Two different questions
Predictive maintenance usually asks one of two things. The first is classification: how likely is this machine to fail within a set window, such as the next 30 days? The second is regression: how much useful life does it have left (remaining useful life, or RUL)? The first suits ranking a fleet. The second suits planning one asset's replacement. Many projects need both.
Calibration: when 20 percent has to mean 20 percent
A ranking only needs the order to be right. A plan that weighs risk against cost and crew time needs the probabilities themselves to be believable. Calibration checks whether, among machines given a 20 percent risk, roughly one in five really fails. If not, the scores can be recalibrated before they are used for decisions.
Explaining why a machine is flagged
Technicians are rightly sceptical of a score with no reason attached. SHAP explanations show which readings pushed a machine's risk up, for example a sensor drifting away from its normal range. That gives the crew something to inspect and gives the analyst a way to spot a model relying on the wrong signal.
The step most projects skip: capacity
A ranked list of the 400 riskiest machines is useless if the crew can do 40 jobs a week. The practical output is a plan built under constraints: how many jobs the team can complete, what each costs, and what a failure costs. Choosing which jobs to do then becomes a selection problem, maximising the risk removed within the available capacity. The result is a short weekly list that fits the team, with the reasoning for each item.
Validating remaining useful life on a public benchmark
When real failure histories are scarce or confidential, a public benchmark gives an independent check. NASA's C-MAPSS dataset contains simulated run-to-failure data for turbofan engines and is widely used to test RUL methods. Results on it say the method works on that kind of data. They do not guarantee the same accuracy on your equipment, so any real project still needs validation against your own history.
What to prepare before starting
- A record of past failures and when they happened, even if rough.
- Sensor or inspection readings with timestamps.
- The cost of a failure compared with the cost of a planned repair.
- How many jobs your crew can realistically complete each week.
See it in a working project
I built a public case study around this topic, with code and documentation on GitHub: industrial-ai-predictive-maintenance. It runs on synthetic data generated for the project, so it demonstrates the method and the engineering rather than results from a real client.
Further reading
- scikit-learn: probability calibration: Why risk scores need to be calibrated before planning around them.
- SHAP documentation: Explaining why a machine is flagged.
- NASA: simulating degradation data for prognostic algorithm development: Background on the C-MAPSS simulated turbofan data used for remaining-useful-life benchmarks.