Noise suppression method and system used for transient electromagnetic data
The noise suppression method using a neural network-based system effectively addresses the challenges of parameter sensitivity and complexity in TEM data processing, achieving superior noise reduction and adaptability in complex environments.
Patent Information
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- INSTITUTE OF GEOLOGY AND GEOPHYSICS CHINESE ACADEMY OF SCIENCES
- Filing Date
- 2025-01-21
- Publication Date
- 2026-04-22
AI Technical Summary
Existing noise suppression methods for transient electromagnetic (TEM) data are sensitive to parameter selection, particularly for long-time sequences or high-dimensional data, and suffer from computational complexity, with variational mode decomposition being sensitive to noise and requiring additional post-processing, while other methods like wavelet transform, Kalman filter, and singular value decomposition have drawbacks such as lack of adaptability or complex operations.
A noise suppression method using a noise suppression network constructed based on a regression-type neural network and LSTM network, trained with stacked forward modeling data and noise data of different weights, to effectively suppress noise in TEM data.
The method achieves excellent noise reduction, improving the efficiency and adaptability of TEM data processing by precisely rejecting various types of noise, including commercial frequency interference, and simplifying the noise reduction process, surpassing traditional methods in practicality and efficiency.
Smart Images
Figure 0007849763000025 
Figure 0007849763000026 
Figure 0007849763000027
Abstract
Description
[Technical Field]
[0001] The present invention relates to the technology of noise suppression, and more specifically to a noise suppression method and system used for transient electromagnetic data. [Background technology]
[0002] Transient electromagnetic (TEM) exploration is a time-domain electromagnetic exploration technique built on the principle of electromagnetic induction and is widely applied to the exploration of mineral and hydrological resources. TEM works by emitting an initial field into the subsurface using a step wave or other pulsed current field source. At the moment the initial field is interrupted, the attenuation characteristics of the secondary induced electromagnetic field generated from the subsurface geological body are measured. Based on the signal received by the analytical instrument, the conductivity and depth of the target geological body are detected, thereby achieving the objective of exploring the subsurface geological body. Transient electromagnetic signals are secondary field signals that attenuate over time. When operating in environments with strong noise interference, such as mining areas or industrial zones, the later signals received by the instrument are greatly affected by various natural and man-made interference noises, resulting in low resolution of deep geological bodies. Therefore, when using TEM, the presence of noise has many unfavorable effects on later data processing and geological analysis. These interference signals can mask necessary and useful signals. Unless noise is effectively suppressed and useful information from the signal is extracted, transient electromagnetic data cannot be better utilized for subsequent data processing and inversion interpretation.
[0003] The patent application, publication number CN117056777A, proposes a transient electromagnetic signal noise reduction method for improving variational mode decomposition based on a sparrow search algorithm. This method globally optimizes the penalty coefficient α and the number of modes K in variational mode decomposition using the sparrow search algorithm to obtain the optimal parameters [K,α]. Variational mode decomposition is performed on the collected transient electromagnetic signal using the optimal combination of parameters obtained after global optimization, thereby realizing a noise reduction process for the original signal. This method avoids the mode mixing problem caused by artificially selecting parameters and can enhance the noise reduction effect on the signal.
[0004] The patent application, publication number CN111650655A, discloses a transient electromagnetic signal noise suppression method with monitoring of non-negative matrix factorization. This method involves, in a training stage, performing a short-time Fourier transform and non-negative matrix factorization on a pure signal to obtain an atomic dictionary of each feature of the characterization signal; then, in a noise suppression stage, processing a noise-containing signal using the atomic dictionary and noise suppression model to obtain a preliminaryly estimated transient electromagnetic signal; and finally, repeating these steps multiple times to accumulate the later data of the preliminaryly estimated transient electromagnetic signal and the original initial and mid-term data of the noise-containing signal, respectively, calculating their arithmetic mean values, and then combining them to obtain the estimated final complete transient electromagnetic signal. This invention can effectively remove noise in actual transient electromagnetic signals and improve the accuracy of transient electromagnetic signal inversion. However, the conventional techniques described above are sensitive to parameter selection and are computationally complex, especially for long sequences or high-dimensional data. Furthermore, variational mode decomposition methods are sensitive to noise in the input signal, and the noise may be decomposed into each mode. This may necessitate the addition of post-processing steps to handle the noise.
[0005] Initial time-domain electromagnetic noise rejection employed methods such as wavelet transform, Kalman filter, and singular value decomposition. Each of these methods has its drawbacks; some lack adaptability, some have weak noise rejection effects, and others have complex operating procedures.
[0006] As described above, the present invention proposes a noise suppression method and system for transient electromagnetic data.
Summary of the Invention
Means for Solving the Problems
[0007] The prior art is sensitive to parameter selection, especially for long - time sequences or high - dimensional data, and the computational complexity is relatively high. Moreover, the variational mode decomposition method is sensitive to noise in the input signal, and noise may be decomposed into each mode. Initial time - domain electromagnetic noise removal has adopted methods such as wavelet transform, Kalman filter, and singular value decomposition. Each of these methods has its own drawbacks, such as lack of adaptability, weak noise removal effect, or complex operation process. To solve the problems in the prior art, the present invention proposes a noise suppression method and system for transient electromagnetic data.
[0008] As one aspect of the present invention, a noise suppression method for transient electromagnetic data is proposed, The method includes: Stacking forward - modeled data and noise data based on different weights, and after stacking, using it as a data set, training a noise suppression network pre - constructed based on the data set to obtain a trained noise suppression network. Here, the data set includes a training set and a test set, and the different weights are obtained based on a preset target signal - to - noise ratio. Inputting the TEM data to be noise - suppressed into the trained noise suppression network to obtain the TEM data after noise suppression. Here, the noise suppression network is constructed based on a regression - type neural network and an LSTM network.
[0009] In some preferred embodiments, the construction method of the data set is as follows: Construct a plurality of different one-dimensional resistivity models, perform forward modeling, and obtain forward modeling data corresponding to the plurality of one-dimensional resistivity models. Artificially simulate noise data of different types and different intensities, and the noise data includes atmospheric noise, commercial frequency noise, and Gaussian noise. Stack the plurality of forward modeling data and the plurality of noise data with different weights, and use the stacked plurality of data as a data set. Divide the data set into a training set and a test set at a set ratio.
[0010] In some preferred embodiments, the acquisition method of the data set is shown by the following formula.
[0011]
Number
[0012] Here, D final is the stacked data set, D i forward is the i-th forward modeling data, D i noise is the i-th noise data, ω forward and ω noise are the weight values of the forward modeling data and the noise data, and are used to adjust the ratios of the forward modeling data and the noise data during stacking respectively. N is the data volume of the forward modeling data.
[0013] In some preferred embodiments, ω forward and ω noise are obtained by the following method.
[0014]
Number
[0015]
Number
[0016] Here, SNR target This is the preset target signal-to-noise ratio.
[0017] In some preferred embodiments, the pre-constructed noise suppression network includes sequentially connected input layers, encoders, decoders, and output layers. The input layer is used to receive input data including the dataset or the noise-suppressing TEM data. Both the encoder and decoder are constructed using two sequentially connected LSTM layers, and the encoder is used to learn the time-series feature data of the input data and convert it into abstract data. The decoder is used to learn the abstract data and reconstruct it into the time-series feature data. The output layer includes sequentially connected fully connected layers and a regression output layer, and the output layer is used to output the reconstructed time-series feature data as denoised TEM data.
[0018] In some preferred embodiments, the method for training the noise suppression network after training is as follows: The training set and the test set are preprocessed, and the preprocessing method includes normalization and folding / reconstruction. The pre-processed training set is input into a pre-built noise suppression network to obtain initial denoising data, the initial parameters of the noise suppression network are adjusted based on the initial denoising data and the actual labels to obtain optimal parameters, and the initial parameters of the noise suppression network are replaced with the optimal parameters. The pre-processed training set is input again in batches into the noise suppression network with optimal parameters to obtain de-noised training data, and multiple iterations are performed. In each iteration, the mean squared error between the de-noised training data and the actual labels is calculated, and the iterations are stopped until the mean squared error converges. The noise suppression network after the iteration is stopped is referred to as the noise suppression network after training.
[0019] In some preferred embodiments, during multiple iterations, the pre-processed test set is simultaneously input to the noise suppression network having optimal parameters to obtain test data, the test data is compared to the corresponding actual labels to obtain a verification error, and the iteration is stopped when the verification error reaches convergence and then begins to expand again.
[0020] In some preferred embodiments, an Adam optimizer is used in each iteration to adjust the parameters of the noise suppression network and minimize the mean squared error between the training data and the actual labels until the mean squared error converges.
[0021] In some preferred embodiments, the method of folding and reconstructing is as follows: The training set and test set of one-dimensional data are folded and reconstructed into two-dimensional data, and this two-dimensional data is used as input to a pre-constructed noise suppression network.
[0022] As another aspect of the present invention, we propose a noise suppression system for transient electromagnetic data based on a noise suppression method used for transient electromagnetic data. The system in question is The system includes a network construction and training module, which stacks forward modeling data and noise data based on different weights to obtain a dataset, trains a pre-built noise suppression network on the dataset, and obtains a trained noise suppression network, where the dataset includes a training set and a test set, and the different weights are obtained based on a preset target signal-to-noise ratio. It includes a noise reduction module, inputs the noise-suppressed TEM data into a trained noise suppression network, and obtains noise-suppressed TEM data. Here, the noise suppression network was constructed based on a recurrent neural network and an LSTM network. [Effects of the Invention]
[0023] The beneficial effects of the present invention are as follows: (1) Excellent noise reduction effect on contaminated late data. The noise suppression network based on the LSTM-Autoncoder neural network that we constructed has an excellent noise reduction effect on contaminated late TEM data, significantly improving the efficiency of using late TEM data and increasing the depth of exploration. (2) Highly versatile and capable of combating various types of noise. Our method excels at processing commercial frequency noise that is difficult to eliminate with traditional methods. We have achieved precise rejection of common commercial frequency noise such as square waves, charge / discharge triangular waves, power frequency interference, step waves, and spike pulses. This innovation demonstrates greater adaptability and practicality in complex real-world environments. (3) Efficiency is improved by intelligent noise reduction. By avoiding the selection of artificial noise reduction parameters such as wavelet noise reduction, variational mode decomposition, and singular value decomposition, and other complex manual operations, its simple and efficient noise reduction efficiency far surpasses that of other methods. [Brief explanation of the drawing]
[0024] Further features, purposes, and advantages of the present application will become clearer by reading the detailed description of non-limiting embodiments made with reference to the following drawings. [Figure 1] This is a flowchart of the noise suppression method used for transient electromagnetic data in the present invention. [Figure 2] This is a schematic diagram of the training process of the noise suppression network used in the noise suppression method for transient electromagnetic data of the present invention. [Figure 3] This is a diagram showing the configuration of the noise suppression network for the noise suppression method used for transient electromagnetic data in the present invention. [Figure 4] This is a diagram illustrating the configuration of a computer system for realizing a server according to an embodiment of the method, system, or apparatus of this application. [Modes for carrying out the invention]
[0025] The present application will be described in further detail below with reference to the attached drawings and embodiments. It is understood that the specific embodiments described herein are not intended to limit the present invention, but are used solely to illustrate related inventions. For the sake of clarity, only the parts relating to the invention are shown in the drawings.
[0026] In cases where there is no conflict, the embodiments and features described herein may be combined with each other. The present invention will now be described in detail with reference to the drawings, in relation to the embodiments.
[0027] As a first embodiment of the present invention, please refer to Figures 1 to 3, and we propose a noise suppression method used for transient electromagnetic data. This method is Forward modeling data and noise data are stacked based on different weights to form a dataset. A pre-built noise suppression network is then trained on this dataset to obtain a trained noise suppression network, where the dataset includes a training set and a test set, and the different weights are obtained based on a preset target signal-to-noise ratio. The noise-suppressed TEM data is input into a trained noise suppression network, and the noise-suppressed TEM data is obtained. Here, the noise suppression network was constructed based on a recurrent neural network and an LSTM network.
[0028] In this embodiment, the network used is preferably a recurrent neural network (RNN), a neural network structure primarily used for time series processing. Based on conventional backpropagation neural networks, a time backpropagation algorithm is introduced to realize time-point connections in the hidden layer. The output of the hidden layer neurons depends not only on the input at the current time but also on the output of the hidden layer neurons at the previous time. This design allows the RNN to make full use of historical information in the time series while maintaining a certain level of memory and processing time-related data.
[0029] When information sequences are long, RNNs are prone to dependency problems over the long term, and it is difficult to efficiently space out distant sequence information using time. To overcome the vanishing gradient and gradient explosion problems, this invention optimizes the network configuration using LSTM. LSTM stabilizes gradient changes by adding cell states and information gating. Cell states are long-span information transmission channels that operate directly throughout the entire information transmission chain and involve only a small amount of linear computation, while gating selectively passes information through. The combined action of these two mitigates the long-term dependency problems of RNNs.
[0030] LSTM first uses a sigmoid function σ and a weight function W. f The forget gate f is calculated t By obtaining this, we control the information that needs to be forgotten in the cell state.
[0031]
number
[0032] Here, h t-1 ,xt The network-wide input is b f This is the first offset, calculated f t This is a forgetting gate, which takes a value between 0 and 1, and represents the cell state C at the previous time step. t-1 Determine the information that needs to be forgotten.
[0033]
number
[0034] Furthermore, Oblivion Gate f t and the cell state C at the previous time point t-1 When multiplied by C, t-1 Determine which information needs to be forgotten.
[0035] Similarly, i t The following equation (equation 6) both determine the information that needs to be held in the total input of the memory gate.
[0036]
number
[0037]
number
[0038] The aforementioned h t-1 ,x t The network-wide input is b i This is the second offset, the sigmoid function σ and the first weight function W. i Long-term memory i t Calculate.
[0039]
number
[0040] The aforementioned h t-1 ,x t The network-wide input is b cThis is the third offset, the tanh function and the second weight function W. c This calculates short-term memory (number 9).
[0041]
number
[0042]
number
[0043] long term memory i t This is multiplied by short-term memory (number 11) to obtain a memory gate, and the part that needs to be retained in the cell's state is determined.
[0044]
number
[0045] Furthermore, the forgetting gate and memory gate are added together to update the cell state value C. t Having obtained,
[0046]
number
[0047] Ultimately, O t The gate is set, and the output value of the neuron is h t We decided,
[0048]
number
[0049] The aforementioned h t-1 ,x t The network-wide input is b o The fourth offset, the sigmoid function, and the third weight function W o Output gate O t The system calculates the necessary values and determines which parts of the cell's state need to be output.
[0050]
number
[0051] And the updated cell state C t The result is processed by the tanh function and output gate O t Multiply by this to obtain the output value of the hidden layer at that point.
[0052] To more clearly explain the noise suppression method used for transient electromagnetic data in the present invention, embodiments of the present invention will be described in detail below with reference to Figure 1. The details are as follows: Forward modeling data and noise data are stacked based on different weights to form a dataset. A pre-built noise suppression network is then trained on this dataset to obtain a trained noise suppression network, where the dataset includes a training set and a test set, and the different weights are obtained based on a preset target signal-to-noise ratio. In the present invention, the method for constructing the dataset is as follows: Construct multiple different one-dimensional resistivity models, perform forward modeling, and obtain forward modeling data corresponding to the multiple one-dimensional resistivity models. Noise data of different types and intensities is artificially simulated, and the noise data includes atmospheric noise, commercial frequency noise, and Gaussian noise. Multiple forward modeling data and multiple noise data are stacked with different weights, and the stacked data is used as a dataset. The aforementioned dataset is divided into a training set and a test set according to a predetermined ratio.
[0053] Here, the method for obtaining the aforementioned dataset is shown by the following formula:
[0054]
number
[0055] Here, D final This is the accumulated dataset, D i forward This is the i-th forward modeling data, D i noise ω is the i-th noise data. forward and ω noise The weights are the forward modeling data and noise data, respectively, and are used to adjust the ratio of forward modeling data to noise data when stacking them, with N being the amount of data in the forward modeling data. By adjusting the weight values, composite data with different signal-to-noise ratios can be obtained. The signal-to-noise ratio (SNR) represents the relative intensity of the signal and noise, and is usually expressed in decibels (dB). The weight values can be adjusted according to the required signal-to-noise ratio. Specifically, assuming a preset target signal-to-noise ratio (in decibels), the corresponding weight values can be calculated using the following formula: ω forward and ω noise It is obtained by the following method,
[0056]
number
[0057]
number
[0058] Here, in this way, different SNRs target By setting this, you can obtain the corresponding weight values and generate composite data with even different signal-to-noise ratios. For example, if you want to obtain data with a higher signal-to-noise ratio, you can use a larger SNR. target You can choose this option, and conversely, if you want to obtain data with a lower signal-to-noise ratio, you can choose a smaller SNR. targetThis allows for flexible control over the signal-to-noise ratio of the synthesized data.
[0059] Specifically, training a neural network requires constructing a dataset with a sufficiently large number of samples. However, since ideal TEM data without natural sounds does not exist, we train an LSTM-Autoncoder neural network by combining TEM data obtained from one-dimensional forward modeling with an artificially simulated noise synthesis dataset. We constructed 1000 different one-dimensional resistivity models and forward-modeled 1000 forward-modeling datasets. These forward-modeling datasets are transient data obtained at a 100MHz sampling rate over a time of 0 to 0.001s, each with 1000 sampling points. We also artificially simulated different types and intensities of noise, including atmospheric noise, commercial frequency noise, and Gaussian noise. We stacked mixed noise of different interference intensities with pure forward-modeling data to obtain 20,000 noise-containing data samples, which were divided into a training set and a test set in a 4:1 ratio. That is, 16,000 samples for the training set and 4,000 samples for the test set.
[0060] Refer to Figure 3; the pre-built noise suppression network of the present invention includes sequentially connected input layers, encoders, decoders, and output layers. The input layer is used to receive input data including the dataset or the noise-suppressing TEM data. Both the encoder and decoder are constructed using two sequentially connected LSTM layers, and the encoder is used to learn the time-series feature data of the input data and convert it into abstract data. The decoder is used to learn the abstract data and reconstruct it into the time-series feature data. The output layer includes sequentially connected fully connected layers and a regression output layer, and the output layer is used to output the reconstructed time-series feature data as denoised TEM data.
[0061] In this embodiment, the two LSTM layers of the encoder have 128 and 64 neurons respectively, and the two LSTM layers of the decoder have 64 and 128 neurons respectively.
[0062] Specifically, please refer to Figure 3. The two LSTM layers of the encoder are LSTM 1 and LSTM 2, respectively, and the two LSTM layers of the decoder are LSTM 3 and LSTM 4, respectively.
[0063] Please refer to Figure 2. The training method for the noise suppression network after training according to the present invention is as follows: The training set and the test set are preprocessed, and the preprocessing method includes normalization and folding / reconstruction. The pre-processed training set is input into a pre-built noise suppression network to obtain initial denoising data, the initial parameters of the noise suppression network are adjusted based on the initial denoising data and the actual labels to obtain optimal parameters, and the initial parameters of the noise suppression network are replaced with the optimal parameters. The pre-processed training set is input again in batches into the noise suppression network with optimal parameters to obtain de-noised training data, and multiple iterations are performed. In each iteration, the mean squared error between the de-noised training data and the actual labels is calculated, and the iterations are stopped until the mean squared error converges. The noise suppression network after the iteration is stopped is referred to as the noise suppression network after training.
[0064] Here, the method for folding and reconstructing is as follows: The training set and test set of one-dimensional data are folded and reconstructed into two-dimensional data, and this two-dimensional data is used as input to a pre-constructed noise suppression network.
[0065] Here, the input dataset is first normalized due to the peculiarities of data exponential decay. The data in the dataset is then logarithmed, and maximum / minimum normalization is performed. In the following equation, x is the dataset data (raw data), and log x mean This is the logarithmic mean of the dataset data, logx max This is the maximum value in the dataset, logx min x' is the minimum value in the dataset, and x' is the normalized data.
[0066]
number
[0067] After normalizing the data, and considering that the training effect is relatively poor when 1D data is input into a neural network, each data point is folded and reconstructed into 2D data before being input into the network. In other words, 1000 x 1 1D data is converted into 40 x 25 x 1 2D data.
[0068] In this invention, after data preprocessing is complete, the model is preliminaryly trained and model parameters are adjusted. The input and output dimensions are fixed, and other parameters are typically adjusted based on training and validation effects to find the optimal values. After parameter adjustment, the model is trained. During training, data is input in batches, and the MSE (Most Common Semantic Exponential) of the training denoised data and the actual labels is calculated in each iteration. The model parameters are then optimized using an Adam optimizer until the MSE converges. Finally, the trained model is saved.
[0069] As a further interpretation of the present invention, the training effect must be verified during the model training process to avoid overfitting. Specifically, during multiple iterations, the pre-processed test set is simultaneously input to the noise suppression network having optimal parameters to obtain test data, which is then compared to the corresponding actual labels to obtain a verification error. When the verification error converges and then begins to expand again, it indicates that the network has reached an overfitting state, and the iteration is stopped.
[0070] As a further explanation of the present invention, when calculating the mean squared error between the training denoised data and the actual labels in each iteration, the parameters of the noise suppression network are adjusted using an Adam optimizer in each iteration to minimize the mean squared error between the training data and the actual labels until the mean squared error converges.
[0071] In the above embodiment, the steps were described in the order described above. However, those skilled in the art will understand that in order to achieve the effects of this embodiment, it is not necessary to perform the steps in this order between different steps; they may be performed simultaneously (in parallel) or in reverse order. All these simple modifications are within the scope of protection of the present invention.
[0072] A second embodiment of the present invention is a noise suppression system for transient electromagnetic data based on a noise suppression method used for transient electromagnetic data, The system in question is The system includes a network construction and training module, which stacks forward modeling data and noise data based on different weights to obtain a dataset, trains a pre-built noise suppression network on the dataset, and obtains a trained noise suppression network, where the dataset includes a training set and a test set, and the different weights are obtained based on a preset target signal-to-noise ratio. It includes a noise reduction module, inputs the noise-suppressed TEM data into a trained noise suppression network, and obtains noise-suppressed TEM data. Here, the noise suppression network was constructed based on a recurrent neural network and an LSTM network.
[0073] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific operating procedures and descriptions of the system described above can be found by referring to the corresponding procedures in the embodiments of the method described above, and will not be explained further here.
[0074] The above embodiments provide a noise suppression system for transient electromagnetic data, explained using only the classification of each functional module as an example. In actual applications, the distribution of the above functions can be carried out by different functional modules as needed. That is, the modules or steps in the embodiments of the present invention can be further disassembled or combined. For example, the modules of the above embodiments may be integrated into a single module, or further divided into multiple submodules to complete all or part of the above functions. The names of the modules and steps in the embodiments of the present invention are for the sole purpose of distinguishing each module or step and are not considered an unreasonable limitation of the present invention.
[0075] An electronic device according to a third embodiment of the present invention is At least one processor, and It includes at least one processor and memory connected to it, Here, The memory stores instructions to be executed by the processor, and these instructions execute the processor to implement the noise suppression method used for transient electromagnetic data as described above.
[0076] A computer-readable storage medium according to a fourth embodiment of the present invention, wherein the computer stores computer instructions, and the computer instructions are executed by the computer to implement the noise suppression method used for transient electromagnetic data described above.
[0077] For the sake of clarity and brevity, the specific operating procedures and descriptions of the above-mentioned storage device and processing device can be found by referring to the corresponding procedures in the embodiments of the method described above, and are therefore omitted here.
[0078] Those skilled in the art will recognize that each example module and method step described in relation to the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both, and the programs corresponding to the software modules and method steps may be in random access memory (RAM), memory, read-only memory (ROM), electrically programmable ROM, electrically erasable and programmable ROM, registers, hard disks, removable media, CD-ROMs, or any other form of storage medium known in the art. To clearly illustrate the compatibility of electronic hardware and software, the configuration and steps of each example are described functionally in general terms above. These functions are performed in electronic hardware or software depending on the specific application and design constraints of the technology. Those skilled in the art will recognize that different methods may be used for each specific application to implement the described functions, but such implementations are not considered to be beyond the scope of the invention.
[0079] Figure 4 below shows a configuration diagram of a computer system for realizing a server according to an embodiment of the method, system, and apparatus of this application. The server shown in Figure 4 is merely an example and does not limit the functionality and scope of use of the embodiments of this application.
[0080] As shown in Figure 4, the computer system includes a central processing unit (CPU) 401 that can perform various appropriate operations and processes according to programs stored in read-only memory (ROM) 402 or programs loaded from memory unit 408 into random access memory (RAM) 403. RAM 403 also stores various programs and data necessary for the operation of the system. The CPU 401, ROM 402, and RAM 403 are connected to each other via a bus 404. An input / output (I / O) interface 405 is also connected to the bus 404.
[0081] The I / O interface 405 includes an input unit 406, which includes a keyboard and mouse; an output unit 407, which includes speakers, which include a cathode ray tube (CRT) and a liquid crystal display (LCD); a storage unit 408, which includes a hard disk; and a communication unit 409, which includes a network interface card such as a LAN (Local Area Network) card and a modem. The communication unit 409 performs communication processing via a network such as the Internet. The driver 410 is also connected to the I / O interface 405 as needed. Removable media 411, such as magnetic disks, optical disks, magneto-optical disks, and semiconductor memory, are installed on the drive 410 as needed, and the computer programs read from the drive are installed on the storage unit 408 as needed.
[0082] In particular, according to embodiments of the present disclosure, the process described above can be implemented as a computer software program with reference to a flowchart. For example, an embodiment of the present disclosure includes a computer program product, which includes a computer program carried on a computer-readable medium, and which includes program code for performing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network via a communication unit 409 and / or installed from removable media 411. When this computer program is executed by a central processing unit (CPU) 401, the above-described functions, limited to the method of the present application, are performed. The computer-readable medium described above may be a computer-readable signal medium, a computer-readable storage medium, or any combination of the two. The computer-readable storage medium may be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of more than these. The computer-readable medium described above may be a computer-readable signal medium, a computer-readable storage medium, or any combination of the two. Computer-readable storage media can be, but are not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any combination of more than these. More specific examples of computer-readable storage media include, but are not limited to, electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), writable and erasable ROM (EPROM or flash memory), optical fibers, compact disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the above.In this application, a computer-readable storage medium may be any tangible medium containing or storing a program that can be instructed to run a system, apparatus, or device, or used in combination with them. In this application, a computer-readable storage medium may be any tangible medium containing or storing a program that can be instructed to run a system, apparatus, or device, or used in combination with them. On the other hand, in this application, a computer-readable signal medium may include data signals propagating in a baseband carrying computer-readable program code, or as part of a carrier wave. Such propagating data signals may take various forms, including, but not limited to, electromagnetic signals, optical signals, or any suitable combination described above. A computer-readable signal medium may be any computer-readable medium other than a computer-readable storage medium that can transmit, propagate, or transmit a program that is used by, or used in combination with, an instruction execution system, apparatus, or device. The program code contained in the computer-readable medium may be transmitted over any suitable medium, including, but not limited to, wireless, wire, optical cable, RF, or any suitable combination described above.
[0083] The computer program code for performing the operations of this invention may be written in one or more programming languages or a combination thereof. The programming languages mentioned above include Java, Smalltalk, C++, object-oriented programming languages such as the "C" language, and general procedural programming languages. The program code may run entirely on the user's computer, partially on the user's computer, run as a standalone package, run partially on the user's computer and partially on a remote computer, or run entirely on a remote computer or server. In the case of a remote computer, the remote computer may connect to the user's computer via any type of network, including a local area network (LAN) or wide area network (WAN), or it may connect to an external computer (for example, via the Internet using an Internet service provider).
[0084] Flowcharts and block diagrams in the drawings illustrating the feasible architectures, functions, and operations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, program segment, or part of code containing one or more executable instructions for realizing a given logical function. It should also be noted that in alternative implementations, the functions shown in the blocks may occur in an order different from the order shown in the drawings. For example, two consecutively shown blocks may actually be executed essentially in parallel, or in reverse order depending on the related functions. Furthermore, each block in a block diagram and / or flowchart, and combinations of blocks in a block diagram and / or flowchart, can be realized in a dedicated hardware-based system that performs a given function or operation, or in a combination of dedicated hardware and computer instructions.
[0085] Terms such as "first," "second," etc., are not intended to describe or represent a specific order or priority, but rather to distinguish similar objects.
[0086] The term “includes” or any similar term is intended to include non-exclusive inclusion, such that a process, method, article or apparatus / device that includes a set of elements also includes other elements not expressly listed, or also includes elements specific to those processes, methods, articles or apparatus.
[0087] Although technical aspects of the present invention have been described above in relation to preferred embodiments shown in the accompanying drawings, it will be readily apparent to those skilled in the art that the scope of protection of the present invention is clearly not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent modifications or substitutions to the relevant technical features, and all such modified or substituted technical proposals fall within the scope of protection of the present invention.
Claims
1. A noise suppression method used for transient electromagnetic data, This method is Forward modeling data and noise data are stacked based on different weights to form a dataset. A pre-built noise suppression network is then trained on this dataset to obtain a trained noise suppression network, where the dataset includes a training set and a test set, and the different weights are obtained based on a preset target signal-to-noise ratio. The TEM data to be noise-suppressed is input into a trained noise suppression network, and the TEM data after noise suppression is obtained. Here, the noise suppression network is constructed based on a recurrent neural network and an LSTM network. The method for obtaining the aforementioned dataset is shown by the following formula: [Number 19] Here, D ｆｉｎａｌ D is a stacked dataset. ｉ ｆｏｒｗａｒｄ D is the i-th forward modeling data. ｉ ｎｏｉｓｅ ω is the i-th noise data. ｆｏｒｗａｒｄ and ω ｎｏｉｓｅ These are the weight values for forward modeling data and noise data, respectively, used to adjust the ratio of forward modeling data to noise data when stacking them, and N is the amount of data in the forward modeling data. ω ｆｏｒｗａｒｄ and ω ｎｏｉｓｅ is obtained by the following method, [Number 20] [Math 21] Here, SNR ｔａｒｇｅｔ A noise suppression method used for transient electromagnetic data, characterized in that the target signal-to-noise ratio is a preset.
2. The method for constructing the aforementioned dataset is as follows: Construct multiple different one-dimensional resistivity models, perform forward modeling, and obtain forward modeling data corresponding to the multiple one-dimensional resistivity models. Noise data of different types and intensities is artificially simulated, and the noise data includes atmospheric noise, commercial frequency noise, and Gaussian noise. Multiple forward modeling data and multiple noise data are stacked with different weights, and the stacked data is used as a dataset. The noise suppression method for transient electromagnetic data according to claim 1, characterized in that the dataset is divided into a training set and a test set in a set ratio.
3. The aforementioned pre-constructed noise suppression network includes sequentially connected input layers, encoders, decoders, and output layers. The input layer is used to receive input data including the dataset or the noise-suppressing TEM data. Both the encoder and decoder are constructed from two sequentially connected LSTM layers, and the encoder is used to learn the time-series feature data of the input data and convert it into abstract data. The decoder is used to learn the abstract data and reconstruct it into the time-series feature data. The noise suppression method for transient electromagnetic data according to claim 1, characterized in that the output layer includes sequentially connected fully connected layers and a regression output layer, and the output layer is used to output the restored time-series feature data as denoised TEM data.
4. The training method for the noise suppression network after the aforementioned training is as follows: The training set and the test set are preprocessed, and the preprocessing method includes normalization and folding / reconstruction. The pre-processed training set is input into a pre-built noise suppression network to obtain initial denoising data, the initial parameters of the noise suppression network are adjusted based on the initial denoising data and the actual labels to obtain optimal parameters, and the initial parameters of the noise suppression network are replaced with the optimal parameters. The pre-processed training set is input again in batches into the noise suppression network with optimal parameters to obtain de-noised training data, and multiple iterations are performed. In each iteration, the mean squared error between the de-noised training data and the actual labels is calculated, and the iterations are stopped until the mean squared error converges. The noise suppression method for transient electromagnetic data according to claim 3, characterized in that the noise suppression network after the iteration is stopped is the noise suppression network after training.
5. The noise suppression method for transient electromagnetic data according to claim 4, characterized in that, during the process of multiple iterations, the pre-processed test set is input to the noise suppression network having optimal parameters to obtain test data, the test data is compared with the corresponding actual label to obtain a verification error, and the iteration is stopped when the verification error reaches convergence and then begins to expand again.
6. The noise suppression method for transient electromagnetic data according to claim 5, characterized in that the parameters of the noise suppression network are adjusted using an Adam optimizer in each iteration to minimize the mean squared error between the training data and the actual labels until the mean squared error converges.
7. The method for folding and reconstructing is as follows: The noise suppression method for transient electromagnetic data according to claim 6, characterized in that the training set and test set of one-dimensional data are folded and reconstructed into two-dimensional data, and the two-dimensional data is used as input to a pre-constructed noise suppression network.
8. A noise suppression system for transient electromagnetic data based on a noise suppression method for transient electromagnetic data according to any one of claims 1 to 7, The system in question is The system includes a network construction and training module, which stacks forward modeling data and noise data based on different weights to obtain a dataset, trains a pre-built noise suppression network on the dataset, and obtains a trained noise suppression network, where the dataset includes a training set and a test set, and the different weights are obtained based on a preset target signal-to-noise ratio. It includes a noise reduction module, inputs the TEM data to be noise-suppressed into a trained noise suppression network, and obtains the TEM data after noise suppression. Here, the noise suppression network is constructed based on a recurrent neural network and an LSTM network. The method for obtaining the aforementioned dataset is shown by the following formula: [Number 22] Here, D ｆｉｎａｌ D is a stacked dataset. ｉ ｆｏｒｗａｒｄ D is the i-th forward modeling data. ｉ ｎｏｉｓｅ ω is the i-th noise data. ｆｏｒｗａｒｄ and ω ｎｏｉｓｅ These are the weight values for forward modeling data and noise data, respectively, used to adjust the ratio of forward modeling data to noise data when stacking them, and N is the amount of data in the forward modeling data. ω ｆｏｒｗａｒｄ and ω ｎｏｉｓｅ It is obtained by the following method, [Number 23] [Number 24] Here, SNR ｔａｒｇｅｔ A noise suppression system used for transient electromagnetic data, characterized by a preset target signal-to-noise ratio.