Calendar-based preventative maintenance is a solution to an information problem: without data on the real state of a piece of equipment, you replace parts at safe intervals so you don't get caught off guard. The problem is that this model systematically wastes parts with remaining useful life, accumulates unnecessary planned downtime costs, and yet does not eliminate unexpected failures — because the calendar is a rough approximation of the actual behavior of each machine in each usage condition. Sensors and predictive models solve the information problem. What no one tells you often enough is that the biggest obstacle to adoption is not technological.
How predictive maintenance works in practice
The logic is straightforward: deteriorating equipment emits physical signals before failing. Nonstandard vibration, anomalous temperature, noise at specific frequencies, irregular electrical current consumption — these signs precede failure by hours, days or weeks, depending on the type of equipment and the failure mechanism.
IoT sensors capture these signals in real time. Sampling rate matters: a vibration sensor that collects data once a minute may miss anomalies that appear and disappear in cycles of seconds. Edge computing platforms process some of this data locally, sending only relevant alerts and summaries to the cloud — which reduces transmission costs and latency in industrial environments with limited connectivity.
The machine learning models that analyze this data can be of different types. Anomaly detection models identify when equipment behavior deviates from historical patterns, without necessarily needing labeled data from previous failures. Useful life prediction models (Remaining Useful Life, or RUL) attempt to quantify how long equipment has before a likely failure, allowing intervention to be scheduled at the optimal time. Diagnostic models attempt to identify not just that something will fail, but what—the specific type of failure, the affected component—to guide the work of the maintenance team.
Where it already works and what it proves
Aviation is the reference sector. Jet engines are monitored in flight by hundreds of sensors; Data reaches maintenance teams on the ground in real time before the aircraft even lands. Rolls-Royce and GE Aviation sell engines as a service — the customer pays per flight hour, not for the turbine — a model that only works with reliable predictive maintenance, because the supplier bears the cost of each unplanned stop.
In manufacturing, results published by manufacturers such as SKF, Siemens and ABB show reductions of 25% to 40% in unplanned maintenance costs and increases in equipment availability in the range of 10% to 20%, depending on the environment and starting point. In power generation — wind turbines, thermal plants, hydroelectric plants — the impact of an unplanned shutdown is so high that the ROI of sensors and models is justified even with relatively low failure rates.
In Brazil, mining, cellulose and paper and petrochemical industries have the most mature cases. Petrobras has had structured predictive maintenance programs on offshore platforms for more than a decade. The segment of medium-sized industrial companies is still mostly on calendar preventive maintenance, which represents both a significant gap and a considerable opportunity for whoever implements it first.
The real problem: legacy systems and trust in teams
Sensor technology is relatively mature — and the cost per monitoring point has fallen by between 60% and 80% in the last five years. ML models for predictive maintenance have extensive literature and consolidated open source tools. Why, then, is adoption still slow outside of large corporations?
The first obstacle is integration with legacy systems. Old industrial PLCs, past-generation SCADA systems, and ERPs that were never designed for real-time data ingestion create a patchwork of connectivity that makes aggregating sensor data consistently expensive and laborious. Many predictive maintenance projects die in the integration phase, not the modeling phase.
The second hurdle is subtler and more difficult: getting operations teams to trust the model rather than the calendar. A technician with fifteen years of experience on a production line has developed intuition about when a machine "feels weird." Presenting him with a dashboard that says "probability of failure: 73% in the next 5 days" without context as to how the model arrived at that number — and asking him to cancel scheduled maintenance based on the calendar because the model says everything is fine — is a recipe for resistance. Confidence in the model is built gradually, with transparency about the variables that feed it and with a history of successes that justifies giving weight to its predictions.
How a leader should look at this
The ROI of predictive maintenance is well established in high availability and critical equipment contexts. The strategic question for a decision maker is not “does it work?” — it works. It's "what is the correct entry point for our operation?"
The answer is almost always to start with critical equipment, with high downtime costs, where historical failure data exists and where the maintenance team is already dissatisfied with the current calendar model. Well-chosen pilot projects generate a track record of success that builds internal trust and creates the business case for expansion.
The trap is the pilot project on non-critical equipment: the result will be irrelevant, the ROI will be difficult to demonstrate, and the internal conclusion will be that "predictive doesn't work for our reality". The other trap is the overly broad project that tries to implement the entire plant at once — the integration cost explodes, the deadline extends, and the organization loses steam before seeing any results.
From a technology perspective, the cost of entry has dropped dramatically. Quality industrial sensors cost a fraction of what they did five years ago. Analytics platforms like AWS IoT, Azure IoT Hub, and open source options like InfluxDB with scikit-learn models are accessible for midsize technical teams. The main investment is in integration and change management with the operations teams — and this investment is where the project wins or loses.
Also read
- Digital twin: simulation that learns from real equipment
- Cybersecurity in 2025: A Strategic Vision for the Future
- How to build a business case for AI in industrial operations
- Inference chips and ASIC: when specialized beats generic
- Specialized chips and the end of the generic CPU era
- Neuromorphic and bio-inspired computing: when the chip starts to learn from the brain