Predictive maintenance: equipment selection criteria and data audit

Prioritizing equipment: 4 selection criteria

Starting by instrumenting every piece of equipment on a line is a common mistake. The ROI of predictive maintenance is concentrated on a few critical assets: identifying which ones is the first decision to make.

The 4 selection criteria in order of importance:

• Criticality of downtime: does a stoppage on this equipment stop the line? does it stop the site? The hourly cost of downtime must be estimated. It is the numerator of the ROI.
• Failure frequency: equipment that fails once every 5 years is not a priority candidate, even if it is critical. Target equipment with 2+ unplanned failures per year.
• Detectability by sensors: is the failure preceded by degradation measurable via vibration, current, temperature or pressure? Sudden electrical failures or overload fractures are not predictively detectable: do not instrument for that type of failure.
• Available historical data: are the equipment's sensor data archived in a historian? Over what period? At least 12 to 18 months of history with documented failure events is the minimum to train a supervised model.

Auditing historical data: what we look for

Time series quality

Before training anything, assess the quality of sensor data on candidate equipment:

• Gap rate: frequent interruptions in the series (maintenance stoppages, network outages, sensor changes) create artifacts in feature engineering. Quantify the percentage of missing data per period.
• Granularity: for vibrations (bearing detection, cavitation), you need at least 1 kHz sampling. A measurement every 10 seconds does not allow spectral analysis. Check what your historian actually stores.
• Sensor drift: a sensor not calibrated for 3 years whose baseline has drifted by 20% trains a model that learns the drift, not the failure. Identify sensors with a suspect history.

Labeling past failures

This is the most time-consuming and most underestimated work. The quality of your failure history directly determines the quality of the supervised model.

What we look for in the CMMS work orders:

• Date and time of the failure (not just the intervention)
• Type of failure (failure mode, not just the replaced part)
• Equipment and sub-assembly involved
• Distinction between sudden failure vs. progressive degradation

In most industrial fleets, 40 to 60% of work orders are insufficiently documented to serve as labels. The work-order requalification session with the maintenance manager is an unavoidable step: allow 2 to 4 days for a 2-year history.

Anomaly detection vs. RUL estimation: the right tool for the right problem

When to use anomaly detection

Anomaly detection (Isolation Forest, autoencoders, LSTM-based) is suited when:

• You have no (or few) labeled failure data
• You are starting the project and want an early signal quickly
• Failure modes are multiple and not fully known
• The degradation duration before failure is variable

Its limitations: it detects that something is abnormal, not what will fail or when. The false positive rate is higher than a supervised model: expect alerts to investigate rather than directly actionable predictions.

When to use RUL estimation

Remaining useful life (RUL) estimation is relevant when:

• You have a documented failure history on the targeted equipment type (minimum 8 to 12 events per failure mode)
• The degradation is progressive and measurable (bearings, wear, cavitation: not overload fractures)
• You need a prediction horizon to plan maintenance interventions

LSTMs and survival models (Weibull, Cox) give good results on rotating equipment with enough data. But they require spectral feature engineering (FFT, envelope) that generic models do not perform automatically.

A common pitfall: instrumenting without prior history

Adding new sensors on equipment with no failure history is the least effective scenario in the short term. You start with zero labels: you will have to wait for several failures on the new device to have enough data for a supervised model.

The more effective strategy: start with equipment already covered by the historian, with a partially documented failure history. Anomaly detection can start quickly. The supervised model will follow once failures accumulate.

New sensors on uncovered equipment are relevant, but plan 12 to 18 months before you can derive a high-performing supervised model from them.