Applications of Machine Learning in Industrial Buildings
Abstract
The advancement of Internet of Things (IoT) devices in computational and storage capabilities enabled their use in data- and resource-intensive applications, including Machine Learning (ML), across various disciplines. This thesis focuses on the challenges of building an ML pipeline in industrial settings, particularly in the stages of data acquisition, data processing, model training, and model interpretation. These challenges involve data privacy, data scarcity, model construction optimization, distributed learning, interpretability of Deep Learning (DL) results, and scarcity of communication and computational resources. To address these challenges, this thesis proposes solutions aimed at facilitating the deployment of ML pipelines in industrial settings. Under the broad theme of building capable ML models, this thesis proposes two methods for feature engineering of time-series data, capturing their underlying dynamics, and incorporating low-complexity ML algorithms and transfer learning to mitigate data scarcity. Additionally, a novel hyper-parameter optimization method is devised, combining reinforcement learning (RL) and transformer architecture, to expedite and enhance its transparency. Following this, an extensive analysis is conducted on integrating data-driven techniques to optimize warehouse energy consumption. Based on data collected from real-world environments, a use-case study evaluates different feature engineering techniques and the physical phenomena connected to Heating, Ventilation, and Air Conditioning (HVAC) operations encountered in enclosed spaces. Recognizing the importance of DL model explainability in aiding decision-making processes, especially, in high-stakes industrial environments, this thesis addresses two use cases: identifying features affected by industrial noise and detecting drifting features caused by cyber-security attacks. Furthermore, the thesis tackles the challenges of communication resource scarcity in the Federated Learning (FL) paradigm by proposing quantization techniques and investigating model weight similarity in heterogeneous FL scenarios. These contributions, available via open-source software, are poised to significantly advance the deployment and effectiveness of ML pipelines in industrial environments.