Sensor Reading Prediction Using Anisotropic Kernel Gaussian Process Regression

We utilize sensors to help us monitor events in the environment around us. To save power consumption, we often prefer to use as few sensors as possible and the sensors can be on for as limited time as possible while keeping the same or similar service performance from the sensors. In this work, we propose a mechanism that can use a small subset of sensor readings and the rest of sensor readings that are not collected can be approximated by the available sensor readings. We adopt Gaussian process regression as the prediction model. One key to have an effective Gaussian process prediction given sensor reading data of high variety relies on how we find an appropriate kernel function for the process. More specifically, given sensor data that have spatial and temporal relationships, we propose an anisotropic kernel for the process that can integrate different relationships as one and we can successfully describe the relationship between each pair of different sensor readings for the reading prediction. The experiments for evaluation are conducted based on a case study on weather data that consist of temperature readings collected in Taiwan. The experiment results show that the proposed Gaussian process regression with anisotropic kernel function can well describe the spatio-temporal relationships between different sensor readings and give effective temperature prediction.