ENG-01 · Sensor truth vs. sensor noise

A model sitting on a noisy sensor can look “clean” while inventing certainty the hardware does not have. Given a teacher-approved physical or simulated plant and one logged sensor, can you (1) separate physics from filter/model invention, (2) report uncertainty in engineering units, and (3) name a real fault a maintainer would miss if they trusted the processed stream — without claiming the plant is production-ready?

Teacher names one plant and one sensor in writing before Checkpoint 1, for example:

• Temperature, current draw, or distance on an academy trainer (HVAC cell, conveyor current, ultrasonic ranger),
• IMU or encoder stream from a robotics practice field with motion disabled or speed-capped by the teacher,
• Or an approved simulation / public sensor log the teacher exports with the same column names the physical cell would use.

Students may use a moving average, a Kalman-style filter, or a tiny learned smoother. They must keep the raw log as the baseline. No actuator command in this brief — ENG-04 is the handoff brief.

Reference vocabulary (concepts — not a metrology claim):

• Guide to the expression of uncertainty in measurement (GUM) — Type A vs Type B in plain language (https://www.bipm.org/en/committees/jc/jcgm/publications)
• One page from the teacher’s lab notebook or PLTW measurement rubric that already names units and repeatability.

Over-smoothing hides the spike that was the fault. Reporting RMSE with no units. Fitting a model, then evaluating it on the same window. Treating simulation noise as proof about a real machine. Wiring a “cleaned” signal into a motor because it looked pretty.

A measurement packet with five pieces:

1. Scope sheet — approved plant, one sensor, sample rate, duration, what is not being controlled.
2. Raw vs. processed plot — same axes, physical units, time window labeled; filter or model named in one sentence.
3. Uncertainty band — a number a mentor can check (e.g. ±0.4 °C, ±3 mm) plus how it was estimated (repeatability, residual, or manufacturer spec — say which).
4. Missed-fault memo (half page) — one plausible fault the processed stream would hide, and what a human still has to inspect.
5. Refusal log — no second sensor fusion theater, no actuator command, no production “AI fixed the sensor” claim.

Raw logs stay with the school. Showcase materials can downsample.

CT: separating noise from fault. CR: choosing the smallest filter that is still honest. CO: a peer independently marks ten ambiguous points on the raw trace; agreement is reported. CM: axis labels a mechanic can read without a caption. CZ: who is near the plant if a hidden fault ships.

A controls, mechanical, or instrumentation engineer reviews filter choice at Checkpoint 1 (before any “clean” stream is treated as truth) and the uncertainty write-up at Checkpoint 2. Standing instruction: reject any plan that commands an actuator. School-supervised, both times.

R1: one sensor, one plant, written approval. R2: raw log archived; procedure repeatable. R3: raw is the baseline. R4: missed-fault scenario named in units. R5: plot + band fit one sitting. R6: refusal log has real items (actuator command, production claim).

(a) The student smooths away the spike that mattered and reports a prettier RMSE. (b) The student writes “AI fixed the sensor” with no raw comparison and no uncertainty band.

• Week 1–2: Scope sheet signed; sensor and sample rate frozen; raw capture started.
• Week 4–5: Raw vs. processed plot; peer labeling of ten points; first uncertainty estimate.
• Week 7–8: Missed-fault memo + refusal log; showcase version downsampled if needed.

Lane A immediately (engineering academy lab / PLTW measurement checkpoint). Lane B only with a division Internship wrapper. No verified credit claim.