Conference Agenda

In the dynamic landscape of industrial manufacturing, adaptability is key to efficiency and precision, especially in the automotive sector. As part of the research project AdaProQ we focus on the implementation of adaptive process chain optimization, combining machine learning (ML) and mathematical optimization to enhance manufacturing processes. At the core of our approach is the development of a prescriptive model capable of adaptively controlling the manufacturing process based on real-time data and predictive analytics.

We introduce the theoretical foundations of adaptive process chain optimization, including key definitions such as process chain, sensor variables, actuating variables, and target variables. Our objective is to train an ML model that serves as the basis for the optimization model, subsequently applying this model for real-time production optimization.

The aim of AdaProQ is the transfer from theory into practice. We demonstrate our approach on a metal forming process at a manufacturing company in the automotive sector. In detail, this encompassed understanding the process, categorizing variables, data preparation and analysis, and targeted data collection during a test day. This enabled the training of ML models and development of an optimization model, validated through a demonstrator.

Concluding, we discuss the success factors, challenges, and future perspectives of our project. Our findings underline the significant potential of adaptive process chain optimization in enhancing automotive manufacturing efficiency and quality, offering insights for ongoing research and development in the field.

The effectiveness of multidimensional poverty analysis can be enhanced through a more cohesive integration of machine learning (ML) techniques and poverty exploration combined with Bayesian estimates. This study focused on multidimensional poverty analytics, employing a blended strategy encompassing heuristic data generation, manifold learning, deep clustering, and Bayesian uncertainty quantification techniques. Specifically, the study utilized data collected from Turkish household budget surveys conducted between 2015 and 2019. To address the class imbalance between poverty classes, the synthetic minority oversampling technique (ubSMOTE) was employed to generate synthetic datasets. All analyses were conducted on the original highly imbalanced and synthetic datasets for comparison. Effective dimension reduction and clustering techniques, namely uniform manifold approximation and projection and Gaussian Mixture Model clustering, were employed to reduce data dimensionality and establish homogeneous household groups. Statistically significant differences (p < 0.001) were observed in the poverty prediction performances of ML models between the original highly imbalanced and synthetic datasets, with random forest outperforming other classifiers (CA: 98.66%). Decision tree and boosting integrated ML models outperformed popular ML algorithms in poverty prediction. In our study key predictors of poverty include income, out-of-pocket health expenditure (OOPHE), and household welfare clusters. Bayesian estimates of poverty using the Gini index with the GB2 distribution revealed that the level of inequality was higher for OOPHE compared to income across all study years. The results of this study contribute to a comprehensive understanding of poverty analysis, a crucial element in governing poverty analytics and achieving sustainable development goals.