"Robust Non-Intrusive Load Monitoring for Industrial settings with high fidelity Simulations and Deep Learning ",
Universita degli Studi di Milano-Bicocca , 2023.
Nowadays, we are observing the fourth industrial revolution 4.0, which integrates new production technologies to increase productivity and production quality. As a result, new Smart Companies are emerging, with data monitoring systems that are increasingly advanced and interconnected. Therefore, there is a growing need to develop advanced energy monitoring techniques to identify machinery behaviors by observing time series of generated data. Modern industrial buildings use complex systems to monitor all devices connected to the grid. Understanding and optimizing these systems requires monitoring all major energy consumers and producers in the building. Unfortunately, high-quality smart meters come at a substantial cost for purchase, installation, and maintenance, so it is desirable to minimize the number of smart meters as much as possible. This thesis presents an approach to Non-Intrusive Load Monitoring (NILM) on real-world Smart Company energy consumption data using a synthetic dataset generated by a realistic simulation framework called the Auto Model Generator. These data are used to train different algorithms with the aim of disaggregating the total power consumption signal. Several state-of-the-art models are tested. Finally, we present a new approach to intelligently train the neural network and achieve more robustness by using data augmentation techniques for time series data and high fidelity simulations of possible industrial scenarios.
Download Bibtex file
Per Mail Request