Open Access

Richard Nordsieck

• The goal of modern manufacturing is to use machines to produce a product according to economic target criteria, e.g. quality characteristics or production speed. To achieve these target criteria, parametrisation processes, in which skilled operators iteratively adjust parameters of the machinery, are conducted. During parametrisation processes, raw material and energy is used. Furthermore, both personnel and production capacity are required. In contrast to this, external factors such as demographic change with resulting skilled labour shortage and climate change highlight the finite nature of these resources while increasing customisation of products leads to a decrease in production batches. Therefore, reducing the number of iterations and thereby the duration of parametrisation process needed is in economic as well as ecological interest. If it is not possible to decrease the complexity of these production processes to allow the operators to cope with the challenges of skilled labourThe goal of modern manufacturing is to use machines to produce a product according to economic target criteria, e.g. quality characteristics or production speed. To achieve these target criteria, parametrisation processes, in which skilled operators iteratively adjust parameters of the machinery, are conducted. During parametrisation processes, raw material and energy is used. Furthermore, both personnel and production capacity are required. In contrast to this, external factors such as demographic change with resulting skilled labour shortage and climate change highlight the finite nature of these resources while increasing customisation of products leads to a decrease in production batches. Therefore, reducing the number of iterations and thereby the duration of parametrisation process needed is in economic as well as ecological interest. If it is not possible to decrease the complexity of these production processes to allow the operators to cope with the challenges of skilled labour shortage and to increase the efficiency of raw material usage significantly, the producing industry in Germany, as it exists today, is unlikely to prevail. This would lead to significant societal challenges due to the high amount of employment by this sector. Given sensing equipment that is able to measure quality characteristics, the goal of reducing the time and resources spent during parametrisation processes can be achieved by providing operators with suggestions for or---if the quality of predictions are good enough and the corresponding parameters can be digitally affected---automatic selection of, a suitable parametrisation. Applying learning systems to this setting seems a reasonable choice given the widespread success they enjoyed in recent years. However, even given the move towards interconnecting and digitalising manufacturing through initiatives such as Industry 4.0, an application of learning system is met with several challenges in practice. Firstly, manufacturing in general, but German manufacturing in particular, is characterised by medium-sized businesses that heavily rely on the manufacture or use of special purpose machinery. This leads to limited amounts of data, since machines and the products they produce are often quasi unique, being produced at a very limited volume. Furthermore, these medium-sized businesses often struggle in holistically digitalising their processes and machines due to short-term economic planning which leads to an incomplete data landscape. Secondly, the data available is strongly skewed. This is due to several years of manual optimisation of the production processes in question which leads to relatively few unsatisfactory products when compared to the overall number of produced products. This thesis contributes to mitigating these practical challenges to manufacturing by proposing a neuro-symbolic fusion of learning systems and the operators' expert knowledge. Its main contributions are twofold: Firstly, the concept of data-based knowledge extraction is introduced to elicit tacit expert knowledge with minimal obstruction. Secondly, two regularisation-based architectures for a neuro-symbolic fusion of the expert knowledge and learning systems are presented. These contributions are developed on requirements gathered from two real-world case studies, of which one is from an industrial manufacturing setting, and evaluated on a synthetic as well as real-world dataset.…