A systematic evaluation of the applicability of machine learning for energy efficient manufacturing

The industry sector is a significant consumer of global energy, making the improvement of its energy efficiency crucial for mitigating climate change. The increasing application of artificial intelligence in manufacturing, particularly machine learning, presents an opportunity to harness data-driven insights for improved energy efficiency. However, the decision to use machine learning (ML) is complicated by the varying applicability, effort and expertise required for different analysis methods.

This dissertation addresses the critical research gap in understanding when ML should be employed for manufacturing energy efficiency analysis. Existing research lacks comprehensive studies assessing the conditions under which ML can be effectively leveraged, and there is a scarcity of comparative evaluations of ML and non-ML methods concerning their performance, complexity, interpretability and data requirements.

As such, this dissertation aims to determine the scenarios in which certain ML classes provide advantages over traditional non-ML methods or other ML classes, identify the specific data requirements for these analyses and explore methods to overcome the challenges associated with those data requirements.

To achieve these aims, a comprehensive methodology is developed and applied across four real-world manufacturing datasets. This methodology systematically varies the number of features and training data instances across three classes of models: ``Simple Non-ML'', ``Simple ML'', and ``Complex ML''. Key findings reveal that ``Simple ML'' models consistently offer the most appropriate balance of performance and ease of development, outperforming ``Simple Non-ML'' models and performing comparably to ``Complex ML'' models while requiring less expertise. Additionally, the research demonstrates that ML models typically require more training data to achieve optimal performance, with a necessary volume approximately 2.5 to 4 times greater than that for non-ML models. Furthermore, the investigation highlights that for effective manufacturing energy consumption modeling, incorporating more than 10 features shows marginal improvements, while fewer than 10 features significantly reduce prediction accuracy. The dissertation also evaluates various data augmentation techniques to address the higher data needs of ML models, concluding that while augmentation is effective in other contexts, it does not yield benefits for manufacturing energy consumption models.

The contributions of this research are twofold. The first contribution comes from the development of a comparison methodology that enables both experts and novices to systematically evaluate different modeling algorithms. This methodology provides a clear framework for identifying the most suitable models for specific manufacturing efficiency challenges. Implemented as part of a low-code ML platform, this methodology enhances accessibility for non-experts, democratizing the application of ML in manufacturing. As a second contribution, the dissertation substantiates the applicability of ML for improving energy efficiency, offering practical guidance on the implementation and data requirements for successful ML usage. These contributions have value for both academia and industry, providing a foundation for future research and enabling manufacturers to make informed decisions on incorporating ML solutions to enhance the energy efficiency and sustainability of their processes.