GPR signal denoising method and system based on multi-scale feature fusion network

Abstract

The invention provides a GPR signal denoising method and system based on a multi-scale feature fusion network, and relates to the field of ground penetrating radar data processing. The method comprises the steps of obtaining sparse target signal data and dense target signal data, and fusing the sparse target signal data and the dense target signal data to obtain a GPR signal data set; the method comprises the following steps: constructing a signal denoising model based on an auto-encoder network, comprising an encoder, a resolver and a decoder: improving the auto-encoder network based on Inceptionblock, introducing an Inceptionblock as a basic component, adding a resolver structure between the encoder and the decoder, each resolver structure is formed by connecting four Inceptionblocks with a 1 * 1 convolution, and introducing Skip connection between the first Inceptionblock and the fourth Inceptionblock; the structure of the resolver is based on an encoder, and then the resolver is connected with a Sigmoid activation function module; inputting a GPR signal data set to be detected into the signal denoising model to obtain a model output result; signal features are effectively reserved, the quality of the denoised signal is improved, and the accuracy of identifying and positioning the underground target is improved.

Description

Technical Field

[0001] The present invention relates to the field of ground penetrating radar data processing, and in particular to a GPR signal denoising method and system based on a multi-scale feature fusion network. Background Art

[0002] Ground Penetrating Radar (GPR), an important nondestructive testing technology for underground exploration, is widely used to study near-surface geophysical structures and detect buried targets in a wide range of fields, including geotechnical engineering, archaeology, and environmental monitoring. GPR primarily transmits high-frequency electromagnetic waves underground and then receives signals reflected from subsurface targets. These signals are analyzed and processed to infer information about the target's distribution, structure, and geology.

[0003] However, during actual detection, GPR signals are subject to various types of noise interference. This noise comes from a wide range of sources, including external electromagnetic interference, instrument-specific electronic noise, and subsurface inhomogeneities. All of these factors contribute to the inclusion of significant noise in the received GPR signals. The presence of noise blurs the signal, obscuring the target's reflection signature. This, in turn, severely impacts the accurate identification and location of underground targets, reducing the reliability and accuracy of detection results.

[0004] While several GPR signal denoising methods exist, such as traditional filtering and wavelet transforms, these methods have limitations in handling complex noise and preserving signal characteristics. Traditional filtering methods can only remove noise within a specific frequency range and are ineffective against non-stationary and complex noise. While wavelet transforms offer advantages in time-frequency analysis, they struggle to achieve ideal denoising results when dealing with multi-scale and multi-feature noise. Therefore, effectively removing noise from GPR signals while preserving their useful characteristics remains a critical challenge in the field of GPR technology. Summary of the Invention

[0005] The main purpose of the present invention is to provide a GPR signal denoising method and system based on a multi-scale feature fusion network to solve the technical problems in the prior art that the removal of complex noise is not thorough enough and useful features are easily lost or unclear.

[0006] To solve the above technical problems, the technical solution adopted by the present invention is: a GPR signal denoising method based on a multi-scale feature fusion network, comprising the following steps: S1: Acquire sparse target signal data and dense target signal data, pre-process them, fuse them according to a preset ratio, and save the fused data into a GPR signal dataset; S2: Build a signal denoising model based on the autoencoder network, input the GPR signal dataset and perform model training, and save the optimal signal denoising model; The signal denoising model includes an encoder, a decomposer, and a decoder. The autoencoder network is improved based on the Inception_block, and the Inception architecture is introduced as a basic component. A decomposer structure is added between the encoder and decoder. The convolution in the Inception architecture is a one-dimensional convolution. The decoder has the same structure as the encoder, consisting of four Inception_blocks followed by a 1×1 convolution. A skip connection is introduced between the first and fourth Inception_blocks. The decomposer structure is based on the encoder structure, and then connected to the Sigmoid activation function module. S3: Input the GPR signal data set to be tested into the optimal signal denoising model to obtain the model test results.

[0007] In the preferred solution, the encoder extracts features from the input data and compresses them to obtain an embedding vector; the decomposer uses the Inception architecture to extract the contour features and multi-scale features of the GPR signal and noise, reuses the input data features, and uses the Sigmoid activation function to generate a feature mask; the decoder reconstructs the denoised GPR signal based on the embedding vector and feature mask, and introduces Skip connections to retain shallow features.

[0008] The preferred solution further includes S4: evaluating the model test results using signal-to-noise ratio and cosine similarity, specifically: S41: The signal-to-noise ratio (SNR) is used to evaluate the denoising effect of the model. The formula is: ; Where, For a clean GPR signal, is the GPR signal after denoising, N is the number of GPR signal sampling points; S42: Cosine similarity is used to evaluate the denoising fidelity performance. The formula for cosine similarity is as follows: ; Where, For a clean GPR signal, is the GPR signal after denoising, and N is the number of sampling points of the GPR signal.

[0009] In the preferred solution, S1: acquiring sparse target signals and dense target signals, where both the sparse target signal data and the dense target signal data are collected by a radar on clay geology, and the radar antenna frequency is set to 400 MHz; Preprocessing specifically includes: S11: Perform multimodal characteristic analysis. Based on the time domain characteristics of the signal, a clustering algorithm is used to perform modal classification on the collected raw GPR data. The formula is: ; Where, is the original GPR data, and its time domain feature vector Including characteristics such as amplitude and frequency; Use K-Means algorithm to cluster the feature vectors: ; Where, is the i-th cluster center, k is the number of clusters, is the new cluster center; S12: Adaptive noise estimation is then performed, combining wavelet transform and empirical mode decomposition (EMD) technology to decompose the signal and estimate the noise level; ; Where, is the scale parameter, is the translation parameter, is the wavelet basis function; The signal Decomposed into several IMFs, the formula is: ; Where, is the i-th IMF, is the residual; If the high-frequency IMF mainly contains noise, the noise level can be estimated by calculating the energy of the high-frequency IMF: ; Where, is the estimated noise variance, is the number of high-frequency intrinsic mode functions (IMFs), is the number of sampling points of the signal; S13: Based on feature matching, the interference signal is removed according to the signal similarity; S14: Finally, dynamic time zero point correction is performed; S15: Normalize the preprocessed GPR data using the Min-Max normalization method to scale the data to the [0, 1] interval. The formula is: ; Where, is the value of the normalized signal at time point t, and Signals The minimum and maximum values ​​in ; S16: Divide the normalized data into a training set, a validation set, and a test set according to a preset ratio.

[0010] In the preferred embodiment, in S2, the model training step specifically includes: The signal denoising model uses two GPR signal datasets with different noise levels and one original GPR signal dataset as input. The model outputs the corresponding clean GPR signal and noise waveform, respectively. The radar noise and radar signal characteristics are learned and distinguished through the structure of the network decomposer. At the end of the iterative training, the current optimal model parameters are saved and the best model is updated. In the model training, the batch size is set to 100, the learning rate is 0.0004, and the number of epochs is 1000. The custom SignalMSE loss function is used. The SignalMSE loss function combines the signal variation and the mean square error. The formula is: ; Where, and are signal weight and MSE weight respectively, N is the number of data points, that is, the number of sampling points of the signal, and are the differences between adjacent sampling points of the target signal and the output signal, is the mean square error, and are the output value of the model and the true value respectively.

[0011] In the preferred solution, S2: In the signal denoising model based on the autoencoder network, each Inception_block extracts features of different scales through successively increasing convolution kernels and splices the feature maps to improve the network performance and efficiency; the Inception module uses the Relu nonlinear activation function, which is expressed as: ; Where, is the input value, is the output value of the ReLu function.

[0012] In the preferred solution, the fusion in S1 is performed according to a preset ratio, and a weighted fusion method based on energy matching is adopted: The energy of the sparse target signal and the dense target signal after preprocessing is calculated respectively. The energy of the signal can be calculated by integrating the square of its amplitude. For discrete signals, the formula is: ; Where, is the value of the signal at the i-th sampling point, is the number of sampling points; According to the calculated energy value, the weighting coefficient of the two signals is determined; the weighting coefficient is dynamically adjusted according to the energy ratio of the signals. The formula is: ; Where, is the energy of the sparse target signal, is the energy of the dense target signal, and is the weighting coefficient; The smoothing factor γ is introduced to smooth the weighted coefficients: ; Where, and is the weighting coefficient after smoothing; and is the weighted coefficient calculated last time; γ is the smoothing factor; The pre-processed sparse target signal and dense target signal are fused according to the determined weight coefficient to obtain a mixed signal. The formula is: ; Where, is the value of the fused mixed signal at the i-th sampling point, is a sparse target signal, It is a dense target signal.

[0013] A GPR signal denoising system based on a multi-scale feature fusion network includes the following modules: The data acquisition and processing module is used to obtain sparse target signal data and dense target signal data, and after pre-processing, fuse them according to a preset ratio and save the fused data into the GPR signal data set; The model building module is used to build a signal denoising model based on the autoencoder network, input the GPR signal dataset and perform model training, and save the optimal signal denoising model; The signal denoising model includes an encoder, a decomposer, and a decoder. The autoencoder network is improved based on the Inception_block, and the Inception architecture is introduced as a basic component. A decomposer structure is added between the encoder and decoder. The convolution in the Inception architecture is a one-dimensional convolution. The decoder has the same structure as the encoder, consisting of four Inception_blocks followed by a 1×1 convolution. A skip connection is introduced between the first and fourth Inception_blocks. The decomposer structure is based on the encoder structure, and then connected to the Sigmoid activation function module. The test module is used to input the GPR signal data set to be tested into the optimal signal denoising model to obtain the model test results.

[0014] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a GPR signal denoising method based on a multi-scale feature fusion network.

[0015] An electronic device comprising a memory and a processor; The memory is used to store computer programs; The processor is used to implement the steps of the GPR signal denoising method based on a multi-scale feature fusion network when executing the computer program.

[0016] The present invention provides a GPR signal denoising method based on a multi-scale feature fusion network. Sparse target signal data and dense target signal data are obtained and fused to obtain a GPR signal dataset. A signal denoising model based on an autoencoder network is constructed, comprising an encoder, a decomposer and a decoder. The autoencoder network is improved based on the Inception_block, and the Inception architecture is introduced as a basic component. A decomposer structure is added between the encoder and the decoder. The decomposer structure consists of four Inception_blocks followed by one 1×1 convolution, and a Skip connection is introduced between the first Inception_block and the fourth Inception_block. The decomposer structure is based on the encoder structure and is then connected to a Sigmoid activation function module. The GPR signal dataset to be tested is input into the signal denoising model to obtain a model output result.

[0017] Technical effects: 1. Utilize the deep learning capability to deeply mine the multi-scale features of data, and combine it with feature classifiers to identify noise and GPR signals, thus achieving noise removal and clutter suppression for GPR signals.

[0018] 2. By accurately distinguishing sparse target signals, dense target signals, and mixed signals, multimodal characteristic analysis is performed, so that signals with different characteristics can be properly processed, improving the accuracy and efficiency of the overall processing. Then, a signal denoising model including an encoder, decomposer, and decoder is constructed. The signal denoising model of the improved autoencoder network comprehensively extracts the multi-scale features of the noise and effectively removes the complex noise in the GPR signal.

[0019] 3. The introduction of Skip connection and decomposer structure retains the shallow features and contour features of the signal while denoising, effectively retains the signal characteristics, avoids the problem of traditional denoising methods easily losing useful features, improves the quality of the denoised signal, and improves the accuracy of identifying and locating underground targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The present invention will be further described below with reference to the accompanying drawings and examples: Figure 1 is a flow chart of the GPR signal denoising method of the present invention; Figure 2 This is the traditional Inception module structure diagram used in this invention; Figure 3 It is a structural diagram of the encoder module of the present invention; Figure 4 It is a structural diagram of the decomposer module of the present invention; Figure 5 It is a structural diagram of the decoder module of the present invention; Figure 6 It is a structural diagram of the signal denoising model constructed by the present invention; Figure 7 Schematic diagram of sparse target signal denoising according to the present invention; Figure 8 This is a schematic diagram of dense target signal denoising according to the present invention; Figure 9 This is a comparative experimental diagram of the signal denoising method of the present invention; Figure 10 This is a visualization diagram of the denoising effect of the GPR noise signal of the present invention; Figure 11 This is a comparative experimental diagram of the GPR signal denoising method of the present invention. DETAILED DESCRIPTION

[0021] Example 1 like Figure 1-11 As shown, a GPR signal denoising method based on a multi-scale feature fusion network includes the following steps: S1: Obtain sparse target signal data and dense target signal data, pre-process them, and fuse them according to a preset ratio to construct a GPR signal dataset.

[0022] S2: Build a signal denoising model based on the autoencoder network, input the GPR signal dataset and perform model training, and save the optimal signal denoising model.

[0023] Among them, the signal denoising model includes an encoder, a decomposer and a decoder: the autoencoder network is improved based on the Inception_block, the Inception architecture is introduced as the basic component, and a decomposer structure is added between the encoder and decoder. The convolution in the Inception architecture is a one-dimensional convolution; the structure of the decoder is consistent with that of the encoder, both consisting of 4 Inception_blocks followed by 1×1 convolution, and a Skip connection is introduced between the first Inception_block and the fourth Inception_block; the decomposer structure is based on the encoder structure, and then connected to the Sigmoid activation function module.

[0024] S3: Input the GPR signal data set to be tested into the optimal signal denoising model to obtain the model test results.

[0025] S4: The model test results are evaluated using signal-to-noise ratio and cosine similarity.

[0026] In this embodiment, deep learning's ability to deeply mine multi-scale features of data is utilized, combined with a feature classifier to distinguish between noise and GPR signals, achieving noise removal and clutter suppression for GPR signals. By accurately distinguishing between sparse target signals, dense target signals, and mixed signals, and performing multimodal characteristic analysis, signals with different characteristics can be appropriately processed, improving the accuracy and efficiency of the overall processing. A signal denoising model, comprising an encoder, decomposer, and decoder, is then constructed. This improved autoencoder network-based signal denoising model comprehensively extracts the multi-scale features of noise, effectively removing the complex noise from GPR signals. The introduction of skip connections and a decomposer structure preserves the shallow and contour features of the signal while denoising, effectively preserving signal characteristics and avoiding the problem of traditional denoising methods easily losing useful features. This improves the quality of the denoised signal and the accuracy of identifying and locating underground targets.

[0027] The signal denoising model is an end-to-end one-dimensional network (1D-Net).

[0028] In the preferred embodiment, step S4 specifically includes the following steps: S41: The signal-to-noise ratio (SNR) is used to evaluate the denoising effect of the model. The formula is: ; Where, For a clean GPR signal, is the denoised GPR signal, and N is the number of GPR signal sampling points.

[0029] S42: Cosine similarity is used to evaluate the denoising fidelity performance. The formula for cosine similarity is as follows: ; Where, For a clean GPR signal, is the GPR signal after denoising, and N is the number of sampling points of the GPR signal. The closer the value is to 1, the closer the angle is to 0, the closer the denoised waveform is to the original signal waveform, and the better the fidelity of the denoising method.

[0030] This example uses the signal-to-noise ratio (SNR) as an evaluation metric and introduces the SignalMSE loss function for model training. This combines the signal variation and mean square error (MSE) to more comprehensively measure the denoising effect. This allows the model to better optimize denoising performance during training, improving the reliability and accuracy of the denoising results.

[0031] In the preferred solution, in step S1, the fusion is performed according to a preset ratio, and a weighted fusion method based on energy matching is adopted.

[0032] The energy of the sparse target signal and the dense target signal after preprocessing is calculated respectively. The energy of the signal can be calculated by integrating the square of its amplitude. For discrete signals, the formula is: ; Where, is the value of the signal at the i-th sampling point, is the number of sampling points.

[0033] According to the calculated energy value, the weighting coefficient of the two signals is determined; the weighting coefficient is dynamically adjusted according to the energy ratio of the signals. The formula is: ; Where, is the energy of the sparse target signal, is the energy of the dense target signal, and is the weighting coefficient.

[0034] The smoothing factor γ is introduced to smooth the weighted coefficients: ; Where, and is the weighting coefficient after smoothing; and is the weighting coefficient calculated last time; γ is the smoothing factor.

[0035] The pre-processed sparse target signal and dense target signal are fused according to the determined weight coefficient to obtain a mixed signal. The formula is: ; Where, is the value of the fused mixed signal at the i-th sampling point, is a sparse target signal, It is a dense target signal.

[0036] In the preferred solution, step S1: obtaining a sparse target signal and a dense target signal, includes: Both the sparse target signal data and the dense target signal data are collected by radar on clay geology, and the radar antenna frequency is set to 400 MHz.

[0037] Preprocessing specifically includes: S11: Perform multimodal characteristic analysis. Based on the time domain characteristics of the signal, a clustering algorithm is used to perform modal classification on the collected raw GPR data. The formula is: ; Where, is the original GPR data, and its time domain feature vector Including characteristics such as amplitude and frequency.

[0038] Use K-Means algorithm to cluster the feature vectors: ; Where, is the i-th cluster center, k is the number of clusters, is the new cluster center.

[0039] S12: Adaptive noise estimation is then performed, combining wavelet transform and empirical mode decomposition (EMD) technology to decompose the signal and estimate the noise level; ; Where, is the scale parameter, is the translation parameter, is the wavelet basis function.

[0040] The signal Decomposed into several intrinsic mode functions (IMFs), the formula is: ; Where, is the i-th IMF, is the residual.

[0041] If the high-frequency IMF mainly contains noise, the noise level can be estimated by calculating the energy of the high-frequency IMF: ; Where, is the estimated noise variance, is the number of high-frequency intrinsic mode functions (IMFs), is the number of sampling points of the signal.

[0042] S13: Based on feature matching, the interference signal is removed according to signal similarity.

[0043] S14: Finally, dynamic time zero point correction is performed.

[0044] S15: Normalize the preprocessed GPR data using the Min-Max normalization method to scale the data to the [0, 1] interval. The formula is: ; Where, is the value of the normalized signal at time point t, and Signals The minimum and maximum values ​​in .

[0045] S16: Divide the normalized data into training set, validation set and test set, and divide the data according to a certain ratio (such as 70% training set, 15% validation set, and 15% test set).

[0046] In this example, a preset ratio of 1:1 was used to construct a measured GPR signal dataset with the same settings: a SIR-3000 ground penetrating radar was used, the radar antenna frequency was set to 400 MHz, and a total of 1000 data channels with 500 sampling points were collected. The collected data were preliminarily processed and stored in A-Scan format, where: 1) Constructing a sparse target signal dataset: In clay geology, there are fewer underground targets, so the characteristics of the reflected echo are short duration and low pulse repetition frequency.

[0047] 2) Constructing a dense target signal dataset: On gravel roads, there are many underground targets, which strongly reflect electromagnetic waves. Therefore, the characteristics of the reflected echo are long duration and high pulse repetition rate.

[0048] 3) The preprocessed sparse target signal set and dense target signal set are fused in a certain ratio.

[0049] This embodiment adopts a 1:1 ratio: 900 data channels are selected from the sparse target signal data set, and 900 data channels are selected from the dense target signal data set, and the data are combined into a mixed signal data set containing 1800 channels.

[0050] Step S2 is to build a model and train the autoencoder network based on Inception_block.

[0051] In the preferred solution, in step S2, the encoder extracts features from the input data and compresses the representation to obtain an embedding vector; the decomposer uses the Inception architecture to extract the contour features and multi-scale features of the GPR signal and noise, reuses the input data features, and uses the Sigmoid activation function to generate a feature mask; the decoder reconstructs the denoised GPR signal based on the embedding vector and the feature mask, and introduces Skip connections to retain shallow features.

[0052] Traditional autoencoders mainly consist of an encoder and a decoder. This embodiment introduces the Inception architecture as the basic component of the network and introduces a decomposer structure between the encoder and decoder to enable the network to better separate signal and noise features, thereby improving the denoising effect.

[0053] like Figure 2 As shown in the figure, the principle of the Inception architecture is that the convolution kernels in each Inception_block are different, and features are extracted through multiple convolution kernels of different sizes, and feature maps of different sizes are spliced ​​together.

[0054] This embodiment replaces all convolutions in the original architecture with one-dimensional convolutions, with kernel sizes of 1, 3, 5, and 11. Using one-dimensional convolutions, the data structure of the GPR signal remains unchanged during feature learning, while extracting data features and retaining more detailed information about the GPR signal. Inception, by running multiple convolutions in parallel, increases the network's denoising generalization when processing signals with different characteristics, extracting multi-scale features of the GPR signal. After performing the parallel convolutions, the feature maps of different scales are concatenated together through cascade operations, achieving multi-scale feature fusion. This enhances the denoising model's ability to represent GPR data and improves the model's predictive performance and robustness.

[0055] In the preferred solution, step S2: constructing a signal denoising model based on an autoencoder network, each Inception_block extracts features of different scales through successively increasing convolution kernels and splices the feature maps to improve network performance and efficiency; the Inception module uses the Relu nonlinear activation function, which is expressed as: ; Where, is the input value, is the output value of the ReLu function.

[0056] like Figure 6As shown in Figure 1, a GPR signal denoising model is constructed by improving the denoising autoencoder network and adding a decomposer structure to separate the noise characteristics of the input data from the signal characteristics. This signal denoising model is an end-to-end network that does not change the data shape from input to output. It can fully process GPR signals, reducing processing steps and feature information loss, thereby improving model prediction performance.

[0057] like Figure 3 As shown in the figure, the improved encoder is composed of 4 Inception_blocks and 1×1 convolutions. Each Inception_block extracts features of different scales through convolution kernels of different sizes, while 1×1 convolution is used for feature fusion and dimensionality reduction. Skip connection is introduced to retain shallow features, prevent feature loss, and enhance the expressiveness of the model.

[0058] like Figure 4 As shown in the figure, the improved decomposer uses four Inception_blocks for synchronous feature extraction, introduces Skip connections to reduce feature loss, and finally uses the Sigmoid activation function to generate the feature mask.

[0059] like Figure 5 As shown in the figure, the decoder has the same structure as the encoder. It reconstructs the signal features through the Inception_block and Skip connection, and generates the denoised GPR signal through the classification space and output layer.

[0060] The specific denoising process is as follows: 1) The encoder structure extracts features from the input data and compresses the input data to obtain an embedding vector In the data encoding phase, the input data is mapped to a low-dimensional vector space while learning the important features of the input data without changing the data shape. After being processed by the multi-layer Inception architecture, the data features are learned from shallow to deep, and the low-dimensional representation vector of the input data is finally obtained. The specific operation formula of the encoder structure is as follows: ; Where, represents the embedding vector, represents the encoder, represents the encoding operation in the encoder, Represents input data, Represents the encoder bias vector.

[0061] 2) The decomposer uses the Inception architecture to extract the contour features and multi-scale features of the GPR signal and noise, and reuses the input data features to reduce data feature loss. This improves the model's expressiveness and efficiency without increasing the number of parameters. At the end of the decomposer structure, the Sigmoid activation function is used to generate the data feature mask V. The formula is: ; Where: represents the decomposer, Represents the input data, b represents the bias vector in the decomposer, and S represents the Sigmoid classifier.

[0062] The Sigmoid function can convert the data feature vector of the network learning into a numerical output between 0 and 1. Due to its numerical output characteristics, the Sigmoid function is introduced into the network decomposer structure to generate a mask: .

[0063] 3) Decoder: In the classification space, the embedding vector is multiplied by the mask element to determine which information in the embedding vector should be retained or discarded. The classification space functions similarly to the forget gate in the LSTM structure. The decoder reconstructs the input embedding information into a GPR signal or noise based on the defined hyperparameter c, thereby removing noise from the GPR signal. The ID-Net classification space and decoder operation formulas are as follows: ; in, Describes the decoder, represents the embedding vector, V represents the feature mask, and c represents the network preset hyperparameters. Indicates the exclusive OR operation, represents the embedding vector after the OR operation with the feature mask, Represents the GPR signal extracted by the network.

[0064] In this embodiment, the noise data set refers to the fused signal data set with noise added, and the GPR signal data set refers to a pure fused noise data set.

[0065] In the preferred embodiment, in step S2, the model training step specifically includes: The GPR signal dataset is used as input to the signal denoising model, and the corresponding clean GPR signal and noise waveform are output respectively. The radar noise and radar signal characteristics are learned and distinguished through the structure of the network decomposer. At the end of the iterative training, the current optimal model parameters are saved and the best model is updated.

[0066] In the model training, the batch size is set to 100, the learning rate is 0.0004, and the number of epochs is 1000. The custom SignalMSE loss function is used. The SignalMSE loss function combines the signal variation and the mean square error. The formula is: ; Where, and are signal weight and MSE weight respectively, N is the number of data points, that is, the number of sampling points of the signal, and are the differences between adjacent sampling points of the target signal and the output signal, is the mean square error, and are the output value of the model and the true value respectively; Compared to the mean absolute error (MAE) and mean square error (MSE) loss functions, the SignalMSE loss function uses the signal's change to measure the difference between the network output and the true value. This is because the signal's change can be considered its derivative. Optimizing the model using derivative information can better preserve signal information such as edges and contours. The SignalMSE loss function, on the other hand, consists of two parts. By integrating these two parts, it balances noise removal and signal preservation, improving denoising and retaining more signal detail, thereby increasing model prediction accuracy.

[0067] During the model training process, this embodiment uses GPRMax software to construct three simulation signal data sets. The first two are noisy GPR signal data sets. These two data sets are used to output clean GPR signals and noise waveforms at the output end, respectively. The last one is a GPR signal data set. It is expected that a clean GPR signal will be output at the output end of the model training. This achieves the simultaneous learning of the characteristics of radar noise and radar signals, and distinguishes the characteristics through the structure of the network decomposer, thereby accelerating the learning speed of the model, making the model better adaptable to different types of data, improving the robustness and generalization ability of the model, and better coping with noise interference in practical applications.

[0068] In this embodiment, the three simulation signal data sets all contain 1950 data, and the data shape is During the training process on the network, the batch size is set to 100, the learning rate is 0.0004, the epoch is 1000, the loss function is set to SignalMSE, and the training environment is on a CPU with 128Gib, The system runs on a 64-bit Windows platform with a GPU. Python version is 3.6.0, and the PyTorch deep learning framework is version 1.10.2.

[0069] When the training process of 1000 epochs is completed, the current optimal model parameters are saved and the best model is updated.

[0070] In this embodiment, the measured GPR signal data set is input into the optimal parameter model for verification, and the number of data sampling points is 500.

[0071] 1) Sparse target signal denoising like Figure 7 As shown in the figure, a randomly selected A-Scan denoised data stream is visualized. The signal-to-noise ratio (SNR) of the noise signal is -1.43dB, indicating that the noise power is greater than the signal power. Under the influence of noise, the GPR signal has lost its waveform texture characteristics. The reflected echo of the electromagnetic wave from the underground object cannot be distinguished from the noise signal. However, after the denoising process of this embodiment, the original GPR signal waveform is essentially restored. The SNR of the GPR noise signal after denoising is calculated to be 7.28dB, and the cosine similarity with the original signal is 0.946.

[0072] 2) Dense target signal denoising like Figure 8 As shown, the signal-to-noise ratio of the noisy GPR signal is 0.66 dB. Its original waveform cannot be distinguished under the influence of noise. After processing by this embodiment, the influence of noise is greatly reduced, and its original signal waveform is restored. The signal-to-noise ratio of the denoised signal is improved to 9.03 dB, and the cosine similarity with the original signal is 0.936.

[0073] like Figure 9 As shown, the method of this embodiment has a better denoising effect: the noise signal-to-noise ratio is -1.57dB, and the original waveform cannot be distinguished. After denoising, the signal-to-noise ratio of this embodiment is improved to 8.52dB, and the signal-to-noise ratio is significantly improved. It can be seen intuitively that the waveform fidelity performance is significantly better than the other two comparison methods. The cosine similarity performance of the method of this embodiment, the EMD method and the VMD method are 0.943, 0.547 and 0.818 respectively.

[0074] As shown in Table 1, which compares the denoising performance of different algorithms on a dense target dataset, this embodiment outperforms the EMD and VMD algorithms in both SNR and cosine similarity. This embodiment improves the SNR by 57.1% compared to the EMD algorithm and by 108.6% compared to the VMD algorithm. In terms of cosine similarity, this embodiment improves the EMD algorithm by 15.8% and the VMD algorithm by 17.5%. Therefore, it can be concluded that this embodiment achieves superior denoising results on dense target datasets compared to the comparison methods.

[0075] Table 1 Comparison of denoising performance of different algorithms on dense target datasets

[0076] 3) Mixed signal denoising The mixed data set contains sparse target data and dense target data, with a total of 1,800 GPR signals.

[0077] like Figure 10 As shown, the noise signal , after being processed by the ID-Net denoising network, the GPR signal The signal-to-noise ratio is improved by 11.9dB. After network processing, not only the signal-to-noise ratio is improved, but the denoised GPR signal waveform is also preserved to the greatest extent. The cosine similarity with the original GPR signal is 0.971, which suppresses the terminal clutter and retains the original waveform of the signal, fully demonstrating the denoising effectiveness of this method.

[0078] like Figure 11 As shown in the figure, a noise signal is randomly selected and three signal denoising methods are used to attenuate the noise. Figure 11 The noise signal has a signal-to-noise ratio (SNR) of -1.4dB. The noise has completely obscured the original signal waveform, making it impossible to distinguish the object's reflected echo. After processing using the method in this embodiment, the noise is largely removed from the GPR signal, preserving the original GPR signal waveform. The SNR is improved by 7dB, and the cosine similarity is 0.934, very close to the original signal waveform.

[0079] As shown in Table 2, the embodiment method performs best in both SNR and cosine similarity. Therefore, it can be concluded that the method of the present invention has a better denoising effect on the mixed geological data set, which proves that the method of the present invention has denoising robustness when processing different geological data.

[0080] Table 2 Comparison of denoising performance of different algorithms on mixed geological datasets

[0081] In use, this embodiment uses an improved autoencoder structure network to construct a signal denoising model to achieve the purpose of denoising GPR signals: a noisy GPR signal is obtained from the input, a denoised signal is expected to be output at the output, and the SignalMSE loss function is used to optimize the network parameter training, thereby completing the signal denoising task while retaining more edge features and texture features of the GPR signal. On the other hand, the signal denoising model can fully learn the multi-scale features of the GPR signal through the Inception structure, fuse and reduce the dimensionality of the features, convert the relationships into model parameters and store them, thereby improving the model's predictive performance. Visualization experiments and comparative experiments have shown that this embodiment can accurately restore the GPR signal waveform in noisy environments with different signal-to-noise ratios, improving the robustness of processing signals with different signal-to-noise ratios.

[0082] Example 2 In combination with Example 1, a GPR signal denoising system based on a multi-scale feature fusion network is provided, including the following modules: The data acquisition and processing module is used to obtain sparse target signal data and dense target signal data, and after pre-processing, fuse them according to a preset ratio and save the fused data into the GPR signal data set; The model building module is used to build a signal denoising model based on the autoencoder network, input the GPR signal dataset and perform model training, and save the optimal signal denoising model; The signal denoising model includes an encoder, a decomposer, and a decoder. The autoencoder network is improved based on the Inception_block, and the Inception architecture is introduced as a basic component. A decomposer structure is added between the encoder and decoder. The convolution in the Inception architecture is a one-dimensional convolution. The decoder has the same structure as the encoder, consisting of four Inception_blocks followed by a 1×1 convolution. A skip connection is introduced between the first and fourth Inception_blocks. The decomposer structure is based on the encoder structure, and then connected to the Sigmoid activation function module. The test module is used to input the GPR signal data set to be tested into the optimal signal denoising model to obtain the model test results.

[0083] In a preferred embodiment, a computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the GPR signal denoising method based on a multi-scale feature fusion network in Example 1.

[0084] In a preferred embodiment, an electronic device includes a memory and a processor; Memory for storing computer programs.

[0085] The processor is configured to implement the steps of the GPR signal denoising method based on a multi-scale feature fusion network in embodiment 1 when executing the computer program.

[0086] The above embodiments are merely preferred technical solutions of the present invention and should not be construed as limiting the present invention. The scope of protection of the present invention shall be the technical solutions set forth in the claims, including equivalent alternatives to the technical features of the technical solutions set forth in the claims. In other words, equivalent alternatives and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A GPR signal denoising method based on a multi-scale feature fusion network, characterized in that: The following steps are involved: S1: Acquire sparse target signal data and dense target signal data, pre-process them, fuse them according to a preset ratio, and save the fused data into a GPR signal dataset; S2: Build a signal denoising model based on the autoencoder network, input the GPR signal dataset and perform model training, and save the optimal signal denoising model; The signal denoising model includes an encoder, a decomposer, and a decoder. The autoencoder network is improved based on the Inception_block, and the Inception architecture is introduced as a basic component. A decomposer structure is added between the encoder and decoder. The convolution in the Inception architecture is a one-dimensional convolution. The decoder has the same structure as the encoder, consisting of four Inception_blocks followed by a 1×1 convolution. A skip connection is introduced between the first and fourth Inception_blocks. The decomposer structure is based on the encoder structure, and then connected to the Sigmoid activation function module. S3: Input the GPR signal data set to be tested into the optimal signal denoising model to obtain the model test results.

2. The GPR signal denoising method based on a multi-scale feature fusion network according to claim 1 is characterized in that: In step S2, the encoder extracts features from the input data and compresses the representation to obtain an embedded vector; the decomposer uses the Inception architecture to extract the contour features and multi-scale features of the GPR signal and noise, reuses the input data features, and uses the Sigmoid activation function to generate a feature mask; The decoder reconstructs the denoised GPR signal based on the embedding vector and feature mask, and introduces Skip connection to retain shallow features.

3. The GPR signal denoising method based on a multi-scale feature fusion network according to claim 1, characterized in that: It also includes S4: using signal-to-noise ratio and cosine similarity to evaluate the model test results, specifically: S41: The signal-to-noise ratio (SNR) is used to evaluate the denoising effect of the model. The formula is: ； Where, For a clean GPR signal, is the GPR signal after denoising, N is the number of GPR signal sampling points; S42: Cosine similarity is used to evaluate the denoising fidelity performance. The formula for cosine similarity is as follows: ； Where, For a clean GPR signal, is the GPR signal after denoising, and N is the number of sampling points of the GPR signal.

4. The GPR signal denoising method based on a multi-scale feature fusion network according to claim 1, characterized in that: S1: Acquire sparse target signals and dense target signals. Both sparse target signal data and dense target signal data are collected by radar on clay geology, and the radar antenna frequency is set to 400MHz. Preprocessing specifically includes: S11: Perform multimodal characteristic analysis. Based on the time domain characteristics of the signal, a clustering algorithm is used to perform modal classification on the collected raw GPR data. The formula is: ； Where, is the original GPR data, and its time domain feature vector Including characteristics such as amplitude and frequency; Use K-Means algorithm to cluster the feature vectors: ； Where, is the i-th cluster center, k is the number of clusters, is the new cluster center; S12: Adaptive noise estimation is then performed, combining wavelet transform and empirical mode decomposition (EMD) technology to decompose the signal and estimate the noise level; ； Where, is the scale parameter, is the translation parameter, is the wavelet basis function; The signal Decomposed into several IMFs, the formula is: ； Where, is the i-th IMF, is the residual; If the high-frequency IMF mainly contains noise, the noise level can be estimated by calculating the energy of the high-frequency IMF: ； Where, is the estimated noise variance, is the number of high-frequency intrinsic mode functions (IMFs), is the number of sampling points of the signal; S13: Based on feature matching, the interference signal is removed according to the signal similarity; S14: Finally, dynamic time zero point correction is performed; S15: Normalize the preprocessed GPR data using the Min-Max normalization method to scale the data to the [0, 1] interval. The formula is: ； Where, is the value of the normalized signal at time point t, and Signals The minimum and maximum values ​​in ; S16: Divide the normalized data into a training set, a validation set, and a test set according to a preset ratio.

5. The GPR signal denoising method based on a multi-scale feature fusion network according to claim 1, characterized in that: In S2, the model training steps specifically include: The signal denoising model uses two GPR signal datasets with different noise levels and one original GPR signal dataset as input. The model outputs the corresponding clean GPR signal and noise waveform, respectively. The radar noise and radar signal characteristics are learned and distinguished through the structure of the network decomposer. At the end of the iterative training, the current optimal model parameters are saved and the best model is updated. In the model training, the batch size is set to 100, the learning rate is 0.0004, and the number of epochs is 1000. The custom SignalMSE loss function is used. The SignalMSE loss function combines the signal variation and the mean square error. The formula is: ； Where, and are signal weight and MSE weight respectively, N is the number of data points, that is, the number of sampling points of the signal, and are the differences between adjacent sampling points of the target signal and the output signal, is the mean square error, and are the output value of the model and the true value respectively.

6. The GPR signal denoising method based on a multi-scale feature fusion network according to claim 1, characterized in that: S2: In the signal denoising model based on the autoencoder network, each Inception_block extracts features of different scales through successively increasing convolution kernels and splices the feature maps to improve the network performance and efficiency. The Inception module uses the Relu nonlinear activation function, which is expressed as: ； Where, is the input value, is the output value of the ReLu function.

7. The GPR signal denoising method based on a multi-scale feature fusion network according to claim 1, characterized in that: In S1, the weighted fusion method based on energy matching is used to fuse the data according to the preset ratio: The energy of the sparse target signal and the dense target signal after preprocessing is calculated respectively. The energy of the signal can be calculated by integrating the square of its amplitude. For discrete signals, the formula is: ； Where, is the value of the signal at the i-th sampling point, is the number of sampling points; According to the calculated energy value, the weighting coefficient of the two signals is determined; the weighting coefficient is dynamically adjusted according to the energy ratio of the signals. The formula is: ； Where, is the energy of the sparse target signal, is the energy of the dense target signal, and is the weighting coefficient; The smoothing factor γ is introduced to smooth the weighted coefficients: ； Where, and is the weighting coefficient after smoothing; and is the weighted coefficient calculated last time; γ is the smoothing factor; The pre-processed sparse target signal and dense target signal are fused according to the determined weight coefficient to obtain a mixed signal. The formula is: ； Where, is the value of the fused mixed signal at the i-th sampling point, is a sparse target signal, It is a dense target signal.

8. A GPR signal denoising system based on a multi-scale feature fusion network, characterized in that: Includes the following modules: The data acquisition and processing module is used to obtain sparse target signal data and dense target signal data, and after pre-processing, fuse them according to a preset ratio and save the fused data into the GPR signal data set; The model building module is used to build a signal denoising model based on the autoencoder network, input the GPR signal dataset and perform model training, and save the optimal signal denoising model; The signal denoising model includes an encoder, a decomposer, and a decoder. The autoencoder network is improved based on the Inception_block, and the Inception architecture is introduced as a basic component. A decomposer structure is added between the encoder and decoder. The convolution in the Inception architecture is a one-dimensional convolution. The decoder has the same structure as the encoder, consisting of four Inception_blocks followed by a 1×1 convolution. A skip connection is introduced between the first and fourth Inception_blocks. The decomposer structure is based on the encoder structure, and then connected to the Sigmoid activation function module. The test module is used to input the GPR signal data set to be tested into the optimal signal denoising model to obtain the model test results.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the GPR signal denoising method based on a multi-scale feature fusion network according to any one of claims 1 to 7 are implemented.

10. An electronic device, characterized in that: including memory and processor; The memory is used to store computer programs; The processor is configured to implement the steps of the GPR signal denoising method based on a multi-scale feature fusion network according to any one of claims 1 to 7 when executing the computer program.

Images