An unplanned machine stoppage on a production line can cost thousands of euros per minute. Against this risk, predictive maintenance (PdM) offers a powerful idea: rather than repairing something once it breaks (corrective maintenance) or replacing parts on a fixed calendar even when they are still healthy (preventive maintenance), the aim is to predict when an asset is going to fail based on its actual condition and to intervene just in time. Artificial intelligence applied to sensor data turns that prediction into a reliable industrial operation. This article explains the full technical architecture — from sensor to model to decision — and clarifies what works, what fails and which regulations apply.
From Corrective to Predictive: Why Calendars Are Not Enough
Interval-based preventive maintenance has a well-known statistical problem: most mechanical failures do not follow an age-related wear curve but instead follow the bathtub curve pattern, with infant mortality and random failures. Replacing a healthy bearing "because it is scheduled" sometimes introduces the very failure it was meant to prevent. Predictive maintenance, by contrast, is based on the actual condition of the asset (CBM, Condition-Based Maintenance) and estimates the Remaining Useful Life (RUL) — the number of cycles or hours the equipment can continue to operate before degrading below the acceptable threshold.
The Data Layer: IoT Sensors and Telemetry
Every predictive model rests on the quality of its telemetry. Typical plant sensors capture vibration (accelerometers — the richest signal for rotating machinery), temperature, electrical current, pressure, acoustic emission and ultrasound. Vibration analysis deserves closer attention: a Fast Fourier Transform (FFT) decomposes the signal into its frequency spectrum, and each mechanical defect — imbalance, misalignment, looseness, bearing race defect — leaves a characteristic signature at specific frequencies. Detecting the anomalous harmonic is, literally, hearing the fault before it occurs.
The architecture collects this data at the edge (processing close to the machine to reduce latency and bandwidth), normalises it and transmits it to a platform where the model is trained and served. Temporal synchronisation and sampling quality are critical: an accelerometer that is poorly mounted or under-sampled produces garbage data, and no algorithm can recover information the sensor never captured.
The Models: From Unsupervised Anomaly Detection to RUL Estimation
There are three families of approaches depending on the available data. Unsupervised anomaly detection (Isolation Forest, autoencoders) is used when there is plenty of normal-operation data but few failure examples — the most common situation in a plant, because machines fortunately fail infrequently. Supervised classification identifies the type of failure when labelled historical data is available. And RUL regression, often using LSTM recurrent networks or sequence models, estimates how many cycles remain; this is the crown of PdM because it converts an alert into a planning window.
The fundamental challenge is class imbalance: if only 0.1% of samples correspond to a failure, a model that always predicts "everything is fine" achieves 99.9% accuracy and is completely useless. For this reason, PdM does not focus on accuracy but on precision and recall for the minority class, and it balances the cost of a false negative (catastrophic stoppage) against the cost of a false positive (unnecessary maintenance).
From Prediction to Return: How to Justify the Project
A predictive system pays for itself through the reduction of unplanned stoppages, but its economic justification must be built with recognisable operational metrics. The reference metric is OEE (Overall Equipment Effectiveness), which multiplies availability, performance and quality; predictive maintenance directly attacks availability by reducing unplanned downtime. Alongside OEE, two classic reliability indicators are tracked: MTBF (Mean Time Between Failures), which measures how often an asset fails, and MTTR (Mean Time To Repair), which measures how long it takes to return the asset to service. PdM aims to raise MTBF by preventing premature failures and to reduce MTTR by turning an emergency into a planned intervention with spare parts and a technician already available.
The business case is quantified by comparing the cost avoided — hours of downtime multiplied by the line's cost per hour, plus secondary damage and delivery penalty costs — against the cost of implementation: sensorisation, platform, integration with the CMMS (Computerised Maintenance Management System) and data science effort. The common trap is promising savings before a measured baseline exists: without knowing the current OEE and the real cost of existing stoppages, there is no honest way to demonstrate improvement. This is why the first deliverable of a PdM project is not a model — it is an instrumented baseline.
Comparative Table of Maintenance Strategies
| Strategy | Trigger | Main Risk | Cost |
|---|---|---|---|
| Corrective (run-to-failure) | Failure has already occurred | Unplanned stoppage, cascading damage | Low until failure; very high at failure |
| Preventive (calendar-based) | Fixed time/cycle interval | Replacing healthy parts; infant mortality | Medium, with wasted useful life |
| Predictive (CBM + AI) | Actual condition / estimated RUL | False negatives if the model is weak | High to implement; low to operate |
| Prescriptive | Prediction + action recommendation | Dependency on data and orchestration | Highest; maximum potential return |
Applicable Regulations: From the AI Act to ISO Maintenance Standards
An industrial predictive system operates within a growing regulatory framework. The European Artificial Intelligence Act (AI Act, Regulation (EU) 2024/1689) classifies systems by risk; most predictive maintenance applications fall into the limited or minimal risk category, but if the system participates in the safety of a machine covered by harmonisation legislation — the new Machinery Regulation (EU) 2023/1230 — it may be classified as high risk, with obligations regarding risk management, data quality, technical documentation, human oversight and robustness. Added to this is the ISO asset management family (ISO 55000) and the machinery condition monitoring standards (the ISO 13374 and ISO 17359 series), which standardise how signals are processed and diagnoses are generated. Ignoring this framework is not merely a legal risk: the traceability and data quality the standards demand are, moreover, what makes the model reliable in the first place.
Common Errors in Predictive Maintenance Projects
- Starting with the algorithm rather than the sensor. Without quality, well-synchronised telemetry, even the best model predicts noise.
- Optimising accuracy with imbalanced classes. The misleading metric hides a model that never detects the actual failure.
- Failing to define asymmetric costs. A false negative (breakdown) and a false positive (unnecessary maintenance) do not cost the same; the decision threshold must reflect this.
- Neglecting retraining. Machines age and change operating regimes; a frozen model degrades (data drift) and loses reliability.
- Skipping human oversight. The AI Act and common sense both require that a technician validates critical recommendations before acting on a safety-relevant asset.
Frequently Asked Questions
What is the difference between preventive and predictive maintenance?
Preventive maintenance intervenes on a fixed calendar or cycle basis, regardless of the equipment's actual condition. Predictive maintenance intervenes based on measured condition and estimated remaining useful life, avoiding both breakdowns and unnecessary replacements.
What is RUL?
Remaining Useful Life is the estimate of the time, cycles or hours that an asset can continue to operate before degrading below an acceptable threshold. It is the target output of regression models in PdM.
Why is vibration analysis so important?
Because in rotating machinery every mechanical defect leaves a characteristic signature at specific frequencies. The Fourier transform allows those harmonics to be isolated and the problem — imbalance, misalignment, bearing defect — to be diagnosed before it produces a visible failure.
Does predictive maintenance fall under the AI Act?
It depends on the application. Many uses are limited risk, but if the system intervenes in the safety of a regulated machine it may be classified as high risk, with obligations covering risk management, data quality, documentation and human oversight.
Conclusion
Predictive maintenance is not an artificial intelligence project — it is an industrial data engineering project in which the model is simply the final layer. Its real value emerges before the algorithm: in a well-mounted accelerometer, in synchronised telemetry and in an honest definition of the cost of being wrong in each direction. When that foundation is properly laid, RUL regression transforms an alert into a planning window that a plant manager can act on without stopping the line. And when the regulatory framework is respected — the AI Act, the Machinery Regulation and ISO 55000 — the system not only predicts better, but becomes auditable and defensible. At Summum Artificial Intelligence we design that complete chain, from sensor to model to decision, with human oversight integrated wherever safety demands it.