OEE Anomaly Detection through Machine Learning

Machine data is one of the fastest growing areas of so-called “big data” in companies – and parts of this machine data can be used to calculate the OEE. The most valuable part of the data is created if it does not meet expectations, if “as-is” is not equal to “to-be”. Then the intervention of the management is expected in a well-managed organization. So how nice would it be if these anomalies in the large amounts of data were automatically recognized and presented to humans for examination? Artificial intelligence or more specifically machine learning is about to crack that nut.

An Anomaly is something Special

If data are well prepared for human processing, deviations from one’s own expectation can be easily recognized. For example, poorly prepared data for humans are data in a table format. Well prepared, they are in a – cleverly chosen – graphics, which is why no spreadsheet program waived graphics functionality. In a graph, a person can easily detect changes in the slope or a deviation from the target value.

Machines deliver their data to a database. That is the worst possible option for humans. Not only that he often has no access to the sensor data. It is also a poorly comprehensible table format. Recognizing an anomaly, i.e. a deviation from an expected behavior, pattern or structure, is an impossibility for man.

Even if the data is rendered graphically, and thus the human – with its understanding of the expected behavior of the time series – can recognize the anomaly, we have better uses of human intelligence than for round-the-clock monitoring of time series data.

OEE Anomaly Detection

Anomaly detection of the OEE data stream can occur on two levels: at the level of the OEE itself and at the levels of the individual data streams for availability, performance and quality.

As an example of anomalies at the level of the OEE, the following measurement can serve.

KI Beispiel 02
Image: OEE-Datastream with Anomaly

At irregular intervals, the OEE drops. There is no loss of availability, but a loss of performance. So there is no data point with a loss reason. Without the visualization, production management would not notice the loss. However, people recognize this anomaly in a graphic without being formally described. Only through visualization does it become visible and tangible. However, we do not yet know the cause. If one compares many of the data streams of the product over a longer period of time, the anomaly may possibly be assigned a pattern and thus an assumption of cause.

During operation of a factory, there are an infinite number of such possible occurrences:

  • The OEE of a product is regularly lower than that of the other products
  • The night shift has more short stoppages than the day shift
  • The OEE is lower than usual regardless of product from 1 to 2 pm
  • The ramp-up curve of a product up to the crest line takes longer than for other products

The usual reporting based on the number of pieces per shift or hour or the OEE of the shift or order means that the vast majority of these OEE optimization options are lost.

But how do you get to these productivity increase potentials without a human being having to continuously monitor the data streams?

OEE-Anomalydetection by Machine Learning

Developments in the field of artificial intelligence or, more precisely, machine learning, are capable of automatically detecting anomalies in large amounts of data and data streams.

Machine learning is a generic term for the “artificial” generation of knowledge from experience: An artificial system learns from examples and can generalize these after completion of the learning phase. That is, the examples are not simply learned by heart, but it alone “recognizes” patterns and laws in the learning data.

In the field of machine learning, a variety of algorithms exist that must be selected and trained for the particular application.

So if the OEE data is in machine-readable form, a properly trained and suitable algorithm evaluates it and the detected anomalies are processed for humans, the technology opens up completely new possibilities for OEE optimization. We are working on that. Look forward to our extended product oee.ai.