Rozprawy doktorskie na temat „ENSEMBLE EMPIRIAL MODE DECOMPOSITION”

Orientador : Luiz Eduardo S. Oliveira<br>Coorientador : David Menotti<br>Dissertação (mestrado) - Universidade Federal do Paraná, Setor de Ciências Exatas, Programa de Pós-Graduação em Informática. Defesa: Curitiba, 20/07/2017<br>Inclui referências : f. 63-72<br>Resumo: As séries temporais financeiras são notoriamente difíceis de analisar e prever dada sua natureza não estacionária e altamente oscilatória. Nesta tese, a eficácia da técnica de decomposição não-paramétrica Ensemble Empirical Mode Decomposition (EEMD) é avaliada como uma técnica de extração de característica de séries temporais provenientes de índices de mercado e taxas de câmbio, características estas usadas na classificação, juntamente com diferentes modelos de aprendizado de máquina, de tendências de curto prazo. Os resultados obtidos em dois datasets de dados financeiros distintos sugerem que os resultados promissores relatados na literatura foram obtidos com a adição, inadvertida, de lookahead bias (viés) proveniente da aplicação desta técnica como parte do pré-processamento das séries temporais. Em contraste com as conclusões encontradas na literatura, nossos resultados indicam que a aplicação do EEMD com o objetivo de gerar uma melhor representação dos dados financeiração, por si só, não é suficiente para melhorar substancialmente a precisão e retorno cumulativo obtidos por modelos preditivos em comparação aos resultados obtidos com a utilização de series temporais de mudanças percentuais. Palavras-chave: Predição de Tendencias, Aprendizado de Máquina, Séries Temporais Financeiras.<br>Abstract: Financial time series are notoriously difficult to analyse and predict, given their nonstationary, highly oscillatory nature. In this thesis, the effectiveness of the Ensemble Empirical Mode Decomposition (EEMD) is evaluated at generating a representation for market indexes and exchange rates that improves short-term trend prediction for these financial instruments. The results obtained in two different financial datasets suggest that the promising results reported using EEMD on financial time series in other studies were obtained by inadvertently adding look-ahead bias to the testing protocol via pre-processing the entire series with EEMD, which do affect the predictive results. In contrast to conclusions found in the literature, our results indicate that the application of EEMD with the objective of generating a better representation for financial time series is not sufficient, by itself, to substantially improve the accuracy and cumulative return obtained by the same models using the raw data. Keywords: Trend Prediction, Machine Learning, Financial Time Series.

With the development of mechanical fault diagnosis technology, complex mechanical systems do not need to be shut down periodically for the maintenance. The working condition of the mechanical systems can be monitored by analyzing the wear metal particles in the systems' lubricating oil. However, the output signals of the monitoring sensor are non-stationary. In some case the particle signals are overlapped with each other. The goal of this thesis is to find a method to decompose those overlapped particle signals, and then count the particle number in the lubricating oil. At the beginning EMD method was introduced in the experiment because of the character of the sensor signals. In this project, because EMD method is sensitive to the noise in the original signals, an improved version of EMD, EEMD method was implemented. Finally, a post processing method was used to get a better result.

Respiratory sounds (RS) are produced by turbulent airflows through the airways and are inhomogeneously transmitted through different media to the chest surface, where they can be recorded in a non-invasive way. Due to their mechanical nature and airflow dependence, RS are affected by respiratory diseases that alter the mechanical properties of the respiratory system. Therefore, RS provide useful clinical information about the respiratory system structure and functioning. Recent advances in sensors and signal processing techniques have made RS analysis a more objective and sensitive tool for measuring pulmonary function. However, RS analysis is still rarely used in clinical practice. Lack of a standard methodology for recording and processing RS has led to several different approaches to RS analysis, with some methodological issues that could limit the potential of RS analysis in clinical practice (i.e., measurements with a low number of sensors, no controlled airflows, constant airflows, or forced expiratory manoeuvres, the lack of a co-analysis of different types of RS, or the use of inaccurate techniques for processing RS signals). In this thesis, we propose a novel integrated approach to RS analysis that includes a multichannel recording of RS using a maximum of five microphones placed over the trachea and the chest surface, which allows RS to be analysed at the most commonly reported lung regions, without requiring a large number of sensors. Our approach also includes a progressive respiratory manoeuvres with variable airflow, which allows RS to be analysed depending on airflow. Dual RS analyses of both normal RS and continuous adventitious sounds (CAS) are also proposed. Normal RS are analysed through the RS intensity–airflow curves, whereas CAS are analysed through a customised Hilbert spectrum (HS), adapted to RS signal characteristics. The proposed HS represents a step forward in the analysis of CAS. Using HS allows CAS to be fully characterised with regard to duration, mean frequency, and intensity. Further, the high temporal and frequency resolutions, and the high concentrations of energy of this improved version of HS, allow CAS to be more accurately characterised with our HS than by using spectrogram, which has been the most widely used technique for CAS analysis. Our approach to RS analysis was put into clinical practice by launching two studies in the Pulmonary Function Testing Laboratory of the Germans Trias i Pujol University Hospital for assessing pulmonary function in patients with unilateral phrenic paralysis (UPP), and bronchodilator response (BDR) in patients with asthma. RS and airflow signals were recorded in 10 patients with UPP, 50 patients with asthma, and 20 healthy participants. The analysis of RS intensity–airflow curves proved to be a successful method to detect UPP, since we found significant differences between these curves at the posterior base of the lungs in all patients whereas no differences were found in the healthy participants. To the best of our knowledge, this is the first study that uses a quantitative analysis of RS for assessing UPP. Regarding asthma, we found appreciable changes in the RS intensity–airflow curves and CAS features after bronchodilation in patients with negative BDR in spirometry. Therefore, we suggest that the combined analysis of RS intensity–airflow curves and CAS features—including number, duration, mean frequency, and intensity—seems to be a promising technique for assessing BDR and improving the stratification of BDR levels, particularly among patients with negative BDR in spirometry. The novel approach to RS analysis developed in this thesis provides a sensitive tool to obtain objective and complementary information about pulmonary function in a simple and non-invasive way. Together with spirometry, this approach to RS analysis could have a direct clinical application for improving the assessment of pulmonary function in patients with respiratory diseases.<br>Los sonidos respiratorios (SR) se generan con el paso del flujo de aire a través de las vías respiratorias y se transmiten de forma no homogénea hasta la superficie torácica. Dada su naturaleza mecánica, los SR se ven afectados en gran medida por enfermedades que alteran las propiedades mecánicas del sistema respiratorio. Por lo tanto, los SR proporcionan información clínica relevante sobre la estructura y el funcionamiento del sistema respiratorio. La falta de una metodología estándar para el registro y procesado de los SR ha dado lugar a la aparición de diferentes estrategias de análisis de SR con ciertas limitaciones metodológicas que podrían haber restringido el potencial y el uso de esta técnica en la práctica clínica (medidas con pocos sensores, flujos no controlados o constantes y/o maniobras forzadas, análisis no combinado de distintos tipos de SR o uso de técnicas poco precisas para el procesado de los SR). En esta tesis proponemos un método innovador e integrado de análisis de SR que incluye el registro multicanal de SR mediante un máximo de cinco micrófonos colocados sobre la tráquea yla superficie torácica, los cuales permiten analizar los SR en las principales regiones pulmonares sin utilizar un número elevado de sensores . Nuestro método también incluye una maniobra respiratoria progresiva con flujo variable que permite analizar los SR en función del flujo respiratorio. También proponemos el análisis combinado de los SR normales y los sonidos adventicios continuos (SAC), mediante las curvas intensidad-flujo y un espectro de Hilbert (EH) adaptado a las características de los SR, respectivamente. El EH propuesto representa un avance importante en el análisis de los SAC, pues permite su completa caracterización en términos de duración, frecuencia media e intensidad. Además, la alta resolución temporal y frecuencial y la alta concentración de energía de esta versión mejorada del EH permiten caracterizar los SAC de forma más precisa que utilizando el espectrograma, el cual ha sido la técnica más utilizada para el análisis de SAC en estudios previos. Nuestro método de análisis de SR se trasladó a la práctica clínica a través de dos estudios que se iniciaron en el laboratorio de pruebas funcionales del hospital Germans Trias i Pujol, para la evaluación de la función pulmonar en pacientes con parálisis frénica unilateral (PFU) y la respuesta broncodilatadora (RBD) en pacientes con asma. Las señales de SR y flujo respiratorio se registraron en 10 pacientes con PFU, 50 pacientes con asma y 20 controles sanos. El análisis de las curvas intensidad-flujo resultó ser un método apropiado para detectar la PFU , pues encontramos diferencias significativas entre las curvas intensidad-flujo de las bases posteriores de los pulmones en todos los pacientes , mientras que en los controles sanos no encontramos diferencias significativas. Hasta donde sabemos, este es el primer estudio que utiliza el análisis cuantitativo de los SR para evaluar la PFU. En cuanto al asma, encontramos cambios relevantes en las curvas intensidad-flujo yen las características de los SAC tras la broncodilatación en pacientes con RBD negativa en la espirometría. Por lo tanto, sugerimos que el análisis combinado de las curvas intensidad-flujo y las características de los SAC, incluyendo número, duración, frecuencia media e intensidad, es una técnica prometedora para la evaluación de la RBD y la mejora en la estratificación de los distintos niveles de RBD, especialmente en pacientes con RBD negativa en la espirometría. El método innovador de análisis de SR que se propone en esta tesis proporciona una nueva herramienta con una alta sensibilidad para obtener información objetiva y complementaria sobre la función pulmonar de una forma sencilla y no invasiva. Junto con la espirometría, este método puede tener una aplicación clínica directa en la mejora de la evaluación de la función pulmonar en pacientes con enfermedades respiratorias

Combustion instability due to feedback coupling between unsteady heat release and natural acoustic modes can cause catastrophic failure in liquid rocket engines and to predict and prevent these instabilities the mechanisms that drive them must be further elucidated. With this goal in mind, the objective of this thesis was to develop techniques that improve the understanding of the specific underlying physical processes involved in these driving mechanisms. In particular, this work sought to develop a small-scale, optically accessible liquid rocket engine simulator and to apply modern, high-speed diagnostic techniques to characterize the reacting flow and acoustic field within the simulator. Specifically, high-speed (10 kHz), simultaneous data were acquired while the simulator was experiencing a 170 Hz combustion instability using particle image velocimetry, OH planar laser induced fluorescence, CH* chemiluminescence, and dynamic pressure measurements. In addition, this work sought to develop approaches to reduce the large quantities of data acquired, extracting key physical phenomena involved in the driving mechanisms. The initial data reduction approach was chosen based on the fact that the combustion instability problem is often simplified to the point that it can be characterized by an approximately linear constant coefficient system of equations. Consistent with this simplification, the experimental data were analyzed by the dynamic mode decomposition method. The developed approach to apply the dynamic mode decomposition to simultaneously acquired data located a coupled hydrodynamic/combustion/acoustic mode at 1017 Hz. On the other hand, the dynamic mode decomposition's assumed constant operator approach failed to locate any modes of interest near 170 Hz. This led to the development of two new data analysis techniques based on the dynamic mode decomposition and Floquet theory that assume that the experiment is governed by a linear, periodic system of equations. The new periodic-operator data analysis techniques, the Floquet decomposition and the ensemble Floquet decomposition, approximate, from experimental data, the largest moduli Floquet multipliers, which determine the stability of the periodic solution trajectory of the system. The unstable experiment dataset was analyzed with these techniques and the ensemble Floquet decomposition analysis found a large modulus Floquet multiplier and associated mode with a frequency of 169.6 Hz. Furthermore, the approximate Rayleigh criterion indicated that this mode was unstable with respect to combustion instability. Overall, based on the positive finding that the ensemble Floquet decomposition was able to locate an unstable combustion mode at 170 Hz when the operator's time period was set to 1 ms, suggests that the dynamic mode decomposition based 1017 Hz mode parametrically forces the 170 Hz mode, resulting in what could be characterized as a parametric combustion instability.

碩士<br>國立臺灣師範大學<br>數學系<br>100<br>Ensemble Empirical Mode Decomposition (EEMD) is an adaptive time-frequency data analysis method. Time series or signals can be decomposed into a collection of intrinsic mode functions (IMFs). Nevertheless, there appears a multi-mode problem where signals with a similar time scale are decomposed into different IMFs. A possible solution to this problem is to combine the multi-modes into a proper single mode, but there is no general rule on how to combine IMFs in the literature. In this paper, we propose to modify EEMD algorithm using the statistical clustering analysis and to provide a framework to combine the IMFs into a condensed set of clustered intrinsic mode functions (CIMFs). The method is applied to two artificially synthesized signals, wind turbine signal at Chunan Miaoli, and a seismic signal during the earthquake at Chi-Chi in 1999. Especially, this seismic signal contains not only the main seismic information but also the seismic motion from a landslide in Tsaoling area. The present method can separate the two signal from different sources correctly, and these applications of other examples demonstrate that, the present method offers great improvement over EEMD for extracting useful information.

碩士<br>國立成功大學<br>航空太空工程學系<br>107<br>The ensemble empirical mode decomposition (EEMD) method is applied for wind data analysis in the current research. However, calculations could take a very long time. Therefore, an attempt is made to accelerate the calculations. MATLAB and Python are used to explore the characteristics of different programming language operations, and a user-friendly graphical interface is also developed, and the execution process will be operated automatically and continuously. The wind data analyzed were collected by the wind turbines located on campus of Case Western Reserve University in the United States. The wind data have been collecting since 2012 and the amount of data keeps growing. Thus, reducing the analyzing time is important. This study not only wants to use the graphics processor to try to shorten the time required for the operation process, but also finds the approximation trend in EEMD and refines the algorithm to shorten the operation time by about 65%.

碩士<br>國立陽明大學<br>醫學工程研究所<br>99<br>In this study, by Huang NE, who proposed the complementary empirical mode decomposition method based on the empirical mode decomposition method. The algorithm inherits the empirical mode decomposition some of the benefits. For example it can solve nonlinear and nonstationary signals. Another Empirical mode decomposition method and the traditional numerical methods, the base need to know prior to analysis, the original signals can be based on the characteristics and nature of the original signal automatically a number of nature apart from the function and the remaining functions for analysis. Complementary empirical mode decomposition method also improved the empirical mode decomposition shortcomings. For example the intermittent signals, noise and the edge of the original signal processing have problems in Empirical mode decomposition method, so Complementary empirical mode decomposition method can solve problems by white noise. Another joining the white noise may affect the analysis of the original signal, so the Complementary empirical mode decomposition method are more general and increase the number of complementary concepts to minimize the impact of white noise, the results of analysis tends to positive solutions. Complementary empirical mode decomposition method can decompose the image into several graphic signals without mode mixing, linear and stable intrinsic mode functions, computing the signal waveform after the intrinsic mode functions to comply with the conditions. Intrinsic mode functions to establish original signal can understand some of original signal information, the remaining factor is the combination of the characteristics of the original image graph. One-dimensional empirical mode decomposition method have a lot of literature to proof of adaptability and it can be used in the ability of nonlinear and nonstationary signals. In this study, the two-dimensional empirical mode decomposition method, the application in image process can be obtained from the original number of information. Local feature in the original image signal area, the edge structure and light shading. Finally, to get the information to do with Bilateral Filter processing, a new method to improve noise issues.

碩士<br>國立成功大學<br>土木工程學系<br>105<br>Non-destructive testing (NDT) has the advantages of economical, fast, and a wide range of applicability. Therefore it is a fairly effective testing method. The testing of piles without pile cap has been considered a well-developed technique. However, affected by complex boundaries conditions, testing of piles with pile cap is more difficult. This study use Impulse response (IR) method for the field test, then use ensemble empirical mode decomposition (EEMD) by Hilbert-Huang Transform to decompose the signals, finally use Fast Fourier Transform (FFT) to transform time domain data to frequency domain data. The purpose of this study is to separate the pile cap, pile bottom, and defect frequency and try to estimate the pile length and defect locations of piles with pile cap. The results show that the error of the estimation of pile length is less than 6%, and the errors of estimating top rectangular-defect were less than 3%. However, this method was less effective to detect circularity-defect and bottom rectangular-defect.

碩士<br>國立東華大學<br>電機工程學系<br>100<br>Cognitive radio(CR) is a novel smart wireless communication technology. To improve spectrum utilization efficiency, CR can detect the communication environment features to automatically adjust the transmitting and receiving parameters of a system, such as power, frequency, and modulation mode etc. The spectrum sensing is a key technology in the CR. In this thesis, we use energy detection for spectrum sensing. Many researches on energy detection are performed on the basis of a precise knowledge of noise power environment to set threshold. In this thesis we use ensemble empirical mode decomposition(EEMD) to decompose signal into licensed users and noise in the noisy environment. The signal power of licensed users and noise power are estimated in the current channel environment. An adaptive threshold is set in an unknown noise environment to improve the performance of energy detector.

碩士<br>國立臺灣科技大學<br>電子工程系<br>99<br>The conventional microphone array near-field Fourier acoustic holography using Fast Fourier Transform (FFT) is able to efficiently reconstruct sound field and acquire an image of noise distribution. However, Fourier transform causes measuring error in practical applications, and people have to select primary frequency for observing sound field holography based on the spectrum of source signal. In this thesis, we use the ensemble empirical mode decomposition (EEMD) owing to its adaptive basis and low mode-mixing, which are able to decompose multiple sound sources in the time domain and acquire instantaneous frequencies by intrinsic mode functions (IMFs). Prior information about the primary frequency is not necessary by this approach that makes the simultaneous observation of each source possible. In addition, EEMD sound source imaging approach may be integrated into near-field equivalent source imaging (NESI) system, which includes a virtual microphone technology generally used for sound field image enhancement. We have implemented and compared both non-stationary sound field spatial transform system and EEMD near-field sound source separating system in Labview language. Finally, several experimental results and detailed discussions are also provided to verify the characteristics of spherical sound source.

碩士<br>朝陽科技大學<br>資訊工程系碩士班<br>99<br>Muscular endurance is not only an indispensable part of physical fitness, but also an important indicator for people’s health. The purpose of this study was to detect the muscle fatigue with the change of the median frequency of the electromygram (EMG) signal measured in exercise. Ten healthy students (five males and five females) were selected as subjects. All of them went a one month sports. A self designed device was used to measure and analytics the EMG signal of vastus lateralis. Three signal decomposition technologies, Wavelet Decomposition, Empirical Mode Decomposition (EMD) and Ensemble Empirical Mode Decomposition (EEMD), were used to find the optimum spectrum band that had the most sensitivity and stability for the phenomenon of muscle fatigue. The results show that the intrinsic mode function (IMF) 1 of EEMD has the best performance.

碩士<br>國立臺灣大學<br>工程科學及海洋工程學研究所<br>99<br>A signal analyzing method, Hilbert-Huang Transform (HHT), was proposed by Norden E. Huang et al. in 1998. By using Emprical Mode Decomposition (EMD), signal could be decomposed into a finite number of intrinsic mode functions (IMFs) based on the local characteristic time-scale of the signal. Devoting these IMFs with Hilbert Transform could obtain meaningful instantaneous information about the signal. In this thesis, Ensemble Empirical Mode Decomposition (EEMD) and the post-processing of EEMD that were improved from the original EMD were involved to reduce the noise contained in the signal. By using the characteristic of EMD, the "Mutual Information" by calculating the entropy of signal from Independent Component Analysis was used to reduce the noisy component at first filtering, and a threshold-filtering selection method adapted to IMFs filtered the signal at second try. Adaptive Center-Weighted Mean Filter was then used to reduce the rest noisy component in the signal. Such attempting of triple-filtering could success removing most noisy component inside the IMFs that was generated by post-processing of Ensemble Empirical Mode Decomposition. The proposed method was tested by 4 test signals and 2 voice signals added with various level of noise under simulation experiment. From the simulation result, compared with wavelets and other existing method, the proposed method had better performance of de-noising in low SNR circumstances. The proposed method could retain more information of the signal with less destruction in de-noising process, and take into account the noise reduction with a better robustness and stability.

碩士<br>國立高雄大學<br>資訊工程學系碩士班<br>100<br>The main purposes of this study were to enhance and improve the speech recognition rate of speech recognition systems subject to some environment noise. In our research, we used Ensemble Empirical Mode Decomposition (Ensemble EMD) to decompose the speech signals with noise to several IMFs, and then find the best weights for each IMF by using real-coded genetic algorithm. Thereafter, the speech signals were recovered by summing the weighted IMFs to reduce the effect of the noise. Since the Ensemble EMD will take much computation time, a parallel computation algorithm under multi-core structure is proposed to speed up the computation of Ensemble EMD. We used parallel instruction coding in the OpenMP library to implement our algorithm.

ElectroCardioGram (ECG) signal records electrical conduction activity of heart. These are very small signals in strength with narrow bandwidth of 0.05-120 Hz. Physicians especially cardiologist use these signals for diagnosis of the heart’s condition or heart diseases. ECG signal is contaminated with various artifacts such as Power Line Interference (PLI), Patient–electrode motion artifacts, Electrode-pop or contact noise, and Baseline Wandering and ElectroMyoGraphic (EMG) noise during acquisition. Analysis of ECG signals becomes difficult to inspect the cardiac activity in the presence of such unwanted signals. So, de-noising of ECG signal is extremely important to prevent misinterpretation of patient’s cardiac activity. Various method are available for de-noising the ECG signal such as Hybrid technique, Empirical Mode Decomposition, Un-decimated Wavelet Transform, Hilbert-Hung Transform, Adaptive Filtering, FIR Filtering, Morphological Filtering, Noise Invalidation Techniques, Non- Local Means Technique and S-Transform etc. All these techniques have some limitations such as mode mixing problem, oscillation in the reconstructed signals, reduced amplitude of the ECG signal and problem of degeneracy etc. To overcome the above mentioned limitations, a new technique is proposed for denoising of ECG signal based on Genetic Algorithm and EEMD with the help of Fuzzy Thresholding. EEMD methods are used to decompose the electrocardiogram signal into true Intrinsic Mode Functions (IMFs).Then the IMFs which are ruled by noise are automatically determined using Fuzzy Thresholding and then filtered using Genetic Particle Algorithms to remove the noise. Use of Genetic Particle Filter mitigates the sample degeneracy of Particle Filter (PF).EEMD is used in this thesis instead of EMD because it solves the EMD mode mixing problem. EEMD represents a major improvement with great versatility and robustness in noisy ECG signal filtering.

碩士<br>國立臺灣大學<br>電子工程學研究所<br>101<br>Cardiovascular disease has been listed as the second rank of the top ten leading causes of death. Electrocardiogram(ECG) has played an important role and has been widely used clinically because it is a non-invasive, real-time, quick and easy-to-implement technique. Cardiovascular disease was diagnosed traditionally by inspection from doctors. For doctors, ECG noise can be easily ignored by visual inspection. Nevertheless, with the advance of science and technology, remote monitoring and diagnosis have become important processes to automatically detecting cardiovascular disease. However, in holter devices, ECG recordings are often corrupted by artifacts in some real practice, such as 50/60Hz power line interference, muscle contraction induced electromyogram(EMG), movement(or breath) induced baseline wandering or motion artifact. These aforementioned noises might result in misleading ECG detection. Thus, pre-processing of ECG noise is a very important task in such ECG analysis systems. In this thesis, an effective approach to eliminate baseline wander and EMG noise from ECG based on modified moving average filter and ensemble empirical mode decomposition (EEMD) was proposed. Modified moving average filter is used to eliminate ECG base line drift. It can be viewed as a pre-processing of the EEMD-based EMG reduction method. If data is interfered by EMG noise, EEMD is first used to decompose ECG data into different frequency components. By combination of proper QRS detection algorithms, only noise part will be extracted without affecting QRS complex or other ECG component. Finally, EMG noise can be estimated and removed from original ECG data. Then, by moving variance detection method, EMG positions can be detected and marked as reference to users. Cross correlation coefficient (Corr-Coef), percentage root-mean-square difference (PRD) and ECG morphology were used to examine the artificial data performance of proposed algorithm. Results showed that proposed de-noising framework successfully eliminate baseline wander and EMG interferences without significantly distorting the ECG waveform.

博士<br>國立臺灣大學<br>工程科學及海洋工程學研究所<br>107<br>An image processing technique, which is based on 2-D ensemble empirical mode decomposition, is developed to decompose the thermal images at wind-wave surfaces with various wind conditions. Four components attributed to different flow processes, including the gravity waves, parasitic capillary waves, Langmuir circulations, and the streamwisely elongated coherent vortices induced by wind shear, and the corresponding contribution fractions are derived. In the lowest wind speed case, coherent vortices contribute the most to the surface signatures, and Langmuir circulations contribute the least. As wind speed grows, the contribution of coherent vortices decreases; in contrast, the contribution of Langmuir circulations keeps growing and becomes competitive with that of other flow processes in moderate wind conditions. The contribution of gravity waves does not change much with the present cases, except the micro-breaking case, in which the contribution of gravity waves is much larger than the others, since the signatures induced by spilling breakers are classified into the attribution of gravity waves. The capillary waves contribute little to the thermal images, but become significant at the present highest wind speed, which may also be attributed to the spanwisely elongated small vortices due to fully wave breaking. The surface streaky signatures attributed to the coherent vortices can be further identified utilizing the image segmentation techniques, and the statistics of spanwise streak spacing can therefore be calculated. It is found that the distributions of streak spacing from low to high wind speed match log-normal probability density distributions. Furthermore, the mean streak spacing decreases as wind speed grows; the non-dimensional mean streak spacing in viscous length scale, however, increases with wind speed, different from that of wall turbulent boundary layer, which is a constant 100 viscous unit.

碩士<br>國立臺灣科技大學<br>電子工程系<br>99<br>In Signal Processing, Empirical mode decomposition (EMD) is one of the useful approaches for processing nonlinear and nonstationary signals. It is adaptive to input signal and could decompose signal straightly in time domain. However, its shortcomings include mode mixing and end effects that usually appear in the decomposed bands. Although a noise-assisted data analysis (NADA) called ensemble empirical mode decomposition (EEMD) has been proposed to circumvent this problem, doing so also results in an inevitably long computation for alleviating the mode mixing. In this paper, we use shaped noise instead of white noise as a disturbance for a fast convergence of EEMD. The signal-spectrum-dependent noise (SSDN) is able to effectively randomize the targeted signal in time domain, and then significantly save the superfluous calculation around the corresponding energy-free frequencies. The experimental results also show that both pink noise and brown noise outperform the white noise in terms of computation for the EEMD of speech-like signal.

碩士<br>國立臺北科技大學<br>商業自動化與管理研究所<br>99<br>Financial time series are inherently nonlinear and non-stationary, it is therefore difficult using statistical models to forecast. ANN(Artificial Neural Networks)does not require strict theoretical assumptions, so it has been widely applied for financial prediction. On ANN learning algorithms, the ELM(Extreme Learning Machine) overcomes the drawback of traditional Back-propagation. This study takes the closing price of Taiwan Capitalization Weighted Stock Index, Shanghai Stock Exchange Composite Index and Hong Kong Hang Seng Index as research subjects during the period of 2001 to 2010. We propose a hybrid forecasting model based on EEMD(Ensemble Empirical Mode Decomposition), GA (Genetic Algorithm) and ELM. Firstly, by using EEMD to decompose stock price into several IMF(Intrinsic Mode Functions) and each IMF component is modeled by individual EELM respectively. Then, we find the optimal parameters with GA. In order to examine the proposed models are better than traditional statistical models, these four models also compare with the ARIMA model. The study concluded the model combined with ELM, GA, EEMD, which has the best prediction performance. The performance of proposed four models is better than ARIMA models, showing the excellence of proposed models.

碩士<br>國立交通大學<br>生醫工程研究所<br>102<br>Excitation-emission matrix (EEM) fluorescence spectroscopy is a noninvasive method for tissue diagnosis and has become important in clinical use. However, the intrinsic characterization of EEM fluorescence remains unclear. Photobleaching and the complexity of the chemical compounds make it difficult to distinguish individual compounds due to overlapping features. Conventional studies use principal component analysis (PCA) for EEM fluorescence analysis, and the relationship between the EEM features extracted by PCA and diseases has been examined. The spectral features of different tissue constituents are not fully separable or clearly defined. Recently, a method called multi-dimensional ensemble empirical mode decomposition (MEEMD) was introduced; this method decomposes data by subtracting local means iteratively and can extract the intrinsic oscillations on multiple spatial scales without loss of information. The aim of this study was to extract the intrinsic characteristics of EEM by using MEEMD. We use simulated signal to examine the decomposition ability of MEEMD on decomposing signal which was similar to EEM but simpler than EEM, and then MEEMD was applied to decompose EEM. PCA was used to compare with MEEMD in this study. The results indicate that although PCA provides the main spectral features associated with chemical compounds, which mainly contributed by collagen, MEEMD can provide additional intrinsic features with more reliable mapping of individual chemical compounds, e.g. collagen and vitamin D. Overall, MEEMD provide a new point of view on EEM analysis and has the potential to extract intrinsic fluorescence features and improve the detection of biological fluorophores.of individual chemical compounds, e.g. collagen and vitamin D. Overall, MEEMD provide a new point of view on EEM analysis and has the potential to extract intrinsic fluorescence features and improve the detection of biological fluorophores.

碩士<br>國立交通大學<br>生醫工程研究所<br>104<br>According to statistics regarding the ten leading causes of death among the Taiwanese people, announced by the Department of Health, Executive Yuan, mortalities due to cardiovascular disease (CVDs) accounted for 40.6% of the total number of deaths in 2014. These health concerns may also induce disease and afflictions such as hypertension, stroke, and diabetes. Therefore, innovative health-care systems with high integrativity have become an important topic of research in recent years. This thesis presents Photoplethysmography (PPG) signal processing System-on-Chip (SoC) with an Ensemble Empirical Mode Decomposition (EEMD) processor. This processor has proven to be an adaptive and efficient method for non-linear and non-stationary signal analysis. EEMD decomposes a signal into several narrowband oscillatory components known as the Intrinsic Mode Functions (IMFs) in the time domain. Analyzing each IMF can obtain the physiological information. However, the EEMD processor involves a large number of complicated and iterative computations. It is necessary to use VLSI technology to implement this proposed system with real-time applications. The System-On-Chip design proposed in this thesis is implemented by using Taiwan Semiconductor Manufacturing Company (TSMC) 90 nm tape-out process. To decrease space complexity and speed up timing performance, many innovative and effective modules were developed in this thesis. The proposed hardware architecture only stored one set of white noise in ROM and others could be generated by changing the first reading address. The four stage pipeline architecture is adopted in cubic spline engine to decrease the latency. The result shows that the proposed EEMD hardware architecture can efficiently save 79.4% of the hardware resources and make hardware implementation feasible. Compared with MATLAB software, the proposed EEMD hardware implementation can speed up 23.9 times.

碩士<br>國立中央大學<br>光機電工程研究所<br>97<br>Post processing of Ensemble Empirical Mode Decomposition (EEMD) can be utilized to decompose the vibration signals of rotating machinery into finite number of Intrinsic Mode Functions (IMFs) without mode mixing problem. The basis of the post processing of EEMD will satisfy the well-defined conditions of IMF. The Autoregressive (AR) model of information-contained IMFs can be used to predict the unmeasured vibration signal, and the coefficients of AR model represent the feature of systematic dynamic behavior. In this paper, the post-processing of EEMD combining the AR model is proposed for diagnosing the looseness faults at different conponents of rotating machinery. The information-contained IMFs are selected to build the AR model. The looseness types are identified by analyzing the coefficients of AR model. The effectiveness of the proposed method is validated through the analysis of the experimental data.

With the development of mechanical fault diagnosis technology, complex mechanical systems do not need to be shut down periodically for the maintenance. The working condition of the mechanical systems can be monitored by analyzing the wear metal particles in the systems' lubricating oil. However, the output signals of the monitoring sensor are non-stationary. In some case the particle signals are overlapped with each other. The goal of this thesis is to find a method to decompose those overlapped particle signals, and then count the particle number in the lubricating oil. At the beginning EMD method was introduced in the experiment because of the character of the sensor signals. In this project, because EMD method is sensitive to the noise in the original signals, an improved version of EMD, EEMD method was implemented. Finally, a post processing method was used to get a better result.

碩士<br>國立東華大學<br>電機工程學系<br>98<br>In this study, we propose an easy-to-use noninvasive arterial stiffness assessment instrument that can be used to record the radial arterial pressure signals from the wrist. The system combines the ensemble empirical mode decomposition (EEMD) algorithm with the signals to derive a modified reflection index (MRI) and modified stiffness index (MSI). The performance of MRI and MSI was verified based on 46 subjects (35 men and 11 women, 20 to 27 years of age). Early self-monitoring of cardiovascular dysfunction and arterial stiffness can be easily and effectively achieved by MRI and MSI. Only few minutes are needed for conducting at home.

碩士<br>國立交通大學<br>電子工程學系 電子研究所<br>104<br>The heavily medical burden caused by population ageing will become a serious challenge for the current and next generation medical care system. There is an urgent need of low-cost disease prevention and home care programs to lower the possible medical burden in the future. The cardiovascular diseases have been on the list of leading cause of death for years in Taiwan. There is about seventeen million people pass away because of cardiovascular around the world. There is urgent need to get the early prevention tool to reduce the risk of cardiovascular disease all over the world. An effective photoplethysmography (PPG) signal processing system based on ensemble empirical mode decomposition (EEMD) method for acquiring the multiple physiological parameters is proposed in this project. The information of arterial pulse can be obtained by near-infrared. A high quality signal can be extracted through the proposed EEMD algorithm. Based on the most advanced semiconductor industry in Taiwan, the regulation of autonomic nervous system (ANS), RI and SI can be derived in real-time and monitored continuously. It makes the at-home care possible and lowers the rate of cardiovascular diseases and medical expenses through long-term monitoring. PPG signal acquired by the PPG capture circuit is sampled through the ADC at sample frequency of 200Hz after being filtered by the band pass filter. The digitized data are decomposed into IMFs with physiological meanings by the EEMD IC. The output IMFs are wirelessly sent to a computer via a Bluetooth module. Then the regulation of autonomic nervous system , RI and SI can be derived and display on the GUI. To overcome the noise and aliasing effect caused by nonstationary signals, many innovative and effective modules were developed in this thesis. The proposed HHT SoC design could be implemented in hardware with limited resources and fabricated under TSMC 90 nm CMOS technology. To assess the potential risk of cardiovascular, the IMFs with physiological meanings can be extracted from PPG. The RI, SI, LF, HF and VHF can be derived as the parameters to help the diagnosis of cardiovascular disease.

碩士<br>國立臺灣大學<br>地質科學研究所<br>99<br>Virtually very little dendrochronology data have been reported internationally from Taiwan, despite the existence of many dendrochronologically appropriate tree species. In this study, the potential for reconstruction of local paleoclimate was investigated using multi-century tree-ring chronologies developed from Picea morrisonicola (the endemic Taiwan Spruce). Significant correlations were found against the mean April-June diurnal temperature range (DTR) and against the mean July-September maximum temperature (Tmax). Both of these climate parameters were reconstructed based on the regression relationships. In a related study, a new frequency decomposition method called empirical mode decomposition (EMD), one part of the Hilbert-Huang Transform (HHT), was investigated as an alternative to standard methods of chronology generation in terms of climate signal. A noise assisted version of EMD called ensemble empirical mode decomposition (EEMD) was used to decompose the tree-ring time series into a series of quasi-periodic modes from high to low frequency. Consecutive modes were combined from high to low frequency and compared with the climate data. The combination with the most significant climate relationships was then used to reconstruct the climate parameters. As with the reconstructions using traditional methods of chronology generation, statistics from the reconstructions of DTR and Tmax also passed tests for model skill. The reconstruction statistics and variance explained were similar for both methods of chronology generation, with EEMD chronology having better results in the DTR reconstruction and the traditional chronology having better results in the Tmax reconstruction. Adjusted latewood ring widths show significant (p&amp;lt;0.01) positive correlation against Alishan July-September Tmax. Linear regression of the Alishan Tmax on the tree-ring chronology produced a calibration model that accounted for 23% of the actual Tmax variance. This model was used to reconstruct the July-September Tmax back to A.D. 1636. The reconstruction shows warm periods during 1718-1726, 1908-1916, and 2002-2008. Evidence from comparisons with NCEP-NCAR reanalysis data indicates that the summer climate variability in Taiwan is regulated by processes associated with changes in the Western Pacific Subtropical High (WPSH). In years with less precipitation the WPSH reduces the southwesterly monsoonal flow by extending further westward than in other years. This appears as an anomalous warm and dry summer accompanied with anti-cyclonic motion over the East China Sea. In addition, eight of the ten warmest summers (July-September Tmax) in central Taiwan occurred during El Niño years, indicating a link between Taiwan summer maximum temperatures and ENSO dynamics. The earlywood mean chronology was calibrated against April-June DTR. A calibration model that accounted for 28% of the actual DTR variance was then produced to reconstruct the DTR. The increasing Tmin, which can be attributed to locally increased cloud cover, contributed to the reduction of DTR. The reconstructed DTR has a cycle of period 28 years, showing the variations in solar irradiance possibly due to cloudiness changes.