小波变换与神经网络融合法在油页岩近红外光谱分析中的应用

便携式近红外光谱分析技术可实现油页岩含油率的原位检测，在油页岩资源现场勘查中发挥着重要作用。但是，由于其测得的原始光谱数据量大、冗余信息多，直接建模会影响速度与精度。因此提出一种小波变换与神经网络融合法，先将油页岩全谱数据进行db8小波3级分解，提取其近似系数形成输入矩阵，然后再进行神经网络建模。为了验证有效性，利用30个油页岩合成样品，从中随机选择20个用于训练，另外10个用于预测，并分别使用全谱数据与小波特征数据进行了10次神经网络建模。结果表明，全谱数据建模速度均值为570.33 s，预测残差平方和及相关系数均值分别为0.006 012及0.843 75;而小波神经网络法对应的以上均值为3.15 s, 0.002 048及0.953 19。由此说明小波神经网络法优于全谱数据建模法，为油页岩含油率的快速、高精度检测提供了一种新方法。

Application of Wavelet Transform and Neural Network in the Near-Infrared Spectrum Analysis of Oil Shale

In the present study, an innovative method is proposed, employing both wavelet transform and neural network, to analyze the near-infrared spectrum data in oil shale survey. The method entails using db8 wavelet at 3 levels decomposition to process raw data, using the transformed data as the input matrix, and creating the model through neural network. To verify the validity of the method, this study analyzes 30 synthesized oil shale samples, in which 20 samples are randomly selected for network training, the other 10 for model prediction, and uses the full spectrum and the wavelet transformed spectrum to carry out 10 network models, respectively. Results show that the mean speed of the full spectrum neural network modeling is 570.33 seconds, and the predicted residual sum of squares (PRESS) and correlation coefficient of prediction are 0.006 012 and 0.843 75, respectively. In contrast, the mean speed of the wavelet network modeling method is 3.15 seconds, and the mean PRESS and correlation coefficient of prediction are 0.002 048 and 0.953 19, respectively. These results demonstrate that the wavelet neural network modeling method is significantly superior to the full spectrum neural network modeling method. This study not only provides a new method for more efficient and accurate detection of the oil content of oil shale, but also indicates the potential for applying wavelet transform and neutral network in broad near-infrared spectrum analysis.