I tell students that measurements are tricky, because we think that when we take measurements we get the truth about some phenomenon. It's intuitive. We take the measurements precisely because we don't know the truth. And sometimes, depending on what's being measured, we might get some data acquisition system that costs a quarter- to half-million dollars US to take the measurements. When we do that, the tendency to believe what we get back is greater.

Using the example of measuring structural displacement of a simple beam, I told them why even if measurements are true, they don't necessarily contain the truth. I think this list is in order of decreasing obviousness (is that a word?), but your mileage varies.

1. The wrong place. If you really want to know about displacement, you measure at a node. (Maybe use FEA to guess where the node is, on a complex structure.) If you measure at an antinode, you get nothing.

2. The wrong time. Sometimes you're interested in the aftermath of an event. (Some complex structures are tested with "twangs" and "bongs.") If you miss the event, you miss the information.

3. The wrong direction. Sometimes 3-D effects just can't be overlooked; sometimes you look for something in the wrong plane.

4. Too quickly. If you don't take measurements over a long enough period, you might miss interesting low-frequency behavior or zero-frequency trends.

5. Too slowly. If you don't take measurements at a high enough frequency, you might miss interesting high-frequency behavior, or (worse yet) have aliasing -- identifying high-frequency behavior incorrectly as something at a lower frequency.

6. Without accounting for uncertainty. Noise is especially at issue here, since noise is commonly found both at the sensor and at the receiver. But uncertainty from other areas (e.g. ambient temperature and pressure, condition of fasteners, "jitter," etc.) is quantifiable in many cases.

7. The wrong expectations. The more complex the system observed, the less likely you'll see just what you expect to see. If you expect one thing, you may miss another; and even if you don't miss it you may be unprepared to take action that you might need to take.

Some students are chagrined when I tell them this -- is it even possible to take meaningful measurements? Sure! But they must be warned anyway, so they know what management will probably ask when they present their findings. Even when we know to consider all this, it never hurts to be reminded. :-)

Here are other considerations in measurement:

• A measurement must to the greatest extent possible be based on known behavior. You'll often find engineers creating math models in advance of measuring data, in an effort to avoid the problems above by predicting behavior.
• A measurement must be repeatable. That means that if you have the same ambient conditions, the same calibration of the same instrument (s), and perform the steps of your experiment (s) in the same order more than once, you should get more-or-less the same result each time. There are, of course, statistical variations. But you can do things to get around most of them.
• A measurement must be verifiable. That means that if someone else uses the same instruments, methods, and conditions you did, they'll get more-or-less the same result you got. Within statistical variations.

Repeatability is particularly important if there's some chance your management will need data to be taken again with some changes in the conditions. Remember the Law of Vain Repetition: "There's never time to do it right; there's always time to do it over."

Calibration

Calibration is essential to accuracy. Measurement devices typically come with accuracy specs consisting of tables showing the degradation of accuracy with time. These tables are generally based on standards such as generated by the National Institute of Standards and Technology (NIST). When you calibrate an instrument, you verify that your measurements have uncertainty only within the specs.

There are several environmental factors that affect your ability to take accurate measurements:

• temperature
• humidity
• pollen or dust
• chemical vapors
• electromagnetic interference
• voltage supply variations

The last two entries on the above list apply to electronic and computer-attached measurement devices. If you use such devices, be sure to perform an internal calibration periodically. You may also find it necessary to perform an external calibration from time to time -- this requires access to accurate traceable standards, and is best handled by a metrology lab or service. Electronic and computer-based devices should provide hands-off calibration, whether internal or external.

The realization that the world is not fully deterministic comes as a shock to many young (and sometimes old) engineers. It's one thing to have had classes in statistics but for the engineer to find out that one-inch ball bearings are not all created equal may come as a shock.

It takes a few such shocks to make some engineers and scientists actually believe the there might be some utility to a well designed experiment or a statistically valid manufacturing control system.

Sources of Instrumentation Error

• Zero error, from incorrect calibration or drift
• Mechanical effects, such as friction, backlash, wear, etc. (this will probably not concern you if your sensors aren't mechanical in nature, but even electronic sensors have physical parts that can wear)
• Environmental effects, such as humidity, vibration, radiation, etc.
• Reading error, such as
  • parallax (i.e., your line of sight isn't normal to your measurement scale)
  • time-varying readings not accounted for properly

Tuve gives a table of selection criteria for (mechanical) sensors in his book -- since his book is dated (and out of print), your mileage will vary. But here are some of the relevant high points, with electronic/PC-based considerations added in bold and italics: