Statistical Quality Control

of Concrete
 Introduction
• Whenever observations are made of material
properties, they will form statistical distributions.
• This occurs because the readings are not
absolutely precise.
 The required properties must be specified for any
construction project, but it may be seen that specifying
zero failures at any acceptance level is not practical.
• The only practical solution is to specify a %age
defect rate (failures) that is considered to be
acceptable.
 Introduction
• The results that are obtained from most
measurements will form a normal
distribution. There are two questions that
must be answered:-
1. How far above the failure level does the
average measurement have to be, in order
to get a given %age of defects?
2. How likely is it to observe a relatively large
number of failures in a small sample taken
from a parent population with a low %age of
defects?
 Sampling
• Figure 1 shows a histogram for a set of
results that might be obtained if 10
concrete cubes from the same mix were
tested for strength (it can be
generated in a spreadsheet using the
random number generator).
• Figure 2 shows the effect of increasing
the number of samples gives a more
regular distribution that approximates
to the curve in Figure 3.
FIGURE 1. Strength Results for 10 Cubes
FIGURE 2. Strength Results for 1000 Cubes
FIGURE 3. Normal Distribution for Strength
 Sampling
• Normal distribution is observed almost
every time a material property is
measured, however, it is not absolutely
accurate but is close enough for
analysis.
• The true distribution for cubes is
normally skewed slightly to the right,
because it is more likely to get a very
strong outlying result than a very weak
one.
 Distributions
• In order to calculate how far above
the failure level the average
measurement has to be in order to get a
given %age of defects, it is necessary
to know the standard deviation,σ of the
measurements.

å
2
( xi - m )
s =
n
• where xi are the observations, n is the number
of observations, ∑xi denotes the sum of all
the values, and m is the mean (average)
defined from:
m=
å xi
n
• The standard deviation is a measure of the
spread of the results.
• Two other statistics are derived from it:
– The standard error (coefficient of variation) =
σ/m
– The variance is σ2
 Distributions
• Figure 4 shows two distributions with
different σ.
• Both have a mean strength of 50 and
the area under the graphs, that
corresponds to the total number of
samples tested, is the same.
• The high σ would occur where there is
poor quality control giving the wide
variation between samples.
FIGURE 4. Two Strength Distributions with
Different Standard Deviations
• The two shaded areas in Figure 4 represent 5% of the area
under each curve so 5% of a large number of samples would
lie within this region.
• The permitted %age failure level is given by equation.
– Mean = permitted %age failure level + β × σ
 permitted %age failure level = Mean - β × σ
• Where β is the “reliability index” and is obtained from Table
1 and the term “β × σ” is referred to as the margin.
• For 5% defectives β is 1.64.
• For the data set with a σ of 5,
– the defective region is below a strength of 50 − 5 × 1.64 = 42 MPa.
• For the data set with σ of 20,
– the defective region is below a strength of 50 − 20 × 1.64 = 17.2 MPa.
Table 1 – β Values for Various Values of %age Failures

%age Failure Permitted β Value

16 1.00
10 1.28
5 1.64
2.5 1.96
2 2.05
1 2.33
 Probability
• Calculating how likely it is to observe a relatively
large number of failures in a small sample taken from
a parent population with a low %age of defects is a
question of probability.
• Calculations of probability are based on equations:
– The probability of multiple events all occurring =
the probabilities of the individual events
multiplied together….......................................(1)
– The probability of any one of several alternative
events occurring = the probabilities of the
individual events added together….............(2)
 Probability
• If the probability of each outcome is the same, equation
(2) means that the probability is multiplied by the number
of ways the outcome can occur.
• If just 3 cubes are tested with a probability of failure of
5% (= 0.05) and a probability of passing of 95% (= 0.95)
for each cube the possible outcomes are given in Table 2.
• The probability of an outcome in Table 2 is calculated as
the probabilities for each sequence calculated from
equation (1) and then the alternatives are added using
equation (2).
• Note, that the probabilities in the table add up to one.
 It is certain that one of the outcomes would occur.
Table 2 - Out come s f or Thre e Cube s , P = Pass, F = Fail
Outcome Sequence Probability
3
All three pass PPP 0.95 = 0.8574
Two pass, one fail PPF 0.952 × 0.05 × 3 = 0.1354
PFP
FPP
One pass, two fail FFP 0.052 × 0.95 × 3 = 0.0071
FPF
PFF
All three fail FFF 0.053 = 0.0001
 Probability
• Now consider a site on which 3 cubes are
tested every day.
• For 1 cube to fail, there is a probability of
0.1354 or 13% (from Table 2) so this is not
considered remarkable.
• However, the probability of this happening on
each of 5 days in a row is:
• 0.13545 = 0.000045 so this is unlikely and may
not be a cause for concern.
 Probability
• This is different from the probability of
having 1 failure on just one of the 5 days.
The probability of this is calculated as:
• Probability of getting 1 failure on one day ×
probability of not getting this on 4 days ×
number of days on which the failure could
occur (i.e., the number of possible sequences).
• = 0.1354 × (1 − 0.1354)4 × 5 = 0.37 or 37%
which is probable.
 Probability
• For more than 1 cube to fail on any given day,
the probability may be calculated from values in
Table 2 as 0.0071 + 0.0001 = 0.0072 or 0.7%, so
this is unlikely.
• However, if the site continues for 100 days of
working, the probability of this happening once is:
• 0.0072 × (1 − 0.0072)99 × 100 = 0.353 or 35% that
is also probable and should not necessarily be a
cause for concern.
 Correlations
• Examples of exam mark against attendance at
lectures for a concrete construction
class shown on a plot (Figure 5).
• What we want to know is whether it is
necessary to attend the lectures to pass the
exam.
• What statistics can tell us is what the
probability is that this is a purely random
distribution with no “correlation” at all.
 Correlations

FIG. 5. Exam Mark Versus Percentage Attendance

 Correlations
• To calculate this, we work out a statistic called r 2
(the coefficient of determination).
• The computer package tells us that r2 = 0.52 and
therefore, r = 0.72. [r = coefficient of
correlation]
• Table 3 tells us that for 80 observations the 1%
significance value of r is 0.283, i.e., if r was
0.283, there would be a 1% chance of a
distribution showing this amount of correlation
occurring by chance.
Table 3. Some Values of r for Two Variables

No. of Data Points r for 5% Significane r for 1% Significane

5 0.754 0.874
10 0.576 0.708
20 0.423 0.537
30 0.349 0.449
40 0.304 0.393
50 0.273 0.354
60 0.250 0.325
70 0.232 0.302
80 0.217 0.283
90 0.205 0.267
100 0.195 0.254
 Correlations
• r is well above this value, so the chance if
this being a random distribution is even
less, i.e., the correlation is proven.
• Note that we have not answered our
original question; it could just be that the
“better” students attend the lectures and
also do well in the exams, we have not
proved cause and effect.
 Correlations
• Exactly the same situation applies in the study of
materials.
• Almost all of the properties of concrete (e.g.,
strength, permeability, frost resistance)
correlate, but this does not prove that they
affect each other.
• The strength is a good indicator of durability,
and has been used for a long time as the only test
that is routinely used to assess the quality of
hardened concrete.
 Correlations
• However, the relationship between
strength and durability is based on
experimentally observed correlations.
• A low w/c ratio causes generally both
high strength and high durability, but
there are many exceptions.
 Conclusions
1. It is not practical to specify “0” failure for a material test, so
it is necessary to use statistics for specifications.
2. Most experimental results will form a normal distribution.
3. The reliability index is measured in standard deviations from
the mean and gives the number of failures that are to be
accepted in a batch.
4. The probability of multiple events all occurring = the
probabilities of the individual events multiplied together.
5. The probability of any one of several alternative events
occurring = the probabilities of the individual events added
together.
6. Correlations do not prove cause and effect.
 Example 1
• Concrete cubes are being tested and
the mean strength is 40 MPa and the
standard deviation is 5 MPa. What is
the strength below which 5% of the
strengths will lie?
• Solution:
40 − (1.64 × 5) = 31.8 Mpa
 Example 2
• Railway sleepers (railroad ties) are
required to have a length of 3 m with
5% defects. If the standard deviation
of the observed lengths is 5 mm, what
average length is required?
• Solution:
3000 + 1.65 × 5 = 3008.2 mm
 Example 3
• If 10 concrete cubes are tested from a
concrete known to have a 5% failure
level, what is the probability of one
failure?

• Solution:
0.05 × 0.959 × 10 = 0.315
 Example 4
• Concrete is delivered to site with a characteristic
strength of 25 MPa and a percentage defect rate of
5%. If 6 cubes are made, what is the probability of:
• a. No failures
• b. 1 failure
• C. If 6 cubes are made each day, what is the
probability of getting one failure per day on each of 3
consecutive days?
• d. If the observed standard deviation of the test
results is 5 MPa, what is the mean strength?
 Solution
• a. 0.956 = 0.735
• b. 0.955 × 0.05 × 6 = 0.232
• c. 0.2323 = 0.012
• d. 25 + (5×1.64) = 33.2 MPa
 Example 5
• a. Explain why the strength of materials such as bricks
and concrete must be specified in terms of a statistical
failure rate?
• b. Bricks are tested for strength. If the average is 15
MPa and the standard deviation is 3 MPa, what is the
strength below which 5% would be expected to lie?
• c. If five of the bricks are tested, what is the
probability of one strength being below the 5% level?
• d. If three further sets of five are tested, what is the
probability of all three sets having one each below the
5% level?
 Solution
• a. Because the strengths will form a
normal distribution that is non-zero at
all strengths.
• b. 15 − (1.64×3) = 10.08
• c. 0.954 × 0.05 × 5 = 0.203
• d. 0.2033 = 0.008
 Sources of variations in concrete strength
test results

• When concrete samples are tested for

strength the variations come from a large
number of different sources, for example,
• Daily variations in concrete supplied to site:
– Changes in cement
– Changes in aggregate
– Changes in batching control
– Changes in temperature
 Sources …

• Variations in concrete between successive

batches (loads):
– Changes in water content
– Changes in haul times
• Sampling variations within a load:
– Changes in aggregate content between two samples
taken from the same load.
– Operator error, for example, effect of sampling
from the end of a load.
 Sources …
• Testing variations:
– Variations in the position of the aggregate in the
cubes.
– Operator error, for example, changes in loading
rate, or dirt on the platens.
• Each of these variations will have a variance
associated with it.
• These are additive, i.e., the total variance is the sum
of the variances from the different sources.
 Note that standard deviations are not additive.
 Making decisions about failing test results

• Consider the set of 30 cube results given in

column 2 of Table 4 that represent 10 sets of 3
cubes taken from 10 different loads (typically,
they might be from 10 days of concrete placing).
• The target mean for these was 50 and the target
standard deviation was 5.
• The simplest form of control chart is shown in
Figure 6 that simply shows individual cube
results.
Table 4. Analysis of a Set of Cube Results in MPa
Table 4. cont’d …
FIG. 6. A Basic Control Chart
 Making …

• The dashed lines are at 1.64 standard deviations

from the mean, and thus 95% of the samples
would be expected to lie between them (see
Table 1).
• Only two out of 30 samples lie outside so the
control might be assumed to be satisfactory.
• The weakness of this chart is that it
cannot show when a decision needs to be
made.
 Making …

• The better method is CUSUM, in which the target mean

strength is subtracted from each of observed strengths
to give a positive or negative difference.
• These differences are then added to give a Cumulative
SUM (see columns 3 and 4 in Table 4) that is plotted in
Figure 7.
• This shows that, although the average of all 30 was close
to target at 49.9, there was a series of poor results from
cubes 15–25.
• The dashed line shows a “mask” that can be moved on
screen to indicate when change is needed.
FIG. 7. A CUSUM Chart
 Identifying the source of the problem

• While the basic CUSUM plot can indicate when action

is needed; it cannot be used to identify the source of
the problem.
• The problem could come from the concrete itself
such as from the testing (the second two lists).
• Problems with the testing are likely to increase the
spread (range) of results between the 3 cubes in a
set that are made from the same batch of concrete
while problems with the concrete itself would not.
 Identifying …

• This spread can be measured as a standard

deviation of each set of three results.
• In Table 4, column 5 shows the standard
deviation of each set, and columns 6 and 7
show the CUSUM analysis of them that is
plotted in Figure 8.
• This does not show any particular trends so it
is likely that the concrete itself was at fault.
FIG. 8. Plot of CUSUM Range
 Multı-varıate analysıs
• CUSUM methods only work well when a large
number of tests on a single type of concrete are
being analyzed.
• If there are many different mixes being made,
there will be insufficient data to develop a
different cusum plot for each individual mix.
• In order to overcome this, the strengths can be
adjusted up or down using historic data on the
different mix types to fit them all to a single
model.
 Multıvarıate ...
• Adjusting results from mixes with cement
replacements or admixtures to match those of
plain cement mixes is, however, very unreliable.
• More powerful systems require a lot more data
collection to understand all the factors that
could affect the strength.
• These are called predictor variables and could be
as follows:
– Water to cement ratio
 Multıvarıate ...
• Cement content
• Cement type
• Cement replacement percentage
• Admixture percentage
• Aggregate content
• Aggregate grading
• Slump
• Temperature
• Haul time
 Multıvarıate ...

• A model can be created based on past data to

predict the strength.
• A simple form of this could take the form of:
– Strength =(x1 × predictor1) + (x2 ×predictor2)+(x3
×predictor3)......etc.......
• Where x1, x2, x3, are constants obtained from a
process called multiple regression on past test
results.
• Further terms such as (predictor 1)2 may need to
be included to make the model accurate.
 Multıvarıate ...

• All test results are entered into the

model and action is needed if the
strength deviates from the predicted
value.
• This indicates that an unexpected
additional factor is affecting the
strength, and is a cause for concern on
site.
 Designing for durability
• When designing major structures, clients
increasingly ask designers to carry out a
durability assessment rather than just relying on
the provisions in local codes and standards.
• For a concrete structure, this could be a
calculation of the depth to which salt (chloride)
penetrates during the design life; because if it
reaches the cover depth for the steel
reinforcement, it causes corrosion.
 Designing ...
• The durability may be increased by using a higher grade of
concrete or increasing the depth of cover to the steel, but both
of these adds to the cost.
• This calculation requires values for the following variables :
1. The cover depth
2. The permeability
3. The rate at which the permeability reduces with time
4. The coefficient of diffusion
5. The rate at which the coefficient of diffusion reduces with
time
6. The salt concentration in the local environment
7. The critical chloride concentration necessary for corrosion
 Designing ...

• None of these are accurately known.

• However, it is possible to estimate a mean
values and a standard deviation for each of
them and from these a 5% defectives level.
• What the designer needs to know is the
values of the permeability and diffusion
coefficient that must be specified in order to
achieve the necessary durability with 95%
certainty.
 Designing ...

• The penetration depth could be calculated

with 6 variables at their 5% defectives level,
but this would only have a probability of
happening of 0.056 (= 1.5 × 10−8) and would give
highly uneconomic requirements.
• The problem is solved by calculating a number
(typically, three) different possible values for
each variable.
 Designing ...

• Figure 9 shows the normal distribution divided

into 3 equal areas that thus have equal
probability.
• The boundaries lie 0.43σ either side of the mean.
• The arrows show the average values for each
area.
• These lie at the mean and 1.1σ to either side and
are representative of each of the 3 different
outcomes.
FIG. 9. Dividing the Normal Probability Curve into Equal Areas
 Designing ...

• If 3 values for 6 variables are obtained there are

36 (= 729) different combinations of them that
could occur with equal probability.
• If the chloride penetration depth is calculated
for all of these and the mean and standard
deviation calculated for them, a value of β can be
obtained as follows from equation:
– Design value = mean + β × standard
deviation
 Designing ...

• Thus, if the mean chloride penetration depth is

40 mm and the standard deviation is 10 mm; a
cover depth of 50 mm gives a β value of 1.
• If a 5% probability of failure is considered
appropriate then Table 1 shows that β should be
1.64.
• For critical elements β must be 3 or more to
reflect the consequences of failure.
 Conclusions
• CUSUM is a method for analyzing trends in test
results, and deciding when to take action.
• CUSUM range can be used to identify the cause
of data trends.
• Multivariate analysis is a more powerful method,
but requires more data collection.
• When designing structures for long design lives, a
statistical approach may be used to calculate a
reliability index.
 Example
• Prepare a CUSUM chart from the following data:
Solution
Quality Control of Concrete
 Introduction
• The required minimum quality level should
be set by the specification, and Quality
Control (QC) or Quality Assurance (QA)
are concerned with so regulating production
that the required quality level is attained at
minimum cost.
• A common mistake is (or was in the past and
still is in some areas) to confuse quality
control with check testing.
• The two have little in common.
 QC concept
• A basic QC concept, is to ‘control the
mass and not the piece’.
• It is far more economical to ensure that
no significantly defective concrete is
produced at a plant than to ensure
that no defective concrete is accepted
at particular delivery points (but the
samples are still taken at delivery
points).
 QC concept
• Excessive variation of concrete
strength signifies inadequate QC.
• Improvement in control may permit
a reduction in the cost of concrete
since the average strength can be
brought closer to specification
requirements.
 The objectives of QC and
QA
• There are two aspects to controlling
concrete quality.
• One of these is the avoidance of
failures and the other the attainment
of low variability.
• Obviously low variability will be of
assistance in avoiding failures and
vice-versa but it helps to consider the
two separately.
 The objectives …

• What is useful is to consider separately

those factors acting continuously and those
acting intermittently.
• It is even possible that some of the same
factors can fit into both categories, for
example, sand grading is unlikely to be
identical from truck to truck but there may
be a more substantial change from time to
time as extraction location or conditions
change.
 Control system
• The continuous basic variability can be thought
of as a feature of the production process.
• It can only be improved by improving that
process or the uniformity of the materials
supplied to it.
• The early detection and reversal of occasional
change in mean is a feature of the control
system.
• So the control system measures the basic
variability and detects change points.
 QC & QA
• Insofar as they differ, QA is concerned
with avoiding problems by pre-
inspection of materials and certification
of implementation of control and
production procedures (e.g. ISO 9000).
– It could be considered to be aimed more at
eliminating change points rather than at
their early detection.
 QA
• This has developed to ensure that specifications
are consistently met. 'Fit for purpose' and
'right first time' are the principles of QA and
the frame of reference for QA is ISO 9000
family of standards.
• To be certified as operating to the ISO 9000
standard is now virtually seen as essential in
today's construction industry.
• Many clients simply will not do business with
companies not certified to ISO 9000.
 QA
• QA has also been described as documented QC,
suggesting that the main difference is only one of
record keeping.
• QA provides an assurance, in the form of certified
records, that the established QC procedures have
been carried out in full.
• The control function consists of monitoring the
situation so as to detect, at the earliest possible,
when either the average quality or the variability
of that quality changes or becomes likely to change.
 Primary criteria for application
of statistical concepts
• The primary criteria for application of
statistical concepts is that the data
to be analyzed has been developed
from random sampling, and
• At least 30 consecutive test results
are available for analysis of
compression strength in ACI 214
procedure.
Standards of Concrete Control (General
Construction Testing) ACI 214

• Excellent: Std. Dev. below 2.76 Mpa

• Very Good: Std. Dev. of 2.76 to
3.45 Mpa
• Good: Std. Dev. of 3.45 to 4.14 Mpa
• Fair: Std. Dev. of 4.14 to 4.83 Mpa
• Poor: Std. Dev. above 4.83 MPa
 Standards of Concrete Control
(Field Control Testing) ACI 214
• Excellent: Coefficient of variation below
3.0%
• Very good: Coefficient of variation of 3.0
to 4.0%
• Good: Coefficient of variation of 4.0 to
5.0%
• Fair: Coefficient of variation of 5.0 to
6.0%
• Poor: Coefficient of variation above 7.0%