Network Effects on Dual Machine Learning Models Predicting Smart Home Sensor Measurements

Supportive smart home systems for older adults are growing in popularity. However, as an example of sensor networks, such systems can generate large amounts of small information packets that are repetitive and contain no new information. This has led to research projects focused on data reduction through dual machine learning models that reduce the network traffic between a home and cloud processing. This report focuses on the sensor-derived measurements that need to be efficiently communicated to the cloud. The report highlights the effects of network latency (or clock drift), jitter, and packet loss on overall performance for dual machine learning, as this has received minimal consideration previously. The current work presents a dual Autoregression machine learning model focused on predicting future sensor measurements of a smart home system. The work emulates a smart home system environment that includes cloud processing and a simulated imperfect network to reflect a real-life implementation. The results show how network effects such as the introduction of latency, jitter, or packet loss have caused as much as 50% of the cloud recording data samples being incorrect, compared to a perfect replication of data when network effects did not exist.