Signal Denoising Of Ground-penetrating Radar Based On Depth Neural Network

Ground-Penetrating Radar(GPR)is an effective means to detect the distribution of subsurface media and is widely used in urban construction,archaeology,and military fields.However,due to the uneven distribution of the subsurface medium,the scattering of electromagnetic waves by the subsurface target,and the equipment error,there is usually noise in the collected GPR signals,which seriousl y affects the detection of subsurface targets.Thus,improving the signal-to-noise ratio of GPR signals and accurately restoring the GPR signals under the influence of noise is one of the urgent problems in GPR application fields.Deep learning networks can complete the learning of data features through data support and extract useful data from cluttered data.To this end,this paper trains the network using GPR signals with the help of deep learning methods and adds noise to the training process so that it can complete the learning of GPR signals and noise features,generate prediction models,and separate noise from the signal by prediction models for noisy GPR signals to achieve the purpose of denoising.The innovative results of this paper include the following:(1)The Inception Denoise-Net(ID-Net)network for GPR signal denoising is designed to address the problem that conventional signal-denoising algorithms need to be more robust to denoising GPR data with different characteristics.In the ID-Net network,an encoder structure is used to learn the multiscale characteristics of GPR signals,and a decomposer structure generates a mask of radar signals and noise and a decoder to reconstruct the radar signal for the purpose of denoising.The network uses the Inception structure as the base component for learning the topological relationship between adjacent sampling points of the GPR signal to improve the model prediction accuracy.The Signal MSE loss function is also designed to enable the model to retain more signal details in the process of GPR signal denoising.Experiments demonstrate that the ID-Net denoising network has denoising robustness for GPR signals with different characteristics and can maintain the original waveform of GPR signals while improving the signal-to-noise ratio of noisy signals.(2)Aiming at the problem of high-frequency noise contained in the GPR detection data and less open standard data sets in the GPR field,this paper proposes a Multi-Stage Attention Net(MSA-Net)for high-frequency noise suppression of GPR signals by combining the transfer learning method with the deep signal denoising network.The MSA-Net network incorporates a multi-level attention structure designed based on the multi-scale and timefrequency features of GPR data,significantly improving the network’s suppression effect on high-frequency noise in GPR signals.Experiments have shown that MSA-Net networks can effectively suppress high-frequency noise in GPR signals and achieve maximum fidelity of GPR signal waveforms.