AI - Challenges

As already described in chapter 3, the algorithms for the successful application of machine learning in production engineering require the highest possible number of good data. Good data is data whose values are basically as correct as possible. This applies both to their mathematical value for the description of a physical quantity and to the temporal continuous correctness of the state of a respective date, also whether it is still a relevant influencing variable on the defined target quantity. Furthermore, the number of available good data is decisive. The more data of this data quality is available, the higher is the later functional quality of the "artificial intelligence" as a central control instrument of an autonomous production. This can be explained by the fact that the algorithms for learning a production behaviour typical for a company should, if possible, fall back on a good learning basis, similar to the learning process of a human being. The more qualitative his education is, the higher his later educational level. The quality of his training information (learning content) is a decisive factor in determining the later quality of his skills.

In a production environment there are at least two relevant types of data families. On the one hand, the group of all technical sensor information, including that of product quality monitoring, of the sensors installed in the production periphery, and on the other hand the group of business management information in the form of calculated data from existing software systems, such as ERP and accounting systems. The availability of these two families of data (technology and business management) allows artificial intelligence to make as accurate statements as possible about the economic efficiency of the production environment under consideration. When selecting the data to be considered, care must be taken to maximize the amount of good data for training the mathematical model of artificial intelligence.

A further factor is the time delay in the availability of information. For example, it is quite easy to provide a physical quantity from the production environment in real time, but business values, such as an operating result, are often not available in real time due to possible booking delays. How does a mathematical model react to incomplete data sets? Or is a statistical probability calculated from the past in order to forecast the real time value of business data as accurately as possible? Thus, there is a certain maximum tolerable fuzziness with regard to data quality for each optimization variable. Where are the limits and relevance of such a fuzziness? Do sufficiently accurate bases for the evaluation of forecast models of key business data exist? In the context of autonomous production technology, it will therefore be useful to find out with what degree of uncertainty in data quality an optimization model of artificial intelligence still works "sufficiently" accurately. The question to be answered is how the concept of "sufficient functional quality" of a mathematical model can be evaluated. For this purpose, there are various approaches to mathematical model validation, which will not be explained in detail in this article.

Thus, there is also an economic optimum in terms of providing data as well as necessary and not as well as possible.