Semi-Automatic Generation and Labeling of Training Data for Non-intrusive Load Monitoring

User awareness is one of the main drivers to reduce unnecessary energy consumption in our homes. This awareness, however, requires individual energy data of the devices we own. A retrofittable way to get this data is to use Non-Intrusive Load Monitoring methods. Most of these methods are supervised and require to collect labeled ground truth data in advance. Labeling on-phases of devices is already a tedious process, but if further information about internal device states are required (e.g. intensity of an HVAC), manual labeling methods are infeasible. We propose a novel data collection and labeling method for Non-Intrusive Load Monitoring. This method uses intrusive sensors directly connected to the monitored devices. A post-processing step classifies the connected devices into four categories and exposes internal state sequences in a semi-automatic way. We evaluated our labeling method with a sample dataset comparing the amount of recognized events, states and classified device category. The event detector achieved a total F1 score of 86.52 % for devices which show distinct states in its power signal. Using our framework, the overall labeling effort is cut by more than half (42%).