基于SVM的混合气体分布模式红外光谱在线识别方法

Method of Infrared Spectrum On-Line Pattern Recognition of Mixed Gas Distribution Based on SVM

针对混合气体组分浓度分析中海量训练样本的获取、分析精度及实时在线分析等问题,将支持向量机这一新的信息处理方法和红外光谱分析法结合,提出了混合气体分布模式的概念。在此基础上,采用先进行混合气体分布模式识别,然后再进行混合气体分析的思路,在大量调查的基础上,研究探索了实际应用中可能出现的混合气体分布模式,确定60种混合气体分布模式,共计6000个混合气体红外光谱数据样本用于模型的训练与检验。采用SMO算法实现了减量和增量的在线学习,最终建立了基于SVM的混合气体分布模式红外光谱在线识别模型。模型由模式识别和结果输出2层组成,模式识别层完成混合气体模式分布模式识别任务;结果输出层由60个SVM校正模型组成,完成具体的浓度分析任务。实验结果表明,该方法对混合气体分布模式的正确识别率不低于98.8%,可在小样本条件下对混合气体的分布模式进行在线识别,可在线实时加入新的混合气体分布模式,具有实际应用价值。

In order to solve the difficulties that the spectrum training data samples of the massive mixed gas cannot be actually obtained, the analysis precision is low and it is not real time online analysis in the analysis of mixed gas component concentration, the support vector machine, a new information processing method, was used in the mixed gas infrared spectrum analysis, and the concept of mixed gas distribution pattern was proposed in the present paper. Based on the thought that the mixed gas distribution pattern recognition is carried out first, and then the analysis work of mixed gas component concentration is done, sixty kinds of mixed gas distribution pattern were determined after investigation and study, and 6 000 mixed gas spectrum data samples were used for model training and testing. Sequential minimal optimization algorithm was applied to realize the decrement and the increase of online learning, and finally the model of infrared spectrum online pattern recognition of mixed gas distribution based on SVM was established. The model structure is composed of 2 levels, pattern recognition level and result output level. The pattern recognition level completes the task of mixed gas distribution pattern reeognitiom while the result output level is composed of 60 SVM calibration models, and it completes the task of mixed gas concentration analysis. Experimental results show that the correct recognition rate of mixture gas distribution pattern is not lower than 98.8%, and that the method can be used for online recognition of mixed gas distribution pattern under the conditions of small samples of a mixed gas, and can add new mixed gas online, and it has the practical application value.