Predictive maintenance for OEMs: turning service into a product

Because the money and the loyalty in industrial equipment sit in the years after the sale, and an unplanned stop is the moment a customer starts shopping for your competitor. McKinsey's analysis across 30 industries found an average margin of 25 percent on aftermarket services against 10 percent on new equipment, measured as earnings before interest and taxes. Service is already the better business; prediction makes it a scheduled one.

Your customers feel the other side of that ledger. Siemens estimates in its True Cost of Downtime 2024 report that the world's 500 largest companies lose almost $1.4 trillion a year to unplanned downtime, equal to 11 percent of their revenues, and that an average large plant still loses 27 hours a month. A plant manager does not need convincing that a stopped line is expensive. They need someone who knows the machine better than they do to warn them early.

That someone should be you. You hold the drawings, the failure history from warranty claims and the field service notes for every unit you have shipped. A plant sees one or two of your machines; you can see the whole fleet, which is the advantage a predictive maintenance system turns into a product.

Monitor the handful of failure modes that cost customers the most and leave a measurable trace, not every tag the controller exposes. Pull a list of warranty claims and emergency service calls from the last few years, rank the parts by repair cost multiplied by frequency, and keep the top five or six. That list, not a sensor catalog, decides the hardware bill.

On the rotary filler fleet we work on, heat, pressure and temperature sensors cover most of the risk, and a camera watches the parts where visual wear says more than any probe. The table below shows how failure modes on a typical packaging machine map to signals.

Because nobody can act on a machine health score, but a service planner can act on a named part with a window. A score that drops from green to amber tells the plant something is wrong; a message that the turret drive bearing on a given serial number is trending toward failure within a few weeks tells your service desk which part to ship and which technician to send.

Part-level models also fit how machine builders already think. Your spare parts catalog, your bill of materials and your service pricing are all organized by part, so predictions organized the same way flow straight into a quote, a parts order or a planned visit. Our packaging machine build predicts wear on specific components for exactly this reason.

The fleet is what makes this possible for an OEM and nearly impossible for a single plant. A failure that a customer sees once a decade happens somewhere in a fleet of hundreds of identical machines every few months. Pooling those histories gives each part model enough failures to learn from, while per-machine baselines absorb the differences between customers' products, shifts and sanitation routines.

You break things on purpose. Before any prediction reaches a customer or a contract, introduce known faults on a test machine or a cooperative customer line, and check that the model flags each one early enough to act on. On our filler program, every part model is tested against faults introduced on purpose before anyone relies on it, and 12 machines run on the platform today.

Induced fault testing answers the three questions a service director will ask: does it catch the fault, how much warning does it give, and how often does it cry wolf. The last one decides adoption. A service desk that receives false alarms for a month will mute the channel, and no model accuracy figure will win it back.

• Record a clean baseline across the products and speeds the machine really runs.
• Introduce one fault at a time: a worn seal, a partially blocked line, a loosened mount, a degraded heater.
• Measure detection, lead time and false alarms per part, not one blended score for the machine.
• Rehearse the full path from alert to work order to parts shipment before the first customer goes live.

Sell commitments, not data. Customers do not want dashboards of sensor traces; they want fewer stops and a named person who calls them first. Most builders end up with tiers like the ones below, starting with the simplest tier a plant manager can approve without a capital request. A custom CRM (Customer Relationship Management) system for machine builders is usually where those contracts, renewal dates and service visits get tracked.

Settle data rights at the point of sale, in the purchase agreement, because retrofitting consent onto an installed machine is slow and awkward. The clause should say what the builder may collect, how it may use pooled fleet data, how the connection is secured and what happens to the data when the service contract ends.

Do not plan on data exclusivity as your moat if you sell into Europe. The European Union's Data Act has applied since 12 September 2025 and gives users of connected machines the right to access the data their use generates and to share it with third parties, including independent repair and maintenance providers. Your advantage has to come from better models and faster service, which is where it belongs anyway.

• Ownership: who owns production data, and what rights the builder holds over machine condition data for service and product improvement.
• Pooling: whether anonymized fleet data may feed models used for every customer.
• Connection: how the machine connects, which way data flows and who on the customer's security team approves it.
• Location: data stays in the region agreed with the customer.
• Exit: what the customer receives, and what is deleted, if monitoring stops.