The question that decides whether it makes sense
Predictive maintenance is the industrial AI use case most often mentioned in equipment vendor presentations and Industry 4.0 events. It’s also one of the most frequently oversold.
The starting question isn’t whether AI can predict breakdowns — under certain conditions, it can. The question is whether the cost of unplanned breakdowns is high enough to justify the sensors, data, and model infrastructure predictive maintenance requires, and whether your plant has the technical conditions for the model to work with enough accuracy.
Start by calculating the real cost of an unplanned stoppage on your most critical line: lost production, emergency repair costs, late-delivery penalties, and the impact on downstream production. That figure, multiplied by how often it happens, defines the value of the problem.
Assessing the opportunity
Predictive maintenance is especially well suited to equipment with these characteristics: failures with high financial consequences (high downtime cost or safety risk), degradation patterns with early warning signs that sensors can detect, and enough failure history to train the model.
It’s less suited to equipment that fails suddenly without warning (instantaneous catastrophic failures), equipment where preventive maintenance costs less than the predictive system, and equipment that fails so rarely the model doesn’t have enough examples of the positive class.
Data requirements
Real-time sensor data. Temperature, vibration, pressure, electrical current, acoustics — the relevant variables depend on the type of equipment. Sensors have to be installed before training data can be collected, which means that for equipment with no existing instrumentation, there’s a data capture phase before model development can begin.
Labeled failure history. This is the hardest requirement to meet. To train a model that predicts failures, you need a history of failure events along with the sensor readings leading up to each one. If maintenance records lack reliable timestamps, don’t specify the cause of failure, or the maintenance management system (CMMS) isn’t kept up to date, building this history is a project in itself.
Enough failure events. Failures are rare. If a machine fails twice a year, two years of data gives you four labeled failure events — not enough to train a model with real discriminative power. AI predictive maintenance needs a minimum number of failure events in the history, and that number varies with the type of failure and the complexity of the model.
Sensor data quality and continuity. Sensors drift out of calibration, fail, or drop data. Gaps in the time series and sensor outliers (as distinct from process outliers) make preprocessing significantly harder.
Cost structure and complexity
The most visible cost is instrumentation: installing sensors on uninstrumented equipment, an industrial communication network (OPC-UA, MQTT), and time-series storage infrastructure. In plants with no IoT foundation, this can be the project’s biggest investment.
The prediction model requires specialized engineering for industrial time series: building time windows, extracting frequency-domain features (FFT for vibration, for example), and calibrating the alert threshold to balance false positives and false negatives.
Recurring costs include maintaining the sensor system, retraining the model when major components are replaced (the degradation pattern changes), and running the alerting system day to day.
Risk assessment
Insufficient data. This is the most common risk. The plant has sensors but no labeled failure history, or the history is too short. The project then needs a data capture period before the model can be developed.
False positives. A model with a high false alarm rate breeds distrust in the maintenance team: if the system warns of breakdowns that never happen, technicians stop acting on alerts. Calibrating the alert threshold to minimize false positives without losing detection is one of the most important technical challenges.
Resistance from the maintenance team. Technicians with years of experience can tell when a machine “sounds wrong” or “acts strange.” If the predictive system conflicts with that judgment, or if technicians don’t understand why the model is raising an alert, adoption can be low. Model explainability — why it’s alerting now — is a critical adoption factor.
Success metrics
- Fewer unplanned stoppages (number and duration) vs. the previous period
- MTBF (mean time between failures) on monitored equipment
- Early detection rate: the share of failures the model catches before they happen
- Average warning time: how far ahead of the failure the alert fires
- False positive rate: alerts that don’t correspond to a real failure
- Planned vs. unplanned maintenance ratio before and after
How it delivers ROI
The sources of value are lower costs from unplanned stoppages (lost production + emergency repair + overtime), longer component life through intervention at the optimal moment (not too early, not too late), and smaller emergency spare parts inventories, because needs are known in advance.
Predictive maintenance can also be justified by lower safety risk: on equipment where a failure could put people at risk, the value of prevention includes the avoided cost of an incident — hard to quantify, but real.
Frequently asked questions
Can predictive maintenance work without real-time sensors? In some cases, process data already available in the SCADA system or the PLC history contains enough signal. For low-complexity equipment, even power consumption data can predict degradation. Whether it’s feasible without new sensors depends on the type of equipment and the type of failure you want to predict.
What’s the difference between predictive maintenance and condition-based maintenance? Condition-based maintenance uses fixed thresholds on process parameters to trigger interventions (if temperature exceeds X, inspect). ML-based predictive maintenance learns the multivariate degradation patterns that precede failure and raises alerts before classic thresholds are triggered. ML’s advantage is that it captures interactions between variables that individual thresholds miss.