Measurement-Chain Physics for Sampled Telemetry

A physical process is continuous. Telemetry is sampled. A lot of "bad data" is not a broken sensor. It is the measurement chain plus sampling showing up in the stream.

This post is a physics framing for common streaming evidence signals. The goal is not to publish a product. The goal is to explain why these math tools correspond to real-world causes.

A simple measurement-chain model

Let the plant be a continuous variable p(t) such as pressure, flow, level, or temperature. You do not measure p(t) directly. You measure a chain with its own dynamics and noise sources, then you sample it.

A workable discrete model is:

x[n] = G·p(nΔt) + b + η_sensor(nΔt) + η_emi(nΔt) + ∑_k A_k sin(2π f_k nΔt + φ_k) + q[n] + ε_drop[n]

The terms, in order: ideal transduction (gain and bias), sensor element and electronics noise, coupled electromagnetic interference, periodic forcing from mechanical or electrical sources, quantization, and missing/hold/repeat artifacts from the transport layer.

A signal-quality layer is essentially asking which terms are dominating x[n] right now, using bounded, deterministic computations.

1) Exponential smoothing behaves like a physical first-order lag

Exponential smoothing:

y[n] = α x[n] + (1 − α) y[n−1]

matches the shape of a first-order low-pass:

τ dy(t)/dt + y(t) = x(t)

with the usual discrete mapping:

1 − α = e^−Δt/τ   ⇒   α = 1 − e^−Δt/τ

Interpretation:

• τ is inertia. Same shape as real lags from thermal mass, pneumatic volume, hydraulic capacitance, and sensor damping.
• If the real sensor already low-passes the signal, additional smoothing adds more lag.

This is why smoothing is never "free." It trades responsiveness for stability.

2) High-pass by subtraction is AC coupling

Define a high-pass component as:

hp[n] = x[n] − lp[n]

This removes slow baseline so fast departures stand out.

Physical split:

• slow components often come from ramps, calibration drift, temperature coefficients, and very slow process motion
• fast components often come from switching transients, vibration, control actions, and transport artifacts

High-pass is a way to isolate "fast energy" from baseline behavior.

3) A moving average is finite-time integration plus delay

Moving average:

y[n] = (1/N) ∑_i=0^N−1 x[n−i]

Interpretation:

• it integrates over a real time span T = NΔt
• it introduces deterministic group delay on the order of:

t_delay ≈ (N−1)/2 · Δt

Delay matters because it can change how downstream logic behaves, especially when decisions react to trends.

4) Variance and CV are fluctuation power normalized by operating point

Variance over a window estimates the second central moment of fluctuations. It is a proxy for fluctuation power in the band that survives the chain and sampling.

A useful normalized measure is the coefficient of variation:

CV = σ / |μ|

Interpretation:

• σ alone has units, so it confounds "big signal" with "big noise"
• CV is dimensionless and reflects relative noise amplitude compared to the operating level
• high CV near small |μ| often indicates the signal is near the measurement floor, where quantization and input noise dominate

5) Drift detection is a discrete slope estimate

First difference:

dX[n] = x[n] − x[n−1]

Approximate physical slope:

dx/dt ≈ dX[n] / Δt

Interpretation:

• real processes often have bounded slopes due to conservation and capacity
• sensor bias drift tends to be slow and can remain monotone for long periods
• persistence is the key discriminator, not a single step

This is why drift logic usually cares about sustained behavior, not one-off changes.

6) Spike detection maps to impulsive energy

Many physical processes cannot jump by a large amount in one sample interval due to inertia and capacity. Single-sample excursions that violate this constraint are often non-plant artifacts. Common sources include transient coupling, ADC settling artifacts, and transport decode glitches.

The point is not that noise is Gaussian. The point is that normal excitation clusters, and impulsive artifacts do not.

7) Oscillation detection is synchronous detection

A targeted correlation method projects the signal onto sine and cosine at frequency f:

C_s = (1/N) ∑_n x[n] sin(2π f nΔt)

C_c = (1/N) ∑_n x[n] cos(2π f nΔt)

Magnitude:

M = √(C_s² + C_c²)

Interpretation:

• this is the same idea as a lock-in amplifier
• narrowband periodic forcing increases M
• it is useful when you care about a small set of suspect frequencies or bands

A full-spectrum approach via FFT tells the same story across all frequencies.

8) Sampling and aliasing can counterfeit everything

Sampling frequency:

f_s = 1 / Δt

Interference above f_s/2 folds into lower frequencies after sampling. One useful expression is:

f_alias = | f − round(f / f_s) · f_s |

mapped into [0, f_s/2].

Consequence:

• high-frequency interference can masquerade as low-frequency drift or oscillation
• certain relationships between interference frequency and sampling rate can produce aliasing near DC, which looks like baseline wander

This is measurement physics, not a DSP corner case.

9) Missing samples are channel behavior and they distort spectra

Missing ratio:

missing_ratio = missing_count / total

Interpretation:

• it reflects transport reliability and scheduling behavior
• it changes the effective sampling clock
• hold-last-value turns smooth motion into stair-steps, creating artificial spectral content at update boundaries

Missingness is both a reliability metric and a waveform contaminant.

Closing

The plant is continuous, the measurement chain adds structure, and sampling imposes hard constraints. A good signal-quality layer does not need exotic models. It needs bounded evidence signals that map to real causes.

← Back to Writing