Chung Cheng University Institutional Repository
Not a member yet
889 research outputs found
Sort by
異質傳感資料基於深度學習的疲勞駕駛預測;Driver Fatigue Prediction using Heterogeneous Sensor Data with Deep Learning
No full text[[abstract]]疲勞駕駛是造成交通事故的其中一個主要原因，若能在駕駛疲勞前預測到不久的未來駕駛將有疲勞發生，提出警告讓駕駛能停車休息，便能有效減少交通事故發生，過去大部分關於疲勞駕駛的研究大多使用單一傳感器資料來進行，一旦這個感測器出問題，很容易就影響整個系統的準確率甚至導致系統無法運作，除此之外，過去關於疲勞駕駛預測的研究通常使用傳統的反饋式類神經來進行預測，但疲勞駕駛的資料通常為一段具有時序性的資料，且有可能前面一段時間與後面一段時間的疲勞資料並沒有非常相關，若使用單純的反饋式類神經來進行疲勞駕駛預測，無法找出資料中的時序性，可能會影響預測的準確率。本篇論文提出一個使用異質資料結合一種效能更好的遞歸神經網路，長短期記憶神經網路來設計疲勞駕駛預測的模擬系統，使用心律變異數以及眼部影像計算眼瞼閉合度 (PERCLOS)同時輸入長短期記憶神經網路預測下一時間點駕駛是否疲勞，並給予駕駛適時的警告。實驗結果證明本文提出的方法能預測到下一時間點的駕駛狀態，正確預測出疲勞的準確率達89\%，整體準確率達90\%，並與傳統反饋式類神經網路及單一資料輸入比較。與本篇論文提出方法所比較的模型中，最佳正確預測出疲勞的準確率為使用心率資料輸入並使用反饋式類神經網路作為預測的模型，正確預測出疲勞的準確率準確率達80\%；最佳整體準確率為使用眼部影像資料輸入並使用反饋式類神經網路作為預測的模型，整體準確率達88\%，由比較的實驗可以看出本篇論文提出的方法有最好的準確率，能提高疲勞駕駛預警系統的效能。
Fatigue driving is one of the main causes of traffic accidents. If we can predict that a driver will be fatigued in the near future, and give warnings and parking suggestions to let the driver stop and rest, traffic accidents can be reduced effectively. In the past, most research work on fatigue driving used data from a single sensor. Once the sensor is faulty, the accuracy of the prediction system will be affected and it might even go unoperational. In addition, past research on driver fatigue prediction mainly used the conventional Back Propagation Neural Network (BPNN) model to make predictions. However, a driver's fatigue can usually be represented by a time series of sequence data. BPNN could not co-relate sequential instances of data in that series, thus making predictions not so accurate.In this Thesis, we propose a system that not only uses data from multiple heterogeneous types of sensors, but also are the more effective Recurrent Neural Network (RNN) with Long Short-term Memory (LSTM) building blocks to predict driver fatigue. We use the Heart Rate Variability (HRV) and the Percentage of Eyelid Closure (PERCLOS) simultaneously as data input to the LSTM RNN model so as to predict whether the driver will be fatigued in some future time slot. Thus, drivers are given timely warnings and/or parking suggestions. Experiments show that the system using heterogeneous sensor data with LSTM-based driver fatigue prediction module can achieve a true positive rate of 89\% and an accuracy of 90\%. We also compared fatigue prediction accuracy of the BPNN-based prediction module with single and heterogeneous sensor data, and the LSTM-based module with single sensor data. From experiments, the highest true positive rate of all compared modules is the BPNN used in the driver fatigue prediction module with HRV data with a true positive rate of 80\%, and the highest accuracy of all compared modules is the BPNN used in the driver fatigue prediction module with PERCLOS data with an accuracy of 88\%. As shown in the experiment results, we demonstrate that our method is more superior
以具適應性種子偵測法進行無序列比對之三代定序錯誤校正;Alignment-Free Error Correction for Third-Generation Sequencing by Adaptive Seed Identification
No full text[[abstract]]因為第三代定序技術所產生的序列為較長的序列、定序的偏差較低與定序分布平均等特性，使得第三代定序技術成為現有基因組裝(de novo assembly)的熱門選項。但是由於它產出的序列錯誤率較高，所以在進行基因組裝前必須進行序列的錯誤修正。目前錯誤修正的方法可分為比列序列分析法，如Canu，和非比對列序分析法；兩者面臨著準確度與速度之間的妥協。我們在之前研發出了一個利用FM-Index的非比對錯誤修正法，稱為FILEC，但是在中至大物種的組裝完整度上仍不讓人滿意。在這篇論文裡，我們提出了新的方法來提高FILEC在高相似度區域的種子準確率。方法首先會把高相似度與低相似度的區域區分開來，並且動態的使用不同的策略找種子；接下來錯誤的種子會被修剪與移除。實驗結果顯示出我們的方法相較於Canu不僅比較快，也保證了組裝的完整度與正確性。雖然在大物種組裝會變的破碎，速度上仍然比Canu快。
The thrid-generation sequencing technology is now commonly used for de novo assembly projects because of longer reads, less sequencing bias, and more uniform coverage. However, it comes at the cost of higher error rate, which requires error correction prior to assembly. The correction algorithms are divided into alignment-based methods like Canu, and alignment-free methods, which face the tradeoff between accuracy and speed. We previously developed an alignment-free algorithm based on FM-index, named FILEC, but the assembly contiguity is unsatisfactory in moderate and large genomes. In this thesis, we propose a new method to improve seeding accuracy of FILEC in repeat regions. The proposed method distinguishes unique and repetitive regions and adaptively uses different seeding strategies. The remaining error seeds were trimmed until the errors were removed. The experiment results showed that our method runs much faster than Canu and guarantees the contiguity and concordance of assembly . In large genome dataset, although the assembly result becomes fragmented, our method is still faster than Canu
中西醫合併大腸癌治療的存活分析;Survival Analysis of Combined Traditional Chinese and Western Medicine Therapy for Colorectal Cancer
No full text[[abstract]]本論文分析中西醫合併治療的大腸癌病患和非中西醫治療的大腸癌病患的存活時間是否有顯著差異。研究首先從台北與大林慈濟醫院取得2003年至2016年的癌症登記資料庫與中醫門診紀錄資料庫，篩選出欲分析的病患資料共2064人，其中包括非中西醫組1777人、中西醫組287人。本論文先透過卡方檢定得知在許多研究變項上治療方式分布有顯著差異，因此無法直接運用Kaplan-Meier存活分析。本論文接著使用Cox比例風險模型來分析多個研究變項對存活率的影響，並且考慮研究變項之間是否存在交互作用。分析結果顯示中西醫相對於非中西醫的治療方式減少約54.6%的死亡風險，且達到顯著差異。此外，本論文也發現2個單味藥、6個方劑、1個方劑組合在治療大腸癌上有顯著的效果。本論文顯示使用中醫改善病患免疫力及減緩西醫治療副作用可以顯著延長大腸癌病患的存活時間。
This thesis studies whether patients with colorectal cancer can prolong their survival through combined traditional Chinese medicine (TCM) and western medicine treatment. We collect the dataset from cancer registration database and Chinese medicine clinic records in Dalin and Taipei Tzu Chi Hospital for data between 2003 and 2016. This dataset includes 2064 valid patients, including 1777 non-TCM patients and 287 TCM patients. This thesis first uses the chi-square test to show that there are significant differences in the distribution of many research variables. Therefore, the Kaplan-Meier survival analysis cannot be directly applied. This thesis then uses the Cox risk regression analysis by stratifying the research variables that do not pass the PH assumption and identifying the research variables with interactions. The results of the analysis show that using the combined Chinese and western medicine treatment reduces 54.6% of the risk of death. In addition, the Cox risk regression analysis also identifies that 2 herbs, 6 formulas and 1 formula combination have significantly good effects in the treatment of colorectal cancer. This thesis shows that using TCM to improve body immunity and reduce side effects of western medicine treatment can significantly improve the survival time of colorectal cancer patients
韌體式訊號產生器之開發;Development of Firmware Signal Generator
No full text[[abstract]]本研究利用Microchip公司所生產的dsPIC30F4011這款微控制器，利用此款晶片具有 DSP engine能快速處理浮點數運算的特色，將演算法寫入晶片，並透過UART接收由Matlab端發送之頻率與波型命令，晶片端接收命令後透過演算法程式產生相符之頻率訊號，再經由數位轉類比IC與後級之主動式電路來產生精準且穩定可靠的波形訊號；透過將Matlab、晶片程式、周邊電路的整合，以低廉的價格、擴充性高、客製化服務為特色，開發出一套專門供給控制實驗室實驗所需的韌體式訊號產生器。
This research used the dsPIC30F4011 microcontroller manufactured by Microchip, Inc. Utilizing this chip features the DSP engine's ability to quickly process floating-point operations. The algorithm was written into the chip and transmitted through the UART to be sent by Matlab. After receiving the frequency and wave command, the chip generates the frequency signal corresponding to the algorithm program, then through the digital to analog IC and the subsequent active circuit to produce accurate and stable and reliable waveform signals; Through the integration of Matlab, the chip program, and peripheral circuits, a set of firmware signal generators that are specifically designed to supply control laboratory are developed at a low price
五軸CNC插補器設計與動態誤差分析;Five-Axis CNC Interpolator Design and Dynamic Error Analysis
No full text[[abstract]]為了達到先進製造的目標，五軸工具機是重要的加工技術之一，而電腦數值控制器是其中的核心技術，本論文所提出之五軸刀具中心點插補演算法可產生國際標準ISO 10791-6中的五軸檢測路徑以及多線段軌跡的高速插補命令，首先根據刀具中心點移動距離與機械驅動軸的最大速度來決定速度規劃中刀具中心點單線段的最大速度，並以刀具中心點與旋轉軸的路徑長度結合機械驅動軸的速度限制條件，推導轉角速度差公式，找出刀具中心點路徑上的轉角速度，接著使用S型加減速方法來規劃平滑之刀具中心點路徑與旋轉軸的進給率，最後利用動態模擬驗證轉角速度差法可大幅減少加工時間。同時本論文針對五軸量測路徑建立輪廓誤差模型，在五軸量測路徑中，分別將路徑插補命令代入伺服模型後，此模型可在五軸同動時準確預測動態誤差，穩態輪廓誤差模型(SSCE)在範例中展示三種特定的動態誤差行為:分別是單圓、雙圓以及平移誤差，並提出五軸伺服調整的方法，可以大幅改善由於各軸動態不匹配所導致的體積誤差，本論文所提出的方法已在實驗室之桌上型五軸雕刻機進行驗證，同樣可以有效的修正刀具中心點路徑之動態輪廓精度。此外，利用插補命令結合伺服系統的動態模型來進行擬驗證轉角速度差演算法的效能評估，動態模擬結果顯示除了可達到與商用控制器相似的輪廓精度外，並可有效改善加工時間。
In this paper, the five-axis tool center point (TCP) feedrate scheduling algorithm for measuring paths of ISO 10791-6 and linear multi-blocks trajectory are proposed to generate a high-speed interpolation command. First, the maximum speed in TCP single block segment is determined according to the segment displacement. The feedrate planning is constrained by axis velocity, acceleration and jerk on the machine tool. Each drive axes constraints and kinematical of the machine are employed to derive the five-axis corner velocity difference (FCVD) formulation. Next step, the S-shape Acceleration/Deceleration (Acc/Dec) profile is adapted to generate smoothing TCP and rotary axis feedrate in blocks. Finally, the FCVD method is demonstrated to substantially reduce the cycle time in dynamic simulation.Simultaneously, the TCP contour error model on five-axis measuring paths is derived. The error model can accurately estimate contour error during five-axis synchronized motions by substituting the commands of the measuring trajectory into the servo dynamics models. The steady state contour error (SSCE) model is demonstrated to illustrate three particular dynamic behaviors: the single-circle with amplitude modulation, double-circle effect and offset behavior. Furthermore, a servo tuning approach to achieve five-axis dynamic matching is utilized to improve contouring performance of the cutting trajectory. The contour errors caused by servo mismatch are reduced remarkably. Finally, experiments are conducted on a desktop five-axis engraving machine to verify the proposed methodology can improve dynamic contouring accuracy of the TCP significantly. Furthermore, the servo dynamic model of the five-axis machine is incorporated with the interpolator to evaluate the performances of the FCVD algorithm. The results is also shown that the FCVD has similar velocity characteristics as commercial controller. The FCVD has similar contour error as commercial controller, but it can achieve less machining time
高階五軸CNC插補器之開發;
No full text[[abstract]]本論文開發一高階五軸CNC插補器。首先，對五軸工具機建立模型，使用齊次座標轉換矩陣的方式建立機台刀具中心點與工作座標之間的相對關係，並透過順向與逆向運動學計算機械座標與刀具中心點座標的位置。接著進入插補程序，經過插補程序會得到以下的資訊(1)各軸的速度、加速度與急衝度的限制。(2)主動軸與從動軸的決定。(3)在多線段相接下的轉角速度。(4)使用即時預讀的方式判斷所計算出的轉角速度是否合理。(5)預測在計算出的轉角速度與設定的後插補時間下所產生的輪廓誤差。(6)選擇平滑化加減速的類型進行規劃。最後，會將本論文開發的五軸CNC插補器與商用控制器(西門子840D-sl)的插補點在相同路徑與參數下做比對。
In this paper, a five axis CNC interpolator is developed. First, the kinematics of five axis machine tool using homogeneous transformation matrix is derived to know the relative relationship between TCP and workpiece. Through the computation of the forward/inverse kinematics, the mechanical coordinate and TCP position can be established. Then from the interpolation procedure, one can obtain the following information (1) TCP and rotary axes velocity, acceleration and jerk limit. (2) Determination of master and slave. (3) Five axis corner speed. (4) Using look-ahead to decide a reasonable corner speed. (5) Prediction of contour error for different corner speed after interpolator. (6) Determination of smooth acc/dec type. Finally, the simulation result validates the proposed algorithm which are almost identical to Siemens interpolation results
日本作為中國民族主義中的他者─以電影文本為例;Japan as the “Other” in Chinese Nationalism: A Comparison of Two Films
No full text[[abstract]]日本在20世紀初期日本對中國發動侵略戰爭，至此之後中日關係便往不好的方向發展，而在21世紀的現在，中國人在提到日本時，往往咬牙切齒，並且將其視為十惡不赦的仇人。而多數影劇在重現中日八年戰爭時，一般多從國家或民族的角度來描繪中日八年戰爭所發生的事情，而戲劇內容也多放在描述戰爭場面，「太行山上」便是其中的代表之一。不過另外一部電影，「鬼子來了」卻是從平民的角度重新詮釋了中日八年戰爭時期，在電影內，中國人與日本人之間的區分似乎不再是如此清晰與絕對。這不禁也讓我們開始思考，日本這個他者在兩部電影中到底是如何被再現的？更進一步的問題則是「我們／他者」之間就是如此絕對嗎？ 從電影的內容上來看，我們可以發現「太行山上」這部電影中，將個體之間的差異與特質抹去，並將民族或國家這種集體與特定特質做連結，例如共產黨不僅受到民眾愛戴而且帶領人民積極抗日，反之國民黨則是在中日戰爭時期還不忘內鬥的團體，日本則被描繪成殘忍、卑鄙與無情的入侵者。而在「鬼子來了」之中，日本人並非全是殘酷無情的入侵者，而中國人民則展現出為求保全性命而委屈求全的一面。透過兩部電影的比較我們可以發現，日本這個中國民族主義中的他者，所以可以這麼簡單的被區分出來，很大一部份是因為我們往往高舉著國家，或是民族的名義來對人做出區分，但是當我們放棄這個旗幟時，我們將會發現這個區分不再如此容易。
In the twentieth century, Japan launched a second Sino-Japanese war, which lasted for eight years (1937-1945). Until the 21st century, the second Sino-Japanese war remains a common genre in the Chinese films, in which Japan always plays the role of invader and hence a negative “Other” vis-a-vis the Chinese “Self.” This thesis conducts a comparative study of two films to investigate how nationalism is deployed and reflected. The first one, “Tai Hang Shan Shang,” is an orthodox Second Sino-Japanese War film that meets the official ideology. In the story, the Chinese Communist Party (CCP), while in competition with the Nationalist Party (the Kuomintang, KMT) for political power and legitimacy, nevertheless fought bravely and full-heartedly against Japan, which was portrayed as cruel, mean and ruthless. In the second one, “Devils on the Doorstep,” however, a holistic perspective is abandoned in favour of individuals. As such, the Japanese military officers and soldiers, as well as some Chinese ordinary villagers, all showed various facets or characteristics, which render a simple and clear line between the Chinese Self and the Japanese Other obscure and difficult to maintain. The thesis hence suggests that a clear demarcation between self/other can only be sustained at the national level and is somehow arbitrary. If the nationalist perspective is replaced with one that focuses on persons with various desires, needs, and relationships, the meaning of those binary concepts underpinning the nationalist Self/Other relationship becomes ambiguous and may further prompt us to rethink the meaning and work of nationalism
基於深度學習方法的單一樂器音樂生成;A Single Instrument Music Generator via Deep Learning
No full text[[abstract]]近年來，人工智慧的浪潮隨著硬體設備的進步逐漸興起。其中，深度學習技術的突破在人工智慧的發展中扮演著重要的角色。深度學習是一種實現機器學習的技術，也是當前人工智慧的主流。深度學習首先於影像方面的應用有卓越的突破，相關的研究及成果如雨後春筍般地湧現。而英國倫敦Google DeepMind所開發的人工智慧AlphaGo在2017年擊敗世界棋王，更進一步證明了深度學習其廣大的發展性及未來在各個領域上應用的潛力。本論文希望藉由深度學習方法研究自然語言處理與其應用。現今，自然語言處理的應用十分廣泛，涵蓋語音辨識、機器翻譯、問答系統等範疇。本論文主要致力於音樂生成的研究，提出了一個基於Long Short Term Memory的編碼-解碼（Encoder-Decoder）模型。Long Short Term Memory設計重點在於可記住長期訊息，因而能夠有效地避免長期依賴問題。藉由合適的模型設計，希望能夠自動生成由單一樂器組成的音樂。關鍵字：深度學習、編碼解碼模型、注意力機制、音樂生成
In recent years, the wave of artificial intelligence has gradually emerged with the advancement of hardware equipment. Among them, the breakthrough of deep learning technology plays an important role in the development of artificial intelligence. Deep learning is a technology that implements machine learning and is also the mainstream of current artificial intelligence. Deep learning has excellent breakthroughs in the application of imaging firstly, and related research and achievements have sprung up. The artificial intelligence AlphaGo developed by Google DeepMind in London, defeated World Chess King in 2017, further demonstrating the deep development of its vast development and future potential in various fields.This paper hopes to study natural language processing and its application through deep learning methods. Today, natural language processing is widely used, covering speech recognition, machine translation, and question and answer systems. This thesis focuses on the study of music generation, and proposes an Encoder-Decoder model based on Long Short Term Memory. Long Short Term Memory is designed to remember long-term messages and thus effectively avoid long-term dependencies. With proper model design, it is desirable to be able to automatically generate music composed of a single instrument.Keywords: Deep Learning, Encoder-Decoder, Attention Mechanism, Musi
將檔案分頁放入休眠檔以縮短 快速開機後的反應時間;Fast Booting Technique for Reducing After Booted Response Time By Extended File-backed page into Hibernation File
No full text[[abstract]]休眠式快速開機在進入休眠前，將應用程式的記憶體大量swap out至swap space、寫回file system或是直接釋放，讓製作出來的休眠檔盡可能地小，使系統能在極短的時間內讀取完休眠檔，回復到休眠前的狀態，達到優化系統開機時間。但是因為休眠前將應用程式的資料大量swap out、寫回file system或直接釋放，如果開機後所執行的應用程式的working set不在休眠檔內，於系統回復後，會發生大量的page fault，從swap space和file system中讀入所需要的page，進而導致應用程式反應時間延長，影響使用者的體驗。本論文基於SBH快速開機技術提出的一個方法，我們的方法分為二個階段，第一個階段是對resume後所存取的file-backed page進行分析。當系統第一次休眠回復後，標記應用程式從file system中讀取進來的檔案分頁，在第二次休眠時，保留這些有被標記到的檔案分頁，擴增到休眠檔中。第二階段為系統真正執行的階段。第二次系統回復後，減少從file system中讀入檔案分頁的數量，達到優化應用程式的反應時間。，用來優化縮短快速開機後的反應時間。
Fast-booting hibernation before entering hibernation mode,the application's memory will swap out to swap space,write back to file system or directly release as more as possible,the hibernation file will be as small as possible,so that system can read the small hibernation file in a very short time and resume, optimizes the system boot time. However, because the application data is swap out to swap space, written back to the file system or directly released before hibernation, if the working set of the application executed after booting is not in the hibernation file, a large number of page faults will occur after the system resume.the system will read required page from swap space or file system, which leads to an extended application response time and affects the user experience.This paper is based on a method proposed by SBH fast boot technology. Our method is divided into two stages. The first stage is to analyze the file-backed page accessed after resume. When the system first time resume, the markup application reads the incoming file-backed page from the file system. When the second time resume, the file-backed page that have been marked are retained and expanded into the hibernation file. The second phase is the stage in which the system is actually execute. After the second time system resume, reduce the number of file-backed page read from the file system to optimize the response time of the application. Used to optimize the response time after fast booting