Question 1 of 7: Calibration and Measurement Fundamentals

Question 1: Calibration and Measurement Fundamentals (20 marks)

Question text not reproduced: the examination questions are © Engineers and Geoscientists BC. Open the official past paper (linked at the top of this page) to read the question, then follow the worked solution below.

a) Sensitivity is raised by making the sensing element respond strongly to the target measurand — but the physical/chemical mechanism that gives a strong response to the target very often responds just as strongly to other, unwanted stimuli (temperature, humidity, other chemical species, mechanical vibration). Raising the gain of the transduction mechanism amplifies everything it couples to, target and interferent alike, so a highly sensitive design tends to lose selectivity (its ability to respond ONLY to the intended measurand) unless additional, separate elements (selective membranes, optical filters, a reference/compensation channel) are added — and those elements themselves usually attenuate some of the raw sensitivity they were meant to protect.

b) Run the calibration a second time with a KNOWN, controlled amount of a likely interferent present (a cross-sensitive gas, a temperature offset, a nearby vibration source) in addition to the normal target-only calibration, and compare the two response curves. Selectivity is then quantified as the ratio of the instrument's response to a given magnitude of the target versus its response to the same magnitude of the interferent — a purely target-only calibration can never reveal this, since it never exposes the instrument to anything it should be rejecting.

c) Repeating the calibration separates RANDOM error from systematic bias: the spread (standard deviation) of readings across repeated runs at the same input quantifies precision/repeatability, while any shift of the average curve between runs reveals short-term instability or drift. A single calibration run cannot distinguish "this instrument is inherently noisy" from "this instrument gave one unlucky reading."

d) Cycle the input over its full working range twice — once increasing, once decreasing — and compare the output reading at the SAME input value on the up-going and down-going branches. The maximum difference between the two branches, expressed as a percentage of full scale, is the hysteresis error; a single-direction calibration sweep can never reveal it.

e) Dynamic response is described by the parameters of the instrument's governing differential equation: for a first-order sensor, the time constant $\tau$; for a second-order (underdamped) sensor, the natural frequency $\omega_n$ and damping ratio $\zeta$ (from which rise time, percentage overshoot and settling time all follow), plus bandwidth (the frequency at which the response falls 3 dB) and phase lag versus input frequency.

f) Subtracting two independently measured (and therefore independently noisy) quantities propagates the uncertainty of BOTH into the difference — the absolute uncertainties combine in quadrature even though the signal of interest (the difference itself) is often much smaller than either raw reading. The RELATIVE uncertainty of the baseline-corrected result can therefore be far worse than either original reading's own relative uncertainty, the classic "small difference of two large, similarly-sized numbers" precision-loss problem. A baseline that has drifted between when it was measured and when the signal was measured adds a further systematic error on top of this random-error penalty.

g) The RMS value aggregates every calibration-point residual into a single number that (i) cannot cancel — because each residual is squared before averaging, a large positive error and a large negative error both increase the RMS, unlike a plain mean, which could hide them by cancellation — and (ii) weights larger errors more heavily than small ones (a residual twice as large contributes four times the squared term). The result is a single, physically meaningful (same units as the measurement), robust summary of the TYPICAL error magnitude across the whole calibrated range, which is exactly what "how successful was this calibration" is asking for.

h) Numerically differentiating a measured signal multiplies its noise spectrum by frequency, so broadband measurement noise — which by definition has significant high-frequency content — is amplified far more than the slowly varying true signal buried underneath it. Even a small amount of noise on the raw reading can produce a derivative estimate that is dominated by amplified noise rather than by the real rate of change, making derivative-based calculations numerically ill-conditioned on real (imperfect) measured data. Integration has the opposite, favourable effect — it averages noise out rather than amplifying it.

i) In principle, forever: an exponential approach to a final value never mathematically reaches it. In practice, "how long" is defined as a settling time to within a stated tolerance band of the final value — for a first-order instrument, about $4\tau$ to $5\tau$ reaches 98-99% of the final value; for an underdamped second-order instrument, $t_s \approx 4/(\zeta\omega_n)$ for a $\pm2\%$ band or $3/(\zeta\omega_n)$ for a $\pm5\%$ band.

j) If the calibration curve is not monotonic, a single output reading can correspond to more than one possible input value, and the instrument has no way to tell which of those candidate inputs actually produced the reading — the input-to-output mapping must be inverted to turn a reading into a measurement, and that inversion is only well defined (single-valued) when the forward curve is monotonic.