34254 research outputs found
Sort by
The Compilation of Taiwan Fear and Greed Index and Its Application in Investment Strategy
No full textBased on the CNN Fear and Greed Index compilation method, this study compiled the Taiwan market's Fear and Greed Index with minor adjustments. The compilation result shows that the correlation coefficient between the self-made index and the original CNN Business version can reach 0.74 between January 2020 to May 2022.
Referring to the research results of Kastenhofer (2021), this study applied the Vector Autoregression model to the self-made Fear and Greed index for forecasting the market return and found that when the market sentiment is extreme fear or greed, the R^2 of the model increases significantly. Also, a leading trend can be found from the index to the market return. In addition, when the index shows that the market sentiment is greed, the lag period of significant response to market return is shorter, while the longer lag period shows this result when the market sentiment is fear.
In terms of the application in investment strategy, this study creates two strategies based on the market sentiment of the index. One is when the market shows a signal of extreme fear, the buying action will be taken continuously. As a result, both the return and Sharpe Ratio of this strategy are better than that of the dollar-cost averaging and buy-and-hold strategies. The other one is that this study uses greed as a signal to invest money in the market and uses extreme fear as a stop-loss signal, it can also have a good performance on return. Furthermore, after considering the results of the vector autoregression model to adjust the stop-loss timing, the strategy has a significantly higher return compared to the original one
Neural Network Accelerator Design Based On Multiple Precision Block Posit Number Representation
No full textRecently, a floating-point-like format, called posit, is proposed which can have larger representation range compared to IEEE-754 floating-point format. Since posit can use 8-bit or 16-bit to represent wide range of numbers, it can be applied to the training and inference of deep neural networks (DNN). In this thesis, we design a multi-precision posit DNN hardware that can support 8-bit and 16-bit posit format. Firstly, we analyze DNN models to determine the required posit bit-width in each layer. Then, similar to the concept of block floating-point, we adopt block posit design where a block of data samples shares the same exponent value in order to reduce the overhead of decoder/encoder hardware during data format conversion. Furthermore, the block posit DNN hardware requires only low-cost fixed-point arithmetic units, leading to significant reduction in area cost
Development of metal-induced Strain-Promoted Alkyne-Azide Cycloaddition
No full textMetal ions possess functions such as osmotic regulation, catalysis, metabolism, and signal transmission in the basic processes of life, and it is distributed in specific regions within tissues or cells based on their functions. However, metal ions must be obtained from the environment and it is possible to simultaneously ingest essential and toxic metal ions, resulting in an imbalance of metal ion level due to excessive uptake and insufficient metabolism, ultimately leading to the occurrence of cancer and other diseases. Formerly, macrocyclic molecules with cavities such as crown ether, aza crown ether, thio crown ether and porphyrin have been developed as acceptors of metal ions and their distribution in vitro cells and tissues has been visualized through fluorescent molecules. Moreover, monitoring in vivo is limited by numerous physiological environments and to overcome these shortcomings, bioorthogonal chemistry was introduced into the biological environment to conduct rapid and selective reactions without influence on the function of endogenous functional groups, including ring Strain Promoted Alkyne Azide Cycloaddition (SPAAC). However, the reactivity induced by high ring strain cannot be controlled in biosystem. Therefore, our study focus on the development of bioorthogonal reaction which can be controllable by metal ions in vivo. Metal ions gets chelated in the cavities of the macrocyclic molecules and shrink the molecule, which further induces ring strain on the molecule. The alkynyl groups on the ring simultaneously trigger the SPAAC and the distribution of metal ions can be visualized clearly through fluorescent molecules. However, construction of the high ring strain alkynyl structure is difficult due to its thermodynamic instability. Our previous work in the laboratory through intramolecular Sonogashira reaction for the final ring closure was futile due to the rigid sp2 and sp3 structures. In the current study, we modified our strategy and initiated the synthesis with Sonogashira reaction followed by acid amine coupling reaction for the macrolactamization reaction to furnish the expected molecule. Initially, we expected that the flexible C-C bond present in the close proximity of amine could react with the carboxylic acid in the counterpart to yield the expected product, but the results were not as expected. Subsequently, we modified the route of synthesis and initiated with peptide coupling followed by ring closing metathesis by introducing two vinyl groups in the system. The reaction conditions were ineffectual and failed to endow the target. We hope that in the future Nicholas reaction or Lindlar catalyst can be used to convert acetylene into ethene and bring the vinyl groups closer to increase the oppurtunity of macrolactamization and establish alkynyl macrocyclic molecules through bromination and elimination reaction
Optimal Circuit Design and Performance Verification of Fan-out Panel Level Package
No full textThe emergence of the fifth generation of mobile communication, accompanied by advancements in fields such as artificial intelligence (AI) and the Internet of Things (IoT), has led to a significant increase in the demand for integrated circuits (ICs). The demand for IC homogenous and heterogeneous integration has also grown, driving the continuous progression of fan-out advanced packaging toward large-scale packaging with multiple chips. Currently, fan-out panel-level packaging (FO-PLP) is a popular choice for development, offering technological advantages in terms of conducting packaging processes on large-area substrates and achieving chip I/O fan-out through the redistribution layer (RDL) process. This enables high IO density, and thin package thickness, and meets the production requirements for large-sized packaging. However, despite its numerous benefits, the large-area process presents challenges, particularly severe warpage at the package edges, raising doubts about whether chips integrated with this packaging can meet performance standards.
Therefore, this thesis aims to achieve a fan-out panel-level packaging design using an SSD controller chip. Considering the challenges posed by the panel-level process-induced warpage, a grid-based approach is adopted to replace the traditional solid plane, and a thorough analysis of the high-speed signal and power network within the package is established to evaluate the electrical requirements of the package routing. Through the analysis, it is confirmed that the impedance variation caused by the grid reference plane leads to signal integrity issues, preventing the signals from meeting the specified requirements. To address this, a design flow is implemented for optimizing the routing of the grid reference plane using flexible printed circuit boards with the same grid requirements. The completion of steps such as signal stack-up verification, grid structure design, grid reference plane routing comparison table, and optimization of the grid reference plane design helps establish an integrated equation that combines the grid dimensions, routing widths, and characteristic impedance. The accuracy of the equation and the optimization results of the grid reference plane routing are verified through RF probe measurements. Additionally, near-field measurements are conducted to validate the signal leakage situation of the grid reference plane, ensuring a comprehensive consideration of the factors in optimizing the grid reference plane design. Finally, the grid reference plane routing design flow is employed to achieve the optimal routing design of the SSD controller package under the requirements of the grid reference plane, meeting the signal specification requirements for the SSD controller package
Application of biodegradable hybrid surfactants to remediate weathered diesel oil contaminated soils
No full textDiesel oil was a hydrophobic liquid petroleum mixture, and its density was lower than that of water. It was classified as a Light non-aqueous phase liquid (LNAPL). When it was accidentally leaked and weathered for a long time, it caused weathered diesel oil pollution that was difficult to treat and biodegrade. One of the sources of total petroleum hydrocarbons (TPH) was weathered diesel oil. The long-term existence of TPH in the environment not only destroyed ecosystems but also had negative effects on animals and plants, such as carcinogenicity and neurotoxicity. In the study, a combination of lecithin, MES, Tween 80, and Triton X-100 was used as hybrid surfactants, and tests were conducted at different concentrations and with different methods of surfactants addition to find the optimal conditions. Lecithin was used as a dispersant to help emulsify oil, while MES, Tween 80, and Triton X-100 were used as solubilizing agents to promote the dissolution and diffusion of pollutants in water. The use of hybrid surfactants yielded better solubilization effects than using a single surfactant. Moreover, the solubilization effect of adding two chemicals in turn was better than that of adding them at the same time after premixing. The removal rate of adding lecithin first and then MES was 64.6%, while the removal rate of adding lecithin first and then MES was 37.7%. The former was 26.9% higher than the latter. Subsequently, The experiments that used the Taguchi method determined that the optimum ratio was 0.5% lecithin, 1% MES, and 7% butanol. Using this optimum ratio, a solubilization efficiency of 78.3% was observed in the column experiment simulating field conditions. In conclusion, the strategy of adding 0.5% lecithin diluted with 7% butanol first, followed by the addition of 1% MES, not only effectively treated weathered diesel-contaminated soil but also ensured that the remaining low concentrations of surfactants could be biodegraded and utilized, preventing secondary pollution and making it an environmentally friendly remediation approach
Prediction and Compensation Verification of Thermal Errors in a Machine Tool Spindle Based on Bi-LSTM
No full textThis study is based on the Bidirectional Long Short-Term Memory neural network
model and deep learning techniques to establish a thermal error compensation model for machine tool spindles. The compensation model utilizes four temperatures and spindle speed as input data to establish their relationship with thermal errors in the X, Y, and Z directions. The training and testing results of the model show that the predicted results have an RMSE below 0.7 \uc2\ub5m and error percentages below 12%, indicating a high level of prediction accuracy. The accuracy and robustness of the model's predictions were verified through experiments involving two sets of compound operating conditions. The experiments demonstrated that the model can maintain a certain level of prediction accuracy in the Z direction, with an error percentage of approximately 21-23%. However, during experimentation, it was discovered that errors in the X and Y directions would deviate due to the machine tool's distinct dynamic characteristics and uncertainties, leading to inaccurate predictions. To address this issue, the study proposes a deep learning approach that utilizes transfer learning. The approach involves fine-tuning and modifying the architecture of the pre-established Bi-LSTM model, which is a pre-trained model. Through experiments conducted over a continuous period of three days, we compare the performance of a pre-trained model with transfer learning and the performance of the same pre-trained model without transfer learning. The results reveal that transfer learning effectively improves predictions in both the X and Y directions, mitigating inaccuracies caused by variations in machine characteristics. Predictions in the Z direction remain at a similar level, with an error percentage of approximately 20%. The maximum error percentages in the X and Y directions have been significantly reduced, with reductions from 320% to 44% and from 173% to 80%, respectively. Moreover, the time required for data collection in the context of transfer learning is reduced by over 85%. Additionally, we compare the model prediction results when the data collection time for transfer learning is reduced from 2.5 hours to 1 hour. Furthermore, the study proposes integrating the Bi-LSTM model with OPC UA (OPC Unified Architecture) for real-time online thermal error compensation. After compensation, the compensation for the Z direction of the workpiece is reduced from approximately 17 \uc2\ub5m to 5 \uc2\ub5m, with an overall accuracy of approximately 62%. However, the compensation effect for the X and Y directions is not significant due to the small measurement errors observed. The actual results demonstrate that the proposed model is feasible and effective in improving machining accuracy
Design of Active Reconfigurable Intelligent Surface for Sub-6 GHz
No full textThis paper proposes a channel model for reconfigurable intelligent surfaces (RIS). The model takes into account the losses incurred by digital phase shifters in different phase states and considers the influence of antenna polarization, making it more suitable for practical wireless communication environments. This model enables quick validation of the gains achievable with RIS when multiple units are employed. Additionally, an active reconfigurable intelligent surface is introduced. The active RIS consists of radiating antennas, RF digital phase shifters, and low-noise amplifiers (LNA), with additional power splitters required in the RIS array. The commercially available chip, MAPS-01044, is used for the phase shifters. Compared to passive RIS architectures, the active RIS enhances overall communication efficiency at the cost of additional power consumption. Finally, the impact of the RIS structural mode on the overall communication strength is validated.
This paper conducts feasibility testing and practical implementation of the proposed active reconfigurable intelligent surface. An automated measurement platform is set up to measure the energy intensity and signal-to-noise ratio of the signal channels. This is done to validate the effectiveness of the architecture proposed in this paper
Occupancy Detection Using FMPSIL Radars
No full textThis thesis uses two different types of test statistics methods based on the generalized likelihood ratio, respectively arithmetic to geometric mean method and the signal-subspace eigenvalues method. These methods are applied to a frequency-modulated phase-and self-injection-locked radar operating in the 5.8 GHz ISM band. The radar system, integrated with the algorithms presented in this thesis, is used to detect the presence of moving objects in the environment, overcoming the limitations of other sensors with high false alarm rates and inability to detect stationary objects.
The frequency-modulated phase-locked radar with ranging capability is able to measure the distance between the radar and the moving objects. Furthermore, by applying a second statistical test, the algorithm can determine the presence of either one or two targets in the environment.
In terms of computation, the radar system's output signal only requires Fourier transform, covariance calculation, and eigenvalue decomposition. Compared to traditional methods, which often take two to three cycles or more, the algorithm used in this thesis only requires one cycle for detection. In human presence experiments, with sensing distances ranging from 3 to 5 meters, the detection accuracy reaches 95%, and the lowest false alarm rate is only 0.4%. With an average signal period of approximately 5 seconds, the required detection time is 5 seconds
A study of Factors that Influence the Purchase of bubble tea among Indians in Taiwan
No full textThe fad of bubble tea is trending around the world, especially among individuals belonging to Generation Z. Even though bubble tea is regarded as an unhealthy beverage, countries like India have also seen a spike in this trend. It is interesting to note that culturally Indians consume hot tea and the rise in this trend is unknown. This study aims to inquire into the factors that motivate Indian consumers' decisions to buy and consume bubble tea in Taiwan.
The finding of this study indicated factors such as price, quality, motivation, health consciousness, packaging and branding, food values, and Country of origin influencing purchasing decisions towards bubble tea in Taiwan. Therefore, the results of this study will thus assist Taiwanese bubble tea brands and other Taiwanese beverage brands in developing and planning strategies for entering the Indian beverage markets, where there is a huge market for foreign beverages as unexplored markets
Concentrations of five heavy metals and C, N stable isotopes in the tissues of four cetacean around Taiwan
No full textThis study collected four species of stranded cetaceans, including Kogia breviceps (pygmy sperm whale), Steno bredanensis (rough-toothed dolphin), Tursiops aduncus (Indo-Pacific bottlenose dolphin), and Tursiops truncatus (common bottlenose dolphin), from the water around Taiwan from 2003 to 2020. The samples were analyzed for the concentrations of iron, zinc, copper, manganese, and cadmium in these cetaceans' muscles, liver, kidneys, and lungs by flame and graphite furnace atomic absorption spectrometry. The study aimed to investigate the variations of heavy metal concentrations and tissue differences among the cetaceans, as well as their correlation with body length. Additionally, carbon (\uce\ub413C) and nitrogen (\uce\ub415N) isotope analysis was conducted on the muscles to explore differences in habitat and diet.
Based on the carbon and nitrogen isotope results, it was found that K. breviceps and S. bredanensis primarily inhabit the continental shelf and deep-sea regions, and K. breviceps with a diet mainly consisting of deep-sea cephalopods. Regarding heavy metal concentrations, there were accumulation differences among the species. Fe concentrations in the muscle and kidney tissues of K. breviceps (613\uc2\ub1189 and 880\uc2\ub1277 mg/kg dry weight, respectively) were higher than the other three species, likely due to their diving capabilities and physiological differences, resulting in varied requirements and utilization of heavy metals. The study revealed that Fe, Zn, Cu, and Mn elements were primarily accumulated in the liver tissues of the four cetacean species, while Cd was found to accumulate in the kidney tissues. Furthermore, positive correlations were observed between Fe concentrations in the liver and kidney tissues of K. breviceps, as well as Cd concentrations in the muscle tissue. Conversely, negative correlations were observed between Zn concentrations in muscle tissue, Mn concentrations in kidney tissue of K. breviceps, and Mn concentrations in the muscle tissues of T. truncatus and S. bredanensis, respectively.
The accumulation patterns of these five heavy metals varied among the four cetacean species, likely due to their physiological requirements and different age-related elemental demands. In comparison to other marine areas in recent years, the concentrations of heavy metals in the coastal waters of Taiwan did not show a higher pollution trend