Papers by Keyword: Partial Least Square (PLS)

Abstract: In this work, we estimate Yunnan housing price from 1999 to 2009. Firstly, we analyze the correlation coefficients between housing price and characteristic variables, identify the characteristic variables. Then, we build the forecasting model using four techniques, support vector regression (SVR), radial basis function neural network (RBFNN), partial least square (PLS) and multiple regression analysis (MRA), based on whole variables and characteristic variables. The results show that PLS technique is the best one for housing price forecasting. Its mean absolute percentage error (MAPE) is only 2.45%. SVR and RBFNN are better techniques to obtain a satisfactory forecasting result with almost 5% MAPE. Furthermore, the performance of MRA and SVR can be obviously improved through variables selection.

Abstract: In the social science research area, there are two important statistical methodologies, one is covariance-based structural equation modeling (CBSEM), and the other one is variance-based partial least square (PLS). Compared with CBSEM, PLS lacks comparatively the reference books and full applications. The main purpose of this study is to develop a paradigm to demonstrate how to assess the reliability, convergent validity, discriminant validity, and path analysis in a proposed research model by using Smart PLS. We hope this study’s result can offer some correct steps when using PLS.

Abstract: Through research of user activity characteristics of the user's social attributes of electric vehicles, electric vehicles, electric vehicles performance and electric vehicles running environment, this paper puts forward electric vehicle running characteristics model, specificly select factors affecting variables, uses structural equation modeling on operating characteristics of the electric vehicle, uses AMOS6.0 software for solve and data fitting adoping least square method, finally gives credible analysis results.

Abstract: The present study is concerning qualitative and quantitative detection of different coal samples based on near infrared spectroscopy. Firstly, near infrared spectroscopy combined with discriminant analysis (DA) was used for the coal cinder characteristics in different spectral pre-processing methods, at the same time it will establish the quantitative of the total moisture. The result indicated that 97.78% recognition ratio for calibration and 93.33% recognition ratio for validation were achieved by DA for the cinder characteristics. The PLS quantitative model of the total moisture of the establishment of the original spectral was the best, the correlation coefficients of calibration and prediction respectively were 0.981 and 0.641, RMSEC, RMSEP and RMSECV respectively were 0.859, 2.51 and 3.44.

Abstract: In the process of product satisfaction solution by SEM (Structural Equation Modeling), the model of product satisfaction has been revised aiming at solving the weights of satisfaction index distribution in the multi-sample situation. According to the characteristic of satisfaction data sampling, the partial least square is introduced, and the algorithmic method of satisfaction weights based on SEM is presented. The proposed method has been validated by an example of digital photo frame.

Abstract: Operating condition recognition of ball mill load is important to improve product quality, decrease energy consumption and ensure the safety of grinding process. A probabilistic one-against-one (OAO) multi-classification method using partial least square-based extreme learning machine algorithm (PLS-ELM) is proposed to identify the operating state of ball mill. The feature of shell vibration spectrum is extracted using KPCA. PLS-ELM model is applied to enhance the reliability and accuracy of the operating conditions identification of the ball mill load. Posterior probability of each class using Bayesian decision theory is defined as a measure as classification reliability. Classification results of the experimental ball mill shown that the accuracy and stability of the proposed method outperform ELM, PLS-ELM and KPCA-ELM model.

Abstract: For the fast and exact detection of printing color, we combine the near infrared (NIR) spectroscopy technique with partial least square (PLS) to build the detection model of printing color. Applying the 144 samples of spectral curve which obtained by the near infrared spectroscopy randomly separated into calibration set and validation set, and base on the 120 calibration set data to establish the prediction model of printing color by PLS, then this model is employed for predicting the color of the 24 validation set. The RMSEC of the 24 blocks’ color parameters L^*, a^*, b^*, E are 0.73, 2.26, 3.03 and 1.09 respectively; The RMSEP are 0.97, 2.77, 3.57 and 1.34 respectively. Those results tell that the NIR spectrum and blocks’ color parameters L^*, a^*, b^*, E could accurately establish a quantitative regression model, applying such model also can accurately predict unknown samples, and the results approximate to the original reference data. The use of near infrared spectroscopy to detect the printed matter nondestructively is feasible, and lays the foundation for the further analysis and establishment of printing chroma model.

Abstract: the detection of the apple internal acidity using near infrared reflection spectroscopy is researched in this paper. This paper also elaborates the principle of partial least squares (PLS), and models spectral data by using the partial least squares method. Focuses on the comparison of the single points measurement and multiple points measurement in the Apple’s maximum diameter measurements, gets the following conclusions: a single point spectral scans can not accurately predict the internal acidity of the apple, the experimental precision of the multi-points measurement is higher then the single points measurement. The authors suggest that the measurement of average spectral values by multi-fiber optic probe should be used for the detection of the apple acidity.

Abstract: Near-infrared (NIR) spectroscopy combined with chemometrics methods has been investigated to discriminate type of honey. 147 NIR spectra of six floral origins of honey samples were collected within 4000~10000cm-1 spectral region. Spectral data were compressed using partial least squares (PLS). Back propagation neural networks (BPNN) models were constructed to distinguish the type of honey. Six spectral data pretreatments including first derivative, first derivatives followed by mean centering(MC), second derivatives, Savitzky-Golay smoothing, standard normal variate transformation (SNV) and multiplicative scattering correction (MSC) were compared to establish the optimal models for honey discrimination. Savitzky-Golay smoothing proved more effective than the other data pretreatments. BPNN models were developed within the full spectral region, 5303～6591cm-1 and 7012～10001cm-1, respectively. Results have shown that the highest(100%) classification rate was achieved within 5303～6591cm-1 wave range. Our results indicated that NIR spectroscopy with chemometrics techniques can be applied to classify rapidly honeys of different floral origin.

Abstract: Ultra-violet, visible and near infrared (UV-VIS-NIR) spectroscopy combined with chemometrics was investigated for fast determination of soluble solids content (SSC) of tea beverage. In this study, a total of 120 tea samples with SSC range of 4.0-9.5 ºBrix were tested. Samples were randomly divided for calibration (n=90) and independent validation (n=30). Spectra were collected by a mobile fiber-type UV-VIS-NIR spectrophotometer in transmission mode with recorded wavelength range of 203.64-1128.05 nm. Various calibration approaches, i.e., principal components analysis (PCA), partial least squares (PLS) regression, least squares support vector machine (LSSVM) and back propagation artificial neural network (BPANN), were investigated. The combinations of PCA-BPANN, PCA-LSSVM, PLS-BPANN and PLS-LSSVM were also investigated to build calibration models. Validation results indicated that all these investigated models achieved high prediction accuracy. Especially, PLS-LSSVM achieved best performance with mean coefficient of determination (R2) of 0.99, root-mean-square error of prediction (RMSEP) of 0.12 and residual prediction deviation (RPD) of 15.16. This experiment suggests that it is feasible to measure SSC of tea beverage using UV-VIS-NIR spectroscopy coupled with appropriate multivariate calibration, which may allow using the proposed method for off-line and on-line quality supervision in the production of soft drink.