基于GA-Elman神经网络的煤矿突水水源判别
Discriminating Mine Water Inrush Sources Based on GA-Elman Neural Network

煤矿突水是矿井生产过程中产生的自然灾害之一，而准确判别突水来源是突水防治工作的重要基础.以煤矿各含水层水化学成分的差异性为依据，选取K⁺+Na⁺，Ca²⁺，Mg²⁺，Cl^-，HCO₃^-，SO₄^2- 6个常量组分作为突水水源的判别因子.为克服Elman神经网络采用梯度下降法所带来的易陷入局部最小值的缺点，采用具有全局搜索能力的遗传算法(GA)通过选择、交叉和变异等步骤训练优化Elman神经网络，建立了收敛速度更快、泛化性更强的GA-Elman神经网络判别模型，结果表明：将具有全局寻优功能的GA和局部精确寻优的Elman神经网络相结合，克服了Elman神经网络初始权值与阈值的随机性、易陷入局部最优的缺点，能够提高Elman神经网络的判别输出精度，为准确、有效判别突水来源提供了可靠的决策依据；经过GA优化过的Elman神经网络在训练过程中的均方误差收敛速度、收敛精度都有很大的提高，在网络模型的判别输出上，判别结果更为稳定、泛化性更好，为该模型在其他领域的推广提供了一定的借鉴性；为进一步确保突水水源判别的准确性、有效性，在密切结合煤矿水文地质条件的前提下，应选取具有代表性和准确性高的水化资料，有效发挥该判别方法对煤矿水害防治及措施制定的指导作用.
Coal mine water inrush is one of the natural disasters in mine safety production, and accurate discrimination of water inrush sources is an important foundation for water inrush prevention and control. Based on the difference in hydro-chemical composition of various aquifers, six major components (K⁺+Na⁺, Ca²⁺, Mg²⁺, Cl^-, HCO₃^- and SO₄^2-) were selected as the discrimination factors of water inrush sources in a study reported in this paper. In order to overcome the disadvantages that Elman neural network is prone to fall into the local minimum value caused by the gradient descent method, a genetic algorithm (GA) with global search ability was adopted to train the Elman neural network by selecting, crossover and mutation. A GA-Elman neural network discriminant model with faster convergence speed and stronger generalization was established. The application results showed that the global optimization function of GA and the local precision optimization of Elman neural network were combined to overcome the randomness of the initial weights and threshold of Elman neural network and its proneness to fall into the local optima, thus improving the discrimination accuracy of Elman neural network output. It was a reliable basis for decision making for accurate and effective discrimination of the water inrush source. The mean square error convergence rate and accuracy of Elman GA neural network optimization were greatly improved in the training process, resulting in greater stability and better generalization in the discrimination output of the network model. It provided a reference for the model to be extended in other areas. In