论文部分内容阅读
目前由于部分电站锅炉所用到的燃煤大多为配煤,在有些情况下,其复杂的物理化学特性导致难以获得高精确度的常规近红外定量分析模型,这给电煤发热量的在线检测带来了一定困难。针对该问题,深入研究了电煤近红外光谱时域和频域特征,提出一种能够通过全局优化策略自动在频域内建立最优近红外定量分析模型的新方法—频域自适应分析法。该方法首先将时域近红外光谱通过快速傅里叶变换转换为频域近红外信号;然后采用有效光谱能量率得到合适的频域信息范围;接着根据近红外光谱频域下的相关系数谱图、方差谱图以及谐波在频域中的坐标合理构建了频域信息量评价参数,利用该参数对模型输入变量的种群位置进行初始化;最后采用频域分区搜索和综合性能评价函数得到最佳建模方案。与此同时,结合电煤煤粉近红外图谱的特性,并以其发热量为待测目标对该方法进行了验证,取得相对较好实验效果,与传统方法主成分回归、偏最小二乘回归、反向传播神经网络以及基于遗传算法的偏最小二乘回归和支持向量机回归相比,该方法预测精度更高,并且有效避免了频域随机搜索潜在的过拟合和虚假有效模型的弊端,具有良好的应用前景。此外,该方法也可推广用于其他类型的光谱定量分析。
At present, most coal-fired boilers used for coal-fired power stations are mostly coal blending. In some cases, their complex physicochemical properties make it difficult to obtain high-precision conventional NIR quantitative analysis models. This gives an online detection zone Some difficulties come. In order to solve this problem, the near-infrared spectroscopy has been studied deeply in time and frequency domain, and a new method of frequency-domain adaptive analysis, which can automatically establish the optimal near-infrared quantitative analysis model through global optimization strategy, is proposed. The method first converts the near-infrared (NIR) spectrum to near-infrared (NIR) signals by fast Fourier transform (FTIR), and then uses the effective spectral energy rate to obtain the appropriate range of frequency domain information. Then, based on the correlation coefficient spectrum in the near- , Variance spectrum and harmonics in the frequency domain, the information parameters in frequency domain are reasonably constructed. The parameters of the model are used to initialize the population position of the model input variables. Finally, the best performance is obtained by using the frequency domain partitioning search and the comprehensive performance evaluation function Modeling program. At the same time, combining with the characteristics of near infrared spectrum of coal and coal, the method was validated by its calorific value as the target to be tested, and achieved relatively good experimental results. Compared with the traditional method of principal component regression and partial least-squares regression , Backpropagation neural network and partial least squares regression based on genetic algorithm and support vector machine regression, this method has higher prediction accuracy and effectively avoids the shortcomings of the random search of potential over-fitting and false effective model in the frequency domain , Has a good application prospects. In addition, this method can also be generalized for other types of spectroscopic quantitative analysis.