The content of rubber additives is an important determinant of rubber quality. Current testing methods for these additives include combustion testing, chemical analysis, chromatography, and infrared spectroscopy. However, these detection techniques present challenges such as intricate pre-processing, time-intensive operations, laborious procedures, and potential inaccuracies in reflecting the genuine additive content. These types of limitations hinder their ability to cater to the growing demand for swift, precise, and non-destructive detection in rubber. This is a significant challenge for the advancement of the rubber industry in China. Furthermore, when analyzing multi-component mixtures, the absorption spectra can overlap and become distorted, leading to unreliable results. In this study, we leverage terahertz time-domain spectroscopy, data fusion, and chemometrics to quantitatively assess additives in five-component mixtures. This offers an innovative approach for detecting and analyzing the content of target components in multi-component mixtures of rubber and its auxiliaries.

In this study, a five-component mixture composed of NBR, silica, zinc oxide, antioxidant H, and antioxidant MB was used as an experimental sample. The terahertz time-domain spectroscopy system was utilized to capture and compute the absorption spectra of the five-component mixture within the range of 0.3?1.6 THz and to analyze its spectral characteristics. The derivative spectral data of the sample were derived by taking the first-order derivatives. Initially, the KS algorithm was employed to segment the sample set data, which was then quantitatively analyzed using partial least squares regression and support vector machine regression models. Subsequently, three data fusion methods were employed to process the data. Specifically, the low-level data fusion directly combined the absorption spectrum data with the derivative spectrum; the mid-level data fusion merged variables after feature extraction via the Monte Carlo uninformative variable elimination and successive projections algorithm; and the high-level data fusion was executed using multiple linear regression. Finally, the predictive accuracy of the models was assessed based on the correlation coefficient and root mean square error.

Through the absorption spectra of five pure substances ― NBR, silica, zinc oxide, antioxidant H, and antioxidant MB ― it is evident that there are noticeable absorption peaks within the range of the analyzed frequency band for all five pure substances (Fig.3). The absorption spectra of the five-component mixtures are averaged individually for each proportion. It is observable that as the content of antioxidant MB in the mixtures increases, the absorbance also rises, suggesting a linear relationship between the absorption spectra of the mixtures and content of antioxidant MB (Fig.4). The full absorption spectra of the five-component mixtures reveal complexity in the mixtures, with overlapping and some distortion (Fig.5). The comparison between the predicted and reference values of the antioxidant MB content in the prediction set reveals that SVR aligns more closely with the actual value than PLSR does when predicting the antioxidant MB content in the five-component mixtures. This indicates that the SVR model predicts more effectively (Fig.6). Both the correlation coefficient and root mean square error demonstrate that SVR predicts with superior accuracy, suggesting a non-linear relationship between the content of antioxidant MB in the five-component mixture and absorbance (Table 2). Based on the SVR model, when comparing the prediction results of absorption spectra to derivative spectra, it is found that the analytical results for the content of antioxidants MB from absorption spectra fluctuate less (Fig.7). In comparing the correlation coefficients and root mean square errors of absorption and derivative spectra using the SVR model, the prediction accuracy for antioxidant MB content from absorption spectra is higher, indicating a superior predictive capability of absorption spectra (Table 3). The comparison between the predicted and reference values of antioxidant MB content for the data fusion prediction set demonstrates that the data fusion model predicts significantly better than a single spectrum, suggesting that the data fusion method enhances the model’s predictive performance (Fig.8). The predictive accuracy of the Monte Carlo-based uninformative variable elimination method for mid-level data fusion surpasses the accuracy of the single spectrum and other data fusions (Table 4).

In the current study, a new method for rapid detection of antioxidant MB content in rubber multi-component mixtures is investigated using terahertz time-domain spectroscopy, MCUVE mid-level data fusion, and SVR. Analysis of the absorption spectra of the five-component mixtures and quantitative analytical models reveals linear and non-linear relationships between the absorbance of the mixtures and antioxidant MB content. Results from quantitative analyses, which combine data fusion methods based on SVR, indicate that prediction accuracy and stability of all four data fusion methods significantly surpass that of a single spectrum. Specifically, the prediction performance of MCUVE mid-level data fusion is the best. In conclusion, the combination of terahertz time-domain spectroscopy, data fusion methods, and SVR modeling addresses the shortcomings of existing rubber and additive detection methods and the accuracy challenges posed by overlapping and distortion phenomena in the absorption spectra of multi-component mixtures. This approach holds significant scientific value and promises substantial market application potential.