Self.stability evaluation model of surrounding rock based on
improved BP neural network
WANG Duo.dian1,2*, QIU Guo.qing1, DAI Ting.ting3, WANG Yue1
1. Engineering Institute of Corps of Engineers,PLA University of Science and Technology,Nanjing Jiangsu 210007,China;
2. Unit 66081 of PLA, Huailai Hebei 050083, China;
3. China Satellite Maritime Tracking and Controlling Department, Jiangyin Jiangsu 214431, China
Abstract:
Command protection engineering is the important component of national protection engineering system. To raise the level of construction of command protection engineering, the Back Propagation (BP) neural network is improved to give research on self-stability evaluation of its’ surrounding rock. Firstly, the network topology is devised,based on the point of surrounding rock . Secondly, the model is improved according to its disadvantages, by introducing the momentum, self-adaptive adjusting learn rate, variable hidden nodes and steep factor, furthermore, Genetic Algorithm(GA) is imported to seek its best initial weight and threshold value. Finally, be used to a certain command protection engineering,the model is proved to be credible and precise.
Command protection engineering is the important component of national protection engineering system. To raise the level of construction of command protection engineering, the Back Propagation (BP) neural network was improved to give research on self.stability evaluation of its surrounding rock. Firstly, the network topology was devised,based on the characteristics of surrounding rock. Secondly, the model was improved according to its disadvantages, by introducing the momentum, self.adaptive adjusting learn rate, variable hidden nodes and steep factor; furthermore, Genetic Algorithm(GA) was imported to seek its best initial weight and threshold value. Finally, an instance was given to validate the algorithm. The results show that the model is scientifically reliable and of better value in engineering.Key words:
Back Propagation (BP) neural network; Genetic Algorithm (GA); evaluation; surrounding rock; self.stability; command protection engineering
0 引言
1986年,Hecht.Nielsen[1]提出了前馈(Back Propagation, BP)学习算法。由于结构简单,可调参数多,训练算法多,可操控性好,BP神经网络获得了广泛的实际应用。据统计,80%~90%的神经网络模型采用了BP网络或者它的变化形式[2]。
BP神经网络目前被广泛应用到各个领域的研究中。由于神经网络原模型的局限性,国内外众多学者对模型进行了改进。采用遗传算法(Genetic Algorithm,GA)改进BP神经网络是目前的重要改进方法[3-6]。部分学者将BP神经网络方法应用到工程隧道围岩分类和稳定性评估中[7-8],取得了不错的效果。
指挥防护工程是供各级指挥员及指挥机关指挥作战使用的工程建筑[9]。指挥防护工程一经建设就要长期使用,担负防护、战备和作战任务。因此,指挥防护工程地质的建设标准、围岩稳定性的评估指标与一般地下工程不尽相同。本文针对指挥防护工程施工掌子面前方围岩的特点,采用遗传算法优化初始权重和阈值,并引入动量项、陡度因子、可变隐层节点等方法优化BP 神经网络,对指挥防护工程围岩自稳能力进行评估。
1 BP网络评估模型建立
1.1 评估网络层节点设计
要使BP神经网络达到一个较好的求值效果,必须建立优秀的网络拓扑结构,包括对隐含层数、神经元的节点数和初始权重和阈值设计,以及对传递函数、学习函数、训练函数、性能函数等的选择。
1)模型的隐含层数设计。
本文将BP网络作为分类器应用到指挥防护工程围岩稳定性评估中。本文采用单隐层BP网络。
2)输入层节点设计(n)。
围岩稳定性评估采用的指标要素是围岩本身特性决定的。通过对综合分级方法的分析,可以提取重要分级指标(如表1),其中:Rc值为围岩单轴饱和抗压强度;Kv值为围岩完整性指标;RQD值为围岩质量因子;Ko值为围岩弹性抗力强度;f值为围岩结构面摩擦系数;Vp值为纵波波速。常规钻地武器打击后的围岩受损程度和侵彻深度还与岩体的密度ρ有密切关系。因此,输入层节点共有7个,分别代表各指标。
由此可得,围岩稳定性评估BP神经网络的权值调节只与3个因素有关,即:7种围岩物理指标无量纲值大小、学习率η和误差信号δ。
1.3 原始BP网络的缺陷与改进
采用最原始网络进行预报时会出现结果与实际不符的现象,或是网络运行时“死机”等问题。这些问题是由BP神经网络结构和误差学习算法的特点所决定,按照上文推导的权值调节公式无法克服。本文改进方法有以下几点。
1)增加动量项,ΔW(t)=ηδX+αΔW(t-1),α为动量因子α∈(0,1)。动量项反映了以前调整的经验,与网络的记忆功能相辅,起到阻尼作用,减小震荡幅度,提高收敛速度。
2)采取自适应调节学习率。若经过一批次权值调整后总误差E上升,η=βη(β>0),若经过一批次权值调整后总误差E下降,η=θη(θ<0)。可以根据误差的波动改变调整的方向,加速调整过程。
3)引入陡度因子跳出局部最小值,当误差曲面进入平坦区域时ΔE值较小,设o=11+e-net/λ,改变输出量,λ为陡度因子,在平坦区时λ>1,退出平坦区后λ=1。 (责任编辑:南粤论文中心)转贴于南粤论文中心: http://www.nylw.net(南粤论文中心__代写代发论文_毕业论文带写_广州职称论文代发_广州论文网)