神经网络在岩体力学参数和地应力场反演中的应用

Application of neural network to back analysis of mechanical parameters and initial stress field of rock masses

BP神经网络已广泛地应用于岩体力学参数和初始应力场的反演分析,但在实际应用中,BP网络存在着网络训练易于过度、收敛速度慢、易陷入局部极小以及隐层节点数难于确定等缺点。采用RBF网络和改进的BP网络,利用基于有限差分格式的快速拉格朗日算法进行正分析计算,依据若干测点的正应力数据,反演了计算区域的岩体力学参数以及初始应力场。算例表明,RBF神经网络与快速拉格朗日算法相结合,在样本容量相同的情况下,反演分析的精度、网络的拓扑结构以及学习、收敛速度,均优于采用BP网络的反演算法。

At present, BP neural network has been widely used in back analysis of material parameters and initial stress field of rock masses in geomechanics. However, BP neural network is prone to over-being-trained, slow in convergence, not global minimum but local ones obtained and number of neurons in hidden layer hard to be determined. Authors using RBF neural network and BP neural network respectively identified mechanical parameters and initial stresses according to measured normal stresses of some specific points. Direct computations based on fast Lagrangian analysis of continuum （FLAC） were performed to get enough training samples for RBF neural network and BP neural network. An example shows that combination of RBF neural network with FLAC is more effective and rapid than application of BP neural network.