Types of maintenance: from reactive to predictive
There are four maturity levels in industrial maintenance:
- Reactive — repair when it fails. The most costly in terms of unplanned downtime
- Preventive — maintenance at fixed intervals based on the calendar or operating hours
- Condition-based predictive — intervention when sensors indicate real degradation
- Prescriptive with AI — the system not only predicts the failure but also recommends the optimal action
How much does an unplanned stop cost? In continuous process industries (paper, chemicals, food), one hour of unplanned downtime can cost between USD 10,000 and 100,000 depending on the sector. Predictive maintenance typically pays back its investment in 6 to 18 months.
What data do you need?
The raw material of predictive maintenance is the equipment's condition signals. The most relevant for motors, pumps, compressors and drivetrains:
- Temperature — bearings, motor windings, cooling
- Vibration — imbalance, misalignment, bearing wear
- Electric current — insulation degradation, overload, rotor eccentricity
- Pressure and flow — pump cavitation, blockages, impeller wear
- Operating hours and cycles — accumulated service life of components
AI algorithms applied
Not every problem requires complex neural networks. For most industrial cases, the most effective algorithms are:
| Algorithm | Application | Complexity |
|---|---|---|
| Statistical process control | Detecting deviation from normal parameters | Low |
| Isolation Forest | Unsupervised anomaly detection | Medium |
| LSTM (recurrent networks) | Failure prediction with temporal patterns | High |
| Computer vision | Inline visual defect detection | High |
In real industrial projects, starting with Statistical Process Control (SPC) and deviation rules gives immediate results without the complexity of training ML models. More sophisticated models are justified when there is enough historical failure data for training.
How to implement it step by step
- Instrument critical equipment — if you do not have vibration or temperature sensors, installing them is the first step
- Historize the signals — at least 3-6 months of normal operating data to establish a baseline
- Define alert thresholds — start with simple rules based on deviations from the baseline
- Configure automated alerts — Grafana or the SCADA alarm system
- Validate with the maintenance team — adjust thresholds to reduce false alarms
- Move toward predictive models — once you have real failure data to train on
Grupo Salba + AI: We have implemented fault detection with computer vision (inline counting and defect detection) and anomaly detection models in continuous processes. Our Datalogger I4 makes it easier to collect labeled data for model training.
Want to implement predictive maintenance?
We evaluate your critical equipment, identify the sensors you need and define the shortest route to predictive maintenance with a calculated ROI.
See the Industry 4.0 service →