TEM measurement data inversion method and system based on unsupervised deep learning
By embedding forward physical rules in the LSTM deep network and using WOA-MVMD algorithm to denoising processing, combined with the Attention mechanism, the problems of low inversion accuracy and long calculation time of TEM measurement data are solved, and efficient and accurate inversion results are achieved .
Patent Information
- Application Number
- CN202411479486.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-23
- Publication Date
- 2024-11-22
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing TEM measurement data inversion methods have low accuracy, long calculation time, difficult to meet high efficiency requirements, and easily lead to misjudgment of inversion results and misunderstanding of geological understanding.
Using an unsupervised deep learning method, the physical rules of forward modeling are embedded in the LSTM deep network structure, the data loss is used as a loss function for training, and dynamic smoothing constraints and Attention mechanisms are introduced, combined with the WOA-MVMD algorithm to denoise To form an efficient inversion model.
It significantly improves the calculation efficiency and inversion accuracy, reduces the error of the resistivity model inversion, and can obtain high-quality TEM data in complex noise environments.
Smart Images

Figure CN119005016A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of electromagnetic data inversion, and in particular relates to a TEM measurement data inversion method, system and readable storage medium based on unsupervised deep learning. Background Art
[0002] Direct measurement of magnetic field using high-temperature superconducting sensors has the characteristics of low noise, high sensitivity, good low-frequency response, high signal-to-noise ratio in the late stage and sensitive to deep low-resistance response. Transient electromagnetic method based on high-temperature superconducting sensors is one of the most commonly used solutions in geophysical exploration. Therefore, it has been widely used in mineral resource exploration, groundwater exploration, coal mine goaf investigation and engineering geological investigation. Effective inversion method is the key to improving the reliability of imaging technology. At present, linear inversion is the main method for interpreting actual data. It predicts the model based on the physical laws of the geoelectric field. However, the inversion result is highly dependent on the initial model and the influence of human factors in geological interpretation, and usually only a local optimal solution can be obtained.
[0003] Nonlinear inversion methods have strong global search capabilities and can alleviate such problems. Many studies focus on introducing nonlinear methods to solve transient electromagnetic data inversion problems and obtain global optimal solutions, including the use of simulated annealing algorithms, firefly algorithms, differential evolution algorithms, and particle swarm optimization algorithms. However, due to their low computational efficiency, these methods have not been widely used in the inversion of actual data. In contrast, neural network methods are relatively fast and achieve a well-fitting nonlinear mapping between input data (observed voltage or magnetic field) and output data (resistivity model) by extracting information from a large set of training samples. These methods take a long time to train, and the trained networks have extremely high inversion efficiency when tested, making them suitable for TEM underground resistivity measurements. In recent years, with the significant improvement of artificial intelligence algorithms and computing performance, deep neural network learning (DL) methods have developed rapidly. Compared with the old neural network methods, deep learning has a stronger ability to construct complex nonlinear mappings. Therefore, using deep learning to solve geophysical inversion problems has gradually become a research hotspot.
[0004] At present, the traditional OCCAM inversion algorithm can only reflect the approximate shape of the stratigraphic model, with low accuracy and long calculation time. It cannot meet the requirements of high-efficiency measurement and may even easily lead to misjudgment of inversion results and misunderstanding of geological knowledge. Summary of the invention
[0005] The present invention provides a TEM measurement data inversion method and system based on unsupervised deep learning, which are used to solve the technical problems of low accuracy and long calculation time, which cannot meet the requirements of high measurement efficiency and may even easily cause misjudgment of inversion results and misunderstanding of geological knowledge.
[0006] In a first aspect, the present invention provides a TEM measurement data inversion method based on unsupervised deep learning, comprising: Obtain forward modeling results and real data; Embed the physical rules of forward modeling into the pre-built LSTM deep network structure; The forward modeling result and the real data are input into the embedded LSTM deep network structure, the LSTM deep network structure is trained with data lack of fit as the loss function, and the loss function is optimized according to the preset dynamic smoothing constraint to obtain the TEM measurement data inversion model, wherein the expression of the loss function is: , In the formula, is the loss function, Mapping for forward modeling, is the symbol of partial derivative, is the observed data, is the transpose symbol, is the dynamic regularization factor, The forward modeling results output by the prediction model; Among them, the dynamic regularization factor calculation formula is: , In the formula, is the initial value of the regularization factor, is the rate of change factor, is the current training number, is the maximum number of training times; The acquired real-time observation data is denoised using the WOA-MVMD algorithm to obtain target real-time observation data, and the target real-time observation data is input into the TEM measurement data inversion model, and the TEM measurement data inversion model outputs an inversion result corresponding to the real-time observation data.
[0007] In a second aspect, the present invention provides a TEM measurement data inversion system based on unsupervised deep learning, comprising: An acquisition module configured to acquire forward modeling results and real data; An embedding module configured to embed the physical rules of the forward modeling into a pre-built LSTM deep network structure; The training module is configured to input the forward modeling result and the real data into the embedded LSTM deep network structure, train the LSTM deep network structure with data lack of fit as the loss function, and optimize the loss function according to a preset dynamic smoothing constraint to obtain a TEM measurement data inversion model, wherein the expression of the loss function is: , In the formula, is the loss function, Mapping for forward modeling, is the symbol of partial derivative, is the observed data, is the transpose symbol, is the dynamic regularization factor, The forward modeling results output by the prediction model; Among them, the dynamic regularization factor calculation formula is: , In the formula, is the initial value of the regularization factor, is the rate of change factor, is the current training number, is the maximum number of training times; The output module is configured to use the WOA-MVMD algorithm to perform denoising on the acquired real-time observation data to obtain target real-time observation data, and input the target real-time observation data into the TEM measurement data inversion model, and the TEM measurement data inversion model outputs an inversion result corresponding to the real-time observation data.
[0008] According to a third aspect, an electronic device is provided, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can perform the steps of the TEM measurement data inversion method based on unsupervised deep learning according to any embodiment of the present invention.
[0009] In a fourth aspect, the present invention further provides a computer-readable storage medium having a computer program stored thereon. When the program instructions are executed by a processor, the processor executes the steps of the TEM measurement data inversion method based on unsupervised deep learning of any embodiment of the present invention.
[0010] The TEM measurement data inversion method and system based on unsupervised deep learning of the present application guides the inversion network training by embedding physical rules into the LSTM deep network structure and using data misfit as the loss function. In addition, dynamic smoothing constraints are added to the loss function to stabilize the training process. On this basis, the Attention mechanism is introduced to extract important feature information of deep learning inversion to reduce the error of resistivity model inversion; and the WOA-MVMD technology is proposed to remove interference noise from the measurement data to obtain high-quality high-temperature superconducting TEM data. The method of the present invention can effectively improve the computational efficiency and inversion accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0011] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0012] Figure 1 A flowchart of a TEM measurement data inversion method based on unsupervised deep learning provided by one embodiment of the present invention; Figure 2 A schematic diagram of TEM synthetic data containing different noises and WOA-MVMD denoising curves of a specific embodiment is provided for one embodiment of the present invention; Figure 3 A schematic diagram of a magnetic field strength fitting curve result of a specific embodiment is provided for an embodiment of the present invention; Figure 4 A schematic diagram of an inversion resistivity model of a specific embodiment is provided for one embodiment of the present invention; Figure 5 A schematic diagram of relative error curve results of a specific embodiment is provided for an embodiment of the present invention; Figure 6 A schematic diagram of a pseudo-two-dimensional resistivity profile obtained by inversion of a specific embodiment provided by an embodiment of the present invention; Figure 7 A structural block diagram of a TEM measurement data inversion system based on unsupervised deep learning provided by one embodiment of the present invention; Figure 8 It is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0013] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0014] See also Figure 1 , which shows a flow chart of a TEM measurement data inversion method based on unsupervised deep learning of the present application.
[0015] like Figure 1 As shown, the TEM measurement data inversion method based on unsupervised deep learning specifically includes the following steps: Step S101, obtaining forward modeling results and real data.
[0016] Step S102, embedding the physical rules of forward modeling into the pre-built LSTM deep network structure.
[0017] Assume that the training sample set includes N groups of resistivity models and its corresponding observational data The neural network parameters are expressed as , the purpose of the inversion network is to train To construct arrive Mapping : .
[0018] The training sample set contains the observed data and its corresponding resistivity model, which serve as the input information and label data information of the neural network respectively. The loss function is the prediction model and Resistivity Model Tags The inversion network is used to generate a prediction model from the observed data and added to the loss function together with the label. Finally, the loss function is used to calculate the network gradient to update the network parameters .
[0019] However, the actual resistivity model is usually not available in field exploration, and the supervised inversion network cannot be trained due to the lack of resistivity model labels. Therefore, it is difficult for the supervised inversion network to produce good inversion results using actual data. By adding forward modeling to the inversion network framework, an unsupervised learning inversion network framework is designed. Compared with the supervised learning inversion network framework, the unsupervised learning training sample set only contains actual observation data, which is used as the neural network input information. The forward modeling module is placed after the prediction model. The prediction model obtains the prediction data through forward simulation, and then the prediction data and the observation data are added to the loss function together. The input of the loss function changes from the model to the data, and finally the data-driven loss function is used to update the inversion network parameters. .
[0020] Forward modeling is based on the physical laws of electric field propagation, which follows Maxwell's governing equations. Maxwell's equations with current sources are expressed as: , , In the formula, is the Nabla operator, is the electric field strength, is the magnetic field strength, is the angular frequency, is the magnetic permeability, is the dielectric constant, is the current density, is the symbol of the imaginary part; The vertical magnetic field excited by the horizontal center loop source on the surface can be obtained: , In the formula, is the magnitude of the transmitting loop current, is the distance from the receiving point to the transmitting source, is a first-order Bessel function; m is the integral variable, , is the conduction current, , and denote the magnetic permeability and electrical conductivity of the uniform half-space medium, respectively. is the angular frequency of the electromagnetic wave, is the input impedance of the first layer. The Green's function describing wave physics is embedded. The Green's function is contained in the coefficient matrix, which determines the resistivity and perpendicular magnetic field The relationship between.
[0021] The vertical magnetic field is solved by using the resistivity model, control equations and boundary conditions. This process is called forward modeling in TEM resistivity measurement. Since there is no analytical solution for the time domain response of layered earth, it can only be solved by numerical calculation methods. The Bessel function is solved by Hankel transform, and then the Gaver-Stehfest transform is used to realize the frequency-time domain conversion to obtain the time domain response.
[0022] Since 3D TEM forward modeling is a heavy computational cost burden, 1D parametric modeling and inversion are often used in field exploration for real-time data interpretation, which has important practical significance. The 1D inversion results are usually spliced together to form a pseudo-2D (or pseudo-3D) underground image. In addition, the 1D inversion results provide a horizontal layered model of resistivity with depth, which can provide a basic understanding of the underground structure of the exploration area in advance, and then facilitate the actual 3D problem.
[0023] Step S103, inputting the forward modeling result and the real data into the embedded LSTM deep network structure, training the LSTM deep network structure with data lack of fit as the loss function, and optimizing the loss function according to a preset dynamic smoothing constraint to obtain a TEM measurement data inversion model.
[0024] In this step, a smoothness constraint is added to the loss function to ensure the convergence of the network training process. The expression of the loss function is: , In the formula, is the loss function, Mapping for forward modeling, is the symbol of partial derivative, is the observed data, is the transpose symbol, is the dynamic regularization factor, The forward modeling results output by the prediction model; Introduce a model gradient calculation strategy with dynamic smoothing constraints. In the early stage of network training, the model gradient calculation process is unstable due to the obvious difference between the prediction model and the label model. Therefore, in the early stage of network training, a larger regularization factor is used. To enhance the smoothness constraint, a smaller regularization factor is used in later training. To reduce the impact of smoothing constraints and achieve accurate imaging of the target area. The dynamic regularization factor calculation formula is: , In the formula, is the initial value of the regularization factor, is the rate of change factor, is the current training number, is the maximum number of training times.
[0025] The Attention mechanism is introduced to make biased selection of data at different levels output by the LSTM network, focusing on the highly correlated and important information with large weights, and allocating low weights to the information with bland features, so as to perform deeper feature extraction on the data, reduce the errors caused by uncertainty and volatility, and improve the stability and reliability of the formation inversion results.
[0026] The attention mechanism includes an encoder and a decoder. Assuming that there are t time steps of input, the output of the encoder at the t time step is expressed as , the output of the decoder at time step i is , then the control vector of the i-th time step is The calculation process is: , , , , In the formula, , , , are model learning parameters, It is the temporary weight of the hidden layer output value to the current input value in the LSTM deep network structure. is the attention weight, To extract important feature information, is the output of the encoder at time step t, is the output value of the decoder at the i-th time step, is the output value of the decoder at the i-1th time step, is the input value, is the kth temporary weight, is the activation function, is the hidden layer state value at the i-th time step.
[0027] Among them, the feature importance evaluation rule is defined as follows: first, calculate the attention weight value of each feature output by the LSTM network. The larger the weight value, the better the effect of the feature on the deep learning stratigraphic model inversion. Then, these features are added together to obtain the feature vector after feature importance evaluation. When performing stratigraphic inversion, we use the feature importance evaluation rule to pay more attention to the feature vectors with higher weight values, and pay less attention to the feature vectors with smaller weight values. This effectively solves the high false alarm rate caused by the similarity between TEM data features and improves the inversion accuracy to a certain extent.
[0028] Step S104, using the WOA-MVMD algorithm to perform denoising on the acquired real-time observation data to obtain target real-time observation data, and inputting the target real-time observation data into the TEM measurement data inversion model, and the TEM measurement data inversion model outputs an inversion result corresponding to the real-time observation data.
[0029] It is difficult to establish a spatial correspondence between resistivity models of noisy data, so a simple deep learning algorithm is difficult to be widely applied to noisy data. A TEM data feature extraction method that is suitable for noisy signals is proposed, that is, the WOA-MVMD algorithm is used to extract effective features from noisy data.
[0030] Decomposing the multivariate time series signal into multiple intrinsic mode functions (IMFs), then finding the optimal component from the decomposed intrinsic mode functions (IMFs) and reconstructing the signal can effectively remove the noise component in the signal and obtain a cleaner signal. The core idea of MVMD is to decompose the input multivariate time series data into its inherent multivariate modulated oscillation. The steps are as follows: A constrained multivariate variational optimization problem is constructed, the goal of which is to minimize the sum of the bandwidths of the multivariate intrinsic mode function set, expressed as: , In the formula, is the number of separate modes set in the MVMD algorithm, is the number of original signals input to the MVMD algorithm, is the center frequency of each mode after MVMD decomposition, is the set of input multivariate time series feature signals, is the set of decomposed time series feature signals, To use the Hilbert transform operator to transform a vector signal The analytical representation of is the symbol of the partial derivative operation with respect to time, and j is the symbol of the imaginary part.
[0031] An augmented Lagrangian function is constructed to remove the constraints of the multivariate variational optimization problem; the multivariate variational optimization problem in the augmented Lagrangian function is solved by an alternating direction multiplier method to obtain the required decomposition mode; and the acquired real-time observation data is denoised according to the decomposition mode to obtain the target real-time observation data.
[0032] The decomposition effect of the MVMD algorithm is mainly related to the number of decomposition layers. and penalty factor In order to solve the influence of artificial empirical parameter settings on the decomposition results of multivariate signals, the present invention proposes a signal denoising method based on whale algorithm optimized multivariate variational mode decomposition (WOA-MVMD). The whale optimization algorithm (WOA) is a new group intelligence optimization search method that simulates the hunting behavior of humpback whale groups in nature (mainly the contraction and encirclement mechanism and the spiral ascent mechanism). It can effectively find the optimal solution and has strong robustness and convergence speed. The WOA-MVMD method applies the whale algorithm to the parameter optimization of the MVMD algorithm. By optimizing the important parameters of the MVMD algorithm (the number of decomposition layers and the penalty factor), it realizes the fast and accurate decomposition of the time series signal and improves the decomposition accuracy and anti-noise ability of the MVMD algorithm. Specifically: Initialize the population size, maximum number of iterations, spatial dimension, and initial position of the whale in the WOA algorithm, and set the value range of the number of eigenmodes, penalty factor, and number of decomposition levels in the MVMD algorithm; Perform MVMD decomposition on the original signal to obtain each IMF component, calculate the sample value of each IMF component, and select the component with the smallest sample entropy value as the fitness function of the global search; Randomly generate probability in the interval [0,1] ,like , then update the whale position in a spiral manner; if And the coefficient vector , then update the whale's position by random search; if and , then update the whale position in a shrinking and encircling manner, and save the current optimal fitness function; The updated whale population position in the previous round of iteration is used as the initial population position in the next round, and the iteration is repeated until the iteration condition is met; When the loop iteration ends, the optimal fitness function and the optimal parameter combination are output.
[0033] In summary, the method of the present application embeds the physical rules (of forward modeling) into the LSTM deep network structure and uses the data lack of fit as the loss function to guide the inversion network training. In addition, a dynamic smoothness constraint is added to the loss function to stabilize the training process. On this basis, the Attention mechanism is introduced to extract important feature information of deep learning inversion to reduce the error of resistivity model inversion; and the WOA-MVMD technology is proposed to remove the interference noise of the measurement data to obtain high-quality high-temperature superconducting TEM data. The method of the present invention can effectively improve the computational efficiency and inversion accuracy.
[0034] In a specific embodiment, in order to verify the effectiveness of the WOA-MVMD denoising algorithm, denoising analysis is performed on TEM synthetic data containing Gaussian random noise. The initial population size of the WOA algorithm is set to 10, the spatial dimension is 2, the maximum number of iterations is 30, the decomposition layer value range is [2, 200], and the penalty factor value range is [10, 2500]. The synthetic magnetic field obtained by theoretical simulation is pure data, while the actual observed magnetic field is noisy data. In order to simulate the actual observed data, the present invention applies 1%, 5% and 10% Gaussian noise to the TEM synthetic data, respectively, and then analyzes the noise reduction effect of the WOA-MVMD method, as shown in FIG. Figure 2 TEM synthetic data with different noises and the noise reduction effect curve of the proposed method.
[0035] Depend on Figure 2 It can be seen that noise has a great impact on TEM data. After decomposition and processing by the WOA-MVMD denoising algorithm, the best component is selected to reconstruct the denoised TEM signal data. Although the TEM data after preprocessing the high-noise synthetic data has a large deviation from the real data, it can basically fit the real data well. The results show that the WOA-MVMD algorithm can significantly improve the signal quality of TEM in the middle and late stages, and can obtain higher quality TEM response vertical magnetic field data, which proves that the proposed denoising method can effectively alleviate noise interference, has good robustness and stability, and verifies the effectiveness and feasibility of the method.
[0036] In order to achieve high-precision inversion of unsupervised TEM resistivity, a two-stage TEM resistivity inversion method (PhTemNet) based on WOA-MVMD and deep learning network was constructed in the present invention.
[0037] In the first stage, WOA is used to optimize the important parameters of the MVMD algorithm (decomposition level and penalty factor), and WOA-MVMD is used to perform time-frequency synchronization analysis on the one-dimensional TEM data sequence, decomposing it into frequency-aligned multivariate intrinsic mode functions. Each original sequence is decomposed into k intrinsic mode functions (IMFs), and the d-dimensional original sequence is decomposed into dimensional frequency-aligned IMFS, i.e. , each dimension decomposition signal has the same central frequency. In the second stage, with the historical multivariate IMFs as input, an unsupervised deep neural network model based on LSTM-Attention is established to perform TEM resistivity inversion. In LSTM-Attenton, the best IMFs are first selected as the input sequence. Based on the unique LSTM unit and the structure of the correlation connection within the hidden layer (the number of LSTM layers in the present invention is 16), the time-related features of the subsequence can be effectively extracted, thereby obtaining the spatiotemporal correlation features h of the subsequence; then, the spatiotemporal correlation features h are input into the Attention layer (the number of Attention layers in the present invention is 2) to enhance the network's learning weights for important time point features; finally, a fully connected layer is connected in series to output the inverted resistivity sequence, and the relative error index is used for analysis.
[0038] It should be noted that LSTM: long short-term memory neural network; LSTM-Attention: Long Short-Term Memory Neural Network-Attention Mechanism; WOA-MVMD-LSTM-Attention: Whale Algorithm-Multivariate Variational Mode Decomposition-Long Short-Term Memory Neural Network-Attention Mechanism.
[0039] In addition, in constructing the deep LSTM inversion network, the present invention utilizes the physical laws based on electric field propagation to perform unsupervised learning, and proposes a dynamic smooth constraint loss function to ensure the convergence of training, thereby realizing physics-driven deep learning inversion of TEM data.
[0040] In this paper, we compare the proposed WOA-MVMD-LSTM-Attention with traditional deep learning methods through numerical experiments. All method simulations adopt unsupervised training mode, that is, no real resistivity model is required for training.
[0041] Geoelectrical model inversion A one-dimensional geoelectric model synthetic data set was established by random combination. The synthetic data sets were generated by 1D geoelectric models with 4, 5, and 6 layers, which are different from the models used to generate the sample data set. After the synthetic model set was established, the corresponding TEM forward response was calculated, with a sampling time range of 1e-6 s to 1e-3s, for a total of 31 time channels, which is the same as the time channels used in the sample set. In order to simulate the actual transient electromagnetic data, 10% Gaussian random noise was added to these synthetic data sets.
[0042] Three inversion algorithms were tested on the same synthetic dataset for inversion comparison, including the traditional deep LSTM algorithm, the deep LSTM algorithm with Attention mechanism (LSTM-Attention), and the deep LSTM-Attention algorithm with WOA-MVMD denoising (WOA-MVMD-LSTM-Attention). For these three deep learning methods, we use the unsupervised learning inversion mode, directly inputting the TEM synthetic data into the trained network without considering other factors, and it will immediately output all the inversion model parameters.
[0043] For synthetic data, such as Figure 3 , Figure 4 , Figure 5 The magnetic field intensity fitting curve, inversion resistivity model and relative error curve results of the three inversion algorithms are shown respectively. It can be seen from these figures that the three deep learning inversion algorithms can fit the resistivity distribution more accurately, and the TEM simulation response corresponding to the inverted resistivity model is consistent with the true value. Due to the influence of smoothness constraints, the inversion resistivity models obtained by all deep learning algorithms show smooth characteristics and deviate greatly from the true layered model, but the proposed WOA-MVMD-LSTM-Attention algorithm is better than the other two algorithms and is closer to the true model. Its inversion resistivity and depth are more consistent with the true one. The inversion results of the 4-layer, 5-layer and 6-layer geoelectric models were analyzed using these three algorithms. The forward magnetic field response of the inversion results of the proposed algorithm has a good fitting effect with the forward magnetic field response of the true model. The maximum relative errors of its forward magnetic field response are 8.15%, 14.50% and 9.85% respectively. The overall performance is far better than the traditional LSTM and LSTM-Attention algorithms. In addition, the three algorithms took 2 to 3 hours to train the three synthetic models. However, once the network training is completed, the proposed deep learning algorithms can predict reliable inversion models in real time.
[0044] It can be concluded that all three algorithms can obtain reliable inversion results from synthetic data containing 10% noise, and the algorithm proposed in this paper can achieve the best inversion effect. This is because 1) the multivariate variational mode decomposition technology using the whale optimization algorithm can remove the noise of TEM synthetic data and obtain higher quality TEM response vertical magnetic field data; 2) the introduction of the Attention mechanism to extract important feature information of deep learning inversion can reduce the error of resistivity model inversion. In addition, the biggest advantage of deep learning inversion is that once the network training is completed, the computational cost can be ignored. The research results show that the WOA-MVMD-LSTM-Attention algorithm has high inversion accuracy and computational efficiency in transient electromagnetic method applications, has good anti-noise performance, and can be applied to the inversion of measured data in complex noise environments.
[0045] Quasi-2D geoelectric model inversion In order to further verify the feasibility of the proposed algorithm for TEM inversion / / the inversion performance of the inversion scheme, a typical quasi-two-dimensional geoelectric model (including high and low anomalies) was designed for inversion testing and compared with the traditional LSTM method. Both methods adopt the unsupervised learning inversion mode.
[0046] The theoretical model and measurement point locations of high and low anomalies are as follows: Figure 6 As shown in the figure. In the theoretical model, the two anomalies are both 200 m×200 m in shape, the high-resistance body resistivity is 1000 Ω·m, the low-resistance body resistivity is 10 Ω·m, and the burial depth is 100 m from the top. They are both located in a uniform half-space with a resistivity of 100 Ω·m. In the figure, the horizontal direction is the measurement line and the vertical direction is the measurement depth. The length of the survey line is 1200 m, the interval between the measurement points is 100 m, and there are 13 measurement points in total. The one-dimensional resistivity model of each point is used to calculate the corresponding synthetic noise-free TEM response to be inverted. In order to truly simulate the field data, 10% Gaussian random noise is added to the electromagnetic data obtained at each measurement point. Among them, the quasi-two-dimensional geoelectric model inversion refers to the use of the inversion algorithm to perform a separate one-dimensional model inversion on all the measurement point data on the survey line, and then generate a two-dimensional profile using the Kriging interpolation method in the Sufer software.
[0047] The feasibility of the proposed method in TEM inversion is further verified by inversion examples of theoretical layered models and quasi-two-dimensional formation models. The proposed method has high inversion efficiency, mainly because: 1) the forward modeling describing physical rules is embedded in the deep LSTM network training, and a dynamic smoothing constraint loss function is introduced to achieve good unsupervised learning; 2) the LSTM network structure with the Attention mechanism is designed to effectively extract important feature information in the inversion; 3) the WOA-MVMD technology is used to remove the noise of the measurement data, which can improve the inversion imaging accuracy of the resistivity model. Therefore, the proposed WOA-MVMD-LSTM-Attention can better handle complex resistivity models, and has more reliable inversion results than traditional deep learning inversion methods, showing good practical application prospects.
[0048] The beneficial effects of the present invention are as follows: a method for inversion of high-temperature superconducting TEM measurement data based on unsupervised deep learning is proposed. The method embeds the physical rules (of forward modeling) into the LSTM deep network structure, and uses data misfit as the loss function to guide the inversion network training. In addition, a dynamic smoothing constraint is added to the loss function to stabilize the training process. On this basis, the Attention mechanism is introduced to extract important feature information of deep learning inversion to reduce the error of resistivity model inversion; and the WOA-MVMD technology is proposed to remove interference noise from the measurement data to obtain high-quality TEM data. The results of numerical simulation and model tests show that unsupervised learning inversion can accurately reconstruct the nonlinear mapping from input (TEM data) to output (resistivity model), and has the ability to handle complex realistic exploration scenarios. The proposed method can improve the efficiency of TEM data interpretation and provide new methods and techniques for high-temperature superconducting TEM data processing and interpretation.
[0049] Compared with the existing resistivity inversion methods based on supervised learning, this method eliminates the dependence of the training set on labels (actual resistivity models). Since it is difficult to obtain actual resistivity models, this method is more suitable for actual measurement scenarios. Compared with the traditional OCCAM method, this method does not rely on the initial model and can achieve fast real-time inversion imaging. This efficient processing speed can meet engineering requirements, especially for large-scale measurement areas.
[0050] See also Figure 7 , which shows a structural block diagram of a TEM measurement data inversion system based on unsupervised deep learning in the present application.
[0051] like Figure 7 As shown, the TEM measurement data inversion system 200 includes an acquisition module 210 , an embedding module 220 , a training module 230 and an output module 240 .
[0052] Wherein, the acquisition module 210 is configured to acquire forward modeling results and real data; An embedding module 220, configured to embed the physical rules of the forward modeling into a pre-built LSTM deep network structure; The training module 230 is configured to input the forward modeling result and the real data into the embedded LSTM deep network structure, train the LSTM deep network structure with data lack of fit as the loss function, and optimize the loss function according to a preset dynamic smoothing constraint to obtain a TEM measurement data inversion model, wherein the expression of the loss function is: , In the formula, is the loss function, Mapping for forward modeling, is the symbol of partial derivative, is the observed data, is the transpose symbol, is the dynamic regularization factor, The forward modeling results output by the prediction model; Among them, the dynamic regularization factor calculation formula is: , In the formula, is the initial value of the regularization factor, is the rate of change factor, is the current training number, is the maximum number of training times; The output module 240 is configured to use the WOA-MVMD algorithm to perform denoising on the acquired real-time observation data to obtain target real-time observation data, and input the target real-time observation data into the TEM measurement data inversion model, and the TEM measurement data inversion model outputs an inversion result corresponding to the real-time observation data.
[0053] It should be understood that Figure 7 Modules and references documented in Figure 1 Therefore, the operations and features described above for the method and the corresponding technical effects are also applicable to Figure 7 The modules in it will not be described in detail here.
[0054] In some other embodiments, the embodiments of the present invention further provide a computer-readable storage medium having a computer program stored thereon, wherein when the program instructions are executed by a processor, the processor is caused to execute the TEM measurement data inversion method based on unsupervised deep learning in any of the above method embodiments; As an implementation mode, the computer-readable storage medium of the present invention stores computer-executable instructions, and the computer-executable instructions are configured as follows: Obtain forward modeling results and real data; Embed the physical rules of forward modeling into the pre-built LSTM deep network structure; The forward modeling result and the real data are input into the embedded LSTM deep network structure, the LSTM deep network structure is trained with data lack of fit as the loss function, and the loss function is optimized according to the preset dynamic smoothing constraint to obtain the TEM measurement data inversion model, wherein the expression of the loss function is: , In the formula, is the loss function, Mapping for forward modeling, is the symbol of partial derivative, is the observed data, is the transpose symbol, is the dynamic regularization factor, The forward modeling results output by the prediction model; Among them, the dynamic regularization factor calculation formula is: , In the formula, is the initial value of the regularization factor, is the rate of change factor, is the current training number, is the maximum number of training times; The acquired real-time observation data is denoised using the WOA-MVMD algorithm to obtain target real-time observation data, and the target real-time observation data is input into the TEM measurement data inversion model, and the TEM measurement data inversion model outputs an inversion result corresponding to the real-time observation data.
[0055] The computer-readable storage medium may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created according to the use of a TEM measurement data inversion system based on unsupervised deep learning, etc. In addition, the computer-readable storage medium may include a high-speed random access memory, and may also include a memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the computer-readable storage medium may optionally include a memory remotely disposed relative to the processor, and these remote memories may be connected to the TEM measurement data inversion system based on unsupervised deep learning via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.
[0056] Figure 8 is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention, such as Figure 8 As shown, the device includes: a processor 310 and a memory 320. The electronic device may also include: an input device 330 and an output device 340. The processor 310, the memory 320, the input device 330 and the output device 340 may be connected via a bus or other means. Figure 8 In the example, the bus connection is used. The memory 320 is the above-mentioned computer-readable storage medium. The processor 310 executes various functional applications and data processing of the server by running the non-volatile software programs, instructions and modules stored in the memory 320, that is, the TEM measurement data inversion method based on unsupervised deep learning in the above-mentioned method embodiment is realized. The input device 330 can receive input digital or character information, and generate key signal input related to user settings and function controls of the TEM measurement data inversion system based on unsupervised deep learning. The output device 340 may include display devices such as display screens.
[0057] The electronic device can execute the method provided by the embodiment of the present invention, and has the functional modules and beneficial effects corresponding to the execution method. For technical details not described in detail in this embodiment, please refer to the method provided by the embodiment of the present invention.
[0058] As an implementation mode, the electronic device is applied to a TEM measurement data inversion system based on unsupervised deep learning, and is used for a client, and includes: at least one processor; and a memory connected to the at least one processor in communication; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can: Obtain forward modeling results and real data; Embed the physical rules of forward modeling into the pre-built LSTM deep network structure; The forward modeling result and the real data are input into the embedded LSTM deep network structure, the LSTM deep network structure is trained with data lack of fit as the loss function, and the loss function is optimized according to the preset dynamic smoothing constraint to obtain the TEM measurement data inversion model, wherein the expression of the loss function is: , In the formula, is the loss function, Mapping for forward modeling, is the symbol of partial derivative, is the observed data, is the transpose symbol, is the dynamic regularization factor, The forward modeling results output by the prediction model; Among them, the dynamic regularization factor calculation formula is: , In the formula, is the initial value of the regularization factor, is the rate of change factor, is the current training number, is the maximum number of training times; The acquired real-time observation data is denoised using the WOA-MVMD algorithm to obtain target real-time observation data, and the target real-time observation data is input into the TEM measurement data inversion model, and the TEM measurement data inversion model outputs an inversion result corresponding to the real-time observation data.
[0059] Through the description of the above implementation modes, those skilled in the art can clearly understand that each implementation mode can be implemented by means of software plus a necessary general hardware platform, or of course by hardware. Based on such an understanding, the above technical solution can essentially or in other words be embodied in the form of a software product that contributes to the prior art. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods of each embodiment or some parts of the embodiment.
[0060] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A TEM measurement data inversion method based on unsupervised deep learning, characterized in that: include: Obtain forward modeling results and real data; Embed the physical rules of forward modeling into the pre-built LSTM deep network structure; The forward modeling result and the real data are input into the embedded LSTM deep network structure, the LSTM deep network structure is trained with data lack of fit as the loss function, and the loss function is optimized according to the preset dynamic smoothing constraint to obtain the TEM measurement data inversion model, wherein the expression of the loss function is: , In the formula, is the loss function, Mapping for forward modeling, is the symbol of partial derivative, is the observed data, is the transpose symbol, is the dynamic regularization factor, The forward modeling results output by the prediction model; Among them, the dynamic regularization factor calculation formula is: , In the formula, is the initial value of the regularization factor, is the rate of change factor, is the current training number, is the maximum number of training times; The acquired real-time observation data is denoised using the WOA-MVMD algorithm to obtain target real-time observation data, and the target real-time observation data is input into the TEM measurement data inversion model, and the TEM measurement data inversion model outputs an inversion result corresponding to the real-time observation data.
2. The TEM measurement data inversion method based on unsupervised deep learning according to claim 1, characterized in that: The embedded LSTM deep network structure contains an attention mechanism, which includes an encoder and a decoder. Assuming that there are t time steps of input, the output of the encoder at the t time step is expressed as , the output of the decoder at the i-th time step is , then the control vector at the i-th time step is The calculation process is: , , , , In the formula, , , , are model learning parameters, It is the temporary weight of the hidden layer output value to the current input value in the LSTM deep network structure. is the attention weight, To extract important feature information, is the output of the encoder at time step t, is the output value of the decoder at the i-th time step, is the output value of the decoder at the i-1th time step, is the input value, is the kth temporary weight, is the activation function, is the hidden layer state value at the i-th time step.
3. The TEM measurement data inversion method based on unsupervised deep learning according to claim 1, characterized in that: The WOA-MVMD algorithm is used to perform denoising on the acquired real-time observation data to obtain the target real-time observation data, including: Determine an optimal hyperparameter combination of the MVMD algorithm, wherein the optimal hyperparameter combination includes a decomposition layer number and a penalty factor; A constrained multivariate variational optimization problem is constructed, the goal of which is to minimize the sum of the bandwidths of the multivariate intrinsic mode function set, expressed as: , In the formula, is the number of separate modes set in the MVMD algorithm, is the number of original signals input to the MVMD algorithm, is the center frequency of each mode after MVMD decomposition, is the set of input multivariate time series feature signals, is the set of decomposed time series feature signals, To use the Hilbert transform operator to transform a vector signal The analytical representation of is the symbol of the partial derivative operation of time, j is the symbol of the imaginary part; Construct an augmented Lagrangian function to remove constraints from multivariate variational optimization problems; The multivariate variational optimization problem in the augmented Lagrangian function is solved by the alternating direction multiplier method to obtain the required decomposed modes. The acquired real-time observation data is denoised according to the decomposition mode to obtain target real-time observation data.
4. The TEM measurement data inversion method based on unsupervised deep learning according to claim 3, characterized in that: The optimal hyperparameter combination of the MVMD algorithm is determined as follows: Initialize the population size, maximum number of iterations, spatial dimension, and initial position of the whale in the WOA algorithm, and set the value range of the number of eigenmodes, penalty factor, and number of decomposition levels in the MVMD algorithm; Perform MVMD decomposition on the original signal to obtain each IMF component, calculate the sample value of each IMF component, and select the component with the smallest sample entropy value as the fitness function of the global search; Randomly generate probability in the interval [0,1] ,like , then update the whale position in a spiral manner; if And the coefficient vector , then update the whale's position by random search; if and , then update the whale position in a shrinking and encircling manner, and save the current optimal fitness function; The updated whale population position in the previous iteration is used as the initial population position in the next iteration, and the iteration is repeated until the iteration condition is met; When the loop iteration ends, the optimal fitness function and the optimal parameter combination are output.
5. A TEM measurement data inversion system based on unsupervised deep learning, characterized in that: include: An acquisition module configured to acquire forward modeling results and real data; An embedding module configured to embed the physical rules of the forward modeling into a pre-built LSTM deep network structure; The training module is configured to input the forward modeling result and the real data into the embedded LSTM deep network structure, train the LSTM deep network structure with data lack of fit as the loss function, and optimize the loss function according to a preset dynamic smoothing constraint to obtain a TEM measurement data inversion model, wherein the expression of the loss function is: , In the formula, is the loss function, Mapping for forward modeling, is the symbol of partial derivative, is the observed data, is the transpose symbol, is the dynamic regularization factor, The forward modeling results output by the prediction model; Among them, the dynamic regularization factor calculation formula is: , In the formula, is the initial value of the regularization factor, is the rate of change factor, is the current number of training times, is the maximum number of training times; The output module is configured to use the WOA-MVMD algorithm to perform denoising on the acquired real-time observation data to obtain target real-time observation data, and input the target real-time observation data into the TEM measurement data inversion model, and the TEM measurement data inversion model outputs an inversion result corresponding to the real-time observation data.
6. An electronic device, characterized in that: include: At least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method described in any one of claims 1 to 4.
7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 4 is implemented.
Citation Information
Patent Citations
Aviation transient electromagnetic data inversion method based on LSTM network
CN113568055A
Direct current resistivity inversion method and system based on unsupervised deep learning
CN116090293A
Photovoltaic power prediction method based on VMD decomposition and lamination deep learning
CN116644970A
Magnetotelluric physical and data hybrid drive inversion method
CN118194728A
Cited By
Anti-collision guardrail stand column burial depth inspection method, system and device based on deep learning
CN121835454A
Deep learning-based crash barrier post depth inspection method, system and device
CN121835454B
