Knowing what one part looks like is harder than it sounds.

The detection ladder from explicit signal to load signature, and the four situations where naive cycle detection produces confident, wrong numbers.

A monitoring system has to know what one part looks like. On a machine that reports a clean cycle-complete signal this is trivial. On a great many real machines it is not, and how a system behaves in those cases is the clearest available test of whether it was built by people who have stood on a shop floor.

This article covers how cycles are inferred when nothing clean is available, and — more usefully — the four situations where naive detection gets it wrong.

The detection ladder

Best available method first.

BESTExplicit signalThe machine reports a completed part. Nothing to infer; count the edge.
GOODCounter registerA PLC or control counter increments. Read the delta, and establish whether it includes rejects.
WORKABLEState transitionA run-to-idle-to-run pattern with consistent timing. Reliable on repetitive machines.
LAST RESORTLoad signatureCurrent or power draw showing a repeating pattern. A model, and must be labelled as an estimate.

Work down this ladder per machine, not per platform. A monitoring system that applies the last rung everywhere will produce counts that look confident and drift from reality; one that applies the first rung where it exists is simply reading a fact.

Failure mode 1 — the multi-part cycle

A program that produces four components in one machine cycle is common on bar-fed turning, multi-cavity moulding and any fixture holding several blanks. A detector counting cycles reports one part where four exist, and both output and performance are wrong by a factor of four.

The correction is a parts-per-cycle multiplier set per part number, not per machine, because the same machine runs different fixtures. If a system offers only a per-machine setting, it cannot handle a floor with mixed fixtures — which is most floors. Where the multiplier applies, it should be visible on the dashboard, so a supervisor can see that the count is derived rather than counted.

Failure mode 2 — dwell inside a cycle

Many legitimate cycles contain pauses: an operator gauging a feature mid-program, a probing routine, a tool waiting for coolant to clear, a controlled dwell for thermal stability. A naive detector treats each pause as the end of one cycle and the start of another, inflating the count and destroying the cycle-time distribution.

The handling is a minimum-cycle threshold and a pause tolerance, both set from that machine's observed behaviour rather than a default. This is one of several reasons a fortnight of data beats a demonstration: the tolerances are learned from what the machine actually does, and a machine with a 45-second cycle and a machine with a 40-minute cycle need entirely different values.

Failure mode 3 — the interrupted cycle

A cycle that starts, stops for chip clearing or a broken tool, and resumes is one part with an interruption, not two cycles. Counted as two, output is inflated and average cycle time collapses; counted as one long cycle with no interruption recorded, availability is overstated and the stop disappears from the Pareto.

The correct treatment records one part, one cycle, and one stop within it — which requires the detector to hold cycle context across an interruption rather than resetting on every state change. If you are evaluating systems, this is a specific and revealing question to ask, because it cannot be answered convincingly by a demonstration on a machine running cleanly.

Failure mode 4 — manual intervention and dry runs

A machine running in manual or jog mode is doing something, but it is not producing. A dry run with no material loaded looks identical to production from every signal except the part it did not make. Setup cycles, first-off inspection and test pieces all present the same problem.

Two things resolve most of it: mapping the mode selector so manual and setup are distinguishable from auto, and treating the job or route card selection at the machine as the authority on whether output counts against an order. Neither is exotic, and both are frequently skipped because they require agreement rather than software.

What to ask, and what to expect

ASK

Which rung of the ladder is this machine using?

A system that cannot answer per machine is applying one method everywhere, and is wrong somewhere.

ASK

How is a multi-part cycle handled?

Per part number, not per machine — otherwise mixed fixtures break the count.

ASK

What happens to an interrupted cycle?

One part, one cycle, one stop recorded inside it. Anything else distorts both output and availability.

EXPECT

Tuning over the first fortnight

Thresholds learned from the machine rather than set from defaults. A system that claims to be perfect on day one has not met your machines.

Questions

Straight answers.

How does a monitoring system detect machine cycles?
Down a ladder, best method first: an explicit part-complete signal, then a counter register, then a state-transition pattern, then a load signature. The right rung is chosen per machine, not applied platform-wide.
What happens with a program that makes four parts per cycle?
It needs a parts-per-cycle multiplier set per part number rather than per machine, because the same machine runs different fixtures. Without it, output and performance are both wrong by a factor of four.
How are pauses inside a cycle handled?
With a minimum-cycle threshold and a pause tolerance learned from that machine's behaviour. Without them, each gauging pause or probing routine registers as the end of a cycle and the count inflates.
What about a cycle interrupted for chip clearing?
It should record one part, one cycle and one stop inside it. Counting two cycles inflates output; ignoring the stop overstates availability. The detector has to hold cycle context across the interruption.
How do you tell production from a dry run or manual mode?
By mapping the mode selector so auto, manual and setup are distinguishable, and by treating job selection at the machine as the authority on whether output counts against an order.
Should cycle detection work perfectly from day one?
No, and a vendor claiming it does has not met your machines. Thresholds are tuned from observed behaviour over the first fortnight, which is why a month of data beats a demonstration.
FREE 2-MACHINE PILOT
Live OEE on two of your machines — this week. ₹0 to start.
Software-first setup in under an hour each · your data stays on-premises · plant-specific ROI model included.