Analyse
Predictive maintenance, starting from what actually breaks
Predictive maintenance works when it is applied to the small number of assets whose failure genuinely hurts, using condition data that genuinely leads the failure. Applied to everything at once with a general-purpose algorithm, it produces alerts nobody trusts. We start from your failure history, not from a platform.
The problem
Maintenance is either too early or too late, and both are expensive
Calendar-based maintenance replaces healthy components on a schedule and still misses the failures that do not follow a calendar. Breakdown maintenance is cheaper right up to the unplanned stoppage that costs a week of production. Neither knows the actual condition of the machine.
- 01
Critical assets fail without warning, repeatedly
- 02
Components are replaced on schedule while still in good condition
- 03
The same failure mode recurs and nobody has established the root cause
- 04
Spares are ordered reactively at premium cost and lead time
- 05
Maintenance history lives in a notebook or in one person's memory
What we build
Concrete deliverables
Nouns, not adjectives. This is what actually gets handed over.
- 01
Criticality and failure-mode review
Which assets matter, how they actually fail, how a failure develops, and what measurable parameter changes first. This determines everything else.
- 02
Condition monitoring instrumentation
Vibration, temperature, current signature, acoustic, pressure and flow sensing, specified per failure mode rather than fitted uniformly.
- 03
Edge signal processing
FFT, envelope analysis and feature extraction on the device, so high-frequency data is used without being transmitted.
- 04
Baselines and thresholds
Per-asset baselines established from a healthy period, with alerting on deviation and trend rather than on fixed absolute limits.
- 05
Anomaly detection where it is justified
Statistical and model-based detection applied to assets with enough history to support it — and not applied where it would only generate noise.
- 06
Maintenance workflow integration
Alerts that become work orders in your CMMS or ERP, with feedback captured so the model and thresholds improve.
How it works
The technical path, step by step
Where the engineering decisions actually get made.
- 01
Start from failure history
We review what has actually broken, how often, and what it cost. Assets with no consequential failure history do not need this.
- 02
Identify the leading indicator
For each failure mode, what changes first — bearing frequencies in vibration, winding temperature, current imbalance, pressure differential.
- 03
Instrument and baseline
Sensors fitted and a healthy baseline established over a representative operating period, including all normal load states.
- 04
Detect deviation
Trend and threshold detection first, because it is explainable and it works. Model-based detection added only where the data supports it.
- 05
Close the loop
Every alert is followed up and the outcome recorded — true, false, or too early. Without this feedback the system degrades into noise within months.
Protocols & technologies
Specifics, because vague answers cost you money later
Chosen per site according to what is installed, not according to what we would prefer to work with.
| Technology | Where it is used | Engineering note |
|---|---|---|
| Triaxial vibration sensors (IEPE / MEMS) | Rotating equipment | MEMS is adequate for trend detection and far cheaper; IEPE where diagnostic detail is needed. |
| FFT / envelope analysis at the edge | Bearing and gear defect frequencies | Raw waveform processed locally; only features transmitted. |
| RTD / thermocouple / thermal | Winding, bearing and process temperature | The simplest and often the most reliable leading indicator. |
| Motor current signature analysis | Motor and driven-load faults | Non-invasive; the sensor goes in the panel, not on the machine. |
| Ultrasonic / acoustic | Leaks, early bearing wear, steam traps | Very effective on compressed air, which is usually a large uncosted loss. |
| Oil condition sensing | Gearboxes and hydraulics | Where the asset justifies it. |
| CMMS / ERP integration | Work order creation | An alert that does not become a task is not maintenance, it is a notification. |
Brownfield
Fitted to running machines, without modification
Condition monitoring sensors mount externally — magnetically or by stud on a bearing housing, clamped around a cable in the panel, or fitted to an existing tapping point. Most installations happen during a routine maintenance window without any change to the machine or its control system.
Works with
- Motors, pumps, fans, blowers and compressors
- Gearboxes, spindles and conveyors
- Chillers, air compressors and utility plant
- Machines already connected through our IIoT layer
- Assets with existing vibration route-based monitoring, made continuous
Use cases
What people actually ask us for
Each one is a situation followed by the outcome it produces — not a feature list.
Critical motor protection
NowA single motor failure stops the whole plant.
AfterContinuous vibration and temperature trending with early deviation alerts and time to plan the intervention.
Compressed air leak detection
NowCompressors run longer than the demand justifies.
AfterUltrasonic survey plus consumption trending, which typically identifies a substantial recoverable loss.
Gearbox condition
NowGearbox failures are catastrophic and have long lead times for replacement.
AfterDefect-frequency trending giving enough warning to order and schedule rather than react.
Pump cavitation
NowPumps degrade and efficiency falls unnoticed.
AfterVibration and pressure signatures identify cavitation and wear before failure.
Spare parts planning
NowSpares held for everything, or for nothing.
AfterCondition-driven ordering, with stock levels informed by measured asset condition rather than by anxiety.
Implementation
How a project runs
Eight stages, each producing something you own. Full detail on the process page.
- 01
Discover
1–2 conversations
- 02
Audit
1–3 days on site
- 03
Design
1–2 weeks
- 04
Engineer
2–8 weeks per phase
- 05
Integrate
1–2 weeks
- 06
Deploy
Your shutdown window
- 07
Monitor
First 4–6 weeks
- 08
Optimise
Ongoing, if you want it
Vibration condition monitoring with edge processing
High-frequency vibration data processed on the gateway so that only features travel upstream — the difference between a practical system and an unaffordable one.
Read the write-upQuestions
Predictive Maintenance — common questions
Do we need AI for predictive maintenance?
Usually not at the start. Trend and threshold detection on the right parameter catches the large majority of developing faults and — crucially — is explainable, so maintenance teams act on it. Model-based anomaly detection adds value once there is enough history including actual failures, which typically means a year or more. Anyone leading with AI before establishing that has the order backwards.
How long before it predicts anything?
A baseline needs a representative healthy period across all normal operating states — typically a few weeks to a couple of months. Deviation detection works from the end of baselining. Anything claiming useful prediction in the first week is detecting something other than machine condition.
Which assets should we start with?
The ones where a failure has actually cost you significantly, and where the failure mode develops gradually enough to be detected. A motor that fails progressively is a good candidate; an electronic board that fails instantly is not. We work through your failure history to build that list rather than instrumenting everything.
Do sensors need wiring to each machine?
Not always. Wireless condition sensors are practical for periodic and trend monitoring and avoid cable runs, at the cost of battery management and lower sample rates. Wired is better for continuous high-resolution monitoring on genuinely critical assets. We usually mix both within one plant.
What happens when it raises a false alarm?
It gets recorded as a false alarm and the threshold or feature is reviewed. This feedback loop is the part most implementations skip, and it is why so many condition monitoring systems are ignored within a year. We build the follow-up step into the workflow from the start.
Which asset keeps failing?
Tell us what has broken in the last two years and what it cost. That list is the right starting point, not a sensor catalogue.