Chung Cheng University Institutional Repository
Not a member yet
889 research outputs found
Sort by
利用單軸加速度規之呼吸訊號擷取方法;Respiratory Rhyme Acquisition with 1-Axis Accelerometer Signal
No full text[[abstract]]「經驗模態分解（empirical mode decomposition；EMD）」已被證明可有效從高雜訊環境中成功分離呼吸訊號，然而呼吸訊號會隨機出現在不同的「本質模態函數（intrinsic mode functions；IMF）」，故篩選本質模態函數是採用經驗模態分解擷取呼吸訊號中最關鍵的技術。前人提出利用「多變量經驗模態分解（multivariate EMD；MEMD）」將三軸加速度器訊號拆解為三維本質模態函數組，並將傾角過大或過小之三維本質模態函數（也就是不可能由人類呼吸產生之訊號成分）剔除成功獲得呼吸訊號；但多變量經驗模態分解之運算量極大，不適合應用在嵌入式系統實作。本論文提出使用單軸加速度訊號之經驗模態分解搭配「趨勢波動分析（detrended fluctuation analysis；DFA）」選取單維本質模態函數，在同樣六個行車環境測試樣本下，皆可成功拆解出駕駛員之呼吸訊號，並大幅降低計算量。本論文同時提出「串流式趨勢波動分析計算（stream DFA）」，更可降低54%趨勢波動分析之計算量，使本論文所提出之單軸加速度器呼吸訊號擷取更適於嵌入式系統產品實作。
Empirical mode decomposition (EMD) is a proven technique to decompose breathing where from noisy signal recorded in severe environments. The breathing signal locates in arbitrary intrinsic mode functions (IMF) from EMD and selecting appropriate IMF to compose the breathing signal is critical for EMD-based breathing signal extraction. Multivariate EMD (MEMD) has been proposed to decompose 3-axis accelerometer signal into 3-dimension IMFs, the angle of which cannot be generated from human breathing is utilized as the feature to select IMF. However, MEMD needs huge computations and is not suitable for embedded implementation. This thesis proposes to use detrended fluctuation analysis (DFA) to select IMF from EMD results on 1-axis accelerometer signal. In our experiments, the accuracy on driver’s breathing extraction is comparable to that based on MEMD. In addition, stream DFA is proposed to reduce 54% computations, which makes the proposed breathing extraction with 1-axis accelerometer much more suitable for embedded systems
網路拍賣與遊樂園門票-以日本 Yahoo!拍賣為例;Buying Amusement Park Tickets in Online Auction: Evidence from Yahoo! Japan
No full text[[abstract]]近年來，科技帶動了網路蓬勃發展，網路拍賣也隨之興盛。雖然網路提供了便捷性，但也加深了買賣家雙方之間的資訊不對稱，本文使用Yahoo!日本拍賣的遊樂園門票為樣本，收集從2014年4月15日至2014年5月14日共一個月之資料，使用迴歸模型分析，發現到賣家有負面評價將會使結標單價下降，而賣家評價較高也容易得到較高的結標單價。另外，相較於eBay，‘延長次數’為Yahoo!日本拍賣之特有變數，‘延長次數’越多可以得到越高的結標單價，賣家可藉由設定較低的起標單價或幫買家支付運費來增加買氣，使得延長次數增加，而有經驗的買家較容易採取不延長賣場時間的策略
安全氣候知覺、風險知覺、組織承諾與職業安全衛生績效關聯之探究;A Study on the Relationships among Safety Climate Perception, Risk Perception, Organizational Commitment, and Occupational Safety and Health Performance
No full text[[abstract]]本研究主要探討安全氣候知覺與風險知覺、職業安全衛生績效間之關係，並利用組織承諾做為調節變項，另驗證風險知覺在安全氣候知覺與職業安全衛生績效間的中介效果。本研究採用問卷調查收集資料，透過親送、E-Mail 或郵寄等方式，以一份公司職安衛人員卷，搭配五份該公司作業人員卷的方式，進行問卷之發送與回收，共計發出職安衛人員 65 份（收集企業組織的職業安全衛生績效）、基層作業人員325 份（收集員工對企業組織在安全氣候知覺、風險知覺及組織承諾等方面的體認），回收的有效問卷分別為職安衛人員 65 份、基層作業人員320 份，共計 65 家公司；並透過描述性統計分析、驗證性因素分析、相關分析、迴歸分析等統計方法探討變數間的關係。本研究以迴歸模式進行主要的假設驗證，經實證分析結果顯示如下： 一、安全氣候知覺對風險知覺具顯著正向影響。 二、風險知覺對職業安全衛生績效不具顯著影響。 三、安全氣候知覺對職業安全衛生績效具顯著正向影響。 四、風險知覺在安全氣候知覺與職業安全衛生績效間不具中介效果。 五、組織承諾(持續承諾)在風險知覺與職業安全衛生績效之間具調節效果。最後，依據研究結果，提出本研究結論與建議，以供後續研究參考。
This study is to explore the relationships among safety climate perception, risk perception, occupational safety and health performance, and organizational commitment. Meanwhile, the mediating effect of risk perception between safety climate and occupational safety and health performance was conducted. In addition, the moderating effect of organizational commitment between risk perception and occupational safety and healthy performance was explored.Therefore, dyad questionnaire included one staff of occupational safety and healthy performance of enterprises and five operational employees. Totally, 65 staffs of occupational safety and healthy performance of enterprises and 320 operational employees valid questionnaires were collected.The statistical methods such as descriptive statistics, confirmatory factor analysis (CFA), and analysis of correlation were used in the study. The results of the study are shown below: 1. Safety climate perception has a significantly positive influence on risk perception. 2. Risk perception has a significantly negative influence on occupational safety and health performance. 3. Safety climate perception has a significant positive impact on occupational safety and health performance. 4. Risk perception has a partially mediating effect between safety climate perception and occupational safety and health performance. 5. Organizational commitment does not have significantly moderating effect between risk perception and occupational safety and healthy performance.Based on the research results, the concrete conclusions and suggestions will provide for HRM practice and future study as references
基於信標之社交系統;A Beacon-Based Social System
No full text[[abstract]]近幾年，社群網路與社交軟體的發展越來越熱烈，像是新浪微博、推特、微信、LINE、Facebook…等，功能不斷的推陳出新，使得人與人之間資訊交換越來越便利、簡單，網路交友也變得是一件稀鬆平常的事情。現今的網路社交軟體不斷的推陳出新，註冊帳號的門檻也越來越低，使得有心人士介入其中，造成許多空頭帳號以及情色廣告資訊氾濫，除此之外，還可能進而造成網路詐騙、人身安全、桃色陷阱…等社會問題，其實，在過去的各國地區中，因網路交友不慎而遭遇不幸的事件不計其數，因此，現今社會擁有這麼多樣化的社群網路與社交軟體的情況下，交朋友一事變得相當重要，必須受到重視，本研究的目標是建立一個區域性社交系統，它擁有即時性、隱匿性與安全性，並可將配對時所遇到的空頭帳號問題去除化，打造優良的社群環境。本系統的環境建立於行動應用程式上，並且結合了信標(Beacon)，藉由它，將範圍內的使用者資訊推播於此範圍中，讓使用者處於信標(Beacon)推播範圍內時，可以與其他使用者進行相似度演算法的社交配對，此範圍可依實際場地狀況進行調校，達到區域性社交之目的，除此之外，我們還利用微型定位技術，對配對成功的使用者進行室內定位，引導雙方進行見面交流。我們提出了一個區域性社交軟體，擁有的即時性與區域性可以提升使用者體驗並同時保障使用者安全，此系統除了可以設置於一般的室內場合，也可以設置於戶外公共場合，例如：公園、車站、廣場都是很好的地點。
In recent years, the social network and social software have grown substantially. The Weibo, twitter, Facebook, Line, and We Chat provides more service that makes information exchange between people more convenient.There were much social software on the network and the threshold of registering accounts is getting lower that making many fake accounts and the proliferation of pornographic advertising, in addition, it may cause problems, such as Internet fraud, personal safety and erotic trap, in fact, some people had suffered because they make wrong friends on the network. Therefore, with such diverse social networks and social software in the community, making friends is an important issue that must be given due attention. The goal of this study is to establish a regional social system that is immediacy and regional, and removes the shortcomings associated with pairing to create an excellent social environment.The environment of the system is built on the mobile application and incorporates a beacon. With this system, the user information is pushed to this range when the user is in the beacon push range, the system will perform user pairing through similarity algorithm that range can be adjusted according to the square meters. In addition, we also use the miniature positioning technology to locate the paired successful users and allow users to meet.Here we propose a regional social system that has immediacy and regional features that enhance user experience and ensure user safety. The system can be set in general indoor, can also be set in public places, such as parks, train station, plaza
根據社群結構改進以標籤傳播之重疊社群檢測方法;Improving Overlapping Community Detection With Label Propagation According to the Community Structure
No full text基於OP_RETURN指令之效能改善與應用;Performance Improvement and Application Based on the OP_RETURN Instruction
No full text[[abstract]]隨著各種數位貨幣近年不斷的推出，區塊鏈(Blockchain)的重要性也逐漸被重視，其中，腳本語言(Scripting Languages)更是扮演著區塊鏈(Blockchain)能成功運行的重要因素之一。 OP_RETURN為腳本語言(Scripting Languages)之中的一項指令，當使用者需要在交易中儲存部分個人資料時，則會使用該OP_RETURN指令，然而，OP_RETURN指令目前最多能儲存80byte的資料量，以目前使用者使用OP_RETURN指令平均儲存23byte的資料量而言，OP_RETURN指令儲存80byte的資料量確實高出許多，這將導致直譯器(Interpreter)在讀取OP_RETURN指令時，額外增加不必要讀取的浪費時間，以及增加系統必需建立不必要的閒置空間。 為了解決上述問題，我們重新定義OP_RETURN指令，提出兩種指令，並重新命名為OP_RETURN27與OP_RETURN26，其兩者指令所能儲存的資料量分別為27byte與26byte，此外，我們提出三種使用者使用指令儲存資料量的演算法，藉由這三種演算法來選出最適合使用者儲存資料的空間。 另一方面，我們使用將近70萬筆的公司註冊資料來做為效能評估之測試資料，結果顯示，在這將近70萬筆的資料中的每筆資料之欄位資料量幾乎小於27 byte，因此我們將部分每筆資料之欄位資料量的預設配給空間設為27 byte，並且事先讓系統設定27 byte空間預留給下一筆欄位使用，藉由我們的方法提高系統預測給予欄位的適當儲存空間之命中率，最後經由數據的統計，經由我們提出的指令和演算法之效能皆優於原本OP_RETURN指令。
With the continuous rollout of a variety of digital currencies in recent years, the importance of blockchain is gradually being emphasized. Among them, scripting languages play an important role in the successful operation of blockchain.OP_RETURN, an instruction in the Scripting Languages, is used when the user needs to store some personal data in the transaction. However, at present the OP_RETURN instruction can store up to 80 bytes of data at most. As the average user’s data storage is 23 bytes, 80 bytes of data storage using OP_RETURN instruction is indeed much higher. This will cause the interpreter to increase the unnecessary time when reading the OP_RETURN instruction and thus create unnecessary idle space in systems.In order to solve the above problem, we redefine the OP_RETURN instruction, propose two kinds of instructions, and rename them as OP_RETURN27 and OP_RETURN26. The amount of data that the two instructions can store is 27 bytes and 26 bytes respectively. In addition, we propose three types of user instructions for storage. The data volume algorithm uses these three algorithms to select the most suitable size for users to store data.On the other hand, we use nearly 700,000 company registration data as test data for the performance evaluation. The results show that the amount of data in each of the nearly 700,000 data column is almost less than 27 bytes. We set the default allocation space for the amount of field data for each piece of data to 27 bytes, and let the system set 27 bytes of space for the next column to be used in advance. By using our method to improve the hit rate of the system to predict the appropriate storage space for the column, and finally through the statistics of the data, the performance of the instructions and algorithms proposed by us is better than the original OP_RETURN instruction
使用深度學習方法之數學文本生成;Mathematic Formula Generator Base On Deep Learning Technology
No full text[[abstract]]近年來隨著硬體設備的不斷進步，在運算能力許可的情況下，人工智慧的發展如日中天。在人工智慧的潮流下，機器學習是一門重要的分支，深度學習則是目前十分熱門的機器學習方法之一。深度學習曾經一度被認為是困難、甚至無法實現的一種機器學習方法，因為深度學習需要有強勁的計算速度以及足夠大的記憶體空間當後背，才能發揮其功效，所幸GPU高度平行運算，以及記憶體平價化，現今我們才能夠有更多資源投入深度學習方面的研究。深度學習首先在影像方面取得了卓越的進步，在2015年10月Google DeepMind公司開發的人工智慧Alpha Go使用圍棋擊敗了職業棋士，證明了深度學習在各個領域上的可能性。本論文希望藉由深度學習的方法，在自然語言處理方面研究其應用。自然語言處理的範疇十分廣泛，今日在語言翻譯、情感分析、句法生成都有其應用。本論文希望在文本生成方面進行深究。數學式為一種容易驗證，並且訓練資料能夠大量生成的一種資料集，加上Recurrent neural networks對於時間序上的記憶性，以及合適的模型設計，希望能夠產生自動化分析數學文本，並且產生其解答的成果
基於熱傳導模型之偵測交通異常種類及 區域之方法;Traffic Anomaly Detection Based on a Heat Diffusion Model
No full text[[abstract]]由於資訊科技技術的進步，許多感測器被部署在智慧城市的道路交通網上，實現了自動化智慧交通管理。本文提出了一種基於能量模型的交通異常區域檢測系統，以自動檢測異常區域。我們觀測到公路交通網的車流交通移動行為，類似物理中的熱傳導模型的方式將高溫區域的熱能轉移到較低的溫度區域。我們運用庫侖定律的概念來獲得道路上各類事件對車流所造成的影響。此外，我們還討論了鄰近地區的關係，以決定這些地區發生的事件的候選發生地點。模型中所使用的權重代表了現實環境中影響車流的傳導能力。隨後，我們提出了一個異常檢測模型，該模型可以及時算出交通事故的確切發生位置。最後，將我們的方法與現有的方法進行比較。實驗結果表明，我們對車流分佈的估計比現有的方法更接近實際的感測器記錄。通過大量的實驗，證明了我們的模型的實用性以及在事故檢測準確度方面優於現有方法。
Thanks to the advancements in sensor technology, many sensors are deployed on road networks in smart cities to achieve the automation development of smart solutions for traffic management. In this paper, we propose a traffic anomalous location detection system based on an energy model to automatically detect anomalous locations where incidents may occur. We utilize the heat diffusion model in physics by observing the level of traffic flow spreading along the road networks in a similar way to the heat energy transferring from a high temperature region to a lower one. The weights used in the model represent the environmental conditions which affect the ``conductibility'' of the traffic flow. Subsequently, we apply the concept of Coulomb's law to acquire the influence caused by various distinct types of incidents on the road. We construct an anomaly detection model that computes the source location and the type of a traffic incident in a timely fashion. The experimental results show that our estimation of the traffic flow distribution is much closer to the actual sensor records than that of existing approaches including an un-directed heat diffusion model and autoregressive integrated moving average model (ARIMA). On the other hand, we analyze and compare the effectiveness of our model with a support vector regression (SVR) model under different traffic flow conditions. Extensive experiments are presented to demonstrate the performance and utility of our model which outperforms the existing approaches in terms of incident location detecting accuracy. Furthermore, our model has the ability to provide various distinct incident types detection which outperforms the detecting accuracy of the support vector machine (SVM) approach
深度感知之影像著色類神經網路;Depth-Aware Image Colorization Network
No full text[[abstract]]在電腦視覺領域中，影像著色一直都因其不確定性而使它成為一個極為困難的問題。所謂不確定性，指的是一個物體的顏色有非常多的可能性，舉例來說，一頂安全帽有可能是黃色，也有可能是藍色，這個特性增加了影像著色中顏色預測的難度。現行的影像著色方法已經可以針對同一物件預測出不同的顏色，提升影像著色的準確度。然而在物件與物件間的交界處，卻時常受到相鄰物件的顏色影響，使不同物件被塗成同一種顏色，產生不自然的著色效果。 本篇論文主要探討在類神經網路的影像著色方法中，結合深度資訊來解決這個問題。近年來，深度學習已成為影像著色方法的主流。另一方面，在許多文獻中，深度資訊已被驗證可以改善電腦視覺以及影像處理相關方法的效能。然而，至今尚未有將深度資訊應用於提升影像著色的方法。因此本篇論文提出一個以類神經網路為基礎，結合深度感知的影像著色方法，並透過許多實驗來驗證提出的方法確實能夠提升影像著色的效能。
Image colorization remains a challenge problem in computer vision. This problem is underconstrained since an object has various colors in the real world. Currently, many studies have achieved good performance in images colorization. However, the color bleeding problem still exists. Different objects share the same color because they are nearby, leading to the result of the boundary between different objects looks unnatural. In this thesis, we study how to combine depth information into neural network-based image colorization. Neural network-based methods are also known as learning-based methods. Recently, learning-based methods have been mainly used in image colorization and have presented promising colorization results. On the other hand, various computer vision works utilizing depth information have achieved good performance. But to the best of our knowledge, depth information was not used in image colorization before. Hence, we propose a depth-aware image colorization network in this thesis. We evaluate the proposed method via several experiments, and the experimental results show that the proposed method can achieve better colorization results
基於深度強化學習之自動駕駛車輛路口控制系統;Intersection Crossing for Autonomous Vehicles based on Reinforcement Learning
No full text[[abstract]]在路口發生的交通事故將會嚴重影響交通流量，且該類型的交通事故佔有很大的發生比例。在人口與車輛數量均增加的趨勢下，發展更安全而且有效率的路口控制系統有其必要性。而且隨著車載通訊與自動駕駛技術的進步之下，我們可以發展更聰明的控制系統，可以減少人為因素所造成的種種交通意外。 因此，我們提出了基於深度強化學習之自動駕駛車輛路口控制系統(Deep Reinforcement Learning-based Autonomous Intersection Management，簡稱DRLAIM)。在這個系統中包含三個主要模型，分別為安全煞車控制模型、基於強化學習之路口控制模型及優先權分配模型。我們希望透過強化學習去訓練我們的系統，讓其在與交通環境的互動中學習到一個好的交通控制策略，並且使用安全煞車控制模型去確保每一輛自動駕駛車輛在行進中的安全。 實驗結果顯示，我們的路口控制模型在使用強化學習訓練之後，道路的吞吐量增加了83%。與快速優先服務策略(Fast First Service，簡稱FFS)相比，DRLAIM的車輛平均等待時間較少。減少的幅度分別為1.2%到11.4%。
Road accidents at intersection significantly impact traffic flow and account for a significant proportion of all accidents. With the trend of increasing population and vehicle numbers, it is necessary to develop a safer and more efficient intersection control system. Furthermore, with the advancement of vehicular communication and self-driving technology, we can develop a smarter control system that can reduce the traffic accidents caused by human factors. For increasing safety and efficiency of traffic environment, we propose a Deep Reinforcement Learning-based Autonomous Intersection Management (DRLAIM) system. This method contains three main models, including brake-safe control model, intersection control model based on reinforcement learning and priority assignment model. We train the system to learn a good intersection control policy by interacting with traffic environment through reinforcement learning and the brake-safe control model is used to ensure the safety of each autonomous vehicle during travel. Experiment results show that after training using reinforcement learning, the throughput of intersection control model increased by 83%. In comparison with the Fast First Service (FFS) policy, the average waiting time of DRLAIM reduced by about 1.2% to 11.4%