The vibration was always there. Hearing the failure in it was the work.
A failing bearing announces itself long before it seizes — if you know how to listen. An edge condition monitor turns raw vibration into early warning, and the hard part is not the sensor but separating the signature of incipient failure from the noise of a machine that is simply doing its job.
DomainConnected product, predictive maintenance
PlatformsConnected Edge
ScopeSensing to early-warning signal
Binding constraintTell degradation from normal running
DisclosureRepresentative programme; customer not named
CONTEXT
Where this started.
Rotating machinery fails progressively. A bearing spalls, an imbalance grows, a gear tooth chips, and each leaves a signature in the machine's vibration long before it stops. A condition monitor exists to catch that early enough to schedule a repair instead of suffering a breakdown.
The vibration is always present; the difficulty is interpretation. A healthy machine under changing load and speed produces a rich, variable signal, and the early signature of failure is a small feature inside it. Sensing has to be faithful, processing has to expose the right features, and the decision has to separate genuine degradation from normal variation without crying wolf.
So the node is built end to end for that judgement: faithful vibration sensing, signal processing that surfaces fault features, and a baseline of normal against which degradation stands out — computed at the edge so a fleet of machines does not stream raw waveforms to the cloud.
CHALLENGES
4 problems, named.
Stated before any of them had an answer.
01
Normal is noisy
A healthy machine under varying load and speed produces a rich signal; the early fault signature is a small feature inside that noise, not an obvious spike.
02
Sensing has to be faithful
Mount, bandwidth and noise floor decide whether the fault feature even survives to the processing; a poor sensing chain loses it before analysis.
03
Baselines drift with operation
What counts as normal changes with load, speed and ambient; a fixed threshold either misses faults or floods false alarms.
04
Edge, not cloud
Streaming raw vibration from a fleet of machines is infeasible; the analysis has to happen at the node, on constrained compute.
ARCHITECTURE
How it was built.
The sensor is the cheap part. Turning a variable, load-dependent vibration into a trustworthy early warning — on the node — is the engineering that prevents the breakdown.
CONTRIBUTION
What Faststream did.
The scope of the work, rather than a capability list.
Vibration sensing chain — mount, bandwidth and noise floor chosen so the fault feature survives to the processing.
Signal processing — spectra, envelope analysis and band energies that surface bearing, imbalance and gear signatures.
Operating-point baselines — a model of normal that adapts to load and speed rather than a fixed threshold.
Degradation detection — trending features over time so a developing fault is caught, not a single noisy reading.
Edge computation — running the analysis on the node so a fleet does not stream raw waveforms to the cloud.
Early alerting — warning far enough ahead to schedule maintenance instead of reacting to a breakdown.
WHAT WAS HARD
The parts that consumed the schedule.
Rarely the subsystem that sounds difficult.
01
Finding the feature in the noise
The early signature is small and buried in a variable signal; the processing that reliably surfaces it, across operating conditions, is the crux.
02
Baselines that adapt
A single machine at different loads and speeds has many 'normals'; modelling that so degradation stands out without false alarms is subtle.
03
Faithful sensing
Get the mount or bandwidth wrong and the fault feature is gone before analysis; the sensing chain is as important as the algorithm.
04
Doing it on the node
Constrained edge compute has to run the analysis a workstation would do easily; the processing is designed to fit, not assumed.
OUTCOME
What resulted.
Failures caught early
Developing faults detected far enough ahead to schedule a repair rather than suffer a breakdown.
Few false alarms
Degradation separated from normal variation, so warnings are trusted and acted upon.
Adapts to operation
Baselines that follow load and speed instead of a brittle fixed threshold.
Runs at the edge
Analysis on the node across a fleet, without streaming raw waveforms to the cloud.
Confidentiality
Customer projects are presented at property, capability, outcome and integration level. Customer names, internal architecture, confidential deliverables and commercial terms are not disclosed. Where a detail would identify a customer it is omitted rather than approximated. More is available under a non-disclosure agreement, within the limits the customer has agreed.
Why is condition monitoring more than fitting a vibration sensor?
Because the vibration is always there; the value is interpreting it. A healthy machine under changing load and speed produces a rich, variable signal, and the early signature of a failing bearing or a growing imbalance is a small feature buried inside it. The engineering is the sensing chain that preserves that feature, the signal processing that surfaces it, and the baseline of normal that lets genuine degradation stand out — not the sensor alone.
02
How do you avoid false alarms?
By modelling normal properly. A machine has many normal states depending on load, speed and conditions, so a fixed vibration threshold either misses faults or floods false alarms. Building baselines per operating point, and trending features over time rather than reacting to a single reading, is what separates real degradation from ordinary variation — and a low false-alarm rate is what makes the warnings trusted and acted upon.
03
Why run the analysis at the edge?
Because streaming raw vibration waveforms from a fleet of machines to the cloud is infeasible in bandwidth and cost. Doing the feature extraction and the degradation decision on the node means only compact results or alerts leave the machine, the warning is immediate, and the system scales across many machines. The trade is that the processing must be engineered to fit constrained edge compute rather than assuming a workstation.