基于扩散模型的室内定位射频指纹数据增强方法

Radio frequency fingerprint data augmentation for indoorlocalization based on diffusion model

射频指纹室内定位方法通过在离线阶段采集足量信号指纹建立密集指纹库保证定位精度。为降低指纹采集成本,提出一种基于扩散模型的射频指纹数据增强方法(FPDiffusion)。首先建立指纹序列的时序图表示,通过高斯加噪方法实现扩散模型的前向过程,反向过程采用U型自编码器网络,根据射频指纹特性设计了网络的损失函数,最后给出了基于稀疏指纹生成稠密指纹的计算过程。实验结果表明,在仅有少量有标签指纹的情况下,FPDiffusion方法在K-近邻(KNN)和卷积神经网络(CNN)算法上的定位误差降低率分别达到76%和28%,在KNN上的定位精度相比高斯过程回归(GPR)和GPR-GAN方法有显著提升。

The radio frequency fingerprint indoor localization method ensures the accuracy by collecting a sufficient amount of fingerprints in the offline state to build a dense fingerprint database.A data augmentation method called FPDiffusion was proposed based on diffusion model to reduce the cost of fingerprint acquisition.Firstly,a temporal graph representation of the fingerprint sequence was constructed,the forward process of the diffusion model was accomplished by adding Gaussian noise,and a U-Net was utilized for the reverse process.The loss function of the network was designed according to the characteristics of radio frequency fingerprints.Finally,the computational process for generating dense fingerprints based on sparse fingerprints was presented.Experimental results demonstrate that FPDiffusion achieves 76%and 28%localization error reduction on K-nearest neighbor(KNN)and convolutional neural network(CNN)respectively,and significantly improves localization accuracy on KNN compared to Gaussian process regression(GPR)and GPR-GAN when only a small amount of labeled fingerprints is available.