It is wellknown that mine gas gushing forecasting is very significant to ensure the safety of mining. A wavelet-based robust relevance vector machine based on sensor data scheduling control for modeling mine gas gushing forecasting is presented in the paper. Morlet wavelet function can be used as the kernel function of robust relevance vector machine. Mean percentage error has been used to measure the performance of the proposed method in this study. As the mean prediction error of mine gas gushing of the WRRVM model is less than 1.5%, and the mean prediction error of mine gas gushing of the RVM model is more than 2.5%, it can be seen that the prediction accuracy for mine gas gushing of the WRRVM model is better than that of the RVM model.
1. Introduction
It is wellknown that mine gas gushing forecasting is very significant to ensure the safety of mining [1–3]. The study results elaborate that the influencing factors of mine gas gushing mainly include buried depth of coal seam, thickness of coal seam, gas content of coal seam, distance of coal seam, daily advance, and day output [4–6]. Artificial neural networks [7–9] have been applied to mine gas gushing forecasting. However, the forecasting results of artificial neural networks are poor due to their shortcomings including over-fitting and local maximum.
A wavelet-based robust relevance vector machine based on sensor data scheduling control for modeling mine gas gushing forecasting is presented in this paper. Relevance vector machine is a kind of improved support vector machine [10–14], which is based on a Bayesian framework. Morlet wavelet function can be used as the kernel function of robust relevance vector machine.
In the study, the experimental data are employed to study the prediction ability for mine gas gushing. The five factors including buried depth of coal seam, thickness of coal seam, lithology of roof, thickness of bed rock, and ratio of sand and rock are used as the input features of the WRRVM model, and mine gas gushing is used as the output of the WRRVM model. Before the training samples are created, the experimental data must be normalized. Then, the training samples are created. Mean percentage error has been used to measure the performance of the proposed method in this study. As the mean prediction error of mine gas gushing of the WRRVM model is less than 1.5%, and the mean prediction error of mine gas gushing of the RVM model is more than 2.5%, it can be seen that the prediction accuracy for mine gas gushing of the WRRVM model is better than that of the RVM model. Therefore, the WRRVM model is very suitable for mine gas gushing prediction.
2. Wavelet-Based Robust Relevance Vector Machine
The regression function of relevance vector machine can be described as follows:
(1)f(x)=∑i=1nwiϕ(xi),
where wi is the weights and ϕ(·) is the mapping function.
Given the vector R=[r1,r2,…,rn]T, the likelihood function of the dataset can be obtained as follows:
(2)p(t∣w,r,σ2)=(12πσ2)n|P|×exp{-12σ2(t-ηw)′P(t-ηw)},p(w∣t,α,r,σ2)=p(w∣α)p(t∣w,r,σ2)p(t∣α,r,σ2)=N(w∣λ,Δ),
where P=diag(r1,r2,…,rn), Δ is the variance matrix, and λ is the mean value vector, which can be described as follows:
(3)Δ=(A+σ-2∑i=1nriϕ(xi)ϕ(xi)′)-1,λ=σ-2Δ(∑i=1nriϕ(xi)ti).
Then, the marginal likelihood function of robust relevance vector machine can be obtained as follows:
(4)p(t∣α,r,σ2)=∫p(t∣w,r,σ2)p(w∣α)dw=(1/2π)n|σ2/B+ηA-1η′|exp{-12t′1|σ2/B+ηA-1η′|t}.
In this study, Morlet wavelet function can be used as the kernel function of robust relevance vector machine, which can be described as follows [15, 16]:
(5)k(x,x′)=∏i=1mcos(1.75×x-x′ai)exp(-∥x-x′∥22ai2).
It is sufficient to prove the inequality:
(6)F[x](w)=(2π)-m/2∫exp(-j(w·x))k(x)dx≥0,
where
(7)k(x)=∏i=1mcos(1.75xai)exp(-∥x∥22ai2).
3. Experimental Analysis for Mine Gas Gushing Forecasting Based on Wavelet-Based Robust Relevance Vector Machine
In this section, the experimental data are employed to study the prediction ability for mine gas gushing of the proposed wavelet-based robust relevance vector machine. The five factors including buried depth of coal seam, thickness of coal seam, lithology of roof, thickness of bed rock, and ratio of sand and rock are used as the input features of the WRRVM model, and mine gas gushing is used as the output of the WRRVM model.
Figure 1 gives the comparison of the prediction values of mine gas gushing between the WRRVM model and the RVM model trained by the five input features including buried depth of coal seam, thickness of coal seam, lithology of roof, thickness of bed rock, and ratio of sand and rock; Figure 2 gives the comparison of the prediction values of mine gas gushing between the WRRVM model and the RVM model trained by the four input features including buried depth of coal seam, thickness of coal seam, lithology of roof, and thickness of bed rock; Figure 3 gives the comparison of the prediction values of mine gas gushing between the WRRVM model and the RVM model trained by the four input features including buried depth of coal seam, thickness of coal seam, lithology of roof, and ratio of sand and rock; Figure 4 gives the comparison of the prediction values of mine gas gushing between the WRRVM model and the RVM model trained by the three input features including buried depth of coal seam, thickness of coal seam, and lithology of roof; Figure 5 gives the comparison of the prediction values of mine gas gushing between the WRRVM model and the RVM model trained by the three input features including buried depth of coal seam, thickness of coal seam, and ratio of sand and rock. Mean percentage error has been used to measure the performance of the proposed method in this paper. As the mean percentage prediction error of mine gas gushing of the WRRVM model is less than 1.5%, and the mean percentage prediction error of mine gas gushing of the RVM model is more than 2.5%, it can be seen that the prediction accuracy for mine gas gushing of the WRRVM model is better than that of the RVM model.
The comparison of the prediction values of mine gas gushing between the WRRVM model and the RVM model trained by the five input features.
The comparison of the prediction values of mine gas gushing between the WRRVM model and the RVM model trained by the four input features including buried depth of coal seam, thickness of coal seam, lithology of roof, and thickness of bed rock.
The comparison of the prediction values of mine gas gushing between the WRRVM model and the RVM model trained by the four input features including buried depth of coal seam, thickness of coal seam, lithology of roof, and ratio of sand and rock.
The comparison of the prediction values of mine gas gushing between the WRRVM model and the RVM model trained by the three input features including buried depth of coal seam, thickness of coal seam, and lithology of roof.
The comparison of the prediction values of mine gas gushing between the WRRVM model and the RVM model trained by the three input features including buried depth of coal seam, thickness of coal seam, and ratio of sand and rock.
4. Conclusion
A wavelet-based robust relevance vector machine based on sensor data scheduling control for modeling mine gas gushing forecasting is presented in the paper. Morlet wavelet function can be used as the kernel function of robust relevance vector machine. As the mean prediction error of mine gas gushing of the WRRVM model is less than 1.5%, and the mean prediction error of mine gas gushing of the RVM model is more than 2.5%, it can be seen that the prediction accuracy for mine gas gushing of the WRRVM model is better than that of the RVM model. Therefore, the WRRVM model is very suitable for mine gas gushing prediction.
Acknowledgments
The work is supported by the National Natural Science Foundation Project of NSFC (NSFC, 50804061), the Chongqing Education Administration Program Foundation of China (no. KJ120514), Chongqing Education Administration Program Foundation of China (no. KJ110513), and the Natural Science Foundation Project of CQ CSTC (no. 2012BB3725). This work was supported in part by the Foundation for University Youth Key Teacher of Chongqing, Outstanding Achievement Transformation Project of CQ CQJW (no. KJzh10207). And special thanks are due to all who have helped to make this study.
ZhangJ.The application of analytic hierarchy process in mine gas prevention system2011157615842-s2.0-8485609898910.1016/j.proeng.2011.11.2341PackhamR.CinarY.MorebyR.Simulation of an enhanced gas recovery field trial for coal mine gas management2011853-42472562-s2.0-7995155314010.1016/j.coal.2010.11.013YangW.LinB.ZhaiC.LiX.AnS.How in situ stresses and the driving cycle footage affect the gas outburst risk of driving coal mine roadway20123113914810.1016/j.tust.2012.04.015WangK. X.FuX. H.ZhouY. A.HEY.WuH.Dynamic development characteristics of amounts of gas and levels of pressure in the Pan-1 coal mine of Huainan20091967407442-s2.0-7044972980510.1016/S1674-5264(09)60135-6WangW.YanJ.The geological factor analysis of influenced Tianchi coal mine gas occurrence20124531732110.1016/j.proeng.2012.08.164WangL.ChengY. P.WangL.GuoP. K.LiW.Safety line method for the prediction of deep coal-seam gas pressure and its application in coal mines20125035235292-s2.0-8485535683910.1016/j.ssci.2011.09.022LamriniB.Della ValleG.TreleaI. C.PerrotN.TrystramG.A new method for dynamic modelling of bread dough kneading based on artificial neural network20122625125242-s2.0-8485771748910.1016/j.foodcont.2012.01.011LafdaniE. K.NiaA. M.AhmadiA.Daily suspended sediment load prediction using artificial neural networks and support vector machines20134785062KhasheiM.BijariM.A novel hybridization of artificial neural networks and ARIMA models for time series forecasting2011112266426752-s2.0-7875160873810.1016/j.asoc.2010.10.015ChuangC. C.LeeZ. J.Hybrid robust support vector machines for regression with outliers201111164722-s2.0-7795789194710.1016/j.asoc.2009.10.017ParkJ.KwonI. H.KimS. S.BaekJ. G.Spline regression based feature extraction for semiconductor process fault detection using support vector machine2011385571157182-s2.0-7915147753710.1016/j.eswa.2010.10.062EnsafiA. A.HasanpourF.KhayamianT.MokhtariA.TaeiM.Simultaneous chemiluminescence determination of thebaine and noscapine using support vector machine regression20107528678712-s2.0-7374908437610.1016/j.saa.2009.12.021ApolloniB.MalchiodiD.ValerioL.Relevance regression learning with support vector machines2010739285528642-s2.0-7795566009910.1016/j.na.2010.06.035ZBL1205.68279FokouéE.SunD.GoelP.Fully Bayesian analysis of the relevance vector machine with an extended hierarchical prior structure20118183962-s2.0-7864970751010.1016/j.stamet.2010.05.005ZBL1213.62043LiuW.TangB.A hybrid time-frequency method based on improved Morlet wavelet and auto terms window2011386757575812-s2.0-7995158156610.1016/j.eswa.2010.12.107TangB.LiuW.SongT.Wind turbine fault diagnosis based on Morlet wavelet transformation and Wigner-Ville distribution20103512286228662-s2.0-7795431319610.1016/j.renene.2010.05.012