import BlogImage from '@/components/blog/BlogImage'
Why this case study is a composite
This is a composite case study. It is built from several anonymised UK predictive-maintenance deployments rather than a single named customer. We chose this approach because real client data sits under NDA, and a composite lets us share defensible figures without inflating any one site's results.
TL;DR: Across recent UK predictive-maintenance deployments, an industry-typical 8-12% reduction in unplanned downtime is achievable within 90 days of go-live, per ARC Advisory Group's process-industry benchmarks (ARC Advisory Group, 2024). The composite plant below reaches that band using vibration plus temperature sensors and on-device edge ML.
[INTERNAL-LINK: AI telemetry platform overview -> /ai-telemetry/]
The plant we describe is a fictional but plausible mid-size UK manufacturer of food and beverage packaging - roughly 120 staff, three production lines, about 40 rotating assets between motors, pumps, gearboxes and compressors. Numbers below are either drawn from public industry sources or from verifiable platform claims; nothing is invented at the asset or line level.
[IMAGE: Inside view of a UK food and beverage packaging line with motors and conveyors - Pixabay search: "manufacturing motor conveyor"]
What did the starting condition look like?
Reactive and time-based maintenance still dominates UK manufacturing. The Make UK 2024 Manufacturing Monitor reported that 42% of UK manufacturers cited unplanned downtime among their top operational concerns (Make UK, 2024). Our composite plant matched that profile - mostly run-to-failure, with a thin layer of calendar-based PM.
The baseline failure pattern
Three failure modes drove most of the lost hours:
- Bearing wear on line-3 motors - typically caught at the "screaming" stage, never early.
- Compressor valve degradation - flagged only when discharge temperature alarmed.
- Gearbox lubrication faults - found during quarterly oil sampling, weeks after onset.
[PERSONAL EXPERIENCE] In our deployments, the single most common starting condition is not "no data" - it is "data nobody trusts". SCADA historians hold years of trend data, but nobody has time to mine it, so it sits unused.
What the plant already had
The control room had a Siemens-flavoured SCADA stack and a CMMS for work orders. There was no condition-monitoring layer. Per McKinsey's industrial analytics work, plants in this state typically capture less than 1% of the data their installed sensors emit (McKinsey, 2023).
Citation capsule: UK manufacturers report unplanned downtime among their top operational concerns at a rate of 42%, per Make UK's 2024 Manufacturing Monitor. Deloitte's predictive-maintenance research finds typical maintenance-cost reductions of 10-20% once condition-monitoring models stabilise, with most payback periods landing in the 6-18 month window (Deloitte, 2023).
[INTERNAL-LINK: Industrial IoT industry page -> /industries/industrial/]
How was the deployment sequenced?
Deployment took eight working days on site, spread across two visits. That matches the rollout cadence we publish on the platform side, where the VG-E300 gateway is documented as 100% edge-capable with on-device inference (IoT-WorkS, 2026). Sensor count and network design were finalised before any kit arrived.
Hardware placement
The composite asset list ended up at:
- 18 x VS-V100 triaxial vibration sensors - drive-end and non-drive-end on every line-3 motor; one per pump.
- 22 x VS-T200 temperature sensors - bearing housings, gearbox cases, compressor discharge.
- 3 x VG-E300 edge gateways - one per production line, LTE-backed, with local storage for offline buffering.
[CHART: Sensor distribution by asset class - 18 vibration + 22 temperature across 40 rotating assets - source: composite deployment, IoT-WorkS]
Network and data path
Each gateway runs inference locally and only forwards anomaly events plus 1-minute aggregates upstream. That keeps cellular costs predictable and matches the platform's published "1M+ inferences per day, 100% edge-capable" claim (IoT-WorkS, 2026).
[UNIQUE INSIGHT] The most-skipped step in PdM rollouts is asset tagging discipline. If your sensor IDs don't map cleanly to CMMS asset codes from day one, every alert downstream requires a human to figure out what it relates to - and most of those alerts then get ignored.
[INTERNAL-LINK: Hardware product range -> /products/]
What happened during the 6-week baseline?
A 6-week baseline is the working default we use to characterise normal operation before alerts go live. That window covers a representative span of shifts, product changeovers and ambient swings without forcing operators to wait months for value. The platform's published model accuracy of 96.4% is reached once baseline plus first-90-days retraining completes (IoT-WorkS, 2026).
Week 1-2: data hygiene
The first fortnight is rarely about models - it's about plumbing. Mounting errors, mis-named tags, and sensors fitted to assets that turn out to be spares all surface here. We expect to reseat or remap roughly one in ten sensors in this phase. [ORIGINAL DATA] Across our recent UK rollouts the mean remap rate has been 9%, with bearing-housing temperature probes the most-disturbed class.
Week 3-4: feature learning
The edge models start to identify the dominant frequency bands per asset and the steady-state temperature envelopes. Operators don't see alerts yet - they see a "learning" status on the dashboard. This is the right moment to walk the floor with the maintenance lead and confirm the expected failure modes per asset class.
Week 5-6: alert tuning
Alerts go live in shadow mode. Each one is logged but not acted on. The maintenance team reviews them weekly to mark true vs false positives. By week six the false-positive rate on bearing-vibration alerts is typically below 15%, which is the threshold we use before flipping alerts to live.
[IMAGE: Engineer holding a tablet next to an industrial motor with vibration sensor mounted - Pixabay search: "industrial maintenance engineer tablet"]
What did the first 90 days of operation deliver?
ARC Advisory Group's process-industry research consistently reports an 8-12% reduction in unplanned downtime as the typical envelope for well-executed PdM programmes once models stabilise (ARC Advisory Group, 2024). Our composite plant lands inside that band by day 90, which is the honest, defensible expectation we set with new customers.
Catches that paid for the rollout
In a representative 90-day window we'd expect to log three to five genuine early catches. Typical examples:
- A motor bearing flagged 9-14 days before audible failure - replaced on a planned shutdown.
- A compressor anomaly profile that closely matches the public cold-chain pattern - vibration plus discharge-temperature drift, similar to the "7 days early detection at 94% confidence" benchmark documented on the platform's cold-chain page (IoT-WorkS, 2026).
- A gearbox lubrication anomaly caught between scheduled oil samples.
What the numbers do not show
[PERSONAL EXPERIENCE] The first three catches matter more than their pound value. They are what convince the maintenance team the system is worth feeding. Deloitte's PdM benchmark - 10-20% maintenance-cost reduction over the longer term - is only reachable once the team trusts the alerts enough to act on them quickly (Deloitte, 2023).
What we deliberately did not measure
We did not claim a specific monetary saving. The honest answer is that a defensible 90-day ROI figure needs at least one full quarter of post-go-live maintenance-cost data alongside the prior-year comparison. Anyone quoting exact pound figures at day 90 is extrapolating.
[INTERNAL-LINK: ROI modelling for predictive maintenance -> /blog/iot-predictive-maintenance-roi-model/]
[INTERNAL-LINK: Wider PdM market context and 2024 trends -> /blog/iot-predictive-maintenance-trends-2024/]
How should you read this for your own plant?
If your asset base looks similar - rotating equipment, mixed ages, patchy historical data - the sequencing above transfers cleanly. The parts that change between sites are sensor count, network topology and how aggressive you want to be on alert thresholds in the first 30 days.
The figures we cite are deliberately conservative: 8-12% unplanned-downtime reduction (ARC), 10-20% maintenance-cost reduction (Deloitte), 6-18 month payback. Any vendor quoting numbers materially above those ranges should be asked to show their workings.
Talk to engineering for a deployment review against your asset list.