翻譯部分英文原文Establishment of grey-neural network forecasting model of coal and gas outburstYang Sheng-qianga, Sun Yana,b, Chen Zu-yuna, Yu Bao-haia, Xu QuanaaState Key Laboratory of Mine Resource and Safe Exploitation, School of Safety Engineering, CUMT, Xuzhou 221008, ChinabDepartment of Public Management, Shanghai Trade Union Polytechnic, Shanghai 201415, ChinaAbstract: Grey correlation analysis was made with respect to factors affecting coal and gas outburst and the input parameters of artificial neural network (ANN) determined. Then five dominant factors were chosen for grey correlation analysis as the input parameters based on the improved BP algorithm, and neural network forecasting model of coal and gas outburst established. The network was trained by using the study samples from the instances of typical coal and gas outburst mines, and coal and gas outburst instances of Yunnan Enhong mine were used as forecasting samples. The comparison between the results from network forecasting with that of the traditional methods indicates that this method can meet the requirement for coal and gas outburst forecast .Keywords: coal and gas outburst ; grey correlation analysis; grey-neural network1. IntroductionIn china, coal has a wide distribution and the landforms of coal fields are complex. The coal production is threatened by water, fire, coal dust, roof fall, gas outburst, and so on. Of these factors, gas outburst is the most serious dangerous one to cause great economic loss and kill coal miners. So, gas outburst forecasting becomes particularly important 1.Because the inherent mechanism of coal and gas outburst is so complicated and lots of uncertain and fuzzy problems exist between effect factors and accidents, both the traditional forecasting technologies based on experience and the statistical forecasting technologies based on mathematical model are restricted in the field application. Grey-neural network forecasting methods of coal and gas outburst is applied in this paper.2. Analysis of effect factors 2.1. Initial velocity of gas ( p )The initial velocity of gas is one of the risk indexes for coal and gas outburst2-3. It shows the blow-off velocity of gas from coal. This index reflects how quickly the gas releases from coal seams. p is related to the gas content of coal, structure and surface property of pore. To a large degree, the movement and destructive power of gas is decided by desorption and blow-off ability of gas in coal during the developing process of coal and gas outburst.2.2. Consistent coefficient of coal ( f )The consistent coefficient of coal is a kind of relative indexes of coal particles mechanical strength. Its value reflects coals physical and mechanical properties and is also an important parameter involved in coal and gas outburst. Generally, the bigger the f is, the more difficult the outburst happens under the same gas pressure and ground stress.2.3. Gas pressureGround stress controls gas pressure field and promotes coal-body to be destructed by gas. The increased pressure in surrounding rock determines ventilation property of coal seams and leads to increase pressure gradient which is favorable for the coal and gas outburst to happen. The content of gas pressure is an important symbol of gas compressive energys value.2.4. Thickness of soft sublayer The deeper the coal seam is, the more frequent the gas outburst happens. Both the outburst times and the scales increase with the increase of coal seam thickness, especially the thickness of the soft sublayer. Because the reason of low mechanical strength of coal and bad ventilation property, much content and pressure of gas exists in the change area of coal bed thickness. 2.5. Coal-body destruction typeGround stresses, including self-weight stress, structural stress, and disturbance stress, get the surrounding rocks or coal-bodys elastic potential energy do work, making the coal-body destroyed and displaced. Coal-body destruction type refers to the coal-body destruction degree of coal-body structure under structural stress. According to the destruction degree, it can be divided into five types:1- non-destructive coal; 2-destructive coal; 3- strong destructive coal;4 - pulverized coal; 5-completely pulverized coal. 2.6. Mining depthViewing from the regional metamorphism of coal, the depth is the main reason lignite changing into anthracite because, with increase of depth, the pressure and temperature increases. The deeper the depth is, the higher the coal rank is. The huge thickness cover makes the gas be formed and protected, most of which are methane and so on. So the outburst intensity of coal will increase with the increase of mining depth.2.7 The Gas Content of Coal seamGas is from the coal seam, strata, gob or production process during mine excavation. The higher the gas content of coal seam is, the more gas will effuse into tunnels and working faces during coal seam excavation. As a result, the threat of the gas accident will be more serious.Based on the above factors, this paper collected some typical data from eight outburst mines in China as the sample sets of grey relation analysis model. Values of various factors are shown in table 1.Table 1. Original data of each influencing factorSampleNumberoutburstscales(t)Initial Velocity(p)Consistencecoefficient(f)Gas Pressure(MPa)Soft stratificationthickness(m)Coal-bodyDestruction TypeMining Depth (km)Gas Content of Coal Seam (m3/t)1 150.00 19.00 0.31 2.76 1.20 3 0.620 10.022 20.60 6.00 0.24 0.95 2.00 5 0.445 13.043 15.10 18.00 0.16 1.20 1.30 3 0.462 10.364 0.00 5.00 0.61 1.17 1.61 1 0.395 9.045 76.50 8.00 0.36 1.25 1.41 3 0.745 9.016 10.20 8.00 0.59 2.80 1.82 3 0.425 10.257 0.00 7.00 0.48 2.00 1.10 1 0.460 9.508 110.20 14.00 0.22 3.95 0.93 3 0.543 8.233. Grey relation analysis In fact, a series of effect factors of coal and gas outburst are non-time series, so its reasonable to generate the interval values from original data. This paper makes the factor that is outburst scales of gas as the controller series and other various factors as sub sequence of grey relation analysis.Getting the interval values from original data according to the formula:,then calculating the absolute differences i(k) of one samples (1)min()()axiiiiyxinterval values between controller series and sub sequence by the formula:. Correlation coefficient series i(k) can be gotten by the Calculation 0()()i ikkFormula and then the 0 0n()max()i iik iki i iixkxkcorrelation degree between controller series and sub sequence can be gotten though the formula: . The Resolution Coefficient is 0.5. The results are shown in table2.1()niikrTable 2. Relational degree taxis of influencing factorsInfluencing factors Correlation degree OrderingInitial Velocity p 0.79 1Consistence coefficient f 0.58 5Gas Pressure p 0.68 4Soft stratification thickness 0.54 7Coal-body Destruction Type 0.70 3Mining Depth 0.78 2Gas Content of Coal Seam 0.55 6The relating sequence of various factors for gas outburst scales shows the affecting degree on the gas outburst scales. According to the calculation results of correlation degree above, correlation degree that various factors for gas outburst scales orders are as follows.Initial velocity mining depth coal-body destruction type gas pressure consistence coefficient gas content of coal seamsoft stratification thickness.4. Model buildingComplicated nonlinear relation exists between various factors and gas outburst scales. Using artificial neural network to forecast the happening coal and gas outburst can reduce the human disturbance, make the result more objective, and show the connection of input and output variables truly4-5.4.1. Determination of input elements The accuracy of neural Network isnt proportional to the number of chosen factors. If factors are overabundant, the learning speed of network will reduce and the learning process will become complicated and difficult to control. At the same time, the number of effect factor cant be too less, otherwise itll make results depend on some parts of factors too much. We chose five dominant factors in Gray correlation analysis: Initial Velocity, Mining Depth, Coal-body Destruction Type, Gas Pressure, and Consistent coefficient as input neurons. In addition, we introduce the improved BP algorithm of Artificial Neural Network and adopt neural network function storehouse of MATLAB to write the procedure and established neural network predicting model of analysis of coal and gas outburst.4.2. Classification of the forecast results According to the actual conditions of coal mine, the coal seam can be divided into two types: outburst coal seam and non-outburst coal seam. In order to improve the actual applicability, this paper divided the outburst situation into three types: small amount outburst (below 50t, called small) ;common outburst(between 50t and 100t, called medium );mass outburst(above 100t, called large). The output value of Neural Network training cant be 0 or 1, so the output values are set into four kinds as follows: 1,0,0,0 stands for“ non“ ;0,1,0,0 stands for “small“ ;0,0,1,0 stands for “medium“ ;0,0,0,1 stands for “ large“.4.3. Determination of network structureAccording to the analysis above, we chose five main effect factors of coal and gas outburst and they can also serve as input nodes of the model. There are four output nodes and the output value is 0 or 1. The number of the hidden layer nodes is very important. If the number of the nodes is too less, the network cant establish complicated judgment boundary; If the number of the nodes is overabundant, the network will lose the summarizing and judging ability 6. The number of 11, 9 and 16 are tried in this paper.Finally the best number of hidden layer is chosen by comparing the network performance. The transfer function between different layers is chosen by sigmoid function and the whole interconnection is used between different layers. Network structure of Neural Network is showed in figure 1.Fig1. BP network structure chart of coal and gas outburst forecast4.4. Collection and normalization of dataTypical coal and gas outburst instances must be chosen as sample and their data must be standardized before network training. This paper chose five effect factors from typical coal and gas outburst mines in China as learning samples shown in table 3.Table 3. Original data of coal and gas outburst instancesSampleNumberInitial Velocity(p)Consistencecoefficient(f)Gas Pressure(MPa)Coal-bodyDestruction TypeMining Depth (km)outburstscales1 19.00 0.31 2.76 3 0.620 Big2 6.00 0.24 0.95 5 0.445 Small3 18.00 0.16 1.20 3 0.462 Small4 5.00 0.61 1.17 1 0.395 Non5 8.00 0.36 1.25 3 0.745 Medium6 8.00 0.59 2.80 3 0.425 Small7 7.00 0.48 2.00 1 0.460 Non8 14.00 0.22 3.95 3 0.543 Big9 11.00 0.28 2.39 3 0.515 Small10 4.80 0.60 1.05 2 0.477 Non11 6.00 0.24 0.95 3 0.455 Medium12 14.00 0.34 2.16 4 0.510 Small13 4.00 0.58 1.40 3 0.428 Non14 6.00 0.42 1.40 3 0.426 Big15 4.00 0.51 2.90 5 0.442 Big16 14.00 0.24 3.95 3 5.52 Small17 4.00 0.53 1.65 2 4.38 Non18 6.00 0.54 3.95 5 5.43 Big19 7.40 0.37 0.75 4 7.40 Medium20 3.00 0.51 1.40 3 4.00 NonInput nodes parameter values of the BP network are different and the values diverge greatly, so the values need to be normalized in order to prevent the information of small values from being weakened by big ones. Generally, various values are normalized between 0 and 1. But it isnt an appropriate method for this case. Because the function sigmoid curve changes is smooth between 0 and 0.1 or between 0.1 and 0.9. So the good normalized value range should be 0.10,0.90. The formula of can satisfy thenormalized requirements. min0.81axXQuantificational data can be normalized using the above method.4.5. Training of BP Neural Network for coal and gas outburst prediction In this paper, the BP tool functions in the ANN toolbox of MATLAB software is applied, and applications of some important tool functions are demonstrated7-10. The inputting layer has five nerve fibres because the inputting samples are 5-dimensional inputting vector. After many times pilot calculation, the high performance network will be gotten if the number of hidden layer is 11. The outputting layer has four neurons because of four outputting data. So the networks structure is 5-11-4. The transfer function between different layers is the S-shaped tangent function. Fig. 2. Training error curve of networkThe training function is traingdx and the learning rate is adapted by network itself, called improved BP algorithm. The training error curve of network is shown in figure 2. From the figure we can get that the network converges after 14,166 times iterative calculations and the network also can identify the study learning sample completely and accurately. Complicated nonlinear relation is set up between various factors and gas outburst scales.4.6. Coal and gas outburst forecasting by trained networkEight coal and gas outburst instances of Yunnan Enhong coal mine were used as forecasting samples. The detail data are showed in table 4.When the network outputting value is close to 1, 0, 0, 0, the scale of coal and gas outburst is “small“, when the network outputting value is close to 0, 1, 0, 0, the scale is “Medium“, when the network outputting value is close to 0, 0, 0, 1, the scale is “big“.Network output is as follows:The expected outputting values of BP network should be all 0,1,0,0, from forecast results, they are all consistent with actual coal mine situation. So the method has certain practicability. The Error curve is shown in figure 3.The forecasting results calculated by grey - neural network and other methods are shown in table 5, in which “You” means having the risk of coal and gas outburst while “Wu” means having no risk of outburst. Fig. 3. Forecast error curve of network Table 4. Instances of coal and gas outburst of Yunan Enhong mineSamplenumberInitial velocityemissionRobustness coefficientGas pressure(MPa)Destroy the type of coalMining depth/×102mstrength ofoutburst (t)1 11 0.37 2.1 3 4.12 112 12.1 0.49 2.0 3 4.12 93 11.5 0.28 1.9 3 4.07 104 11.8 0.36 2.3 3 4.03 155 10.8 0.30 2.2 3 3.96 96 12.4 0.38 1.8 3 4.10 9.37 11.8 0.57 1.6 3 4.08 36.88 10.0 0.55 1.5 3 4.05 10.8Table 5. Comparison of forecast results between grey-neural network and other forecast methodsSamplenumberInitial velocityemissionRobustness coefficientGas pressure(MPa)Destroy the type of coalMining depth/×102mK S C BP1 11 0.37 2.1 3 4.79 20 You You Small2 12.1 0.49 2.0 3 4.12 18.0 You You Small3 11.5 0.28 1.9 3 1.4 16.8 You You Small4 11.8 0.36 2.3 3 2.4 14.2 You Wu Small5 10.8 0.30 2.2 3 2.8 19.4 You You Small6 12.4 0.38 1.8 3 4.7 17.5 You You Small7 11.8 0.57 1.6 3 3.9 16.1 Wu You Small8 10.0 0.55 1.5 3 1.8 20.5 Wu You SmallSingle-target method and comprehensive target method are respectively instead by S and C.From table 5, part of the prediction results by single-target method and comprehensive target method respectively are discrepant with the reality and they can't reflect the outburst risk degree. The reason is that those methods could not post the complicated relationship between influencing factors of coal and gas outburst. But the prediction results by BP nerve network are more accurate than others. So the method presented in the paper has certain practicability.5. Conclusions1) Seven main effect factors of coal and gas outburst are analyzed, and actual data in eight typical coal and gas outburst mines in China are collected. Through the Grey correlation analysis, we get the correlation order of effect factors affecting coal and gas outburst.2) Neural network forecasting model of coal and gas outburst is built. According to the result of grey correlation analysis, Input Elements in Grey-neural Network Forecasting Model are determined. 3) 20 Chinese typical examples of coal and gas outburst are collected to train the network model. The trained gray- neural network model has been applied in Yunnan Enhong coal mine and then checked usefulness and accuracy, showing that the grey - neural network model is suitable for predicting coal and gas outburst. Acknowledgements Project supported by the Natural Science Hall of Yunnan Province (No 2005IT02) and by the Natural Science Key-Foundation (No 50834005). References1 G. Wen, S. Xu, A new progress of preventing and curing coal mine gas outburst. Mining Safety & Environmental Protection, 27 (2002). 2 S. Ma, E. Wang, Coal and Gas forecast methods. Mining & Processing Equipment, 5 (2001) 71-72. 3 C. Jiang, Forecast model and indexes of coal and gas outburst. Journal of China University of Mining & Technology, 27 (1998) 373-376. 4 J. Deng, Grey Systems Theory Course. Wuchang: Huazhong University of Science and Technology, 1989. 5 J. Deng, Grey Control System. Wuchang: Huazhong University of Science and Technology, 1997. 6 B. Xu, Z. Liu, MATLAB Engineering Mathematics and Applications. Beijing:Tsinghua University Press, 2000. 7 H. Yu, Intelligent Diagnosis on Neural Network. Beijing: Metallurgical Industry Press, 2000. 8 Y. Pan, C. Zhai, An evaluation of the road traffic safety based on the grey cluster and neural network. Journal of Chongqing Jiaotong University, 4 (2005) 101-104.中文譯文煤與瓦斯突出灰色-神經(jīng)網(wǎng)絡(luò)預(yù)測模型的建立楊勝強 1， 孫巖 1 2，陳祖云 1， 于寶海 1, 徐泉 11.礦產(chǎn)資源與安全開采國家重點實驗室，安全工程學(xué)院，中國礦業(yè)大學(xué)，徐州 221008，中國2.公共管理系，上海工會理工學(xué)院，上海 201415，中國摘要：對煤與瓦斯突出影響因素進行灰關(guān)聯(lián)分析，以此確定人工神經(jīng)網(wǎng)絡(luò)的輸入?yún)?shù)。并應(yīng)用改進的 BP 算法，選擇灰關(guān)聯(lián)分析的 5 個優(yōu)勢因子作為輸入?yún)?shù)，建立了煤與瓦斯突出預(yù)測的神經(jīng)網(wǎng)絡(luò)模型。選用典型突出礦井的煤與瓦斯突出實例作為學(xué)習樣本，對網(wǎng)絡(luò)進行訓(xùn)練學(xué)習，并以云南恩洪煤礦的煤與瓦斯突出實例作為預(yù)測樣本，將經(jīng)過網(wǎng)絡(luò)預(yù)測的結(jié)果與傳統(tǒng)方法的計算結(jié)果進行對比。結(jié)果表明該灰色-神經(jīng)網(wǎng)絡(luò)模型能夠滿足煤與瓦斯突出預(yù)測的要求。關(guān)鍵詞：煤與瓦斯突出；灰關(guān)聯(lián)分析；灰色-神經(jīng)網(wǎng)絡(luò)1.簡介我國煤炭分布范圍廣泛、埋藏地形復(fù)雜，煤炭生產(chǎn)一直受到各種災(zāi)害，如瓦斯、涌水、火災(zāi)、煤塵及冒頂?shù)鹊耐{，其中尤以瓦斯事故后果最為嚴重。由于瓦斯事故帶來的人員傷亡和經(jīng)濟損失在煤礦事故中占據(jù)十分重要的位置，因此，瓦斯事故的預(yù)測方面的研究就顯得十分重要。煤與瓦斯突出的內(nèi)在機理極為復(fù)雜，突出影響因素與突出事故之間相關(guān)規(guī)律存在一定的不確定性和模糊性，基于經(jīng)驗的傳統(tǒng)預(yù)測技術(shù)和基于數(shù)學(xué)建模的統(tǒng)計預(yù)測方法的應(yīng)用已受到了很多的限制。文章采用基于灰色關(guān)聯(lián)分析的神經(jīng)網(wǎng)絡(luò)方法來對煤與瓦斯突出進行預(yù)測。2.煤與瓦斯突出影響因素分析2.1 瓦斯放散初速度（p）煤的瓦斯放散初速度是預(yù)測煤與瓦斯突出危險性的指標之一2-3，該指標反映了含瓦斯煤體放散瓦斯快慢的程度。p 的大小與煤的瓦斯含量、孔隙結(jié)構(gòu)和孔隙表面性質(zhì)與大小有關(guān)。在煤與瓦斯突出的發(fā)展過程中，瓦斯的運動和破壞力，在很大程度上取決于含瓦斯煤體在破壞時瓦斯的解吸與放散能力。2.2 煤的堅固性系數(shù)（f）煤的堅固性系數(shù)是煤顆粒本身力學(xué)強度的一種相對指標，其數(shù)值大小也是煤層物理力學(xué)性質(zhì)的重要反映，是煤與瓦斯突出現(xiàn)象所涉及到的重要參數(shù)之一。通常情況下，在相同的瓦斯壓力和地應(yīng)力條件下，煤的堅固性系數(shù)越大，越不容易發(fā)生突出。2.3 瓦斯壓力地應(yīng)力控制瓦斯壓力場，促進瓦斯破壞煤體，圍巖中應(yīng)力的增加，決定了煤層的透氣性，造成瓦斯壓力梯度增高，對突出有利。瓦斯壓力的大小是煤體含瓦斯壓縮能高低的重要標志。2.4 軟分層煤體厚度煤層越厚特別是軟分層越厚，瓦斯突出越頻繁，突出次數(shù)和突出強度隨著煤層厚度，特別是軟分層的厚度的增加而增加。因為在煤層厚度變化區(qū)域煤的力學(xué)強度低，透氣性差，瓦斯含量和瓦斯壓力較大。2.5 煤體破壞類型地應(yīng)力（包括自重應(yīng)力、構(gòu)造應(yīng)力和采動應(yīng)力）使圍巖或煤體的彈性潛能做功，使煤體破壞和位移。煤的破壞類型是指煤體結(jié)構(gòu)受構(gòu)造應(yīng)力作用后的煤體破壞程度，根據(jù)其破壞程度，一般分為 5 類：1-非破壞煤；2-破壞煤；3-強烈破壞煤；4-粉碎煤；5 -全粉煤。2.6 開采深度從煤質(zhì)區(qū)域變質(zhì)的角度看，在由褐煤向煙煤無煙煤演變過程中主要是由于蓋層的增厚，使其地層溫度升高，壓力增大。因此蓋層愈厚，煤的變質(zhì)程度亦愈高，巨厚的蓋層使其以甲烷為主的變質(zhì)氣體產(chǎn)物得以大量產(chǎn)生并予以保護。因此，隨著開采深度的增加，煤層突出強度就會增強。2.7 煤層瓦斯含量瓦斯是在礦井采掘過程中，從煤層、巖層、采空區(qū)放出的和生產(chǎn)過程中產(chǎn)生的。煤層的瓦斯含量越高，開采煤層時涌入井巷和工作面的瓦斯就越多，瓦斯災(zāi)害的威脅也越大。根據(jù)以上所考慮的各種因素，論文收集了在中國的 8 個比較具有代表性的突出礦井的實測數(shù)據(jù)作為灰色關(guān)聯(lián)分析的模型樣本集。各因素的取值見表 1。表 1 各影響因素的原始數(shù)據(jù)樣本序號突出強度/t放散初速度 p堅固性系數(shù) f瓦斯壓力/Mpa軟分層厚度/m煤體破壞類型開采深度 /km煤層瓦斯含量/(m3/t)1 150.00 19.00 0.31 2.76 1.20 3 0.620 10.022 20.60 6.00 0.24 0.95 2.00 5 0.445 13.043 15.10 18.00 0.16 1.20 1.30 3 0.462 10.364 0.00 5.00 0.61 1.17 1.61 1 0.395 9.045 76.50 8.00 0.36 1.25 1.41 3 0.745 9.016 10.20 8.00 0.59 2.80 1.82 3 0.425 10.257 0.00 7.00 0.48 2.00 1.10 1 0.460 9.508 110.20 14.00 0.22 3.95 0.93 3 0.543 8.233.煤與瓦斯突出影響因素的灰色關(guān)聯(lián)分析事實上，煤與瓦斯突出的影響因素實質(zhì)為一個非時間序列，因此采用原始數(shù)據(jù)區(qū)間化比較合理。文章把煤與瓦斯突出的強度作為灰色關(guān)聯(lián)分析的母序列，而把其他因素作為灰色關(guān)聯(lián)分析的子序列。依據(jù)下述公式從原始數(shù)據(jù)中獲得間隔值： ，然后利用公(1)min()()axiiiiyx式 計算樣本中母序列和子序列之間的絕對值。相關(guān)系數(shù) i(k)可通過0()()i ikxk計算公式得到： ，母序列與子0 0mni()a()xi iik iki i iixkxk序列之間的關(guān)聯(lián)度可通過公式： 獲得。其中分辨系數(shù) =0.5。結(jié)果見表 2。1()niikr表 2 各影響因素的關(guān)聯(lián)度影響因素 關(guān)聯(lián)度 順序放散初速度 p 0.79 1堅固性系數(shù) f 0.58 5瓦斯壓力 p 0.68 4軟分層厚度 0.54 7煤體破壞類型 0.70 3開采深度 0.78 2煤層瓦斯含量 0.55 6關(guān)聯(lián)度計算分析可得各子序列與母序列之間的關(guān)聯(lián)度,即各影響因素對瓦斯突出強度的關(guān)聯(lián)度排序為：放散初速度開采深度 煤體破壞類型瓦斯壓力 堅固性系數(shù)瓦斯含量軟分層厚度。4.煤與瓦斯突出預(yù)測灰色- 神經(jīng)網(wǎng)絡(luò)模型的建立煤與瓦斯突出與其影響因素之間存在著復(fù)雜的非線性關(guān)系，采用人工神經(jīng)網(wǎng)絡(luò)進行煤與瓦斯突出預(yù)測，能減少人為的干擾，從而更具有客觀性，并且具有極強的非線性逼近能力，能真實刻畫出輸入變量與輸出變量之間的關(guān)系4-5。4.1 灰色-神經(jīng)網(wǎng)絡(luò)模型輸入元的確定神經(jīng)網(wǎng)絡(luò)的準確性與影響因素選擇的數(shù)量并不成正比。如果影響因素的選擇過分充裕，會導(dǎo)致神經(jīng)網(wǎng)絡(luò)的工作速度降低，工作過程會變得十分復(fù)雜并難以控制。與此同時，影響因素的數(shù)量也不能過少，否則會導(dǎo)致結(jié)果過多取決于部分影響因素。根據(jù)前面灰色關(guān)聯(lián)分析可知，煤與瓦斯突出危險性影響因素最主要有以下 5 個優(yōu)勢因子：瓦斯放散初速