改进的SVM在矿井气体定量分析中的应用

自行搭建了气体采集系统,根据井下的气体情况,采集了包括甲烷、乙烷、丙烷、正丁烷和二氧化碳五种气体的中红外光谱数据共236组,其中校正集186组,验证集50组。在对光谱数据进行预处理之后,利用主成分分析技术将得到的主吸收峰区域的红外光谱数据进行降维处理,通过特征提取得到3个特征值作为矿井气体光谱数据的输入量。该方法有效减少了模型的计算量,加快了模型的收敛速度。然后,利用改进支持向量机分别对这五种气体建立了定量分析模型。为提高该算法的预测精度,利用遗传算法和粒子群优化算法分别对SVM参数进行参数寻优。最后,选择优化效果更好的粒子群算法,并通过验证集对这五种气体进行了浓度预测分析。实验结果表明:五种气体浓度预测结果的平均误差均小于1.78%,最大误差均小于4.98%,且对于50组的气体预测耗时均小于103 s。表明所提出的改进的SVM算法能够准确、快速地预测矿井气体浓度,对实现矿井气体检测有着积极的意义。

Application of improved SVM in quantitative analysis of mine gas concentration

A quantitative analysis model of mine gas concentration based on improved support vector machine (SVM) was adopted. Five mine gases were used for experiment, which included methane, ethane, propane, n-butane and carbon dioxide. Mid-infrared spectral data of these five gases and mixed gases were collected with Fourier infrared spectrometer. 236 groups of these mixed gases were divided into 186 groups for calibration set and 50 groups for validation set. Principal component analysis (PCA) was used to reduce the dimensionality of the infrared spectral data, and 3 eigenvalues were extracted as input, which helped to improve convergence speed and reduce calculation time. Particle swarm optimization (PSO) and genetic algorithm (GA) were used to optimize parameters of support vector machine (SVM) method respectively, and PSO was adopted for its better optimization effect over GA. The mixed gases were detected through this algorithm, and experiment results show that the average errors of concentration predictions of five gases are all less than 1.78%, and the maximum errors of concentration predictions of five gases are all less than 4.98%. The time cost for concentration prediction is all less than 103 s for the 50 groups. This suggested that the improved SVM method based on PSO can be used to predict the gas concentration accurately, and can meet the requirement of real-time detection of mine gases, which has great value in the study of concentration prediction of mine gases.