FAILURE HYPOTHESES
Instrumentation begins with competing explanations.
If a motor stalls, possible mechanisms might include undervoltage, excessive load, thermal derating, control saturation, encoder loss, current-limit activation, or mechanical binding. Logging only motor command confirms what software wanted, not what the machine experienced.
List the likely failure explanations and ask which measurements separate them. Supply voltage, current, commanded torque, measured speed, temperature, and fault-state flags may discriminate the set far better than one thousand unrelated counters.
SMALLEST USEFUL SET
More channels are not free if nobody can reason about them.
Telemetry consumes bandwidth, storage, engineering time, calibration effort, test points, ADC channels, CPU time, and attention. The smallest useful set is the one that preserves the variables needed to reconstruct dominant states and transitions.
Prefer signals with explanatory leverage: rail voltage, current, temperature, timing, command, measured response, mode state, reset reason, and fault flags often explain more than dozens of cosmetic metrics.
TIME
Measurements without shared time are hard to combine.
A current spike and a reset are only causally interesting if their timing relationship is known. Use synchronized timestamps or a common monotonic clock wherever possible. Record sample rate, buffer latency, filtering, and whether events were captured at source time or later after transport.
Timing uncertainty belongs in the measurement model just like voltage accuracy does.
RAW + DERIVED
Keep enough raw evidence to challenge the summary.
Derived metrics such as RMS vibration, packet loss percentage, or estimated battery state can be extremely useful. Preserve raw or minimally processed windows around important events when feasible so later analysis can test whether the derivation hid structure.
Do not let a dashboard become the only remaining record of what the sensor actually produced.
TRIGGERS
High-rate data can be event-driven.
Continuous high-rate recording is often unnecessary. Keep a rolling buffer and freeze pre-event plus post-event data when a threshold, reset, fault flag, watchdog warning, thermal limit, or operator marker occurs.
This “black box” pattern preserves expensive detail around rare failures without filling storage with hours of normal operation nobody will inspect.
TEST POINTS
Digital telemetry does not replace physical access.
Expose critical power rails, references, buses, actuator drive, clocks, and fault lines where an oscilloscope or meter can reach them. Firmware can misreport its own world. Independent physical measurement is how that disagreement becomes visible.
Label the points and preserve a board map. Mystery pads are archaeology, not instrumentation.
CALIBRATION
Instrumentation has its own error budget.
Current shunts need tolerance and amplifier gain accounted for. Thermistors need conversion models. ADC references drift. Timestamp clocks wander. Sensors saturate. Log sensor range, calibration, units, scale factors, and clipping state so later analysis knows when the instrumentation itself stopped being trustworthy.
STATE
Continuous signals need discrete mode context.
The same current or vibration can mean different things during startup, normal operation, degraded mode, shutdown, calibration, or fault recovery. Record system mode, software image, configuration version, and relevant control states alongside physical measurements.
Otherwise analysts are forced to infer operational context from waveforms that may have several valid explanations.
OPERATOR MARKERS
Give humans a way to mark “it just did the thing.”
A physical button, software marker, or synchronized note can tag an event while the operator is observing it. That creates a bridge between subjective symptom and objective telemetry without requiring the operator to remember a timestamp later.
FIELD METHOD
Review instrumentation like architecture.
1. List dominant failure explanations.
2. Map each measurement to the hypotheses it can distinguish.
3. Remove channels that answer no real question.
4. Verify timestamp relationships and sample rates.
5. Trigger high-rate capture around important events.
6. Keep independent physical test access.
7. Revisit the set after every major failure. Instrumentation should evolve with what the machine teaches.
BOTTOM LINE
Instrument uncertainty, transitions, and competing causes.
Version one is where unknowns are cheapest to expose. A little thoughtful instrumentation can convert months of intermittent folklore into one afternoon of evidence.
Telemetry is useful when it changes which explanation survives.
SOURCE TRAIL
Technical starting points.
NASA — Systems Engineering Handbook
NIST — measurement and SI reference material