Prediction of Gas Emission Based on Information Fusion and Chaotic Time Series

Mar.2006J. China Univ, of Mining Tech(English Edition)VoL 16 No. 1Prediction of gas emission based on infor-mation fusion and chaotic Time seriesAO Li, YU Hong-zhenCollege of Information and Electrical Engineering, China University of Mining Technolog, Xuzhou, Jiangsu 221008, ChinaAbstract: In order to make more exact predictions of gas emissions, information fusion and chaos time series are com-bined to predict the amount of gas emission in pits. First, a multi-sensor information fusion frame is established. Theframe includes a data level, a character level and a decision level. Functions at every level are interpreted in detail inthis paper. Then, the process of information fusion for gas emission is introduced. On the basis of those data processedat the data and character levels, the chaos time series and neural network are combined to predict the amount of gasemission at the decision level. The weights of the neural network are gained by training not by manual setting, in orderto avoid subjectivity introduced by human intervention. Finally, the experimental results were analyzed in Matlab 6.0and prove that the method is more accurate in the prediction of the amount of gas emission than the traditional methodKey words: gas emission; information fusion; chaos time series; neural networkCLC number: TD 712+51 Introduction2 Information fusionGas emission has a great effect on the producThe principle of information fusion is to utilizetion of a coal mine. It is also a major reason for the fully all the information from sensors. It may seeminstallation or expansion of ventilation in coal mineske the ability of the human brain to makeThe accuracy of gas emission prediction is very im- outside information synthetically Information fusionportant to the security of the production of the mine theory has a great advantage over other methods inand the improvement of economic effectivenessimproving the reliability of a system. Different kindsDifferent mines and working faces have differ- of sensors form the basis of information fusionent rules concerning gas emission and gas prediction multi-information is the processing objective of inmodels should be changed when exterior conditions formation fusion and coordination of optimizationchange. There are many factors that affect the amount and comprehensive processing is the core of informaof gas emission It is therefore difficult to make ac- tion fusion.curate predictions of gas emissionIn our investigation, the quantity of coal minegas emission is estimated by using the principles of 3 The Prediction Modelchaos theory and multi-sensor information fusiontheory. Multi-sensor information fusion is an evolv-Our model of gas emission is established according technology concerned with the problem of how to ing to real conditions in pits. This model can be dicombine information from different sensors in order vided into three levels: data level, character level andto achieve greater accuracy. In recent years, multidecision level. the details of this model are shown assensor information fusion has been extensively invesFig. 1. Following is the interpretation of the functionstigated and widely used in many fields. But few use it at each levelto predict amounts of gas emission In order to make 3.1 Data levelgas prediction more accurate and prevent dangerousFirst, the observation data from all gas emissionconditions, we combine multi-sensor information detection sensors are fused at this level. Then, wefusion with chaos time series to predict the quantityextract characteristics from those fused data andof gas emissionmake a judgment. The objects of sensors are the samesuch as. the de中国煤化工 objects areReceived 06 September 2004; accepted 11 November 2004Project BK2001073 supported by Natural Science Foundation of JiangsuHCNMHGCorrespondingauthorteL:+86-516-82115872;E-mail:janeever@163.comGao Li et alPrediction of Gas Emission Based on Information Fusion and Chaotic Time Seriesnot the same we need to fuse those data at the charAfter classifying every sensor, classificationacter or decision levelsults can be fused at the decision level. Data should3.2 Character levelhave sufficient disposal at the base of the classification of sensor data. On the basis of historic data, preAt this level, ditterent kinds of sensors extract diction of gas emissions can be made at the decisiondifterent characteristics which are then fused into a level of fusion. Historic data and some expert ex-ector Classification recognition is used to deal with perience are stored in the data warehouse. The pre-his fused data. For example, we extract characterdiction results come from the mining data and can betics from gas sensors in the monitoring systemfed back to the decision level3.3 Decision levelData FusionCharacter FusionDecision FusionationExtractionRecognitionOutput of GasharacterClassiticationDecision fusion叫 Sensor2ExtractionRecognition :ClassificationSensor nExtractionFig 1 Multi-sensor information fusion frame4 Prediction of gas emissiontracted from processed data. According to Figure 1we fuse those data which have been processed at theel character level and decision level. The4.1 Data detection based on multi-component of the amount of gas emission is calcuThe method of firedamp content is usedlated by formula 2 and the amount of gas emission istice. We have adopted this method which can be de- calculated by formula 1. After calculation, the inforscribed asmation about the amount of gas emission is stored in∑Q(=1,2,…n(1) the data warehouse as historic data. The prediction ofgas emission is based on this historic datawhere Q represents the amount of gas emission in 4.3 Chaos time series and BP neural networkthe digging pit, o represents the amount of gasThe data and character levels provide the inforemission of coal per tonne and can be represented asmation about amount of gas emission At the decision(2) level, mining data are used to make predictions of gasemissions Chaos time series and neural networks areAccording to formula(2), those weights should integrated to serve as the data mining method. Acbe detected by sensorscording to our experiment, the information of theamount of gas emissions stored in the data warehouse1)Coal Ply of Gas mi; Digging Ply m,show chaotic behavior. Time series derived from such2)Original Gas in coal level:xo;systems seem stochastic in nature, however, these3)Remnant Gas in ground coal:x,;chaos time series can be predicted4)The amount of gas emission of gas i: Cshort time spans. The algorithm of several chaaos indices in time series is proposed to search for chaos in4.2 Fusion processload time series. The advantage of this method is thatSignals detected by sensors are amounts of the predicted error can be controlled by adjustingnon-electricity and have different characteristics. some parameters if the user has access to long timeThese non-electricity amounts need to be changed seriesinto electricity amounts and these in turn should beIn order to make a comparison, we use data fromchanged into digital signals by aA/D converter so Xu. State space reconstruction is used to make athat they can be processed by computers. This procdement of this time series in reference [8]. Its fracess needs to take place before the information can be tal dimension is 33. So the time series in reference [8implemented. At the data level, we dispose of some represents a chicaInoise signals. Detected signals include some noisedicted over a s中国煤化工 d Bpand equilibration and filters are adopted to deal with ral networks arCN MH Ge method ofnoise.Then, some expected characteristics are ex- adjusting weights is introduced as followsJ. China Univ, of Mining Tech (English Edol. 16 No. 11)Weight adjustment between middle level and seen from Fig. 2 and is an improvement over the re-output levelults in reference [8]wherein is the learning coefficient and8 theδ=(z*y)2z(1-z)where z is the output of the model and v theoutput in practice2)Weight adjustment between input level andmiddle levelw=w+n8,x6,=y(-y)05101520253035404550Sample numberi represents the input level.j represents the output levelgas emission amount ediction ofFig. 2 Experimental result of the pre5 Case study6 ConclusionsThe structure of the bp neural network consistsWe started the process with the establishment ofof three nodes in the input layer, ten nodes in the a framework of multi-sensor fusionhidden layer and one node in the output layerThen, the prediction was made by a chaotic timelearning error is 0.0005 and the learning coefficient series and a neural network at the decision level. the0. 2. We use matlab 6.0 to evaluate this designresult of the experiment indicates that the predictionFirst,we trained data from the former 70 sets. of gas emission, based on multi-sensor fusion andhen, in order to eliminate the error made by assignchaotic time series is more accurate than the olding artificial weights, the adjustment of weights is method. The weights are gained by training, not byimproved according to formulas (3)and (5)manud avoids the subjective effectsneural network model is obtained at the end of net- troduced by human intervention. Findinork training. This model is an effective predictionpriate number of hidden levels is still a difficulttool in chaos time seriesThe result of the experiment is shown in Fig. 2problem and requires us to do more research in thehere the symbol indicates real data and the curve futureindicates prediction data. The result can be clearlyReferences[1] You Y, SHWX, WSA. The study and application of the electric power system short-term load forecasting using a new modeProceedings of the CSEE, 2002, 22(9): 15-18.(In Chinese)[2] Lv J H, Zhang S C. Application of adding-weight one- rank local-region method in electric power system short-term loadforecast. Control Theory and Publications, 2002, 19(5): 767-780(In Chinese)[3 Hu SR Neural Network Application. Changsha: National University of Defense Technology Publishing Company, 1993. (InChinese)[4 Zang L, Application of neural network to analysis and treatment of chaotic observation time series. Journal of Data Acquisition &e Processing, 2001, 16(4): 486-489. (In Chinese[5] He Y, Wang G H, LV D J, et al. Mulitsensor Information Fusion and Application. Beijing: Publishing House of ElectronicsIndustry, 2000. (In Chinese)[6] Jian X C, Zheng J L. Analysis of algorithm and error evaluation criteria in chaotic neural network predictions. Joural ofinghua University ( Sci& Tech), 2001, 41(7): 43-46.(In Chinese)[7 Gao L, Yu H Z, Wang Y F. Coal blending scheduling in coal preparation plant based on multi-sensor information fusionJournal of China University of mining Technology, 2004, 33(1): 99-102(In Chinese)[8 Xu ZH F. Research on Intergrated System of Gas Real-Time Detecting Information Based on Intranet. Xuzhou: China University of Ming and Technology, 2001. (In Chinese中国煤化工CNMHG