神经网络定位中多点多映射概率样本增强方法研究

Research on construction method for training set in neural network location technology

神经网络室内定位算法的定位精度取决于训练样本的数量和样本对数据特征的表征性,针对样本采集工作量大的问题,本文提出了一种多点多映射概率样本的增强方法。该方法首先根据无线信号的传播特性以及空间相关性,利用少量实测点数据,快速构建多点反应环境影响的特征,然后利用概率法生成表征无线信号波动的多映射正态分布样本数据。实验结果表明,利用多点多映射概率样本增强方法生成的训练样本集,有效数据量大,信号表征性好,且人工采集工作量低,由此训练的神经网络定位算法定位精度高。

The precision of the indoor positioning algorithm basing on the neural network depends on the number and the characterization of training samples.This paper proposes a method for enhancing samples with multi-point and multi-mapping probability feature to simplify the collection work of samples.First,according to the propagation characteristics and spatial correlation of wireless signals,the characteristics of multi-point concerning environmental influences are quickly constructed by small amount of measured data.Then,multi-mapping normal distribution samples indicating fluctuations characterize of wireless signal are generated using probabilistic methods.The experimental results show that the training samples generated by this enhancement method has many advantages such as large amount of effective data,good signal representation and low manual collection workload.Using those training samples to train the neural network can improve the positioning precision.