體重、BMI和年紀是衡量該洗腎病人體力狀況的三種非常重要的參數，給洗腎患者量測體重的工具是坐式體重機。本研究實驗設計是使用Azure Kinect DK(K)對患者從走向坐式體重機到量完體重離開的完整過程做全身及步態分析並萃取其全身骨架動態特徵，實驗分析出的結果可以拿來量化相同病症之患者的體力狀況，或是某些跟行走困難有關的病徵，設計了跌倒風險值(Risk of falling, Rf)來預警跌倒狀況，也設計最小下巴高度(Minimum Chin Height, Ch)和駝背量(Humpbacked value, HV)預警中風狀況。為了確認K所量測的數據之信效度，針對實驗環境中的長寬高(Length、Width、Height, L、W、H)實際值(Actual Value, AV)與K所算出的L、W、H做比較，L的誤差範圍是-0.87cm ~ 1.97cm，W誤差範圍是-1.79cm ~ 1.12cm，H誤差值範圍是0.09cm~ 2.55cm，並使用比約30幀率（Frames per second, Fps）K更高的約90Fps市售產品Intel RealSense Depth Camera D435(D435)錄下深度影像，算出一個步伐(Stride, S)及其所需步時所換算出的步速(Pace, P)，比對K所算出的結果得出兩者差值。一位男性受試者模擬各種行走姿態，得出K及D435兩者比較的P差值變化範圍是1.6561cm/s ~ 2.7460cm/s、S差值變化範圍是-3.3756cm ~ 0.7862cm、起坐時間(Duration of stand up to sit down, DSS)差值變化範圍是-0.15s ~ -0.04s。而K也與市售產品Zeno Walkway(ZW)進行S、P與重心(Gravity Center, GC)參數的比較，其中ZW重心是二維重心(Two Dimensions (2D) GC, 2GC)，K須將其三維重心(Three Dimensions (3D) GC, 3GC)投影成二維，兩者量測的S大小差之變化範圍是-10.008cm ~ 5.438cm、P差之變化範圍是5.033cm/s ~ 8.371cm/s、2GC差之變化範圍是0.0001cm ~ 38.0749cm 。通過以上三種(AV、D435、ZW)取證以確定本K系統基於GPU所量測的數據值及其所計算出的參數值的信效度是可以滿足本研究的臨床應用需求。
Sunneng Sandino Berutu
IoT Power Monitor System Development and Using Deep Learning Technology for Disturbances Classification
陳永欽 YEONG-CHIN CHEN
The power monitoring system is essential in retaining the quality of the electrical equipment and help the consumer in reducing the cost of electricity usage. Power disturbances are a problem that requires to be addressed in order to maintain the quality of electrical appliances. The causes of this issue include unstable loads and the power environment.
The Internet of Things (IoT) has been widely used for power monitoring systems. This technology assists users in obtaining information in real-time. Another advantage is the low cost of developing a system based on the IoT. However, based on surveying the existing system, the process of data sending and the functional complexity of the system are required to develop. Therefore, the author developed a system by introducing the data transfer technique and involving several algorithms to monitor the power. This system can measure electricity usage, power factor, fundamental frequency, and power disturbances. In addition, this system can also filter out the harmonic and noise.
Many researchers have adopted a deep learning technology to detect and classify the power quality disturbances (PQDs) problem. However, the author found several shortages such as a high computation time cost and preserving all the original information in a two-dimension (2D) image. Therefore, the author proposed two approaches in solving these shortages. The first is to introduce data processing by employing the wavelet transform (WT) method to reduce the computation time cost in the training phase. This study used a one-dimension (1D) deep convolutional neural networks (CNNs) algorithm in the PQDs classification task. In this section, the synthetic disturbance signals were generated from the mathematical formulas. Then the signal was compressed using the WT method. Next, a dataset contains the compressed PQDs signal was fed into the model. The second section introduces a new approach for data preprocessing in classifying the PQDs. The 2D deep CNNs algorithm was employed for the classification task. To fulfill the requirement in utilizing this algorithm, the PQDs signals were converted into the 2D image. Therefore, this study had proposed a new approach to a signal to an image. In this approach, the signal was divided into several cycles by adopting the zero-cross algorithm. Then, each cycle was transformed into a matrix. The matrices were combined into a new matrix. Finally, this matrix was converted into the 2D image grayscale. The 2D image resulted could retain the waveform and all the information of the signal.
The experiment results show that the system can measure the parameters of power monitoring. The proposed system with a low-cost development had a similar performance to the measurement devices made in the factory. Furthermore, the experiment results in classifying the PQDs using the 1D deep CNN section have demonstrated that the compressed disturbance signals reduced the computation time cost at the training phase and reached a high accuracy at the testing phase. Meanwhile, the evaluation of the PQDs classification utilizing the 2D deep CNN has shown that the proposed approach performance was superior to the 1D deep CNN and the existing approaches. For further study, the author aims to develop and examine the combination of the IoT and deep learning technology in the identification and classification of the PQDs using the actual data input.
Performance Analysis of Reconfigurable Intelligent Surfaces Integrated with Non-Orthogonal Multiple Access Technique and Cognitive Radio Networks
DINH-THUAN DO DINH-THUAN DO
The future of 5th Generation and 6th Generation wireless networks is inescapable. Every second of human life depends on technology. Every day, millions of wireless IoT devices are getting activated and getting connected to wireless networks. That requires the spectrum efficiency and the bandwidth to the users should be increased for seamless connectivity considering low latency and being highly reliable. But this is not possible in traditional techniques that are been used till now. The propagation medium plays a major role in the field of communication networks. The strength of the signal transmitted depends on the properties of the medium. In a practical scenario of wireless communications, the strength of the transmitted signal depreciates during the transmission. Therefore, the quality of the signal received will not be the same as transmitted. One of the primary reasons for this situation is the rigid obstacles that are present between the source and destination.
The recent invention of Reconfigurable Intelligent Surfaces (RIS) has opened wide opportunities for explorers to learn how to control the radio waves in any propagation medium. The RIS can help eradicate the adverse consequences of natural wireless propagation. Non-Orthogonal Multiple Access (NOMA) can provide multiple user connectivity using different power coefficients in the identical time, and frequency. Depending on the position of the user, the power levels are allocated, and the weak user gets more power compared to the strong user, enabling the best communication possible. The important point to be noted is that this technique can handle massive device connectivity while providing efficient spectrum utilization, low latency, and high reliability.
Cognitive radio networks have been a great addition to the recent technologies as their ability to enhance spectrum utilization and sharing of networks among two groups, primary users, and secondary users. Though the working of the CRN proves to be efficient enough, there are major issues that need to be addressed like limiting the interference caused due to the two networks and power consumption during the transmission.
Therefore, motivated by these potential emerging technologies, in this research, I study the integration of CRN with RIS-aided NOMA networks and the RIS-aided NOMA system. The analytical expression has been derived for the outage performance and throughput of the two proposed systems. In the CRN-assisted NOMA-aided RIS network, the primary findings involve the analysis of the outage probability and throughput of the system. Whereas in NOMA-aided RIS system, the primary findings involve analysis of outage performance of the users individually and throughput of the system, and the effect of role of the number of meta-surfaces in RIS, transit power at the base station, and power allocation factors of the NOMA users. Results yield that implementation of CR in RIS-aided NOMA system proves to efficient while considering the interference caused from the primary network. Whereas, from traditional RIS-aided NOMA system, it is manifested that number of meta-surfaces and transmit SNR play a vital role in enhancing the system performance. Both the proposed systems have proven itself that NOMA outperforms the traditional Orthogonal Multiple Access (OMA) technique.
Analysis of Autoencoder's Features From Addiction Data of Questionnaire
蔡志仁 ZHI-REN TSAI;呂威甫 WEI- FU LU
The aim of the study is to provide an evaluation for the features of four addiction types from different three student targets by unsupervised learning techniques and the importance of four cases of a questionnaire. However, the interview of questionnaires method takes a too long time. To speed up the diagnosis process of four addictions we choose unsupervised learning. First, the qualitative indexes are separated by expert threshold scoring. We developed a prediction model of an autoencoder features generator plus an XGBoost cclassifier for four kinds of addictions by using famous terms in technology
i.e., Online game case, Facebook case, smartphone case, and Internet case. So, the research to solve How to sort the different importance of four cases of questionnaires for the different three student targets to obtain the intervention priorities of identifying the four addiction people. In this method, to further determine or to verify which case of addiction the questionnaire is the most important for the different three targets, we conducted research in the three groups of elementary school students, junior high school students, and senior high school students as three targets by giving the four addictions of the psychological questionnaire, respectively. Finally, the answers are obtained automatically through AI self-learning way. The experiment shows that the self-learning method could generate the
intervention priorities of three targets precisely and effectively. Moreover, the visualization of autoencoder’s features can easily evaluate the original four types of questionnaires design about the four addictions to observe the human behaviors in this experiment.
Keywords: online game, Facebook, smartphone, Internet, autoencoder, XGBoost.
Distributed Processing of Skyline Queries and the Applications for Upgrading Products to Maximize Profits
陳良弼 CHEN, ARBEE L. P.
A skyline query determines the data points in a dataset that are not dominated by others. It is widely used for many applications which require multi-criteria decision-making. However, skyline query processing is considerably time-consuming for a high-dimensional large-scale dataset. This study consists of two main tasks. The first is to design an efficient parallel computing technique to address the computational problem of skyline queries for large high-dimensional datasets. It is based on MapReduce frameworks to process large datasets. The second study focuses on the recommendation of upgrading products based on the skyline data points.
A large number of efficient MapReduce skyline algorithms have been proposed in the literature. However, there are still opportunities for further parallelism. Our method to divide a large dataset is called by LShape partitioning strategy and we propose an effective filtering algorithm named propagation filtering. We verify that our algorithms outperformed the state-of-the-art approaches by extensive experiments, especially for high-dimensional large-scale datasets.
The manufacturer often needs to make a proper decision to gain maximal profits from the product upgrading. The goal of upgrading products is to maximize the profit by increasing the total number of expected customers with a certain upgrading cost. Our algorithms in this study are based on the dominating relationships among products. It has been proved that finding the dominating relationship of products with three or more features is NP-hard. We first propose an optimal algorithm for upgrading a single product and then modify it to be an efficient heuristic algorithm with a percentage error approaching 20%. We also extend this heuristic algorithm for simultaneously upgrading multiple products.