OEE and the value of time

Aaron Blutstein 25 August 2026  

Overall Equipment Effectiveness (OEE) has traditionally helped manufacturers quantify where productive capacity is being lost, but the growing availability of machine, process and condition data is creating an opportunity to recognise some of the conditions associated with those losses earlier. Artificial intelligence may extend that capability, provided it is applied to an engineering problem rather than treated as an end in itself. Smart Machines & Factories reports.

Overall Equipment Effectiveness (OEE) has proved useful because it brings three familiar sources of production loss into a common measure. Availability reflects the proportion of planned production time during which equipment is actually running, performance accounts for losses associated with running below the ideal production rate, including slow cycles and small stops, and quality reflects the proportion of output produced correctly first time. Used properly, OEE therefore provides production and engineering teams with a consistent way of identifying where productive capacity is being lost, although the headline percentage is considerably less informative than the losses from which it is calculated.

One limitation is that OEE principally describes production performance rather than equipment health. By the time a four-hour breakdown appears as an availability loss, the maintenance department is already well aware that the machine has stopped, while repeated minor stops or a gradual reduction in production rate may have affected output for several shifts before the accumulated performance loss attracts attention. Real-time OEE reporting shortens the delay in seeing these effects, but it does not by itself establish the mechanical, electrical or process condition responsible for them.

Condition monitoring addresses a different part of the problem and has been doing so since long before the current interest in artificial intelligence. Vibration analysis, oil analysis, thermography, electrical measurements and process parameters can all provide evidence of developing faults, while established diagnostic and prognostic techniques can help engineers assess deterioration and, where sufficient information exists, estimate future asset condition. Predictive maintenance should not therefore be regarded as synonymous with AI, since the principle of using information about asset condition to inform maintenance decisions is already well established.

What has changed is the quantity and variety of information that can economically be collected from production equipment and the ability to analyse relationships within those data. PLCs, drives, machine controllers, condition monitoring systems, vision equipment and additional sensors can provide information about an asset while it is operating, allowing production behaviour to be examined alongside indicators of mechanical or process condition. A motor drawing progressively more current, for example, may be unremarkable in isolation, as might a small change in cycle-time consistency, but a relationship between several such changes may warrant investigation even though none has reached an established alarm level.

Looking beneath the OEE figure

Machine learning can be useful where relevant behaviour is represented by relationships between several variables or where normal operating conditions vary according to product, speed, load or process state. This does not make established condition monitoring obsolete, nor does it mean that an algorithm will necessarily identify a developing fault more reliably than an experienced analyst using an appropriate technique. Its potential lies in examining larger and more complex sets of operating data than would normally be practical for a person to monitor continuously, then highlighting patterns or changes that may merit engineering attention.

The connection with OEE becomes particularly interesting when deterioration affects production before it causes a breakdown. A packaging machine, for example, may develop a handling problem that causes increasingly frequent brief interruptions without suffering a prolonged stoppage. Operators clear the fault and restart the machine, with each interruption perhaps lasting only 20 or 30 seconds, yet the accumulated effect over a shift can become a significant performance loss. If the increasing frequency of those events coincides with changes in drive load, vibration or another relevant condition indicator, examining the information together may provide useful evidence of a developing problem and give engineers a reason to investigate it.

Quality losses can develop in a similar manner because a process does not necessarily move directly from producing acceptable components to producing rejects. Tool wear, temperature variation, pressure changes and other forms of process drift can move measured characteristics gradually towards specification limits while production remains acceptable. Where there is a repeatable relationship between that movement and machine or process data, statistical analysis or machine-learning techniques may help identify conditions associated with the drift before they result in scrap. Whether that approach is worthwhile will depend upon the process, the available measurements and the cost of the losses involved rather than simply upon the availability of an AI system.

Forecasting an OEE value is not, however, the same as predicting the physical conditions that will cause future production losses. OEE forecasts can be produced from historical production and machine data, and machine-learning techniques are among the methods that have been investigated for doing so. Such a forecast may indicate that availability, performance or quality is likely to deteriorate, but it does not necessarily explain the engineering reason for that deterioration. For a maintenance team, evidence about the developing condition of an asset may therefore be considerably more useful than knowing only that a future OEE figure is expected to be lower.

There are also practical limitations that become particularly apparent in factories containing equipment of different ages and from different suppliers. Machine learning depends upon appropriate data, and no analytical method can compensate for unreliable sensors, poorly defined production states or inaccurate production parameters. Ideal cycle time, for example, should represent the theoretical fastest rate at which the process can produce rather than an easily achievable target or budget rate. Setting it incorrectly can distort the performance component of OEE before any more advanced analysis has taken place.

The same caution applies to downtime classification and operating context. A data set showing that a machine has stopped does not necessarily establish whether the cause lies with the machine, an upstream process, unavailable material, an operator intervention or a downstream blockage. Relationships discovered in historical data can also be misleading when changes in product mix, production schedule or operating practice are not represented properly, which means that engineering interpretation remains at least as important as the sophistication of the analytical method.

For that reason, an OEE loss analysis provides a sensible starting point for deciding where additional condition or process information might have value. Where one asset repeatedly constrains a production line through availability losses, additional monitoring may help engineers understand the failure mechanisms and recognise signs of deterioration earlier. Another asset may rarely suffer a significant breakdown but consistently lose output through reduced speed or minor stops, in which case production states, cycle behaviour, drive loads or process variables could be more informative than measurements aimed solely at assessing component condition.

Earlier evidence for engineering decisions

Experienced operators and maintenance engineers already carry out a considerable amount of pattern recognition, often noticing changes in sound, vibration, temperature, product behaviour or machine response that are difficult to reduce to a single alarm value. Their accumulated knowledge may include an understanding that a particular machine becomes troublesome after extended operation, that a certain combination of apparently unrelated symptoms has previously preceded a stoppage or that a process behaves differently with one product variant. Analytical systems are most valuable when they complement this knowledge by examining information that cannot reasonably be observed continuously, rather than attempting to replace engineering judgement with a numerical prediction.

The way an analytical result is presented consequently matters almost as much as the calculation behind it. A statement that a machine has a particular percentage probability of failing within a specified period can create an impression of precision without giving the maintenance team enough information to decide what should be inspected. Evidence showing that vibration behaviour has changed over several shifts while motor load has increased and cycle-time variation has become greater provides a more useful basis for investigation, because the engineer can compare those observations with the machine’s physical condition, operating history and known failure modes before deciding whether intervention is justified.

Any investment in this area also needs to be judged against production outcomes rather than the quantity of data collected. Condition monitoring, machine learning and additional sensing have a stronger business case where their use can be associated with fewer unplanned stops, reduced performance losses, more stable quality or better use of planned maintenance opportunities. The current UK approach to industrial AI reflects the same practical difficulty, with manufacturers facing issues around legacy equipment, fragmented data, integration, assurance and demonstrating value as they attempt to move applications from trials into routine production.

For many plants, the useful development will consequently be less dramatic than the suggestion that forecasting OEE will somehow reveal what is happening inside the equipment. OEE can continue to show where productive capacity has actually been lost, while forecasts may indicate how that performance is likely to develop and condition, process and production data can provide evidence about the reasons behind it. If engineers approaching a planned shutdown on Friday can see on Tuesday that a critical machine is behaving differently from its established operating pattern, they still have to determine what the data mean and whether intervention is justified, but they have gained something maintenance departments rarely complain about having too much of: time.