Predictive Maintenance & Condition Monitoring
Detect bearing wear, gearbox degradation, and blocked filters from live sensor data — often with nothing more than a microphone and signal processing.
Book a 30-min technical fit callPhD APC/MPC · CERN · IAV · SkySails
From Reactive to Predictive
Equipment failures are rarely sudden. Bearings degrade over weeks and their acoustic signature shifts as pitting develops; gearboxes give measurable warnings long before they seize; filters restrict flow gradually, visible in pressure and acoustic data long before the machine trips.
Reactive maintenance means unplanned downtime, emergency sourcing and lost production. Fixed-interval maintenance replaces parts that still have service life. Predictive maintenance acts when the data says so — not on a calendar and not in a crisis.
What Can Be Monitored
Wear monitoring — bearings, gears, impellers
Bearing defect frequencies (BPFO, BPFI, BSF, FTF) appear as narrow spectral peaks long before audible noise develops; gear-mesh harmonics shift as teeth wear. Envelope analysis separates these indicators from broadband noise at signal-to-noise ratios where the ear hears nothing unusual.
Abnormal conditions — filters, pumps, motors
A clogged filter raises differential pressure and changes the acoustic signature of the flow path. Pump cavitation produces characteristic broadband noise. Motor current signature analysis (MCSA) detects eccentricity, broken rotor bars and bearing defects from the supply current alone — no mechanical access needed.
Technical Approach
Signal acquisition
Low-cost sensors are usually enough: a microphone for airborne sound, a MEMS accelerometer for structure-borne vibration or a current clamp for motor signature analysis. Often one sensor on the housing suffices.
Feature extraction
FFT, envelope and cepstral analysis and time-domain metrics — RMS, kurtosis, peak-to-peak — isolate fault frequencies and track their amplitude against a healthy baseline.
Baseline and anomaly detection
Record a healthy-machine reference, then monitor the deviation with threshold rules, statistical process control or a lightweight anomaly model. No cloud infrastructure or special hardware required.
Integration and alerting
Results go to the existing SCADA, PLC HMI or a stand-alone dashboard, with configurable thresholds, trends and plain-language fault descriptions — not just a red light.
Why It Works
No expensive sensors needed
A USB microphone or a low-cost MEMS accelerometer can reveal bearing faults, gear damage and filter blockage with the right signal processing. The value is in the analysis, not the hardware.
Maintenance on your schedule
Replace the part in the next planned production window — not during an emergency shutdown at 2 a.m.
Non-invasive installation
Sensors are surface-mounted or clamped: no machine modification, no PLC change, no production interruption during installation.
Interpretable alarms
Fault frequencies have a physical meaning, so an alarm shows which component is affected and why — not just an anomaly score.
Technology Stack
Sensor types
Accelerometers (MEMS or piezoelectric) for structure-borne vibration, microphones for airborne sound, current clamps for MCSA, ultrasonic thickness sensors for corrosion and erosion, thermal cameras for hot spots, oil-debris sensors for lubrication. In most cases one well-placed low-cost sensor is enough.
Signal processing
The FFT splits the vibration spectrum into frequency components. Bearing defect frequencies follow from the bearing geometry and appear as predictable peaks. Envelope analysis demodulates high-frequency resonances to isolate impulsive fault energy; cepstral analysis separates source and transmission path in gear diagnostics.
Machine-learning methods
Anomaly detection (isolation forest, autoencoders) tracks deviation from a healthy baseline without labelled failure data. Classifiers (SVM, random forest, gradient boosting) identify the fault type once labelled examples exist. Remaining-useful-life models need run-to-failure data.
Deployment
Results reach SCADA / PLC HMI via OPC UA, MQTT or REST. Edge deployment for latency-sensitive cases; central deployment for fleet-wide models across sites. Alarm levels: early warning (trend), advisory (schedule soon), urgent (inspect this week).
Reactive vs. Preventive vs. Predictive
| Aspect | Reactive | Preventive | Predictive |
|---|---|---|---|
| When action is taken | After failure | On a fixed schedule | When the data indicates a need |
| Main cost driver | Emergency repair and downtime | Over-maintenance and discarded parts | Monitoring system |
| Downtime | Unplanned, long | Planned, sometimes unnecessary | Planned, minimal |
| Parts replaced | When failed | By time interval | By actual condition |
| Operational risk | High | Medium | Low |
Relevant Design Patterns
Book a 30-min technical fit call
If you suspect a reliability problem or want to set up monitoring before the next shutdown — a 30-minute call is enough to assess which signals to collect and what analysis will work.
Book a 30-min technical fit call →