油页岩含油率近红外光谱原位分析方法研究

为实现油页岩含油率的原位检测，采用便携式近红外光谱分析技术，针对吉林扶余油页岩基地2号钻井的66个岩芯样品开展了原位检测的分析建模方法研究。采用自制便携式近红外光谱仪器获得反射率、吸光度、K-M函数三种数据形式光谱数据，结合主成分-马氏距离(PCA-MD)剔除异常样品、无信息变量消除法(UVE)波长筛选及二者组合的四种建模数据优化方法，采用相同的数据预处理方法进行偏最小二乘(PLS)和反向传播神经网络(BPANN)两种方法的建模分析研究，确定最佳分析模型及方法。结果表明(1)不论是否采用四种不同的数据优化方法，两种建模方法所用建模数据库适合采用反射率或K-M函数的光谱数据形式;(2)两种建模方法，采用四种不同的数据优化方法，对相同数据库建模精度的影响不同：采用PLS建模方法、以PCA-MD和UVE+ PCA-MD两种方法进行数据优化、可以提高K-M函数光谱数据形式数据库的建模分析精度，采用BPANN建模方法、以UVE、PCA-MD 与UVE组合的 三种方法进行数据优化、对三种数据形式数据库的建模精度均有所提高;(3)除以反射率光谱数据并进行PCA-MD数据优化外，采用BPANN方法的建模精度好于PLS法。其中采用反射率光谱数据形式、只进行UVE数据优化外的BPANN建模精度最高，预测相关系数为0.92、标准偏差为0.69%。

Analyzing and Modeling Methods of Near Infrared Spectroscopy for In-situ Prediction of Oil Yield from Oil Shale

In order to in-situ detect the oil yield of oil shale, based on portable near infrared spectroscopy analytical technology, with 66 rock core samples from No.2 well drilling of Fuyu oil shale base in Jilin, the modeling and analyzing methods for in-situ detection were researched. By the developed portable spectrometer, 3 data formats (reflectance, absorbance and K-M function) spectra were acquired. With 4 different modeling data optimization methods: principal component-mahalanobis distance(PCA-MD) for eliminating abnormal samples, uninformative variables elimination (UVE) for wavelength selection and their combinations: PCA-MD+UVE and UVE+PCA-MD, 2 modeling methods: partial least square(PLS)and back propagation artificial neural network (BPANN), and the same data pre-processing, the modeling and analyzing experiment were performed to determine the optimum analysis model and method. The results show that the data format, modeling data optimization method and modeling method all affect the analysis precision of model. Results show that whether or not using the optimization method, reflectance or K-M function is the proper spectrum format of the modeling database for two modeling methods. Using two different modeling methods and four different data optimization methods, the model precisions of the same modeling database are different. For PLS modeling method, the PCA-MD and UVE+PCA-MD data optimization methods can improve the modeling precision of database using K-M function spectrum data format. For BPANN modeling method, UVE, UVE+PCA-MD and PCA-MD+UVE data optimization methods can improve the modeling precision of database using any of the 3 spectrum data formats. In addition to using the reflectance spectra and PCA-MD data optimization method, modeling precision by BPANN method is better than that by PLS method. And modeling with reflectance spectra, UVE optimization method and BPANN modeling method, the model gets the highest analysis precision, its correlation coefficient (R_P) is 0.92, and its standard error of prediction (SEP) is 0.69%.