Sensors as physical translators

Requirement

A controller or a logger needs a number for something physical: position, speed, force, or temperature. A motor drive needs to know how fast its shaft turns, a battery charger needs the cell temperature, a test rig needs the force in a bracket, a printer needs the carriage position. The circuit that must act on these quantities reads electrical signals and nothing else, while the quantities themselves live in the mechanical and thermal world. Something has to stand between the two, exposed to the physical quantity on one side and producing a readable signal on the other. That device is a sensor, and the requirement it must meet is specific: deliver a number for the quantity over the whole span it will actually cover, fine enough to catch the changes that matter, fast enough to keep up with them, and with an honestly stated error.

Variables

Eight quantities frame every sensor decision; they are defined here and used through the rest of the entry. The input quantity x is the physical thing being measured, in its own units: meters, newtons, degrees Celsius. The output signal y is what the sensor produces, in its own units: volts, counts from a converter, millimeters of a liquid column. The range is the span of x over which the sensor gives a usable y. The sensitivity S is how much y changes per unit change in x. The noise N is the random wander of y while x is held fixed, quoted in output units. The resolution is the smallest change in x that can actually be read; it is derived from S and N below rather than copied off a datasheet. Drift is the slow shift of y at a fixed x over hours and months. The response time is how long y takes to settle after x changes.

Design choices

Sensors work the same way underneath: the measured quantity is allowed to alter some property of the device itself, and that altered property is what gets read. Temperature changes the resistance of a metal wire; force flexes a beam and stretches a foil pattern glued to it; position moves a wiper along a resistive track; speed sweeps a magnet past a coil. A device that converts a quantity in one physical form into a signal in another form is called a transducer, and a sensor is a transducer pointed at a measurement task. Plot the output y against the input x, point by point, and you get the curve that characterizes the device; that curve is its transfer function, and everything a sensor does well or badly is a statement about that curve.

Selection falls straight out of the requirement, in three checks. The candidate's range must cover the span x will actually reach, because outside it the transfer function is no longer trustworthy. Its resolution, computed as in the next section, must be finer than the smallest change in x the application cares about, or the reading carries no information at the scale that matters. And its response time must be shorter than the time over which x meaningfully moves, or the signal reports where the quantity was rather than where it is. A candidate that fails any one of the three is out, whatever its other virtues.

A raw y is still not a measurement. It becomes one through calibration: exposing the sensor to known values of x, recording y at each, and fitting the map that converts future readings back into input units, with the known values traceable to a standard. That procedure, and why traceability is what makes two laboratories' numbers comparable, is the territory of the units, standards, and calibration entry; this entry imports it rather than re-teaching it. The burden here is what the calibrated map buys and where it runs out.

The translator picture that names this entry should also be cut down to size here. A translated sentence can be checked by anyone fluent in both languages; a sensor's translation can be checked only against a reference standard, and outside the calibrated range there is no fluency to appeal to, just extrapolation into a region the map was not fitted over.

Calculations

Two numbers, both measurable, decide what a sensor can actually resolve. Fit the calibration points with a line, or with a smooth curve whose local slope you read off; the fitted slope is the sensitivity S, output change per unit input change. Hold x fixed and watch y wander; the spread of that wander is the noise floor N, in output units. A change in the input of size delta-x moves the output by S times delta-x, and that movement is readable only when it climbs out of the random wander, so the smallest input change worth claiming is delta-x = N / S. That derived quantity is the usable resolution, and it can differ badly from the number of digits on the display.

A worked calibration makes it concrete. A temperature probe is placed in five stirred water baths held at 20, 40, 60, 80, and 100 degrees Celsius, each checked against a reference thermometer, and its amplified output reads 0.80, 1.62, 2.38, 3.22, and 4.00 millivolts. A fitted straight line through those points has slope S = 0.040 millivolts per degree. The points miss the line by at most about 0.024 millivolts, which is 0.6 degrees in input units; that is the map's uncertainty, and it gets stated with the map. With the probe left in one bath, the reading wanders by about N = 0.008 millivolts, so the usable resolution is N / S = 0.2 degrees. The display shows three decimal places of millivolts, a least count of 0.001 millivolt that converts to about 0.025 degrees; that figure flatters the probe, since the noise-limited resolution N / S = 0.2 degrees is some eight times coarser, and the arithmetic is the one to trust.

Failure modes

Sensors fail in patterned ways, each with a cause. Drift happens because the transducer's own material ages and shifts with temperature, so the map fitted in January quietly mis-translates in July; the guard is scheduled recalibration, or at minimum a periodic zero check against a known input. Loading happens because the sensor must exchange energy with the system it measures, so measuring perturbs the thing measured: a room-temperature probe dropped into a small cup of hot water cools the very water it reports on. Saturation happens because past the ends of the range the transfer function flattens, and a flat map no longer inverts: two different inputs produce the same output, and the reading pins at a rail while the quantity keeps moving. Aliasing belongs to a different link in the chain. Many measurement chains sample the signal at intervals rather than reading it continuously, and a signal sampled at less than twice its highest frequency is indistinguishable from a slower one, so a fast oscillation masquerades as a slow wander. That failure is the sampling stage's, and no improvement to the transducer touches it.

Limits and boundary conditions

A sensor reports its own state: the resistance of its wire, the strain of its beam, the expansion of its liquid. The target quantity is an inference through the calibrated map, and the inference holds only where and how the map was fitted: within the calibrated range, under the mounting, supply, and ambient conditions of the calibration, and within the drift interval since it was done. Between calibration points the map interpolates, which is defensible on a smooth transfer function; beyond the end points it extrapolates, which is a guess. And the noise floor N bounds resolution from below no matter how finely the output is digitized.

Build with it

Measure the pull force a small winch applies to its line. Stated requirement: forces from 0 to 40 newtons, changes worth catching down to 0.5 newtons, varying over roughly a second. Stated candidates: (a) a strain-gauge load cell rated 0 to 200 newtons, with a noise floor equivalent to about 0.05 newtons and millisecond settling; (b) a pointer spring scale, 0 to 50 newtons, marked every 1 newton, needle settling in about two seconds; (c) an elastic cord read against a ruler, ample range, but with visible creep under sustained load and slow recovery. Choose among them on the three criteria of the Design choices section: range, resolution, and response time. Then calibrate your choice against a known reference by hanging reference masses and converting to force, where a mass m in kilograms pulls with m times g and g = 9.8 newtons per kilogram, fit the map, and state three numbers: the sensitivity S from the fitted slope, the map's uncertainty from its residuals, and the usable resolution N / S from your own noise measurement with the load held fixed. Finally, name one failure mode from the list above that this setup is exposed to, and the guard you would apply. Success is a selection defended against the three stated criteria rather than by habit, a calibration map with its uncertainty stated and the usable resolution derived from your measured N and S, and a named failure mode with its guard.