Comparison Cannot Be Perfect, Measurements Inherently Include Error

Definition:
Measurement is collection of quantitative data. A measurement is made by comparing a quantity with a standard unit. Since this
comparison cannot be perfect, measurements inherently include error.
Examples:
The length of a piece of string can be measured by comparing the string against a meter stick
Introduction
In a measurement process, measurement data may involve gross errors, which significantly
exceed true values, when a measurement quantity is measured repeatedly without significant
changes in measurement conditions. If the suspicious data remained during data processing, it
will cause incorrect assessment on measurement accuracy due to the distorted measurement
results have been used. Correct identification of gross measurement errors is an important issue
to achieve reliable measurement results.
Gross errors may be reduced by using suitable measurement devices and under appropriate
physical conditions. However, it would be quite difficult to avoid gross errors in measurement
processes having large quantities of data.
Gross error identification has been based on statistics. A number of criteria have been used such
as the 3σ criterion, the Chauvenet criterion, the Grubbs criterion, and the Dixon criterion [1]. The
existing methods [2-10] have been based on a typical distribution such as the Gaussian
distribution and require prior knowledge on the data that the measurement data conform to the
Gauss distribution or conform approximately to it. In a practical measurement, it may be difficult
to have a large quantity of data. Furthermore, the distribution may be found not conforming to
the typical distribution. These may make the statistical methods [2-10] not applicable for gross
error identification and removal.
In order to address the above issues to proceed with gross error identification and subsequently
removal, a new method using the grey system theory is proposed. The advantages of the
proposed new method is that the measurement data is not required to conform to a particular
probability density distribution and the sampling size of the data does not need to be large. The
principle of the gross error identification is presented and an identification criterion proposed. A
case study is provided to demonstrate the effectiveness of the proposed grey system method. The
proposed method should be a convenient and useful tool to identify gross errors in a precision
measurement process
Resolution is the minimum increment to which a measurement can be made. For us, this is a design parameter of
the X-Y stage motion encoders or galvo.
Accuracy is how close a measurement is to the "true" value. "Truth" is the traceability of the measurement to a
primary standard at NIST, a USA governmental agency, National Institute for Standards & Technology. For us, this
is conformance to our customer's metrology.
Precision is repeatability. For us, precision is half the increase in laser line width or hole diameter when the laser
is run repeatedly over the same part a number of times.
All are plus/minus numbers except resolution.
Correct identification and elimination of gross errors in plant measurements is a key factor
for the performance of industrial on-line optimization. For rigorous nonlinear plant models,
the gross error detection problem is very challenging. This paper presents two data
reconciliation and gross error detection strategies for nonlinear models: one based on a serial
elimination algorithm using a linearized model and another one based on the Tjoa-Biegler
contaminated normal distribution approach. The comparison is based upon the results
obtained with a commercial on-line data reconciliation and optimization package using a
rigorous model of two industrial refinery processes.
1. INTRODUCTION
Data reconciliation is widely used to adjust the plant data and provide estimates for
unmeasured variables and parameters. Data reconciliation improves the accuracy in measured
variables and model parameters by exploiting redundancy in the measured data. Traditional
data reconciliation assumes that only random errors exist in process data. If gross errors also
occur, they need to be identified and eliminated. Otherwise, the reconciled solution will be
highly biased. Since data reconciliation is often used to provide better starting points to the
economic optimization, it is very important that gross errors will not significantly affect the
quality of reconciled data.
Statistical hypothesis testing techniques have been employed to detect persistent gross
errors [1]. However, correctly identifying all gross errors is still a challenging task, even for
steady state models. The existing strategies based on statistical tests sometimes wrongly
report a different number or location of gross errors than in reality. This problem occurs for
the following two reasons. First, a gross error may propagate in data reconciliation and
contaminate the reconciled data. Second, redundancy in measured data for a chemical or
refinery process may be low compared to the size of the process.
Most gross error detection and identification strategies have been designed for linear data
reconciliation models, such as plant mass flow balances. A linear data reconciliation problem
can be written as a constrained weighted least-squares minimization problem:
min F - (x- xD T y-1 (x- ~) - (x~ - xi )2
X i--1 O- i (1)
s.t. B(~ + a)- c

Comparison Cannot Be Perfect, Measurements Inherently Include Error

Uploaded by

Copyright:

Available Formats

Comparison Cannot Be Perfect, Measurements Inherently Include Error

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Comparison Cannot Be Perfect, Measurements Inherently Include Error

Uploaded by

Copyright:

Available Formats

Definition:

to achieve reliable measurement results.

processes having large quantities of data.

Gauss distribution or conform approximately to it. In a practical measurement, it may be difficult

error identification and removal.

All are plus/minus numbers except resolution.

rigorous model of two industrial refinery processes.

quality of reconciled data.

refinery process may be low compared to the size of the process.

can be written as a constrained weighted least-squares minimization problem:

min F - (x- xD T y-1 (x- ~) - (x~ - xi )2

s.t. B(~ + a)- c

You might also like