# Data qualityData qualityThe degree to which data is fit for purpose: accurate, complete, consistent, timely, valid and unique. Poor quality data undermines analytics, reporting and AI.View full definition → metrics that matter on the shop floor
At 2:14 a.m., a torque sensor on Line 3 starts logging readings with a timestamp lag of 40 seconds. Nobody notices. Three days later, a batch of automotive brake brackets ships with a mislabeled heat-treatment code. The error surfaces during a customer audit, not before. The root cause was never the torque sensor itself. It was a data qualitydata qualityThe degree to which data is fit for purpose: accurate, complete, consistent, timely, valid and unique. Poor quality data undermines analytics, reporting and AI.View full definition → failure that nobody was measuring: timeliness drift that broke the link between sensor readings and the batch record.
This lesson is about the metrics that would have caught it first.
Manufacturing data has a split personality. On one side: high-frequency, machine-generated sensor data (temperature, vibration, torque, cycle time), often via SCADA (Supervisory Control and Data Acquisition) systems or PLCs (Programmable Logic Controllers). On the other: transactional data (work orders, batch genealogy, purchase orders, quality inspections), usually living in an ERP (Enterprise Resource Planning) or MES (Manufacturing Execution System, the software layer that tracks production in real time between ERP and the plant floor).
These two worlds run at wildly different speeds and get reconciled poorly. A batch record in the MES might say "heat treatment completed 14:32." The furnace sensor log might say the last valid reading was 14:28, with a gap after that. Nobody checks the gap. That gap is where mislabeled batches hide.
Definition: the percentage of expected data points actually captured, with no missing fields or dropped records.
Example: a vibration sensor on a CNC (Computer Numerical Control) spindle should log a reading every 5 seconds during a production run. Over an 8-hour shift, that's 5,760 expected readings. If only 5,200 arrived, completeness is:
Completeness = (Records received / Records expected) × 100
= (5,200 / 5,760) × 100
= 90.3%A 90% completeness rate sounds fine until you realize the missing 10% clusters right around a tool change, exactly when defects are most likely. Aggregate completeness hides the failure. Always check completeness *by time window*, not just as a daily average.
Benchmark (industry estimate, as of 2025): best-in-class manufacturers target above 98% completeness for safety-critical sensor streams (source: general guidance from NIST's Manufacturing Extension Partnership, treat as directional, not a hard standard).
Definition: the lag between when an event occurs and when it's recorded and available for use.
This is the metric that would have flagged our opening scenario. A 40-second lag on a torque reading might seem trivial, but if the MES uses that reading to auto-approve a batch release, a lag means the approval logic is working with stale data.
Timeliness is usually measured as latency (time from event to data availability) against a defined SLA (Service Level Agreement, an agreed performance threshold). For real-time quality gates, sub-second to few-second latency is often required. For daily production reporting, same-day latency (T+0 or T+1) is common.
Definition: how closely recorded data matches ground truth.
Accuracy is the hardest to measure because you rarely have perfect ground truth to compare against. Common proxies:
A mislabeled batch is fundamentally an accuracy failure: the label said one heat-treatment code, the process record said another. Catching it requires reconciling two independent sources, not trusting either one alone.
Definition: the percentage of records that are redundant copies of the same event.
Duplicates happen constantly in manufacturing IT: a network retry sends the same sensor reading twice, or a work order gets re-entered manually after an ERP sync failure. Duplicates inflate counts (a batch might appear to have passed inspection twice) and can mask a missing real record behind a false positive.
Duplication rate = (Duplicate records / Total records) × 100Benchmark (estimate): under 1% is considered healthy for automated sensor feeds; above 3% usually indicates an integration or network retry problem worth investigating.
None of these metrics work in isolation. A practical shop floor data governancedata governanceData governance is the set of policies, roles, and processes that ensure data is accurate, secure, well-defined, and used responsibly across an organization.View full definition → routine, run daily or per shift, looks like:
1. Completeness check by machine and time window, flagged if below 95%.
2. Timeliness check on critical control points (batch release, quality gate), flagged if latency exceeds SLA.
3. Accuracy reconciliation between MES and ERP for batch identifiers, quantities, and process codes.
4. Duplication scan on transactional tables before they feed into reporting or customer-facing quality certificates.
This is the essence of data governancedata governanceData governance is the set of policies, roles, and processes that ensure data is accurate, secure, well-defined, and used responsibly across an organization.View full definition →: the policies, roles, and processes that ensure data is trustworthy enough to act on. In regulated manufacturing (automotive under IATF 16949, medical devices under FDA 21 CFR Part 820 in the US or the EU Medical Device Regulation), these checks aren't optional. Auditors will ask for evidence of them.
Knowledge check
1. In the opening scenario, what was the true root cause of the mislabeled batch shipping to a customer?
2. Why does the lesson describe manufacturing data as having a 'split personality'?
3. A vibration sensor is expected to log a reading every 5 seconds during a shift, but many readings arrive late, clustered, or out of order relative to when they were supposed to occur. Which metric would primarily flag this issue, distinct from simply counting missing records?
4. Select ALL correct answers about why the gap between MES batch records and sensor logs (e.g., '14:32' vs. last valid reading at '14:28') is dangerous on the shop floor.
Select all the correct answers.
5. Select ALL correct answers about calculating and interpreting completeness for shop floor sensor data.
Select all the correct answers.
Back to our brake bracket batch. Here's how the four metrics would have caught it before shipment:
A simple automated rule (pseudocode, the kind of logic a data engineer would implement in a quality dashboard):
if mes_heat_code != erp_shipment_code:
flag_batch(batch_id, reason="heat_code_mismatch")
hold_shipment(batch_id)This single reconciliation rule, run before shipment release, is cheaper than a failed customer audit and far cheaper than a recall.
For a primer on data qualitydata qualityThe degree to which data is fit for purpose: accurate, complete, consistent, timely, valid and unique. Poor quality data undermines analytics, reporting and AI.View full definition → dimensions applicable beyond manufacturing, the DAMA-DMBOK (Data Management Body of Knowledge) is the reference framework used across industries, including manufacturing IT teams building governance programs.
🎬 [VIDEO: "What is Data QualityData QualityThe degree to which data is fit for purpose: accurate, complete, consistent, timely, valid and unique. Poor quality data undermines analytics, reporting and AI.View full definition →? (Completeness, Accuracy, Consistency Explained)" - youtube.com - a plain-language walkthrough of core data qualitydata qualityThe degree to which data is fit for purpose: accurate, complete, consistent, timely, valid and unique. Poor quality data undermines analytics, reporting and AI.View full definition → dimensions, useful as a refresher for non-technical managers]