Ground-air transient electromagnetic data rapid inversion method based on long and short term memory network
By constructing a CNN-BiLSTM-Attention hybrid neural network model, the problems of low efficiency and insufficient accuracy of traditional ground-to-space transient electromagnetic inversion methods are solved, and efficient and accurate inversion of underground medium parameters is achieved, which is applicable to mining subsidence areas, aquifers and mineral exploration.
Patent Information
- Application Number
- CN202511981188.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-25
- Publication Date
- 2026-05-12
AI Technical Summary
Traditional ground-to-air transient electromagnetic inversion methods are computationally inefficient, rely on human experience, and are prone to getting trapped in local optima, making it difficult to meet the needs of real-time and engineering-oriented detection. Existing fully connected neural networks have weak feature extraction capabilities, rely on insufficient modeling in the long term, and have limited inversion accuracy and generalization ability.
A hybrid neural network model of CNN-BiLSTM-Attention is adopted. The model combines a convolutional neural network (CNN) to extract local spatial features of transient electromagnetic signals, a bidirectional long short-term memory network (BiLSTM) to model long-term temporal dependencies, and an attention mechanism to dynamically focus key information. Finally, a Seq2Seq decoder is used to generate formation parameters.
It achieves second-level inversion of transient electromagnetic data between the ground and the air, improving the inversion speed and accuracy. It has anti-interference capabilities, can accurately reflect the underground electrical structure and stratigraphic details, and is suitable for automated inversion in various geological scenarios.
Smart Images
Figure CN122018021A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of geophysical exploration technology, specifically to a method for rapid inversion of ground-space transient electromagnetic data based on long short-term memory networks. Background Technology
[0002] The Shaanxi-Gansu-Ningxia region has a wide distribution of mining subsidence areas, posing a significant risk of geological disasters. High-precision and high-efficiency geophysical exploration methods are urgently needed for effective disaster prevention and mitigation. The ground-to-air transient electromagnetic method, combining the advantages of ground transmission and air reception, boasts advantages such as large detection depth, high resolution, and strong terrain adaptability, making it an effective method for detailed exploration of mining subsidence areas in complex environments. However, this method involves large amounts of data and complex temporal characteristics. Traditional inversion methods rely on initial model selection and iterative optimization, resulting in low computational efficiency, long interpretation cycles, and a tendency to get trapped in local optima, making it difficult to meet the needs of real-time and engineering-scale exploration.
[0003] Existing inversion methods are mostly based on regularization or linear approximation theory, which are sensitive to noise and rely on human experience, especially when dealing with multi-point, large-scale data, resulting in low efficiency. In recent years, although some studies have attempted to use fully connected neural networks for electromagnetic inversion, they generally suffer from weak feature extraction capabilities and long-term reliance on insufficient modeling, resulting in limited inversion accuracy and generalization ability. Therefore, there is an urgent need for an intelligent inversion method that can balance efficiency and accuracy to achieve rapid and reliable interpretation of ground-to-air transient electromagnetic data. Summary of the Invention
[0004] To address the problems mentioned in the background section, this invention provides the following technical solution: a method for rapid inversion of ground-to-space transient electromagnetic data based on long short-term memory networks, comprising the following steps:
[0005] Construct a training dataset for ground-to-air transient electromagnetic events, including generating resistivity and thickness parameters through a stochastic layered model and generating corresponding transient electromagnetic response data based on forward modeling.
[0006] The training dataset is preprocessed, including normalization, temporal reshaping, and dataset partitioning;
[0007] Construct a CNN-BiLSTM-Attention hybrid neural network model, which includes a CNN feature extraction module, a BiLSTM temporal modeling module, an attention mechanism module, and a Seq2Seq decoder module connected in sequence.
[0008] The hybrid neural network model is trained using the training dataset until the loss function converges;
[0009] The ground-to-air transient electromagnetic observation data to be inverted is input into the trained model, and the resistivity and thickness parameters of the underground medium are output.
[0010] Preferably, the CNN feature extraction module includes at least one-dimensional convolutional layers and pooling layers for extracting local spatial features of transient electromagnetic signals.
[0011] Preferably, the BiLSTM timing modeling module includes a bidirectional LSTM layer for modeling the bidirectional long-range timing dependencies of transient electromagnetic signals.
[0012] Preferably, the attention mechanism module is the Bahdanau attention mechanism, which is used to dynamically weight the BiLSTM output and focus on key time channel information.
[0013] Preferably, the Seq2Seq decoder module includes an LSTM layer and a fully connected layer for autoregressive generation of formation parameter sequences.
[0014] Preferably, the construction of the training dataset includes logarithmic spatial uniform sampling within a preset resistivity and thickness range, and injecting segmented differential noise to simulate actual test conditions.
[0015] Preferably, the loss function of the model includes mean squared error (MSE) and mean absolute error (MAE), which are used for optimization during the training phase and evaluation during the validation phase, respectively.
[0016] Preferably, the method is applicable to the inversion of ground-to-space transient electromagnetic data in goaf areas, aquifers, or mineral exploration.
[0017] Preferably, the inversion time of the model for single measurement point data during the inference phase is less than 0.5 seconds.
[0018] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the method.
[0019] Compared with existing technologies, this invention provides a fast inversion method for ground-to-space transient electromagnetic data based on long short-term memory networks, which has the following advantages:
[0020] 1. This rapid inversion method for ground-to-space transient electromagnetic data based on long short-term memory networks utilizes a trained network for inference, reducing the inversion time for a single measurement point to the second level and compressing the traditional interpretation cycle of several days to the hour level, greatly improving the response speed of geological disaster monitoring and exploration.
[0021] 2. This method for fast inversion of ground-to-space transient electromagnetic data based on Long Short-Term Memory (LSTM) networks effectively captures the local spatial features of transient electromagnetic signals through a CNN module, models the bidirectional long-term decay process using a BiLSTM module, and dynamically focuses on key time channels using an attention mechanism, achieving multi-level feature fusion from local to global and from shallow to deep layers.
[0022] 3. This fast inversion method for ground-to-air transient electromagnetic data based on long short-term memory networks introduces piecewise differentiated noise into the training set to simulate actual measurement conditions. Combined with dual loss supervision of MSE and MAE, the model possesses strong anti-interference capabilities and adaptability to unknown data, resulting in more stable and reliable inversion results.
[0023] 4. Compared with the traditional global apparent resistivity imaging method, the rapid inversion method of ground-space transient electromagnetic data based on long short-term memory network can more accurately reflect the underground electrical structure and stratigraphic details, especially in the identification and thickness estimation of low-resistivity thin layers.
[0024] 5. This method for rapid inversion of ground-to-space transient electromagnetic data based on long short-term memory networks is applicable to various geological scenarios (such as goaf, aquifer, and mineral exploration). It does not require manual initial model setting or repeated parameter tuning, has end-to-end automated inversion capability, and is easy to promote and use in practical engineering. Attached Figure Description
[0025] Figure 1 This is a technical roadmap for the present invention;
[0026] Figure 2 This is a flowchart of the training set construction process for this invention;
[0027] Figure 3 This is a schematic diagram of the convolutional layer operation of the present invention;
[0028] Figure 4 This is a diagram of the convolutional neural network structure of the present invention;
[0029] Figure 5 This is a structural diagram of the LSTM unit of the present invention;
[0030] Figure 6 This is a diagram of the BiLSTM unit structure of the present invention;
[0031] Figure 7 This is a sequence-to-sequence model diagram of the present invention;
[0032] Figure 8 This is a diagram illustrating the Attention mechanism architecture of this invention.
[0033] Figure 9 This is the four-layer dielectric model of the present invention;
[0034] Figure 10 This is the loss function curve of the four-layer model of this invention;
[0035] Figure 11 This is the loss function curve of the two-layer model of this invention;
[0036] Figure 12This invention presents the interpretation results of transient electromagnetic data between the ground and the air in a coal mine.
[0037] Figure 13 This is a diagram illustrating the interpretation results of measurement line 1 in this invention;
[0038] Figure 14 This is a structural diagram of the CNN-BiLSTM-Attention model of this invention. Detailed Implementation
[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0040] Please see Figure 1-14 This invention provides a technical solution: a method for fast inversion of ground-to-space transient electromagnetic data based on long short-term memory networks.
[0041] Example 1: Inversion of the four-layer geoelectric structure based on a theoretical model
[0042] This embodiment demonstrates the application effect of the method of the present invention on theoretical model data and verifies its basic inversion capability.
[0043] S101: Training Set Construction
[0044] Training samples were constructed using a stochastic layered model generator function. The number of strata was set to 4, the resistivity sampling range was 10-200 Ω·m, and the layer thickness sampling range was 10-100 m. Uniform sampling was performed in logarithmic space to generate 10,000 sets of strata model parameters (ρ1,h1,ρ2,h2,ρ3,h3,ρ4).
[0045] S102: Forward Simulation
[0046] For each geological formation model, a one-dimensional forward modeling program was used to calculate its ground-to-air transient electromagnetic response. The transmitter length was set to 1000 m, and the receiver was located at an altitude of 30 m in the air, simulating the induced electromotive force (dBz / dt) over 50 time channels. 5% Gaussian random noise was added to the forward modeling data to simulate the measured conditions.
[0047] S103: Data Preprocessing
[0048] The input (transient electromagnetic response) and output (formation parameters) were normalized to the [0,1] interval using MinMaxScaler. The input data was reshaped into a three-dimensional tensor form (10000,1,50), and the output data was converted to (10000,1,7). The training and validation sets were divided in an 8:2 ratio.
[0049] S104: Network Construction and Training
[0050] The CNN-BiLSTM-Attention network is constructed as follows:
[0051] CNN module: Conv1D(64,kernel_size=3)→MaxPooling1D(2)→Conv1D(128,kernel_size=3)→MaxPooling1D(2)→Flatten
[0052] BiLSTM module: Bidirectional(LSTM(256,return_sequences=True))×2
[0053] Attention module: Dense(128, activation='tanh') → Dense(1) → Softmax → Weighted summation
[0054] Decoder: LSTM(256) → Dense(128) → Dense(7)
[0055] Training parameters: Optimizer Adam (lr=0.001), batch size 32, number of iterations 200, loss function MSE.
[0056] S105: Inversion Test and Results
[0057] The loss function converged after 73 training rounds. Two sets of test samples were selected for inversion, and the results are shown in Table 2. The relative errors between the inverted resistivity and layer thickness and the true values are both less than 25%, with the inversion error of resistivity ρ4 being only 0.6%, verifying the reliability of the method on the theoretical model.
[0058] Example 2: Inversion of measured data from goaf areas in northern Shaanxi coal mines
[0059] This embodiment demonstrates the application effect of the method of the present invention in actual engineering.
[0060] S201: Training Set Construction
[0061] Based on geological data from a coal mine in northern Shaanxi, a four-layer model was used to construct a training set. The resistivity ranged from 10 to 200 Ω·m, and the layer thickness ranged from 10 to 200 m. 100 measuring points were evenly selected within the survey area, and 10,000 models were generated for each measuring point, resulting in a total of 1,000,000 training samples.
[0062] S202: Network Training
[0063] Using the same network structure as in Example 1, some parameters were adjusted: Dropout=0.2, batch size=64. The loss function converged after 87 training rounds on a workstation, taking 14.8 hours.
[0064] S203: Inversion of measured data
[0065] The transient electromagnetic data from 100 measuring points in the coal mining area were input into a trained network for inversion. The average inversion time for each measuring point was only 0.2 seconds, and the inversion of the entire measuring line was completed in only 20 seconds.
[0066] S204: Comparative Analysis of Results
[0067] The inversion results of this invention are compared with the global apparent resistivity imaging results and the ground source loop inversion results. Figure 12-13 The results of this invention agree well with the results of ground-based transient electromagnetic inversion, and can clearly identify:
[0068] Shallow sandstone high-resistivity layer (30-80 Ω·m, thickness approximately 150 m)
[0069] The central water-bearing sandstone and mudstone layer has low resistivity (15-20 Ω·m, about 70 m thick).
[0070] Deep high-resistivity substrate (>25 Ω·m)
[0071] Compared with global apparent resistivity imaging, the results of this invention provide more accurate estimation of the thickness of low-resistivity layers and higher resolution, effectively overcoming the layer blurring problem caused by volume effects.
[0072] Figure 10 In the middle: (a) Mean squared error (MSE); (b) Mean absolute error (MAE).
[0073] Figure 11 In the middle: (a) Mean squared error (MSE); (b) Mean absolute error (MAE).
[0074] Figure 12 (a) Results of global apparent resistivity imaging; (b) Results of long short-term memory network inversion.
[0075] Figure 13(a) Full-domain apparent resistivity imaging results of ground-to-air transient electromagnetic data; (b) Inversion results of ground-to-air transient electromagnetic data using long short-time memory network; (c) Inversion results of ground-based fixed-source loop transient electromagnetic data.
[0076] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for fast inversion of ground-to-space transient electromagnetic data based on long short-term memory networks, characterized in that, Includes the following steps: Construct a training dataset for ground-to-air transient electromagnetic events, including generating resistivity and thickness parameters through a stochastic layered model and generating corresponding transient electromagnetic response data based on forward modeling. The training dataset is preprocessed, including normalization, temporal reshaping, and dataset partitioning; Construct a CNN-BiLSTM-Attention hybrid neural network model, which includes a CNN feature extraction module, a BiLSTM temporal modeling module, an attention mechanism module, and a Seq2Seq decoder module connected in sequence. The hybrid neural network model is trained using the training dataset until the loss function converges; The ground-to-air transient electromagnetic observation data to be inverted is input into the trained model, and the resistivity and thickness parameters of the underground medium are output.
2. The method for fast inversion of ground-to-space transient electromagnetic data based on long short-term memory networks according to claim 1, characterized in that, The CNN feature extraction module includes at least one-dimensional convolutional layers and pooling layers, used to extract local spatial features of transient electromagnetic signals.
3. The method for fast inversion of ground-to-space transient electromagnetic data based on long short-term memory networks according to claim 1, characterized in that, The BiLSTM timing modeling module includes a bidirectional LSTM layer for modeling the bidirectional long-range timing dependencies of transient electromagnetic signals.
4. The method for fast inversion of ground-to-space transient electromagnetic data based on long short-term memory networks according to claim 1, characterized in that, The attention mechanism module is the Bahdanau attention mechanism, which is used to dynamically weight the BiLSTM output and focus on key time channel information.
5. The method for fast inversion of ground-to-space transient electromagnetic data based on long short-term memory networks according to claim 1, characterized in that, The Seq2Seq decoder module includes an LSTM layer and a fully connected layer, which are used to autoregressively generate formation parameter sequences.
6. The method for fast inversion of ground-to-space transient electromagnetic data based on long short-term memory networks according to claim 1, characterized in that, The construction of the training dataset includes logarithmic space uniform sampling within a preset resistivity and thickness range, and injection of segmented differential noise to simulate actual test conditions.
7. The method for fast inversion of ground-to-space transient electromagnetic data based on long short-term memory networks according to claim 1, characterized in that, The model's loss function includes mean squared error (MSE) and mean absolute error (MAE), which are used for optimization during the training phase and evaluation during the validation phase, respectively.
8. The method for fast inversion of ground-to-space transient electromagnetic data based on long short-term memory networks according to claim 1, characterized in that, The method is applicable to the inversion of ground-to-space transient electromagnetic data in goaf areas, aquifers, or mineral exploration.
9. The method for fast inversion of ground-to-space transient electromagnetic data based on long short-term memory networks according to claim 1, characterized in that, The model takes less than 0.5 seconds to invert single-point data during the inference phase.
10. A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the method as described in any one of claims 1 to 9.