SUBJECT CODE

TYPE THE SUBJECT NAME HERE

UNIT NO 2

SYSTEM IDENTIFICATION

● Parametric methods- Least square error

Recursive Least square Error
III VI

EI8691
COMPUTER CONTROL OF PROCESSES
 EI8691
COMPUTER CONTROL OF PROCESSES

LEAST-SQUARES TECHNIQUE

We will be interested in the identification of systems with constant parameters

which will form the parameter vector θ, which will be related to the measurements
by the linear vector relation

where d, is a p-vector of data corresponding to measurement y, and

θ is a p-vector of parameters which we wish to find based on the data y, and diT
 EI8691
COMPUTER CONTROL OF PROCESSES

LEAST-SQUARES TECHNIQUE

Let us define our estimate of the parameter vector as which may

differ from the actual parameter values of θ. The set of equations then are

where e, is the error induced because we have employed the estimate of

θ. The set of n equations of can be written as the vector equation

where D is the n x p data matrix with an ith row of diT and

y is commonly called the n x 1 measurement vector and
e is an n x 1 vector of errors.
 EI8691
COMPUTER CONTROL OF PROCESSES

LEAST-SQUARES TECHNIQUE

.There are many practical problems that we can put into the form of which is the
curve-fitting problem

An example of least-squares curve fitting involves the set of experimental data

shown if figure which we wish to fit with a straight line, or
 EI8691
COMPUTER CONTROL OF PROCESSES

In this case the parameter vector is two-dimensional and contains the slope
and y intercept of the straight line, or

The n relations can be written in matrix form as

In this example the number of measurements (n) must be greater than the
number of parameters (p), which in this case is two.
 EI8691
COMPUTER CONTROL OF PROCESSES

The technique of least squares is not restricted to models that are linear
in the independent variable x, but the models must be linear in the parameters,
as is this one in m and b.
Polynomial curve fits are common, but note that these fits are linear in the
coefficients which are the parameters.
The model of the estimator is
y = Dθ
while in reality

where the e vector represents the errors in the estimates of the points y,.
 EI8691
COMPUTER CONTROL OF PROCESSES

We want to minimize the sum of the squares of the errors or minimize the
scalar function

We see that the error vector is the difference of y and the predicted value of y

These two terms are the same, and hence it is sufficient that one of them
be zero and hence
 EI8691
COMPUTER CONTROL OF PROCESSES

RECURSIVE LEAST-SQUARES TECHNIQUE

• The weighted least square calculation for θ is referred as a batch

calculation, because by the definition of several entries the formula
presumes that one has a batch of data of length N from which the normal
equations are solved
• There are times when the data are acquired sequentially rather than in a
batch and otherwise one wishes to examine the nature of solution as more
data are included to see whether perhaps some improvement in the
parameter estimation continues to be made or whether any sudden change
in θ or persistent drift in one or more of the parameters are needed
• In short RLS is used if one wishes something to do a visual or experimental
examination of new estimates as one or several more data points are
computed
 EI8691
COMPUTER CONTROL OF PROCESSES

RECURSIVE LEAST-SQUARES TECHNIQUE

Features of Recursive identification methods

• They are central part of adaptive system where the action is based on the
most recent model

• Their requirement on primary memory is quite modest since not all data
are stored

• They can be easily modified into real time algorithms armed at tracking
time varying parameters

• They acan be the first step in fault detection algorithm which is used to find
out whether the system has changed significantly

• Most adaptive systems are based on recursive identification It requires

small primary memory since only modest amount of information is stored
this amount will not increase the time
 EI8691
COMPUTER CONTROL OF PROCESSES

RECURSIVE LEAST-SQUARES TECHNIQUE

Consider the model

y(t) =b + e(t)
where e(t) denotes disturbance of variance lambda
The least square estimates of θ=b is given by
y(t) = b
ɸ(t) = 1
θ = 1/N [ y(1)+…….y(N)]
This expression is simply the arithmetic mean of all measurements

This expression can be formulated as a recursive algorithm

 EI8691
COMPUTER CONTROL OF PROCESSES

RECURSIVE LEAST-SQUARES TECHNIQUE

The estimates of θ at time is equal to previous estimates at time (t-1) plus a

corrective term The correction term is proportional to the derivation of
predicted value from what is actually observed at time t, namely y(t).

Morover the prediction error is weighed by the the factor 1 / t. which means
that the magnitude of changes of estimates will decrease with time Instead
the old information as condensed in the estimate will become more reliable

The variance of θ(t) is given neglecting lambda

P(t) = (ɸTɸ)-1 = 1 / t.
 EI8691
COMPUTER CONTROL OF PROCESSES

RECURSIVE LEAST-SQUARES TECHNIQUE