Support Vector Machine Regression Model Research Articles

Parametric design and optimization of high speed train nose

Aiming at shortening the design period and improve the design efficiency of the nose shape of high speed trains, a parametric shape optimization method is developed for the design of the nose shape has been proposed in the present paper based on the VMF parametric approach, NURBS curves and discrete control point method. 33 design variables have been utilized to control the nose shape, and totally different shapes could be obtained by varying the values of design variables. Based on the above parametric method, multi-objective particle swarm algorithm, CFD numerical simulation and supported vector machine regression model, multi-objective aerodynamic shape optimization has been performed. Results reveal that the parametric shape design method proposed here could precisely describe the three-dimensional nose shape of high speed trains and could be applied to the concept design and optimization of the nose shape. Besides, the SVM regression model based the multi-points criterion could accurately describe the non-linear relationship between the design variables and objectives, and could be generally utilized in other fields. No matter the simplified model or the real model, the aerodynamic performance of the model after optimization has been greatly improved. Based on the SVR model, the nonlinear relation between the aerodynamic drag and the design variables is obtained, which could provide guidance for the engineering design and optimization.

Investigation into the phenomena affecting the retention behavior of basic analytes in chaotropic chromatography: Joint effects of the most relevant chromatographic factors and analytes’ molecular properties

The aim of this study was to systematically investigate the phenomena affecting the retention behavior of structurally diverse basic drugs in ion-interaction chromatographic systems with chaotropic additives. To this end, the influence of three factors was studied: pH value of the aqueous phase, concentration of sodium hexafluorophosphate, and content of acetonitrile in the mobile phase. Mobile phase pH was found to affect the thermodynamic equilibria in the studied system beyond its effects on the analytes’ ionization state. Specifically, increasing pH from 2 to 4 led to longer retention times, even with analytes which remain completely protonated. An explanation for this phenomenon was sought by studying the adsorption behavior of acetonitrile and chaotropic additive onto stationary phase. It was shown that the magnitude of the developed surface potential, which significantly affects retention – increases with pH, and that this can be attributed to the larger surface excess of acetonitrile. To study how analytes’ structural properties influence their retention, quantitative structure–retention modeling was performed next. A support vector machine regression model was developed, relating mobile phase constituents and structural descriptors with retention data. While the ETA_EtaP_B_RC and XlogP can be considered as molecular descriptors which describe factors affecting retention in any RP-HPLC system, TDB9p and RDF45p are molecular descriptors which account for spatial arrangement of polarizable atoms and they can clearly relate to analytes’ behavior on the stationary phase surface, where the electrostatic potential develops. Complementarity of analytes’ structure with that of the electric double layer can be seen as a key factor influencing their retention behavior. Structural diversity of analytes and good predictive capabilities over a range of experimental conditions make the established model a useful tool in predicting retention behavior in the studied chromatographic system.

Towards an effective cross-task mental workload recognition model using electroencephalography based on feature selection and support vector machine regression

Electroencephalographic (EEG) has been believed to be a potential psychophysiological measure of mental workload. There however remain a number of challenges in building a generalized mental workload recognition model, one of which includes the inability of an EEG-based workload classifier trained on a specific task to handle other tasks. The primary goal of the present study was to examine the possibility of addressing this challenge using feature selection and regression model. Support vector machine classifier and regression models were examined under within-task conditions (trained and tested on the same task) and cross-task conditions (trained on one task and tested on another task) for well-trained verbal and spatial n-back tasks. A specifically designed cross-task recursive feature elimination (RFE) based feature selection was used to handle the possible causes responsible for the deterioration of the performance of cross-task regression model. The within-task classification and regression performed fairly well. Cross-task classification and regression performance, however, deteriorated to unacceptable levels (around chance level). Trained and tested with the most robust feature subset selected by cross-task RFE, the performance of cross-task regression was significantly improved, and there were no significant changes in the performance of within-task regression. It can be inferred that workload-related features can be picked out from those which have been contaminated using RFE, and regression models rather than classifiers may be a wiser choice for cross-task conditions. These encouraging results suggest that the cross-task workload recognition model built in this study is much more generalizable across task when compared to the model built in traditional way.

Estimation of cylinder orientation using autonomous underwater vehicle mapping of bistatic scattered fields

When an aspect-dependent target is insonified by an acoustic source, distinct features are produced in the resulting bistatic scattered field. These features were exploited in a process for estimating target aspect angles that was demonstrated on a real-world data set using models produced using simulation data. Bistatic scattering data was collected during an experiment in November 2014 in Massachusetts Bay using a ship-based acoustic source producing 7-9kHz LFM chirps and a steel pipe target. The true target orientation was unknown, as the target was dropped from the ship with no rotation control. The Autonomous Underwater Vehicle (AUV) Unicorn, fitted with a 16-element nose array and data acquisition payload, was deployed in broadside data collection behaviors around the target, and the ship was moved to create two target aspects. Scattering data was collected for each target aspect angle. A Support Vector Machine (SVM) regression model was trained using simulated scattering bistatic field data from the OASES-SCATT simulation package. This model was used to estimate the target aspect angles using the real data collected by the AUV during the experiment. The aspect angle estimates were consistent with experimental observations of relative source positioning based on ship position.

Binary mixtures of waxy wheat and conventional wheat as measured by NIR reflectance

Waxy wheat contains very low concentration (generally <2%) of amylose in endosperm starch, in contrast to conventional wheat whose starch is typically 20% amylose, with the balance being the branched macromolecule, amylopectin. With the release of a commercial hard winter waxy wheat cultivar in the United States, the grain trade, milling, and processing industries seek to have a rapid technique to ensure the purity of identity preserved waxy wheat lots. Near infrared (NIR) reflectance spectroscopy, a technique widely used in the cereals industry for proximate analysis, is a logical candidate for measuring contamination level and thus is the subject of this study. Two sets of wheat samples, harvested, prepared and scanned one year apart, were used to evaluate the NIR concept. One year consisted of nine pairs of conventional:waxy preparations, with each preparation consisting of 29 binary mixtures ranging in conventional wheat fraction (by weight) of 0–100% (261 spectral samples). The second year was prepared in the same fashion, with 12 preparations, thus producing 348 spectral samples. One year's samples were controlled for protein content and moisture level between pair components in order to avoid the basis for the conventional wheat fraction models being caused by something other than spectral differences attributed to waxy and nonwaxy endosperm. Likewise the second year was controlled by selection of conventional wheat for mixture preparation based on either protein content or cluster analysis of principal components of candidate spectra. Partial least squares regression, one and two-term linear regression, and support vector machine regression models were examined. Validation statistics arising from sets within the same year or across years were remarkably similar, as were those among the three regression types. A single wavelength on second derivative transformed spectra, namely 2290nm, was effective at estimating the mixture level by weight, with standard errors of performance in the 6–9% range. Thus, NIR spectroscopy may be used for measuring conventional hard wheat ‘contamination’ in waxy wheat at mixture levels above 10% w/w.

Multi-method ensemble selection of spectral bands related to leaf biochemistry

Multi-method ensembles are generally believed to return more reliable results than the application of one method alone. Here, we test if for the quantification of leaf traits an ensemble of regression models, consisting of Partial Least Squares (PLSR), Random Forest (RFR), and Support Vector Machine regression (SVMR) models, is able to improve the robustness of the spectral band selection process compared to the outcome of a single technique alone. The ensemble approach was tested using one artificial and five measured data sets of leaf level spectra and corresponding information on leaf chlorophyll, dry matter, and water content. PLSR models optimized for the goodness of fit, an established approach for band selection, were used to evaluate the performance of the ensemble. Although the fits of the models within the ensemble were poorer than the fits achieved with the reference approach, the ensemble was able to provide a band selection with higher consistency across all data sets. Due to the selection characteristics of the methods within the ensemble, the ensemble selection is moderately narrow and restrictive but in good agreement with known absorption features published in literature. We conclude that analyzing the range of agreement of different model types is an efficient way to select a robust set of spectral bands related to the foliar properties under investigation. This may help to deepen our understanding of the spectral response of biochemical and biophysical traits in foliage and canopies.

Small Time-Scale Network Tidal Water Level Forecasting Based on Quantum Neural Network

Zhao, K.; Wang, H., and Guo, T., 2015. Small time-scale network traffic prediction on the basis of Quantum Neural Network.This paper proposed a quantum neural network model consist of quantum bit, universal quantum gates and quantum weighted, gave the learning algorithm based on improved PRP conjugate gradient, in order to improve the convergence speed of the network and the network performance, and further proved the global convergence of the algorithm in theory. Compared with the existing localized support vector machine (SVM) regression and flexible neural tree model, the QNN model proposed in this paper have a higher prediction precision and low computational complexity for small scale tidal water level prediction, its convergence speed and robustness are better than BP network and quantum weighted neural network.

Estimation of potentially toxic elements contamination in anthropogenic soils on a brown coal mining dumpsite by reflectance spectroscopy: a case study.

In order to monitor Potentially Toxic Elements (PTEs) in anthropogenic soils on brown coal mining dumpsites, a large number of samples and cumbersome, time-consuming laboratory measurements are required. Due to its rapidity, convenience and accuracy, reflectance spectroscopy within the Visible-Near Infrared (Vis-NIR) region has been used to predict soil constituents. This study evaluated the suitability of Vis-NIR (350–2500 nm) reflectance spectroscopy for predicting PTEs concentration, using samples collected on large brown coal mining dumpsites in the Czech Republic. Partial Least Square Regression (PLSR) and Support Vector Machine Regression (SVMR) with cross-validation were used to relate PTEs data to the reflectance spectral data by applying different preprocessing strategies. According to the criteria of minimal Root Mean Square Error of Prediction of Cross Validation (RMSEPcv) and maximal coefficient of determination (R2 cv) and Residual Prediction Deviation (RPD), the SVMR models with the first derivative pretreatment provided the most accurate prediction for As (R2 cv) = 0.89, RMSEPcv = 1.89, RPD = 2.63). Less accurate, but acceptable prediction for screening purposes for Cd and Cu (0.66 ˂ R2 cv) ˂ 0.81, RMSEPcv = 0.0.8 and 4.08 respectively, 2.0 ˂ RPD ˂ 2.5) were obtained. The PLSR model for predicting Mn (R2 cv) = 0.44, RMSEPcv = 116.43, RPD = 1.45) presented an inadequate model. Overall, SVMR models for the Vis-NIR spectra could be used indirectly for an accurate assessment of PTEs’ concentrations.

The Weighted Support Vector Machine Based on Hybrid Swarm Intelligence Optimization for Icing Prediction of Transmission Line

Not only can the icing coat on transmission line cause the electrical fault of gap discharge and icing flashover but also it will lead to the mechanical failure of tower, conductor, insulators, and others. It will bring great harm to the people’s daily life and work. Thus, accurate prediction of ice thickness has important significance for power department to control the ice disaster effectively. Based on the analysis of standard support vector machine, this paper presents a weighted support vector machine regression model based on the similarity (WSVR). According to the different importance of samples, this paper introduces the weighted support vector machine and optimizes its parameters by hybrid swarm intelligence optimization algorithm with the particle swarm and ant colony (PSO-ACO), which improves the generalization ability of the model. In the case study, the actual data of ice thickness and climate in a certain area of Hunan province have been used to predict the icing thickness of the area, which verifies the validity and applicability of this proposed method. The predicted results show that the intelligent model proposed in this paper has higher precision and stronger generalization ability.

A laser induced breakdown spectroscopy quantitative analysis method based on the robust least squares support vector machine regression model

Data fluctuation in multiple measurements of Laser Induced Breakdown Spectroscopy (LIBS) greatly affects the accuracy of quantitative analysis. Our proposed method achieved better prediction accuracy and modeling robustness.

Application of Support Vector Regression for Simultaneous Modelling of near Infrared Spectra from Multiple Process Steps

Near infrared (NIR) spectroscopy in combination with partial least-squares regression (PLS) is widely applied in process control for non-destructive measurement of quality parameters during production. PLS assumes an approximate linear relationship between the parameter to be estimated and the intensity of its absorption bands. Spectra, however, may contain non-linearities for various reasons such as differences in viscosity, temperature, pH, particle size and chemical composition of the sample matrix. In such cases, PLS might not predict the parameter of interest sufficiently well, and one must find other methods for the calibration task. Support vector machine regression (SVR) has been gaining interest within chemometrics in recent years. SVR is capable of modelling highly non-linear data, also when data are of very high dimensions. The aim of this study was to develop calibration models of NIR spectra from four different process steps in a raw-sugar factory. The models were intended for monitoring two quality parameters at the individual process steps. Our goal was to obtain global calibration models covering all four process steps in order to obtain simple calibration maintenance. SVR was used for the calibration task, since all of the above-mentioned sources of non-linearities were present. SVR results were benchmarked against PLS. SVR had a better prediction performance than PLS (1) for models built on individual process steps, (2) for global models covering all four process steps and (3) when the global models were evaluated on the individual process steps. Moreover, the majority of SVR models had prediction errors close to reference uncertainty and hence were close to being optimal. Finally, the global SVR models predicted the individual process steps better than the corresponding local PLS models. We conclude that the nonlinear modelling method SVR was able to model non-linearities caused by pooling NIR spectra from multiple different process steps. Implementation of the global SVR models would have several advantages over the local PLS models. First, they would allow simple calibration maintenance because only one model per quality parameter would have to be maintained. Second, they would allow more precise estimation of the quality parameters and therefore better process monitoring.

MALARIA SEVERITY CLASSIFICATION THROUGH JORDAN-ELMAN NEURAL NETWORK BASED ON FEATURES EXTRACTED FROM THICK BLOOD SMEAR

This article presents an alternative approach useful for medical prac- titioners who wish to detect malaria and accurately identify the level of severity. Malaria classiers are usually based on feed forward neural networks. In this study, the proposed classier is developed based on the Jordan-Elman neural networks. Its performance is evaluated using a receiver-operati characteristic curve, sensitiv- ity, specicity, positive predictive value, negative predictive value, confusion matrix, mean square error, determinant coefficient, and reliability. The effectiveness of the classier is compared to a support vector machine and multiple regression models. The results of the comparative analysis demonstrate a superior performance level of the Jordan-Elman neural network model. Further comparison of the classier with previous literature indicates performance improvement over existing results. The Jordan-Elman neural networks classier can assist medical practitioners in the fast detection of malaria and determining its severity, especially in tropical and subtropical regions where cases of malaria are prevalent.

Energy Field Filling of NEIC Broadband Radiated Energy Catalogue Based on Support Vector Machine Regression Model

Earthquake prediction is one of the most difficult problems in modern natural science. Undoubtedly, various seismic parameters included in broadband radiated energy catalogue of NEIC is very important data source of investigate correlation between different earthquake within specific time space scope. In the paper, a fitting mode of seismic energy based on Support Vector Machine is established, using which we can fill the absent energy filed in the broadband radiated energy catalogue.

Research on Nonlinear Modeling Method of Support Vector Machine with Wavelet Derivation Kernel Function

The model capability of Support Vector Machine (SVM) relies on the selection of kernel function. To obtain a better application modeling of SVM, the wavelet kernel function that satisfies Merce condition is introduced to use the kernel function of SVM, achieving a good effect. In the paper, on the basis of wavelet kernel function, a wavelet derivation kernel function is proposed in the application of SVM for higher accuracy. An actual example on nonlinear function approximation shows that SVM regression model has a satisfactory approximation effect, and also support an effective nonlinear modeling method.

고분자 전해질 연료전지용 막가습기의 상대습도 추정을 위한 소프트센서 개발

It is important to accurately measure and control the relative humidity of humidified gas entering a PEM (polymer electrolyte membrane) fuel cell stack because the level of humidification strongly affects the performance and durability of the stack. Humidity measurement devices can be used to directly measure the relative humidity, but they cost much to be equipped and occupy spaces in a fuel cell system. We present soft sensors for predicting the relative humidity without actual humidity measuring devices. By combining FIR (finite impulse response) model with PLS (partial least square) and SVM (support vector machine) regression models, DPLS (dynamic PLS) and DSVM (dynamic SVM) soft sensors were developed to correctly estimate the relative humidity of humidified gases exiting a planar-type membrane humidifier. The DSVM soft sensor showed a better prediction performance than the DPLS one because it is able to capture nonlinear correlations between the relative humidity and the input data of the soft sensors. Without actual humidity sensors, the soft sensors presented in this work can be used to monitor and control the humidity in operation of PEM fuel cell systems.

Epithelial-mesenchymal transition in prostate cancer is associated with quantifiable changes in nuclear structure.

Prostate cancer progression is concomitant with quantifiable nuclear structure and texture changes as compared to non-cancer tissue. Malignant progression is associated with an epithelial-mesenchymal transition (EMT) program whereby epithelial cancer cells take on a mesenchymal phenotype and dissociate from a tumor mass, invade, and disseminate to distant metastatic sites. The objective of this study was to determine if epithelial and mesenchymal prostate cancer cells have different nuclear morphology. Murine tibia injections of epithelial PC3 (PC3-Epi) and mesenchymal PC3 (PC3-EMT) prostate cancer cells were processed and stained with H&E. Cancer cell nuclear image data was obtained using commercially available image-processing software. Univariate and multivariate statistical analysis were used to compare the two phenotypes. Several non-parametric classifiers were constructed and permutation-tested at various training set fractions to ensure robustness of classification between PC3-Epi and PC3-EMT cells in vivo. PC3-Epi and PC3-EMT prostate cancer cells were separable at the single cell level in murine tibia injections on the basis of nuclear structure and texture remodeling associated with an EMT. Support vector machine and multinomial logistic regression models based on nuclear architecture features yielded AUC-ROC curves of 0.95 and 0.96, respectively, in separating PC3-Epi and PC3-EMT prostate cancer cells in vivo. Prostate cancer cells that have undergone an EMT demonstrated an altered nuclear structure. The association of nuclear changes and a mesenchymal phenotype demonstrates quantitative morphometric image analysis may be used to detect cancer cells that have undergone EMT. This morphometric measurement could provide valuable prognostic information in patients regarding the likelihood of [future] metastatic disease.

A diagnosis model for amnestic mild cognitive impairment based on structural network characteristics of the brain

The aim of this work is to select topological characteristics of structural network which were higher correlated with cognitive performance and estimate the classification models to classify normal aging and amenstic mild cognitive impairment( aMCI) patients. In two groups of diffusion tensor image( DTI) dataset,a group has 52 normal aging subjects and another group has 39 aMCI patients. We employed the diffusion tensor image to construct the structural network in each group and used the method of graph theory to extract the characteristics of structural network. We selected significant features by the correlation analysis between the characteristics of brain network and subject's mini-mental state examination( MMSE) score. These features were used to establish five kinds of classification model for evaluating the classification efficiency of the models.For normal aging dataset and aMCI dataset,18 characteristics of the structural network were selected as features,which are significantly associated with cognitive ability and locate in 9 brain areas according to the automated anatomical labeling( AAL) template in each group. However,the features and the related regions are different for the two datasets. Among 5 algorithms,sequential minimal optimization learning algorithm for support vector machine regression model was more accurate due to the higher specificity of 88. 46%,the higher sensitivity of 83. 05% and the higher accuracy of 85. 71%. Brain structural network metrics which were correlated greatly with the cognitive performance could be taken as biological marker pointers to establish the classification model to classify normal aging subjects and aMCI patients. Furthermore,these structural network features can provide the information of connection changes between corresponding brain regions.

Oil Palm Leaf Nutrient Estimation by Optical Sensing Techniques

<abstract> <bold><sc>Abstract.</sc></bold> This study evaluated four optical sensor systems in an effort to characterize a robust nondestructive technique to estimate tree-specific oil palm leaflet nutrient deficiency and aid in determining the required fertilizer dosage rates. The four sensor systems used were a handheld reflectance sensor (GreenSeeker 505), a chlorophyll meter (SPAD 502 Plus), a fluorescence sensor (Multiplex 3), and a spectroradiometer (FieldSpec 3). Spectral responses were collected using the above sensors for leaf samples from 164 randomly selected trees from three commercial farms in Malaysia. The spectral data were processed using either stepwise multiple linear regression (SMLR) or rank features techniques for feature extraction. The feature datasets were analyzed using support vector machine (SVM), soft independent modeling of class analogies (SIMCA), and artificial neural network (ANN) classifiers. The SVM and ANN classifiers performed better compared to SIMCA for most nutrient prediction cases. Overall, with SMLR as a feature extraction technique and SVM as a classifier, a spectroradiometer can be used to effectively predict N, P, K, Mg, Ca, and B nutrient levels in oil palm leaflets. The average overall prediction accuracies for the above conditions were in the range of 59% to 85%. The spectroradiometer dataset was further analyzed using SVM regression models for estimating nutrient levels. Either SMLR-selected features or partial least square (PLS) latent variables were used for establishing the above regression models. The models fitted on the SMLR features performed better for N, Mg, Ca, and B, with correlation coefficients (r) for test datasets in the range of 0.45 to 0.78. The best fitted SVM regression models were for boron estimation, with a correlation coefficient of 0.78 and root mean squared error of prediction (RMSEP) of 2.82.

Traffic Speed Time Series Short Term Forecasting Using Aggregated Model

This paper integrated superiority from time series model and least square support vector machine regression model with data aggregation for traffic speed short term forecasting. Based on the results of traffic data variations analysis, the practicability that speed data can be aggregated to several periods was confirmed, and aggregated model can be developed to forecast the speed with auto regression (AR) model and support vector machine regression (SVR). Then the speed data in case study were integrated to 4 periods at the location of Remote Traffic Microwave Sensors (RTMS) 2047 on 2nd Ring Road Expressway in Beijing. Arguments with coefficients from AR models then act as the independent variables of LSSVR in aggregated model. Short term traffic speed was predicted by aggregated model, and the results indicated that taking advantages of time periods variation rule inside the aggregated model would help save the model running time cost under the premise of accuracy with better prediction ability than LSSVR in certain conditions.

Estimating soil organic carbon content with visible-near-infrared (vis-NIR) spectroscopy.

The selection of a calibration method is one of the main factors influencing measurement accuracy with visible-near-infrared (Vis-NIR, 350-2500 nm) spectroscopy. This study, based on both air-dried unground (DU) and air-dried ground (DG) soil samples, used nine spectral preprocessing methods and their combinations, with the aim to compare the commonly used partial least squares regression (PLSR) method with the new machine learning method of support vector machine regression (SVMR) to find a robust method for soil organic carbon (SOC) content estimation, and to further explore an effective Vis-NIR spectral preprocessing strategy. In total, 100 heterogeneous soil samples collected from Southeast China were used as the dataset for the model calibration and independent validation. The determination coefficient (R(2)), root mean square error (RMSE), residual prediction deviation (RPD), and ratio of performance to interquartile range were used for the model evaluation. The results of this study show that both the PLSR and SVMR models were significantly improved by the absorbance transformation (LOG), standard normal variate with wavelet detrending (SW), first derivative (FD), and mean centering (MC) spectral preprocessing methods and their combinations. SVMR obtained optimal models for both the DU and DG soil, with R(2), RMSE, and RPD values of 0.72, 2.48 g/kg, and 1.83 for DU soil and 0.86, 1.84 g/kg, and 2.60 for DG soil, respectively. Among all the PLSR and SVMR models, SVMR showed a more stable performance than PLSR, and it also outperformed PLSR, with a smaller mean RMSE of 0.69 g/kg for DU soil and 0.50 g/kg for DG soil. This study concludes that PLSR is an effective linear algorithm, but it might not be sufficient when dealing with a nonlinear relationship, and SVMR turned out to be a more suitable nonlinear regression method for SOC estimation. Effective SOC estimation was obtained based on the DG soil samples, but the accurate estimation of SOC with DU soil samples needs to be further explored. In addition, LOG, SW, FD, and MC are valuable spectral preprocessing methods for Vis-NIR optimization, and choosing two of them (except for FD + SW and LOG + FD) in a simple combination is a good way to get acceptable results.