請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/74329
標題: | 機器學習方法用於抗雜訊頻譜分析 Machine Learning Method for Noise Robust Frequency Analysis |
作者: | Chih-Hao Wang 王治皓 |
指導教授: | 丁建均 |
關鍵字: | 深度學習,背景雜訊處理,異常資訊分析,機器學習,物件的偵測與識別,時頻分析, deep learning algorithm,machine learning algorithm,background noise processing,abnormal information analysis,object detection and recognition,time-frequency analysis, |
出版年 : | 2019 |
學位: | 碩士 |
摘要: | 我們主要的目標是偵測微型目標。在做微型目標的偵測時,首先,我們分析大約每隔 50 ~ 2000 ms 的間隔往各個不同方向發射的 narrow beam 的雷達信號,一個類似於方波或其他固定波形的信號。接著,我們再根據接收到的回傳的雷達信號的延遲,運用都卜勒效應 (Doppler effect) 來判斷物體的速度。然而,由於微型目標的體積小且距離遠,以致於它們的雷達截面積 (Radar Cross Section, RCS) 通常很小,不易做辨識。
而用來偵測微型目標的雷達信號,更是容易受到干擾。一般而言,雷達信號的訊噪比 (signal to noise ratio, SNR) 大約在 10dB 至 15dB 之間。這在訊號處理的問題當中,算是不高的,容易造成辨識的錯誤。 在此篇論文中,首先我們先使用時頻分析得到雷達信號的時頻圖。接著利用前處理的方式,如內插與異常資訊分析來改善我們的資料。藉由使用脊部濾波器(ridge filter)找出我們欲分析的信號,最後利用自相關與傅立葉轉換得到的頻譜來分析旋轉頻率。這個問題類似於找出音樂信號的基頻,會有倍頻的出現,我們設立四個條件來決定最終旋轉頻率。 另外我們也試著使用深度學習的方式(UNet)來分析雷達信號,由於訓練模型的需要,我們自行模擬資料產生許多輸入資料來幫助我們訓練模型,並對UNet架構進行修改使其更適合使用在雷達信號分析。同時我們也使用兩個經典的機器學習的方式,SVM與KNN來和UNet預測的結果做比較。我們希望使用深度學習的方式來減少雜訊對分析雷達信號的影響。 Our main purpose is to detect micro-target. In the detection of micro-targets, first, we analyze the radar signal of the narrow beam, which is transmitted at intervals of 50 to 2000 ms in different directions. The radar signal is similar to a square wave or other fixed waveform. Next, we use the Doppler effect to determine the velocity of the object based on the delay of the received radar signal. However, because micro-target is small and far away, its Radar Cross Section (RCS) is usually small and difficult to identify. Radar signals used to detect micro-targets are more susceptible to interference. In general, the signal-to-noise ratio (SNR) of radar signals is between 10dB and 15dB. This is not high in the problem of signal processing, and it is easy to cause identification errors. In this thesis, first of all, we use the time-frequency analysis to obtain the time-frequency diagram of the radar signal. We use pre-processing methods such as interpolation and abnormal information analysis to improve our data. By using a ridge filter, we can find out the signal that we want to analyze. Finally, using the spectrum obtained by autocorrelation and Fourier transform to analyze the rotation frequency. This problem is similar to finding the fundamental frequency of a music signal, and there will be lots of harmonic frequencies. We set four conditions to determine the final rotation frequency. Besides, we also try to use the deep learning method (UNet) to analyze the radar signal. We generate a lot of input data to help us train the model and modify the UNet architecture to make it more suitable for radar signal analysis. At the same time, using two classic machine learning methods, SVM and KNN to compare with the results of the UNet prediction. We hope to use deep learning to reduce the impact of noise on analyzing radar signals. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/74329 |
DOI: | 10.6342/NTU201902987 |
全文授權: | 有償授權 |
顯示於系所單位: | 電信工程學研究所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-108-1.pdf 目前未授權公開取用 | 2.95 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。