Data Augmentation Strategies for High-Frequency NILM Datasets

The nonintrusive load monitoring (NILM) process aims to monitor the different appliances connected to a power grid. It accomplishes this by analyzing one or multiple signals measured at the main breaker level of the power grid. In order for it return a response with high fidelity, it needs to be trained and tested using data pools of individual and aggregated appliance signals. Although some datasets are publicly available, they lack a sufficient quantity of labeled and time-stamped aggregated signals and a sufficient quantity of individual signals per appliance. Although recent works around this topic are available, they, however, work with low sampling frequencies, which does not allow us to extract high-frequency signatures from the signal waveform and in particular from the transient phases. In this article, we aim to solve the two problems mentioned above. We solve the first one using an algorithm that takes any number of individual appliance signals and construct a fully labeled and time-stamped aggregated signal. Next, we solve the second problem by introducing an approach that allows us to construct a large number of synthetic single appliance signals using a relatively smaller number of real single appliance signals. In addition, we perform all of this work on signals with high sampling frequencies, which will allow the extraction of the high-frequency features for later use.