Data Quality in Manufacturing: Why Garbage Time-Series Breaks Analytics

Industrial analytics projects usually fail quietly: the model trains, the dashboard looks great, and then the numbers do not match the operators' reality. Most of the time the cause is data quality - bad values that look plausible.

The Usual Suspects

• Stuck values: a sensor that reports a constant reading (cable fault, frozen transmitter). Looks like a stable process, is actually no data.
• Out-of-range and clipped values: readings pinned at the sensor limit, often during the exact events you want to analyze.
• Time skew: devices with unsynchronized clocks make correlation meaningless. Use NTP (or PTP where needed) and record timestamps at the source.
• Unit and scaling drift: a value that silently changes scale (mA to percent, raw to engineering units) between commissioning and today.
• Duplicated or dropped samples: gateway retry logic and buffering can insert or lose points in ways that distort averages.

What to Do About It

1. Store quality flags alongside values (good, suspect, manual, invalid) at the source - historians and platforms support this, most pipelines ignore it.
2. Run automated sanity checks: range, rate-of-change, stuck-value detection, and timestamp monotonicity. Alert on violations.
3. Record a data dictionary: for every tag, the unit, scale, deadband, and acquisition path. Version it like code.
4. Validate before you train: plot a month of raw data, ask the process engineer if it looks right. Human review of the extremes beats any automated check.