A Deep Learning Empowered Framework for Enabling Energy Conservation and Machine Diagnosis via Non-Intrusive Load Monitoring

Despite the immense success and powerful capability of machine learning and its subsets, there remain areas in which the technology has not been as thoroughly researched. The recent advent of Industry 4.0 has enabled new applications of machine learning - deep learning in particular, which require a vast amount of training data. The objective of this thesis is to investigate the more recent innovations in machine learning in the field of smart meter data through non-intrusive load monitoring (NILM): a technique for analysing gradual change in the energy draw and deducing what is used and how. Specifically, we explore machine learning-enabled challenges toward the health monitoring and proactive repair of electrical appliances, increasing the operational lifetime and inherently privacy-preserving. Machine learning can allow us to identify hidden indicators in data that the electrical appliance is deviating from its known, normal working pattern. Our dataset originates from retrofitted power outlets with metering functionality, and we are looking to investigate similar techniques based on energy consumption alone. Many existing techniques involved in IoT condition monitoring enjoy access to feature-rich sensor data, with a large basis of data on which to train. The research within is a natural complement to the data already collected by the smart appliances we are to support. This thesis explores the key challenges in implementing machine learning algorithms and the lack of research on those that look at appliances with cyclic load patterns (e.g., laundry appliances or dishwashers) before offering proposed discoveries. This thesis proposes and evaluates two deep learning algorithms and one ensemble-based machine learning algorithm in solving three distinct challenges. The first proposed model aims to identify anomalous behavior in the power signatures of household electrical appliances using an incremental clustering algorithm which first classifies the cycle, then trains an autoencoder to reconstruct the signature. The second model in this work aims to predict the idle time until the next usage following a period of activity of an electrical appliance, which can inform other systems to conserve power, prolonging the appliance. The model is based on temporal point processes, a classical statistical field. We architect an LSTM neural network capable of outputting direct time deltas. Finally, we look to identify specific faults in an appliance, this time knowing more of its nature, based on known failure conditions. We propose a gradient boosting model to classify a machine failure. Both anomaly-related models competently account for the expectedly skewed performance in the literature due to a steep imbalance in the data.