Generalized regression neural network association with terahertz spectroscopy for quantitative analysis of benzoic acid additive in wheat flour

Predicting wheat gluten concentrations in potato starch using GPR and SVM models built by terahertz time-domain spectroscopy

The purpose of this study was for the first time to explore the feasibility of terahertz (THz) spectral imaging for the detection of gluten contents in food samples. Based on the obtained 80 THz spectrum data, Gaussian process regression (GPR) and support vector machine (SVM) models were established to predict wheat gluten concentrations in 40 potato starch mixture samples. The prediction performances of GPR and SVM obtained were R2 = 0.859 and RMSE = 0.070, and R2 = 0.715 and RMSE = 0.101 in the gluten concentration range of 1.3%-100%, respectively, showing that the linear SVM algorithm had better prediction performance. The results indicated that THz spectral imaging combined with GPR could be used to predict the gluten content in food samples. It is thus hoped that this research should provide a novel technique for gluten content detection to ensure gluten-free food samples.

A Review of Recent Advances for the Detection of Biological, Chemical, and Physical Hazards in Foodstuffs Using Spectral Imaging Techniques

Traditional methods for detecting foodstuff hazards are time-consuming, inefficient, and destructive. Spectral imaging techniques have been proven to overcome these disadvantages in detecting foodstuff hazards. Compared with traditional methods, spectral imaging could also increase the throughput and frequency of detection. This study reviewed the techniques used to detect biological, chemical, and physical hazards in foodstuffs including ultraviolet, visible and near-infrared (UV-Vis-NIR) spectroscopy, terahertz (THz) spectroscopy, hyperspectral imaging, and Raman spectroscopy. The advantages and disadvantages of these techniques were discussed and compared. The latest studies regarding machine learning algorithms for detecting foodstuff hazards were also summarized. It can be found that spectral imaging techniques are useful in the detection of foodstuff hazards. Thus, this review provides updated information regarding the spectral imaging techniques that can be used by food industries and as a foundation for further studies.

Chemometric Analysis of a Ternary Mixture of Caffeine, Quinic Acid, and Nicotinic Acid by Terahertz Spectroscopy

Caffeine, quinic acid, and nicotinic acid are among the significant chemical determinants of coffee quality. This study develops a chemometric model to quantify these compounds in ternary mixtures analyzed by terahertz time-domain spectroscopy (THz-TDS). A data set of 480 THz spectra was obtained from 80 samples. Combinations of data preprocessing methods, including normalization (Z-score, min-max scaling, Mie baseline removal) and dimensionality reduction (principal component analysis (PCA), factor analysis (FA), independent component analysis (ICA), locally linear embedding (LLE), non-negative matrix factorization (NMF), isomap), and prediction models (partial least-squares regression (PLSR), support vector regression (SVR), multilayer perceptron (MLP), convolutional neural network (CNN), gradient boosting) were analyzed for their prediction performance (totaling to 4,711,685 combinations). Results show that the highest quantification performance was achieved at a root-mean-square error of prediction (RMSEP) of 0.0254 (dimensionless mass ratio), using min-max scaling and factor analysis for data preprocessing and multilayer perceptron for prediction. Effects of preprocessing, comparison of prediction models, and linearity of data are discussed.

Fractional Variation Network for THz Spectrum Denoising without Clean Data

Deep learning can remove the noise of the terahertz (THz) spectrum via its powerful feature extraction ability. However, this technology suffers from several limitations, including clean training data being difficult to obtain, the amount of training data being small, and the restored effect being unsatisfactory. In this paper, a novel THz spectrum denoising method is proposed. Low-quality underwater images and transfer learning are used to alleviate the limitation of the training data amount. Then, the principle of Noise2Noise is applied to further reduce the limitations of clean training data. Moreover, a THz denoising network based on Transformer is proposed, and fractional variation is introduced in the loss function to improve the denoising effect. Experimental results demonstrate that the proposed method estimates the high-quality THz spectrum in simulation and measured data experiments, and it also has a satisfactory result in THz imaging.

Artificial neural networks for material parameter extraction in terahertz time-domain spectroscopy

Terahertz time-domain spectroscopy (THz-TDS) is a proven technique whereby the complex refractive indices of materials can be obtained without requiring the use of the Kramers-Kronig relations, as phase and amplitude information can be extracted from the measurement. However, manual pre-processing of the data is still required and the material parameters require iterative fitting, resulting in complexity, loss of accuracy and inconsistencies between measurements. Alternatively approximations can be used to enable analytical extraction but with a considerable sacrifice of accuracy. We investigate the use of machine learning techniques for interpreting spectroscopic THz-TDS data by training with large data sets of simulated light-matter interactions, resulting in a computationally efficient artificial neural network for material parameter extraction. The trained model improves on the accuracy of analytical methods that need approximations while being easier to implement and faster to run than iterative root-finding methods. We envisage neural networks can alleviate many of the common hurdles involved in analyzing THz-TDS data such as phase unwrapping, time domain windowing, slow computation times, and extraction accuracy at the low frequency range.

A Generalized Regression Neural Network Model for Predicting the Curing Characteristics of Carbon Black-Filled Rubber Blends

In this study, a new generalized regression neural network model for predicting the curing characteristics of rubber blends with different contents of carbon black filler cured at various temperatures is proposed for the first time The carbon black contents in the rubber blend and cure temperature were used as input parameters, while the minimum and maximum elastic torque, scorch time, and optimal cure time, obtained from the analysis of 11 rheological cure curves registered at 10 various temperatures, were considered as output parameters of the model. A special pre-processing procedure of the experimental input and target data and the training algorithm is described. Less than 55% of the experimental data were used to significantly reduce the total number of input and target data points needed for training the model. Satisfactory agreement between the predicted and experimental data, with a maximum error in the prediction not exceeding 5%, was found. It is concluded that the generalized regression neural network is a powerful tool for intelligently modelling the curing process of rubber blends even in the case of a small dataset, and it can find a wide range of practical applications in the rubber industry.

An effective approach for breast cancer diagnosis based on routine blood analysis features

Breast cancer is a widespread disease and one of the primary causes of cancer mortality among women all over the world. Computer-aided methods are used to assist medical doctors to make early diagnosis of the disease. The aim of this study is to build an effective prediction model for breast cancer diagnosis based on anthropometric data and parameters collected through routine blood analysis. The proposed approach innovatively exploits principal component analysis (PCA) technique cascaded by median filtering so as to transform original features into a form of containing less distractive noise not to cause overfitting. Since a generalized regression neural network (GRNN) model is adopted to classify patterns of the transformed features, the computational load imposed in the training of artificial neural network model is kept minimized thanks to the non-iterative nature of GRNN training. The proposed method has been devised and tested on the recent Breast Cancer Coimbra Dataset (BCCD) that contains 9 clinical features measured for each of 116 subjects. Outperforming all of the existing studies on BCCD, our method achieved a mean accuracy rate of 0.9773. Experimental results evidence that this study achieves the best prediction performance ever reported on this dataset. The fact that our proposed approach has accomplished such a boosted performance of breast cancer diagnosis based on routine blood analysis features offers a great potential to be used in a widespread manner to detect the disease in its inception phase. Graphical abstract.