 8/12/2019 Presentation Acc 2012

    7/19

    7

    In our approach we map:

    i.e.,

    Then we have

    Attitude Estimator

  • 8/12/2019 Presentation Acc 2012

    8/19

    8

    Attitude Estimator

    The solution is then the intersection of state polytope in R9with

    SO(3).

    Summarizing:

    SO(3)

    k k+1

    SO(3)

    R9 R9

    No conservatism is added by this operation

  • 8/12/2019 Presentation Acc 2012

    9/19

    9

    The uncertainty in the model is associated with the bias and noise in

    the rate gyro measurements

    Attitude Estimator

  • 8/12/2019 Presentation Acc 2012

    10/19

    10

    Attitude Estimator

    Main advantages:

    The state is guaranteed to be inside the polytope.

    There are no singularities.

    The solution is global, in the sense that it converges for any initial

    conditions.

    The estimates are stable, i.e. the set containing the state does not grow

    unbounded.

    It is based solely on the kinematics, thus it is platform independent and it

    is suitable even if a dynamic model is unknown or inaccurate .

    Main disadvantages:

    Computational cost.

  • 8/12/2019 Presentation Acc 2012

    11/19

    11

    Rate Gyro Bias Estimator

    Goal: Reduce the uncertainty in the bias to improve the attitude

    estimates

    Strategy: to design an SVO for the rate gyro bias

    Polytope: known from the rate gyro

    properties

    Polytope: Computed using R(k) and R(k-1)

    Vector: sensor measurement

  • 8/12/2019 Presentation Acc 2012

    12/19

    12

    Increasing the Convergence Speed

    Divide-to-conquer strategy which consists on:

    Divide the polytope containing the bias, B(k), into sub-polytopes, so

    that only one sub-polytope contains the true bias

    Design SVOs initialized with each of the sub-polytopes

    The SVOs initialized with polytopes not containing the bias will, at

    some time instant, degenerate into empty sets

    When all but one SVO have degenerated, the remaining sub-

    polytope is divided and the process repeats itself

  • 8/12/2019 Presentation Acc 2012

    13/19

    13

    Increasing the Convergence Speed

    With this strategy we bring together the SVO methodology and the ideaof model falsification

    Disadvantages:

    Higher computational cost

    Multi-core and multi-processor systems can be exploited for implementing

    this strategy since each core can be assign to one SVO

  • 8/12/2019 Presentation Acc 2012

    14/19

    14

    Simulation Results

    Angular velocity:

    Noise bounded by 0.115 deg/s

    Vector measurements:

    Noise bounded by 0.01Trajectory

  • 8/12/2019 Presentation Acc 2012

    15/19

    15

    Simulation Results

    The representation of the uncertainty in R9does not provide an intuitive

    measure of the uncertainty on the attitude estimate.

    To obtain uncertainty bounds on more suitable attitude

    representations, different approaches can be explored:

    Exploit the rotation matrix polytope to obtain another polytope containingthe rotation vector.

    To compute bounds on the Euler angles using nonlinear optimization.

    Remark: these operations are conservative but do not add conservatism

    to the estimator.

  • 8/12/2019 Presentation Acc 2012

    16/19

    16

    Simulation Results

    Upper and lower bounds on the attitude estimates

  • 8/12/2019 Presentation Acc 2012

    17/19

    17

    Simulation Results

    Bias upper and lower bounds

  • 8/12/2019 Presentation Acc 2012

    18/19

    18

    Conclusions

    A solution for the problem of attitude estimation is proposed based on set-

    valued observers.

    The observer has no singularities, since the attitude is given by the rotation

    matrix and is global, in the sense that, it is valid for any initial conditions.

    The nonlinearities of the plant are tackled by adding conservatism to the

    estimates.

    The online estimation of the rate gyro bias reduces the uncertainty in the

    angular velocity measurements and consequently in the model.

    Parallel processing can be exploited to increase the estimation rate.

  • 8/12/2019 Presentation Acc 2012

    19/19

    19

    Thank you