Marine magnetotelluric data denoising method and related equipment

Abstract

The invention relates to the technical field of geophysical exploration signal processing, and discloses an ocean magnetotelluric data denoising method and related equipment, and the method comprises the steps: constructing a training data set, and training a model to obtain a dual-path noise recognition model and a generative reconstruction model; performing preprocessing and windowing on the new ocean magnetotelluric original data to be processed to obtain a plurality of data windows; performing parallel noise identification on each data window to obtain a periodic noise mask and a random pulse noise mask; carrying out fusion processing to obtain a comprehensive noise mask, and marking noise data points needing to be reconstructed in the data window by the comprehensive noise mask; and performing signal reconstruction on the data window based on the generative reconstruction model and the comprehensive noise mask to obtain a processed clean data window, and performing splicing processing to obtain a complete de-noised ocean magnetotelluric time sequence. According to the method, the defects of an existing MMT data processing technology in the aspect of coping with long-period and multi-type mixed noise are overcome.

Description

Technical Field

[0001] This invention relates to the field of geophysical exploration signal processing technology, specifically to a method and related equipment for denoising marine magnetotelluric data. Background Technology

[0002] Marine magnetotellurics (MMT) is one of the core technologies for deep-sea geophysical exploration, and the quality of its data directly affects the success or failure of seabed oil and gas resource exploration, natural gas hydrate investigation, and deep geological structure research. However, raw MMT time series data is often severely contaminated by various complex noises.

[0003] Noise from terrestrial electromagnetic exploration primarily originates from high-frequency, strong electromagnetic interference generated by human activities (such as power frequency interference and high-voltage lines). However, the noise characteristics of the marine environment are quite different. Deep seawater has a strong shielding and attenuation effect on high-frequency electromagnetic waves, significantly reducing the strong electromagnetic interference commonly found on land. Instead, more complex noise types unique to the marine environment are generated, mainly including: Random impulse noise originates from transient electromagnetic radiation from passing ships, discharge of marine organisms, and abrupt changes in the contact state between electrodes and sediments. It is highly random and sudden, and manifests as isolated high-amplitude spikes in the time domain.

[0004] Long-period noise: This type of noise is particularly challenging in MMT data, as it has diverse sources, complex forms, and periods that can extend from tens of minutes to several hours. Its sources include disturbances to the attitude of seabed instruments caused by tidal movements, temperature drift of the instruments themselves, noise generated by reading and writing disks, and repetitive electromagnetic interference in specific marine environments.

[0005] Meanwhile, ocean electromagnetic observation is characterized by high sampling rates (typically greater than 100Hz) and ultra-long observation periods (lasting from several days to several months), with a single mission generating tens of gigabytes of data. This massive data characteristic poses a severe challenge to traditional processing methods.

[0006] Currently, most methods used in the industry to process noise in MMT data are directly adopted or modified from land-based electromagnetic processing techniques, such as far-reference channel suppression, wavelet transform, and median filtering. These methods have fundamental shortcomings when dealing with ocean-specific noise: Low-frequency information loss: For long-period noise, traditional frequency domain filtering methods (such as band-stop or low-pass filters) struggle to achieve a steep transition between the passband and stopband when processing low-frequency signals. Since much important deep geological information is also reflected in the low-frequency part of the signal, using filters inevitably removes both the effective signal and the noise, resulting in permanent loss of deep information and severely affecting the depth and accuracy of subsequent inversions.

[0007] Limitations of Model Assumptions: Traditional methods are mostly based on ideal prior models of signals or noise. For example, simple fitting and stripping methods (such as polynomial fitting) can only handle simple trend noise, and are powerless against complex non-sinusoidal periodic noise. In addition, after removing noisy data segments, the commonly used linear or spline interpolation methods are essentially mathematical filling and cannot restore the geophysical laws inherent in the signal. When the noise segment is long, the interpolation results will be severely distorted, destroying the continuity and spectral structure of the signal.

[0008] Processing efficiency bottleneck: Traditional methods rely on a lot of manual interaction and parameter tuning, which is inefficient when faced with massive amounts of high sampling rate data and makes it difficult to achieve standardized and automated processing.

[0009] In summary, the existing technological system lacks a new automated paradigm that can intelligently and efficiently process the long-period and random impulse noise unique to the ocean, restore the physical characteristics of the signal with high fidelity after denoising, and simultaneously handle massive data processing tasks. Summary of the Invention

[0010] The purpose of this invention is to provide a method and related equipment for denoising marine magnetotelluric data, so as to solve the technical problem that existing MMT data processing technology is insufficient in dealing with long-period, multi-type mixed noise.

[0011] This invention is achieved through the following technical solution: In a first aspect, the present invention provides a method for denoising marine magnetotelluric data, comprising: A training dataset was constructed and the model was trained to obtain a dual-path noise recognition model and a generative reconstruction model; The new raw marine magnetotelluric data to be processed is preprocessed and windowed to obtain multiple data windows; Parallel noise identification is performed on each data window based on a dual-path noise identification model to obtain periodic noise masks and random impulse noise masks. A composite noise mask is obtained by fusing periodic noise mask and random impulse noise mask, and the composite noise mask marks the noise data points that need to be reconstructed in the data window; The data window is reconstructed based on a generative reconstruction model and a comprehensive noise mask to obtain a clean data window after processing. All clean data windows were stitched together to obtain the complete denoised marine magnetotelluric time series.

[0012] Preferably, constructing the training dataset includes obtaining pairs of noisy marine magnetotelluric time series X_noisy and corresponding clean marine magnetotelluric time series Y_clean; and segmenting the pairs of noisy marine magnetotelluric time series X_noisy and corresponding clean marine magnetotelluric time series Y_clean into fixed-length windows with overlapping parts to form training sample pairs.

[0013] Furthermore, the dual-path noise recognition model is used to identify periodic noise and random impulse noise. The specific process of training the dual-path noise recognition model includes training the periodic noise recognition model and training the random impulse noise recognition model: The periodic noise recognition model training includes constructing a one-dimensional convolutional neural network (1D-CNN), taking a pair of noisy marine magnetotelluric time series X_noisy as input, and using a periodic noise mask Mask_p generated by calculating the difference between |X_noisy-Y_clean| and combining it with frequency domain analysis as a label for training. In the periodic noise mask Mask_p, the long-period noise positions are marked as 1, and the rest are marked as 0. The training of the random impulse noise identification model includes constructing an autoencoder network, using the corresponding clean ocean magnetotelluric time series Y_clean as input and training target for training, so that the model learns the characteristics of normal ocean magnetotelluric signals, and the autoencoder identifies random impulse noise through reconstruction error; The generative reconstruction model is used to reconstruct signals in noisy regions based on contextual information. The specific process of training the generative reconstruction model includes: constructing a sequence generation model, which is a bidirectional long short-term memory network or a Transformer self-attention mechanism deep learning model; taking X_masked, obtained by masking a pair of noisy marine magnetotelluric time series X_noisy, as input, and using the corresponding clean marine magnetotelluric time series Y_clean as label for training, so that the model can recover the clean signal based on contextual information.

[0014] Preferably, the preprocessing and windowing method is consistent with the window segmentation method when constructing the training dataset, that is, segmenting into fixed-length windows with overlapping parts.

[0015] Preferably, each data window is input into a trained 1D-CNN model to obtain a predicted periodic noise mask M_p; each data window is input into a trained autoencoder model, and a random impulse noise mask M_i is obtained by calculating the reconstruction error and setting a threshold.

[0016] Preferably, the fusion process involves performing a logical OR operation on the periodic noise mask M_p and the random impulse noise mask M_i to generate the final composite noise mask M_total, which marks all noise data points that need to be reconstructed within the current window.

[0017] Preferably, the signal reconstruction of the data window is performed based on the generative reconstruction model and the comprehensive noise mask. The signal reconstruction includes masking the data window based on the comprehensive noise mask M_total, and inputting the processed sequence into the generative reconstruction model to obtain a clean data window filled with noise regions.

[0018] Preferably, the splicing process involves performing a smooth weighted average of the overlapping areas of all clean data windows before splicing.

[0019] Secondly, the present invention also provides a denoising system for marine magnetotelluric data, comprising: The data model training module is used to build a training dataset and train the model to obtain a dual-path noise recognition model and a generative reconstruction model. The data preprocessing windowing module is used to preprocess and window new raw marine magnetotelluric data to be processed, resulting in multiple data windows. The parallel noise identification module is used to perform parallel noise identification on each data window based on the dual-path noise identification model to obtain periodic noise mask and random impulse noise mask; The mask fusion processing module is used to perform fusion processing based on periodic noise mask and random impulse noise mask to obtain a comprehensive noise mask, wherein the comprehensive noise mask marks the noise data points to be reconstructed in the data window; The signal reconstruction processing module is used to reconstruct the signal of the data window based on the generative reconstruction model and the comprehensive noise mask, so as to obtain a clean data window after processing. The data window stitching module is used to stitch together all clean data windows to obtain a complete denoised marine magnetotelluric time series.

[0020] Thirdly, the present invention also provides a mobile terminal, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method for denoising marine magnetotelluric data.

[0021] Compared with the prior art, the present invention has the following beneficial technical effects: This invention provides a denoising method for marine magnetotelluric (MMT) data. By constructing a dual-path noise identification model, it specifically trains and identifies long-period noise and random impulse noise, which are unique to and difficult to process in the ocean, respectively. This accurately generates two types of noise masks, achieving precise localization of complex marine noise and avoiding the limitations of single models in adapting to different noise types. A generative reconstruction model combined with a comprehensive noise mask is used for signal reconstruction. The model, trained on massive amounts of data, grasps the inherent physical laws and time dependencies of MMT signals, and performs high-fidelity recovery of noisy regions based on contextual information. This effectively avoids the loss of low-frequency effective signals caused by traditional frequency domain filtering, ensuring the integrity of deep geological information. The entire denoising process is divided into two stages: offline model training and online data processing. During online processing, preprocessing windowing, parallel noise identification, automated reconstruction, and stitching eliminate the need for extensive manual interaction and parameter tuning, significantly improving processing efficiency. By using noise masking fusion processing, all noise data points that need to be reconstructed are accurately marked, ensuring that the reconstruction is performed only on the noise area, thus preserving the effective information in the original signal to the greatest extent. At the same time, when splicing data windows, overlapping areas are smoothed and weighted averaged to ensure the continuity and integrity of the output time series, which greatly improves the accuracy and reliability of deep-sea geophysical exploration.

[0022] Furthermore, this invention achieves a paradigm shift from model-driven to data-driven approaches, solving the challenge of separating long-period noise from effective signals. Traditional methods rely on prior mathematical models, which cannot handle complex ocean noise with unknown models. This invention utilizes supervised learning to directly learn the complex nonlinear characteristics of noise from a large amount of real data, particularly learning to distinguish the subtle differences between long-period noise fluctuations and the true electrical response in the time domain. This allows for the accurate identification and removal of noise without compromising the effective signal in the same frequency band. This represents a fundamental subversion of traditional filtering methods.

[0023] Furthermore, this invention reconstructs signals with high fidelity, preserving deep geological information to the maximum extent. This invention abandons traditional mathematical interpolation and employs advanced generative models (such as Bi-LSTM) for signal reconstruction. This model learns the inherent laws of massive amounts of pure signals, and its reconstructed data is not only numerically continuous but also conforms to the physical propagation characteristics of magnetotelluric signals in terms of phase, spectrum, and time dependence. This ensures that the reconstructed data is physically accurate rather than mathematically smoothed, providing highly reliable input for subsequent inversion interpretation, directly impacting the success or failure of deep resource exploration.

[0024] Furthermore, the processing flow of this invention is highly automated and intelligent, significantly improving the efficiency of processing massive amounts of data. Once the model has completed offline training, the entire denoising process requires no manual intervention or tedious parameter adjustments for new massive amounts of data. This "train once, apply anywhere" model liberates data processing personnel from repetitive and inefficient labor, enabling rapid and standardized processing of terabyte-level data, greatly shortening project cycles and reducing the risk of human error.

[0025] Furthermore, this invention provides a highly adaptable and standardized intelligent processing framework, filling a technological gap in China. Through its data-driven modeling process, this invention can quickly construct a dedicated denoising model for the unique noise of a new survey area using only a small amount of calibration data. This method eliminates the heavy reliance on manual experience in traditional processing, ensuring optimal processing results under different exploration environments. It achieves standardization and engineering in data processing, and its core technological ideas can also provide a reference for other geophysical exploration fields. Attached Figure Description

[0026] Figure 1 This is a flowchart of a method for denoising marine magnetotelluric data in an embodiment of the present invention; Figure 2 This is a schematic diagram of the original time series of marine magnetotelluric data for a certain region in an embodiment of the present invention; Figure 3 This is a schematic diagram of the time series of marine magnetotellurics data for a certain area after denoising by the present invention in an embodiment of the present invention; Figure 4 This is a schematic diagram comparing the apparent resistivity and phase curves of marine magnetotellurics in a certain area before and after denoising by the present invention in an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the principle of the marine magnetotelluric data denoising system in an embodiment of the present invention; In the diagram: 1. Data model training module; 2. Data preprocessing windowing module; 3. Parallel noise identification module; 4. Mask fusion processing module; 5. Signal reconstruction processing module; 6. Data window splicing module. Detailed Implementation

[0027] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0028] The purpose of this invention is to provide a method and related equipment for denoising marine magnetotelluric (MMT) data, addressing the shortcomings of existing MMT data processing techniques in handling long-period, multi-type mixed noise. The core idea of ​​this invention is to transform the MMT denoising problem into a data-driven supervised learning task. Using existing paired time-series samples ("before denoising" - noisy) and "after denoising" - clean, a set of specialized deep learning models is trained in stages. Then, this trained model is applied to automatically process new noisy data.

[0029] The present invention will now be described in further detail with reference to the accompanying drawings: See Figure 1 This invention provides a method for denoising marine magnetotelluric data, comprising: Step 1: Construct a training dataset and train the model to obtain a dual-path noise recognition model and a generative reconstruction model; Specifically, constructing the training dataset includes obtaining pairs of noisy ocean magnetotelluric time series X_noisy and corresponding clean ocean magnetotelluric time series Y_clean; and segmenting the pairs of noisy ocean magnetotelluric time series X_noisy and corresponding clean ocean magnetotelluric time series Y_clean into fixed-length windows with overlapping parts to form training sample pairs.

[0030] The dual-path noise recognition model is used to identify periodic noise and random impulse noise. The specific process of training the dual-path noise recognition model includes training the periodic noise recognition model and training the random impulse noise recognition model. The periodic noise recognition model training includes constructing a one-dimensional convolutional neural network (1D-CNN), taking a pair of noisy marine magnetotelluric time series X_noisy as input, and using a periodic noise mask Mask_p generated by calculating the difference between |X_noisy-Y_clean| and combining it with frequency domain analysis as a label for training. In the periodic noise mask Mask_p, the long-period noise positions are marked as 1, and the rest are marked as 0. The training of the random impulse noise identification model includes constructing an autoencoder network, using the corresponding clean ocean magnetotelluric time series Y_clean as input and training target for training, so that the model learns the characteristics of normal ocean magnetotelluric signals, and the autoencoder identifies random impulse noise through reconstruction error; The generative reconstruction model is used to reconstruct signals in noisy regions based on contextual information. The specific process of training the generative reconstruction model includes: constructing a sequence generation model, which is a bidirectional long short-term memory network or a Transformer self-attention mechanism deep learning model; taking X_masked, obtained by masking paired noisy marine magnetotelluric time series X_noisy, as input, and using the corresponding clean marine magnetotelluric time series Y_clean as label, for training, so that the model can recover the clean signal based on contextual information. Step 2: Preprocess and window the new raw marine magnetotelluric data to be processed to obtain multiple data windows; Specifically, the preprocessing and windowing methods are consistent with the window segmentation method used when constructing the training dataset, that is, segmenting into fixed-length windows with overlapping parts.

[0031] Step 3: Perform parallel noise identification on each data window based on the dual-path noise identification model to obtain periodic noise mask and random impulse noise mask; Specifically, each data window is input into a trained 1D-CNN model to obtain a predicted periodic noise mask M_p; each data window is input into a trained autoencoder model, and by calculating the reconstruction error and setting a threshold, a random impulse noise mask M_i is obtained.

[0032] Step 4: A composite noise mask is obtained by fusing the periodic noise mask and the random impulse noise mask. The composite noise mask marks the noise data points that need to be reconstructed in the data window. Specifically, the fusion process involves performing a logical OR operation on the periodic noise mask M_p and the random impulse noise mask M_i to generate the final composite noise mask M_total. The composite noise mask M_total marks all noise data points that need to be reconstructed within the current window.

[0033] Step 5: Reconstruct the signal of the data window based on the generative reconstruction model and the comprehensive noise mask to obtain the clean data window after processing; Specifically, signal reconstruction is performed on the data window based on the generative reconstruction model and the comprehensive noise mask. The signal reconstruction includes masking the data window based on the comprehensive noise mask M_total, and inputting the processed sequence into the generative reconstruction model to obtain a clean data window filled with noise regions.

[0034] Step 6: Stitch together all clean data windows to obtain the complete denoised marine magnetotelluric time series.

[0035] Specifically, the splicing process involves performing a smooth, weighted average of the overlapping areas of all clean data windows before splicing them together.

[0036] according to Figure 2 The figure shown is the original time series of marine magnetotelluric data for a certain region; according to... Figure 3 As shown, the time series of marine magnetotellurics data for a certain area, after denoising using this invention, is based on... Figure 4 As shown, a comparison of the apparent resistivity and phase curves of marine magnetotellurics in a certain area before and after denoising using this invention is presented.

[0037] In summary, this invention provides a denoising method for marine magnetotelluric (MMT) data. By constructing a dual-path noise identification model, it specifically trains and identifies long-period noise and random impulse noise, which are unique to and difficult to process in the ocean, respectively. This accurately generates two types of noise masks, achieving precise localization of complex marine noise and avoiding the limitations of single models in adapting to different noise types. A generative reconstruction model combined with a comprehensive noise mask is used for signal reconstruction. The model, trained on massive amounts of data, grasps the inherent physical laws and time dependencies of MMT signals, and performs high-fidelity recovery of noisy regions based on contextual information. This effectively avoids the loss of low-frequency effective signals caused by traditional frequency domain filtering, ensuring the integrity of deep geological information. The entire denoising process is divided into two stages: offline model training and online data processing. During online processing, preprocessing windowing, parallel noise identification, automated reconstruction, and stitching eliminate the need for extensive manual interaction and parameter tuning, significantly improving processing efficiency. By using noise masking fusion processing, all noise data points that need to be reconstructed are accurately marked, ensuring that the reconstruction is performed only on the noise area, thus preserving the effective information in the original signal to the greatest extent. At the same time, when splicing data windows, overlapping areas are smoothed and weighted averaged to ensure the continuity and integrity of the output time series, which greatly improves the accuracy and reliability of deep-sea geophysical exploration.

[0038] Example 1 This embodiment 1 uses an 8-day MMT data segment containing long-period noise and random ship interference from a certain survey area as an example to specifically illustrate the implementation process of the present invention. The data was collected at a sampling rate of 150 Hz.

[0039] Phase 1: Model training (already completed offline beforehand) Using 100 sets of 24-hour data pairs {X_noisy, Y_clean} from the same test area, we completed the training of three models: 1D-CNN, autoencoder, and Bi-LSTM.

[0040] Phase Two: New Data Processing Data windowing: Divide the 8 days of data to be processed into several data windows with a window length of 24 hours and a step size of 12 hours (50% overlap).

[0041] Process the first window (0-24 hours): Noise Identification: The 24-hour data was input into the pre-trained 1D-CNN and autoencoder respectively. The 1D-CNN successfully identified the locations of several long-period fluctuations and output a mask M_p. The autoencoder detected large reconstruction errors at several time points, identifying them as ship interference pulses and outputting a mask Mi.

[0042] Mask fusion: Combine the periodic noise mask M_p and the random impulse noise mask M_i to obtain the comprehensive noise mask M_total.

[0043] Signal reconstruction: The 24-hour data is processed using a comprehensive noise mask M_total to mark the regions to be reconstructed. This sequence is then input into a pre-trained Bi-LSTM model. The model uses the clean signals before and after as context to fill in the marked regions (including long-period fluctuations and pulse spikes), outputting a clean 24-hour data segment.

[0044] Loop processing: Repeat the above steps for all subsequent windows in turn.

[0045] Data stitching: After processing all windows, a series of clean 24-hour data segments are obtained. For overlapping regions (e.g., 12-24 hours), the results from the first and second windows are linearly weighted and averaged to achieve a smooth transition. Finally, all parts are stitched together to obtain a complete, high-quality 8-day output time series. This series can then be used for subsequent impedance tensor estimation and inversion interpretation.

[0046] In summary, this embodiment achieves a paradigm shift from model-driven to data-driven approaches, solving the challenge of separating long-period noise from effective signals. Traditional methods rely on prior mathematical models, which cannot handle complex ocean noise where the model is unknown. This embodiment utilizes supervised learning to directly learn the complex nonlinear characteristics of noise from a large amount of real data, particularly learning to distinguish the subtle differences between long-period noise fluctuations and the true electrical response in the time domain. This allows for the accurate identification and removal of noise without compromising the effective signal in the same frequency band. This represents a fundamental overhaul of traditional filtering methods.

[0047] This embodiment reconstructs signals with high fidelity, preserving deep geological information to the maximum extent. This invention abandons traditional mathematical interpolation and employs advanced generative models (such as Bi-LSTM) for signal reconstruction. This model learns the inherent laws of massive amounts of pure signals, and its reconstructed data is not only numerically continuous but also conforms to the physical propagation characteristics of magnetotelluric signals in terms of phase, spectrum, and time dependence. This ensures that the reconstructed data is physically accurate rather than mathematically smoothed, providing highly reliable input for subsequent inversion interpretation, directly impacting the success or failure of deep resource exploration.

[0048] This embodiment features a highly automated and intelligent processing flow, significantly improving the efficiency of processing massive amounts of data. Once the model has completed offline training, the entire denoising process requires no manual intervention or tedious parameter adjustments for new massive datasets. This "train once, apply anywhere" model frees data processing personnel from repetitive and inefficient labor, enabling rapid and standardized processing of terabyte-level data, greatly shortening project cycles, and reducing the risk of human error.

[0049] This embodiment provides a highly adaptable and standardized intelligent processing framework, filling a technological gap in China. Through its data-driven modeling process, this embodiment can quickly construct a dedicated denoising model for the unique noise of a new survey area using only a small amount of calibration data. This method eliminates the heavy reliance on human experience in traditional processing, ensuring optimal processing results under different exploration environments. It achieves standardization and engineering in data processing, and its core technological concepts can also provide a reference for other geophysical exploration fields.

[0050] This embodiment utilizes existing expert processing experience (i.e., paired data before and after denoising) to train a deep learning model, enabling it to automatically and accurately identify and locate long-period noise and random impulse noise mixed in the signal; avoiding the use of traditional filtering methods that damage effective low-frequency components. Using a generative model, based on the contextual physical laws of the signal, it performs high-fidelity reconstruction of the noise-polluted area, preserving the integrity of geophysical information to the maximum extent.

[0051] Example 2 according to Figure 5 As shown, this embodiment 2 provides a denoising system for marine magnetotelluric data, including: Data model training module 1 is used to construct a training dataset and train the model to obtain a dual-path noise recognition model and a generative reconstruction model. Data preprocessing windowing module 2 is used to preprocess and window new raw marine magnetotelluric data to be processed, resulting in multiple data windows; Parallel noise identification module 3 is used to perform parallel noise identification on each data window based on the dual-path noise identification model to obtain periodic noise mask and random impulse noise mask; The mask fusion processing module 4 is used to perform fusion processing based on periodic noise mask and random impulse noise mask to obtain a comprehensive noise mask, wherein the comprehensive noise mask marks the noise data points that need to be reconstructed in the data window; Signal reconstruction processing module 5 is used to reconstruct the signal of the data window based on the generative reconstruction model and the comprehensive noise mask to obtain a clean data window after processing. Data window stitching module 6 is used to stitch together all clean data windows to obtain a complete denoised marine magnetotelluric time series.

[0052] Example 3 The present invention also provides a mobile terminal, including a memory, a processor, and a computer program stored in the memory and executable on the processor, such as a denoising program for marine magnetotelluric data.

[0053] When the processor executes the computer program, it implements the steps of the above-described method for denoising marine magnetotelluric data, for example: A training dataset was constructed and the model was trained to obtain a dual-path noise recognition model and a generative reconstruction model; The new raw marine magnetotelluric data to be processed is preprocessed and windowed to obtain multiple data windows; Parallel noise identification is performed on each data window based on a dual-path noise identification model to obtain periodic noise masks and random impulse noise masks. A composite noise mask is obtained by fusing periodic noise mask and random impulse noise mask, and the composite noise mask marks the noise data points that need to be reconstructed in the data window; The data window is reconstructed based on a generative reconstruction model and a comprehensive noise mask to obtain a clean data window after processing. All clean data windows were stitched together to obtain the complete denoised marine magnetotelluric time series.

[0054] Alternatively, when the processor executes the computer program, it implements the functions of each module in the above system, for example: Data model training module 1 is used to construct a training dataset and train the model to obtain a dual-path noise recognition model and a generative reconstruction model. Data preprocessing windowing module 2 is used to preprocess and window new raw marine magnetotelluric data to be processed, resulting in multiple data windows; Parallel noise identification module 3 is used to perform parallel noise identification on each data window based on the dual-path noise identification model to obtain periodic noise mask and random impulse noise mask; The mask fusion processing module 4 is used to perform fusion processing based on periodic noise mask and random impulse noise mask to obtain a comprehensive noise mask, wherein the comprehensive noise mask marks the noise data points that need to be reconstructed in the data window; Signal reconstruction processing module 5 is used to reconstruct the signal of the data window based on the generative reconstruction model and the comprehensive noise mask to obtain a clean data window after processing. Data window stitching module 6 is used to stitch together all clean data windows to obtain a complete denoised marine magnetotelluric time series.

[0055] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the mobile terminal.

[0056] For example, the computer program can be divided into a data model training module 1, a data preprocessing windowing module 2, a parallel noise identification module 3, a mask fusion processing module 4, a signal reconstruction processing module 5, and a data window splicing module 6. The specific functions of each module are as follows: Data model training module 1 is used to construct a training dataset and train the model to obtain a dual-path noise recognition model and a generative reconstruction model. Data preprocessing windowing module 2 is used to preprocess and window new raw marine magnetotelluric data to be processed, resulting in multiple data windows; Parallel noise identification module 3 is used to perform parallel noise identification on each data window based on the dual-path noise identification model to obtain periodic noise mask and random impulse noise mask; The mask fusion processing module 4 is used to perform fusion processing based on periodic noise mask and random impulse noise mask to obtain a comprehensive noise mask, wherein the comprehensive noise mask marks the noise data points that need to be reconstructed in the data window; Signal reconstruction processing module 5 is used to reconstruct the signal of the data window based on the generative reconstruction model and the comprehensive noise mask to obtain a clean data window after processing. Data window stitching module 6 is used to stitch together all clean data windows to obtain a complete denoised marine magnetotelluric time series.

[0057] The mobile terminal can be a computing device such as a desktop computer, laptop, handheld computer, or cloud server. The mobile terminal may include, but is not limited to, a processor and memory.

[0058] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the mobile terminal, connecting various parts of the mobile terminal via various interfaces and lines.

[0059] The memory can be used to store the computer program and / or module. The processor implements various functions of the mobile terminal by running or executing the computer program and / or module stored in the memory and calling the data stored in the memory.

[0060] The memory may primarily include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a function (such as sound playback, image playback, etc.); the data storage area may store data created based on the use of the mobile phone (such as audio data, phonebook, etc.). Furthermore, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disks, RAM, plug-in hard disks, SmartMediaCards (SMC), Secure Digital (SD) cards, FlashCards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0061] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for denoising marine magnetotelluric data, characterized in that, include: A training dataset was constructed and the model was trained to obtain a dual-path noise recognition model and a generative reconstruction model; The new raw marine magnetotelluric data to be processed is preprocessed and windowed to obtain multiple data windows; Parallel noise identification is performed on each data window based on a dual-path noise identification model to obtain periodic noise masks and random impulse noise masks. A composite noise mask is obtained by fusing periodic noise mask and random impulse noise mask, and the composite noise mask marks the noise data points that need to be reconstructed in the data window; The data window is reconstructed based on a generative reconstruction model and a comprehensive noise mask to obtain a clean data window after processing. All clean data windows were stitched together to obtain the complete denoised marine magnetotelluric time series.

2. A method for denoising marine magnetotelluric data according to claim 1, characterized in that, The construction of the training dataset includes obtaining pairs of noisy marine magnetotelluric time series X_noisy and corresponding clean marine magnetotelluric time series Y_clean; and segmenting the pairs of noisy marine magnetotelluric time series X_noisy and corresponding clean marine magnetotelluric time series Y_clean into fixed-length windows with overlapping parts to form training sample pairs.

3. The method for denoising marine magnetotelluric data according to claim 2, characterized in that, The dual-path noise recognition model is used to identify periodic noise and random impulse noise. The specific process of training the dual-path noise recognition model includes training the periodic noise recognition model and training the random impulse noise recognition model. The periodic noise identification model training includes constructing a one-dimensional convolutional neural network, taking a pair of noisy marine magnetotelluric time series X_noisy as input, and using a periodic noise mask Mask_p generated by calculating the difference between |X_noisy-Y_clean| and combining it with frequency domain analysis as a label for training. In the periodic noise mask Mask_p, the long-period noise positions are marked as 1, and the rest are marked as 0. The training of the random impulse noise identification model includes constructing an autoencoder network, using the corresponding clean ocean magnetotelluric time series Y_clean as input and training target for training, so that the model learns the characteristics of normal ocean magnetotelluric signals, and the autoencoder identifies random impulse noise through reconstruction error; The generative reconstruction model is used to reconstruct signals in noisy regions based on contextual information. The specific process of training the generative reconstruction model includes: constructing a sequence generation model, which is a bidirectional long short-term memory network or a Transformer self-attention mechanism deep learning model; taking X_masked, obtained by masking paired noisy marine magnetotelluric time series X_noisy, as input, and using the corresponding clean marine magnetotelluric time series Y_clean as label for training, so that the model can recover the clean signal based on contextual information.

4. The method for denoising marine magnetotelluric data according to claim 1, characterized in that, The preprocessing and windowing method is consistent with the window segmentation method used when constructing the training dataset, that is, it is segmented into fixed-length windows with overlapping parts.

5. A method for denoising marine magnetotelluric data according to claim 1, characterized in that, Each data window is input into the trained 1D-CNN model to obtain the predicted periodic noise mask M_p; each data window is input into the trained autoencoder model, and the random impulse noise mask M_i is obtained by calculating the reconstruction error and setting a threshold.

6. A method for denoising marine magnetotelluric data according to claim 1, characterized in that, The fusion process involves performing a logical OR operation on the periodic noise mask M_p and the random impulse noise mask M_i to generate the final composite noise mask M_total. The composite noise mask M_total marks all noise data points that need to be reconstructed within the current window.

7. A method for denoising marine magnetotelluric data according to claim 1, characterized in that, The signal reconstruction of the data window is based on the generative reconstruction model and the comprehensive noise mask. The signal reconstruction includes masking the data window based on the comprehensive noise mask M_total, and inputting the processed sequence into the generative reconstruction model to obtain a clean data window filled with noise regions.

8. A method for denoising marine magnetotelluric data according to claim 1, characterized in that, The splicing process involves performing a smooth, weighted average of the overlapping areas of all clean data windows before splicing them together.

9. A denoising system for marine magnetotelluric data, characterized in that, include: The data model training module is used to build a training dataset and train the model to obtain a dual-path noise recognition model and a generative reconstruction model. The data preprocessing windowing module is used to preprocess and window new raw marine magnetotelluric data to be processed, resulting in multiple data windows. The parallel noise identification module is used to perform parallel noise identification on each data window based on the dual-path noise identification model to obtain periodic noise mask and random impulse noise mask; The mask fusion processing module is used to perform fusion processing based on periodic noise mask and random impulse noise mask to obtain a comprehensive noise mask, wherein the comprehensive noise mask marks the noise data points to be reconstructed in the data window; The signal reconstruction processing module is used to reconstruct the signal of the data window based on the generative reconstruction model and the comprehensive noise mask, so as to obtain a clean data window after processing. The data window stitching module is used to stitch together all clean data windows to obtain a complete denoised marine magnetotelluric time series.

10. A mobile terminal, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method for denoising marine magnetotelluric data as described in any one of claims 1-8.

Images