OEE is useful only when it remains a loss tree

The NIST maintenance report presents OEE as availability multiplied by performance rate and quality rate. Its definitions relate downtime to required availability, design cycle time and output to operating time, and quality defects to production input. These are different denominators, so the components cannot be added or averaged to obtain OEE.[S02] For this chapter, the practical question is whether the production records can explain every step from planned opportunity to accepted output.

GFT proposes a local notation before any dashboard is configured. Let T be planned production minutes, D be counted downtime minutes, R be running minutes after subtracting D, c be the approved ideal cycle in minutes per unit, N be total completed units and G be units accepted without rework at the declared checkpoint. These variables describe a single operation and a single product basis. They are not attended labour minutes, stitch counts or a sum of unrelated machine capacities.

This convention makes an important distinction. The final ratio measures effective output against opportunity, while the components explain the route to that result. A low score alone cannot reveal whether the loss arose because the machine stopped, the operation ran slowly or completed work failed inspection. The loss tree therefore needs both a reconciled total and a credible allocation. Recalculating the total is an arithmetic check; validating the allocation requires observation.

For a proposed garment application, define the output event at an operation that produces a countable result, such as a completed seam operation or an inspected finished garment. Do not call a machine stitch a garment. For parallel machines, identify whether the denominator is machine minutes or elapsed line minutes. When the boundary is the whole line, use an approved line output model. Multiplying machine scores does not create a defensible line score because the same garment can pass through those machines.

The proposed architecture in Exhibit 6 connects a component to its evidence and decision. It adapts NIST's loss categories into a supervisory method. The layer labels are recommendations, not a claim that GFT has implemented this design.[S02]

Read the architecture from the observation towards the decision as well as from the score downwards. A supervisor who sees a persistent speed residual should first ask whether counts, time and cycle basis agree. Only then should the residual become a work-method investigation. This order avoids presenting a calculation error as an operator problem. It also allows a paper log and a later digital system to use the same definitions.

Retain unavailable or invalid results explicitly. If planned time is absent, the OEE denominator is unresolved. If running time or completed output is absent, some components cannot be divided meaningfully. Report the observed stop and output state with the calculation limitation rather than allowing spreadsheet defaults to invent a result. The purpose is an inspectable account of production, including periods when the component calculation is not defined.

GFT proposes retaining a compact calculation packet for every reporting period. It should contain the approved schedule, the cycle-standard version, the event intervals, the completed-output ledger, the acceptance ledger and the calculation result. A reviewer should be able to reconstruct the score without interviewing its original author. A dashboard screenshot cannot replace this packet because it omits the assumptions that determine the result. Where records disagree, show the discrepancy before approving the period for comparison.

The packet should also distinguish machine time from human work. An operator may inspect, prepare or handle a component while the machine is not cycling. Under the proposed method, that observation prompts a check of the cycle boundary and work content; it does not automatically become avoidable downtime. A machine signal needs an operational definition before it can describe a manual or mixed activity. This preserves the distinction between observing electrical activity and measuring effective production.

Research by Sonmez and colleagues explicitly identifies uncertainty in OEE components when production speeds and stoppage durations are captured manually or semi-automatically. Its abstract also discusses ignored minor stops, idling and speed losses.[S07] GFT's practical implication is to treat performance as a measured residual that needs investigation, rather than automatic proof of insufficient effort.

Under the proposed convention, running time still contains brief interruptions that the protocol leaves below the separately timed stop threshold. Performance compares the ideal time represented by completed work with that running interval. The threshold therefore changes the division between availability and performance. It should be selected for reliable observation and useful response, recorded in the protocol and kept stable through the comparison. It is not a device for improving either component.

Consider an illustrative operation with 450 planned minutes, 45 minutes of recorded downtime, 300 completed units and an assumed ideal cycle of 1 minute per unit. Running time is 405 minutes and ideal production time is 300 minutes. Performance is 300 / 405, approximately 74.07 per cent. Its residual is 105 minutes. That residual could contain brief stops, slow handling, count error or an unsuitable cycle assumption; the arithmetic alone cannot divide it among those explanations.

In the same illustrative case, assume observation identifies 15 minutes of brief stops within that residual. Moving those minutes into downtime gives running time of 390 minutes. Availability changes from 90.00 to approximately 86.67 per cent, while performance changes from approximately 74.07 to 76.92 per cent. Their product remains 300 / 450, approximately 66.67 per cent. A changed component score has not recovered any output. Preserve both the old and revised classification when definitions change.

Do not subtract the same brief stop twice. If it has already reduced the running-time denominator, it must not also be deducted as a separate performance-minute loss. Conversely, if it remains in running time, the performance residual already includes its effect. The event register may tag the interruption in either treatment, but the reconciliation must identify which branch carries its duration. The source log should be richer than the additive loss account without making the account non-additive.

GFT recommends an observation check before modifying the ideal cycle. Confirm the product, operation, machine, attachment, method and output unit. A cycle based on a different seam length or different handling requirement is a different comparison. An apparent result above the approved maximum should trigger a count, clock and standard review. Silently capping the result would hide the discrepancy; quietly loosening the standard would reward it. The next chapter treats the relationship with standard minutes in detail.

For the proposed capture method, make the threshold observable with the tools actually available. A rule requiring precise timing that a paper log cannot support will create apparent precision rather than better decisions. Trial the rule against direct observation and ask whether different observers classify the same event consistently. If they do not, simplify the state definitions or improve capture before attaching consequences to the component score. The measurement-uncertainty research supports exposing these limitations rather than concealing them in a point estimate.[S07]

Treat missing short-stop causes differently from missing short-stop time. If output, cycle basis and running time are trustworthy, the performance residual can be calculated even while its causes remain unknown. If the clock or count is unreliable, the residual itself is uncertain. GFT recommends marking these situations separately: the first needs diagnostic observation, while the second needs record repair. Combining them in an uncoded loss bucket would obscure which kind of work could make the account more useful.

Use short observation samples to test the residual's composition, with their limits recorded. A sample of stable running does not establish how often interruptions occur across a shift. A sample taken during a difficult change does not describe normal method speed. GFT proposes retaining the observation conditions alongside the estimate and reporting an unresolved balance where the sample cannot explain all lost time. False precision in cause allocation is less useful than an honest residual with a named follow-up.