Analytics is only as good as the readings underneath it. Sensors drift, radios drop packets, clocks wander and firmware bugs produce impossible values. If those problems reach a dashboard or a model unnoticed, people make decisions based on fiction. This checklist is a practical way to find and contain data quality issues in an IoT system.

Why IoT Data Goes Wrong

Business data usually starts life in a controlled application. IoT data starts in the physical world, so it inherits the world’s messiness:

• Sensors age, get dirty, are installed badly or are moved.
• Power and connectivity are intermittent.
• Device clocks are inaccurate or reset after a power loss.
• Firmware changes alter units, scaling or sampling behaviour.
• Devices are repurposed without updating their metadata.

The goal is not perfect data. It is knowing how much to trust each reading and making that trust visible.

Six Dimensions Worth Checking

Completeness. Are expected readings actually arriving? Gaps may be real outages or silent failures.

Validity. Do values fall within physically possible ranges and use the expected units?

Timeliness. Does data arrive soon enough for its use, and are timestamps trustworthy?

Consistency. Do related readings agree, and do devices of the same type behave similarly?

Uniqueness. Are duplicates being counted twice?

Accuracy. Do readings match a trusted reference, such as a calibrated instrument?

The Checklist

Check	What to look for	Where to apply it
Range check	Values outside physical or specified limits	Device and pipeline
Unit check	Mixed units or scaling after a firmware change	Pipeline, using version fields
Missing data	Gaps longer than the expected reporting interval	Pipeline monitoring
Stuck values	A reading that never changes when it should	Analytics layer
Spikes	Sudden jumps inconsistent with physical behaviour	Pipeline and analytics
Timestamp sanity	Future dates, very old dates, clock resets	Ingestion
Duplicates	Repeated messages with the same key	Ingestion and storage
Cross-sensor agreement	Neighbouring sensors diverging unexpectedly	Analytics layer
Metadata accuracy	Device location, type and calibration records	Device registry

Run these automatically and report results per device, per fleet and over time.

Distinguish Missing From Zero

One of the most damaging mistakes is treating absent data as a zero reading. A device that stops reporting is not measuring zero. Store missing values explicitly, and make aggregations aware of them. When you compute an average, know whether it covers the full period or only the moments when the device was awake.

Handle Bad Data Without Hiding It

When a check fails, you have several options, and the right one depends on the use case.

• Reject the reading if it is clearly impossible and keep a count of rejections.
• Flag the reading and keep it, so downstream consumers can choose whether to use it.
• Correct it only when you have a documented, defensible method, such as applying a known calibration offset.
• Interpolate gaps cautiously, and label the filled values so nobody mistakes them for measurements.

Avoid silently deleting or smoothing anomalies. A spike might be a sensor fault, or it might be the most important event of the day.

Quality Scores and Ownership

A simple quality indicator per device, such as the share of expected readings received and passing validation in a period, turns an abstract concern into something a team can act on. Pair it with an owner. Data quality improves when a named team is responsible for the fleet’s health and has a standing review, not when everyone assumes someone else is watching.

Also keep calibration and maintenance records close to the data. If a sensor was replaced or recalibrated on a certain date, analysis should know about the change rather than interpreting a step change as a real-world event.

A Hypothetical Example

Suppose a team tracks temperature in storage rooms. One sensor suddenly reports a flat value for several days. The dashboard shows a calm line, and nobody notices. A stuck-value check, which flags readings that do not vary over a window where variation is normal, would have caught it. In a second case, a firmware update changes a sensor from reporting in one temperature scale to another without a version marker, and averages across the fleet shift oddly. Including a schema version in every message and checking units in the pipeline would have made the cause obvious in minutes.

Building the Habit

1. Start with the three checks that catch the most embarrassing failures: missing data, range and timestamp sanity.
2. Add a quality score to your main dashboard so the trust level is visible next to the numbers.
3. Review quality failures in the same meeting as product metrics.
4. After each incident, add the check that would have caught it.
5. Version your schemas and record firmware changes alongside data.

Frequently Asked Questions

Should bad data be deleted? Generally keep it, flagged, in raw storage so you can investigate and reprocess later. Exclude it from downstream calculations where appropriate.

Can machine learning detect bad sensor data? Anomaly detection can help, but start with simple rules, which are easier to explain and debug. Add models when rules stop being enough.

How do I know if a sensor is drifting? Compare it against neighbouring sensors, a reference instrument, or its own historical behaviour under similar conditions. Periodic calibration checks are the dependable approach.

Who should fix data quality problems? The owner depends on the cause: firmware teams for device bugs, platform teams for pipeline issues, field operations for installation and maintenance problems.

How strict should validation be at the device? Light checks on the device save bandwidth and catch obvious faults, but keep authoritative validation in the cloud where rules can change without a firmware update.

Next Step

Choose one fleet and run the missing-data, range and timestamp checks over the last month of readings. Count the failures per device. The few devices at the top of that list are usually where your biggest quality gains are hiding.