44 Sensor Noise and Averaging Limits

sensors

sensor

best

practices

44.1 Start With the Measurement Story

Change the Noise Before Adding More Samples

Picture a light sensor whose value jumps beside a motor. Averaging may calm the display while hiding the real cause. The team must learn whether the change is random noise, drift, interference, or a fixed error.

An analog to digital converter, or ADC, turns a varying voltage into a number. Bandwidth means the range of change a measurement path can follow. Signal-to-noise ratio means the useful signal level compared with the noise level. Narrowing bandwidth may reduce fast noise but can also hide a real event.

Record raw values at a fixed input. Move the motor, change sample rate, shorten the wires, add shielding, average several window sizes, and repeat after warm-up. Compare spread, bias, response time, and missed events.

This runway does not prove that averaging fixes every sensor. The deeper sections explain noise sources, signal-to-noise ratio, converter steps, filters, grounding, calibration, uncertainty, and honest release evidence.

Noise averaging only helps when the signal story is understood. Begin by separating random noise from drift and interference, then use SNR and repeated samples to prove whether averaging improves the measurement enough.

44.2 Learning Objectives

After this page, you should be able to:

Identify thermal, flicker, and ADC quantization noise in sensor front-ends.
Estimate whether front-end noise, ADC step size, sensor tolerance, or environment dominates the error budget.
Explain why averaging improves white noise but not slow drift or fixed bias.
Choose between averaging, bandwidth limiting, shielding, grounding, reference checks, and calibration.
Record enough evidence to defend a sensor reading as signal plus uncertainty rather than a single unqualified number.

44.3 After Sensor Lab Best Practices

Sensor Lab Best Practices teaches the defensive workflow: validate readings, filter noise, fuse sensors, monitor health, and use hysteresis. This page explains when those safeguards work and when they only hide the wrong failure mode.

Use it before adding larger averaging windows, changing thresholds, or declaring a sensor healthy. If variation is random and fast, bandwidth limiting and averaging can help. If the value drifts, moves with supply voltage, changes with placement, or disagrees by a fixed offset, the answer is evidence, calibration, or hardware cleanup rather than more samples.

44.4 Overview: Every Reading Comes With Noise

No sensor reading is perfectly still. Even with the input held constant, the value wobbles because of physical noise that no design can remove entirely, only manage. Good lab practice is really about managing the ratio between the signal you want and the noise you cannot avoid — the signal-to-noise ratio (SNR).

Three fundamental noise sources set the floor for most sensor front-ends. Thermal (Johnson) noise is the hiss produced by any resistance simply because it is warm. Flicker (1/f) noise grows as frequency falls, so it dominates slow, near-DC measurements. Quantization noise is the rounding error introduced by the ADC. Real circuits can add capacitive or inductive pickup, shot noise in junctions, and amplifier voltage or current noise on top of those floors. Knowing which one limits you tells you which fix will actually help.

Noise management begins by deciding what independent information is available, not by choosing an averaging window. Inspect Figure 44.1 to compare the evidence sources that a fusion algorithm can legitimately combine.

Figure 44.1: Sensor fusion pipeline combining noisy same-sensor readings and different sensor measurements into a state estimate with uncertainty.

On the left of Figure 44.1, Same Sensor means repeated measurements over time, whereas Different Sensors contributes complementary data; these reduce uncertainty for different reasons. Domain Constraints such as physics models or map priors are marked optional and should be included only when they are defensible. The Fusion Algorithm—examples include Kalman, particle, and EKF methods—weights those inputs by reliability and produces both a State Estimate and + Uncertainty represented by a covariance matrix. That paired output is the key lesson for the lab: noise reduction without an uncertainty estimate can make a number look stable without proving it is trustworthy, so the logging and SNR practices below retain conditions as well as values.

Intuition: you can raise SNR two ways — make the signal bigger (more gain, more excitation, better placement) or make the noise smaller (narrower bandwidth, averaging, lower-noise parts). Both matter, and the best design uses both.

This is why best-practice sensor logs record conditions, not just values. A temperature reading of 24.8 °C is more useful when the log also states the sensor model, sample interval, enclosure state, airflow, supply voltage, and whether the reading is raw, filtered, or calibrated. Without that context, a later reviewer cannot tell whether a 0.4 °C change is a real environmental shift or a power-rail, placement, or averaging artifact.

Noise work also decides when more code will not help. If the signal itself is small because the sensor is mounted too far from the phenomenon, averaging only hides the weak placement. If the ADC input is picking up motor switching spikes, a software moving average may smear the spike into several samples; the real fix is grounding, shielding, RC filtering, or sampling away from the switching edge. Lab discipline means changing one variable at a time and keeping enough evidence to explain why the fix worked.

44.4.1 Overview Knowledge Check

The mathematical gist. A 12-bit ADC over 3.3 V has a 0.8057 mV code step and a uniformly distributed rounding floor of $q/\sqrt{12}=232.6\ \mu$ V RMS. That is about $572\times$ the chapter’s 0.407 µV thermal-noise result for 10 kΩ over 1 kHz, while the ideal converter ceiling is 74.0 dB. Bits reduce quantisation; they do not remove reference noise, sensor tolerance, drift, or aliasing.

44.5 Quantify Noise and Averaging

Thermal noise from a resistance has a voltage spectral density that depends only on temperature and resistance. The calculation is not just theory: it tells you whether the front-end noise floor is likely to matter compared with ADC resolution, sensor tolerance, and environmental variation.

Noise density  = sqrt(4 × k × T × R)   volts per sqrt(Hz)
Total RMS noise = density × sqrt(bandwidth)
SNR (dB)        = 20 × log10( Vsignal_rms / Vnoise_rms )
Resolution      = input-referred RMS noise / sensor sensitivity
k = 1.38e-23 J/K (Boltzmann), T in kelvin, R in ohms

Worked example: thermal noise and averaging

A 10 kΩ source at room temperature (~300 K):
  density = sqrt(4 × 1.38e-23 × 300 × 10000)
          ≈ 13 nV per sqrt(Hz)

Over a 1 kHz measurement bandwidth:
  total noise = 13 nV × sqrt(1000) ≈ 0.41 µV RMS

Now average N independent samples. White noise falls
as sqrt(N), so SNR improves by 10 × log10(N):
  N = 16  ->  sqrt(16) = 4x less noise  ->  +12 dB SNR
  Every doubling of N  ->  +3 dB.

12-bit ADC at 3.3 V:
  1 count = 3.3 / 4096 = 0.806 mV
  if the sensor changes 10 mV per degree C,
  one count is about 0.081 degree C.

Noise-limited resolution:
  if the integrated input noise is 20 uV RMS
  and sensitivity is 10 mV per degree C,
  the smallest trustworthy change is about 0.002 degree C
before sensor accuracy, drift, and environment are added.

Two practical takeaways fall out of the math: narrowing the bandwidth to only what the signal needs directly cuts integrated thermal noise, and averaging buys SNR but with diminishing returns — going from 16 to 256 samples is another 12 dB, but it costs 16× the time. The ADC example also shows why quantization is not always the bottleneck; if the sensor's own accuracy is ±0.5 °C, chasing 0.01 °C ADC steps is wasted effort unless the rest of the measurement chain supports it.

A practical lab sequence is to collect a still-air baseline first. Hold the sensor at a stable reference, sample at the intended rate, and compute mean, standard deviation, minimum, maximum, and obvious outliers. Then repeat after changing exactly one variable: add a fan, move the cable near a motor lead, change the supply, or alter the averaging window. If the standard deviation falls as roughly 1/sqrt(N), the noise is mostly white and averaging is doing real work. If the mean walks while the standard deviation looks small, you are seeing drift or placement error, and the fix is calibration or environment control.

Use hysteresis and plausibility checks at the decision boundary. For a humidity controller, a filtered value of 60.1% should not switch a relay on and off every sample if the sensor noise is ±1%. A better rule might turn on above 62% and turn off below 58%, while rejecting impossible jumps such as 20 percentage points in one second unless a separate diagnostic explains them. These small guardrails turn a classroom reading into a repeatable field measurement.

The next workbench makes the front-end crossover visible instead of leaving it as separate equations. Begin with the wide noisy preset, where the integrated analog floor dominates. Change only the upper band edge from 1010 Hz to 110 Hz: the shaded band contracts, SNR improves, and quantization becomes the larger remaining floor—but the converter itself remains exactly 16 bit.

A noise total is meaningful only after every contributor is referred through the same signal path. Figure 44.2 locates six contributors before the final bandwidth integral.

Noise matrix with sensor, bias resistor, amplifier input, feedback, and output columns, followed by an integrate row that root-sum-squares power spectral density over bandwidth. — Figure 44.2: Input-referred noise budget matrix locating shot, flicker, resistor thermal, amplifier voltage and current, and mechanical thermal noise.

The shot and mechanical thermal rows of Figure 44.2 enter at the sensor, while the amplifier row adds both en and in times source impedance. Only the final integrate row combines S1 through S4 as input-referred RMS; adding spectral-density amplitudes directly would overstate correlated structure and mishandle bandwidth.

44.5.1 Build an Input-Referred Noise Budget

A front end contains several noise sources, and their units must agree before they can be combined. Shot noise is a current-noise process associated with discrete charge crossing a junction. For DC current $I_{DC}$ , its one-sided current power spectral density is

S_{i,\text{shot}}(f)=2qI_{DC}\quad\text{A}^2/\text{Hz},

so the amplitude density is $i_{n,\text{shot}}=\sqrt{2qI_{DC}}\ \text{A}/\sqrt{\text{Hz}}$ . The square root matters: $2qI_{DC}$ is a power density, not an amplitude density.

At low frequency, junctions and FETs can add flicker noise. A useful empirical model is

S_{i,1/f}(f)=\frac{K I_{DC}^{\alpha}}{f^{\beta}}, \qquad \beta\approx1,

with $K$ , $\alpha$ , and the corner frequency taken from measurement or a datasheet. It is common to write a combined model $S_i(f)=2qI_{DC}+K I_{DC}^{\alpha}/f^{\beta}$ , but $K$ is device-specific; the expression is not a universal junction constant.

Now refer every contribution to the amplifier input. An amplifier voltage-noise density $e_n(f)$ is already in volts per square-root hertz. Its current-noise density $i_n(f)$ becomes a voltage through source impedance $Z_s(f)$ , while a source resistance contributes Johnson noise $4kTR$ . Over measurement band $f_L$ to $f_H$ ,

v_{n,\text{in}}^2= \int_{f_L}^{f_H} \left[e_n^2(f)+i_n^2(f)|Z_s(f)|^2+4kTR+S_{v,\text{sensor}}(f)\right]df.

Independent terms add as powers inside the integral, not as worst-case amplitudes. Mechanical thermal noise belongs at the transducer input and is converted through the sensor sensitivity before it joins this voltage budget. Aliased out-of-band noise must be limited by an analog filter; a digital average cannot undo noise already folded into baseband.

As an instrumentation-amplifier scale example, use approximately $4\ \text{nV}/\sqrt{\text{Hz}}$ voltage noise and $200\ \text{fA}/\sqrt{\text{Hz}}$ current noise in the flat region, then confirm the exact frequency, gain, and source conditions from the selected device datasheet. With a 10 k $\Omega$ resistive source and a 10—100 Hz rectangular noise bandwidth:

Input-referred term	Density	RMS over 90 Hz
Amplifier voltage noise	$4\ \text{nV}/\sqrt{\text{Hz}}$	$38\ \text{nV}$
Amplifier current noise through 10 k $\Omega$	$2\ \text{nV}/\sqrt{\text{Hz}}$	$19\ \text{nV}$
10 k $\Omega$ resistor at 300 K	$12.9\ \text{nV}/\sqrt{\text{Hz}}$	$122\ \text{nV}$

The root-sum-square result is

v_{n,\text{in}}\approx\sqrt{38^2+19^2+122^2}\ \text{nV} \approx129\ \text{nV RMS}.

For a $10\ \mu\text{V RMS}$ input signal, the input-referred SNR is

\operatorname{SNR}_{dB}=20\log_{10}\!\left(\frac{10\ \mu\text{V}}{0.129\ \mu\text{V}}\right) \approx37.8\ \text{dB}.

Finish the budget by dividing $v_{n,\text{in}}$ by sensor sensitivity to obtain resolution in pascals, newtons, degrees, or the required engineering unit. Then add drift, calibration uncertainty, nonlinearity, and ADC quantisation separately. SNR describes random-noise separation over a stated bandwidth; it is not the whole accuracy claim.

44.5.2 Bias-Dependent 1/f Noise: The Hooge Relation

Johnson noise exists even with no applied bias. Flicker noise is different: in many semiconductor and resistive devices its measured voltage noise grows with the DC voltage across the device and rises as frequency falls. A common empirical form is the Hooge relation

\frac{S_V(f)}{V^2}=\frac{\alpha_H}{Nf},

where $S_V$ is voltage-noise power spectral density in $\mathrm{V^2/Hz}$ , $V$ is the DC bias voltage, $N$ is the effective number of mobile carriers, $f$ is frequency, and $\alpha_H$ is a dimensionless empirical coefficient for the material, geometry, contacts, and fabrication process. Taking the square root gives the amplitude spectral density:

e_n(f)=\sqrt{S_V(f)}=|V|\sqrt{\frac{\alpha_H}{Nf}}\quad\mathrm{V/\sqrt{Hz}}.

This dimensional form is safer than memorising a transcription with a stray resistance term. Resistance enters through the bias and device physics. Under constant-current bias, $V=IR$ , so

e_n(f)=|I|R\sqrt{\frac{\alpha_H}{Nf}}.

For a uniform resistor, $R=\rho L/A$ while $N=nAL$ . At fixed material and length, making the cross-section smaller raises $R$ and reduces $N$ , both of which increase flicker noise. Across unrelated parts, however, resistance alone cannot predict $N$ or $\alpha_H$ ; the datasheet or a measured spectrum must supply the device-specific evidence.

Because $S_V\propto1/f$ , integrate power—not amplitude—over the measurement band:

v_{n,\mathrm{rms}} =\sqrt{\int_{f_L}^{f_H}\frac{\alpha_HV^2}{Nf}\,df} =|V|\sqrt{\frac{\alpha_H}{N}\ln\!\left(\frac{f_H}{f_L}\right)}.

Walk the estimate. Suppose a biased sensing resistor has $V=1.0\ \mathrm{V}$ , $\alpha_H=2\times10^{-3}$ , and $N=10^9$ effective carriers. At $1\ \mathrm{Hz}$ , $e_n\approx1.41\ \mathrm{\mu V/\sqrt{Hz}}$ . Integrated from $0.1$ to $10\ \mathrm{Hz}$ , the estimate is $1.0\sqrt{(2\times10^{-3}/10^9)\ln(100)}\approx3.0\ \mathrm{\mu V\ RMS}$ . Halving the bias halves this noise voltage; moving the lower cutoff down by a decade adds another equal logarithmic band of noise power. That is why a long average can stop improving even though the white-noise calculation predicts another $\sqrt{N}$ gain.

Use the relation as a comparison model, not a universal guarantee. Measure several bias levels: Johnson noise remains essentially bias-independent, while device flicker noise should scale roughly with $V^2$ in power if the Hooge model applies. Then record the corner where the falling 1/f spectrum meets the white floor, because that corner decides whether chopping, auto-zeroing, a shorter observation window, or a different sensor technology will help.

44.5.3 Practitioner Knowledge Check

44.6 Averaging White Noise, Not Drift

The most important nuance in noise reduction is that averaging only works on noise that is uncorrelated between samples. That is true of thermal and quantization noise when the samples are far enough apart or the front-end bandwidth is broad enough. It is not true of flicker (1/f) noise or slow drift, whose energy piles up at low frequencies and moves slowly relative to your averaging window.

White noise: averaging wins

Thermal and quantization noise are flat across frequency. Successive samples are independent enough, so averaging N of them reduces RMS by sqrt(N) as promised. This is the case where longer sample windows are justified, as long as the added response lag is acceptable.

1/f noise and drift: averaging stalls

Flicker noise and thermal drift change slowly. Over a long average they look like a wandering baseline, not random hiss, so more samples stop helping and a long integration can even track the drift. A stable-looking average can still be wrong if the baseline moved during the run.

Beating low-frequency noise

Chopper-stabilised and auto-zero amplifiers, ratiometric measurement, and keeping each measurement short relative to the drift timescale attack 1/f and drift where plain averaging cannot. In firmware, periodic zero checks and reference-channel reads serve the same purpose.

Bias is not noise

A constant offset averages to itself. No number of samples removes a systematic error — that is a calibration problem, not a noise problem, and confusing the two wastes effort. Treat offset, gain error, and nonlinearity as calibration evidence, not as random variation.

Nonlinearity can rectify vibration into offset

A sensor whose response is only approximately linear can turn an AC disturbance into a DC bias. If an accelerometer channel behaves like v = S*a + E*a^2 and the unwanted vibration is a = A*sin(wt), the squared term contains sin^2(wt). Its average is not zero, so part of the vibration becomes a steady offset. This is vibration rectification: the plot may look calmer after averaging, while the reported baseline has still moved.

Where do S and E actually come from? Any smooth sensor transfer function is only linear locally — write it as a Taylor series about the operating point and the coefficients fall out on their own. Take a resistive half-bridge, one of the most common analog sensor readouts: a fixed resistor R stacked over a sensing resistor Rsense = R0 + ΔR, giving Vout = Vin·Rsense / (Rsense + R). Expanding Vout in powers of ΔR around the rest value R0:

Vout = Vin·R0/(R+R0)          [offset]
     + ΔR·Vin·R/(R+R0)^2      [linear term  -- this is "S"]
     - ΔR^2·Vin·R/(R+R0)^3    [quadratic term -- this is "E"]
     + higher-order terms ...

quadratic term / linear term  =  ΔR / (R + R0)

The first term is a fixed offset, the second is exactly the sensitivity S the vibration-rectification model above calls linear, and the third is exactly the error term E that rectifies vibration into a DC shift. The ratio at the bottom says how fast that error grows relative to the useful signal: proportional to the fractional change in the sense resistor, ΔR/(R+R0). Biasing the bridge near mid-range and keeping ΔR small relative to R+R0 is therefore not just good practice — it is the literal condition for the linear-only model to hold. Push Rsense through a large swing and the same Taylor expansion that produced E also bends the whole transfer curve away from a straight line, which is the same nonlinearity the datasheet's percent-of-full-scale linearity spec is quietly bounding.

The cure is not a longer moving average. Keep the sensor in its linear region, lower the mechanical vibration reaching it, narrow the bandwidth to the required signal, and record a zero-reference check before trusting small offsets.

So the diagnostic question in the lab is always: is this variation random and fast, or slow and structured? Random and fast yields to bandwidth limiting and averaging. Slow and structured needs chopping, ratiometric technique, reference checks, or calibration. Applying the wrong remedy is the most common way a careful-looking measurement stays wrong.

A useful field diagnostic is the "pause and compare" test. Log raw values, filtered values, and the decision output at the same time. If the raw value jumps but the filtered value recovers, the filter is absorbing random noise. If raw and filtered values both creep in the same direction, suspect drift, heating, supply movement, or the reference condition itself. If two redundant sensors disagree by a fixed amount, treat that as calibration mismatch; if they disagree only during motor activity, treat it as coupling or grounding. That evidence-first habit is what makes sensor best practices auditable.

44.6.1 Under-the-Hood Knowledge Check

44.7 Release Checklist

Before treating a sensor result as reliable, record these checks:

Start by stable-reference mean, standard deviation, min, max, and outlier count. Then sample rate, averaging window, filter bandwidth, and response-lag trade-off. Next evidence that noise falls near the expected 1/sqrt(N) trend when averaging is claimed. After that drift, supply, placement, grounding, and reference-condition checks when the mean moves. Finally calibration evidence for fixed bias, gain error, nonlinearity, or redundant-sensor mismatch.

44.8 See Also

Start by Sensor Lab Best Practices for validation, filtering, fusion, health monitoring, and hysteresis labs. Then Sensor Data Processing for filtering, calibration, and persistence patterns. Next Sampling, Aliasing, and Anti-Alias Boundaries for acquisition checks before digital filtering. Finally Sensor Circuits and Signal Conditioning for analog filtering, grounding, and ADC front-end design.

44.9 Next

Return to Sensor Lab Best Practices after the noise and drift evidence is documented, then use Sensor Calibration when the remaining error is systematic.