Condition-Based Maintenance — Sensor Selection, Data & Pilot Checklist

Welcome — the practical hunger

If you want to move from calendar-based maintenance to actions driven by real machine condition, the right sensors and trustworthy data are the foundation. This guide helps you choose practical sensors, plan sampling and signal conditioning, capture essential metadata, validate data quality, and run a short pilot that gives your analysts something they can actually use.

Quick decision flow

1. Clarify the failure modes you want to detect (bearing wear, imbalance, overheating, insulation failure, lubrication contamination, electrical anomalies).
2. Map each failure mode to sensor types and minimal measurement needs (what to measure, required sensitivity, plausibility of detection).
3. Choose an initial set of conservative sensors and sampling settings for a pilot (you can trim later).
4. Define essential metadata and data validation rules before you collect data.
5. Run a short pilot, validate, iterate.

Sensor types and practical use-cases

• Vibration (accelerometers) — Best for rotating machinery faults (bearings, imbalance, misalignment). Use where early mechanical degradation matters.
• Temperature — Good for bearing/junction/gearbox heating trends and safety thresholds. Often lower sample-rate than vibration.
• Current/Power — Detects electrical issues, motor load changes, stalling, and some mechanical load signatures via motor current signature analysis.
• Ultrasound — Useful for detecting early air/steam leaks, electrical discharge, or early bearing issues in noisy environments.
• Oil/Fluid Analysis — Periodic lab or inline sensors for contamination, wear particles, and lubricant condition.

Typical measurement considerations (practical guidance, not absolute rules)

Sampling choices are highly use-case dependent. Use these as starting points and validate in the pilot.

• Vibration accelerometers: For bearing diagnostics and FFT-based analysis, sample rates commonly range from 2 kHz to 10 kHz (or higher for high-speed spindles). For trend-based RMS or overall vibration monitoring, lower rates (e.g., 1 kHz) may be acceptable.
• Temperature: Sample intervals from seconds to minutes. For slow-developing heating, once per minute or once every 5–15 minutes may suffice; for short transient detection, sample faster.
• Current/power: For FFT or motor signature analysis, sample at least 1–2 kHz. For load trending, lower sample rates can work (e.g., 1 Hz or slower depending on dynamics).
• Ultrasound: Capture high-frequency bursts; typical ultrasonic detectors operate in tens to hundreds of kHz bands—use vendor guidance for expected burst durations and repeatability.
• Oil analysis: Periodic sampling frequency is program-specific (weekly to monthly) or event-driven based on runtime or filtered particle counters.

Mounting and installation best practices

• Prefer hard stud mounting for accelerometers where possible — it provides the best signal fidelity for bearing and housing vibration. If not possible, use recommended adhesives or pre-qualified magnetic mounts for short-term tests.
• Document mounting orientation and location on the machine (e.g., drive-end bearing housing, 3 o'clock position). Orientation affects axis naming in analytics.
• Isolate sensors from extraneous structural resonance when you are measuring specific machine elements—place sensors as close as practical to the component of interest.
• Use proper cable routing and strain relief. Keep cables away from power lines to reduce electrical noise. Seal and protect connectors in harsh environments.
• Record sensor calibration date, model, and serial number at installation.

Signal conditioning & pre-filtering

• Anti-aliasing filters are essential. Ensure the acquisition hardware applies a suitable low-pass filter before down-sampling.
• For vibration FFTs, use bandwidths sufficient to capture expected defect frequencies. If unsure, sample at a higher rate during the pilot and down-sample in processing.
• Consider on-sensor preprocessing (RMS, crest factor, envelope) for low-bandwidth telemetry links, but always retain a periodic raw-sample option where possible for deeper analysis.

Essential metadata to capture (make these required fields)

Metadata allows data to be interpreted and models to remain valid over time. Store these as structured fields, not free-text where possible.

• timestamp (ISO 8601, UTC)
• machine_id (unique asset identifier)
• sensor_id (unique sensor serial or tag)
• sensor_type (accelerometer, thermocouple, current transducer, ultrasonic, oil)
• channel or axis (e.g., X/Y/Z)
• sampling_rate_hz
• units (g, m/s², °C, A, Pa, etc.)
• mounting_location (bearing DE, NDE, gearbox input shaft, etc.)
• mounting_method (stud, adhesive, magnetic)
• machine_operating_state (running, idling, startup, shutdown, speed_rpm)
• load or process context (production batch id, throughput, torque, where relevant)
• calibration_date and firmware_version

Example JSON metadata snippet

{
  "timestamp": "2026-08-01T13:45:30Z",
  "machine_id": "MX-PLC-001",
  "sensor_id": "ACC-00123",
  "sensor_type": "accelerometer",
  "channel": "X",
  "sampling_rate_hz": 5000,
  "units": "m/s^2",
  "mounting_location": "drive_end_bearing",
  "mounting_method": "stud",
  "machine_operating_state": "running",
  "speed_rpm": 3600,
  "calibration_date": "2025-12-01"
}

Data quality validation checks (automated where possible)

• Timestamp sanity: No large jumps, duplicates, or inconsistent timezones; monotonic checks for streamed sequences.
• Sampling consistency: Expected sampling_rate_hz matches actual samples per second; flag variable-rate gaps.
• Signal clipping or saturation: Detect repeated max/min values or flat-lines indicating clipping or disconnected sensors.
• Noise floor & SNR: Compute baseline noise and flag abnormally high noise or very low variance (stuck sensor).
• Missing fields: Required metadata present (machine_id, sensor_id, sampling_rate_hz, units, mounting_location).
• Cross-sensor correlation: Check that nearby sensors show consistent events for true machine events versus isolated sensor faults.
• Heartbeat telemetry: A small, frequent status message from each sensor helps detect disconnections early.

Pilot design checklist (make this your minimum viable pilot)

1. Select a small set of representative machines (3–10) with known failure modes or historical issues.
2. Install conservative sensors (e.g., accelerometer + temperature on rotating machines; current + temperature on motors).
3. Collect continuous raw or burst vibration for 30–90 days while also recording operating context (speed, load, production state).
4. Define success criteria up front (e.g., detect bearing fault confirmed by maintenance within 30 days, or achieve >X% data completeness).
5. Validate data quality after first week; adjust sampling, mounting, or channels as needed.
6. Document data governance: owner, retention policy, access rules, and how model outputs feed work orders or alerts.

Common pitfalls and how to avoid them

• Pitfall: Missing or inconsistent metadata. Fix: Enforce metadata schema at ingestion and provide templates for installers.
• Pitfall: Choosing sensors only on cost rather than data utility. Fix: Prioritize sensors that reliably detect the targeted failure modes in your environment.
• Pitfall: Insufficient sampling for intended analytics. Fix: During pilot, capture higher-fidelity data where possible; reduce later if unnecessary.
• Pitfall: Data sitting unused due to unclear ownership. Fix: Assign analytics owners and define how alerts translate to workflow actions.

Next practical steps

Start with a short, well-scoped pilot. Use the metadata schema above, run automated data-quality checks within the first week, and be prepared to adjust mounting and sampling early. Treat the pilot as a learning loop: capture what you need to detect the failure modes, then simplify where possible.

Why this matters

Good sensor choices and disciplined data practices reduce false alarms, improve model trust, and make condition-driven maintenance feasible and repeatable across sites. Poor choices and missing metadata are the most common reasons pilot models fail to generalize.

Further resources

• Reference vendor mounting guides for specific accelerometer models.
• Standards and best practices from vibration analysis bodies (for deeper technical detail).
• Consider adding a short interactive sensor-install checklist or an ingestion validation tool (see capability notes).