Predictive maintenance
Bearing damage, imbalance and cavitation surface in the residual weeks before failure — reported with a cause, not just a threshold crossing.
Vibration · AcousticsIndustrial edge AI / Sweden
Noether finds machine faults by subtracting what physics already explains.
Hybrid physical and learned models for predictive maintenance and virtual sensing, running beside the machine with no cloud in the control loop.
- < 10 ms
- Target inference at the machine
- 0
- Process bytes sent to a cloud loop
- 1
- Machine in the first pilot
- 4 weeks
- From signal to edge deployment
01 / The problem
A machine can whisper for weeks before a conventional alarm hears it.
Threshold alarms watch amplitude. Amplitude is often the last thing to move: lower the threshold and normal operating variation triggers false alarms; leave it high and damage is already advanced when it trips.
Conventional thresholdWaits for the whole signal to grow
Simple to deploy, but normal machine vibration masks the earliest fault signature.
Physics-based residualRemoves healthy behaviour first
The remaining signal is smaller, structured, and concentrated around what changed.
02 / The approach
Physics carries the structure. Machine learning handles the remainder.
- 01
Predict
Describe healthy behaviour from rotordynamics, heat transfer, or fluid mechanics.
- 02
Subtract
Compare the physical prediction with the sensor stream and isolate the residual.
- 03
Learn
Train a compact model on the residual, where wear and imbalance are easier to separate.
- 04
Deploy
Quantise the result for an NPU or MCU beside the machine, operating network-down.
03 / The evidence
A bearing fails over ninety days. Only one of these plots notices.
Both plots come from the same simulated 1490 rpm pump. Drag operating days forward. The measured amplitude barely changes while periodic impacts emerge in the residual.
- Measured RMS
- 0.460 —
- Residual kurtosis
- 2.9 —
The bearing is sound. Measurement and model agree; the residual is noise.
Fig. 1 — Kurtosis measures how impulsive a signal is rather than how large. A spalled race strikes sharply and periodically, so kurtosis climbs long before amplitude does. The values are computed from the plotted samples as you drag.
What changes
By day seventy, measured amplitude has moved by roughly one percent, inside the tolerance a practical threshold needs. Residual kurtosis has moved by several hundred percent. The evidence was present in the signal; the physical model made it visible.
Pump + bearingPhysics modelResidual modelCause + lead time
04 / Applications
One method, four useful industrial questions.
Virtual sensors
Estimate quantities you cannot instrument — temperature inside the tool, torque along the shaft — from the sensors already fitted.
State estimationProcess optimisation
Trade energy against quality in real time under hard physical constraints, so the optimiser can never walk the process outside its safe envelope.
Energy · YieldExplainable anomaly detection
When physics is the reference, every deviation is interpretable: what departed, by how much, and from which conservation law.
Auditable by design05 / First engagement
One machine. One failure mode. Four weeks.
The deliverable is not a report. It is a model running on your hardware against your process, plus an honest answer about whether the residual carries a useful signal.
- Week 1
One machine, one failure mode
I pick a single asset that costs you money when it stops, and write down the physics that governs it. No instrumentation project — I start from the sensors you already have.
- Weeks 2–3
Hybrid model on your data
The physical model is fitted to your process and the residual model trained on your historian. You see which deviations it can already account for and which it cannot.
- Week 4
Running on the edge box
The model is quantised and deployed on hardware at the machine, streaming live. You keep the box and the model whether or not we continue.
06 / Builder
Built by a physicist who also writes the firmware
Noether sits where theoretical physics meets hardware-near programming: the differential equations, the signal processing and the code on the microcontroller from the same pair of hands. That combination is exactly what hybrid edge-AI demands, and it is rare — most teams have one half and outsource the other.
Anyone can train a network on cloud data. Far fewer can write down a process's Lagrangian, discretise it, and make it fit in 512 kB of RAM.
Pilot partners / Nordic industry
Bring me a machine that costs you money when it stops.
I am looking for two or three pilot partners in Nordic manufacturing and process industry. You get a hybrid model running on your own machine; I get validation data from a real plant instead of a benchmark dataset.
info@noether.se