Note: This page has been translated by MathWorks. Click here to see

To view all translated materials including this page, select Country from the country navigator on the bottom of this page.

To view all translated materials including this page, select Country from the country navigator on the bottom of this page.

After you use the Extended Kalman Filter, Unscented Kalman
Filter or Particle Filter blocks for online state estimation of a
nonlinear system, validate the estimation before deploying the code in your application. If
the validation indicates low confidence in the estimation, then see Troubleshoot Online State Estimation for next steps. After
you have validated the online estimation results, you can generate C/C++ code for the blocks
using Simulink^{®} Coder™ software.

To validate the performance of your filter, perform state estimation using measured or simulated output data from these scenarios.

Obtain output data from your system at different operating conditions and input values — To ensure that estimation works well under all operating conditions of interest. For example, suppose that you want to track the position and velocity of a vehicle using noisy position measurements. Measure the data at different vehicle speeds and slow and sharp maneuvers.

For each operating condition of interest, obtain multiple sets of experimental or simulated data with different noise realizations — To ensure that different noise values do not deteriorate estimation performance.

For each of these scenarios, test the filter performance by examining the residuals and state estimation error.

The residual, or output estimation error, is the difference
between the measured system output `yMeasured[k]`

,
and the estimated system output `yPredicted[k|k-1]`

at
time step `k`

. Here, `yPredicted[k|k-1]`

is
the estimated output at time step `k`

, which is predicted
using output measurements until time step `k-1`

.

The blocks do not explicitly output `yPredicted[k|k-1]`

,
however you can compute the output using the estimated state values
and your state transition and measurement functions. For an example,
see Compute Residuals and State Estimation Errors.

The residuals must have the following characteristics:

Small magnitude — Small errors relative to the size of the outputs increase confidence in the estimated values.

Zero mean

Low autocorrelation, except at zero time lag — To compute the autocorrelation, you can use the Autocorrelation block from DSP System Toolbox™ software.

When you simulate the output data of your nonlinear system and
use that data for state estimation, you know the true state values.
You can compute the errors between estimated and true state values
and analyze the errors. The estimated state value at any time step
is output at the **xhat** port of the blocks. The
state estimation errors must satisfy the following characteristics:

Small magnitude

Zero mean

Low autocorrelation, except at zero time lag

You can also compute the covariance of the state estimation
error, and compare it to the state estimation error covariance that
is output by the blocks in the **P** port of the
blocks. Similar values increase confidence in the performance of the
filter.

This example shows how to estimate the states of a discrete-time Van der Pol oscillator and compute state estimation errors and residuals for validating the estimation. The residuals are the output estimation errors, that is, they are the difference between the measured and estimated outputs.

In the Simulink™ model, the Van der Pol Oscillator block implements the oscillator with nonlinearity parameter, mu, equal to 1. The oscillator has two states. A noisy measurement of the first state `x1`

is available.

The model uses the Unscented Kalman Filter block to estimate the states of the oscillator. Since the block requires discrete-time inputs, the Rate Transition block samples `x1`

to give the discretized output measurement `yMeasured[k]`

at time step `k`

. The Unscented Kalman Filter block outputs the estimated state values `xhat[k|k]`

at time step `k`

, using `yMeasured`

until time `k`

. The filter block uses the previously written and saved state transition and measurement functions, `vdpStateFcn.m`

and `vdpMeasurementFcn.m`

. For information about these functions, see Nonlinear State Estimation Using Unscented Kalman Filter and Particle Filter.

To validate the state estimation, the model computes the residuals in the Generate Residual block. In addition, since the true state values are known, the model also computes the state estimation errors.

To compute the residuals, the Generate Residual block first computes the estimated output `yPredicted[k|k-1]`

using the estimated states and state transition and measurement functions. Here, `yPredicted[k|k-1]`

is the estimated output at time step `k`

, predicted using output measurements until time step `k-1`

. The block then computes the residual at time step `k`

as `yMeasured[k]`

- `yPredicted[k|k-1]`

.

Examine the residuals and state estimation errors, and ensure that they have a small magnitude, zero mean, and low autocorrelation.

In this example, the Unscented Kalman Filter block outputs `xhat[k|k]`

because the **Use the current measurements to improve state estimates** parameter of the block is selected. If you clear this parameter, the block instead outputs `xhat[k|k-1]`

, the predicted state value at time step `k`

, using `yMeasured`

until time `k-1`

. In this case, compute `yPredicted[k|k-1] = MeasurementFcn(xhat[k|k-1])`

, where `MeasurementFcn`

is the measurement function for your system.

Extended Kalman Filter | Kalman Filter | Particle Filter | Unscented Kalman Filter