Predictive maintenance projects fail for a dull reason: the data was never asked whether it could support the promise. Here is the check we run in the first week of every assessment. Answer it honestly for your own plant.
The ten questions
- Do you keep a breakdown register, and does it record the cause as well as the date?
- Does the register cover at least two years?
- Is each breakdown tied to a machine identifier that is used consistently everywhere else?
- Do you know what each machine produced, per shift, over the same period?
- Do your PLCs or machines log anything at all, even hourly?
- If they do, has anyone ever opened those logs?
- Do you record spare-part issues against a machine?
- Can you say which three machines cost you the most in stoppages last year, with a number?
- Does the maintenance head want this project, or is it only the owner who wants it?
- Would you accept a rule-based alert system for the first six months if the data cannot support a model yet?
What the answers mean
Eight or more yes answers: the data can probably support a model for your top failure modes. A pilot on one line is realistic. Five to seven: start with a rule-based system and fix the gaps while it runs; a model follows in six to nine months. Fewer than five: a data foundations engagement first, and do not let anyone sell you sensors until the register problem is solved.
The question people get wrong
The ninth one. We have never seen a predictive maintenance project succeed where the maintenance head was a bystander. The system tells the team where to look first; if the team does not trust it, it is an expensive dashboard.