Clamp-On OEE: Measuring Machine Utilization Without Machine Integration

01The measurement gap

Ask a plant manager for the utilization of a given machine and the answer usually arrives as a range with a shrug attached: “seventy, maybe seventy-five percent.” The number that would replace the shrug — run time divided by available time, measured continuously — requires a data source on the machine. For a machine built in the last decade that means a PLC integration: mapping tags, standing up an OPC-UA server, and threading the data through the plant network. For a machine built before that, it means retrofitting a source that does not exist.

Both paths are real projects with real owners and real backlogs. The result is that the machines most in need of measurement — older assets, mixed brands, the lathe in the corner that runs one job a week — are the least likely to get it. The measurement gap is not a technology gap. It is a prioritisation gap, and it is closed by removing the integration, not by scheduling it.

02What a clamp-on sensor can observe

A machine that is doing work vibrates, and it moves. Those two facts are enough to distinguish work from no-work without any electrical connection to the machine. The sensor used here is a coin-cell Bluetooth Low Energy tag — the reference implementation uses a RuuviTag broadcasting the RAWv2 (data format 5) manufacturer payload¹ — magnet-mounted to the spindle housing, motor, or frame.

Each advertisement carries, among other fields:

• Three-axis acceleration, in milli-g, sampled by the tag’s onboard accelerometer. At rest this reads roughly 1 g on the gravity axis and near zero on the others; under machining load it develops broadband noise on all three.
• A motion counter — an 8-bit value the accelerometer’s activity-detection interrupt increments on each detected jolt. Its rate of change is a robust, low-power proxy for “something is happening,” and it survives the tag’s duty-cycled advertising better than any single acceleration sample.
• Temperature, relative humidity, and battery voltage — environmental context, and, on a machine, a slow secondary signal (a spindle motor warms under load).

The tag advertises roughly once per second and is passive: it holds no connection, so any number of gateways can hear it, and losing one gateway loses no data. A fixed BLE gateway — a small single-board computer, or a phone in kiosk mode for a pilot — scans continuously and does the interpretation.

03Classification: the state machine

The gateway maintains a short rolling window of advertisements per tag (30 seconds in the reference build) and, on each evaluation, decodes that window into motion statistics and derives a state.

The running test

An asset is running when either condition holds over the window:

The counter difference is taken modulo 256 to absorb the 8-bit wrap. Thresholds are per-asset: a surface grinder and a stamping press live at different vibration floors, and the pilot sets each from a few hours of observed baseline.

Idle, off, and hysteresis

The output is a stream of closed state intervals — {asset, state, start, end} — plus one open interval re-sent on every flush so the dashboard always shows the current state. Every interval id is deterministic, so a re-sent flush is idempotent and the server never re-aggregates.

04From state intervals to OEE

State intervals are accumulated into hourly buckets. Each bucket holds the seconds spent in each state, a cycle count where one is available, and per-metric summaries (max and mean of temperature, vibration RMS, battery). From the bucket:

This is deliberately conservative about what it claims to have measured:

A utilization figure that is correct is more useful than an OEE figure that is complete but partly guessed. The dashboard shows the OEE number only when its inputs are real, and labels the rest as availability.

05Data architecture

The backend is BiseraDB, Onshore’s document data engine. Two design choices keep it fast and cleanly multi-tenant:

A namespace per site

Each customer site gets its own namespace (mfg_<site>). No collection ever holds more than one shop’s data, so a query is scoped by construction and a tenant can be exported or deleted as a unit.

Time-bucketed documents

Readings are not stored one row per sample. One document holds one sensor’s data for one hour — the minute-level samples as an array, the hour’s state totals and rollup as fields, keyed <sensor>:<YYYYMMDDHH>. A year of a 25-machine shop is on the order of tens of thousands of documents, not millions of rows.

Every dashboard read is then a single indexed equality plus a sort and a limit — “this sensor’s last 168 hourly documents, newest first.” On a real collection an indexed query of this shape returns in roughly 13 ms; the same query without the index, as a full scan, takes ten to seventeen seconds. The bucketing is what makes the difference structural rather than incidental.

Ingest is a single keyed HTTPS POST of already-aggregated hour documents, state intervals, and downtime events. There is no inbound connection to the plant and no long-lived socket; a gateway behind a restrictive firewall or on a cellular link works unchanged.

06A reference deployment

The method runs continuously on Onshore’s own infrastructure. A RuuviTag is mounted in the server room; a gateway on a nearby machine scans it, classifies, buckets, and flushes to the production ingest endpoint every 60 seconds; the data lands in a dedicated namespace and is visible on the same dashboard a customer would use.

• Setup: the tag was activated from the vendor’s phone app and picked up by the gateway with zero configuration on any monitored equipment. Time to first data point on the dashboard: about ten minutes, most of it walking.
• Steady state: a stationary tag classifies as idle when seen and off when out of range — correct behaviour. The value from a fixed sensor is the environmental trend (temperature, humidity, battery), not a machine state; the same pipeline serves both.
• Resilience: the gateway stores and forwards. A dropped uplink backfills the missed hours when it recovers, and — per the gap-attribution rule — the outage is charged to off, not to whatever state was last seen.

Nothing in the path is specific to the reference site. A pilot swaps the tag onto a real machine, sets that machine’s thresholds from a short baseline, and points the gateway at the same endpoint.

07Limits of the method

An honest account of where a clamp-on sensor stops being enough:

• It sees that a machine stopped, not why. Vibration cannot distinguish a planned tool change from a jam from an operator break. Downtime reasons come from a person — a tablet at the machine, or an AI proposal a person confirms — not from the sensor.
• Quality is out of reach without a count source. Scrap and rework require a data feed the sensor does not provide. Until one exists, the method reports availability and (where a cycle target is set) performance, and is explicit that the OEE quality term is assumed.
• Placement and range matter. A tag mounted where the running signal is weak, or a gateway too far from the tag, degrades classification. Both are set during the pilot from observed signal strength and baseline vibration; neither is a field guess.
• It is Tier-1 monitoring, by design. The sensor observes. It never commands, moves, or overrides a machine or robot controller. Motion, trajectory, and safety stay entirely with the certified control system. For robots this is the whole scope: state, cycles, faults, and e-stop, read from the controller’s own status port, one pane of glass across a mixed fleet — and nothing more.
• Battery is a maintenance item. A coin cell in a duty-cycled tag lasts on the order of a year or more; the tag reports its own voltage, and a low-battery alert is part of the dashboard, but it is a consumable.

08Conclusion

The reason most floors estimate utilization instead of measuring it is not that the measurement is hard — it is that the usual way in costs an integration project per machine. Reading the machine from the outside removes that cost. A thirty-dollar sensor, a state machine with sensible hysteresis, and a storage model that keeps every query cheap produce a utilization number that is defensible, an OEE number where the inputs are real, and a mixed-brand, mixed-age floor on a single screen — in an afternoon rather than a fiscal quarter.

The method is deliberately modest about its claims. It measures availability directly, measures performance where a cycle target exists, and does not pretend to measure quality it cannot see. Within those bounds it turns the machines that were invisible into a signal — which is the part that was actually missing.