论文部分内容阅读
为了解决混合气体多组分间特征吸收峰相互重叠引起的特征选择困难问题,提出了新型红外光谱特征选择方法,并对该方法的性能进行了分析与评价。首先,充分结合思维进化计算的并行机制、异化操作与蝙蝠算法的局部搜索能力,设计了思维进化蝙蝠算法。接着,通过实验采集两个混合气体数据库,利用思维进化蝙蝠算法对其目标组分的特征峰进行筛选。然后,从算法的收敛速度和筛选出的特征峰两个方面,将思维进化蝙蝠算法与基本蝙蝠算法、遗传算法、粒子群优化算法及并行萤火虫群优化算法等进行比较。最后,讨论了思维进化蝙蝠算法与无信息变量消除法相结合对结果的影响。实验结果表明:CO的特征峰范围包括2 090~2 110 cm-1和2 115~2 125 cm-1,共包含32个波长点;N2O的特征峰范围为2 225~2 250 cm-1,共包含26个波长点。利用筛选出的特征波长点建立的浓度反演模型,测试集均方根误差为0.155,决定系数可达0.908。实验结果表明:思维进化蝙蝠算法收敛速度快、全局搜索能力强,适用于存在重叠特征峰的混合气体的特征选择,对应的浓度反演模型的泛化性能也有显著提升。
In order to solve the problem of feature selection caused by the overlapping of the characteristic absorption peaks of mixed gases, a new method of infrared spectral feature selection was proposed and the performance of the method was analyzed and evaluated. First of all, the algorithm of thinking evolutionary bat is designed based on the parallel mechanism of thinking evolutionary computation, the alienation operation and the local searching ability of bat algorithm. Then, two mixed gas databases were collected experimentally, and the characteristic peak of the target component was screened by using the evolutionary bat algorithm. Then, compared with the basic bat algorithm, the genetic algorithm, the particle swarm optimization algorithm and the parallel firefly swarm optimization algorithm, the evolutionary bat algorithm is compared with the convergence speed of the algorithm and the selected characteristic peaks. Finally, we discuss the impact of combining evolutionary bat algorithm with no-information variable elimination on the results. The experimental results show that the characteristic peaks of CO range from 2 090 to 2 110 cm-1 and from 2 115 to 2 125 cm-1, and contain 32 wavelengths. The characteristic peaks of N2O range from 2225 to 2250 cm- A total of 26 wavelength points. Using the filtered concentration inversion model, the root mean square error of the test set is 0.155 and the determination coefficient is up to 0.908. The experimental results show that the thought evolutionary bat algorithm has the advantages of fast convergence and strong global search ability, and is suitable for the feature selection of gas mixtures with overlapping peaks. The generalization performance of the corresponding concentration inversion model is also significantly improved.