recursiveOE

Online parameter estimation of output-error polynomial model

Description

Use the recursiveOE System object™ for parameter estimation with real-time data using an output-error model structure. If all the data you need for estimation is available at once and you are estimating a time-invariant model, use the offline estimation function oe.

To perform parameter estimation with real-time data:

1. Create the recursiveOE object and set its properties.

2. Call the object with arguments, as if it were a function.

To learn more about how System objects work, see What Are System Objects?

Creation

Syntax

oeobj = recursiveOE

oeobj = recursiveOE([nb,nf,nk])

oeobj = recursiveOE([na,nb,nf],B0,F0)

oeobj = recursiveOE(___,Name=Value)

Description

oeobj = recursiveOE creates a System object for online parameter estimation of a default single-output output-error model. The default model structure has a polynomial of order 1 and initial polynomial coefficient values eps.

oeobj = recursiveOE([nb,nf,nk]) sets the orders of polynomials B and F to nb and nf, respectively. This syntax also sets the input-output delay to nk.

oeobj = recursiveOE([na,nb,nf],B0,F0) specifies the initial coefficient values of polynomials B and F by setting the InitialB property to B0, and the InitialF property to F0. Specify initial values to potentially avoid local minima during estimation. If the initial values are small compared to the default InitialParameterCovariance property value and you have confidence in your initial values, specify a smaller InitialParameterCovariance.

oeobj = recursiveOE(___,Name=Value) specifies one or more properties of the model structure or recursive estimation algorithm using name-value arguments. For example, oeobj = recursiveOE(2,EstimationMethod="NormalizedGradient") creates an estimation object that uses a normalized gradient estimation method.

Before R2021a, use commas to separate each name and value, and enclose Name in quotes. For example, oeobj = recursiveOE(2,"EstimationMethod","NormalizedGradient") creates an estimation object that uses a normalized gradient estimation method.

Input Arguments

nb — Order of polynomial B(q) + 1

1 (default) | nonnegative integer

Order of polynomial B(q) + 1, specified as a nonnegative integer.

nf — Order of polynomial C(q)

1 (default) | nonnegative integer

Order of polynomial C(q), specified as a nonnegative integer.

nk — Input-output delay

1 (default)

Input-output delay, specified as a nonnegative integer. nk is number of input samples that occur before the input affects the output.

Properties

Unless otherwise indicated, properties are nontunable, which means you cannot change their values after calling the object. Objects lock when you call them, and the release function unlocks them.

If a property is tunable, you can change its value at any time.

For more information on changing property values, see System Design in MATLAB Using System Objects.

B — Estimated coefficients of polynomial B(q)

Read-only: [] (default) | row vector

This property is read-only.

Estimated coefficients of the polynomial B(q), returned as a row vector. The elements of this vector appear in order of ascending powers of q^-1.

B is initially empty when you create the object and is populated after you run the online parameter estimation.

F — Estimated coefficients of polynomial A(q)

Read-only: [] (default) | row vector

This property is read-only.

Estimated coefficients of the polynomial F(q), returned as a row vector. The elements of this vector appear in order of ascending powers of q^-1.

F is initially empty when you create the object and is populated after you run the online parameter estimation.

InitialB — Initial coefficients of polynomial B(q)

[0 eps] (default) | row vector

Initial values for the coefficients of polynomial B(q) of order nb – 1, specified as a row vector of length nb + nk, with nk leading zeros. nk is the input-output delay. Specify the coefficients in order of ascending powers of q^-1.

If the initial guesses are much smaller than the default InitialParameterCovariance, 10000, the initial guesses are given less importance during estimation. In that case, specify a smaller initial parameter covariance.

Tunable: Yes

InitialF — Initial coefficients of polynomial F(q)

[1 eps] (default) | row vector

Initial values for the coefficients of polynomial F(q) of order nf, specified as a row vector of length nf +1, with 1 as the first element. Specify the coefficients in order of ascending powers of q^-1.

The coefficients in InitialF must define a stable discrete-time polynomial with roots within a unit circle.

If the initial guesses are much smaller than the default InitialParameterCovariance, 10000, the initial guesses are given less importance during estimation. In that case, specify a smaller initial parameter covariance.

InitialOutputs — Initial values of output buffer in finite-history estimation

0 (default) | column vector of real values

Initial values of the output buffer in finite-history estimation, specified as 0 or a (W + nf)-by-1 vector, W is equal to WindowLength.

Use this property to control the initial behavior of the algorithm.

When you set InitialOutputs to 0, the object populates the buffer with zeros.

If you set the initial buffer to 0 or if the buffer does not contain enough information, you see a warning message during the initial phase of your estimation. The warning usually clears after a few cycles. The number of cycles required to buffer sufficient information depends on the order of your polynomials and your input delays. If the warning persists, evaluate the content of your signals.

Tunable: Yes

Dependencies

To enable this property, set History to 'Finite'.

InitialInputs — Initial values of input buffer in finite-history estimation

0 (default) | column vector of real values

Initial values of the inputs in the finite history window, specified as 0 or as a (W-1+max(nb)+max(nk))-by-N_uu matrix, where W is equal to WindowLength and N_u is the number of inputs.

Use this property to control the initial behavior of the algorithm.

When you set InitialInputs to 0, the object populates the buffer with zeros.

If the initial buffer is set to 0 or does not contain enough information, you see a warning message during the initial phase of your estimation. The warning should clear after a few cycles. The number of cycles it takes for sufficient information to be buffered depends upon the order of your polynomials and your input delays. If the warning persists, you should evaluate the content of your signals.

Dependencies

To enable this property, set History to 'Finite'.

ParameterCovariance — Estimated covariance

Read-only: [] (default) | N_p-by-N_p symmetric positive-definite matrix

This property is read-only.

Estimated covariance P of the parameters, stored as an N_p-by-N_p symmetric positive-definite matrix, where N_p is the number of parameters to be estimated. The software computes P assuming that the residuals (difference between estimated and measured outputs) are white noise and the variance of these residuals is 1.

The interpretation of P depends on your settings for the History and EstimationMethod properties.

• If you set History to 'Infinite' and EstimationMethod to:

  • 'ForgettingFactor' — R₂ * P is approximately equal to twice the covariance matrix of the estimated parameters, where R₂ is the true variance of the residuals.

  • 'KalmanFilter' — R₂ * P is the covariance matrix of the estimated parameters, and R₁ /R₂ is the covariance matrix of the parameter changes. Here, R₁ is the covariance matrix that you specify in ProcessNoiseCovariance.

• If History is 'Finite' (sliding-window estimation) — R₂P is the covariance of the estimated parameters. The sliding-window algorithm does not use this covariance in the parameter-estimation process. However, the algorithm does compute the covariance for output so that you can use it for statistical evaluation.

ParameterCovariance is initially empty when you create the object and is populated after you run the online parameter estimation.

Dependencies

To enable this property, use one of the following configurations:

• Set History to 'Finite'.

• Set History to 'Infinite' and set EstimationMethod to either 'ForgettingFactor' or 'KalmanFilter'.

InitialParameterCovariance — Covariance of initial parameter estimates

10000 (default) | positive scalar | vector of positive scalars | symmetric positive-definite matrix

Covariance of the initial parameter estimates, specified as one of these values:

• Real positive scalar α — Covariance matrix is an N-by-N diagonal matrix in which α is each diagonal element. N is the number of parameters to be estimated.

• Vector of real positive scalars [α₁,...,α_N] — Covariance matrix is an N-by-N diagonal matrix in which α₁ through α_N] are the diagonal elements.

• N-by-N symmetric positive-definite matrix.

InitialParameterCovariance represents the uncertainty in the initial parameter estimates. For large values of InitialParameterCovariance, the software accords less importance to the initial parameter values and more importance to the measured data during the beginning of estimation.

Tunable: Yes

Dependency

To enable this property, use one of the following configurations:

• Set History to 'Finite'.

• Set History to 'Infinite' and set EstimationMethod to either 'ForgettingFactor' or 'KalmanFilter'.

EstimationMethod — Recursive estimation algorithm

'ForgettingFactor' (default) | 'KalmanFilter' | 'NormalizedGradient' | 'Gradient'

Recursive estimation algorithm used for online estimation of model parameters, specified as one of the following:

• 'ForgettingFactor' — Use forgetting factor algorithm for parameter estimation.

• 'KalmanFilter' — Use Kalman filter algorithm for parameter estimation.

• 'NormalizedGradient' — Use normalized gradient algorithm for parameter estimation.

• 'Gradient' — Use unnormalized gradient algorithm for parameter estimation.

Forgetting factor and Kalman filter algorithms are more computationally intensive than gradient and unnormalized gradient methods. However, the former algorithms have better convergence properties. For information about these algorithms, see Recursive Algorithms for Online Parameter Estimation.

Dependencies

To enable this property, set History to 'Infinite'.

ForgettingFactor — Forgetting factor for parameter estimation

1 (default) | scalar in the range (0, 1]

Forgetting factor λ for parameter estimation, specified as a scalar in the range (0, 1].

Suppose that the system remains approximately constant over T₀ samples. You can choose λ to satisfy this condition:

$$T_0=\frac{1}{1-\lambda }$$

• Setting λ to 1 corresponds to "no forgetting" and estimating constant coefficients.

• Setting λ to a value less than 1 implies that past measurements are less significant for parameter estimation and can be "forgotten". Set λ to a value less than 1 to estimate time-varying coefficients.

Typical choices of λ are in the range [0.98, 0.995].

Dependencies

To enable this property, set History to 'Infinite' and set EstimationMethod to 'ForgettingFactor'.

Tunable: Yes

EnableAdapation — Option to enable or disable parameter estimation

true (default) | false

Option to enable or disable parameter estimation, specified as one of the following:

• true — The step function estimates the parameter values for that time step and updates the parameter values.

• false — The step function does not update the parameters for that time step and instead outputs the last estimated value. You can use this option when your system enters a mode where the parameter values do not vary with time.

  Note

  If you set EnableAdapation to false, you must still execute the step command. Do not skip step to keep parameter values constant, because parameter estimation depends on current and past I/O measurements. step ensures past I/O data is stored, even when it does not update the parameters.

Tunable: Yes

DataType — Floating point precision of parameters

Read-only: 'double' (default) | 'single'

This property is read-only.

Floating point precision of parameters, specified as one of the following values:

• 'double' — Double-precision floating point

• 'single' — Single-precision floating point

Setting DataType to 'single' saves memory but leads to loss of precision. Specify DataType based on the precision required by the target processor where you will deploy generated code.

You must set DataType during object creation using a name-value argument.

ProcessNoiseCovariance — Covariance matrix of parameter variations

0.1 (default) | nonegative scalar | vector of nonegative values | N-by-N symmetric positive semidefinite matrix

Covariance matrix of parameter variations, specified as one of the following:

• Real nonnegative scalar, α — Covariance matrix is an N-by-N diagonal matrix, with α as the diagonal elements.

• Vector of real nonnegative scalars, [α₁,...,α_N] — Covariance matrix is an N-by-N diagonal matrix, with [α₁,...,α_N] as the diagonal elements.

• N-by-N symmetric positive semidefinite matrix.

N is the number of parameters to be estimated.

The Kalman filter algorithm treats the parameters as states of a dynamic system and estimates these parameters using a Kalman filter. ProcessNoiseCovariance is the covariance of the process noise acting on these parameters. Zero values in the noise covariance matrix correspond to estimating constant coefficients. Values larger than 0 correspond to time-varying parameters. Use large values for rapidly changing parameters. However, the larger values result in noisier parameter estimates.

Tunable: Yes

Dependencies

To enable this property, set History to 'Infinite' and set EstimationMethod to 'KalmanFilter'.

AdaptationGain — Adaptation gain

1 (default) | positive scalar

Adaptation gain, γ, used in gradient recursive estimation algorithms, specified as a positive scalar.

Specify a large value for AdaptationGain when your measurements have a high signal-to-noise ratio.

Tunable: Yes

Dependencies

To enable this property, set History to 'Infinite' and set EstimationMethod to either 'Gradient' or 'NormalizedGradient'.

NormalizationBias — Bias in adaptation gain scaling

eps (default) | nonegative scalar

Bias in adaptation gain scaling used in the 'NormalizedGradient' method, specified as a nonnegative scalar.

The normalized gradient algorithm divides the adaptation gain at each step by the square of the two-norm of the gradient vector. If the gradient is close to zero, this division can cause jumps in the estimated parameters. NormalizationBias is the term introduced in the denominator to prevent such jumps. If you observe jumps in estimated parameters, increase NormalizationBias.

Tunable: Yes

Dependencies

To enable this property, set History to 'Infinite' and set EstimationMethod to 'NormalizedGradient'.

History — Data history type

Read-only: 'Infinite' (default) | 'Finite'

This property is read-only.

Data history type, which defines the type of recursive algorithm to use, specified as one of the following:

• 'Infinite' — Use an algorithm that aims to minimize the error between the observed and predicted outputs for all time steps from the beginning of the simulation.

• 'Finite' — Use an algorithm that aims to minimize the error between the observed and predicted outputs for a finite number of past time steps.

Algorithms with infinite history aim to produce parameter estimates that explain all data since the start of the simulation. These algorithms still use a fixed amount of memory that does not grow over time. To select an infinite-history algorithm, use EstimationMethod.

Algorithms with finite history aim to produce parameter estimates that explain only a finite number of past data samples. This method is also called sliding-window estimation. The object provides one finite-history algorithm. To define the window size, specify the WindowLength property.

For more information on recursive estimation methods, see Recursive Algorithms for Online Parameter Estimation.

You must set History during object creation using a name-value argument.

WindowLength — Window size

Read-only: 200 (default) | positive integer

This property is read-only.

Window size for finite-history estimation, specified as a positive integer indicating the number of samples.

Choose a window size that balances estimation performance with computational and memory burden. Sizing factors include the number and time variance of the parameters in your model. WindowLength must be greater than or equal to the number of estimated parameters.

Suitable window length is independent of whether you are using sample-based or frame-based input processing (see InputProcessing). However, when using frame-based processing, your window length must be greater than or equal to the number of samples (time steps) contained in the frame.

You must set WindowLength during object creation using a name-value argument.

Dependencies

To enable this property, set History to 'Finite'.

InputProcessing — Input processing method

Read-only: 'Sample-based' (default) | 'Frame-based'

This property is read-only.

Input processing method, specified as one of the following:

• 'Sample-based' — Process streamed signals one sample at a time.

• 'Frame-based' — Process streamed signals in frames that contain samples from multiple time steps. Many machine sensor interfaces package multiple samples and transmit these samples together in frames. 'Frame-based' processing allows you to input this data directly without having to first unpack it.

The InputProcessing property impacts the dimensions for the input and output signals when using the recursive estimator object.

• Sample-based

  • y,u, and estimatedOutput are scalars.

• • Frame-based with M samples per frame

    • y,u, and estimatedOutput are M-by-1 vectors.

You must set InputProcessing during object creation using a name-value argument.

Usage

Syntax

[B,F,estimatedOutput] = oeobj(y)

Description

[B,F,estimatedOutput] = oeobj(y) updates and returns the parameters and output of recursiveOE model oeobj online based on real-time output data y and input data u.

Input Arguments

y — Output data

real scalar

Output data acquired in real time, specified as a real scalar.

u — Input data

real scalar

Input data acquired in real time, specified as a real scalar.

Output Arguments

estimatedOutput — Estimated output

real scalar

Estimated output, returned as a real scalar. The output is estimated using input-output estimation data, current parameter values, and the recursive estimation algorithm specified in the recursiveOE System object.

Object Functions

To use an object function, specify the System object as the first input argument. For example, to release system resources of a System object named obj, use this syntax:

release(obj)

Common to All System Objects

step	Run System object algorithm
release	Release resources and allow changes to System object property values and input characteristics
reset	Reset internal states of System object

Specific to recursiveOE

clone	Create duplicate System object
isLocked	Determine if System object is in use

Examples

Estimate Output-Error Polynomial Model Online

Create a System object for online parameter estimation of a Output-Error polynomial model using recursive estimation algorithms.

obj = recursiveOE;

The Output-Error model has a default structure with polynomials of order 1 and initial polynomial coefficient values, eps.

Load the estimation data. In this example, use a static data set for illustration.

load iddata1 z1;
output = z1.y;
input = z1.u;

Estimate Output-Error model parameters online using step.

for i = 1:numel(input)
[B,F,EstimatedOutput] = step(obj,output(i),input(i));
end

View the current estimated values of polynomial F coefficients.

obj.F

ans = 1×2

    1.0000   -0.7618

View the current covariance estimate of the parameters.

obj.ParameterCovariance

ans = 2×2

    0.0024    0.0002
    0.0002    0.0001

View the current estimated output.

EstimatedOutput

EstimatedOutput = 
-4.1866

Create System Object for Output-Error Model with Known Orders and Delays

Specify Output-Error polynomial model orders and delays.

nb = 1;
nf = 2;
nk = 1;

Create a System object for online estimation of Output-Error polynomial model with the specified orders and delays.

obj = recursiveOE([nb nf nk]);

Create System Object for Output-Error Model with Known Initial Parameters

Specify Output-Error polynomial model orders and delays.

nb = 1;
nf = 2;
nk = 1;

Create a System object for online estimation of Output-Error model with known initial polynomial coefficients.

B0 = [0 1];
F0 = [1 0.7 0.8];
obj = recursiveOE([nb nf nk],B0,F0);

Specify the initial parameter covariance.

obj.InitialParameterCovariance = 0.1;

InitialParameterCovariance represents the uncertainty in your guess for the initial parameters. Typically, the default InitialParameterCovariance (10000) is too large relative to the parameter values. This results in initial guesses being given less importance during estimation. If you have confidence in the initial parameter guesses, specify a smaller initial parameter covariance.

Specify Estimation Method for Online Estimation of Output-Error Model

Create a System object that uses the unnormalized gradient algorithm for online parameter estimation of an Output-Error model.

obj = recursiveOE([1 2 1],'EstimationMethod','Gradient');

More About

Output-Error Model Structure

The general output-error model structure is:

$$y(t)=\frac{B(q)}{F(q)}u(t-n_k)+e(t)$$

The orders of the output-error model are:

$$\begin{array}{l}nb\text{: }B(q)=b_1+b_2{q}^{-1}+\mathrm{...}+b_{nb}{q}^{-nb+1}\\ nf\text{: }F(q)=1+f_1{q}^{-1}+\mathrm{...}+f_{nf}{q}^{-nf}\end{array}$$

Extended Capabilities

C/C++ Code Generation

Generate C and C++ code using MATLAB® Coder™.

Usage notes and limitations:

• For Simulink-based workflows, use Recursive Polynomial Model Estimator.

• For limitations, see Generate Code for Online Parameter Estimation in MATLAB.

Supports MATLAB Function block: No

Version History

Introduced in R2015b