矿井突水水源判别的ESN正则化模型

针对标准回声状态神经网络（ESN）因病态解而导致水源判别模型准确率低，精度差的问题，提出了将6种正则化方法与ESN神经网络相结合，并应用于矿井突水水源的判别，与标准ESN模型的判别结果进行对比分析。结果表明:ESN模型易出现过拟合问题，判别准确率只有49%~88%；而采用阻尼最小二乘奇异分解法（DSVD）与广义交叉验证法（GCV）相耦合的正则化方法能够较好的解决模型病态解问题，使模型的准确率提高到100%，最佳判别精度比标准ESN模型提高了64%，稳定性提高了61%；且该方法对不同规模的储备池结构表现出较强的适应性，不仅简化了模型的映射关系，提高计算效率，还增强模型的泛化能力。因此，基于GSVD_GCV正则化的ESN水源判别模型可作为一种快速有效判别矿井突水来源的新方法。 

ESN regularization model for discriminating mine water inrush source

Aiming at the problem that the standard echo state neural network(ESN) is over-fitting due to the abnormal solution, six kinds of regularization methods are combined with ESN neural network and applied to discriminate mine water inrush source. The models were evaluated and compared with the standard ESN model. The results show that the ESN water source discrimination model is prone to over-fitting, and the accuracy of discrimination is only 49%~88%. The damping least squares singular decomposition method(DSVD) combined with generalized cross validation method(GCV) called as the regularization method can improve the accuracy of the model, the accuracy of the model is improved to 100%, the best accuracy is about 64% higher than that of the standard ESN model, and the stability is improved by about 61%, and the method is adaptable to the reserve pool, which can simplify the complex mapping of the model, improve the computational efficiency, and enhance the generalization ability of the ESN discrimination model. Therefore, the ESN water source discrimination model based on GSVD-GCV regularization can be used as a new method to determine the source of water inrush in a rapid and effective way.