Errors and Calibration

Author: John M. Cimbala, Penn State University

Latest revision: 28 August 2008
Random vs. Systematic Errors
• There are two general categories of error: systematic (or bias) errors and random (or precision) errors.
• Systematic errors (also called bias errors) are consistent, repeatable errors. For example, suppose the first
two millimeters of a ruler are broken off, and the user is not aware of it. Everything he or she measures will
be too short by two millimeters – a systematic error.
• Random errors (also called precision errors) are caused by a lack of repeatability in the output of the
measuring system. The most common sign of random errors is scatter in the measured data. For example,
background electrical noise often results in small random errors in the measured output.
Systematic (or Bias) Errors
• Systematic errors are consistent, repeatable errors in a set of measurements
• The systematic error is defined as the average of measured values minus the true value.
• Systematic errors arise for many reasons. Here are just a few:
o calibration errors – perhaps due to nonlinearity or errors in the calibration method.
o loading or intrusion errors – the sensor may actually change the very thing it is trying to measure.
o spatial errors – these arise when a quantity varies in space, but a measurement is taken only at one
location (e.g. temperature in a room - usually the top of a room is warmer than the bottom).
o human errors – these can arise if a person consistently reads a scale on the low side, for example.
o defective equipment errors – these arise if the instrument consistently reads too high or too low due to
some internal problem or damage (such as our defective ruler example above).
• A nondimensional form of bias error is the mean bias error, defined as MBE = systematic error / true value.
Random (or Precision) Errors
• Random errors are unrepeatable, inconsistent errors in the measurements, resulting in scatter in the data.
• The random error of one data point is defined as the reading minus the average of readings.
• Example:
Given: Five temperature readings are taken of some snow: -1.30, -1.50, -1.20, -1.60, and -1.50oC.
To do: Find the maximum magnitude of random error in oC.
Solution: The mean (average) of the five temperature readings is -1.42oC. The largest deviation from this is
the reading of -1.20oC. The random (or precision) error for this data point is defined as the reading minus
the average of readings, or -1.20 - (-1.42) = 0.22oC. Thus, the maximum absolute value of random error
is 0.22oC. You can verify that the magnitude of the random error for any of the other data points is less
than this.
Comment: Five readings are not sufficient to measure a good average reading, especially in this example
since the individual readings vary so widely.
A
Accuracy
• Accuracy is the overall closeness of agreement between a measured value and the
true value.
• Accuracy error is formally defined as the measured value minus the true value.
• The accuracy error of a reading (which may also be called inaccuracy or
uncertainty) represents a combination of bias and precision errors.
• The overall accuracy error (or the overall inaccuracy) of a set of readings is
defined as the average of all readings minus the true value. Thus, overall
accuracy error is identical to systematic or bias error. B
Precision
• Precision is a term used for instruments that characterizes the random error of the
instrument output.
• Precision error is defined as the reading minus the average of readings. Thus,
precision error is identical to random error.
• Instrument precision is usually associated with the number of digits displayed on
the output, i.e., its resolution. Be careful though – an instrument can be very
precise, but not very accurate, and vice-versa.
• The drawings to the right show the difference between precision and accuracy. Two people, A and B, shoot
guns at targets. Both people shoot eight times. Each plus sign marks the spot where a bullet hits the target.
 • It is proper to say:
ƒ A is more precise than B, but not as accurate.
ƒ On the other hand, B is more accurate than A, but not as precise.

• Example:
Given: Consider the same five temperature readings as in the previous example, i.e., -1.30, -1.50, -1.20,
-1.60, and -1.50oC. Also suppose that the true temperature of the snow was -1.45oC.
To do: What is the accuracy error of the third data point in oC? What is the overall accuracy error?
Solution: The accuracy error (that is, the inaccuracy) of this data point is defined as the reading minus the
true value, or -1.20 - (-1.45) = 0.25oC. The overall accuracy error is the same as the systematic error or
bias error, which is the average reading minus the true value, i.e., -1.42 - (-1.45) = 0.03oC.

• Example: Four stopwatches are used to measure a time span. The exact, true, or actual time span is
45.623451 … s. What can you say about the accuracy and precision of each stopwatch?

TIMEXAM TIMEXAM TIMEXAM TIMEXAM

46. s 43. s 44.189 s 45.624 s
(a) (b) (c) (d)
ƒ Stopwatch (a) is accurate, but not very precise since it has poor resolution (only two digits), but does the
best job possible within its limits – its reading matches the true reading rounded off to two digits.
ƒ Stopwatch (b) is not very accurate and not very precise since it displays an inaccurate reading compared
to the true value, and also has poor resolution (only two digits).
ƒ Stopwatch (c) has excellent precision (resolution) with 5 digits, but the reading is very inaccurate.
ƒ Stopwatch (d) is both accurate (reading close to true value) and precise (5 digits of resolution).
Other Errors
• There are many other errors, which all have technical names, as defined here:
o zero error: The instrument does not read zero when the input is zero. Zero error is a type of bias error
that offsets all measurements taken by the instrument, but can usually be corrected by some kind of zero
offset adjustment. Zero balance is a term used by manufacturers to indicate the maximum expected zero
error of their instrument.
o linearity error: The output deviates from the calibrated linear relationship between the input and the
output (see further discussion in the next section on calibration). Linearity error is a type of bias error,
but unlike zero error, the degree of error varies with the magnitude of the reading.
o sensitivity error: The slope of the output vs. input curve is not calibrated exactly in the first place. Since
this affects all readings by the instrument, this is a type of systematic or bias error.
o resolution error: The output precision is limited to discrete steps (e.g., if one reads to the nearest
millimeter on a ruler, the resolution error is around +/- 1 mm). Resolution error is a type of random or
precision error.
o hysteresis error: The output is different, depending on whether the input is increasing or decreasing at
the time of measurement. [This is a separate error from instrument repeatability error.] For example, a
motor-driven traverse may fall short of its reading due to friction, and the effect would be of opposite
sign when the traverse arrives at the same point from the opposite direction; thus, hysteresis error is a
systematic error, not a random error.
o instrument repeatability error: The instrument gives a different output, when the input returns to the
same value, and the procedure to get to that value is the same. The reasons for the differences are usually
random, so instrument repeatability error is a type of random error.
o drift error: The output changes (drifts) from its correct value, even though the input remains constant.
Drift error can often be seen in the zero reading, which may fluctuate randomly due to electrical noise
and other random causes, or it can drift higher or lower (zero drift) due to nonrandom causes, such as a
slow increase in air temperature in the room. Thus, drift error can be either random or systematic.
Calibration
• There are two types of calibrations: static calibration and dynamic calibration.
• Static calibration is performed when time is not relevant in the measurement.
• Normally, some output (a voltage, current, etc.) is plotted as it varies with some known reference input, as
sketched.
output

data points

calibration curve fit

input
Here, several data points are taken at known input values, and a calibration curve fit is drawn through the
points. The curve fit can be any function (linear, parabolic, exponential, etc.), but most instruments are
designed to have a linear behavior. In the example above, a straight line has been fitted through the data.
• Dynamic calibration is required when time is relevant to the measurement.
• Normally, the time response of the system or instrument can be found by suddenly increasing or decreasing
the input, and then recording the time response of the output. There are three basic types of time response:

o Zero order (ideal) – In this case, the output increases instantaneously with the input. Here, the input is
suddenly increased at time t0.

output

0
0 t0 time, t
No real system has a perfect, ideal response like this, but this is the goal towards which designers of
instruments shoot.

o First order – For the same instantaneous increase of the input at time t0, a first order system rises
smoothly to its final value, as sketched, with no overshoot.

output

0
0 t0 time, t
As can be seen, there is a time lag in the instrument – some time elapses before the output shows its final
steady-state value. A good example is sticking a thermometer into a pot of boiling water. It takes some
time for the thermometer to read the correct temperature of the water, and the thermometer behaves as a
first-order measurement system.
o Second order – For the same instantaneous increase of the input at time t0, a second order system will
also rise to its final value after some time lag, but there may or may not be overshoot, depending on how
much damping is available in the measurement system. There are generally three possibilities:

ƒ Underdamped – With too little damping, the signal will rapidly overshoot, and then oscillate many
times before settling to the final reading.

output

0
0 t0 time, t
A good example is seen when a person jumps onto a common bathroom scale. The scale behaves as
a second-order measurement system, oscillating several times before settling down.

ƒ Overdamped – With too much damping, the signal will not overshoot at all, but in general will take
much more time to settle to the final reading.

output

0
0 t0 time, t
From these sketches it is hard to distinguish between a first-order system and a second-order system
with overdamping.

ƒ Optimally damped – With just the right amount of damping, the signal will overshoot a little (to
about 5% overshoot), and then quickly settle to the final reading with very small amplitude
oscillations.

output

0
0 t0 time, t
The optimally damped system settles down the quickest of the three. For best time response, many
measurement systems can be adjusted so as to behave with optimal damping. We discuss dynamic
systems such as these in more detail later in the course.