Artificial Intelligence for OEE

Artificial intelligence, machine learning and neural networks are major topics in the future of industrial productivity. Much is discussed at the headline level. In this article, we’ll give you an insight into how the details behind the technology work and how we use them to increase equipment productivity (OEE) using asset stability analysis as an example.

Why neural networks?

Artificial neural networks are a branch of artificial intelligence and are modeled similar to the structures inside the human brain. The specifics of neural networks make them interesting in all applications where there is little explicit knowledge about the problem to be solved. If one had this explicit knowledge, the problem could be solved with classical programming methods in the sense of if-then-else. However, since this form of modeling is not possible or at least extremely time-consuming for many problems, neural networks are used in these cases. This advantage of the technology is bought with the fact that in the retrospective it is not possible to explain why the algorithm made which decision.

Technological basis

The technological base for calculating neural networks for oee.cloud is Tensorflow, an open source product from Google. Tensorflow is accessed through the Keras library in the form of an API. In this way, neural networks – in the application of oee.cloud for the analysis of time series data – can be programmed, trained, tested and also used in production. In particular, recurrent (= feedback) neural networks, of which a larger number are available, are suitable for the analysis of time series data. Their correct selection and configuration is the success factor for the use of AI for the respective use case.

Data technology basis

For the training of a neural network training data are needed, where in principle the rule, the more training data the better, is valid. In order to tell the algorithm what to look for, these training data must first be classified by a human – for example, in our case, stable and unstable. This task can be very specific for each application and therefore become very expensive. There are service providers available on the Internet to whom this task can be assigned.

Training a neural network

Before the training begins, training and test data are separated, so you give the algorithm only a part of the labeled data for training. The learning of a neural network takes place in several training rounds, whereby the number of rounds strongly depends on the selected algorithm and the heterogeneity of the data. When the training is complete, you test how the trained algorithm performs on the labeled test data. So one comes to a statement how well the algorithm can handle the data, with which he was not trained.

In the log visualized below, it is documented that around 11,500 labeled records were available, that the algorithm was trained on around 70% of the data in 12 rounds, and it applied to the approximately 3,500 test data for just over 90% the same decision regarding stable and unstable like the human who was training it.

Protocol of the learning process

The result is good, but not outstanding. More training data and a fine-tuning of the specific algorithm parameters can further increase performance.

In a visualization of the learning curve, you can see that the greatest learning success already occurs in the first learning round, there is then a slight increase in quality up to the fifth round, and no further learning effect of the algorithm occurs through the remaining training rounds.

Visualization ion the learning curve

Use of the neural network for OEE optimization

If one now visualizes the phases that are recognized as unstable in combination with the unit number time series, it is easy for humans to understand at which points the algorithm has classified the data as unstable.

Red marked areas als defined as unstable by the RNN

With this trained neural network all future data can now also be analyzed and as long as there is no structural change in the data, e.g. If the product portfolio on the equipment changes, the algorithm will always make about 90% of the same decision as the person who trained it.

With the help of this analysis, one can now enter into the improvement process of the equipment productivity. What special features were there at the times when the system was unstable? Was the processed material within the specifications?, Was there a technical problem with the system?, Should the staff be better trained? At this point, human intelligence takes over from artificial intelligence and implements the analysis results for the increase of the Overall Equipment Effectiveness.