Predictive maintenance (PdM)
Predictive maintenance uses sensor data and machine learning to predict equipment failures before they happen, scheduling maintenance only when the model indicates it's needed.
Predictive maintenance (PdM) sits between reactive maintenance (fix it when it breaks) and preventive maintenance (replace it on a schedule). The promise is straightforward: instead of replacing a $50,000 bearing every 12 months because the manual says to, replace it only when vibration signatures, temperature trends, or current draw indicate it's actually wearing out. Real-world deployments typically cut unplanned downtime 20-40% and maintenance cost 10-25% — but require investment in sensors, data pipelines, and ML expertise.
What PdM looks like in practice
A typical PdM deployment: install vibration, temperature, and current sensors on critical equipment. Stream the data into a time-series database. Train an ML model on historical data — usually a combination of physics-based features (RMS vibration, kurtosis, FFT peaks) and learned representations (autoencoders, anomaly detection). When the model flags a deviation, trigger a maintenance work order in CMMS or ERP. Iterate: every failure that the model missed and every false alarm becomes training data.
What makes PdM hard
Labeled failure data is rare — most equipment doesn't fail often. Sensor placement matters enormously: a vibration sensor on the wrong axis tells you nothing. Maintenance teams have to trust the model enough to act on it, which requires explainable outputs. Most early PdM projects fail not because the ML is wrong, but because the integration into existing workflows is wrong.
Algorithms commonly used
For binary failure prediction with labeled data: gradient-boosted trees (XGBoost, LightGBM), random forests, neural networks. For anomaly detection without failure labels: isolation forests, autoencoders, one-class SVMs. For remaining-useful-life estimation: survival models, recurrent neural networks, transformer-based time-series models. Pick the simplest model that works.
Frequently asked questions
How much data do I need before PdM is viable?
Practical rule: at least 3-5 instances of each failure mode you want to predict, plus 6+ months of normal-operation data. Below that, start with anomaly detection (unsupervised) and accumulate labeled failures over time.
What sensors are typical for PdM?
For rotating equipment: triaxial vibration, temperature, motor current signature analysis. For pumps: pressure, flow, motor current. For HVAC: temperature, humidity, current, runtime hours. For batteries: voltage, internal resistance, temperature.
Does S2 Data Systems build PdM systems?
Yes. S2 builds end-to-end PdM solutions — sensor data ingestion (from Monnit, Wovyn, or any vendor), feature engineering, ML model development, model serving, alerting integration into CMMS/ERP, and ongoing model retraining. We are hardware-agnostic and integrate with existing SCADA, historian, or IoT platforms.
Building on Predictive maintenance (PdM)? Talk to S2 Data Systems.
Book a 30-minute call. We’ll scope how this fits into your IoT data and analytics roadmap.
Book a strategy call