堆叠监督自动编码器的近红外光谱建模

近红外光谱中包含了物质中有机分子含氢基团的特征信息,具有维度高、冗余大等特点.传统的基于浅层校正模型,比如主成分回归、偏最小二乘回归、人工神经网络、支持向量回归等,无法提取近红外光谱数据深层的信息.提出一种基于堆叠监督自动编码器的近红外光谱建模方法,不仅可以拟合光谱数据与理化值之间复杂的非线性关系,还可以提取数据深层的...

Near-Infrared Spectral Modeling Based on Stacked Supervised Auto-Encoder

The near-infrared spectrum contains the characteristic information of the hydrogen-containing groups of organic molecules in the substance, and it has the characteristics of high dimensionality and large redundancy. Traditional near-infrared spectroscopy techniques are based on shallow correction models, such as principal component regression, partial least squares regression, artificial neural networks, support vector regression etc., which cannot extract the deep information of the spectral data. This paper proposes a near-infrared spectroscopy modeling method based on stacked supervised autoencoders, which can fit the complex non-linear relationship between spectral data and target physicochemical values and extract the deep feature information of the data. First, the optimal preprocessing method is selected by comparing the effects of different spectral preprocessing on the model prediction results. Then the correlation coefficient method is used to extract the characteristic bands of the preprocessed spectrum. The method uses the processed near-infrared spectrum data as the input signal. Then use the target physicochemical values to perform supervised pre-training on multiple supervised autoencoders, and stack multiple pre-trained supervised autoencoders. The stacked supervised autoencoder is obtained, the pre-trained parameters are used as the initialization parameters of the stacked supervised autoencoder, and then the target physicochemical values are used to supervise and fine-tune the stacked supervised autoencoder. Finally the optimal parameters of the model are obtained. Established partial least squares regression prediction model, artificial neural networks prediction model, stack auto-encoder prediction model and stack supervised auto-encoder prediction model on the corn water content data and the total acid content data of yellow wine respectively, verifying the feasibility of stack supervised auto-encoder modeling. The root means square error and residual prediction deviation are employed to evaluate model performance. The accuracy of four modeling methods of partial least squares regression, backpropagation- artificial neural networks, stack auto-encoder, and stack supervised auto-encoder are compared and analyzed. The analysis results show that the model established by stack supervised auto-encoder has a good prediction effect. The two evaluation indexes of the corn water content data set reached 0.061 1 and 4.271; the two evaluation indexes of rice wine’s total acid content data reached 0.126 6 and 4.006, excellent for the other three methods.