基于深度自编码的大曲高光谱数据解混研究

Decontamination of Daqu hyperspectral data based on depth autoencoder

利用高光谱技术检测大曲发酵品质时,获取的水分等高含量物质的高光谱数据可能掩盖对大曲质量评价至关重要的微量物质高光谱数据。为方便后续更微量物质的光谱曲线分解,需先排除水分等高含量物质的数据干扰,该文通过建立无监督的深度自编码模型,可实现大曲水分高光谱曲线分解。通过实验设计,采集与水混合后的成品曲粉光谱数据。首先编码部分,将混合大曲光谱曲线压缩为低维表示,即端元;解码部分,将光谱的低维表示解压重构为原始光谱曲线,结合比较不同目标函数,反向传递重构误差,更新解码权重;最终,通过端元解码出水与曲粉各自的光谱曲线,运用欧氏距离与皮尔逊相关系数方法从特征距离和相关系数两方面同时评价解混效果。实验显示,利用深度自编码解混模型,选择L-C目标函数得到的解混效果最好,解混曲线与纯曲粉曲线的欧式距离与皮尔逊相似度分别为0.3427和0.9967。研究表明,利用深度自编码网络能够对大曲高光谱数据进行解混,可为实现大曲高光谱微量物质检测提供理论支持和技术支撑。

When hyperspectral technology is used to detect the fermentation quality of Daqu,the obtained hyperspectral data of high-content substances,such as water,may conceal the hyperspectral data of trace substances,which are essential for the quality evaluation of Daqu.In order to eliminate the interference of water and other high-content substances and decompose the spectral curve of trace substances,the unsupervised deep autoencoder model was established to decompose the water hyperspectral curve of Daqu.Through the experimental design,the spectral data of the mixture of Daqu powder and water were collected.Firstly,in the coding part,the spectral curve of mixed Daqu was compressed into a low dimensional representation,that is,the end member;in the decoding part,the low dimensional representation of the spectrum was decompressed and reconstructed into an original spectral curve,and the decoding weight could be updated by comparing different objective functions and transferring the reconstruction error in reverse;finally,the spectral curves of water and Daqu powder were decoded through the end member,then the Euclidean distance and Pearson correlation coefficient method were used to evaluate the effect of the solution from two aspects of characteristic distance and correlation coefficient at the same time.The experiment showed that using deep autoencoder model and choosing the L-C objective function could get the best result,the Euclidean distance and Pearson similarity of the demixed curve and the pure powder curve were 0.3427 and 0.9967,respectively.The results showed that the deep autoencoder network can be used to demix the hyperspectral data of Daqu,which can provide theoretical and technical support for the detection of trace substances of Daqu hyperspectral.