Compressive data gathering for large-scale wireless sensor networks-包括大規(guī)模的負載平衡-sjtu-msra-

1、 Figure 14: Temperature data of the lowest slot generally in ascending order with only some small uctuations. There are also a few signicant spikes indicating abnormal temperature readings. Based on proposed compressive data gathering scheme, we are able to reconstruct such noisy sparse signals with

2、 spikes from M (M < N random measurements. Fig. 15(b(c and Fig. 16(b(c show the reconstruction results from M = 0.5N and M = 0.3N measurements at two time instances. The average reconstruction precision is over 98%. More importantly, the abnormal readings are accurately captured. To cope with the

3、 situation that temporal correlation becomes weak when the time interval increases, we can refresh the ordering of di periodically. In particular, for every one or two hours, the sink requests M (M = N random measurements in one data gathering process. When M = N , the set of equations in (2 is solv

4、able and the sink is able to obtain the exact values of di . Then, the sink can re-sort di and use this new ordering for data reconstruction in the subsequent hour or two. We would like to point out that both conventional compression and distributed source coding are unable to exploit this type of s

5、parsity which is observed only at certain reshufed ordering. In conventional compression, explicit data communication is required between correlated nodes. If correlated nodes are not physically close to each other, the communication between them may take multiple hops. This introduces high overhead

6、s and makes compression procedure costly. In distributed source coding, nodes are classied into main nodes and side nodes. The sink allocates appropriate number of bits to each node according to the correlation pattern. However, if the correlation pattern is based on changing sensor ordering, the si

7、nk needs to carry out these two tasks and communicate the results to every single node periodically. In contrast, the data gathering process in CDG is unaected when the ordering of di changes. The knowledge of correlation is only used during data reconstruction. Recall that CDG solves an l1 -minimiz

8、ation problem dened in (7 to reconstruct data. In previous sections, we have discussed how to select the matrix such that sensor readings d can be represented by a sparse vector x in domain. This section shows how d can be re-organize to be a sparse signal. This unprecedented exibility of CDG demons

9、trates how CDG can achieve a compression ratio of two to three at bottleneck nodes when other conventional mechanisms fail. 6. CONCLUSION AND FUTURE WORK We have described in this paper a novel scheme for energy ecient data gathering in large scale wireless sensor networks based on compressive sampl

10、ing theory. We believe this is the rst complete design to convert the traditional compress-then-transmit process into a compressive gathering (compress-with-transmission process to address the two major technical challenges that todays large scale sensor networks are facing. In the development of the proposed scheme, we have carried out the analysis of capacity for wireless sensor network when compressive data gathering is adopted. We have shown that CDG can achieve a capacity gain of N/M over baseline transmission. We have also designed ns-2 simulations to