随机子空间深度回归方法在紫外光谱水质分析中的应用

紫外光谱法进行TOC浓度分析时存在数量多、维数高等问题。针对此问题，本文提出了一种基于随机子空间深度回归的分析方法，该算法首先采集TOC标准溶液的紫外光谱数据进行预处理，得到吸光度数据，然后在高维数据空间随机选取低维子空间来构造不同的特征子集，并采用深度信念网络对各子集进行特征提取，最后将得到的低维特征进行组合后送入BP神经网络中进行训练，建立TOC浓度反演模型。在构建的水质分析平台上的实验结果表明：本文提出的基于随机子空间深度回归的水质分析方法对每种TOC浓度反演结果的相对误差均在1%以内，且反演结果的稳定性和准确性也要优于常规的水质分析方法。

Random Subspace Deep Regression and its application in Water Quality Analysis of TOC in UV spectroscopy

here are the problems of large quantities and high dimensionality in the analysis of TOC concentration by ultraviolet spectrometry. To solve these problems, this paper proposed a TOC analysis method based on random subspace deep regression. Firstly, the proposed method preprocessed the ultraviolet spectral data of TOC standard solutions to obtain the absorbance data. Then, in the high-dimensional space, it randomly selected the low-dimensional subspace to construct different feature subsets, and extracted the features of each subset by using the deep belief network. Finally, it established the TOC concentration inversion model by BP neural network, which was trained with the discriminant features. Experimental results based on the water quality analysis platform show that the relative errors of TOC concentration inversion results by the proposed method were less than 1%, and its stability is superior to the traditional water quality analysis methods.