Prediction of Next Sensor Event and its Time of Occurrence using Transfer Learning across Homes

We present results on the prediction of sequential sensor events and time of occurrence using transfer learning with Recurrent Neural Network with Long Short-Term Memory, between five apartments. Our dataset has been collected from real homes with one resident each and contains data from a set of 13–17 sensors, depending on the apartment, including motion, magnetic, and power sensors. We compare the prediction accuracy and the required dataset size for the prediction when each apartment is modelled individually, and when transfer learning is used. Transfer learning is used in two configurations — a) training with data from four apartments and fine-tuning and testing on each of the target apartments, and b) training with one apartment and fine-tuning and testing on each of the target apartments. In our best prediction models, a top accuracy of 87% is attained when predicting the next sensor event, and 81% when predicting both the next sensor event and the mean time elapsed to the next sensor event. There is a variability of 10% in the attained prediction accuracy across apartments. For a small number of events in the target dataset, having a network pre-trained with data from four apartments and fine-tuned with the target apartment provides the best prediction models.