Estimation Method Based on Deep Neural Network for Consecutively Missing Sensor Data

The phenomenon of missing sensor data is very common in wireless sensor networks (WSN). It has a dramatic effect on the usability, stability and efficiency of the WSN-based applications. There exist many methods for the missing sensor data estimation. However, the accurate and efficient consequent estimation of missing sensor data remains a challenging problem. To solve this problem, we propose a new method named consecutive sensor data deep neural network (CSDNN). In this method, firstly, we analyze the correlation coefficients among different types of sensor data and choose a certain number of nearest neighbors of the target sensor nodes. Secondly, to estimate a certain type of sensor data from a target sensor node, we utilize the different types of sensor data that are from the same target sensor node and have strong correlation with the missing ones, and the same type of sensor data from the aforementioned nearest neighbors. We treat these data as the input of the deep neural networks (DNN). Thirdly, we construct the DNN model, discuss the optimized DNN structure for the missing data problem, and test the accuracy of CSDNN for different types of environmental sensor data. The results show that the CSDNN method allows to accurately estimate the consecutively missing sensor data.