3. Data quality dimensions for plausibility assessment

Plausibility review is not a single check but a structured judgment across several independent quality dimensions. A posttest dataset can look perfectly well-formed — every field populated, every timestamp present — and still be implausible because two channels disagree, because a value is physically impossible, or because the data arrived too late to describe the condition it claims to represent. This section defines the four dimensions that anchor the review: accuracy, consistency, completeness, and timeliness. Each is assessed with its own evidence and its own failure signature, and each maps to a different quarantine decision.

The dimensions are deliberately orthogonal. Accuracy asks whether a value is close to the true physical quantity. Consistency asks whether related values agree with each other and with known relationships. Completeness asks whether the required observations exist for the required interval. Timeliness asks whether the observation was captured and delivered within the window in which it is meaningful. A dataset can score well on three dimensions and fail the fourth; that is precisely why the review must evaluate them separately rather than collapse them into one pass/fail verdict.

Accuracy: closeness to the physical truth

Accuracy is the degree to which a recorded value matches the actual quantity at the moment of measurement. In posttest review you rarely have ground truth, so accuracy is assessed through plausibility envelopes: engineering limits derived from the test article's design envelope, sensor range, and expected operating conditions. A channel reading outside its physical range is inaccurate by definition; a channel reading inside its range but drifting monotonically when the stimulus was steady is inaccurate by behavior.

Practical accuracy checks combine three techniques. First, range and rate-of-change gates: define a valid minimum and maximum per channel, plus a maximum slew rate. For a pressure transducer rated 0–10 bar sampled at 100 Hz, a jump from 2.1 bar to 9.8 bar within one sample interval implies a slew of 770 bar/s, which exceeds any physical response of the test article and flags the sample as suspect. Second, cross-sensor comparison: redundant or correlated sensors should track within a stated tolerance. Third, reference-anchor checks: pre-test calibration offsets, known ambient conditions at t=0, and zero-load baselines provide fixed points against which the early record can be validated.

Define accuracy envelopes before looking at the data. Deriving limits after seeing an anomaly invites rationalization; a pre-registered envelope makes the review defensible and repeatable.

Consistency: agreement across channels, time, and metadata

Consistency covers internal agreement — between channels, between redundant sensors, between the data and its metadata, and across time. It is the dimension most often missed because each individual channel may pass its own range check while the ensemble is incoherent.

Three consistency relationships matter most in posttest datasets:

  • Physical coupling. Channels linked by physics must move together. If a valve command channel shows the valve closing at t=12.40 s while the downstream flow channel shows no decay until t=12.95 s, either the command timestamp, the flow timestamp, or the flow channel itself is wrong. The 550 ms lag must be explained by actuator dynamics or treated as a consistency defect.
  • Redundancy agreement. Where two sensors measure the same quantity, their difference should stay within a documented tolerance. A persistent bias that grows with temperature suggests one sensor's compensation is failing, not that the quantity changed.
  • Metadata coherence. Sample rates, units, channel scaling factors, and test-configuration identifiers must match the acquisition configuration actually used. A channel logged at 50 Hz but labeled 100 Hz will corrupt every downstream time-based calculation even though the samples themselves are accurate.

Completeness: presence of required observations

Completeness is measured against a defined requirement, not against an abstract ideal. The review must first establish what the test plan required — which channels, at what rate, over what interval, with what maximum allowable gap — and then measure the delivered dataset against that requirement.

Useful completeness metrics include channel coverage (fraction of required channels present), temporal coverage (fraction of the required interval with valid samples), and gap statistics (count, duration, and location of missing runs). A dataset with 99.9% temporal coverage can still be unusable if the missing 0.1% falls exactly on the transient of interest. Completeness defects also include silent substitution: a channel filled by last-value-hold or interpolation during a dropout is present in the file but absent in reality, and must be flagged rather than counted as complete.

Timeliness: fitness of the observation window

Timeliness has two distinct meanings in posttest review, and conflating them causes misclassification. The first is acquisition timeliness: was each sample captured close enough to the event it describes? Latency between a physical event and its recording — buffering, filter group delay, bus arbitration — shifts the record in time. A 40 ms anti-alias filter group delay is harmless for a slow thermal trend but fatal for correlating a fast pressure spike with a command edge. The second is delivery timeliness: did the dataset reach the review process while the test article, calibration records, and operators were still available to resolve ambiguities? A dataset delivered after the rig is reconfigured may be accurate and complete yet impossible to validate.

Timeliness is assessed by comparing event markers, command logs, and independent timing references against channel timestamps, and by recording the interval between test completion and data availability. Both are reported as explicit numbers, not impressions.

Four data quality dimensions mapped to evidence, failure signatures, and quarantine actions A structured diagram showing accuracy, consistency, completeness, and timeliness as four parallel assessment lanes. Each lane lists the evidence used to assess it, a typical failure signature, and the resulting quarantine decision. Posttest Plausibility Review — Four Quality Dimensions Each dimension is assessed independently, then combined into a channel-level quarantine decision DIMENSION EVIDENCE USED TYPICAL FAILURE SIGNATURE QUARANTINE ACTION Accuracy closeness to truth Range & slew-rate gates, cross-sensor comparison, calibration anchors, zero-load baseline at t=0 Value outside physical envelope; impossible slew (770 bar/s); monotonic drift under steady load Quarantine channel; retain for forensic review, exclude from analysis Consistency agreement across data Physical coupling pairs, redundant sensor deltas, metadata vs. acquisition configuration Unexplained 550 ms lag between valve command and flow decay; growing redundancy bias Quarantine the weaker channel; resolve timestamp or scaling before release Completeness presence vs. requirement Test-plan channel list, required interval, gap statistics, last-value-hold detection Missing 0.1% falls on the transient of interest; dropout masked by interpolation Quarantine affected interval; mark substituted samples as invalid, not complete Timeliness fitness of the window Event markers vs. channel timestamps, filter group delay, test-to-delivery interval 40 ms group delay misaligns pressure spike from command; delivery after rig reconfiguration Quarantine time-alignment claims; escalate late-delivery datasets for re-validation A channel passes plausibility only when all four dimensions pass; a single failure routes the channel to quarantine with the failing dimension recorded.

Combining the dimensions into a decision

The four dimensions are assessed independently but combined into one channel-level verdict. The practical rule is conservative: a channel is released only when it passes all four; a single failure routes it to quarantine with the failing dimension recorded as the reason. Recording the dimension matters because it determines the remediation path. An accuracy failure usually requires sensor or calibration investigation. A consistency failure may be resolved by correcting a timestamp or scaling factor without discarding the data. A completeness failure is often interval-scoped, so only the affected window is quarantined. A timeliness failure may be recoverable if the delay is a known, constant filter characteristic that can be compensated.

Consider a worked example from a hydraulic rig test. Channel P_UPSTREAM is logged at 100 Hz over a 60 s window. The review finds: all samples within the 0–10 bar envelope (accuracy pass); a 550 ms lag against the valve command channel that exceeds the documented 120 ms actuator response (consistency fail); 99.97% temporal coverage with a 20 ms gap at t=31.2 s, outside the transient window (completeness pass); and a 40 ms filter group delay that is documented and constant (timeliness pass, with compensation applied). The verdict is quarantine, reason consistency, action: verify the command-channel timestamp source and the flow-channel scaling before re-review. The data are not discarded — they are held pending a targeted correction, which is exactly the outcome a dimension-aware review is designed to produce.

Do not let a strong score on one dimension mask a failure on another. A dataset that is 100% complete and perfectly accurate but time-shifted by an undocumented delay will produce confident, wrong conclusions. Report each dimension's result separately before issuing the combined verdict.

Finally, the review should record the assessment method alongside the result for each dimension — the envelope used, the tolerance applied, the required interval, the timing reference. This makes the plausibility judgment auditable and allows a later reviewer to reproduce it exactly, which is the difference between a defensible quarantine decision and an opinion.

← Previous 📖 Contents Next →