采场覆岩光纤监测数据LSSVM填补方法

LSSVM method of missing data imputation of optical fiber monitoring with mining-induced overburden

完备的光纤监测数据是智能开采中矿压显现前兆信息识别、上覆岩层变形预测的基础,而实际得到的监测数据大多是不完整的。为有效填补光纤监测数据的缺失值,文中以采场覆岩光纤监测物理模拟实验中光纤传感器采集的数据为基础,分析了缺失数据的特征,建立了多测点单属性小样本缺失数据的最小二乘支持向量机(LSSVM)缺失数据填补方法。并将LSSVM与BP神经网络、3次样条插值等方法,在Fv11,Fv12光纤的6个不同数据集上,按照离散型、连续型、混合型3种数据缺失类型并产生不同缺失率,进行对比实验。针对离散型随机产生20%缺失数据,LSSVM,BP神经网络、3次样条插补方法的均方根误差(RMSE)平均值分别为0.0032,0.0056,0.0069,最大偏离量(MDE)平均值分别为0.012,0.022,0.028;针对连续型随机产生36%缺失数据,3种不同方法的RMSE平均值分别为0.0061,0.0077,0.0090,MDE平均值分别为0.021,0.028,0.041;前2类实验结果表明LSSVM方法均优于其他2种缺失值插补方法。当随机产生兼具离散和连续型缺失且缺失比例不同时,缺失比例小于30%时LSSVM方法略优于其他2种方法,当缺失率大于36%时LSSVM明显优于其他2种方法。综合所有实验结果表明,LSSVM插补方法对单属性小样本缺失数据填补是一种简单有效的填补方法。

Continuity,integrity and accuracy of monitoring data for optical fiber is a very foundation that precursory information is identified and predicted about overburden deformation and dynamic phenomenon during mining.In order to implement the imputation of missing value effectively which has the characteristic of multipoint sampling and local small sample during collection of fiber sensing,a novel algorithm model was proposed named LSSVM to deal with missing data imputation.The data set was obtained from the fiber sensing in physical simulation experiment of monitoring overburden deformation,with the characteristics of missing data analyzed.In this paper,a comparison has been made of the LSSVM with BP network imputation model and cubic spline interpolation imputation algorithm on two different fibers Fv11 and Fv12 about 6 data sets.The missing data was generated randomly by the different missing ratios.Firstly,random 20%missing data was generated for discrete type.The experimental results show that the average of RMSE are 0.0032,0.0056,0.0069,and the average of MDE are 0.012,0.022,0.028 by using LSSVM,BP network and Cubic spline method.Secondly,random 36%missing data was generated for continuous type.The experimental results show that the average of RMSE are 0.0061,0.0077,0.0090,and the average of MDE are 0.021,0.028,0.041 about three methods.Accordingly the LSSVM model is better than the other two methods.Finally,the missing data for both discrete and continuous missing characteristics was generated with different missing rates.The results show that the LSSVM model is slightly better than the other two methods with the missing ratio by less than 30%.When the missing ratio is 50%,LSSVM is significantly better than the other two methods.All the experimental results show that the proposed method is effective for filling missing data.