【摘要】 采集74份标准水样进行紫外可见波段全光谱扫描,结合Savitzky-Golay(SG)平滑算法、标准正态变换(SNV)、一阶微分(1st D)等6种方法对提取的光谱数据进行去噪处理,然后采用半监督近邻传播算法(SAP)、连续投影算法(SPA)、无信息变量消除算法(UVE)进行特征波长的选择。基于全光谱法建立了偏最小二乘(PLS)模型,基于特征波长建立了极限学习机(ELM)模型,另外把PLS回归模型得到的主成分作为支持向量机回归(SVR)、BP和RBF神经网络的输入建立了PCA+SVR、PCA+BP和PCA+RBF模型。结果表明:使用主成分分析结合RBF神经网络建立的PCA+RBF预测模型效果最优,其相对误差最稳定并保持在较低水平,测量上限高达数百mg/L,为实现水体中硝酸盐氮的在线检测和其他水质参数的检测奠定了基础。更多还原

【Abstract】 A total of 74 standard water species are collected. By ultraviolet visible spectrum full spectrum scanning, combined with the Savitzky-Golay smoothing algorithm(SG), standard normal transformation(SNV), first order differential(1 st D) and so on, 6 kinds of methods are used to remove noise spectrum data processing, and then using a semi-supervised nearest propagation algorithm(SAP), successive projection algorithm(SPA) and uninformed variable elimination algorithm(UVE) to select the characteristic wavelengths. Partial least squares(PLS) is used to build models with the full spectra, and extreme learning machine(ELM) is applied to build models with the selected wavelength variables. In addition, PCA+SVR, PCA+BP and PCA+RBF models are established by using principal components of PLS regression model as the input of support vector machine regression(SVR), BP and RBF neural networks. The results showed that the PCA+RBF model based on principal component analysis and RBF neural network prediction model has the best effect, and its relative error is stable and remains a low level. The upper limit of measurement is up to several hundred mg/L, which lays a certain foundation for the detection of nitrate nitrogen and other water quality parameters in the future.更多还原