低功耗藍牙裝置應用於室外定位之研究

Abstract

本研究將針對室外倉庫之物料管理進行定位之相關研究，由於室外定位技術目前主流是透過全球衛星定位系統(GPS)來達到定位及導航的目的，而手機之GPS誤差約在5至10公尺左右，因此本研究使用Beacon定位來提高其定位精準度。現有文獻大多利用Beacon來進行室內定位及通訊，其透過高密度佈點及短距離傳輸的特性來達到估算距離之目的，然而透過Beacon來進行室外定位之研究卻非常稀少，因此本研究嘗試以Beacon來進行室外定位試驗測試。本論文使用低耗電藍牙(BLE)作為定位之技術，其原理為將量測之訊號強度(RSSI)依照其訊號強弱按照其特定換算方式判定距離遠近，由於訊號在傳送過程中會因為多重路徑效應影響而產生訊號衰減情況，因此本研究將透過手機大量收集在不同環境因子下的訊號強度以建立一濕度、溫度、訊號強度對應距離關係之回歸模型，亦嘗試以類神經網路來改善量測時所造成的誤差導致利用回歸模型推算距離時的誤判情況。本研究以倒傳遞類神經網路中的Bayesian Regularization演算法搭配矩陣型態的輸出資料格式訓練結果得到89.6%的預測正確率，平均誤差0.426公尺。最後以三角定位演算法估算手機座標位置，由類神經網路預測結果發現預測目標與實際目標座標誤差0.6356公尺，明顯改善由回歸模型推估的誤差2.143公尺。

In this paper, we focus on outdoor material positioning. Due to the development of wireless sensor technology, outdoor positioning and indoor positioning are widely popular with research scholar. Outdoor positioning are usually based on GPS(Global Positioning System) but the positioning error of mobile phone fall between 5 to 10 meters. Most of existing literature are using beacon in indoor positioning. It use high density and short distance transmission to estimate the distance between transmitter and receiver. Because there is a little literature about using beacon in outdoor positioning, we try to use beacon in outdoor positioning in this paper. BLE(Bluetooth Low Energy) is one of the most common positioning technology how it works is measuring the signal strength to estimate its distance between transmitter and receiver. In this paper we use collection of signal strength data with temperature and relative humidity to build a regression model. We also try back-propagation of supervised neural network to reduce the error of estimating the distance between transmitter and receiver from regression model, and we get 0.426(m) average error with 89.6% accuracy. Subsequently, we use three known coordinates of beacons in triangulation algorithm to find out the coordinate of phone. In this paper, we use Bayesian Regularization of back-propagation neural network to forecast coordinates of phone, and we get 0.6356 (m) average error better than 2.143(m) average error of using regression model.