Solid powder (flour, gunpowder, pharmaceuticals, etc.) is the most common raw material for industrial production. The proportion consistency of raw material components is one of the most important indicators to guarantee the production process and product qualification rate. Among them, the moisture content of solid powder is an important factor for the disproportionate component ratio. Before actual production, solid powder undergoes transportation and storage, which inevitably exchanges moisture with the environment, resulting in unknown moisture content changes. Therefore, the rapid detection technology of the moisture content of solid powder is significant for controlling industrial production process and product quality. Currently, methods adopted for moisture content testing include weightlessness method, microwave, and near infrared spectroscopy techniques. The weightlessness method is time-consuming and does not meet the needs for rapid testing, while microwave technology requires frequent calibration and maintenance, and is not suitable for measuring flammable and explosive substances. Near-infrared (NIR) spectroscopy is a less penetrating technique reflecting only surface moisture content and is usually not applicable to measurements of low moisture content. Terahertz spectroscopy is a new detection technology developed in recent years, and the detection method based on it features non-destructiveness, rapidness, and efficiency. Water molecules are highly refractive and absorptive in the terahertz bands, making terahertz spectroscopy a technique with unique potential for water content measurement. HMX is a high-performance explosive that is employed as an energy material for missiles, solid propellants, and other strategic weapons. The rapid moisture content detection of HMX powder is vital for the quality control of related weapons.

Firstly, it is necessary to design a pretreatment method for water-containing specimens suitable for terahertz spectroscopic measurements and to design specimen molds applicable for transmission terahertz time-domain spectroscopy systems. The time-domain spectra of samples with different moisture content gradients are measured by the system, and the refractive indices and absorption coefficients of the samples with different moisture concentrations are obtained by preprocessing the time-domain spectral data. Meanwhile, the computed terahertz refractive indices and absorption coefficients are smoothed by the Savitzky-Golay method to reduce prediction errors. Secondly, a water content analysis method based on support vector machine regression (SVR) is investigated. A modeling method incorporating specimen mass and terahertz refractive index is proposed to solve the problem of poor generalization of regression models based on refractive indices or absorption coefficients. The correlation coefficient R² of the prediction set and the root mean square error (RMSE) of the set are utilized as model evaluation indices. On this basis, the genetic algorithm (GA) and particle swarm algorithm (PSO) are leveraged to optimize the regularization parameter (C) and kernel function (γ) for SVR modeling, which further improves the model accuracy.

As shown in Fig. 5, water content shows a significant correlation with both terahertz refractive index and absorption coefficient. Additionally, we build different regression models based on terahertz spectra combined with quality parameters and employ the R²and RMSE of the prediction set as the model evaluation coefficients. The regression modeling results based on terahertz spectroscopy are shown in Fig. 6(a), with a refractive index-based modeling R² of 0.689 and RMSE of 0.221%, and an absorption coefficient-based modeling R² of 0.957 and an RMSE of 0.072%. A fourth set of data is adopted for external validation and the results are shown in Fig. 6(b) to verify the generalization of the above model. The predicted R² based on refractive index is 0.597 with RMSE of 0.243%, and the predicted R² based on absorption coefficient is 0.888 with RMSE of 1.120%. The results show that the accuracy of the external validation results is low, which proves that the above regression model has poor generalization. On this basis, we propose two kinds of regression models for the fusion of mass and terahertz spectral data. The first one is that the mass is directly fused with the terahertz spectral data as feature data, and the second is that the mass is fused with the terahertz spectral data as a scaling factor for scaling. The regression results of the fused model are shown in Fig. 7, with R² of 0.695 and RMSE of 0.219% for the direct fusion modeling based on refractive index, and R² of 0.974 and RMSE of 0.059% for the direct fusion modeling based on absorption coefficient. Scale factor fusion modeling based on the refractive index has an R² of 0.980 and an RMSE of 0.048%, and that based on absorption coefficient has an R² of 0.975 and an RMSE of 0.054%. Additionally, the model is externally validated, and the results are shown in Fig. 8, which indicates that the mass proportional fusion method not only improves the modeling accuracy but also enhances the generalization more effectively. Finally, two optimization algorithms are proposed to optimize the parameters of the SVR regression model based on the fusion model, and the optimized results are shown in Fig. 9. In the figure, both the GA and PSO can further improve the modeling accuracy, with the optimal accuracy of PSO-SVR modeling based on the fused refractive index.

Taking HMX as an example, we investigate a moisture content detection method for solid powder based on terahertz spectroscopy. Currently, terahertz spectroscopy measurements require careful consideration of pre-processing methods to avoid moisture exchange during both the sample preparation and measurement processes. Although different moisture levels can lead to variations in refractive indices and absorption coefficients, models based solely on these parameters tend to exhibit limited generalization. This limitation arises from the fact that variations in HMX volume fraction can also influence terahertz refractive indices and absorption coefficients. Therefore, we propose a modeling approach that integrates mass with refractive indices to enhance the model generalization. Finally, our findings provide valuable insights into moisture content detection in solid powder based on terahertz spectroscopy.