Presence Detection in Homes using Aggregated Sensor Data

Sammanfattning

This thesis presents the development of a binary classifier, which classifies homes as occupied or not. This was done using two different types of neural networks; standard feedforward networks and LSTM-networks. The input to these networks was sensor data collected from devices created by Minut AB. Data from previously used automatic alarm feature was used as ground truth. A major part of the project consisted of preparing and filtering the ground truth and input data. Once this was done a set of suitable hyperparameters was found by tuning the hyperparameters one by one with the other ones fixed. In general the tuning of the hyperparameters was too noisy to make any certain conclusions about which values were optimal. The classifiers succeeded in identifying some patterns indicative of home occupancy, outperforming a baseline model, which randomly guesses occupancy status. Despite this, the classifiers' performance did not yield the high accuracy required for their intended applications in heating system regulation and other home automation tasks. The feedforward network model got the best results, but LSTM-networks could potentially be equally good for this task, since the LSTM-networks were trained on smaller amounts of data and data quality appeared to affect the result more than model choice. Areas to improve include preprocessing, the method for choosing hyperparameters and quality of the ground truth data.