Data Reconstruction In PM2.5 Sensor Networks

In the past years, air pollution is becoming serious which has attratced widespread attention. For real-time monitoring of air quality, government and individual deploy a large number of PM2.5 sensors. The sensors generate massive real-time PM2.5 sensor data. Because of various network transmission errors, central server does not record portion of the sensor data. The data missing produce great difficulties for PM2.5 data query and PM2.5 data analysis. The PM2.5 sensors data reconstruction will be full of significance. In this paper, we focus on reconstruction of the missing data in PM2.5 sensor networks.More interests are aroused in the data missing problem in Wireless Networks, and some algorithm to reconstruct is presented. But the existing method only reconstruct the number of missing data directly, which may lead to large deviation. In some application, probability distribution reconstruction is more meaningful than number reconstruction directly. The existing algorithm need the low rank structure or the sparsity of the original data. For PM2.5, the complexity of the location of sensor deployment result in the low rank and sparsity of the PM2.5 data. Our paper provides two different solutions for different applications.For sparsity deployment of PM2.5 sensor, we analyze the data by some test and we drew some conclusion as follows:PM2.5 data accord with gaussian distribution; data of PM2.5 has no structure of low rank; the distributed parameter generated by the sliding windows neither has the structure of low rank. On the basis of the above results, we define two kinds of similarity measurement:Tanimoto-coeficient and e_cosine_coeficient. Then, we proposed the matrix factorization based on local low rank and the similarity measurement which are defined by our paper.For densest deployment of PM2.5 sensor, we found the low rank structure of data. Our paper proposed a new reconstruction algorithm which is warm start, fast convergent and with temporal smoothing constraint.The test using real data show the efiency of our algorithm.