Search

CN-121995336-A - MSTA-UNet-based full-waveform inversion method for multi-scale ground penetrating radar

CN121995336ACN 121995336 ACN121995336 ACN 121995336ACN-121995336-A

Abstract

The invention discloses a multi-scale ground penetrating radar full-waveform inversion method based on MSTA-UNet, which comprises the steps of S1 decomposing original GPR data into low, medium and high frequency three-way data through wavelet transformation, S2 constructing an MSTA-UNet network, extracting deep features of each frequency band through a multi-branch module, S3 fusing cross-frequency band features through a Fusion attribute module, and restoring resolution by deconvolution of a decoder, and S4 outputting inversion results containing dielectric constants by adopting a mixed loss function training model of a data domain and a physical domain FWI. According to the invention, cycle skip is relieved through multi-scale frequency division, the defect of 'black box' is overcome by fusing data and physical driving, the detection depth and resolution are balanced, the inversion precision, generalization capability and calculation efficiency are improved, and the method is suitable for shallow geophysical detection scenes such as urban underground pipeline detection and road disease detection.

Inventors

  • SHENG GUANQUN
  • LU DONGPING
  • TAN YUNZHI

Assignees

  • 三峡大学

Dates

Publication Date
20260508
Application Date
20260121

Claims (10)

  1. 1. The MSTA-UNet-based full-waveform inversion method for the multi-scale ground penetrating radar is characterized by comprising the following steps of: S1, acquiring original Ground Penetrating Radar (GPR) data, performing multi-resolution preprocessing through wavelet transformation, decomposing, reconstructing and separating to obtain low-frequency, medium-frequency and high-frequency data, and taking the low-frequency, medium-frequency and high-frequency data as three independent input channels; s2, constructing an MSTA-UNet multi-scale feature extraction network, and carrying out multi-branch feature extraction module, downsampling and feature deepening treatment on frequency division data to obtain deep coding features of each frequency band; S3, weighting and fusing coding features of each frequency band through a Fusion attribute module, and recovering spatial resolution through deconvolution by a decoder after bottleneck layer processing to realize cross-frequency band feature self-adaptive enhancement and decoding; S4, constructing a mixed loss function containing data domain loss and physical domain Full Wave Inversion (FWI) loss, and inputting test data to output inversion results of physical parameters of the underground medium including dielectric constants after iterative training of network parameters.
  2. 2. The MSTA-UNet-based full-waveform inversion method for the multi-scale ground penetrating radar according to claim 1, wherein the wavelet basis function in S1 is 'db 4' wavelet, and the number of decomposition layers of discrete wavelet transformation is And obtaining low, medium and high frequency data with 3-5 layers through inverse wavelet transform (IDWT) and length correction.
  3. 3. The MSTA-UNet-based full-waveform inversion method for the multi-scale ground penetrating radar according to claim 1, wherein the four convolution kernel sizes of the multi-branch feature extraction module in S2 are 1×1,3×3,5×5 and 7×7, respectively, and the feature extraction process is as follows: realizes characteristic splicing and fusion, wherein in the process, Is an input feature; Representing a convolution operation with a convolution kernel of size k; concat denotes feature stitching; Is a fused feature.
  4. 4. The MSTA-UNet-based multi-scale ground penetrating radar full-waveform inversion method according to claim 1, wherein the data field loss in S4 is a weighted combination of Mean Absolute Error (MAE) and Structural Similarity (SSIM), and the total data loss is expressed by the following formula: calculating, wherein, the calculation method comprises the steps of, The total loss of data is calculated as a function of the total loss of data, 、 To balance the coefficients of MAE and SSIM weights, For average absolute error loss, SSIM is structural similarity.
  5. 5. The MSTA-UNet-based full-waveform inversion method for multi-scale ground penetrating radar according to claim 1, wherein said physical domain FWI loss in S4 is used to construct forward modeling operator by Finite Difference Time Domain (FDTD) Substituting physical parameters of underground medium predicted by network into Maxwell equation set to generate simulated radar profile Calculating the simulated section And observing radar data Is a L2 norm residual of (c).
  6. 6. The MSTA-UNet-based multi-scale ground penetrating radar full-waveform inversion method according to claim 1, wherein the weighted Fusion of Fusion attribute modules in S3 satisfies: Wherein, the For the feature after the cross-band fusion, 、 、 Deep coding features of low, medium and high frequency bands respectively, Weight parameters automatically learned for attention mechanisms and satisfying 。
  7. 7. The MSTA-UNet-based full-waveform inversion method of the multi-scale ground penetrating radar according to claim 1, wherein the activation function of the decoder in S3 is a ReLU function, and the decoding process is as follows: The realization is that, wherein, For the output profile of the i-th layer decoder, reLU is a linear rectification activation function, For the convolution weights of the decoder, For convolution operation symbols, upsample is an up-sampling operation, Is an output characteristic diagram of the i-1 layer decoder, For a corresponding hierarchy of fused hopped connection features, Is a convolution offset for the decoder.
  8. 8. The MSTA-UNet-based full-waveform inversion method for the multi-scale ground penetrating radar according to claim 1, wherein the optimizer in S4 is an Adam optimizer, gradients are calculated through a back propagation algorithm, and network parameters are updated iteratively until the model converges.
  9. 9. The MSTA-UNet-based full-waveform inversion method of the multi-scale ground penetrating radar according to claim 1, wherein the deep feature coding in S2 is a four-level coding process, the downsampling process is implemented by a max pooling operation, and the max pooling is expressed by the formula: calculating, wherein, the calculation method comprises the steps of, Output feature map location after pooling The numerical value of the position is calculated, For the window area to be maximally pooled, For the amount of coordinate offset within the window, For input to the feature map of the pooling operation, To output the coordinate position of the feature map.
  10. 10. The MSTA-UNet-based full-waveform inversion method of the multi-scale ground penetrating radar according to claim 1, wherein the dielectric constant in the physical parameters of the underground medium is used for identifying the layering of the underground medium and judging the abnormal body, and the inversion error of the dielectric constant in a shallow detection scene with the detection depth of 50 cm-10 m is not more than 5%.

Description

MSTA-UNet-based full-waveform inversion method for multi-scale ground penetrating radar Technical Field The invention relates to the technical field of ground penetrating radar detection and geophysical exploration, in particular to a multi-scale ground penetrating radar full-waveform inversion method based on MSTA-UNet. Background Ground Penetrating Radar (GPR) is used as an efficient and lossless shallow geophysical detection technology, and has wide application in the fields of urban underground space development, infrastructure detection and the like. However, accurate reconstruction of subsurface medium physical parameters from complex radar echo data is a core technical challenge in this field. Although the traditional Full Waveform Inversion (FWI) method can theoretically provide high-resolution underground structure information, the method is essentially a highly nonlinear uncertainty problem, has extremely strong dependence on an initial model, is easy to sink into a local minimum, has high calculation cost, and is difficult to meet the real-time requirement. In recent years, a deep learning driven inversion method provides a new thought for the problem, but the method has the defects of black boxes generally, has insufficient physical interpretability, has generalization capability limited by the completeness of a training data set, and is difficult to adapt to complex and changeable geological detection environments. Disclosure of Invention The invention aims to provide a multi-scale ground penetrating radar full-waveform inversion method based on MSTA-UNet, which solves the problems that the existing full-waveform inversion is strong in nonlinearity, easy to occur cycle skip and remarkable in dependence on an initial model. In order to solve the problems, the technical scheme of the invention is as follows: A multi-scale ground penetrating radar full-waveform inversion method based on MSTA-UNet comprises the following steps: S1, acquiring original Ground Penetrating Radar (GPR) data, performing data preprocessing by utilizing multi-resolution analysis characteristics of wavelet transformation, and constructing a multi-band input data set, wherein a wavelet basis function is selected to perform Discrete Wavelet Transformation (DWT) on the original GPR data, the Discrete Wavelet Transformation (DWT) is performed to obtain approximate coefficients and detail coefficients, and low-frequency data, intermediate-frequency data and high-frequency data are obtained through coefficient reconstruction and separation and serve as three independent input channels; S2, constructing an MSTA-UNet multi-scale feature extraction network, and carrying out deep feature coding on each frequency band data, namely after inputting frequency division data into an initialized convolution layer, capturing features by a multi-branch feature extraction module containing four convolution kernels with different sizes, and obtaining deep coding features of each frequency band by alternating maximum pooling and downsampling and feature deepening of a convolution+multi-scale module; S3, introducing a jump connection Fusion mechanism and bottleneck layer processing to realize cross-frequency band characteristic self-adaptive enhancement and decoding, wherein each frequency band coding characteristic is subjected to weighted Fusion through a Fusion attribute module, the bottleneck layer generates a global bottleneck characteristic, and a decoder recovers spatial resolution through a cascade decoding module and deconvolution; S4, constructing a mixed loss function containing data domain loss and physical domain Full Wave Inversion (FWI) loss, iteratively updating network parameters by using a back propagation and optimizer, inputting test data after training, and outputting an inversion result of the physical parameters of the underground medium. Further, the wavelet basis function in S1 is db4 wavelet, and the number of decomposition layers of discrete wavelet transformationAnd obtaining low, medium and high frequency data with 3-5 layers through inverse wavelet transform (IDWT) and length correction. Further, the four convolution kernel sizes of the multi-branch feature extraction module in S2 are 1×1,3×3, 5×5 and 7×7, respectively, and the feature extraction process is as follows: realizes characteristic splicing and fusion, wherein in the process, Is an input feature; Representing a convolution operation with a convolution kernel of size k; concat denotes feature stitching; Is a fused feature. Further, the data field loss in S4 is a weighted combination of Mean Absolute Error (MAE) and Structural Similarity (SSIM), and the total data loss is expressed by the following formula: calculating, wherein, the calculation method comprises the steps of, The total loss of data is calculated as a function of the total loss of data,、To balance the coefficients of MAE and SSIM weights,For average absolute error loss (measuring pixel level