博碩士論文 107327010 完整後設資料紀錄

DC 欄位 語言
DC.contributor光機電工程研究所zh_TW
DC.creator胡鎧麟zh_TW
DC.creatorKai-Lin Huen_US
dc.date.accessioned2020-8-18T07:39:07Z
dc.date.available2020-8-18T07:39:07Z
dc.date.issued2020
dc.identifier.urihttp://ir.lib.ncu.edu.tw:88/thesis/view_etd.asp?URN=107327010
dc.contributor.department光機電工程研究所zh_TW
DC.description國立中央大學zh_TW
DC.descriptionNational Central Universityen_US
dc.description.abstract摘要 本論文主旨在探討有關脈搏訊號研究的方法並結合現今主流的幾套時頻分析理論,以自行開發的一套特徵提取演算法分別依據倒傳遞神經網路及卷積神經網路的訓練需求,提取以光體積變化描計圖法量測到的脈搏訊號在時域或時頻域上的特徵,以監督式學習的方法訓練由資料庫中提供的脈搏訊號樣本,透過倒傳遞神經網路以及卷積神經網路的測試準確度來探討是否能根據時域以及時頻域的特徵成功地辨識被歸類在不同心血管疾病的人;並利用另一組資料庫驗證特徵提取以及神經路設定的合理性。最後討論加入主成分分析法及Dropout的概念優化倒傳遞神經網路並檢驗訓練結果的精準度是否有所上升。 實驗結果證明,倒傳遞神經網路在給定適當的收斂條件下的測試精確度能夠提升到60%至71%之間;而利用卷積神經網路進行訓練得到的測試準確度大約在61%至67%左右。由於卷積神經網路得到的結果較不理想,因此我們利用相同的設定對第二組資料庫進行訓練,最後發現對該筆資料庫訓練後的測試結果準確度能到達74%至78%左右,因此辨識不同健康狀況者所提出的幾種不同提取特徵的方式具有一定合理性與功能性,但因為第一個資料庫提供的樣本過少且訊號量測品質不穩導致最後的測試準確度並不如第二個資料庫的訓練結果來的高。 最後以主成分分析法對對原特徵群做降維的處理後配合Dropout的概念重新對倒傳遞神經網路進行訓練後發現測試準確度可以穩定地維持在70%至72%之間,可以視為一個收斂且可靠的結果。 關鍵字: 小波轉換、總體經驗模態分解、光體積變化描計圖法、倒傳遞神經網路、卷積神經網路zh_TW
dc.description.abstractAbstract The aim of this thesis is to discuss the details and algorithms of signal processing techniques such as continuous wavelet transform, discrete wavelet transform and ensemble empirical mode decomposition; and how they can be applied to human health condition classification by analyzing human pulse with aforementioned methods. By applying these methods and a new algorithm proposed in this thesis on human pulse dataset measured by Photoplethysmography(PPG), the features in time domain and time-frequency domain of PPG pulse signal can be extracted successfully as the inputs to two kinds of artificial neural network: Backpropagation Neural Network(BPNN) and Convolution Neural Network(CNN) to examine whether the 88 samples could be clustered into three different groups according to the cardiovascular conditions described in the datasets. The experiment shows that the testing accuracy of BPNN with inputs vector composed by six features in time domain, eight features in time-frequency domain and proper stopping criterions to avoid over-fitting is about 67% to 73%. On the other hand, the testing accuracy of CNN with inputs are time-frequency mapped with frequency range between 2Hz to 10Hz and time interval is one second can achieve about 61% to 64%. Comparing the testing accuracy of two datasets gives us a conclusion that the algorithm used in this paper to extract features affects the classification result is much less than the quality of PPG pulse signal acquired in different circumstances and samples of dataset for analysis. Keyword: Wavelet transform, Ensemble empirical mode decomposition, Photoplethysmography, Backpropagation neural network, Convolution neural networken_US
DC.subject小波轉換zh_TW
DC.subject總體經驗模態分解zh_TW
DC.subject光體積變化描計圖法zh_TW
DC.subject倒傳遞神經網路zh_TW
DC.subject卷積神經網路zh_TW
DC.subjectWavelet transformen_US
DC.subjectEnsemble empirical mode decompositionen_US
DC.subjectPhotoplethysmographyen_US
DC.subjectBackpropagation neural networken_US
DC.subjectConvolution neural networken_US
DC.title以PPG脈搏訊號提取時頻特徵做心血管疾病診斷的神經網路zh_TW
dc.language.isozh-TWzh-TW
DC.type博碩士論文zh_TW
DC.typethesisen_US
DC.publisherNational Central Universityen_US

若有論文相關問題,請聯絡國立中央大學圖書館推廣服務組 TEL:(03)422-7151轉57407,或E-mail聯絡  - 隱私權政策聲明