Incremental learning with multi-level adaptation
Neurocomputing
| Hi-index | 0.00 |
With the rapid development of microelectro mechanical systems (MEMS), wireless communications, and information networking and integrated circuits technologies, wireless sensor network technology came into being. For the monitoring of large areas or dangerous region or enemy region, commonly used, such as randomly dropping the sensor nodes, and coverage is one of the main problems to solve for the WSN. Control of the coverage for WSN, can be seen as when the energy of sensor network nodes, wireless communications network bandwidth, network computing ability are limited, through placing the sensor nodes and choosing routing at the network and other means to optimize the allocation of resources of WSN, avoid blind spots in the coverage area, or get in a local minimum value. Covering algorithm needs to consider network connectivity, energy efficiency, dynamics of the network, and other issues. According to the configuration of wireless sensor nodes, coverage is divided into certainty coverage and random coverage. In the actual environment, network topology is unpredictable, and network topology changes complexly, so it seems particularly important to achieve a dynamic coverage. It is a NP problem to achieve optimal coverage for monitoring the area of unfixed and intensive signals, so we made an effective distributed approximation algorithm to ensure that the sensor nodes change rapidly to achieve optimal coverage. Assuming each sensor node can directly communicate with other nearby sensors, have the same sensing range, and have data-processing functions, and move free according to the “order” of the nodes surrounding. Assume that each sensor can know its position, and the positioning algorithm has been adopted. Using a neural network algorithm can solve the nodal redistribution problem of WSN according to the detected region changes. In the artificial neural network, the learning algorithm can be divided into erection and growth structure learning algorithm. Growth learning algorithm starts training from a small network structure, then gradually increases until it achieves the final request. We choose the GNG algorithm, and the improved GNG-U algorithm which is improved on the basis of the GNG algorithm. To the nonstationary input signal, if the distribution suddenly changes, using the algorithm can quickly delete unused nodes, effect nodes will unite together fast, and redistribution, reduce redundant nodes finally reach optimal coverage rapidly. Combine the improved GNG-U algorithm with wireless sensor network, the network can rapidly re-coverage the target region especially for special environments. The simulation results show that, compared with the growing neural gas algorithm, the GNG-U and improve GNG-U algorithm can reduce redundant nodes, improve mobility of the network, accelerate the rate of convergence, and arrive at the optimal re-coverage dynamically.