Count the failures before you buy the prediction

Predictive maintenance is one of the most persuasive proposals a manufacturer's board will see. The equipment is expensive, unplanned downtime is visible in every operations review and the vendor's case usually shows savings that pay back within two years. My argument is that most of these cases rest on an assumption nobody checks: that the plant has enough recorded, correctly coded failures for a model to learn from. In my experience it rarely does, and the most valuable money in a predictive maintenance programme is often spent on the maintenance records before any model is bought.

I do not argue that prediction never works. It works well for some failure modes, on some machines, in plants with good histories. I argue about sequence and about evidence.

What fifty-eight failures can teach a model

In 2023 our operations team at Tessandier Aerospace proposed a fleet-wide spindle-prediction system for 64 five-axis machining centres across three plants. A spindle failure stops one of those machines for days and costs between C$90,000 and C$160,000 to repair. The vendor's case claimed annual savings of C$6.1 million, the programme had been mentioned in the operating plan presented to our board, and a discount was available for signing quickly.

Before signing, my team ran the model in shadow mode on 22 machines for six months. It produced 47 alerts. Eleven corresponded to faults that developed within thirty days, and three failures happened without any alert. That is not a useless model, but it is not the model in the business case either.

The reason became clear when we counted. Across six years and 64 machines we had 58 recorded spindle failures, and 27 of the work orders had no usable cause. Some said only replaced spindle; others mixed a crash caused by a programming error with a bearing that wore out. A model trained on that history was learning, in part, how our planners wrote work orders. I had seen the same thing years earlier at a mining-equipment maker, where our first warranty models predicted the writing habits of dealers better than they predicted failing hydraulic pumps.

Fifty-eight failures across a fleet is a normal number for well-maintained equipment. It is also a small sample on which to estimate a time to failure, especially when the failures are spread across machine models, spindle types and duty cycles. A vendor's model trained on other customers' machines can help, but only if those machines resemble ours, and the vendor's sales team is not the right party to say whether they do.

Condition monitoring is not a consolation prize

We changed the programme in three ways. First, we rebuilt failure coding with the maintenance planners in all three plants, so that a work order records the failure mode, the evidence and whether the cause was wear, a crash or a lubrication fault. Second, we installed condition monitoring on all 64 machines: vibration and temperature measurements with fixed limits, reviewed by planners when a limit is crossed. Third, we kept the prediction model on the 14 Longueuil machines with the cleanest histories, and retrained it as the recoded records accumulated.

Condition monitoring is often presented as the modest option a company settles for when it cannot afford prediction. I think that is backwards for most machining plants. A fixed limit on bearing vibration is explainable to a planner in one sentence and auditable after the event. It does not estimate time to failure, but it tells a planner that a machine needs attention, which is what most of the saving depends on. In 2025 unplanned spindle downtime across the fleet was 27% lower than in 2022, and the annual saving was about C$2.2 million. Most of that came from monitoring and better records, not from prediction.

The retrained model on the Longueuil machines improved its precision from 23% to 52%. I think that is a respectable result, and it would not have been possible without the work on records. It is also the result of a smaller ambition than the one in the original case. Our chief operating officer argued at the time that my standard of evidence would stop every improvement in the plants, and I take that criticism seriously. The answer was to approve the monitoring quickly and make prediction wait for evidence, not to stop both.

There is a cost I underestimated. Monitoring produces alerts, and alerts land on maintenance planners who already have full days. At one point Longueuil's planners were receiving forty condition alerts a week and had begun closing them unread. A system that floods the people who must act on it is worse than no system, because it teaches them to ignore the signal. We now cap alerts per planner and tune limits machine by machine before switching them on.

The question for the investment committee

I would ask any board reviewing a predictive maintenance case to request four facts before approving it. How many failures of the relevant mode has the plant recorded, on which machines, over what period? What proportion of those records give a usable cause? How did the model perform in a shadow trial on the company's own equipment, including the failures it missed? What would condition monitoring alone achieve, and at what cost?

None of these questions requires technical expertise from directors. They require management to separate the saving that comes from noticing problems early, which monitoring delivers, from the saving that comes from estimating when a failure will occur, which prediction promises. In my experience the first is usually larger than the second, and much more certain.

Tessandier's investment policy now asks the failure-count question before any predictive programme is approved. It has slowed two proposals and stopped one. The one we stopped concerned furnace heating elements, where we had recorded nine failures in eight years. A model would have learnt nothing from nine failures. A replacement schedule based on hours in service, which the furnace maker had recommended all along, was cheaper and easier to defend to a customer's auditor.