The detection and transmission of a physical variable over time, by a node of a sensor network to its sink node, represents a significant communication overload and consequently one of the main energy consumption processes. In this article we present an algorithm for the prediction of time series, with which it is expected to reduce the energy consumption of a sensor network, by reducing the number of transmissions when reporting to the sink node only when the prediction of the sensed value differs in certain magnitude, to the actual sensed value. For this end, the proposed algorithm combines a wavelet multiresolution transform with robust prediction using Gaussian process. The data is processed in wavelet domain, taking advantage of the transform ability to capture geometric information and decomposition in more simple signals or subbands. Subsequently, the decomposed signal is approximated by Gaussian process one for each subband of the wavelet, in this manner the Gaussian process is given to learn a much simple signal. Once the process is trained, it is ready to make predictions. We compare our method with pure Gaussian process prediction showing that the proposed method reduces the prediction error and is improves large horizons predictions, thus reducing the energy consumption of the sensor network.

中文翻译：

---

基于小波高斯过程的无线传感器网络时间序列预测

随着时间的推移，传感器网络的节点对其物理节点的物理变量的检测和传输代表了很大的通信过载，因此是主要的能耗过程之一。在本文中，我们提出了一种用于预测时间序列的算法，该算法有望通过仅在感测值的预测不同时才向汇聚节点报告时减少传输次数，从而减少传感器网络的能耗在一定程度上达到实际感应值。为此，所提出的算法将小波多分辨率变换与使用高斯过程的鲁棒预测相结合。数据在小波域中处理，利用变换功能捕获几何信息并在更简单的信号或子带中分解。随后，对于小波的每个子带，通过高斯过程一个近似分解信号，以这种方式给出高斯过程以学习非常简单的信号。一旦对过程进行了培训，就可以进行预测了。我们将我们的方法与纯高斯过程预测进行了比较，表明所提出的方法减少了预测误差，并且改善了大范围的预测，从而降低了传感器网络的能耗。