Journal of Chemometrics最新文献

Multivariate statistical process monitoring is commonly used to detect abnormal process behavior in real time. Multiple process variables are monitored simultaneously, and alarms are issued when monitoring statistics exceed a predetermined threshold. Traditional approaches use a parametric threshold based on the assumptions of independence and multivariate normality of the process data, which are often violated in complex processes with high sampling frequencies, leading to excessive false alarms. Some approaches for improved threshold selection have been proposed, but they assume independence of the monitoring statistics, which are often autocorrelated. In this paper, we compare the performance of nonparametric estimators for computing thresholds from autocorrelated monitoring statistics through simulation. The false alarm rate and in-control average run length of each estimator under different distributions, sample sizes, and autocorrelation levels and types are found. Estimator performance is found to depend on sample size and the strength of autocorrelation. The class of kernel density estimation (KDE) methods tends to perform better than estimators that use bootstrapping, and the proposed adjusted KDE methods that account for autocorrelation are recommended for general use. A case study to monitor a wastewater treatment facility further illustrates the performance of nonparametric and parametric thresholds when applied to real-world systems.

Deployment of process analytical technology tools such as Raman or IR spectroscopy and associated multivariate calibration models for process monitoring and control plays an important role in process automation and advanced manufacturing of pharmaceuticals. Preprocessing or preparation of the spectroscopic data is an important step in developing a multivariate calibration model. There are several ways available to preprocess the data and each may influence the calibration model performance differently. Here we investigated the influence of preprocessing procedures on the development and performance of the chemometric models to predict the glucose concentration in a bioreactor. Box–Behnken design of experiment (DOE) was used to generate the Raman spectroscopy data. Four factors were considered critical in the DOE—glucose, glutamine, glutamic acid, and antifoam concentration. Raman spectroscopy data were collected both with and without aeration conditions, independently from three cell culture media. For each medium, data consisted of calibration set (27 conditions) and model validation set (9 conditions) separately. Additionally, Raman data was also collected for certain DOE runs with increasing concentration of cell densities ranging from 0.5 × 10 E06/mL to 30 × 10 E06/mL under aerating conditions. Data from the three cell culture media were used separately to develop calibration models that used four different preprocessing procedures, namely, baseline correction (BLC), Savitzky–Golay smoothing (SGS), Savitzky–Golay derivative (SGD) and orthogonal signal correction (OSC). The preprocessing procedures were applied individually and in combinations to evaluate the calibration model parameters and the performance metrics. We further developed glucose calibration models based on partial least squares (PLS) regression with 1–3 principal components. The models developed with OSC procedure gave superior performance metrics with just one principal component across all three media. Models developed with other preprocessing procedures required two or more principal components to give comparable performance. Overall, the choice of preprocessing procedures affected the model performance.

Chemical separations data are typically analyzed in the time domain using methods that integrate the discrete elution bands. Integrating the same chemical components across several samples must account for retention time drift over the course of an entire experiment as the physical characteristics of the separation are altered through several cycles of use. Failure to consistently integrate the components within a matrix of $��M�\times �N�$ samples and variables creates artifacts that have a profound effect on the analysis and interpretation of the data. This work presents an alternative where the raw separations data are analyzed in the frequency domain to account for the offset of the chromatographic peaks as a matrix of complex Fourier coefficients. We present a generalization of the factorization, permutation testing, and visualization steps in ANOVA-simultaneous component analysis (ASCA) to handle complex matrices and use this method to analyze a synthetic dataset with known significant factors and compare the interpretation of a real dataset via its peak table and frequency domain representations.

Artemisia argyi smoke, generated from the combustion of A. argyi, is widely utilized in traditional Chinese medicine for moxibustion and fumigation therapies. The released smoke during the combustion of A. argyi is rich in massive compounds and can be affected by storing periods. However, there is a lack of a comprehensive understanding on the chemical composition of released smoke, and the effects of annual storage on the released smoke were still unclear. Herein, a strategy that integrated chromatography–mass spectrometry (GC-MS) with advanced chemometric software, AntDAS-GCMS, was developed for comprehensively characterizing tens of compound in the released smoke of A. argyi and evaluating the quality variation across annual storage periods. Both particle and gas phases of the released smoke during the combustion were collected for GC-MS analysis; the raw data files were then imported to our recently developed data analysis software AntDAS-GCMS for automatically retrieving underlying components. Components that show significant difference among annual storage periods were screened to provide a total of 471 components. Finally, 61 compounds were identified. Both supervised and unsupervised chemometric methods suggest that the 2- and 4-year storage periods were close to clinic used sample (3-year storage), whereas a too short storage (like 1 year) or too long storage (6 years) were quite different from the 3-year storage samples. In conclusion, this strategy provides a novel solution for evaluating smoke samples from traditional Chinese medicine.

The soluble solids content (SSC) in apples directly affects their quality. This study aimed to detect SSC nondestructively using hyperspectral technology combined with chemometrics. However, data generation may not follow a specific pattern, and even small perturbations in the data can have a significant impact on the constructed model. To improve the anti-interference capability of individual models, this study proposed a stacking ensemble learning method that adopted partial least squares (PLS), support vector machine (SVM), extreme gradient boosting (Xgboost), random forest (RF) as basic-learners, and RF serving as a meta-learner. Experimental results showed that the performance of the established model on the test set were as follows: the root mean square error (RMSE) was 0.4325, mean absolute error (MAE) was 0.3245, mean absolute percentage error (MAPE) was 0.0271, coefficient of determination ($��{�R�}^{�2�}�$) was 0.9250. These results indicate that the stacking ensemble learning approach could appropriately fuse the predictive results of each basic-learner and improve the prediction accuracy of individual models. To verify the superiority of the proposed stacking ensemble learning method, the selection of its basic-learners, meta-learner, and combination strategy were compared and analyzed. This study not only provides a theoretical reference for the further development of related nondestructive detection equipment but also offers guidance for fusion algorithms as well.

This study introduces two robust multivariate Shewhart-type control charts based on grouped observations to detect changes in the covariance matrix, with a focus on monitoring sulfur dioxide levels during quality control processes. We compute the covariance matrix of observations, and apply the least absolute shrinkage and selection operator to penalize it in the in-control process. Logarithms are then applied to eigenvalues derived through singular value decomposition (SVD) of the shrunken covariance matrix, ensuring robustness to non-normality in the multivariate data. The proposed methods offer significant advantages, particularly in their ability to maintain robustness to non-normality without relying on strict distributional assumptions. Performance comparisons using the average run length demonstrate that the proposed charts exhibit superior robustness to normality assumptions compared with existing methods. However, potential limitations include the computational complexity of the shrinkage and SVD processes, which may affect the scalability of large datasets. An application to the white wine production process illustrates the effectiveness of the proposed methods for analyzing complex multivariate chemical data. These findings indicate that the introduced charts enhance the detection of shifts in the covariance matrix of physicochemical properties, thereby improving the reliability of quality control processes in non-normal environments. This study provides valuable tools for quality engineers and practitioners in industries dealing with multivariate analytical data, contributing to improved process monitoring and control, ensuring higher quality standards, and ensuring consistent product outcomes in fields such as food science and industrial chemistry.

Gastrodia elata is a traditional Chinese medicine with medicinal and edible values. In this paper, two kinds of datasets were acquired: partial spectra (artificially obtained peak segment spectra) and full spectra (4000–400 cm⁻¹). Competitive adaptive reweighted sampling algorithm (CARS) and successive projection algorithm (SPA) were utilized to extract the characteristic variables of the two datasets, and Partial Least Squares Discriminant Analysis (PLS-DA) models, Support Vector Machines (SVM) models, Random Forests (RF) models, and Residual convolutional neural networks (ResNet) were established. It was found that among the PLS-DA models whole-MSC-CARS-PLS-DA was optimal, with a Root Mean Square Error of Prediction (RMSEP) of 0.0658; among the SVM models Partial-Standard Normal Variable (SNV-SPA-SVM was the best, with a kernel parameter of 0.1768 and the lowest number of support vectors; among the RF models Partial-SNV-RF is optimal, but not as effective as the first two models. The loss value of the ResNet model built based on effective information is 0.001, and the model building time is short and directly uses the original data. Therefore, the ResNet model based on feature bands is the most suitable for practical application compared with other models.