摘要
基于BP(Back Propagation)神经网络法,收集对比130个公开的地下封闭爆炸数据,区分黏土、砂性土和岩石三组围岩介质,简化地下任意点爆炸围岩介质峰值力的多因素影响,分析爆炸比例距离、纵波波速、密度和饱和度等特征参数产生的权重影响,考察爆冲的输出峰值压力特征,提出土中任一点围岩介质峰值压力的简易预测方法,并通过与经验方法对比,验证简易方法的实效性。利用Matlab建立不同隐含层单元数的BP(Back-Propagation)神经网络,当隐含层分别选择6、7、6个神经元时整体网络性能最佳,该条件下对比测试样本的BP神经网络、经验公式和多元回归分析方法(MVRA)的预测效果,BP神经网络方法得到最小平均绝对误差。在各围岩介质数据的误差对比分析中,BP神经网络法得到砂性土的预测误差相对最小,相比经验公式和多元回归分析优势明显。在同一围岩介质参数的敏感度分析中,纵波波速对峰值压力产生最显著影响。将工程实例参数带入对比BP神经网络和MVRA,考虑不同介质反射系数得到峰值压力预测值和拱顶爆炸荷载峰值实测值的相对误差可小于20%。该估算方法可为类似地下结构防护设计值提供一种简化参考。
Here,based on the BP neural network method,130 open data sets of underground closed blasts were collected and compared to distinguish three groups of surrounding rock medium including clay,sandy soil and rock,and influences of multi-factor on surrounding rock medium peak pressure of arbitrary point explosion in soil were simplified.The weight influences produced by characteristic parameters of blast scaled distance,longitudinal wave velocity,density and saturation level were analyzed to investigate blast impact’s output peak pressure characteristics and propose the simple prediction method for surrounding rock medium blast peak pressure of arbitrary point in soil.Through comparing with the empirical method,the effectiveness of the simple method was verified.BP neural networks with different hidden layer elements were built using the software MATLAB.It was shown that when the hidden layer has 6,7,6 neurons,respectively,the overall network performance is the best;under this condition,comparing BP neural network,empirical formula and multi-variable regression analysis(MVRA)prediction results for measured samples,the BP neural network method obtains the minimum average absolute error;in the error contrastive analysis for surrounding rock medium data,the BP neural network method’s prediction error for sandy soil medium is relatively minimum to have obvious advantages compared with empirical formula and MVRA;in the sensitivity analysis for the same surrounding rock medium parameter,longitudinal wave velocity affects peak pressure most obviously;the actual engineering example’s parameters are brought into the BP neural network method and MVRA for comparison,the relative error between the peak pressure predicted values considering different mediums’reflect coefficients and the actual measured blast load peak value at the dome can be less than 20%.The proposed estimation method provided a simplified reference for the protection design of similar underground structures.
作者
郭璇
马思远
郭一帆
张晓新
GUO Xuan;MA Siyuan;GUO Yifan;ZHANG Xiaoxin(Key Laboratory of Urban Underground Engineering of Ministry of Education,Beijing Jiaotong University,Beijing 100044,China;School of Civil Engineering,Beijing Jiaotong University,Beijing 100044,China;Huachengboyuan Engineering Technology Group Co,Ltd,Beijing 100052,China)
出处
《振动与冲击》
EI
CSCD
北大核心
2019年第3期199-206,共8页
Journal of Vibration and Shock
基金
国家自然科学基金(51678038)
霍英东教育基金会(122009)
北京交通大学基本科研业务费(2017JBM083)
中国国家留学基金委(201707095041)
关键词
BP神经网络
围岩介质
峰值压力
预测值
爆炸荷载
BP(back propagation)neural network
surrounding rock medium
peak pressure
prediction value
blast loads
