Due to growing demands for quality, sustainability, and digitalization, data science and artificial intelligence are gaining importance across industries. The extensive product range in many sectors often poses considerable challenges. For example, machine learning (ML) models may struggle with limited data per production variant. The present paper proposes a methodology that integrates the fields of data science and physical simulations. The results from finite element method (FEM) simulations are utilized to transform the process data in such a manner that it can be compared across processes for different production variants and employed for machine learning (ML) methods and statistical analyses. The method is illustrated using an example of aluminum production. A key advantage of this approach is that it can effectively model even production variants with very low quantities. The following discussion will present how this method can be used to enhance production processes, specifically to identify parameters that directly influence product quality, which would not be evident using alternative approaches. Furthermore, the work explores the potential for precisely controlling these parameters using ML models and discusses some major challenges.
In addition to topics such as data heterogeneity and missing values, there are numerous other issues that arise in the industrial context when optimizing processes using data-derived methods. A key issue is often the determination of an appropriate target variable that is comparable across all samples in the database, which is critical to the success of any analysis. A particular challenge arises when defects occur at the beginning of the process chain but are only detected at the end, given that products undergo 10 to 20 different systems in the meantime, undergoing several forming steps that influence defect detectability. A further challenge that is present in many industries is process data that describes entire batches. The challenge arises from the fact that all products in a batch often have the same process data, while the product quality within batches often varies considerably. This necessitates a comprehensive analysis to calculate a target variable for a batch. The chapter provides a detailed discussion of these challenges associated with the analysis of industrial process data. Additionally, it offers a perspective on the utilization of data analysis for the purpose of process optimization.
Due to growing demands for quality, sustainability, and digitalization, data science and artificial intelligence are gaining importance across industries. The extensive product range in many sectors often poses considerable challenges. For example, ma-chine learning (ML) models may struggle with limited data per product type. This paper presents a method that combines data science and physics-based simula-tions for target variable estimation, with the aim of creating a comparable database for effective statistical analysis and ML models — based on a use case from the aluminum production industry. A key advantage of this approach is that it can effectively model even production variants with very low quantities. The following discussion will pre-sent how this method can be used to enhance production processes, specifically to identify parameters that directly influence product quality. Furthermore, the work ex-plores the potential for precisely controlling these parameters using ML models and discusses some major challenges. The article demonstrates that integrating data science, technological knowledge and physics-based simulations is an effective methodology for estimating target variables, facilitating precise modelling even for production variants with minimal quantities and improving production processes.
In view of the increasing digitalization of industry, data science and data visualization are assuming an increasingly pivotal role in numerous industry sectors, facilitating enhanced monitoring and analysis of production processes. This development is primarily driven by growing demands for product quality and sustainability. In a multitude of industrial contexts, large volumes of high-resolution sensor data are continuously recorded throughout production. To make this huge amount of data usable for technologists, an interactive visual analytics tool for the analysis of production data has been developed. The tool enables users to explore complex process data, identify relevant patterns, and derive data-driven insights for quality optimization. This, in turn, leads to reduced scrap rates, as well as improved plant utilization and delivery performance. The present paper demonstrates the effectiveness of the proposed approach using a selected industrial use case and highlights key insights gained from its application in practice.