# Measurement fundamentals: ADC resolution, sensor specs and uncertainty

> Every logged number is an estimate. Know how ADC resolution and sampling limit what you can see, how to read a sensor data sheet, and how to put an honest uncertainty on a result – including whether a setup change really made a difference.

Area: Data Acquisition & Analysis · Starting level: L2 Practitioner · Updated 2026-11-16
Source page: https://race-engineer-academy.vercel.app/lesson/daq-measurement-uncertainty

## L1 · Foundation

**Measured value ≠ true value.** The chain is: physical process (true value) → sensor → signal conditioning → analogue-to-digital conversion → stored number (measured value). Every step adds error. The best you can do is estimate a range within which the true value most probably lies.

**ADC resolution.** An ADC with $n$ bits divides its input range into $2^n$ steps. A 12-bit input on 0–5.5 V has $5500/4096 \approx 1.34$ mV per step. If a 100 mm linear potentiometer outputs 0–5 V, it uses about 3720 of those steps, so one step ≈ 0.027 mm. Wire the same pot to a 0–15 V input ($15000/4096 \approx 3.66$ mV per step) and only ~1365 steps are used: resolution degrades to ~0.073 mm. **Match the sensor's output span to the input range.**

**Why that matters:** damper velocity is a derivative. At 100 Hz, one step of 0.073 mm between samples equals 7.3 mm/s – already inside the low-speed damper range (typically 0–25/50 mm/s). Your velocity histogram would be quantised into useless bins. Note the trap: raising the sampling rate makes the velocity quantum *larger* ($\Delta v = \Delta x \cdot f_s$), so the fix is more ADC bits, a matched input range or a differentiation stencil spanning several samples – not more Hz.

**Sensor data sheet vocabulary:** transfer function (output vs input), sensitivity (slope, e.g., mV/g), offset (output at zero input), measurement range, tolerance/uncertainty, non-linearity, hysteresis, noise and resolution.

## L2 · Practitioner

**Types of error.**

- **Gross errors** – wrong calibration, wrong channel, broken wire, sensor outside range. They can't be corrected afterwards; catch them with plausibility checks.
- **Systematic error (bias)** – consistent offset or scale error from the sensor, method or installation. Remove by calibration against a reference. Classic example: comparing your tyre-pressure gauges with the tyre supplier's gauges and noting each gauge's bias.
- **Random error** – scatter from noise, environment and small uncontrolled effects. Reduce by averaging; quantify with statistics.

**Reading the data sheet.** Segers walks through a capacitive accelerometer with output $V = 2.5 + 0.4a$ (V, a in g) and a tolerance of ±1 % of sensitivity: at 2.5 g, ideal output is 3.500 V but anywhere between ~3.490 and 3.510 V is within spec – about ±0.025 g. Non-linearity (±0.3 %) is inside the tolerance. Decide which accuracy you actually need: water temperature to 0.01 °C makes nobody smarter; damper position can never be too accurate.

**Standard uncertainty of an average.** With $n$ independent samples of standard deviation $\sigma$:

$$\sigma_{\bar{x}} = \frac{\sigma}{\sqrt{n}}$$

Result $= \bar{x} \pm \sigma_{\bar{x}}$ (≈68 % confidence) or $\pm 2\sigma_{\bar{x}}$ (≈95 %). Segers' roll-ratio example: mean 1.12, σ = 3.35 (kerbs and near-zero front roll explode the ratio), n = 10 421 samples → 1.12 ± 0.03 (68 %).

**Is the setup change real?** The most important trackside application. Two runs, 5 clean laps each, lap-time scatter σ = 0.25 s. The standard error of the difference of means is $0.25\sqrt{2/5} \approx 0.16$ s. With so few laps the 95 % multiplier is not 2 but Student's $t$ ≈ 2.3 (8 degrees of freedom), so you need a difference above ~0.37 s to be 95 % confident. A 0.1 s "gain" from a damper click on five laps is not evidence. Use more laps, fuel- and tyre-corrected times, or a more sensitive metric (minimum speed at a specific corner, understeer angle) with lower scatter.

## L3 · Advanced

**Independence is the hidden assumption.** $\sigma/\sqrt{n}$ assumes independent samples. Consecutive samples at 200 Hz are highly correlated – the car does not change state in 5 ms. In the roll-ratio example the *effective* number of independent observations is closer to the number of corners per lap times laps (tens to hundreds), not ten thousand. The real uncertainty is therefore several times larger than the naive value. Practical approach: compute one value per corner (or per lap), then use the scatter of those values.

**Ratio channels are statistically nasty.** When the denominator approaches zero (front roll angle near zero, lateral g near zero), the ratio's distribution has heavy tails and the mean is meaningless. Better: fit a straight line of rear roll vs front roll over gated data and report the slope with its standard error, or use the median. Because *both* axes are noisy, ordinary least squares biases the slope low (regression dilution); Deming or total-least-squares regression treats the two axes symmetrically.

**Uncertainty propagation.** For $y = f(x_1,\dots,x_k)$ with independent errors:

$$\sigma_y^2 \approx \sum_i \left(\frac{\partial f}{\partial x_i}\right)^2 \sigma_{x_i}^2$$

Example: understeer angle depends on steering angle, steering ratio, wheelbase, lateral g and speed. At 1.5 g and 30 m/s on a 2.7 m wheelbase the Ackermann term is $57.3\cdot 2.7\cdot 1.5\cdot 9.81/900 \approx 2.5°$. A 0.02 g accelerometer error changes it by ~0.03°; a 0.3 % speed error by ~0.015°. A steering-angle zero error of 3° (sensor re-installed, or the rack centre moved by an asymmetric toe change) with a 12:1 ratio shifts the result by 0.25° – about seven times more. Sensor zeroing matters more than accelerometer grade.

**Systematic beats random.** In practice the largest errors in motorsport data are systematic: motion ratio wrong by 3 %, tyre radius wrong by 1 %, accelerometer tilted by 1°. Averaging does nothing for them. Invest in calibration procedures before buying higher-resolution hardware.

## L4 · Expert

**Design the test so the statistics can work.** Lap times inside a run are not independent draws: fuel burn makes each lap a few hundredths faster, tyre degradation and track evolution push the other way, and a driver settling in adds a trend of its own. If setup B always runs second, the trend is attributed to B.

**Run order.** Use **A-B-A** (or A-B-B-A) running: compare B with the mean of the two A runs, so a linear drift in track or tyres cancels. Within each run, regress lap time on lap number (or on fuel mass and tyre age), compare the fitted values at equal fuel and tyre age, and use the *residual* scatter as $\sigma$ – the raw scatter of a trending run overstates it, while ignoring the trend biases the mean.

**Correlated samples.** For data with lag-1 autocorrelation $\rho$ the effective sample size is roughly
$$n_{eff} \approx n\,\frac{1-\rho}{1+\rho}$$
(in general $n/(1+2\sum_k \rho_k)$). Laps with $\rho = 0.3$ are worth only ~54 % of their count; samples at 200 Hz are worth almost nothing individually.

**Small samples.** With $n_A + n_B - 2$ degrees of freedom below ~20, replace the factor 2 by Student's $t$: 2.31 for 5 + 5 laps, 2.23 for 6 + 6. For skewed metrics (ratios, minimum speeds with outliers) bootstrap the per-lap or per-corner values to get a confidence interval.

**How many laps?** To show a gain $\Delta$ at the 2-standard-error level with lap scatter $\sigma$ in both runs,
$$n \ge 2\left(\frac{2\sigma}{\Delta}\right)^2$$
laps per run. With $\sigma$ = 0.25 s and an expected 0.10 s gain that is 50 clean laps per run – impossible in a test session. The professional answer is not more laps but a more sensitive metric: minimum speed or time through the affected corner (scatter often 0.3–0.5 km/h or 0.01–0.02 s), understeer angle in a gated phase, or a back-to-back on a correlated simulator or rig.

## Key formulas

**ADC resolution**

$$\Delta V_{LSB} = \frac{V_{range}}{2^{n}}$$

- V_{range}: input range (V)
- n: number of bits
- \Delta V_{LSB}: voltage per step

**Quantisation of a derivative**

$$\Delta v_{min} = \Delta x_{LSB}\, f_s$$

- \Delta x_{LSB}: displacement resolution (mm)
- f_s: sampling rate (Hz)
- \Delta v_{min}: smallest non-zero velocity step (mm/s)

**Standard uncertainty of a mean**

$$\sigma_{\bar{x}} = \frac{\sigma}{\sqrt{n}}$$

- \sigma: standard deviation of samples
- n: number of independent samples

**Significance of a difference between two runs**

$$\Delta\bar{t} > t_{0.975,\nu}\sqrt{\frac{\sigma_A^2}{n_A} + \frac{\sigma_B^2}{n_B}}$$

- \Delta\bar{t}: difference of mean lap times (s)
- \sigma_{A,B}: lap-time scatter of each run (s)
- n_{A,B}: number of clean laps
- t_{0.975,\nu}: Student-t factor for 95 % (≈2 for many laps, 2.3 for 5 + 5 laps)

**Laps per run needed to resolve a gain**

$$n \ge 2\left(\frac{2\sigma}{\Delta}\right)^2$$

- n: clean laps per run
- \sigma: lap-time scatter (s)
- \Delta: expected lap-time gain (s)

## Trackside takeaways

- Match the sensor output span to the logger input range to use the full ADC resolution.
- Derivatives amplify quantisation: displacement resolution × sampling rate = velocity step.
- Calibrate away systematic errors; averaging only reduces random ones.
- Consecutive samples are not independent – compute statistics per corner or per lap.
- Before declaring a setup change better, check the gain exceeds about twice the standard error of the difference (Student-t ≈ 2.3 for five-lap runs).
- Run A-B-A to cancel drift; small gains need corner-level metrics, not 50-lap runs.

## By car category

### GT3

Tyre-pressure gauge bias between team and tyre supplier is a classic systematic error – with mandatory minimum pressures in many GT series, a 0.05 bar gauge error can mean a penalty.

### One-make Cup

With identical cars, half-tenth differences between setups are common claims; insist on enough laps and tyre/fuel-corrected comparisons.

### FIA F3

Short runs on a single compound make lap-time statistics thin; use corner-specific metrics (min speed, understeer angle) to judge changes.

### Hypercar (LMH/LMDh)

Teams invest heavily in calibration rigs (pot calibration, load cell checks) because derived aero and tyre models depend on sub-millimetre accuracy.

### LMP2

Ride-height lasers on a low car need excellent resolution; a 1 mm systematic error is a large fraction of the aero working window.

## References

- Segers, Analysis Techniques for Racecar Data Acquisition, 2nd ed. (SAE 2014): §19.1–19.3 Introduction; A/D conversion; Sensor Selection and Application
- Segers, Analysis Techniques for Racecar Data Acquisition, 2nd ed. (SAE 2014): §19.4 Measurement Uncertainty (roll ratio uncertainty example)
