Research On Signal Denoising Of Ground Penetrating Radar Based On Deep Learning

Ground-penetrating radar is a non-destructive,economic,and fast geophysical exploration instrument that plays a vital role in geological exploration,soil structure analysis,and petroleum exploration.However,in the process of ground-penetrating radar data acquisition,there are various interference factors that affect the accuracy of data acquisition.Therefore,the problem of ground-penetrating radar signal denoising has been of great concern.Since ground-penetrating radar noise signals usually exhibit the complex characteristics of co-frequency mixing with the target signal,the traditional algorithms are not ideal for processing ground-penetrating radar data acquired in complex geological structures,and noise reduction are difficult.To effectively improve the noise reduction level of ground-penetrating radar signals,deep neural networks are introduced to perform efficient and intelligent noise reduction of ground-penetrating radar signals and achieve better processing results than traditional methods.(1)A Modified Convolutional autoencoder denoising network(MCDAE)for groundpenetrating radar signals is proposed for the complex ground-penetrating data noise suppression problem,and a Gaussian noise dataset and a random noise dataset are introduced to evaluate the noise removal effect of the model.The MCDAE network encoder and decoder consist of 13 layers of network structure.The encoder extracts data features from the ground-penetrating radar data,and the decoder is symmetric to the encoder and is responsible for reconstructing the compressed data features into the corresponding groundpenetrating radar signals.To verify the effectiveness of the proposed algorithm,the trained network model is processed and analyzed against the ground-penetrating radar real-world data,and the denoising results are compared with those of classical,traditional algorithms and deep learning algorithms.The experimental results show that the MCDAE network model effectively suppresses the noise signal and reconstructs the denoised signal better.(2)A multitasking encoder network model(Mt Net)with a fused attention mechanism is proposed to address the problem of incomplete removal of low-intensity noise by the MCDAE network model.The channel and spatial attention mechanisms are incorporated in encoding and decoding,respectively,and the composite attention mechanism of channel and space is used to optimize the feature extraction method and improve the feature characterization ability.In the intermediate separation and mapping stage,the Transformer model is used to learn the mapping relationship between the composite waveform features extracted by the encoder and the separated ground-penetrating radar signal and noise signal time series features through the Transformer model,and finally,the loss function is used to further optimize the model to make it more suitable for ground-penetrating radar noise suppression.To verify the effectiveness of the proposed algorithm,the trained network model is processed and analyzed against the ground-penetrating radar real-world data and compared with several classical,traditional algorithms and deep learning denoising algorithms.The results show that the multitask encoder model incorporating the attention mechanism can effectively suppress the complex random noise in the ground-penetrating radar data and,at the same time,has good recovery capability for the ground-penetrating radar signal.(3)The time-frequency domain learn-ing neural network(Tef Net)is proposed from the perspective of conversion domain.The idea of conversion domain is used to realize the ground-penetrating radar signal time-frequency domain denoising research.The Tef Net network is constructed for the high-intensity interference generated by the ground-penetrating radar noise data.Firstly,the original ground-penetrating radar data are converted to the time-frequency domain using short-time Fourier transform and input to the network,then the data representation and data flow in the time-frequency domain are obtained by using convolution combined with the tandem structure of multi-scale residual blocks(MRDB),and the mapping of noisy ground-penetrating radar data and target data is calculated,and finally the output data are converted to the time-domain denoised signal by short-time Fourier inverse transform.The trained network model is analyzed to denoise the ground-penetrating radar data,and the results show that Tef Net can effectively remove the noise generated in the ground-penetrating radar data acquisition process and has a strong ability to reconstruct the original signal.