瓦斯炭黑是橡胶中重要的添加剂之一, 其含量对橡胶性能有重要的影响。 丁腈橡胶是工业生产中应用比较广泛的合成橡胶, 研究丁腈橡胶中瓦斯炭黑的含量具有重要意义。 利用太赫兹时域光谱技术, 对八种不同含量瓦斯炭黑与丁腈橡胶组成的混合物样本中的瓦斯炭黑含量进行测试, 获取了混合物样本在0.3~1.4 THz频段的吸收光谱数据。 分别利用偏最小二乘(PLS)和支持向量回归(SVR)建立混合物中瓦斯炭黑的定量分析模型, 使用均匀梯度法来选择模型的校正集和预测集, 获得瓦斯炭黑预测集的相关系数与均方根误差。 偏最小二乘模型相关系数与均方根误差分别为0.985 8和2.098 9%, 支持向量回归模型相关系数与均方根误差分别为0.998 0和0.785 4%。 实验结果表明, 支持向量回归定量分析模型的预测结果优于偏最小二乘模型。 为进一步证明支持向量回归模型的稳定性, 多次使用随机选择法选择它的校正集和预测集, 并求得其相关系数与均方根误差。 结果表明, 无论是利用均匀梯度法还是随机选择法对支持向量回归定量分析模型的校正集和预测集进行选择, 求得的相关系数和均方根误差均优于偏最小二乘模型。

Gas carbon black (GCB) is one of the important additives in rubber. Its content has an important influence on the performance of rubber. Nitrile butadiene rubber (NBR) is a synthetic rubber used widely in industrial production. It is important to study the content of GCB in NBR. In this paper, the content of GCB in eight kinds of samples consisted of GCB and NBR with different proportion is detected via terahertz time-domain spectroscopy (THz-TDS). Absorption spectra data of these samples is obtained in the frequency ranging from 0.3 to 1.4 THz. Two quantitative analysis models of GCB are established respectively using partial least squares (PLS) method and support vector regression (SVR) method. The uniform gradient method is used to select the calibration set and the prediction set of two models. The correlation coefficient (r) and the root mean square error (RMSE) of two models were calculated. The r and RMSE for the prediction set of PLS model were 0.985 8 and 2.098 9%. The r and RMSE for the prediction set of SVR model were 0.998 0 and 0.785 4%. Experimental results showed that the predictive result of SVR model was better than that of PLS model. In order to prove the stability of the SVR model, we used the random selection method several times to select its calibration set and prediction set, and got their r and RMSE. The results showed that all the r and RMSE of SVR model are better than that of PLS model, whether the uniform gradient method or the random selection method is used to select the calibration set and the prediction set of the SVR model.