透過您的圖書館登入
IP:44.200.23.133
  • 學位論文

應用於蜜蜂尋跡系統之平行處理、目標測距、與資料分群

Parallel Processing, Ranging, and Clustering for Bee Tracing Systems

指導教授 : 張帆人
共同指導教授 : 姜義德(Yi-Te Chiang)

摘要


數年前,臺灣大學與中正大學的研究團隊完成了蜜蜂尋跡系統。其關鍵技術包含了諧波雷達、微型收發器、以及虛擬隨機序列測距。由於諧波雷達輸入電腦之資料量非常龐大,在相關運算上相當耗時而成為瓶頸,以致於雷達天線的轉速受到限制。天線緩慢掃描的後果,讓野外實驗得到的蜜蜂位置數據嚴重不足。再者,原系統之目標判讀有賴人工,不僅費時而且容易誤判。本論文針對以上缺失,提出了新的運算架構(平行處理)與兩種演算法(目標測距和資料分群)。   平行處理部分,由於蜜蜂尋跡系統其運作原理甚為繁複,為減少資料運算的耗時、以便即時呈現運算結果,因而提出平行處理架構,該架構有效的利用了四個運算元、解決上述遭遇的困難。   目標測距的部分,本文將會使用人工類神經網路(Artificial Neural Network)對訊號進行分類,以辨識其是否包含目標資訊,訊號的特徵萃取將採用高階累積量(Cumulant)以及相似係數(Resemblance Coefficient)。本文也會深入探討虛擬隨機序列(Pseudorandom Noise)遭受嚴重雜訊干擾時之統計特性,以凸顯利用累積量作為特徵,在本系統訊號識別上的優勢。   在每次些許差異的測試環境下,訊號偶會遭遇到反射現象,本文亦利用統計學中之漢佩爾辨識法(Hampel Identifier),克服了這個問題。緊接著的問題是,雷達之天線束指向性有限,因此同個目標往往造成重複出現的假象。針對此問題我們利用了DBSCAN(Density Based Spatial Clustering of Applications with Noise)演算法,並基於我們的應用做出些改善,最後確能有效的對這種散佈點的例子進行分群。   以上運算架構以及演算法,使得改善後的蜜蜂雷達尋跡系統掃描速度提升了14倍,掃描40度的範圍只需要三秒鐘,並在辨識正確率高達99.6%的情況下,自動產生蜜蜂位置資訊,以供昆蟲專家研究蜜蜂行為。此外本文亦深入的探討偽隨機序列於雜訊影響下的統計結果、以及累積量於此的使用。基於本應用的改善版DBSCAN,也是十分新穎的嘗試。這些種種不應視為只是本系統的特例,其廣泛性應可推廣到未來其他方面的應用。

並列摘要


Several years ago, the bee tracing system was set up by researchers from both National Taiwan University and National Chung Cheng University. Technics involve: harmonic radar, micro transponder, and pseudorandom noise ranging. However, the heavy load of correlation calculations caused bottlenecks, and further limited the rotation speed of the antenna. Such issue led to the lack of bees’ location information. Moreover, the target identification in original system relied on manual work, which is neither effective nor precise. In this paper, we hereby introduce a new processing structure (parallel processing) and a series of algorithms (detection, ranging and clustering) with respect to the mentioned problem in original system. The parallel processing structure is mainly aim at solving the high computational complexity of our system, so that the result could be real-time presented. This structure properly utilized all four cores by parallel programming, thus increasing the scanning speed by 14 times, and the scanning of a 40° range will cost only 3 seconds. For the detection, we are going to use the Artificial Neural Network as our classifier, and the feature extraction is based on: 1) “Cumulant”, a higher-order statistic quantity, and 2) “Resemblance Coefficient”, an effective approach for radar signal feature. Its accuracy could be as high as 99.6%. We will also further discuss the statistical properties of Pseudorandom Noise under the serious noise situation, giving prominence to the utilization of cumulant here. Sometimes we may suffer from reflective propagations, we will introduce Hampel Identifier method to deal with such ranging problem. What comes next is the repeating of single target’s signals due to limited directivity of our antenna. We modified the DBSCAN algorithm, creating a novel method which is especially useful for clustering the spreading points such as our case.

參考文獻


[39]. J. P. M. de Sa and J. P. Marques de Sa, Pattern recognition: Concepts, methods and applications. Germany: Springer-Verlag Berlin and Heidelberg GmbH & Co. K, 2001.
[52]. R. R. Wilcox and R. Wilcox, Applying contemporary statistical techniques. United States: Elsevier Science, 2003.
[3]. Z.-M. Tsai et al., "A high range resolution 9.4/18.8 GHz harmonic radar for bees searching," in Microwave Symposium Digest (MTT), 2012 IEEE MTT-S International, Montreal, QC, Canada, 2012, pp. 1-3.
[4]. Z.-M. Tsai et al., "A high-range-accuracy and high-sensitivity harmonic radar using pulse pseudorandom code for bee searching," IEEE Trans. Microw. Theory Tech., vol. 61, no. 1, pp. 666-675, Jan. 2013.
[5]. P.-H. Jau et al., "Signal processing for harmonic pulse radar based on spread spectrum technology," IET Radar, Sonar & Navigation, vol. 8, no. 3, pp. 242–250, Mar. 2014.

延伸閱讀