AdaBoost算法在矿井突水水源的荧光光谱识别中的研究
|
摘要
矿井突水是影响矿井安全生产的重要因素之一,如果矿井发生突水,能够快速、准确地判别突水水源类型是治理矿井突水灾害保证生产安全的重要环节,因此,建立一个能够快速识别矿井突水水源的模型具有重要的意义。水化学分析法作为在传统的矿井突水水源类型识别方法里应用最为广泛的识别方法,通过获得相应的pH值、离子浓度、电导率等参数,然后利用这些参数来建立突水水源的类型识别模型对矿井突水的类型进行判别。针对这种传统矿井突水水源识别方法在判别时间上耗时长和识别准确率低等不足,鉴于LIF技术具有分析速度快、灵敏度高等优点,提出了将线性判别分析(LDA)算法作为弱分类器的自适应提升(AdaBoost)算法用于激光诱导荧光(LIF)光谱识别矿井突水水源的新方法。用于实验的九种水样(每种水样各取50个样本)由淮南地区某矿的老空水、灰岩水以及按不同比例混合的老空水与灰岩水的七种混合水构成。将405nm激光器发射的激光打入被测水体并采集荧光光谱数据,然后对采集到450组荧光光谱数据进行分析,取其中360组光谱数据(每种水样各40组)用作训练集,取剩余90组光谱数据用作测试集。分别选取三种算法针对水样的激光诱导荧光光谱的分类进行了建模并将三种结果进行对比。首先利用决策树算法对光谱进行分类识别,在节点个数为8时决策树对测试集的分类效果最好,分类准确率达到91.11%。然后针对决策树算法分类效果的不足,利用决策树算法作为弱分类器的AdaBoost算法,当选取节点个数为9的决策树作为弱分类器的时,对训练集的分类准确率为97.78%。最后针对基于决策树的AdaBoost算法的泛化性能不足和为了获得更好的分类效果,提出了基于LDA算法作为弱分类器的AdaBoost算法,在设置迭代次数为150后对水样光谱数据分类准确率可以达到100%。通过实验结果可以发现,集成学习算法的分类能力比传统的分类算法对水样的光谱的分类识别能力更强,相较于同为九个节点的决策树算法,采用节点数为9的决策树作为弱学习器的AdaBoost算法对测试集的分类准确率从88.89%提升到了97.78%,对训练集的分类准确率从99.72%提升到了100%;然后可以发现相对于使用决策树作为弱分类器的AdaBoost算法,采用LDA算法作为AdaBoost算法的弱分类器对水样的光谱的测试集的分类准确率从97.78%提升到了100%,对训练集的分类准确率达到100%,具有更好的识别效果,并且具有更好的泛化性能。实验结果证明采用Adaboost-LDA算法为激光荧光光谱的模式分类用于矿井突水水源的判别和预警是可行且有效的。
The water inrush is one of the most important elements that can influence the mining safety,and being able to recognize the category of water inrush sources accurately and rapidly will greatly enhance the mining safety condition when water inrush happens accidentally.Therefore,it is extremely important and necessary to create a model system that can recognize water inrush sources effectively.The water chemistry analytical method is the widest used method to recognize water inrush sources among traditional methods;in this method,we build a model system by using ph,ionic concentration,conductivity and so on,then use that model system to recognize water inrush sources.However,the water chemistry analytical method has disadvantages that usually be time-costing and of low accuracy.This essay will deal with this problem and introduce the AdaBoost method that uses LDA as weak classifier based on LIF technology because of the rapidnessand high sensitivity of LIF technology.In this research,there are nine kinds of waters from a certain mine in the Huainan City considered and fifty independent samples in each kind of water,limestone water,high pressure water from floor of coal seam and gob areas,and seven different proportion mixture of those two kind of water.Emit laser from the 405 nm laser emitter into laboratory water samples and collect experiment statistics of fluorescence spectrum,analyze these 450 water samples by select 360samples(40samples of each kind of water source)as a training set first and set other 90 samples as a training set.In this essay,we use three different kinds of arithmetic to build three different model systems and compare results from each model system.First of all,we use decision-making tree to recognize and classify different fluorescence spectrum,we get the best outcome and the accuracy rate is 91.11% at that time when the node number is 8.Then,we use the AdaBoost arithmetic and set the decision-making tree as the weak classifier according to the shortage of the decision-making tree,and we get the best accuracy rate of classifying training sets of 97.78% when selecting a decision-making tree whose node number is 9as the weak classifier.And last,we introduce a AdaBoost arithmetic base on setting LDA arithmetic as the weak classifier to get better classifying results according to the generalization shortage of AdaBoost arithmetic which bases on decision-making tree,and finally we get the spectrum accuracy rate of 100% when iterate150 times.As we can get from our experiment,classifying arithmetic that integrates the learning arithmetic is much better than other traditional classifying arithmetic,for instance,compared with the decision-making tree arithmetic,AdaBoost arithmetic which sets decision-making tree as its weak classifier can enhance the accuracy rate of classifying testing set from 88.89% to97.78% and enhance the accuracy rate of classifying training set from 99.72%to 100% when the node number is 9;then compared with the AdaBoost arithmetic which sets decision-making tree as its weak classifier,the AdaBoost arithmetic which uses LDA as its weak classifier can enhance the accuracy rate of classifying sample water fluorescence spectrum testing set from97.78%to 100% and enhance the accuracy rate of classifying sample water fluorescence spectrum training set to 100%as well,and we can get better recognition outcomes and make our model system have better generalization by using such strategy at the same time.Therefore,it is extremely fair to say that using AdaBoost-LDA arithmetic to classify fluorescence spectrum to recognize and alarm water inrush sources is effective and feasible.
The water inrush is one of the most important elements that can influence the mining safety,and being able to recognize the category of water inrush sources accurately and rapidly will greatly enhance the mining safety condition when water inrush happens accidentally.Therefore,it is extremely important and necessary to create a model system that can recognize water inrush sources effectively.The water chemistry analytical method is the widest used method to recognize water inrush sources among traditional methods;in this method,we build a model system by using ph,ionic concentration,conductivity and so on,then use that model system to recognize water inrush sources.However,the water chemistry analytical method has disadvantages that usually be time-costing and of low accuracy.This essay will deal with this problem and introduce the AdaBoost method that uses LDA as weak classifier based on LIF technology because of the rapidnessand high sensitivity of LIF technology.In this research,there are nine kinds of waters from a certain mine in the Huainan City considered and fifty independent samples in each kind of water,limestone water,high pressure water from floor of coal seam and gob areas,and seven different proportion mixture of those two kind of water.Emit laser from the 405 nm laser emitter into laboratory water samples and collect experiment statistics of fluorescence spectrum,analyze these 450 water samples by select 360samples(40samples of each kind of water source)as a training set first and set other 90 samples as a training set.In this essay,we use three different kinds of arithmetic to build three different model systems and compare results from each model system.First of all,we use decision-making tree to recognize and classify different fluorescence spectrum,we get the best outcome and the accuracy rate is 91.11% at that time when the node number is 8.Then,we use the AdaBoost arithmetic and set the decision-making tree as the weak classifier according to the shortage of the decision-making tree,and we get the best accuracy rate of classifying training sets of 97.78% when selecting a decision-making tree whose node number is 9as the weak classifier.And last,we introduce a AdaBoost arithmetic base on setting LDA arithmetic as the weak classifier to get better classifying results according to the generalization shortage of AdaBoost arithmetic which bases on decision-making tree,and finally we get the spectrum accuracy rate of 100% when iterate150 times.As we can get from our experiment,classifying arithmetic that integrates the learning arithmetic is much better than other traditional classifying arithmetic,for instance,compared with the decision-making tree arithmetic,AdaBoost arithmetic which sets decision-making tree as its weak classifier can enhance the accuracy rate of classifying testing set from 88.89% to97.78% and enhance the accuracy rate of classifying training set from 99.72%to 100% when the node number is 9;then compared with the AdaBoost arithmetic which sets decision-making tree as its weak classifier,the AdaBoost arithmetic which uses LDA as its weak classifier can enhance the accuracy rate of classifying sample water fluorescence spectrum testing set from97.78%to 100% and enhance the accuracy rate of classifying sample water fluorescence spectrum training set to 100%as well,and we can get better recognition outcomes and make our model system have better generalization by using such strategy at the same time.Therefore,it is extremely fair to say that using AdaBoost-LDA arithmetic to classify fluorescence spectrum to recognize and alarm water inrush sources is effective and feasible.
引文
[1] MENG Lei,DING En-jie,WU Li-xin(孟磊,丁恩杰,吴立新).Journal of China Coal Society(煤炭学报),2013,38(8):1397.
[2] XU Xing,WANG Gong-zhong(徐星,王公忠).Coal Technology(煤炭技术),2016,35(7):144.
[3] ZHANG Yan(张雁).Coal Science and Technology(煤炭科学技术),2014,42(10):98.
[4] SHAO Liang-shan,LI Yin-chao,XU Bo(邵良杉,李印超,徐波).Journal of Safety and Environment(安全与环境学报),2017,17(5):1730.
[5] LIU Jian-min,WANG Ji-ren,LIU Yin-peng,et al(刘剑民,王继仁,刘银朋,等).Journal of Safety and Environment(安全与环境学报),2015,15(1):31.
[6] YIN Xiao-xi,CHEN Lu-wang,XIE Wen-ping,et al(殷晓曦,陈陆望,谢文苹,等).Hydrogeology&Engineering Geology(水文地质工程地质),2017,44(5):33.
[7] YAN Peng-cheng,ZHOU Meng-ran,LIU Qi-meng,et al(闫鹏程,周孟然,刘启蒙,等).Spectrocopy and Spectral Analysis(光谱学与光谱分析),2016,36(1):243.
[8] JIANG Wei-jian,GUO Gong-de,LAI Zhi-ming,et al(江伟坚,郭躬德,赖智铭,等).Journal of Shangdong University·Engineering Science(山东大学学报·工学版),2014,44(2):43.
[9] ZHU Cheng-zhang,XIANG Yao,ZOU Bei-ji,et al(朱承璋,向遥,邹北骥,等).Journal of Computer-Aided Design&Computer Graphics(计算机辅助设计与图形学学报),2014,26(3):445.
[10] LI Yi-jing,GUO Hai-xiang,LI Ya-nan,et al(李诒靖,郭海湘,李亚楠,等).Systems Engineering-Theory&Practice(系统工程理论与实践),2016,36(1):189.
[11] Tao CHEN,Shijian LU.IEEE Transactions on Vehicular Technology,2016,65(6):4006.
[12] Ignacio Martin-Diaz,Daniel Morinigo-Sotel,Oscar Duque-Perez,et al.IEEE Transactions on Industry Applications,2017,53(3):3066.
[13] Andreas Ranftl,Fernando Alonso-Fernandez,Stefan Karlsson.The Institution of Engineering and Technology,2017,6(6):468.
[2] XU Xing,WANG Gong-zhong(徐星,王公忠).Coal Technology(煤炭技术),2016,35(7):144.
[3] ZHANG Yan(张雁).Coal Science and Technology(煤炭科学技术),2014,42(10):98.
[4] SHAO Liang-shan,LI Yin-chao,XU Bo(邵良杉,李印超,徐波).Journal of Safety and Environment(安全与环境学报),2017,17(5):1730.
[5] LIU Jian-min,WANG Ji-ren,LIU Yin-peng,et al(刘剑民,王继仁,刘银朋,等).Journal of Safety and Environment(安全与环境学报),2015,15(1):31.
[6] YIN Xiao-xi,CHEN Lu-wang,XIE Wen-ping,et al(殷晓曦,陈陆望,谢文苹,等).Hydrogeology&Engineering Geology(水文地质工程地质),2017,44(5):33.
[7] YAN Peng-cheng,ZHOU Meng-ran,LIU Qi-meng,et al(闫鹏程,周孟然,刘启蒙,等).Spectrocopy and Spectral Analysis(光谱学与光谱分析),2016,36(1):243.
[8] JIANG Wei-jian,GUO Gong-de,LAI Zhi-ming,et al(江伟坚,郭躬德,赖智铭,等).Journal of Shangdong University·Engineering Science(山东大学学报·工学版),2014,44(2):43.
[9] ZHU Cheng-zhang,XIANG Yao,ZOU Bei-ji,et al(朱承璋,向遥,邹北骥,等).Journal of Computer-Aided Design&Computer Graphics(计算机辅助设计与图形学学报),2014,26(3):445.
[10] LI Yi-jing,GUO Hai-xiang,LI Ya-nan,et al(李诒靖,郭海湘,李亚楠,等).Systems Engineering-Theory&Practice(系统工程理论与实践),2016,36(1):189.
[11] Tao CHEN,Shijian LU.IEEE Transactions on Vehicular Technology,2016,65(6):4006.
[12] Ignacio Martin-Diaz,Daniel Morinigo-Sotel,Oscar Duque-Perez,et al.IEEE Transactions on Industry Applications,2017,53(3):3066.
[13] Andreas Ranftl,Fernando Alonso-Fernandez,Stefan Karlsson.The Institution of Engineering and Technology,2017,6(6):468.