Artificial intelligence, machine learning and neural networks are important topics for the future development of industrial productivity. There is a lot of discussion at the headline level. In this article, we provide an insight into how the details behind the technology work and how we use them to increase plant productivity (OEE) using the example of analyzing the stability of the plant flow.
Why neural networks?
Artificial neural networks represent a branch of artificial intelligence and are based on the human brain. The properties of neural networks make them interesting in all applications where there is little explicit knowledge about the problem to be solved. If you had this explicit knowledge, the problem could be solved with classical programming methods as defined by if-then-else. However, since this form of modelling would not be possible or at least extremely complex for many problems, neural networks are used in these areas. This advantage of the technology comes at the price of the fact that it is not possible to explain in retrospective why the algorithm made which decision.
Technological basics
The technical basis for calculating neural networks for oee.ai is Tensorflow, an open-source product from Google. Tensorflow is accessed via the Keras library in the form of an API. In this way, neural networks — in the case of oee.ai for analyzing time series data — can be programmed, trained, tested and also used productively. Recurrent (= feedback) neural networks, of which a variety of are available, are particularly suitable for analyzing time series. Selecting and configuring them correctly is the success factor for using AI for the respective problem.
Basics of data technology
Training data is required to train a neural network, although in principle, the more training data the better. In order to tell the algorithm what to search for, this training data must first be classified by a human — for our use case, for example, into stable and unstable. This task can be very specific to every application and can therefore be very complex. There are service providers available on the Internet to whom this task can be outsourced.
Training a neural network
Before training begins, training and test data are separated from each other, meaning that you only give the algorithm some of the labelled information for training. Learning a neural network takes place in several training rounds, the number of rounds being heavily dependent on the chosen algorithm and the heterogeneity of the data. Once the training is complete, you test how the trained algorithm performs on the labeled test data. This is how well the algorithm can handle the data with which it was not trained.
In the protocol visualized below, it is visible that around 11,500 labelled data sets were available, that the algorithm was trained on around 70% of the data in 12 rounds and, applied to the approximately 3,500 test data, it makes the same decision regarding stable and unstable as humans.
The result is already good, but not yet outstanding. Performance can be further increased by using more training data and fine-tuning the specific algorithm parameters.
In a visualization of the learning curve, you can see that the greatest learning success is already in learning round one, there is then a slight increase in quality up to the fifth round and that the remaining training rounds no longer have any further learning effect of the algorithm.
Using the neural network for OEE optimization
If you now visualize the phases identified as unstable in combination with the quantity time series, it becomes easy for humans to understand where the algorithm has classified the data as unstable.
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. the product portfolio on the system changes, the algorithm will always make around 90% of the same decision as the person who trained it.
With the help of this analysis, you can now start the process of improving plant productivity. What special features were there at the time when the plant was unstable? Was the material to be processed within the specification? , Were there any technical problems with the system? , Do employees need to be better trained? At this point, human intelligence takes over from machine intelligence and translates the analysis results into increasing plant productivity.