|摘要: ||無線感測網路(Wireless Sensor Network, WSN)主要被運用於區域資料的蒐集，透過資料的蒐集可以使得操作者對於其區域有更深的了解並根據資料做出更好的決策。WSN雖然可以針對一整片區域進行資料的蒐集，但是對於某些任務來說，必須要更準確的了解每個感測器所處於的位置，因此必須要對感測器做定位的動作以增加資料的可用性，而定位的準確性與所需花費的成本也為其最關鍵的問題。|
就目前定位計算的分類大致可以分為兩類: Statistics-based與AI-based。Statistics- based的方法計算較為簡單，但限制較多且定位的結果較不準確，如Kalman filter。而AI-based則反之，雖然計算較為複雜，但卻有限制少與高準確率的優點，如PSO、Neural Network。在AI的演算法中，我們主要評估不同Neural Network定位方法間彼此的差異性，根據定位方法的不同，NN的輸入也將有所差異，這樣的差異也將使得效能有所分別。文中除了比較NN定位方法間的差異外，我們還納入了PSO-based定位方法以評估不同AI間的區別。
Wireless Sensor Network is used mainly in regional data collection, through data collection can make the operator to understand the area and make the better decisions based on the information. Although it can collect the data of the area, but for the certain tasks, it must understands the location of each sensor, therefore, the sensors must be located to increase the availability of information, the accuracy of location and the cost are the most critical issues.
The classification of the current location calculation can be divided into two categories: Statistics-based and AI-based. The calculation of statistics-based method is simple, but it is more restrictive and inaccurate location, such as Kalman filter. The AI-based is contrary, although the calculation is more complex, but the limits are few and the accuracy is high, such as PSO and Neural Network. In the AI algorithms, we evaluate mainly the different location methods of Neural Network, according to the different location methods, input of Neural Network will be different, there will cause the performance difference. The paper in addition to compare the different between Neural Network location methods, we also included a PSO-based location method to evaluate the different between different AI.
For WSN, in addition to have a good performance, cost is also a very important part, if there is no balance of cost and performance, it will cause the reference value of location is not enough and life reduced of WSN.
We premise a location method of NN-based without additional cost, new method uses online training method to cause the trained network model correlated with the topology, we included all possible scenarios in topology to our training data, this mean the topology will be training data, online training method and topology training data will make our trained network model is completely relevant with topology. In order to improve the performance, we propose the estimate distance that combine two kinds estimate distance of RSSI and hop count to make more accurate, in the input part of network model, we also do the inverse cause the inputs have weight, combining two estimate distance and the concept of weight will cause the method have the better performance without addition cost.