基于贡献因子BP神经网络的地磁适配性研究

Research on adaptability evaluation of underground geomagnetic positioning based on BP neural network with contribution factor

文章针对井下地磁空间分布特点及小样本数据,提出一种基于贡献因子后向传播(back propagation,BP)神经网络的井下地磁适配性评价方法;通过对井下地磁空间分布7个特征参数的回归分析,确定其在适配性评价过程的贡献因子,将贡献因子作为BP神经网络先验输入权值,进行地磁适配性评价。试验选取45个人防工程小区域样本,计算地磁标准差、相关系数、地磁粗糙度等7个特征参数,对贝叶斯判别法、线性距离判别法、二次函数判别法、普通BP神经网络和基于贡献因子BP神经网络5种评价方法进行适配性评价精度对比。试验结果表明,贝叶斯判别法、线性距离判别法、二次函数判别法3种方法对训练样本的判别准确率为80%左右,但测试样本准确率仅50%左右,判别精度不高;而基于贡献因子BP神经网络对训练样本的判别准确率达到了95%,测试样本准确率接近73%,明显优于传统适配性评价方法,且一定程度上克服了普通BP神经网络易陷入局部收敛和收敛速度慢的缺点。基于贡献因子BP神经网络的评价方法能够有效避免人工构造评价规则的盲目性和样本数量较少的缺点,评价过程快捷,可为实现井下地磁定位导航的智能化提供基础。

This paper proposes a new method of the adaptability evaluation based on back propagation(BP)neural network with the contribution factor,which is suitable for the small sample of underground geomagnetic data.The BP evaluation of geomagnetic adaptability is carried out using the regression analysis of multiple characteristic parameters of underground geomagnetic spatial distribution,determining the contribution factor of the adaptability evaluation process,and calculating the BP prior input weight.In the experiment,45 plots samples of civil air defence engineering were selected,and seven characteristic parameters of geomagnetic spatial distribution including standard deviation,correlation coefficient and roughness were calculated.And the accuracy comparison of the adaptability evaluation of the five methods including Bayesian discriminant method,linear distance discriminant method,quadratic function discriminant method,generic BP neural network and BP neural network with the contribution factor was carried out.The experimental results showed that the accuracy rate in training samples of the Bayesian discriminant method,the linear distance discriminant method and the quadratic function discriminant method was about 80%,while the accuracy rate in test samples was only about 50%,and the discriminating precision was low.The accuracy rate in training samples of BP neural network with the contribution factor reached more than 95%,and the accuracy rate in test samples reached 73%,which was obviously superior to the traditional adaptability evaluation methods,and overcame the shortages of the generic BP neural network that was easy to fall into local convergence and had slow convergence rate to some extent.This method can effectively avoid the blindness of artificial construction evaluation rules and the shortcomings of the small sample scale,and the evaluation process is quite fast.It can provide a foundation for the intelligence of underground geomagnetic positioning navigation.