(中南大學(xué) 資源與安全工程學(xué)院，長沙 410083)
摘 要: 根據(jù)某露天礦臺階爆破實測數(shù)據(jù)，利用基于回歸分析的經(jīng)驗公式和普通BP神經(jīng)網(wǎng)絡(luò)模型以及基于擬牛頓法的改進BP神經(jīng)網(wǎng)絡(luò)(QN-BP)模型對爆破振動峰值速度進行預(yù)測。兩種模型的訓(xùn)練結(jié)果表明：QN-BP模型經(jīng)過122次迭代即可收斂，訓(xùn)練平均誤差為3.7%；而普通BP模型收斂需要10萬次以上迭代，訓(xùn)練平均誤差4.2%。通過QN-BP模型、BP模型和經(jīng)驗公式的預(yù)測結(jié)果與實測值的對比，三者的平均相對誤差分別為6.05%、10.21%和23.42%。
關(guān)鍵字: 爆破振動；BP神經(jīng)網(wǎng)絡(luò)；擬牛頓法；預(yù)測
（School of Resources and Safety Engineering, Central South University, Changsha 410083, China）
Abstract:According to the measured data of an open pit bench blasting, the experience formula based on regression analysis and ordinary BP neural network model and improved BP neural network model based on Quasi-Newton method (QN-BP) were used to forecast the peak speed of blasting vibration. The training results of two kinds of models show that QN-BP model can be convergence after 122 times iterative, whose average training error is 3.7%. The ordinary BP model need more than 100 000 times iterative to be convergence, whose average training error is 4.2%.By comparing the forecast values with the measured value, the average relative error of the three results(QN-BP, BP and experience formula) are 6.05%,10.21% and 23.42%, respectively.
Key words: blasting vibration; BP neural network; Quasi-Newton method; forecast
