Table of Contents
Machine learning improves OEE only by reducing the losses behind it. Split OEE into availability, performance and quality, find the largest losses on the bottleneck, and apply the matching model: failure prediction for breakdowns, stop-pattern analysis for minor stops, predictive quality for scrap. Before any of that, the loss data itself must be trustworthy.
Key takeaways
- OEE is the product of availability, performance and quality, so each factor maps to a different set of losses and a different kind of model (OPC Foundation).
- Fix the data first. In a study of 884 machines at 23 companies, almost half of the recorded OEE losses could not be classified.
- Work on the bottleneck, and expect it to move. On an engine line, the previous run's bottleneck was the next run's bottleneck only 20.8% of the time (Subramaniyan et al., 2018).
- A run/stop signal alone predicts little. A random forest on one packing machine's state log caught only 21% of breakdowns in the following hour (Filios et al., 2022).
- Evidence on maintenance strategy is mostly survey-based: a NIST survey links less reactive maintenance to 52.7% less unplanned downtime, and its authors call the comparison anecdotal (Thomas and Weiss, 2021).
How is OEE calculated?
OEE (overall equipment effectiveness) multiplies three ratios. Vorne's reference site gives the common form (OEE.com):
- Availability = run time ÷ planned production time. Stops, planned or not, reduce it.
- Performance = (ideal cycle time × total count) ÷ run time. It should never exceed 100%.
- Quality = good count ÷ total count.
In Vorne's worked example, 88.81% availability, 86.11% performance and 97.80% quality give an OEE of 74.79%. Because the factors multiply, 90% on each gives about 73%, not 90%.
ISO 22400-2 uses the same structure with different names: availability, effectiveness and quality rate. The OPC UA specification for machine tools takes its KPI definitions from ISO 22400-2 and notes that planned busy time and planned run time per item "usually come from an MES or ERP and not from the machine" (OPC Foundation). ISO 22400 defines a set of 34 KPIs for manufacturing operations management, OEE among them (Kang et al., 2016).
What is a good score? The 85% "world class" figure comes from Seiichi Nakajima, who introduced TPM, OEE and the six big losses in the early 1970s, and it reflects Japanese automotive manufacturing of that time. Vorne says most manufacturers are closer to 60% (OEE.com). Compare a line with its own baseline, not with a benchmark.
Which losses can machine learning reduce?
The six big losses split OEE into causes, two per factor (OEE.com). Each needs different data and a different model.
| Loss | OEE factor | What machine learning can do | Data needed |
|---|---|---|---|
| Equipment failure | Availability | Detect abnormal behaviour and predict failures early enough to plan the repair | Condition signals (vibration, temperature, current), alarms, failure history from the CMMS |
| Setup and adjustments | Availability | Predict changeover duration for each product sequence, to support scheduling | Changeover logs with start, end, product from and product to |
| Idling and minor stops | Performance | Find patterns behind frequent short stops; forecast stop time per shift | Machine state log with timestamps and reason codes, upstream and downstream states |
| Reduced speed | Performance | Estimate the achievable speed per product and material; flag slow running | Cycle times, setpoints, material lots, ideal cycle time per product |
| Process defects | Quality | Predict which units or batches will fail inspection | Process data linked to each unit and its inspection result |
| Reduced yield | Quality | Predict start-up scrap and recommend start-up settings | Start-up events, settings and scrap counts per start-up |
Not every loss needs a model. Many minor stops come from jams, misaligned sensors and blocked product flow, which an engineer fixes once they are counted. Machine learning earns its place when causes interact, when signals arrive before the loss, or when there are too many machines and products to analyse by hand.
For failures, our guide to anomaly detection methods compares the options. For scrap and rework, see predictive quality.
Why fix the OEE data before modelling?
A model learns from the loss records it is given. Hedman, Subramaniyan and Almström analysed automatic OEE data from 884 machines at 23 companies. The average OEE was 65%, and almost half of the recorded losses could not be classified because loss categories were missing or poorly described. Of the stop time that was classified, 90% related to supporting activities performed by operators, not to the automatic process itself (Hedman et al., 2016).
Unclassified stops are invisible to a model. Before training anything:
- Capture states automatically. Read run, stop and fault states from the PLC rather than from shift sheets. PackML, now also available as an OPC UA companion specification, defines a standard machine state model and tag set that includes OEE data (OPC Foundation).
- Define the minor-stop threshold. Vorne describes minor stops as typically a minute or two, and the cut-off between a minor stop and a breakdown as a policy decision, such as requiring a reason code for any stop over two minutes (OEE.com).
- Check the ideal cycle time for every product. If it is too slow, performance looks good and speed losses disappear from the data.
- Clean the reason codes. Short, mutually exclusive categories that operators can pick in seconds.
Operators' free-text comments and maintenance work orders also hold causes. NIST researchers argue that out-of-the-box language processing pipelines need re-imagining for this kind of technical text (Brundage et al., 2021). Our post on root cause analysis with an LLM over maintenance history covers that side.
Why start with the bottleneck?
A line's throughput is constrained by one or more bottleneck machines, and prioritising maintenance and improvement work on them substantially improves the throughput of the whole system (Subramaniyan et al., 2018). Raising OEE on a machine that waits for its neighbours does not raise output.
The bottleneck also moves. Subramaniyan et al. studied MES data from an automotive engine line in Sweden with 13 machines, over 215 consecutive production runs of 17 hours each. The bottleneck of one run was the true bottleneck of the next run only 20.8% of the time. Their algorithm forecast each machine's active periods with ARIMA and predicted the next run's bottlenecks with a recall of 62.53%, against 24.69% for the naive method of assuming last run's bottleneck. Not perfect, but enough to plan maintenance and support staff on the right machines.
This is a cheap first use of machine data. It needs only machine states and timestamps, which you have once state capture is automatic.
Can machine learning predict stops before they happen?
Sometimes, with the right inputs. Filios et al. logged the run/stop state of one packing machine at a food and beverage plant from October 2018 to October 2019: 73,576 records, including more than 32,000 minor stops of 10 seconds to 5 minutes, 2,545 breakdowns of 5 to 40 minutes and 289 longer stops (Filios et al., 2022).
Their results show both the promise and the limit:
- Forecasts of the next day's total minor-stop time had a mean absolute percentage error of about 20%.
- A random forest predicting a breakdown in the next hour reached 72% accuracy, but caught only 21% of breakdowns, and about half of its alarms were false.
- The authors note that the absence of the stop causes weakened the models, since only the run/stop signal was available.
Accuracy misleads when events are rare, as our predictive quality guide explains. Add context before adding model complexity: reason codes, alarms, upstream and downstream states, process variables and the product being run. Rebuilding failure labels from weak records is covered in labelling failures in time-series data.
For breakdowns, the economic case rests largely on surveys. In a NIST survey of US manufacturers with 71 usable responses, the half relying less on reactive maintenance reported 52.7% less unplanned downtime and 78.5% fewer defects. Among the rest, those relying more on predictive than preventive maintenance reported 18.5% less unplanned downtime, but direct maintenance costs relative to shipments 212.3% higher. The authors say these comparisons "should be treated as anecdotal evidence" because the groups are small (Thomas and Weiss, 2021).
How do you know the model improved OEE?
- Split by time. Ordinary cross-validation on time-ordered data "would lead to training on future data and evaluating on past data" (scikit-learn). Train on older months and test on a later period.
- Beat a naive baseline, such as last run's bottleneck or last week's stop rate, as Subramaniyan et al. did.
- Count what operators feel: stops anticipated, false alerts per shift, and hours of downtime avoided.
- Report availability, performance and quality separately. A gain in one can hide a loss in another.
- Freeze the definitions. Reclassifying a stop as planned raises OEE without any real change on the line.
Step by step: a first OEE project
- Choose one line and identify its bottleneck from machine states.
- Automate data capture: states, counts and reason codes from the PLC, through OPC UA or the MES. See OPC UA for data scientists.
- Fix the definitions: planned production time, ideal cycle time per product and the minor-stop threshold.
- Build a loss Pareto per OEE factor over several months that cover the product mix.
- Pick the largest loss that engineers cannot already explain, and choose the model from the table above.
- Validate on a later period against a naive baseline.
- Run in shadow mode for several weeks, then bring the outputs into shift meetings and maintenance planning.
- Name an owner, monitor the model for drift, and plan the move from pilot to plant-wide rollout before extending to other lines.
What we have seen at KABAS AI
Our manufacturing page lists real-time production monitoring, automated shift reports and KPI tracking, and quality trend analysis with anomaly detection among the capabilities of DALILE. It lists predictive maintenance scheduling, capacity planning and bottleneck analysis, and parameter optimisation among those of Virtual Lab. Both run on the customer's own infrastructure and can work offline. Our data science team covers data collection and preparation, trend detection, predictive models and the MLOps pipelines that keep them monitored.
Our practical advice follows the evidence above. The first deliverable of an OEE project is a loss Pareto that production, maintenance and quality all trust, not a model. Agree on definitions and reason codes before modelling, start on the bottleneck, and keep models advisory: they inform maintenance and planning decisions, and people act through existing procedures. Keep stop histories and production rates on your own servers, since they describe your capacity and costs. Before you start, check the data with an AI readiness assessment, and see MES, ERP and AI for where OEE data usually lives. Contact us to discuss your line.
Sources
- Vorne Industries, Calculating OEE, OEE.com
- Vorne Industries, The Six Big Losses, OEE.com
- Vorne Industries, World-Class OEE, OEE.com
- OPC Foundation, OPC 40501-1 OPC UA for Machine Tools Part 1, Annex C.2 KPI Calculation
- OPC Foundation, OPC 30050 PackML, 4.2 PackML Overview
- Kang, Zhao, Li and Horst, A Hierarchical Structure of Key Performance Indicators for Operation Improvement in Production Systems, International Journal of Production Research 54(21), 2016 (NIST record)
- Hedman, Subramaniyan and Almström, Analysis of Critical Factors for Automatic Measurement of OEE, Procedia CIRP 57, 2016
- Subramaniyan et al., A data-driven algorithm to predict throughput bottlenecks in a production system based on active periods of the machines, Computers & Industrial Engineering 125, 2018
- Filios et al., Agnostic Learning for Packing Machine Stoppage Prediction in Smart Factories, ITU Journal on Future and Evolving Technologies 3(3), 2022 (arXiv 2212.06288)
- Thomas and Weiss, Maintenance Costs and Advanced Maintenance Techniques in Manufacturing Machinery: Survey and Analysis, International Journal of Prognostics and Health Management 12(1), 2021 (NIST record)
- Brundage et al., Technical Language Processing: Unlocking Maintenance Knowledge, Manufacturing Letters 27, 2021 (NIST record)
- scikit-learn, TimeSeriesSplit
Frequently asked questions
What is a good OEE score?
An OEE of 85% is often called world class. The figure comes from Seiichi Nakajima's work on TPM in 1970s Japanese discrete manufacturing, and Vorne, an OEE equipment vendor, says most manufacturers are closer to 60%. Benchmarks from other industries say little about your line, so compare each line with its own baseline and track the trend.
Do we need an MES to measure OEE?
Not necessarily, but you need automatic data. The machine can supply run and stop states, run time and part counts, while planned production time and the ideal cycle time usually come from an MES, an ERP or the production plan. Machine learning needs each stop with its start time, duration and reason, which totals written by hand at the end of a shift do not provide.
Can machine learning increase OEE directly?
No. OEE is a result, not a control variable. A model predicts or explains one loss, such as a breakdown, a run of minor stops or a batch of scrap, and OEE only improves when someone acts on that prediction through maintenance, planning or process changes.
How much data do we need to start?
There is no universal number. The history should cover the full product mix and contain enough events of the loss you want to predict. Minor stops are plentiful, so a few months may be enough to study them, while breakdowns are rarer and often need a year or more of history linked to maintenance records.
Should we calculate OEE for every machine?
Measure it everywhere, but improve the bottleneck first, because it limits the output of the whole line. Raising OEE on a machine that waits for its neighbours does not increase output. Bottlenecks can also move from one production run to the next, so check where the constraint is regularly.
Can these models run on-premise?
Yes. The usual models for OEE work, such as gradient boosted trees, survival models and time-series forecasts on machine states, are small and run on ordinary servers inside the plant network. Keeping them on-premise also keeps production rates, stop history and quality data inside the company.



