论文部分内容阅读
目的:基于多参考相关系数法和误差反向传递人工神经网络(BP-ANN)建立矿物药紫石英的近红外光谱定性模型,用于紫石英的生品、煅制品、醋煅品和伪品的快速鉴别。方法:采集紫石英、紫石英煅制品、紫石英醋煅品及紫石英伪品这4类不同紫石英样品的近红外光谱,对光谱进行二阶导数和9点平滑预处理,计算多项相关系数。在Matlab 2014a软件中将多项相关系数作为输入数据,以BP-ANN建立4类不同紫石英样品的快速鉴别模型。结果:建立了紫石英近红外光谱BP-ANN鉴别模型,模型验证结果显示,15批验证样品中14批样品预测结果正确,仅1批样品预测有误,准确率达93.33%。结论:建立的紫石英近红外光谱BP-ANN鉴别模型能通过一次性整合的运算区分紫石英生品、煅制品、醋煅品及其伪品,鉴别结果准确可靠。此外,模型在近红外光谱相关系数法基础上,以多个参考光谱为对照计算所得的多组相关系数作为网络特征输入数据,实现了光谱数据的压缩。
OBJECTIVE: To establish a near-infrared spectroscopy qualitative model of mineral medicine purple quartz based on multi-reference correlation coefficient method and error feedback artificial neural network (BP-ANN), which is used to produce raw materials, calcined products, calcareous products and counterfeit products of purple quartz Quick identification. Methods: Near infrared spectra of four different samples of amethyst, including purple quartz, purple quartz calcined product, purple quartz calcareous product and purple quartz counterfeit sample, were collected. The spectra were subjected to second derivative and 9-point smoothing to calculate the correlations coefficient. In Matlab 2014a software, a number of correlation coefficients were taken as input data, and four kinds of rapid identification models of different amethyst samples were established by BP-ANN. Results: The identification model of BP-ANN was established by near infrared spectroscopy. The results of model validation showed that the predictive results of 14 batches of samples from 15 batches of batches were correct. Only one batches of batches of pigs were predicted incorrectly with the accuracy of 93.33%. Conclusion: The established BP-ANN discriminant model of purple quartz near infrared spectroscopy can distinguish the raw products of purple quartz, calcined products, vinegar calcareous products and their counterfeit products through the one-time integrated operation, and the identification results are accurate and reliable. In addition, based on the correlation coefficient method of near-infrared spectrum, the model uses the multiple correlation coefficients calculated by multiple reference spectra as the input data of the network features to achieve the compression of spectral data.