International On-Line Coal Analyzer Technical Conference

St. Louis, Missouri, November 8-10, 2004

What Is Involved In Effective Calibration Of A PGNA Analyzer

In The Factory And The Field
Steve Foster1
______________________________________________________________________________

ABSTRACT: This paper takes a critical look at those things necessary to calibrate a nuclear based
elemental analyzer in both the factory and the field to ensure a successful installation. Customers who buy
modern on-line nuclear elemental analyzers need to ensure that the performance of their purchase can be
performance verified. The verification of performance is a topic unto itself. This paper takes a look at the
proper calibration techniques of the analyzer that help lead to a successful performance verification.
Modern PGNA analyzers are factory calibrated with generic artificial coal standards that range from
anthracite to lignite chemistries with 0-4% sulfur and 5-40% ash levels. This paper takes a detailed look at
what goes into a meaningful orthogonal factory calibration as well as the use of reference standards in the
field to enable the customer to start using his analyzer quickly. It also stresses the need to have dynamic
field data available in order to ensure the analyzer is adding value to the operation as it relates to the user's
customers and to facilitate ongoing analyzer improvement. Fundamental issues in factory and field
calibration and some guidelines for successful calibration and therefore a successful installation are
examined. The material is written around on-belt analyzer technology for coal operations.

Introduction
According to the dictionary definition below, a calibration is an adjustment against some kind of
standard. When most customers think of calibration they think of adjusting the gain or offset of
an instrument so that an instrument output coincides with some standard, such as the results from
a lab on a representative sample gathered from somewhere in the operation.

cal·i·brate (k l -br t )
v. cal·i·brat·ed, cal·i·brat·ing, cal·i·brates

1. To check, adjust, or determine the graduations of a quantitative

measuring instrument by comparison with a standard.
2. To determine the caliber of a tube.
3. To make corrections in or adjust a procedure or process.

Vendors of nuclear based elemental analyzers use the word calibrate to mean at least two very
different things. In the typical nuclear elemental analyzer factory the word calibrate means not
only the adjustment of gains and offsets but also, and perhaps more importantly, it refers to the
use of well characterized orthogonal static reference standards to determine how basic elemental
results will be reported to the customer. In the field, calibration has historically referred only to
the adjustment of a gain and offset, but now can also include the optimization of the way in
which elemental information is reported.

This paper will take an in-depth look at both aspects of calibration of a nuclear elemental
analyzer, beginning with a basic look at how these analyzers work.
1
Steve Foster is Vice President of Marketing, Sales, and Service at Sabia, Inc. 7944 Convoy Court, San Diego, CA
92111. Tel: 858-279-4000. Email: sfoster@sabiainc.com.
 2 FOSTER – INTERNATIONAL ON-LINE COAL ANALYZER TECHNICAL CONFERENCE

How the Analyzers Work

When a bulk material such as coal is bombarded with thermal neutrons, (<1 electron volt neutron
energy), from a Californium 252 nuclear source, many of the neutrons are captured by elemental
atoms within the coal. When this happens the atom becomes temporarily unstable. In order to
re-stabilize the atom sheds a spectrum of high-energy gamma rays. The specific energies of
gamma rays given off are a unique set for each of the elements within the periodic table. This
principle makes it possible to create a signal to enable the on-line elemental analysis of coal
possible with PGNA.

In order to create an electronic signal used for the determination of the weight percent of the
elements of interest within the coal the unique elemental signature gamma rays resulting from
the capture of neutrons by elemental atoms are detected by a scintillating crystal such as Sodium
Iodide (NaI). As the gamma rays penetrate the detector they deposit their energy as high-speed
electrons within the crystal. These electrons create ionization, which can be detected as UV light
pulses. The light pulses are in turn detected by photo-multiplier tubes (a vacuum tube electronic
component operating at a high voltage, typically 500 to 1000 VDC) and turned into electrical
pulses which are immediately amplified, shaped and then converted into digital signals, and
collected into a spectrum over some predetermined period of time (typically one minute) which
can then be processed by a computer at very high speeds.

Nucleus of Nucleus of Nucleus of Elemental

Thermal Stable Elemental Excited Elemental Atom Stabilized by
Neutron Atom Atom with Captured Release of Gamma Spectrum
Neutron Neutron still captured inside

γ
Figure 1. The Nuclear Physics of PGNAA

The resulting gamma-ray spectrum collected over a one-minute period is actually a histogram of
all the incoming gamma-ray energy levels ranging from zero to ten Mev (Million electron volts).
In coal anywhere from five to fifteen elements are represented in the spectrum. A typical
spectrum is shown below which over one minute collects several million pulses.
3 FOSTER – INTERNATIONAL ON-LINE COAL ANALYZER TECHNICAL CONFERENCE

Typical PGNAA Gamma-Ray Spectra

5000000

@ 2.223 Mev
4500000

Hydrogen
4000000

3500000
Total Counts

3000000

2500000

2000000

1500000

1000000

500000

0
0.5 1.5 2.5 3.5 4.5 5.5 6.5 7.5 8.5
Mev

Figure 2. A Typical Gamma Ray Spectrum – High and Low Energies

Typical PGNAA Gamma-Ray Spectra

400000

350000
Silicon

300000
Calcium

Calcium
Iron

250000
Total Counts

Silicon

200000

150000

100000
Iron

Aluminum

50000

0
0.5 1.5 2.5 3.5 4.5 5.5 6.5 7.5 8.5
Mev

Figure 3. A Typical Gamma Ray High Energy Spectrum

 4 FOSTER – INTERNATIONAL ON-LINE COAL ANALYZER TECHNICAL CONFERENCE
Optimum use of the information in the gamma-ray spectrum requires integration of the area
under the peak. Most commonly, vendors use a full spectrum analysis such as Library Least
Squares that utilizes the instrument response to pure elements used as a library against which the
incoming spectral data can be compared on a minute-by-minute basis. Typically a multiple
linear regression technique is used which solves a linear matrix equation with matrix inverse
math. With the high speed and data capacity of computers available today the time required for
this mathematical treatment (de-convolution of the spectra) of the data takes only seconds and
becomes transparent to the end user. Prior to presentation of the final answers to the customer
the results of the multiple linear- regression are normalized with respect to each other. The
technology has made significant strides and now offers the marketplace impressive precisions
and accuracies. Today’s PGNA analyzers are calibrated in the factory with an orthogonal set of
synthetic coal reference standards arrive at the site initially calibrated for the universe of
possibilities in coal. This means that the analyzers should be nearly immune to changes in coal
types. An analyzer properly calibrated in the factory can handle any number of different seams
at a site, or if installed at a utility, can handle any number of different supplier’s coal types
coming into the coal yard. It should be noted that PGNA analyzers utilizing thermal neutrons for
signal derivation are limited to a BTU calculation using an empirical ash/moisture dilution
relationship that assumes a constant moisture and ash free heating value (MAF BTU) and relies
on an accurate measurement of moisture by a microwave moisture meter. Microwave moisture
meters are susceptible to variations in bound moisture of the coal.

MLR Example Calculation – Taking the Mystery out of PGNA

Step 1. The analyzer collects data each minute from the interrogated material and prepares to
perform a multiple linear regression (MLR) using its library of elements against the coal sample
gamma spectra. See Figure 4 below:

A PGNA Spectra and Elemental Responses

400000

350000
Spectra
si response
300000
fe response
ca response
250000

200000
Counts

150000

100000

50000

0
130
146
162
178
194
210
226
242
258
274
290
306
322
338
354
370
386
402
418
434
450
466
482
498

-50000
Gamma Energy Level

Figure 4. A Sample Signal Spectra, with Elemental Library Responses for Si, Fe, & Ca
 5 FOSTER – INTERNATIONAL ON-LINE COAL ANALYZER TECHNICAL CONFERENCE

Step 2. The analyzer computer performs a multiple linear regression. The regression software
calculates the amount of each library response that will most nearly approximate the shape of the
signal spectra when added together. See the results of the MLR in Figure 5 below:

A Visual Comparison of Sample Spectra with MLR Results Using Si,

Fe, and Ca

400000

350000

300000

250000

200000
Counts

Spectra
Solution
150000

100000

50000

0
130
145
160
175
190
205
220
235
250
265
280
295
310
325
340
355
370
385
400
415
430
445
460
475
490
505
-50000
Gamma Energy Level

Figure 5. The MLR results have been applied to the Elemental Response Library and
shown here graphically juxtaposed with the original Sample Gamma Spectra. Below is
the Output of the MLR program:

MLR Output:

Gain
Silicon: 0.596
Iron: 1.936
Calcium: 0.498
Overall Offset: -199.800

These multipliers (gains shown above), returned by the Multiple Linear Regression evaluation of
the data determine the levels of each element present in the sample and are therefore used to
report to the customer the levels of the various elements in the coal.
 6 FOSTER – INTERNATIONAL ON-LINE COAL ANALYZER TECHNICAL CONFERENCE

The table below shows the theoretical range of sensitivity for each of the elements in the periodic
table for a PGNA instrument.

Table 1. Expected PGNA Sensitivity to Elements of Interest*

Sensitivity in Weight % ** Elements
<0.01% Cl,Sc,Ti,Ni,Cd,Hg,Sm,Gd,Dy,Ho
0.01-0.1% S,V,Cr,Mn,Fe,Co,Cu,Rh,Ag,In,Hf,Ir,Au,Nd,Eu,Er,Yb,H
0.1-0.3% N,Na,Al,Si,K,Ca,Ga,Se,Y,Cs,La,W,Re,Os,Pt,Pr,Tm
0.3-1.0% Li,Be,Mg,P,Zn,As,Mo,Te,I,Ta,Pb,Ce,Tb,Lu,Th,U
1.0-3.0% C,Ge,Br,Sr,Zr,Ru,Pd,Sb,Tl
>3.0% Other Elements
* Note: Table taken from “On-Line Prompt Gamma Neutron Activation Analyzers, Published in the Process/Industrial
Instrument and Controls Handbook, Editor-Gregory K. McMillan, Fifth Edition, McGraw Hill, 1999.
** Three sigma detection limit in 10 minutes within an elementary simple rock matrix, ≥150mm thick

Some Calibration Basics

A basic understanding of the use of gains and offsets to optimize the data reported to the
customer by an analyzer is fundamental to understanding many of the concepts put forth in the
paper. The graphs below demonstrate some of the basic concepts:

Raw Sulfur Data Prior to Calibration

Lab Sulfur vs Analyzer Sulfur
Data Needs a Calibration Correction

1.9

1.7 Unity Line

1.5
Lab Sulfur (%)

1.3

1.1

0.9
RMSD = 0.203
0.7
Trendline r2 = 0.971

0.5
std.error = 0.041
0.3
0.3 0.5 0.7 0.9 1.1 1.3 1.5 1.7 1.9
Analyzer Sulfur (%)

Figure 6. A Plot of the Raw Sulfur Results, An Analyzer Compared with the Lab
 7 FOSTER – INTERNATIONAL ON-LINE COAL ANALYZER TECHNICAL CONFERENCE

Lab Sulfur vs Analyzer Sulfur

Data is Gain Corrected Only

1.9

1.7

1.5

1.3
Lab Sulfur (%)

1.1

0.9
RMSD = 0.053

0.7
r2 = 0.971

0.5
std.error = 0.041
0.3
0.3 0.5 0.7 0.9 1.1 1.3 1.5 1.7 1.9
Analyzer Sulfur (%)

Figure 7. The Raw Data with a New Gain of 0.76 Applied to Correct the Reported Sulfur

Lab Sulfur vs Analyzer Sulfur

Data is Offset Corrected Only

1.9

1.7

1.5

1.3
Lab Sulfur (%)

1.1

0.9
RMSD = 0.083

0.7
r2 = 0.971

0.5
std.error = 0.041
0.3
0.3 0.5 0.7 0.9 1.1 1.3 1.5 1.7 1.9
Analyzer Sulfur (%)

Figure 8. The Raw Data with an Offset -0.18 Applied to the SABIA Sulfur Result
 8 FOSTER – INTERNATIONAL ON-LINE COAL ANALYZER TECHNICAL CONFERENCE

Lab Sulfur vs Analyzer Sulfur

Data is Gain and Offset Corrected

1.9

1.7

1.5

1.3
Lab Sulfur (%)

1.1

0.9
RMSD = 0.040

0.7
r2 = 0.971

0.5
std.error = 0.041
0.3
0.3 0.5 0.7 0.9 1.1 1.3 1.5 1.7 1.9
Analyzer Sulfur (%)

Figure 9. The Raw Data with a New Gain of 0.76 and an offset of +0.034 Applied to Correct the
Reported Sulfur – Notice that the Trend line in Red Coincides with the Unity Line

Guarantees
Depending on the specific application, the performance verification of a PGNAA system can be
labor intensive. In almost most cases the analyzer installation must be planned in such a way
that results from existing mechanical sampling systems can be compared in a meaningful way
with the results of the nuclear analyzer. Historically, guarantees by vendors have been made in
terms of “RMSD”, “Static Precision”, or more recently, “GRUBBS Estimator”. At issue, is the
need to quantify the performance of the instrument and to do it in such a way that it can be
related to existing sampling/lab systems. A coal analyzer guarantee might be stated as follows:

Table 2. Hypothetical PGNAA Analyzer Guarantee

Parameter Static Precision RMSD Comments
10 minute basis, 1 hour analysis
30 points, 1σ, time, 20 points,
Wt. % Wt.%
Sulfur 0.10 0.05 Range not to
exceed 1.5%
Ash 0.5 0.5 Range not to
exceed 3.0%

Static Precision = standard deviation of consecutive readings where;

 9 FOSTER – INTERNATIONAL ON-LINE COAL ANALYZER TECHNICAL CONFERENCE

n 2
∑ ( xi − x)
i =1
S=
n−1

S = Static Precision
n = the number of paired comparisons
xi = the analyzer value in the ith comparison
x = the average of the n analyzer values

RMSD = Root Mean Square Deviation, as the name implies, it is the square
Root of the mean of the squared differences between the analyzer and lab values,
where;

∑ ( xi − yi )
2

i =1
RMSD =
n

n = the number of paired comparisons

xi = the analyzer value in the ith comparison
yi = the reference value in the ith comparison

Calibrating in the Factory with Generic Artificial Standards

The creation of a synthetic factory standard for the calibration of a nuclear based elemental coal
analyzer in the factory requires careful planning and execution of the plan. A meaningful set of
artificial reference standards has the following traits:

• The standards must cover a wide range of quality for all the parameters of interest, for
example, to ensure that the analyzer can perform properly over a wide range, the
standards should allow for an ash range of 40% and a sulfur range o 4%.
• The standards must be assembled in such a way as to ensure that their elemental
composition is known with great accuracy. This is accomplished by starting with “four
nines” source material (this means the source material vendor stands behind the stated
composition to within four decimal places). In order to assemble the materials, in the
final synthetic standards, a detailed plan for assembly which minimizes human error to
the greatest extent possible is an absolute requirement.
• The materials must be combined in such a way as to eliminate self correlation. In other
words, no two elements can move together throughout the range of samples. For
example, if the sulfur % of coal increases through the sample set in a way analogous with
another elemental parameter such as calcium % of Coal.
• The reference samples should be packaged in such a way as to allow the operator in the
factory and in the field to vary the loading in the case of an on-belt Analyzer.
10 FOSTER – INTERNATIONAL ON-LINE COAL ANALYZER TECHNICAL CONFERENCE

With a set of about 16 carefully designed sets ( a set of standards would consist of several
containers of the synthetic standard material – typically enough to place several layers of
material on an the conveyor for an on-belt t analyzer and extend well beyond the analysis
region in both directions) of standards used in a carefully conducted factory calibration a
vendor of nuclear elemental analyzers can collect the data necessary to deliver an analyzer
that after field calibration will meet the appropriate performance guarantees for any type of
coal or combination of coals the customer would want to analyze. If the standards have been
designed correctly (a correct design includes taking into account the range of quality of all
variables of interest to the customer or that might impact the nuclear analyzer performance
for the entire region of interest, in this case, the United States of America) and built to
exacting standards then the vendor will be able to deliver an analyzer that with a robust
calibration for the universe of coal possibilities, for varying conveyor bed depth, and for
varying nuclear source strength such that the analyzer is precise and accurate enough to meet
the contract performance guarantee. Of course, even the best factory calibration must be
accompanied by a good field calibration – see the section on field calibration below.

The photo below in Figure 10 shows how bags can be used to place the material on the
conveyor belt in multiple layers. Plastic Tubes filled with the reference materials can be used
in the same way. Early attempts at using synthetic reference standards utilized larger
containers filled with the sample material, but this precluded the ability to take varying belt
loading into account in the factory calibration.

Figure 10. Using Carefully Weighed Bags as Containers for Synthetic Standards
 11 FOSTER – INTERNATIONAL ON-LINE COAL ANALYZER TECHNICAL CONFERENCE

In a factory calibration utilizing synthetic calibration standards the data is typically gathered to
vary the following parameters:

• Widely varying elemental parameters. Ash range of 40% and Sulfur range of 4%. This
is accomplished through the creation of the standards themselves.
• Varying bed depth on the analyzer conveyor test fixture.
• Varying source strength. The data should be gathered at different source strengths,
including full strength.

Each configuration is run for 30 minutes to an hour to gather sufficient data for use in
optimization of the analyzer.

Evaluation of the data from running the 16 factory standards involves the use of sophisticated
commercially available math/statistical tools. Because the source thermal neutrons in a PGNA
device bombard everything within their reach, the sample spectrum contains much more than just
the elemental information for the coal sample. This additional information can be used to
optimize the performance of the analyzer. Some of the variables reported in the sample gamma
spectrum can be used by the analyzer for self-correction of the resulting information displayed to
the customer. Utilization of this information enables the factory technician to set the way in
which an analyzer calculates the end result such that the analyzer is more robust, more accurate
and more precise than would otherwise be possible. This enhanced robustness includes the
ability to compensate for varying belt loading, an amazing range of elemental qualities, such as
0 % to 4 % sulfur levels and 0 % to 40 % ash levels, as well as the decay of the thermal neutron
source, typically Californium 252.

Once it is determined what relationships within the static sample data are statistically sound, they
are incorporated into the basic factory “calibration” of the analyzer, which also includes
derivation of basic gains and offsets for the parameters of interest that are displayed for the
customer. With the factory calibration process complete, the analyzer is set with gains at unity
and offsets of zero. Afterwards, any modification that requires a gain of less than 0.7 or greater
than 1.3 should be approached with suspicion, since the responsiveness (and therefore the
precision) of the instrument is being altered.

Some basic rules of thumb for this calibration effort include:

1. Be Careful with Outliers. Never discard apparent “outliers” without a sound reason. The
basic principle here is that most data that looks like an outlier is telling the observer
something fundamental about the model used to portray the composition of the coal.

In general, experience has shown that if you are dealing with an established ASTM lab you
should trust the data. This sometimes means that you have to put in some extra hours to
Figure out what the analyzer is doing or what human error has been made in the process.

2. Never Verify Analyzer Performance with the Same Data used to Calibrate. Never use the
data used to calibrate to also verify the performance of the analyzer. For example, if the
 12 FOSTER – INTERNATIONAL ON-LINE COAL ANALYZER TECHNICAL CONFERENCE
operator has a statistically sound set of data, say 60 data points from the field covering a wide
range of quality he can sort the data the following way:

a. Use a large portion of the data set to calibrate the unit. Select this set such that all
coal types and elemental quality ranges are represented and such that the same is
true of the remaining verification set. The more data available the better.
b. Perform a calibration using the majority of the data available, being careful not to
throw out outliers without very good cause and using sound judgment relative to
the results of the statistical tools available. Remember to be certain that both the
validation data set and the calibration data set cover the range of quality and are
as evenly dispersed throughout the data as possible.
c. Once the operator is confident he has arrived at a calibration that is based on
sound principles as it relates to the physics of the analyzer and good rules of
statistics, the next step is to test the calibration. This is done by observing the
results the analyzer comes up with, using the new calibration on the verification
set of data set aside by the operator. If the verification data set has enough range
of quality and coal types and the results are good, the calibration is probably
valid. See the example below:

Ash Calibration
Using Separate Calibration and Validation Data

14.00

13.00

12.00
Lab Ash (%)

11.00 Ash
10.00 Validation Set

9.00

8.00

7.00
7.00 8.00 9.00 10.00 11.00 12.00 13.00 14.00
Sabia Ash (%)

Figure 11. Using Separate Calibration and Validation Data

3. Don’t Over fit the Data. Experience has shown that the mathematical tools available for
processing of comparative data are so powerful that they can put any data set on a straight
line if given enough degrees of freedom. The tool “solver” found in Excel is notorious for
this. Below are two examples of calibration data from an analyzer. In Figure 14. the
information as reported for the individual oxides is simply added and regressed against the
lab.
 13 FOSTER – INTERNATIONAL ON-LINE COAL ANALYZER TECHNICAL CONFERENCE

Lab Ash vs Analyzer Ash

Finised Product Belt 2"x0" Topsize
48" belt, 1200 tons/hr

20

18
ASTM Lab Inter-Reproducibility Limits

16
Lab Ash (%)

14

12

10
r2 = 0.871
8

std.error = 0.480
6
6 8 10 12 14 16 18 20
Analyzer Ash (%)

Figure 12. Using Excel to Regress a Logical Assembly of Data

Lab Ash vs Analyzer Ash

Finised Product Belt 2"x0" Topsize
48" belt, 1200 tons/hr

20

18
ASTM Lab Inter-Reproducibility Limits

16
Lab Ash (%)

14

12

10
r2 = 0.943
8

std.error = 0.320
6
6 8 10 12 14 16 18 20
Analyzer Ash (%)

Figure 13. Using Excel Solver in a Haphazard Way, with unrelated Variables to Regress
Calibration Data
 14 FOSTER – INTERNATIONAL ON-LINE COAL ANALYZER TECHNICAL CONFERENCE

Calibrating in the Field

It is not realistic with current analyzer technology to expect that an analyzer can be calibrated in
the factory and shipped to a site and operated without an accompanying field calibration.
Calibration in the field is optimum if it includes an initial static calibration followed by a
dynamic calibration with ongoing lab and analyzer data with samples from a good mechanical
sampler and results from a good lab. Careful preparation of static samples enables the vendor
and the customer to quickly make any adjustments necessary to the robust factory calibration.
These adjustments take into account differences in belt composition and surrounding structural
materials. Once the static standards are prepared the initial site calibration can be completed in a
matter of a few hours. Creation of the static samples is defined in the next section.

Although the analyzer has now been calibrated in the factory with a definitive set of artificial
standards representing the universe of possibilities of coal in the U.S., it is always prudent to
consider creation of a set of standards using actual coal from the site to be used in adjusting the
factory calibration at the site. As mentioned previously, this calibration adjustment is necessary
to take into account possible differences in belt composition between the factory and the site as
well as differences in the structure of the conveyor system itself.

In order to build a useful set of site standards at least 5 or 6 samples are necessary. Each sample
must contain enough material to cover the analysis region of the belt (about 1.5 x 55 gallon
drums of coal) and must be of a topsize that lends itself to the loading of the calibration sample
containers. Each sample container must be completely filled without voids. A 2”x0” topsize
spread lends itself nicely to this effort.

The quality of the samples must follow the standard rule of calibration samples as follows:

• All of the parameters of interest to the customer should be present in the samples
• The quality parameters should cover as wide a range as possible. This is very difficult to
achieve in normal, day-to-day operation. To the extent the range covered in the coal
samples is short of the range of interest, the calibration is subject to potential errors.
• The quality of the samples should be spread out evenly among the six samples. For
example, if the customer has coal that ranges in Sulfur from 1% to 2.5%, then he would
like samples with Sulfur levels of 1.0%, 1.25%, 1.5%, 1.75%, 2.0%, and 2.5%.
• Great care should be taken to characterize the coal as accurately as possible, ensuring that
the sample in the lab truly represents the coal in the drums and analysis done in a lab that
provides consistently accurate results.

Just as in the factory, the six static reference standards created with coal from the customer’s site
are run in varying layers to provide the vendor data that enables him to utilize a meaningful
correction for variances in belt loading. Typically, the samples are one in a “1 layer”, “2 layer”,
“3 layer”, and “4 layer” scenario. As in the factory, each scenario needs to be run for 30 minutes
to 1 hour. A 24-hour period, if carefully planned out is more than enough time to complete the
gathering of this data.

With this additional data in place, the vendor can make any adjustments necessary to the
analyzer calibration and the customer is now off and running with a meaningful calibration for
 15 FOSTER – INTERNATIONAL ON-LINE COAL ANALYZER TECHNICAL CONFERENCE
his on-belt analyzer. Now site static calibration should be followed by a calibration with
dynamic data.

With a calibrated analyzer in place and operating the customer is now in a position to provide the
vendor with dynamic data as a routine part of his ongoing operation. The vendor in turn can take
this lab data and compare it with the analyzer performance to further optimize results.

It is always recommended that customers apply a PGNA analyzer in such a way as to get
ongoing dynamic data as part of the routine operation. Please note, that if a special, non-
recurring effort is necessary to obtain performance verification, or ongoing calibration data the
application has probably not been planned well. If at all possible, analyzer applications should
have some means of gathering representative mechanical samples of the coal being analyzed as
part of the routine operation. Otherwise, analyzer performance may not be optimum and
disputes are difficult to resolve. Below are some of the classic calibration pitfalls experienced by
customers and vendors over the years:

Figure 14. An OnBelt Analyzer for Run-of-Mine Coal

 16 FOSTER – INTERNATIONAL ON-LINE COAL ANALYZER TECHNICAL CONFERENCE

Inadequate Range in Data for Gain Adjustment

Lab Sulfur vs Analyzer Sulfur

Exercise Caution Applying Gain Corrrections
To Data With Limited Range

1.9

1.7

1.5
Lab Sulfur (%)

1.3

1.1

0.9

0.7

0.5 r2 = 0.530

std.error = 0.028
0.3
0.3 0.5 0.7 0.9 1.1 1.3 1.5 1.7 1.9
Analyzer Sulfur (%)

Figure 15. The Data in this Graph implies it is Pointed in a Certain Direction – Be Careful
There is not enough range in this data to apply a gain correction with confidence. In this
particular case, the Excel regression is calling for a gain correction of 1.45. This should be a red
flag to the operator. Gain corrections ±.3 beyond unity should be regarded with suspicion. It is
probably best in this case to apply an offset correction until additional data is available with more
range.

Not Enough Data to Base a Calibration On

Below is a calibration of the data in Figure 17. The vendor has recommended an offset
adjustment to the analyzer as reflected in Figure 18.
 17 FOSTER – INTERNATIONAL ON-LINE COAL ANALYZER TECHNICAL CONFERENCE

Lab Sulfur vs Analyzer Sulfur

Bias Corrected Data

1.9

1.7

1.5

1.3
Lab Sulfur (%)

1.1

0.9

0.7

0.5 r2 = 0.530

std.error = 0.028
0.3
0.3 0.5 0.7 0.9 1.1 1.3 1.5 1.7 1.9
Analyzer Sulfur (%)

Figure 16. The Data of Figure 17 has been bias corrected.

Now additional data becomes available to the customer. Added to the original data, the customer
decides to ask for a calibration adjustment.
 18 FOSTER – INTERNATIONAL ON-LINE COAL ANALYZER TECHNICAL CONFERENCE

Lab Sulfur vs Analyzer Sulfur

Additional Data - 4 Data Points
Operations Wants a Gain and Offfset Correction

1.9

1.7

1.5
4 New Data Points
These added to the first
Lab Sulfur (%)

1.3
group continue to imply
1.1
the data is pointing in
upward direction.
0.9

0.7

r2 = 0.788
0.5
std.error = 0.030
0.3
0.3 0.5 0.7 0.9 1.1 1.3 1.5 1.7 1.9
Analyzer Sulfur (%)

Figure 17. 4 Additional Data Points are Added – Operations Wants a New Calibration

Lab Sulfur vs Analyzer Sulfur

Additional Data - 4 Data Points
A Gain of 1.42 and an Offset of -0.27 Have Been Applied

1.9

1.7

1.5
Lab Sulfur (%)

1.3

1.1

0.9

0.7

0.5 r2 = 0.788

std.error = 0.030
0.3
0.3 0.5 0.7 0.9 1.1 1.3 1.5 1.7 1.9
Analyzer Sulfur (%)

Figure 18. A New Gain and Offset Have Been Implemented

 19 FOSTER – INTERNATIONAL ON-LINE COAL ANALYZER TECHNICAL CONFERENCE

Lab Sulfur vs Analyzer Sulfur

Additional Data - 9 More Data Points
This Graph Shows Pre-Calibration

1.9

1.7

1.5
Lab Sulfur (%)

1.3

1.1

0.9

0.7

0.5 r2 = 0.856

std.error = 0.034
0.3
0.3 0.5 0.7 0.9 1.1 1.3 1.5 1.7 1.9
Analyzer Sulfur (%)

Figure 19. A Look at 9 More Data Points with Calibration Backed Out – Oops!!

Lab Sulfur vs Analyzer Sulfur

Additional Data - 9 More Data Points
This Graph Shows What Calibration Does

1.9

1.7

1.5
Lab Sulfur (%)

1.3

1.1

0.9 Trains Have Been Shipped

Out of Spec!!
0.7

0.5 r2 = 0.856

std.error = 0.034
0.3
0.3 0.5 0.7 0.9 1.1 1.3 1.5 1.7 1.9
Analyzer Sulfur (%)

Figure 20. The Erroneous Calibration Creates Some Out-of-Spec Trains

 20 FOSTER – INTERNATIONAL ON-LINE COAL ANALYZER TECHNICAL CONFERENCE
The dilemma of a calibration based on insufficient data is that it destroys the confidence of the
operations people in the technology and the vendor who has yielded to the pressure to put a poor
calibration into the analyzer is tempted to blame the customer. This is a “Lose/Lose” situation.

Standard error, R2, and RMSD

Standard Error can be a deceiving parameter for judging the quality of a calibration. Below is
the same graph shown in Figure 6. Please note that the standard error is a very impressive 0.041,
while the RMSD (Root Mean Squared Deviation) is a very bad 0.203. This is because the
standard error is a statistic about its own trend line, while the RMSD is a statistic about the Unity
Line, or the Zero Bias Line. For customers involved in a performance evaluation, RMSD has to
be the statistic of most importance in determining value to the operation.

Raw Sulfur Data Prior to Calibration

Lab Sulfur vs Analyzer Sulfur
Data Needs a Calibration Correction

1.9

1.7 Unity Line

1.5
Lab Sulfur (%)

1.3

1.1

0.9
RMSD = 0.203
0.7
Trendline r2 = 0.971

0.5
std.error = 0.041
0.3
0.3 0.5 0.7 0.9 1.1 1.3 1.5 1.7 1.9
Analyzer Sulfur (%)

Figure 21. Standard Error vs. RMSD

In much the same way, R2 can be very deceiving when trying to decide how well an analyzer is
performing.
 21 FOSTER – INTERNATIONAL ON-LINE COAL ANALYZER TECHNICAL CONFERENCE

Lab Sulfur vs Analyzer Sulfur

Additional Data - 4 Data Points
Operations Wants a Gain and Offfset Correction

1.9

1.7

1.5
4 New Data Points
These added to the first
Lab Sulfur (%)

1.3
group continue to imply
1.1
the data is pointing in
upward direction.
0.9

0.7

r2 = 0.788
0.5
std.error = 0.030
0.3
0.3 0.5 0.7 0.9 1.1 1.3 1.5 1.7 1.9
Analyzer Sulfur (%)

Figure 22. In this Example from Figure 18 the R2 has Improved from .0533 to 0.788

In the example of Figure 23, borrowed from Figure 6. the dramatic improvement in R2 helps
convince the operator that the data is real. This is a mistake. In general, R2is of limited value in
making decisions about the validity of the data. As seen in Figure 21., the end result is a very
bad calibration.

A Good Rule of Thumb: You shouldn’t feel too comfortable with your dynamic data calibration
until you lave lots of data, dozens and dozens of points, and there is excellent range to all the
data.
 22 FOSTER – INTERNATIONAL ON-LINE COAL ANALYZER TECHNICAL CONFERENCE

Chasing the lab, Poor Samples, Human Error, etc..

The graph from Figure 24., shown below as Figure 25. also shows how two data points have
played an inordinate role in influencing those involved, to implement a faulty calibration in an
analyzer.

Lab Sulfur vs Analyzer Sulfur

Additional Data - 4 Data Points
Operations Wants a Gain and Offfset Correction

1.9

1.7

1.5
4 New Data Points
These added to the first
Lab Sulfur (%)

1.3
group continue to imply
1.1
the data is pointing in
upward direction.
0.9

0.7

r2 = 0.788
0.5
std.error = 0.030
0.3
0.3 0.5 0.7 0.9 1.1 1.3 1.5 1.7 1.9
Analyzer Sulfur (%)

Figure 23. The 2 Data Points in the Lower Part of the Graph Outside of the ASTM limits end up
Erroneously Influencing the Calibration (See Figure 21. for the Bad Result)

Assuming that the analyzer result was correct on these two points, the error could have possibly
been in the lab analysis, the quality of the mechanical sample. The error could have been as
simple as incorrect transcription of numbers (the elusive human error). Great care must be taken
to ensure that faulty data does not play too important a role in calibration of an analyzer.

Conclusion
The calibration of a modern PGNA nuclear based elemental analyzer can seem complicated. It
doesn’t have to be. Attention to detail from the vendor and the customer, combined with good
communication and teamwork can ensure that one of these modern instruments can add
tremendous value to an operation - value that puts very real dollars on the bottom line of many
operations. In addition, current technology is such that once an analyzer is calibrated correctly it
can run for months on end without correction to the calibration. Pay attention to some of the
simple rules in this paper and hopefully some problems can be avoided.
 23 FOSTER – INTERNATIONAL ON-LINE COAL ANALYZER TECHNICAL CONFERENCE

References
Kirchner, A.T., “On-line analysis of coal”, IEA Coal Research report IEACR/40, September
1991.

Proctor, R.J., “On-line Prompt Gamma Neutron Activation Analyzers”, Process/Industrial and
Controls Handbook, Editor-Gregory K. McMillan, Fifth Edition, McGraw Hill, 1999.

Rose, Charles D., “Methods for Assessing the Accuracy of On-Line Coal Analyzers”, Journal of
Coal Quality, p.19-28, January-March 1991.

Rose, Charles D., “Calibration and Testing of On-Line Coal Analyzers”, Journal of ASTM, April
2004, Vol .1, No.4.