Self-learning dual-wave fusion transient electromagnetic imaging method for resistivity anomaly target

Abstract

The application discloses a self-learning dual-wave fusion transient electromagnetic imaging method for resistivity anomaly targets, and comprises the following steps: constructing a dual-channel parallel convolutional neural network, respectively extracting transient electromagnetic response features of a narrow pulse width triangular wave and a wide pulse width trapezoidal wave, fusing dual-waveform features through a multi-head attention, and preliminarily predicting the resistivity distribution of an underground structure; reconstructing the predicted resistivity by using a domain perception variational autoencoder, and judging whether a test sample deviates from a training data distribution through a reconstruction error; inputting the predicted resistivity of an outlying sample into a physical forward model to generate an electromagnetic response, and taking a residual error between the predicted response and a measured response as a physical loss to guide online fine-tuning of a fusion inversion network, so that self-learning imaging with a domain adaptation capability is realized; and outputting a fine imaging structure of a resistivity anomaly target. The application realizes dual-waveform feature level complementary fusion, unsupervised domain drift detection and physical guided online self-learning, and significantly improves full-depth imaging resolution, generalization capability and detection reliability.

Description

Technical Field

[0001] This invention relates to the field of geophysical transient electromagnetic detection and inversion imaging technology, specifically to a self-learning dual-wave fusion transient electromagnetic imaging method for resistivity anomaly targets. Background Technology

[0002] Subsurface spatial structure imaging has significant application value in non-destructive testing of underground cavities, maintenance of urban underground pipe networks, and subway line planning. Therefore, rapid, reliable, and non-invasive geophysical imaging technology is particularly crucial. The small-loop transient electromagnetic method based on deep learning has been widely applied to imaging near-surface resistivity anomalies due to its low cost and large detection depth. Currently, conventional inversion methods typically use the electromagnetic response of a single transmitted waveform as network input, directly predicting subsurface resistivity distribution using architectures such as convolutional neural networks and long short-term memory networks. For example, Wu et al. (S. Wu, Q. Huang, and L. Zhao, “Convolutional neural network inversion of airborne transient electromagnetic data,” Geophys. Prospecting, vol. 69, nos. 8-9, pp.1761-1772, Aug. 2021) proposed a convolutional neural network-based inversion method for airborne transient electromagnetic data, directly mapping the electromagnetic response to a resistivity model using deep learning. However, conventional transient electromagnetic systems generally use wide-pulse trapezoidal wave excitation currents, which are rich in low-frequency components and beneficial for deep target detection, but lack high-frequency components, resulting in insufficient resolution for complex shallow structures. Narrow-pulse triangular waves, while rich in high-frequency components and sensitive to shallow targets, have limited detection capabilities for deep targets. This physical limitation of single-waveform excitation restricts fine imaging across the entire depth range. To overcome the limitations of single-waveform, multi-waveform fusion methods have been proposed. Li et al. (D. Li, Y. Ji, Y. Yu, and S. Wang, “Fine TEM detection system for urban underground space based on time-domaintrapezoidal-semisinusoidal combined current waveforms,” IEEE Trans. Instrum.Meas., vol. 74, pp. 1-15, Feb. 2025) proposed a small-loop transient electromagnetic detection system using a trapezoidal-semisinusoidal combined waveform, which weights and superimposes the inversion results of different waveforms.Wang et al. (S. Wang, Y. Wang, Y. Yu, H. Luan, Y. Wang, and Y. Ji, “Dual-waveform combination of electromagnetic detection system based on small-loop TEM method for shallow refinement,” IEEE Trans. Instrum. Meas., vol. 73, pp. 1-14, Nov. 2024) studied a dual-waveform combination detection method based on small-loop transient electromagnetics, which improves shallow refinement capability by weighted fusion of inversion results from different waveforms. However, the above method only performs simple fusion at the level of inversion results and fails to achieve deep integration of the electromagnetic properties of different waveforms in the feature extraction stage. This fails to effectively alleviate the non-uniqueness problem of transient electromagnetic inversion, resulting in limited improvement in inversion accuracy.

[0003] On the other hand, to enhance the physical consistency of data-driven methods, physical information neural networks have been introduced into the field of geophysical inversion. Liu et al. (B. Liu et al., “Physics-driven deep learning inversion for direct current resistivity survey data,” IEEE Trans. Geosci. Remote Sens., vol. 61, Apr. 2023, Art. no. 5906611) proposed a physics-driven deep learning inversion method, which embeds physical constraints during training by jointly constructing a loss function from data errors and physical equation residuals. Wu et al. (S.Wu, Q. Huang, and L. Zhao, “Physics-guided deep learning-based inversion for airborne electromagnetic data,” Geophys. J. Int., vol. 238, no. 3, pp. 1774-1789, Jul. 2024) further applied physics-guided deep learning inversion to airborne electromagnetic data. However, the aforementioned physical information methods all employ offline training, requiring iterative numerical calculations (such as Hankel transform) on a large number of training samples to obtain physical loss, resulting in extremely high computational costs. More importantly, the neural networks trained offline are difficult to change, and when faced with field measurement data that differs significantly from the training data distribution, they lack adaptive learning capabilities, making it difficult to achieve accurate detection of targets in real and complex geological environments.

[0004] Due to the highly nonlinear and ill-conditioned nature of transient electromagnetic inversion problems, existing methods generally suffer from the following shortcomings: (1) Single-waveform excitation is difficult to simultaneously meet the requirements of high resolution in shallow areas and large detection depth in deep areas; (2) Existing multi-waveform fusion methods lack deep feature-level integration of complementary electromagnetic response characteristics, resulting in insufficient multi-solution suppression capabilities; (3) Most existing physical information methods adopt offline training modes, which cannot adaptively adjust for outlier samples in actual measurements, thus limiting generalization ability and robustness. Therefore, there is an urgent need for an adaptive transient electromagnetic imaging method that can deeply fuse complementary multi-waveform information at the feature level and possess online physical self-learning capabilities. Summary of the Invention

[0005] The purpose of this invention is to provide a self-learning dual-wave fusion transient electromagnetic imaging method to solve the problems of existing transient electromagnetic deep learning inversion methods, such as strong dependence on training data distribution, poor generalization ability under complex geological conditions, and insufficient single waveform resolution.

[0006] To achieve the above objectives, the technical solution proposed by this invention is as follows: A self-learning dual-wave fusion transient electromagnetic imaging method for resistivity anomaly targets is characterized by the following steps: Step 1: Acquisition and Preprocessing of Dual-Waveform Transient Electromagnetic Data A transient electromagnetic detection system is used to deploy measurement points in the area to be measured. First waveform current and second waveform current are emitted respectively, and the corresponding dual-waveform transient electromagnetic secondary field response data for each measurement point are collected and preprocessed with normalization. The first waveform is a narrow-pulse-width triangular wave with a relatively uniform spectrum distribution over a wide frequency range, rich in high-frequency components, used to enhance shallow-depth resolution; the second waveform is a wide-pulse-width trapezoidal wave with energy concentrated in the low-frequency band, used to improve deep-depth detection capabilities. Preferably, after equal logarithmic interval sampling of the collected dual-waveform transient electromagnetic secondary field response data, maximum-minimum normalization is performed based on the maximum and minimum values ​​of each time channel in the training dataset.

[0007] Step 2: Construct a multi-head attention dual-wave fusion inversion neural network: Construct a multi-head attention dual-wave fusion inversion neural network, the network comprising: The first feature extraction channel is used to extract the electromagnetic features of the first waveform response; The second feature extraction channel is used to extract the electromagnetic features of the second waveform response; The multi-head attention fusion module is used to adaptively weight and fuse features extracted from dual channels; The resistivity prediction module is used to output a predicted resistivity model based on the fused features.

[0008] Preferably, both the first and second feature extraction channels employ a convolutional neural network structure, with each channel consisting of multiple convolutional layers and fully connected layers connected in series. The calculation process of the multi-head attention fusion module is as follows: first, the feature matrices extracted from the two channels are linearly transformed to obtain the query matrix Q, the key matrix K, and the value matrix V; then, the attention weights are calculated by scaling the dot product; finally, the outputs of multiple self-attention heads are concatenated and multiplied by the weight matrix to obtain the fused features.

[0009] Step 3: Joint offline pre-training A training dataset was constructed using different resistivity models generated through simulation and their corresponding dual-waveform transient electromagnetic response data. This dataset was then used to jointly pre-train the multi-head attention dual-waveform fusion inversion neural network and the domain-aware variational autoencoder (VADE) offline. The VDE took the predicted resistivity model output by the multi-head attention dual-waveform fusion inversion neural network as input and learned the distribution characteristics of the resistivity model in the training data. The total loss function for joint pre-training included the inversion loss and the variational autoencoder loss.

[0010] Preferably, the domain-aware variational autoencoder consists of an encoder and a decoder. The encoder outputs the mean and variance of the predicted resistivity latent distribution and obtains the latent variables through reparameterized sampling. The decoder reconstructs the resistivity based on the latent variables. The loss function of the variational autoencoder is the weighted sum of the reconstruction error and the KL divergence between the latent variable distribution and the prior distribution.

[0011] The total loss function for the joint offline pre-training is expressed as: in, To predict the mean square error between resistivity and true resistivity, For variational autoencoder loss, These are the weighting coefficients.

[0012] Step 4: Preliminary Inversion and Domain Determination The actual collected and preprocessed dual-wavelength transient electromagnetic response data is input into the pre-trained multi-head attention dual-wavelength fusion inversion neural network to obtain a preliminary resistivity model. This preliminary resistivity model is then input into the pre-trained domain-aware variational autoencoder for reconstruction, and the reconstruction error is calculated. Step 5: Outlier Identification Based on the reconstruction error, it is determined whether the current test sample is an outlier. When the reconstruction error exceeds a preset threshold, the sample is determined to be an outlier sample whose distribution is inconsistent with the training data.

[0013] Preferably, the preset threshold is the average reconstruction error of all samples in the training dataset plus twice the standard deviation, or it is determined by a certain quantile of the reconstruction error of the training dataset.

[0014] Step Six: Construct a Physical Information-Guided Self-Learning Loss Function For measurement points identified as outliers in step five, the preliminary resistivity model obtained in step four is input into the transient electromagnetic physics forward model to calculate the corresponding predicted electromagnetic response. The physical loss function is constructed using the residual between the predicted electromagnetic response and the actual measured electromagnetic response, and a network parameter regularization term is added to form a self-learning total loss function.

[0015] Preferably, the transient electromagnetic physics forward model is a one-dimensional, two-dimensional, or three-dimensional time-domain electromagnetic field numerical calculation model. This model does not contain learnable parameters, the input is a resistivity model, and the output is the time-domain electromagnetic response under the corresponding first and second waveform excitations.

[0016] More preferably, the physical loss function is the mean square error between the measured electromagnetic response and the physical forward modeling predicted electromagnetic response, and its expression is: in For the first i The sample at the th j Measured electromagnetic response over time. The electromagnetic response is obtained by physical forward modeling from the predicted resistivity, where N is the number of samples and D is the number of time channels.

[0017] The self-learning total loss function is a weighted sum of the physical loss function and the L2 regularization term of the network parameters, and its expression is: in For the L2 regularization term of the network weights, is the regularization coefficient.

[0018] Step Seven: Online Self-Learning and Fine-Tuning The multi-head attention dual-wave fusion inversion neural network is fine-tuned online using the self-learning total loss function to update the network parameters and obtain an adaptive resistivity model that adapts to the distribution of outlier samples.

[0019] Preferably, the online fine-tuning adopts the Adam optimization algorithm, with a learning rate lower than that of the offline pre-training stage and a batch size less than or equal to that of the offline pre-training stage. During the fine-tuning process, only the measurement point data identified as outliers in step five or the small batch data containing such samples are iteratively updated until the physical loss function converges or the preset maximum number of iterations is reached.

[0020] Step 8: Output the final resistivity imaging results: For measuring points identified as non-outliers in step five, the preliminary resistivity model obtained in step four is directly used as the final resistivity model. For measuring points identified as outliers in step five, the adaptive resistivity model obtained after fine-tuning in step seven is used as the final resistivity model. The final resistivity models of all measuring points are then fitted by interpolation to obtain the spatial distribution and resistivity image of underground resistivity anomalies. Preferably, the interpolation fitting uses cubic spline interpolation or Kriging interpolation methods to map the one-dimensional depth-resistivity sequence of each measuring point to a two-dimensional or three-dimensional grid, generating a resistivity distribution profile or stereoscopic image for identifying high-resistivity cavities, low-resistivity aquifers, or resistivity anomalies such as metal pipe networks.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) By designing a dual-channel parallel neural network and combining it with a multi-head attention mechanism, the transient electromagnetic response of narrow pulse width triangular wave and wide pulse width trapezoidal wave was deeply integrated at the feature level. The complementary advantages of the two waveforms in terms of shallow high resolution and deep detection depth were fully explored, and the resolution and accuracy of resistivity imaging were significantly improved.

[0022] (2) Introducing a domain-aware variational autoencoder and using reconstruction error as a domain discrimination index can effectively identify actual samples that are inconsistent with the distribution of training data. Outlier detection can be achieved without real labels, enhancing the applicability to real complex environments.

[0023] (3) A self-learning module guided by physical information is proposed. The network parameters are finely adjusted online using forward physical constraints for outlier samples, so that the electromagnetic response of the prediction model gradually approaches the measured data. This significantly reduces the inversion error caused by domain offset, avoids false anomalies, and improves imaging accuracy and environmental adaptability. Attached Figure Description

[0024] Figure 1 This is a flowchart of a self-learning dual-wave fusion transient electromagnetic imaging method for resistivity anomaly targets.

[0025] Figure 2 This is a schematic diagram of a multi-head attention dual-wave fusion inversion neural network and a domain-aware variational autoencoder.

[0026] Figure 3 This is a schematic diagram of the workflow of the physical information-guided self-learning module. Detailed Implementation

[0027] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the following embodiments are only for explaining the present invention and do not constitute any limitation on the scope of protection of the present invention. All equivalent transformations or modifications made based on the concept of the present invention fall within the scope of protection of the present invention.

[0028] Example 1: Self-learning dual-wavelength fusion transient electromagnetic imaging for underground cavity detection This embodiment uses the detection of underground cavities in urban roads as an application scenario, and employs the method of the present invention to perform fine imaging of possible cavity anomalies under a municipal road.

[0029] The main steps include the following: Step 1: Data Acquisition and Preprocessing.

[0030] Several measuring points are arranged along the measurement line in the area to be measured. A transient electromagnetic detection system is used to sequentially transmit a narrow-pulse triangular current and a wide-pulse trapezoidal current at each measuring point. The transmitting coil can be a multi-turn square coil, and the receiving coil can be a circular coil with a large equivalent receiving area. Each waveform is transmitted multiple times and then superimposed and averaged to improve the signal-to-noise ratio. The induced voltage data after the transmitting current is turned off is extracted, and multiple time channels are extracted at equal logarithmic intervals. The extracted data is then normalized to maximum and minimum values ​​of each time channel in the training dataset using a max-min normalization method.

[0031] Step 2: Construct a multi-head attention dual-wave fusion inversion neural network (MDFI-NN).

[0032] The network comprises two structurally identical convolutional branches, processing triangular and trapezoidal wave responses respectively. Each branch consists of three convolutional layers and two fully connected layers cascaded together, with progressively increasing kernel size and number. Each convolutional layer is followed by max pooling and ReLU activation. The features extracted from the two branches are concatenated and fed into a multi-head attention module. Adaptive weights are calculated using the scaled dot product of query, key, and value to achieve deep fusion of the two wave features. Finally, the predicted resistivity value is output through a fully connected layer.

[0033] Step 3: Offline pre-training.

[0034] A source domain simulation dataset containing tens of thousands of one-dimensional layered resistivity models and their corresponding two-wave transient electromagnetic responses was constructed. Simultaneously, a domain-aware variational autoencoder (DA-VAE) was built. The encoder takes the predicted resistivity from a multi-head attention two-wave fusion inversion neural network as input and outputs the mean and variance of the latent variable distribution. The decoder reconstructs the resistivity from the latent variables. The pre-training total loss function is... ,in This represents the mean square error between the predicted and actual resistivity values. The sum of the reconstruction error and the KL divergence. 10 can be taken - ². The Adam optimizer was used for training, with an initial learning rate set to 10. - ³, set the batch size to 256, until the loss converges.

[0035] Step 4: Domain discrimination.

[0036] Measured and preprocessed triangular and trapezoidal wave data are input into a pre-trained multi-head attention dual-wave fusion inversion neural network to obtain a preliminary resistivity model. This model is then input into a domain-aware variational autoencoder for reconstruction, and the reconstruction error (mean square error between predicted and reconstructed resistivity) for each sample is calculated. The average reconstruction error of the training dataset is used as a threshold, and samples with reconstruction errors greater than the threshold are marked as outliers.

[0037] Step 5: Physical information guides self-learning.

[0038] For outlier samples, the preliminary resistivity model obtained in step four is fed into a transient electromagnetic forward modeling program to calculate the corresponding triangular and trapezoidal electromagnetic responses. Physical loss function. The mean square error of the measured response and the forward response is used as the regularization term. Let be the sum of squared L2 norms of all learnable parameters of the network. The total self-learning loss is... , 10 can be taken -4 Keeping the network structure unchanged, the MDFI-NN was iteratively fine-tuned using the Adam optimizer, with optimizations made to both the learning rate and batch size. After fine-tuning, updated resistivity predictions were obtained.

[0039] Step 6: Output the final imaging results.

[0040] The above process is performed on all measuring points. The obtained one-dimensional resistivity model is then interpolated to obtain a two-dimensional resistivity profile, which is used to identify the structure and spatial distribution of abnormal targets such as underground high-resistivity cavities, low-resistivity aquifers, and underground metal pipe networks.

[0041] The above specific embodiments are merely illustrative examples, and the scope of protection of this invention is defined by the claims. All equivalent changes or modifications made according to the spirit and essence of this invention are within the scope of protection of this invention.

Claims

1. A self-learning dual-wavelength fusion transient electromagnetic imaging method for resistivity anomaly targets, characterized in that, Includes the following steps: Step 1: Using a transient electromagnetic detection system, set up measurement points in the area to be measured, transmit the first waveform current and the second waveform current respectively, collect the corresponding dual-waveform transient electromagnetic secondary field response data of each measurement point, and perform normalization preprocessing; the first waveform is a narrow pulse width triangular wave, and the second waveform is a wide pulse width trapezoidal wave. Step 2: Construct a multi-head attention dual-wave fusion inversion neural network, the network comprising: a first feature extraction channel for extracting electromagnetic features of the first waveform response; a second feature extraction channel for extracting electromagnetic features of the second waveform response; a multi-head attention fusion module for adaptively weighting and fusing the features extracted by the two channels; and a resistivity prediction module for outputting a predicted resistivity model based on the fused features. Step 3: Construct a training dataset using different resistivity models generated by simulation and their corresponding dual-waveform transient electromagnetic response data, and perform joint offline pre-training on the multi-head attention dual-wave fusion inversion neural network and the domain-aware variational autoencoder; wherein, the domain-aware variational autoencoder takes the predicted resistivity model output by the multi-head attention dual-wave fusion inversion neural network as input and learns the distribution characteristics of the resistivity model in the training data; the total loss function of the joint pre-training includes the inversion loss and the variational autoencoder loss; Step 4: Input the actual collected and preprocessed dual-waveform transient electromagnetic response data into the pre-trained multi-head attention dual-wave fusion inversion neural network to obtain a preliminary resistivity model. Then, input the preliminary resistivity model into the pre-trained domain-aware variational autoencoder for reconstruction and calculate the reconstruction error. Step 5: Determine whether the current test sample is an outlier based on the reconstruction error. When the reconstruction error exceeds a preset threshold, the sample is determined to be an outlier that is inconsistent with the distribution of the training data. Step 6: For the measurement points identified as outliers in Step 5, input the preliminary resistivity model obtained in Step 4 into the transient electromagnetic physics forward model, calculate the corresponding predicted electromagnetic response, and construct the physical loss function based on the residual between the predicted electromagnetic response and the actual measured electromagnetic response. At the same time, add the network parameter regularization term to form the self-learning total loss function. Step 7: Use the self-learning total loss function to fine-tune the multi-head attention dual-wave fusion inversion neural network online, update the network parameters, and obtain an adaptive resistivity model that adapts to the distribution of outlier samples. Step 8: Output the final resistivity imaging results to obtain the spatial distribution and resistivity distribution of underground resistivity anomalies.

2. The self-learning dual-wave fusion transient electromagnetic imaging method as described in claim 1, characterized in that: In step one, the spectrum of the narrow pulse width triangular wave is relatively uniformly distributed in the range of 0-5000 Hz and is rich in high-frequency components; the energy of the wide pulse width trapezoidal wave is concentrated in the low-frequency range of 0-1000 Hz; after the collected dual-waveform transient electromagnetic secondary field response data is sampled at equal logarithmic intervals, the maximum-minimum normalization process is performed according to the maximum and minimum values ​​of each time channel in the training dataset.

3. The self-learning dual-wave fusion transient electromagnetic imaging method as described in claim 1, characterized in that: In step two, both the first and second feature extraction channels adopt a convolutional neural network structure, with each channel consisting of multiple convolutional layers and fully connected layers connected in series. The calculation process of the multi-head attention dual-wave fusion module is as follows: first, the feature matrices extracted from the dual channels are linearly transformed to obtain the query matrix Q, the key matrix K, and the value matrix V; then, the attention weights are calculated by scaling the dot product; finally, the outputs of multiple self-attention heads are concatenated and multiplied by the weight matrix to obtain the fused features.

4. The self-learning dual-wave fusion transient electromagnetic imaging method as described in claim 1, characterized in that: In step three, the domain-aware variational autoencoder consists of an encoder and a decoder. The encoder outputs the mean and variance of the predicted resistivity latent distribution and obtains latent variables through reparameterized sampling. The decoder reconstructs the resistivity based on the latent variables. The loss function of the variational autoencoder is the weighted sum of the reconstruction error and the KL divergence between the latent variable distribution and the prior distribution.

5. The self-learning dual-wave fusion transient electromagnetic imaging method as described in claim 4, characterized in that: In step three, the total loss function for the joint offline pre-training is: in, To predict the mean square error between resistivity and true resistivity, For variational autoencoder loss, To balance the weighting coefficients of the two.

6. The self-learning dual-wave fusion transient electromagnetic imaging method as described in claim 1, characterized in that: In step five, the preset threshold is the average reconstruction error of all samples in the training dataset plus twice the standard deviation, or it is determined by a certain quantile of the reconstruction error of the training dataset.

7. The self-learning dual-wave fusion transient electromagnetic imaging method as described in claim 1, characterized in that: In step six, the transient electromagnetic physics forward model is a one-dimensional, two-dimensional or three-dimensional time-domain electromagnetic field numerical calculation model. This model does not contain learnable parameters, the input is a resistivity model, and the output is the time-domain electromagnetic response under the corresponding first and second waveform excitations. The physical loss function is the mean square error between the measured electromagnetic response and the physical forward modeling predicted electromagnetic response, and its expression is: in, For the first i The sample at the th j Measured electromagnetic response over time. The electromagnetic response is obtained by physical forward modeling from the predicted resistivity, where N is the number of samples and D is the number of time channels.

8. The self-learning dual-wave fusion transient electromagnetic imaging method as described in claim 1, characterized in that: In step six, the self-learning total loss function is a weighted sum of the physical loss function and the L2 regularization term of the network parameters, and its expression is: in, For the L2 regularization term of the network weights, is the regularization coefficient.

9. The self-learning dual-wave fusion transient electromagnetic imaging method as described in claim 1, characterized in that: In step seven, the online fine-tuning adopts the Adam optimization algorithm, with a learning rate lower than that of the offline pre-training stage and a batch size less than or equal to that of the offline pre-training stage. During the fine-tuning process, only the measurement point data identified as outliers in step five or the small batch data containing such samples are iteratively updated until the physical loss function converges or the preset maximum number of iterations is reached.

10. The self-learning dual-wave fusion transient electromagnetic imaging method according to claim 1, characterized in that: In step eight, the interpolation fitting uses cubic spline interpolation or kriging interpolation to map the one-dimensional depth-resistivity sequence of each measuring point to a two-dimensional or three-dimensional grid, generating a resistivity distribution profile or stereoscopic image, which is used to identify resistivity anomalies such as high-resistivity cavities, low-resistivity aquifers, or metal pipe networks.

Images