The Data Quality Problem in Industrial IoT
Sensors start producing data on the first day. Data you can base a decision on takes longer, and needs a validation layer most budgets leave out.
What the budget assumes
An industrial IoT project is normally costed in three parts: sensors, the network that carries their readings and a dashboard to show them. The unstated belief is that seeing more will mean deciding better.
That holds only if the readings are right, and in practice sensors drift, calibration lapses and connections drop and resume. A dashboard draws all of it as clean lines and numbers, with nothing to mark the doubtful ones.
Errors that do not look like errors
Teams expect bad data to be obvious, like a flat line or a wild spike, and those faults are caught. The damaging errors are the plausible ones.
Exhibit 1: Common sensor data faults, how they appear and how to detect them.
| Fault | How it appears on a dashboard | How to detect it |
|---|---|---|
| Drift | A slow trend that looks like real change | Compare with a reference instrument at set intervals |
| Calibration offset | Steady readings, consistently too high or low | Scheduled calibration with records |
| Connectivity dropout | A gap, or the last value repeated | Flag stale timestamps and missing intervals |
| Wrong tag or unit | Sensible numbers attached to the wrong asset or scale | Commissioning checks and periodic audits |
| Fouling or physical damage | Damped or noisy signal | Compare with neighboring sensors and expected ranges |
A temperature sensor reading a few degrees low gives values that look entirely reasonable, and maintenance intervals or process settings may be based on it for months before anyone checks.
What a validation layer contains
- A calibration schedule for each sensor class, with records of when it was done.
- Automated checks on incoming readings for range, rate of change and staleness.
- Comparison between related sensors, so one that disagrees with its neighbors is flagged.
- A baseline for each sensor, with alerts when its pattern departs from it.
- A defined route for a flagged sensor: who is told, and what happens to its data meanwhile.
- A named owner for data quality, separate from the people who consume the data.
None of this is exotic, since process industries have long managed instrument reliability for control systems. IoT projects led from the IT side sometimes miss that the same care applies.
Five questions before trusting a dashboard
- How would we know if this sensor were wrong?
- What does the screen show when a sensor goes offline?
- Who is responsible for the accuracy of this reading?
- When was it last checked against something independent?
- Which decisions depend on it?
A deployment without validation produces a great deal of data, whereas a deployment with it produces fewer readings that people are willing to act on.
Analytics built on top inherit whatever quality the data has, as shown in predictive maintenance ROI: pilot versus program. Connected sensors also widen the attack surface, covered in cybersecurity exposure in connected industrial operations.
Key takeaways
- Ask when each critical sensor was last calibrated, and whether a schedule exists.
- Check how the dashboard displays a sensor that has stopped reporting.
- Pick one decision made from sensor data last quarter and verify the reading behind it.
- Add calibration, anomaly detection and data ownership to the next IoT budget request.
Check whether your sensor data can be trusted
Synergy Nexus Digital can audit your sensor network for drift, gaps and calibration, and design a validation layer for it.
Cybersecurity Exposure in Connected Industrial Operations
Every sensor and camera added to a plant network is another way in. Most industrial security designs predate them.
Computer Vision in Industrial Safety: Beyond the Hype
Cameras that detect missing PPE and zone breaches extend what a safety team can see. They do not replace the program behind it.
The Real Return on Industry 4.0 Automation Investment
Labor savings are the easiest automation benefit to put in a spreadsheet. Consistency is usually the bigger one.