Predictive Maintenance has a rare-failure problem

Predictive Maintenance uses sensor data and analytics to forecast equipment failure, so that repairs happen when the entire plant chooses rather than when the singular machine decides. The idea is decades old. But no doubt, the urgency is new. According to Siemens’ True Cost of Downtime report, an idle line in a large automotive plant now costs $2.3 million an hour, which is more than $600 a second. This is twice the 2019 figure. Between 2019 and 2023, US price inflation ran at 19%, while the hourly cost of automotive downtime rose by 113%.

To our detriment, most industry discussion of the technology stops at that number and moves straight to sensors and software. The same report, however, contains a detail that further complicates the already-precarious picture.

The cost of a stoppage has outrun its frequency

Siemens found that the average large plant now loses 27 hours a month to unplanned downtime - down from 39 in 2019. In caremaking and heavy industry, hours lost have roughly halved. Better equipment and sturdier routines have clearly helped, and failures are rarer.

Each active failure, though, lands on a tighter and more interdependent line. A stalled conveyor or a faulty weld cell can starve paint and assembly within minutes, and the supplier trucks queuing outside are not willing to wait. Fewer events at a higher price per event, is a very different problem from the one maintenance teams faced a decade ago.

Fewer breakdowns, thinner evidence

But here lies the awkward aspect. Machine learning, much like human beings, learns from examples, and a model has to have seen a failure - or something close to it - before it can recognise the next one. It's pattern recognition. But a transfer press that fails twice in five years gives an algorithm almost nothing to study. So a plant that has engineered its worst failures out of existence has, in a sense, engineered away its own training set.

Practitioners routinely flag missing failure history alongside noisy and siloed data as the commonest reasons projects disappoint. Research by PwC and Mainnovation found that 60% of maintenance professionals rank access to reliable, high-quality data as the most important success factor. David Schultz, Director for the SMIIoT Division at the International Society of Automation (ISA), makes the same point from a slightly different direction, noting that AI’s promise in manufacturing depends on a strong foundation of integrated and contextualised data.

What the sensors actually listen for

The older approach services equipment by the calendar or the cycle count. It either replaces parts that still have life in them or misses a bearing wearing faster than the schedule assumed. Condition-based maintenance swaps the calendar for evidence, with sensors tracking vibration, temperature and pressure and flagging drift.

Read the article:

BMW Regensburg deploys GenAI to reinvent quality checks

At BMW’s Regensburg plant, the load carriers that move vehicles through assembly send data to a Predictive Maintenance cloud platform run by BMW Group itself. The algorithm hunts for irregularities such as fluctuations in power consumption, abnormal conveyor movements and barcodes that are not sufficiently legible, any of which can precede a malfunction. None of it is exotic. Each signal is cheap to collect. The value lies in watching all of them, all of the time, across thousands of assets.

Sharing the scar tissue

One answer to thin data is to widen the pool. Stellantis has spent two years on a generative AI maintenance project, and it treats every plant as a contributor. Arnaud Deboeuf, the carmaker’s chief manufacturing officer, explained to AMS how it works.

“Every night, we download all the problems that we have had in the plants that day, and the processes for solving the problems,” he explained. “All this data is downloaded into the system so, whenever you raise a problem, the system provides specific information on how this has been managed previously in another plant, another country with the same equipment.”

That is a knowledge base rather than a forecast, yet it speaks directly to the data problem. A fault that strikes once a decade at one site may well have struck last month at another. Deboeuf is clear that retrieval is only the start.

“This is the first step, which is already in place and deployed across 28 locations worldwide, all connecting to this predictive maintenance system. What we now want to develop is how can we have predictive maintenance system that also provides the solution.” The ambition, in other words, is to move from recalling past fixes to proposing new ones.

How Predictive Maintenance powers manufacturing uptime

Automotive manufacturers are integrating predictive maintenance systems to boost shopfloor productivity and reduce downtime. Waites' Bill Kilbey explains how digital tools transform equipment reliability across OEM and supplier operations.

Trust is the real bottleneck

Pilots, we see, tend to succeed. But production rollouts tend to stall. The cause is rarely a weak model and more often the first wave of false alarms. Sensors in hot, vibrating environments age and drift, and a model that leans on small deviations can raise an alert because the hardware has degraded rather than the machine. Technicians who chase three phantom warnings will not chase the fourth. Accuracy is only half of the answer. An alert that does not become a work order, complete with location, likely fault and the spare parts needed, is just noise with a timestamp. Integration with the maintenance management system, and a named owner for the data, tend to matter more than the choice of algorithm.

Starting where a stoppage hurts most

The sensible route is narrow before broad. Rank assets by the cost of a stop rather than by how easy they are to instrument. Press lines, body-shop robots, paintshop conveyors and end-of-line test equipment usually top that list. Validate alerts against real outcomes for a full quarter before trusting them, and keep preventive routines where failures are cheap and predictable.

Suppliers are on the same journey. Antolin uses IoT to capture and monitor process data for Predictive Maintenance. Fersa Bearings is building digital twins to predict failures, and it says the combination with human intuition and knowledge is what sharpens its decisions.

New product lines pose a harder test. Battery and e-drive equipment is young, and its failure modes are thinly documented. A 2026 review of digital twins in battery systems, published in Frontiers in Batteries and Electrochemistry, observes that robust models need long-term data spanning ageing states and failure modes, and that access to such data is limited. Hybrid twins, which pair physical models of how equipment should behave with live data, offer a way to begin before the failures arrive.

The sharper question for a plant director this quarter is not which platform to buy. It is which three machines would do the most damage by stopping on a Tuesday afternoon, and whether anyone, at any site in the network, has seen them fail before.