A main spindle fails on Thursday afternoon. The replacement bearing isn’t in stock, the service technician won’t arrive until Monday, and the production line is down for four days. Scenarios like this are why predictive maintenance has been on every digital transformation agenda for years. The promise: The machine will alert you before it breaks down, and maintenance will take place exactly when it’s needed—not sooner, not later.
Technically, this promise can be fulfilled. But the market rarely explains what it requires. Predictive maintenance isn’t a sensor you screw on, nor is it software you install. It’s a data project. Anyone without structured, historical machine and process data is buying a predictive model without a basis for prediction.
During plant visits, the same pattern emerges time and again: The sensors are in place, the dashboards are colorful, but the failure history from recent years exists only in paper maintenance logs or in Excel spreadsheets. However, a model designed to predict failures has never learned what a failure looks like in the data. The project then fails not because of the AI, but because of the data foundation.
This article explains what predictive maintenance achieves, how the technology works, what data requirements must be met, what a realistic first step looks like, and where the economic and regulatory limits lie.
KEY POINTS AT A GLANCE
|
IN SHORT
|
The term is widely used in the market, often to describe anything that has sensors and dashboards. This leads to false expectations. A clear distinction between maintenance strategies lays the foundation for any investment decision.
Predictive maintenance forecasts the time of failure or the remaining service life of a component based on continuously collected condition data and historical failure patterns. Maintenance is performed exactly when the condition of the component requires it. This approach thus differs from both reactive maintenance (repairing after a failure) and preventive maintenance (fixed intervals) and goes beyond mere condition monitoring (observing the condition but not predicting it).
| Strategy | Maintenance Triggers | Typical Consequence | Data Requirements |
|---|---|---|---|
| Reactive Maintenance | Failure has occurred | Unplanned downtime, express shipping of replacement parts, consequential damage to adjacent components | None |
| Preventive Maintenance | Fixed interval (time or number of units) | Predictable but expensive: As a rule of thumb, components with 30 to 50% remaining service life are replaced | Low (operating time counter) |
| Condition Monitoring | Threshold value of a condition signal exceeded | Early warning, but without a prognosis. Response time is often short | Real-time sensor data |
| Predictive Maintenance | Predicted failure time or remaining service life | Maintenance performed within the scheduled window, replacement parts ordered in a timely manner, downtime minimized | Sensor data plus historical failure and maintenance events |
The key to cost savings lies in downtime costs. In automotive assembly, one hour of unplanned line downtime typically costs five-figure amounts, and significantly more at some plants. In mechanical engineering with interlinked machining centers, downtime, rework, and schedule delays add up just as quickly. Predictive maintenance addresses precisely these costs: it’s not that maintenance itself becomes cheaper, but rather that unplanned failures occur less frequently.
Every predictive model follows the same process: capture signals, extract features, compare patterns with historical failures, and derive a forecast. Understanding this process allows you to realistically evaluate vendors’ claims.
The signal sources depend on the wear mechanism. Vibration sensors on bearings and spindles, temperature sensors on drives, current consumption on motors, pressure profiles in hydraulic systems, and torque curves on screwdriving systems and test benches. For bearing diagnostics, sampling rates in the kilohertz range are common because damage frequencies are only visible in high-resolution vibration spectra. Many modern control systems and tools already provide this data via OPC UA or manufacturer-specific telegrams, without the need for additional sensors.
Characteristic values are calculated from the raw signals: root mean square values, spectral components, waveforms, and trend slopes. Anomaly detection flags deviations from the learned normal state. This works even without a failure history and is therefore often the first productive step; strictly speaking, however, it is still condition monitoring, not prediction.
The actual prediction is made when the model has learned from historical degradation trends how a signal pattern develops from the first sign of trouble through to failure. The target variable is called “Remaining Useful Life.” For rolling bearings, the achievable lead time for a prediction is typically 2 to 6 weeks before failure; for tool wear in machining, it tends to be a matter of hours to days. The lead time determines how much planning flexibility maintenance actually gains.
Important to note: A tool that is wearing out often produces inferior parts before it fails. Those who analyze machine condition and component quality using the same process data benefit twice over.
First, the uncomfortable truth: The choice of model is rarely the problem. As a rule of thumb, 70 to 80% of the project effort goes toward data collection, data cleaning, and data integration—not the model itself.
Four patterns recur time and again in failed or stalled projects. First: There is no documented failure history. Maintenance events are recorded in paper logs or in maintenance software without any connection to the machine data. However, a model can only learn what a failure looks like if failures are flagged in the data. Second: The timestamps of different systems are not synchronized. Sensor data, control messages, and maintenance orders can no longer be unambiguously matched retroactively. Third: Data is stored in silos created by individual equipment manufacturers and is accessible only through proprietary software. Fourth: Historical data has been deleted for performance reasons; ironically, the very time periods containing the most instructive failure events are missing.
All four scenarios share the same root cause: the absence of a comprehensive process data management system that captures machine and process data independently of the manufacturer, archives it with a clear temporal and plant reference, and ensures its long-term availability. This infrastructure is not an AI task, but a data task, and it pays off regardless of the predictive model: for traceability according to IATF 16949 Section 8.5.2, for data-driven decisions according to ISO 9001:2015 Section 9.1, and for any subsequent analysis.
PRACTICAL TIP: MANUFACTURING OSThe data foundation for predictive maintenance is established in process data management, not in an AI project. CSP’s Manufacturing OS lays this foundation:
At Mercedes-Benz, IPM provides the end-to-end database for axle assembly, which enables early identification of process weaknesses before they lead to breakdowns or quality issues. BMW and Knorr-Bremse rely on the same database for comparable applications. → The white paper “Production Data” explains how this database is structured. |
WHEN PREDICTIVE MAINTENANCE WORKS
|
The most common strategic mistake is launching on a large scale. Anyone who tries to connect 40 systems at once will get bogged down in interfaces before the first benefits become apparent. The most reliable approach is to start with a focused pilot project.
A suitable machine is one where downtime has a significant impact (a bottleneck machine, costly downtime), has known wear patterns, and is equipped with existing or easily retrofittable sensors. A machining center with a documented spindle history makes a better pilot than a machine that never fails: without failure events, there’s nothing to learn.
Before a model is trained, three data streams must be consolidated: the system’s status signals, process data from completed jobs, and the maintenance history. All of this must include synchronized timestamps and unique equipment identifiers. Where historical data resides in legacy systems, it must now be backed up and made accessible—not deleted. In practice, this step takes the longest and determines everything that follows.
A realistic pilot begins with anomaly detection under normal operating conditions: This provides initial early warnings after just a few weeks and builds trust. Predictions of remaining service life follow as soon as sufficient documented degradation trends are available. Realistic timeframe for the entire pilot: 3 to 6 months, with 1 to 2 facilities.
A forecast without organizational action is worthless. Maintenance requires defined response procedures: Who evaluates the report, who orders the replacement part, and during which maintenance window is the replacement performed? Only once this cycle is functioning during the pilot is it worthwhile to roll it out to additional units, prioritized according to downtime costs.
“Most companies want to start with the model and stop at the data. It only works the other way around: If you’ve kept two years’ worth of clean process data, predictive maintenance is practically a given. If you don’t have it, you won’t get it for all the money in the world.”
— Amadeus, Chief Technology Evangelist, CSP
Predictive maintenance isn’t a sure thing and isn’t always worth the investment. Three key points need to be considered before making any investment decision.
What predictive maintenance cannot do: Purely random failures without measurable degradation—such as spontaneous electronic malfunctions—cannot be predicted. Nor can a model replace the maintenance organization: A forecast with a 3-week lead time is useless if the replacement part has an 8-week delivery time. And every model produces false alarms. In pilot projects, initial false alarm rates are often in the double-digit percentage range and only decrease as the data history grows.
| Cost Considerations | Typical Order of Magnitude | What it depends on |
|---|---|---|
| Retrofitting sensors | Typically 1,500 to 5,000 euros per measurement point, including installation | Number of measurement points, available control data (OPC UA significantly reduces requirements) |
| Data connectivity and infrastructure | Depends on the project; often the largest single expense | Heterogeneity of the systems, existing process data management, legacy systems |
| Model and software | Plan for licensing or development, ongoing maintenance | In-house development vs. off-the-shelf software, number of asset types |
| Organization | Training, response processes, roles | Maintenance maturity level |
This is offset by the following effects: In pilot projects, maintenance costs typically decrease by 10 to 30%, while equipment availability increases by 5 to 10%. The rule of thumb for prioritization: The more expensive an unplanned equipment outage is, the faster the predictive maintenance system pays for itself. For a bottleneck system with five-figure downtime costs per hour, the pilot project often pays for itself with the first prevented outage.
Regulatory Classification: Predictive maintenance provides decision support, not autonomous decision-making. The EU AI Act requires transparency and human oversight for AI systems in critical application contexts; in safety-critical industries, fully autonomous approval decisions by AI are not permitted under regulatory guidelines. The approval of a system after maintenance and the evaluation of a forecast remain the responsibility of humans. The EU Product Liability Directive 2024 also broadens the definition of “manufacturer” to include AI-supported decisions: Anyone who incorporates predictions into approval processes must be able to demonstrate the data basis and the human review used to make the decision. This, too, requires a complete, historized database, in accordance with the risk-based approach outlined in ISO 9001:2015, Section 6.1, and the documentation requirements of IATF 16949, Section 7.5.
Predictive maintenance is a maintenance strategy that predicts the maintenance needs of a machine or component before a failure occurs. It is based on continuously collected condition data—such as vibration, temperature, or current consumption—which is compared with historical failure patterns. Maintenance is performed on a condition-based schedule within a planned window, rather than at fixed intervals or only after a failure occurs. The goal is to achieve higher equipment availability while reducing maintenance costs.
Preventive maintenance replaces components at fixed intervals—such as based on operating hours or unit counts—regardless of their actual condition. This approach is predictable but wastes remaining service life: As a rule of thumb, components are replaced when they have 30 to 50% of their service life remaining. Predictive maintenance, on the other hand, is condition-based: sensor data and predictive models determine the optimal time for maintenance. This reduces both unnecessary replacements and unplanned outages, but requires a robust data foundation.
Three categories of data are required. First, condition signals from the equipment—such as vibration spectra, temperatures, current consumption, or torque curves—with a sufficient sampling rate. Second, a documented history of failure events and maintenance events, because the model can only learn from marked failures. Third, contextual data such as production orders and parameter changes, to correctly classify changes in condition. All three categories must be linkable via synchronized timestamps and a unique equipment identifier. Typically, 6 to 12 months of historical data, including failure events, are required before a remaining useful life prediction becomes reliable.
The most suitable equipment includes machines with measurable wear mechanisms and high downtime costs: machining centers with spindles and bearings, interlinked assembly lines, screwdriving systems, pumps, and hydraulic units, as well as bottleneck machines whose failure brings the entire line to a halt. Systems with purely random failures without measurable degradation—such as purely electronic defects—as well as non-critical systems with low downtime costs, for which reactive or preventive maintenance remains more cost-effective, are less suitable.
The costs consist of sensors, data connectivity, software, and organizational implementation. For retrofitting sensors, the typical cost range is 1,500 to 5,000 euros per measurement point, including installation; if control systems already provide the data via OPC UA, this cost item decreases significantly. In many projects, the largest single cost item is data connectivity and data cleansing, which, as a rule of thumb, account for 70 to 80% of the project effort. A pilot project involving 1 to 2 systems can realistically be implemented in 3 to 6 months and often pays for itself in the case of bottleneck systems with the first prevented shutdown.
Predictive Maintenance focuses on the machine and predicts when a component needs to be serviced to prevent a failure. Predictive Quality focuses on the product and predicts whether a component will meet quality requirements while it is still being manufactured. Both approaches use largely the same process data but pursue different objectives: equipment availability on the one hand, and reduction of scrap on the other. In practice, they complement each other because a tool subject to wear often produces inferior parts before it fails.