Integrated testing, diagnostics and trustworthy measurements

Choose where the comparison can challenge the model

State the required quantity, range, response time and relevant environment. Define which installed version and load path the test represents. A reference applied near a sensor may not create the same physical quantity at another location. A test should connect its fixture and reference uncertainty to the quantity being compared.

Check more than a single unloaded point. Offset, scale, hysteresis, saturation and timing defects leave different patterns as conditions change. A simulated input at the converter can test the downstream chain but cannot establish that the installed transducer responds correctly. Conversely, a mechanical check does not verify the byte conversion in the receiving application.

Use a discrepancy to choose the next observation

The figure compares two supplied channel records against nominal reference forces. A nearly constant residual and a residual growing with input suggest different hypotheses. A growing difference is consistent with a scale mismatch, but fixture load sharing, environmental response or interpretation errors may produce similar patterns. A pattern narrows an investigation without proving a unique cause.

Separate a correction from its supporting evidence. Keep the original raw and reported values, reference method, timebase, configuration and transformation version. If a correction is justified, test conditions not used to choose it. Replaying the same three points used to fit a line mainly checks that the implementation reproduces that fit.

Invalid and missing observations are informative

A disconnected channel, overflow or stale conversion should have an identifiable status rather than silently becoming zero. Zero can be a valid physical value. A held last value needs its age so that a changing system is not represented by an apparently current constant. These requirements concern data meaning as much as software reliability.

A sequence 100, 101, 103 indicates a missing record under a declared consecutive sequence convention; it does not reveal what the missing value was. Diagnostics also need tests that check their own detection and reporting path. A passed fault flag cannot establish that every fault mode is covered or that a particular response is safe for every physical system.

Investigate disagreement across a range

At nominal reference forces 0, 25, 50 N, channel A reports .1, 25.1, 50.1 N and B reports 0, 27.5, 55.0 N. Residuals relative to those nominal references are A: .1, .1, .1 N and B: 0, 2.5, 5 N. The B−A differences are −.1, 2.4 and 4.9 N. Matching near zero alone would conceal the growing discrepancy.

A 1.10 scale factor is consistent with B’s supplied numbers, but it must not be installed as a correction solely from this pattern. Check physical load equivalence, independent reference evidence, channel configuration and count-to-unit conversion. Then preserve a bounded retest with uncertainty and additional conditions. The observations are constructed and carry no acceptance limit or instrument qualification.

Try the reasoning

1. Can a simulated converter input validate the installed mechanical sensing path?

   Show the answer and reasoning

   No. It bypasses the transducer and its coupling to the physical system.

2. Does a missing sequence number establish the missing sample value?

   Show the answer and reasoning

   No. It identifies a gap; any reconstruction adds a model and should be labelled.

3. Why test fresh conditions after fitting a correction?

   Show the answer and reasoning

   They can challenge the correction beyond the observations that were used to choose it.

What to carry forward

Keep physical references, raw records and software interpretation connected when diagnosing an integrated measurement.