Skip to content
Park Hyoin PARKHYO.IN
Go back

I thought it was predictive maintenance, but it was condition-based maintenance

Edit page

While working on time-series analysis of equipment energy efficiency, I ended up building an anomaly detection model. I don’t work at the level of looking at individual parts, but since this seemed like something widely used in manufacturing AX, I decided to look into it.

It started with the keyword “predictive maintenance” and trying to check whether I could apply it to what I was currently working on. But after looking into it, what I was trying to do wasn’t predictive maintenance.

Table of contents

Open Table of contents

Maintenance is divided into four types

TypeWhat it doesProblem
Breakdown Maintenance (BM)Fix it when it breaksLow cost, but sudden line stoppages cause losses
Preventive Maintenance (PM)Inspect and replace parts on a fixed scheduleReduces failures, but wastes resources replacing parts that are still fine
Condition-Based Maintenance (CBM)Monitor condition via sensors and service when values exceed a threshold—
Predictive Maintenance (PdM)Learn patterns in condition data to predict when failure will occur and optimize maintenance timing—

It was only after reading this far that I understood. I thought I was trying to introduce predictive maintenance into equipment anomaly detection, but what I was actually thinking of was condition-based maintenance.

Alerting when a value crosses a threshold and predicting in advance when it will cross that threshold are different things. I realized that to do AX properly, I needed to understand predictive maintenance more precisely and in more detail.

Predictive maintenance proceeds in five stages

5 stages of predictive maintenance — from data collection to linking with maintenance decisions
  1. Data collection — Gather sensor data, equipment operation history, and maintenance history
  2. Anomaly detection — Detect patterns that differ from the normal state
  3. Fault diagnosis — Determine which component has what type of problem. In terms of the work I do, this could mean a situation where fuel consumption at a specific piece of equipment becomes higher than usual
  4. Remaining useful life (RUL) prediction — Estimate the time remaining until failure
  5. Linking to maintenance decisions — Connect the prediction results to maintenance plans, the maintenance management system, MES, etc., so they translate into actual work orders

How far is this actually used in practice?

Apparently, many projects referred to as “predictive maintenance” in the field actually only implement up to stage 2, anomaly detection.

Remaining useful life prediction requires enough accumulated failure history data, and the better-maintained the equipment, the less failure data there is, which makes this difficult. So it’s common to start at the pilot stage with anomaly detection and early warnings, then expand into diagnosis and lifespan prediction as data accumulates.

Accuracy is a metric you can’t use

Since failures occur rarely, accuracy is often not a suitable performance metric. Even a model that always predicts “normal” can achieve 99% accuracy.

In predictive maintenance, other things matter more.

  • How few failures were missed (recall)
  • How frequent false alarms are
  • How many days in advance of failure the alarm was raised (lead time)

In my case, I think it should be possible to detect anomalies by training on sensor data, operation records, and maintenance history.

Things to study further

  • Remaining useful life (RUL) prediction models — How to approach this when failure data is scarce. NASA’s C-MAPSS turbofan dataset is frequently used as a benchmark in this field
  • Comparing anomaly detection techniques — Whether statistical methods (control charts, thresholds), machine learning (Isolation Forest, One-Class SVM), or deep learning (autoencoders, LSTM) is right for time-series equipment data
  • How to evaluate lead time — To turn “how many days before failure the alarm was raised” into a metric, I need to figure out how to define the ground-truth point in time. I need to check whether there’s a standardized definition like there is for recall and precision
  • Handling imbalanced data — In data where failures make up less than 1%, raising recall also raises false alarms. I need a method for deciding where to cut off that trade-off
  • ISO 13374 · ISO 17359 — International standards for condition monitoring and diagnostics. I’d like to compare what terms the standards use for the 5 stages outlined above

Edit page