Equipment maintenance broadly follows three models: fix it when it breaks (reactive), strip it down on a calendar schedule (preventive), and judge by actual condition (predictive). The first two either impose heavy downtime costs or waste spare parts and labor. Predictive maintenance aims to intervene after a fault signature appears but before function is lost.
Many projects, however, open with discussions about algorithms and platforms, then stall on the fundamentals: what to measure, where to mount it, and what counts as abnormal. This article follows the actual implementation order.
Step 1: Not Every Asset Deserves Monitoring
Start by classifying equipment so that resources go to what is critical. A practical set of classification dimensions:
- Impact: does a stoppage halt the entire line (main drives, air compressors, critical pumps)?
- Failure frequency and consequence: has it failed often historically, and does failure carry safety or quality risk?
- Spare parts and repair lead time: are spares on long procurement cycles, and does repair require the vendor on site?
Proven practice: use ABC classification to screen out the top 10 to 20 percent of assets as critical and monitor those by condition; keep the rest on scheduled maintenance. Coverage is not the first objective — driving down unplanned downtime on critical assets is.
Step 2: Choose the Right Variables
Different failure modes map to different physical quantities. Measure the wrong one and no amount of data will help:
- Vibration: bearing wear, misalignment, imbalance, gear damage, looseness. The first choice for rotating equipment and the most widely applied measure in predictive maintenance.
- Temperature: overheating bearings and windings, poor lubrication, cooling failure, loose electrical connections. Low cost, suitable for broad deployment.
- Current/power: abnormal motor load, stalls, efficiency loss. Readable from existing meters or drives, so no extra hardware is needed.
- Pressure/flow: performance degradation in pumps and fans, pipeline blockage, leakage.
- Oil and wear particles: deep monitoring for large gearboxes and hydraulic systems. Higher cost, generally reserved for the most critical assets.
Step 3: How to Mount and Select Sensors
- Measurement location: vibration sensors belong on a rigid surface close to the bearing, never on a guard or thin panel. Temperature probes belong near the heat source with thermal protection.
- Mounting method: mount vibration sensors rigidly, by stud or magnetic base. Adhesive mounting attenuates high-frequency content and degrades early fault detection.
- Range and frequency response: select the vibration frequency range from machine speed; slow-speed equipment needs good low-frequency performance. Leave margin on temperature relative to actual conditions.
- Environmental fit: in high-temperature, humid, dusty or oily-mist environments, pay attention to ingress protection and cable protection. Hazardous areas require intrinsically safe products.
- Signal type: prefer sensors with standard signal output or a digital interface, so they connect directly to gateways and industrial computers without adding converters later.
Step 4: How to Process the Data
Raw collected values do not indicate a fault on their own. Three processing stages sit in between:
- 1Cleansing: remove spikes and outliers caused by dropouts, and unify sampling rate and units.
- 2Feature extraction: RMS, peak, kurtosis and spectral features are common for vibration; mean, trend slope and fluctuation band for temperature and current.
- 3Baselining: build a per-asset baseline from periods of known normal operation, then compare current values against it.
Important reminder: do not take the generic limits from a vendor manual and use them directly as alarm lines. Identical models vary widely under different loads and mounting conditions. The correct approach is to collect two to four weeks of normal data first, establish an individual baseline per asset, and only then set thresholds.
Step 5: Thresholds and Alarm Strategy
Thresholds are usually set at two or three levels to avoid alarm storms that make maintenance teams numb to them:
- Advisory: deviation from baseline by a set margin, recorded in the trend but not pushed. Used for early observation.
- Warning: sustained deviation meeting defined conditions, pushed to the asset owner to schedule a planned inspection.
- Alarm: safety or functional limit exceeded, pushed immediately and handled per the response plan.
Two points carry the most weight in an alarm strategy: persistence logic (a single spike does not alarm; several consecutive out-of-range points do) and consolidation (related alarms from one asset merge into a single work order). Get these two right and alarm credibility improves noticeably.
Step 6: Close the Loop So Data Actually Drives Action
The most common failure mode in predictive maintenance is building dashboards without a closed loop. A complete loop includes:
- Alarms that generate work orders automatically, with a named owner and a deadline.
- Outcome feedback (confirmed anomaly, false alarm, resolved) used to continuously refine thresholds.
- A monthly review of alarm accuracy and missed detections, feeding adjustments back into baselines and rules.
- Effective criteria documented into that asset's maintenance standard and folded into routine inspection rounds.
Suggested Three-Stage Rollout
- Stage one (foundation): instrument critical rotating assets with temperature and vibration, establish trend curves and two-level alarms, and solve the "do we have data at all" question.
- Stage two (analysis): bring in process variables such as current and pressure, run multi-parameter correlation analysis, and form a health assessment per asset.
- Stage three (prediction): once enough history and failure samples have accumulated, introduce models for remaining useful life estimation and fault type classification.
Frequently Asked Questions
We have no historical failure data. Can predictive maintenance still work? Yes, but the starting point is condition monitoring rather than prediction. Establish a normal baseline first and judge anomalies from relative change; that alone covers most progressive faults such as bearing wear, scaling and blockage. Failure samples accumulate with use, and classification or life-prediction models come later, once the sample set is large enough.
Which has higher priority, vibration or temperature monitoring? For mechanical faults on rotating equipment, vibration comes first, since it is more sensitive to early bearing and misalignment problems. Temperature suits a broad, low-cost baseline layer and works well for poor lubrication, cooling failure and electrical heating. With a limited budget, a common combination is vibration on critical assets and temperature on the rest.
How should the return on a predictive maintenance investment be measured? Track three indicators: count and duration of unplanned stoppages on critical assets, spare parts inventory turnover, and the share of labor hours spent on emergency repair. Record baseline values before the project starts and compare them after six months of operation rather than counting how many sensors were installed.