Motivation. Technological advances that come with Industry 4.0, in e.g. sensoring and communication, unlock unprecedented large volumes of manufacturing data. This opens new horizons for data-driven methods like Machine Learning (ML) in analysis of manufacturing processes for a wide range of industries. An important scenario here is monitoring of manufacturing processes, including e.g. analysing the quality of the manufactured products and predicting the health state of machines and equipment. Consider an example of welding quality monitoring at Bosch, where welding is performed with automated machines that connect pieces of metal together by pressing them and passing high current electricity through them. Development of ML approaches for welding quality monitoring used in Bosch follows an iterative workflow that includes data collection (Step 1), task negotiation (Step 2), data preparation (Step 3), ML model development (Step 4), result interpretation and model selection (Step 5), model deployment (Step 6). Challenges. Development of such ML approaches is complex and costly and at Bosch it is mostly affected by the following 3 challenges: - Transparency: Steps 2 and 5 of welding quality monitoring require collaborative work of experts from different areas. The asymmetric knowledge backgrounds, including complexity of engineering practices in manufacturing and sophistication of ML algorithms that constrain the transparency of ML results and models, make the communication time consuming and error-prone. - Data preparation: Step 3 requires to integrate data from dozens of sources and this is a labor-intensive effort that requires necessary understanding of multi-faceted domain knowledge and plentiful data complications. - Generalisability of ML quality models: each ML model developed in Step 4 is typically tailored to a specific dataset and one process and thus its reuse for other data or processes, which is often needed, requires a significant effort.
By reasoning over ontologies we automate ML modelling. In particular, when
welding experts annotate raw welding data with features from the Domain (see Data
annotation box in Fig. 1), the reasoner derives from these features the relevant feature groups,
and then with the help of the ML Pipeline it derives relevant feature processing
algorithms, and relevant ML algorithms. Moreover, we use domain ontologies to annotate
resulting ML models and their selected features, thus enhancing their explainablility.
Evaluation at Bosch. We implemented and evaluated our approach at Bosch for
welding quality monitoring and prediction. The purpose was to enable predictive measures
such as equipment (re-)configuration or maintenance and thus to prevent welding
quality failures. The evaluation was done in the offline mode on two car manufacturing lines.
These lines generated large volumes of heterogeneous data such as historic data with
sensor measurements, welding machine configurations, manufacturing specifications,
and the quality estimates of finished welding operations. The raw data came in SQL,
Excel, plain text files, RUI-files and resulted in 263 fields with 53.2 million records, and
1.4 billion items after data integration. The accuracy of welding quality prediction by
ML models constructed with the help of SemML is good: the mean absolute
percentage error is within 1.61%–2.50% comparing to 3.19%–7.74% for simple baselines [