A predictive model is only as honest as the signal feeding it. In a steel plant, sensors sit next to ladle heat, scale, water spray, electromagnetic noise and constant vibration, so readings drift, freeze, spike or disappear without anyone noticing. When that data trains a model or sets a baseline, the result is false alarms, missed faults and a team that stops trusting the system. This article explains how to find and fix sensor data quality problems, and how Oxmaint maintenance management software links sensor health to calibration and work orders.
Steel Plant Sensor Data Quality
Fix the Signal Before You Trust the Prediction
Bad sensor data quietly teaches AI models the wrong normal. Catch drift, dropouts and mislabeled signals early, and route every sensor problem into a tracked maintenance task.
Raw signalNoisy, drifting, incomplete
Health checksRange, rate, flat-line, gaps
Context taggingAsset, state, grade, unit
Trusted datasetReady for models and trends
What poor sensor data does to maintenance decisions
Missed warningA drifting temperature probe reads low while a bearing runs hot, so the alert never fires.
False alarmElectrical noise or a loose connector looks like a vibration spike and sends a technician out for nothing.
Wrong baselineA model retrained on drifted readings accepts degraded behavior as normal.
Lost trustAfter several bad alerts, operators ignore the system, including the one alert that mattered.
Research on industrial condition monitoring consistently ties reliable condition-based maintenance to reliable sensor calibration, which is why calibration strategy belongs in the design from day one.
Six dimensions to measure for every signal
Quality is not a single score. Measure each dimension separately, because each one fails for a different reason and needs a different fix.
01
Accuracy
Does the reading match a trusted reference within tolerance? Checked by calibration and cross-checks against neighboring sensors.02
Completeness
What share of expected samples arrived? Gaps often follow network drops, power events or gateway restarts.03
Timeliness
Is data arriving fast enough for the failure mode? A bearing fault may need seconds, a refractory trend only hours.04
Consistency
Do units, scaling and time stamps agree across systems? Mixed units and clock offsets break comparisons.05
Validity
Are values physically possible and within the sensor range? Impossible jumps usually indicate a fault, not a process event.06
Traceability
Can you link each tag to a physical sensor, an asset and a calibration record? Without it, nobody can fix the source.Recognize the failure pattern, then find the cause
Each data problem has a signature. Use the table to move from what you see on a trend chart to the most likely cause and the maintenance action.
| Pattern on the trend | Likely cause | Where it shows up in a steel plant | Maintenance action |
|---|---|---|---|
| Slow, steady offset from neighbors | Sensor drift or aging element | Thermocouples near furnaces and ladle areas | Calibration check and replacement planning |
| Flat line at one value | Frozen signal, dead sensor or stuck transmitter | Remote conveyor and pump skids | Inspect sensor, wiring and power supply |
| Sudden spikes with no process cause | Electrical noise, loose connection, grounding fault | Drives, welders and high-current areas | Check shielding, terminals and grounding |
| Periodic gaps in data | Gateway, network or power interruption | Wireless nodes in congested or shielded areas | Check connectivity and gateway logs |
| Step change after a repair | Sensor replaced, moved or recalibrated | After roll changes, relines or motor swaps | Record the event and reset the baseline |
| Reading outside physical limits | Wrong scaling, failed channel or open circuit | 4-20 mA loops and analog input cards | Verify loop, scaling and channel health |
| Vibration trend creeping upward on one side only | Loose mounting or degraded sensor attachment | Vibration sensors on gearboxes and fans | Check mounting torque and contact surface |
Why steel plants are hard on sensors
Environmental stress
- Radiant heat and thermal cycling near furnaces, ladles and casters
- Scale, dust and water spray that foul probes and connectors
- Heavy vibration that loosens mounts and cabling
- Corrosive atmospheres in pickling and coating areas
Operational complexity
- Large drives and arc equipment that generate electrical noise
- Frequent changes in grade, speed and load
- Planned outages followed by repairs that move or replace sensors
- Legacy instruments mixed with newer wireless nodes
None of this makes monitoring impractical. It means the plant needs a routine for checking sensors, not just reading them.
From calendar calibration to condition-aware calibration
Fixed calibration intervals are a sound starting point, but they can waste effort on healthy sensors and miss the ones that drift early. A better approach combines both.
Calendar only
- Every sensor is calibrated on the same interval
- Drift between visits goes unnoticed
- Calibration records sit in spreadsheets or paper
- Replacements happen after failure
Calendar plus condition checks
- Intervals set by criticality and drift history
- Automated checks flag sensors that disagree with neighbors
- Calibration results stored against each instrument
- Replacements planned from trend and age
Prioritize calibration by consequence
| Priority | Sensor role | Suggested approach |
|---|---|---|
| High | Feeds safety interlocks or critical asset alarms | Tighter intervals, redundant checks, documented sign-off |
| Medium | Feeds predictive models on important assets | Scheduled calibration plus automated drift checks |
| Low | Informational or non-critical trending | Longer intervals, replace on failure |
Quality risks differ for each kind of sensor
A single quality rule cannot cover every measurement. Each sensor type fails in its own way, so checks should reflect how the device works.
| Sensor type | Common quality problems | Useful checks | Typical steel plant use |
|---|---|---|---|
| Vibration accelerometers | Loose mounting, damaged cable, wrong sensitivity setting, saturation | Mounting inspection, bias voltage check, comparison with a portable analyzer | Fans, gearboxes, pumps and motors |
| Temperature probes and thermocouples | Drift, poor contact, damaged sheath, wrong compensation | Reference check, comparison with nearby probes, contact inspection | Bearings, cooling circuits and motor windings |
| Current and power transducers | Wrong scaling, saturated range, noisy supply | Compare with drive readings and nameplate values | Hoist, conveyor and pump motors |
| Pressure and flow transmitters | Blocked impulse lines, zero shift, air in lines | Zero and span checks, line purge, loop test | Hydraulics, lubrication and cooling water systems |
| Oil condition sensors and samples | Contaminated sample points, irregular sampling, sensor fouling | Standard sample procedure, lab comparison | Gearboxes, hydraulic units and lubrication systems |
| Wireless nodes | Low battery, weak signal, intermittent reporting | Battery tracking, signal strength logging, gateway checks | Remote conveyors, utility areas and temporary monitoring |
Many data problems begin on the day of installation
A sensor installed carelessly will never produce good data, no matter how advanced the analytics. Use a short commissioning routine for every new or replaced sensor.
1
Confirm location and mounting
Place the sensor where it senses the fault you care about, secure it as the manufacturer specifies and photograph the final position.2
Protect cabling and connections
Use shielded cable where needed, route away from power cables and seal connectors against dust, water and heat.3
Check the signal end to end
Compare the reading at the sensor, in the controller and in the data platform to confirm units and scaling match.4
Record the baseline
Capture normal readings across running states, so later drift can be judged against real values.5
Register the instrument
Create the instrument record with tag, asset link, install date, calibration status and responsible owner.Handling missing and noisy data without hiding real faults
Cleaning data is necessary, but careless cleaning can erase the very fault signature a model should detect. Treat each correction as a decision that needs a reason.
Risky practices
- Filling long gaps with straight-line estimates across a shutdown
- Smoothing so heavily that short vibration bursts disappear
- Deleting outliers without checking if they are real events
- Mixing data from before and after a sensor swap
Safer practices
- Mark gaps clearly and exclude long ones from training
- Filter by signal type, keeping the frequencies that carry fault information
- Review outliers against operating events before removing them
- Start a new baseline period after any sensor change
Keep a record of every cleaning rule that is applied, so an engineer can explain later why a value was excluded or adjusted.
Who is responsible for sensor data quality
Data problems often persist because nobody clearly owns them. Assign responsibilities in writing.
- Instrumentation technician: installs, tests, calibrates and repairs sensors, and updates the instrument record.
- Reliability engineer: defines quality checks, reviews drift reports and decides what data is fit for analysis.
- Maintenance planner: schedules calibration and repair tasks so they fit planned windows.
- OT and IT staff: keep gateways, networks and time synchronization healthy.
- Operations: report process changes and unusual conditions that explain unexpected signal behavior.
What a useful sensor health report shows
A weekly report keeps data quality visible to the people who can act on it. Keep it short and focused on the signals that feed critical decisions.
| Report element | What it tells you | Action it supports |
|---|---|---|
| Signals with gaps above threshold | Where communication or power problems are recurring | Network, gateway and power checks |
| Signals flagged for drift | Which instruments disagree with their neighbors or expected behavior | Calibration or replacement tasks |
| Overdue calibrations | Which critical instruments are past their verification interval | Scheduling and escalation |
| Instruments with repeat faults | Locations or models that fail more than expected | Redesign, relocation or supplier review |
| Alerts caused by sensor faults | How much of the alert noise is a data problem, not an equipment problem | Targeted data quality improvements |
Quick wins for the first 30 days
- List the sensors that feed your ten most critical assets and confirm each has an owner.
- Check that every one of those tags maps to an asset record and a calibration date.
- Turn on simple flat-line and gap detection for those signals.
- Create preventive maintenance tasks for the instruments that have no calibration schedule.
- Review the last quarter of false alerts and trace how many came from sensor faults.
Turn Sensor Problems Into Tracked Work
Link every sensor to its asset, schedule calibration as preventive maintenance and create work orders when data quality checks fail.
Is this signal ready for AI predictive maintenance?
Run this checklist for each critical signal before it trains or feeds a model.
Yes / NoThe tag is mapped to one physical sensor and one asset record.
Yes / NoUnits, scaling and sample rate are documented and consistent.
Yes / NoThe last calibration or verification date is recorded and within interval.
Yes / NoAutomated checks cover range, rate of change, flat-line and gaps.
Yes / NoOperating state such as running, idle or ramping is captured with the signal.
Yes / NoSensor replacements and relocations are logged with dates.
Yes / NoPast failures on the asset have recorded modes to validate against.
Yes / NoA named person owns the sensor and responds to data quality alerts.
The sensor health workflow
1
Detect
Automated checks flag a flat-line, spike, gap or disagreement with neighboring sensors.2
Quarantine
Suspect data is marked so it does not train models or trigger asset alarms.3
Create the task
A work order is raised against the instrument with the symptom and location.4
Verify and repair
A technician checks wiring, mounting and calibration, then repairs or replaces.5
Release and record
Data is released after verification, and the findings stay in the instrument history.How Oxmaint supports this workflow
- Asset and instrument records that connect sensors to the equipment they protect
- Preventive maintenance schedules for calibration and loop checks
- Corrective work orders with mobile entry for field findings
- Inspection records and history that show when a sensor was last verified
- Reports on overdue calibration and repeat sensor failures
Measure sensor data quality like any other reliability metric
Data availabilityShare of expected samples received per critical signal.
Calibration compliancePercent of critical instruments verified within interval.
Time to repair a bad signalHours from detection to released, verified data.
False alert shareAlerts later traced to sensor faults rather than equipment faults.
Repeat sensor failuresInstruments failing more than once in a set period.
Traceable tagsPercent of tags mapped to a sensor, asset and calibration record.
Review these monthly in the reliability meeting. Improving them is usually the quickest way to improve prediction quality.
Sensor data quality questions
How do I know if a sensor is drifting?
Compare it with neighboring sensors and expected process relationships, and watch for slow offsets that persist across operating states.Should bad data be deleted before training a model?
Flag and quarantine it rather than deleting, so you keep the evidence and can explain why it was excluded.How often should critical sensors be calibrated?
Follow manufacturer guidance and your quality procedures, then adjust by drift history and consequence of failure.Can a CMMS help with sensor quality?
Yes, by scheduling calibration, tracking instrument history and creating work orders. You can book a demo to see the workflow.Where should a plant start?
Start with sensors on your most critical assets, map them to asset records and set up calibration tasks in Oxmaint.Give Your AI Models Data They Can Trust
Track every sensor, calibration and signal problem in one maintenance system, so predictions rest on verified readings.
