Presenter: Mads Kjærgaard Nielsen, AU

Link to personal website

Title and abstract: Using machine learning for production machine and material analysis: Bridging statistical learning and physics to generate actionable insights

- Mads Nielsen is a PhD student at BTECH collaborating with CORE affiliated Johan Clausen at BTECH on the use of advanced statistics for production machine energy optimization.

In times of increasing material and energy costs, uncertainty, and demands for reduction of greenhouse gas emissions, optimizing production machines in complex manufacturing systems can be central both for the protection of profit margins and the reduction of energy consumption. In industries such as the processing industry (which can span the production of plastics, biomaterials, and foodstuffs both for animal and human consumption), production machinery is typically part of a complex manufacturing setup where each machine can have many integrated adjustable parameters. Moreover, other factors linked to continuous production processes, e.g., different types of raw materials, high batch-/recipe variability in production orders, only increase the complexity of the problem. What is especially interesting, is that from physics, we know that the high number of adjustable parameters on the machinery and the variable composition of inputs and products create an inherently nonlinear and somewhat stochastic production process.

While stylized physics-based models can be applied to optimize such production machines, there are several core challenges a company faces when applying such models. It is a fact that some areas or materials still are underexplored theoretically, and often relatively minor differences in machine or product characteristics observed in practice can be difficult to account for in said models. Moreover, an important observation is that Danish SMEs in this area often are challenged with having the resources and expertise to extensively apply and validate physics-based models unless a strong business case can be established. Therefore, finding the right balance between process engineering depth and business value creation is paramount.

The main objective of this research is, when possible, to exploit modern statistical models to create a way to efficiently analyze production systems, which can handle the inherently nonlinear relationships and uncover the material-process-product relationships. The largest challenge has been to identify the right complexity of the modelling approach such that the participating companies gain tangible benefits from the findings and future use of the modelling workflow while still being feasible for the company to internalize and use in the future. Using data from real production environments, either as full- or pilot-scale systems, models were developed using relatively simple methods for machine learning together with clear variable interpretation methods (e.g., SHAP, counterfactual, and 3D surface visualizations). In the case of smaller datasets (i.e., fewer features and samples), parametric models including response surface modelling with L2 regularization have been used on observational data. Lastly, in especially low output systems, (e.g., polymer pultrusion) the output rate constraints the dataset size and only allows the use of basic statistics. By drawing upon physics-based knowledge of the production process we can contextualize and validate the statistical model outputs, and we can explore when it merely confirms expected results or adds new insights. Ultimately the aim of this work is to develop a framework that can be used to identify the match between modelling complexity and scale a given manufacturing setting calls for, thereby enabling companies with heterogenous needs to more efficiently use production data to optimize their complex production processes. 

Host: Marcel Turkensteen

Organisers: Surabhi Verma and Hartanto Wong