Geoscience signal processing method based on deep learning acceleration variational modal decomposition optimization

Abstract

The present application discloses a geoscience signal processing method based on deep learning acceleration variational mode decomposition optimization, relates to the technical field of geophysical signal processing, and comprises the following steps: constructing a CNN-LSTM deep learning agent model, training by using a training data set, and establishing a mapping relationship among a geoscience signal, a VMD candidate parameter combination and a real fitness value; collecting a geoscience signal sample, generating a VMD candidate parameter combination by using an iterative optimization algorithm, inputting the VMD candidate parameter combination into the CNN-LSTM deep learning agent model, obtaining a predicted fitness value, feeding back the predicted fitness value to the iterative optimization algorithm, and obtaining an optimal VMD parameter combination; and performing VMD decomposition on a geoscience signal to be processed according to the optimal VMD parameter combination, obtaining IMF components, screening and reconstructing the IMF components, and obtaining a denoised geoscience signal. The method can quickly determine a global optimal VMD parameter combination and improve the quality of geoscience signal processing.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of geophysical signal processing, in particular to a geoscience signal processing method based on deep learning acceleration of variational mode decomposition optimization. BACKGROUND

[0002] In the field of geoscience signal processing, variational mode decomposition (VMD) is often used to extract intrinsic mode functions (IMF) from noisy geoscience signals. The VMD parameter combination composed of mode number and quadratic penalty factor needs to be determined in the preset VMD parameter space. VMD decomposition is performed, and the VMD decomposition quality is evaluated by energy concentration, spectral entropy, and reconstruction error as indicators. Parameter optimization is performed through grid search or swarm intelligence iterative method, and then decomposition and reconstruction are performed to obtain the denoised geoscience signal.

[0003] Conventional VMD parameter optimization relies on repeated execution of VMD decomposition, which has high computational complexity. Iterative optimization algorithms require a large number of iterations to converge, and the parameter optimization process is time-consuming, which is difficult to meet the demand for high timeliness of on-site data processing. Due to significant differences in noise spectrum and observation conditions, a single evaluation indicator and fixed threshold cannot maintain consistent discrimination in various scenarios, resulting in insufficient stability and portability of VMD parameter selection. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides a geoscience signal processing method based on deep learning acceleration of variational mode decomposition optimization, which solves the problem of low computational efficiency and long time consumption of VMD decomposition parameter optimization.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] The present application provides a geoscience signal processing method based on deep learning acceleration of variational mode decomposition optimization, which includes collecting geoscience signal samples, generating VMD candidate parameter combinations in a preset VMD parameter space, calculating true fitness values through VMD decomposition, and constructing a training data set.

[0008] A CNN-LSTM deep learning agent model is constructed, and the deep learning agent model is trained offline using the training data set to establish a mapping relationship between geoscience signals, VMD candidate parameter combinations, and true fitness values.

[0009] The collected geoscience signal samples are input into the CNN-LSTM deep learning agent model to obtain predicted fitness values, which are fed back to the iterative optimization algorithm to obtain the optimal VMD parameter combination.

[0010] Based on the optimal VMD parameter combination, the original geoscience signal to be processed is decomposed by VMD to obtain IMF components, the IMF components are screened and reconstructed to obtain a denoised geoscience signal.

[0011] As a preferred scheme of the geoscience signal processing method based on deep learning acceleration variational modal decomposition optimization, the steps of constructing the training data set are specifically,

[0012] Collect geoscience signal samples and record sampling information and numbers, set the mode number and the range of the quadratic penalty factor according to the preset VMD parameter space, and generate a VMD candidate parameter combination list;

[0013] The geoscience signal sample and the VMD candidate parameter combination are decomposed by VMD to obtain an IMF component set;

[0014] According to the IMF component set, the real fitness value is calculated according to the entropy index;

[0015] The geoscience signal sample, the VMD candidate parameter combination and the corresponding real fitness value are used as elements to construct a training sample for all collected geoscience signal samples and all VMD candidate parameter combinations, and a training data set is generated.

[0016] As a preferred scheme of the geoscience signal processing method based on deep learning acceleration variational modal decomposition optimization, the preset VMD parameter space refers to the upper and lower limits and sampling range of the values of the two key parameters of variational modal decomposition, mode number and quadratic penalty factor.

[0017] As a preferred scheme of the geoscience signal processing method based on deep learning acceleration variational modal decomposition optimization, the CNN-LSTM deep learning agent model refers to a CNN-LSTM deep learning agent model composed of CNN and LSTM.

[0018] As a preferred scheme of the geoscience signal processing method based on deep learning acceleration variational modal decomposition optimization, the steps of constructing the CNN-LSTM deep learning agent model are specifically,

[0019] The geoscience signal sample and the VMD candidate parameter combination in the training data set are preprocessed and input into the CNN-LSTM deep learning agent model;

[0020] In the CNN-LSTM deep learning agent model, the local features of the geoscience signal are extracted by a one-dimensional convolution layer, the local features of the geoscience signal are obtained through a long short-term memory network layer to obtain the time sequence dependence of the geoscience signal, and the time sequence dependence of the geoscience signal is output through a full connection layer as a predicted fitness value.

[0021] As a preferred scheme of the geoscience signal processing method based on deep learning accelerated variational mode decomposition optimization, the preprocessing refers to arranging the geoscience signal sample / geoscience signal into a fixed-length geoscience signal sequence, normalizing the VMD candidate parameter combination in a value range, and combining the geoscience signal sequence and the normalized VMD candidate parameter combination according to the input order specified by the CNN-LSTM deep learning agent model.

[0022] As a preferred scheme of the geoscience signal processing method based on deep learning accelerated variational mode decomposition optimization, the deep learning agent model is trained offline using a training data set to establish a mapping relationship between the geoscience signal, the VMD candidate parameter combination, and the true fitness value. The specific steps are as follows:

[0023] With the true fitness value as a reference, the error between the predicted fitness value and the true fitness value is calculated through a loss function, the parameters of the CNN-LSTM deep learning agent model are adjusted, and the training is repeated until the error is stable, and the trained CNN-LSTM deep learning agent model is obtained.

[0024] The trained CNN-LSTM deep learning agent model is saved, and the mapping relationship between the geoscience signal, the VMD candidate parameter combination, and the true fitness value is established.

[0025] As a preferred scheme of the geoscience signal processing method based on deep learning accelerated variational mode decomposition optimization, the geoscience signal sample collected is input into the CNN-LSTM deep learning agent model through an iterative optimization algorithm to obtain a predicted fitness value. The specific steps are as follows:

[0026] The population and search boundary of the iterative optimization algorithm are initialized, and the geoscience signal sample collected is generated through the internal mechanism of the iterative optimization algorithm to generate a VMD candidate parameter combination.

[0027] The geoscience signal sample collected and the VMD candidate parameter combination are input into the trained CNN-LSTM deep learning agent model to output a predicted fitness value.

[0028] As a preferred scheme of the geoscience signal processing method based on deep learning accelerated variational mode decomposition optimization, the predicted fitness value is fed back to the iterative optimization algorithm to update the individual position of the population and generate a new VMD candidate parameter combination.

[0029] The trained CNN-LSTM deep learning agent model is repeatedly predicted and fed back using the new VMD candidate parameter combination until the iterative optimization algorithm converges, and the optimal VMD parameter combination is obtained.

[0030] As a preferred scheme of the geoscience signal processing method based on deep learning acceleration variational mode decomposition optimization of the application, wherein: based on the optimal VMD parameter combination, the original geoscience signal to be processed is subjected to VMD decomposition, the IMF component is obtained, the IMF component is screened and reconstructed, and the denoising geoscience signal is obtained, and the specific steps are,

[0031] The optimal VMD parameter combination is used as the VMD decomposition parameter, and the original geoscience signal to be processed is subjected to one-time VMD decomposition to obtain an IMF component set;

[0032] Based on the IMF component set, the Pearson correlation coefficient of each IMF component and the original geoscience signal is calculated, and the effective IMF component is screened according to the Pearson correlation coefficient threshold;

[0033] The effective IMF component is superimposed and reconstructed to obtain a denoising geoscience signal;

[0034] The original geoscience signal refers to the original noisy transient electromagnetic signal to be processed.

[0035] The application has the beneficial effects that: by constructing a CNN-LSTM deep learning agent model, the fitness value is quickly predicted; by generating a VMD candidate parameter combination through an iterative optimization algorithm, the global optimal VMD parameter combination is quickly determined, and the signal processing quality is improved; by collecting different geoscience signal samples to construct a training data set, the deep learning model and the iterative optimization algorithm are flexibly replaced, and the method is universal and scalable. BRIEF DESCRIPTION OF DRAWINGS

[0036] In order to more clearly illustrate the technical solutions of the embodiments of the application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0037] Figure 1 The flowchart of the geoscience signal processing method based on deep learning acceleration variational mode decomposition optimization.

[0038] Figure 2 The schematic diagram of the geoscience signal processing method based on deep learning acceleration variational mode decomposition optimization.

[0039] Figure 3 The flowchart of deep learning acceleration variational mode decomposition.

[0040] Figure 4 The three-dimensional waterfall diagram of the IMF component obtained by VMD decomposition.

[0041] Figure 5 The figure shows the effect comparison of the original geoscience signal before and after noise removal processing. DETAILED DESCRIPTION

[0042] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0043] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application. However, the present application can be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the scope of the present application, so the present application is not limited to the specific embodiments disclosed below.

[0044] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the present application. "In one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an embodiment that is separate or alternative to other embodiments.

[0045] Reference Figures 1-5 For one embodiment of the present application, the embodiment provides a geoscience signal processing method based on deep learning acceleration variational modal decomposition optimization, including the following steps:

[0046] S1, collect geoscience signal samples, generate VMD candidate parameter combinations in a preset VMD parameter space, calculate the true fitness value through VMD decomposition, and construct a training data set.

[0047] Collect geoscience signal samples and register sampling information and number, set the mode number and quadratic penalty factor range according to the preset VMD parameter space, and generate a VMD candidate parameter combination list.

[0048] Further, through numerical simulation or collection from historical projects, 10000 geoscience signal samples of different signal-to-noise ratios and different noise types are obtained, each geoscience signal sample has a length of 1024 sampling points, and constitutes a geoscience signal sequence.

[0049] It should be noted that the noise types include but are not limited to power frequency interference and random noise.

[0050] Further, within the preset VMD parameter space, grid sampling is performed to generate 200 VMD candidate parameter combinations, and a VMD candidate parameter combination list is output.

[0051] The preset VMD parameter space refers to the two key parameters of variational modal decomposition, mode number And quadratic penalty factor The upper and lower boundaries and sampling range of the preset value.

[0052] It should be noted that the modal number is too small to cause the IMF component to be insufficiently separated, and the modal number is too large to introduce false IMF components. Based on engineering experience, the modal number is in the range of [3, 15].

[0053] The quadratic penalty factor is too small to cause the modal bandwidth to be too large, and the quadratic penalty factor is too large to cause the modal bandwidth to be too small, resulting in loss of effective signal energy. Based on engineering experience, the value of the quadratic penalty factor is in the range of [500, 10000].

[0054] The geoscience signal sample and the VMD candidate parameter combination are subjected to VMD decomposition to obtain an IMF component set. Specifically, the VMD constrained variational problem objective function is represented as:

[0055] ;

[0056] wherein, indicates the index of the IMF component being updated, indicates the time, indicates the differential of , the indicates the th IMF component after VMD decomposition, indicates the set of all IMF components, indicates the center frequency of the th IMF component, is the set of center frequencies of all IMF components, indicates the Dirac impulse function, indicates convolution calculation, indicates the original geoscience signal, indicates an imaginary number, satisfying , indicates the squared two-norm, indicates a minimum search on the set of all IMF components and the set of center frequencies of all IMF components, indicates a constraint condition.

[0057] By introducing a penalty parameter and a Lagrange multiplier operator, the VMD constrained variational problem is converted into an unconstrained problem, and the augmented Lagrange function is represented as:

[0058] ;

[0059] wherein, indicates the augmented Lagrange function, indicates the quadratic penalty factor, indicates the Lagrange multiplier operator.

[0060] Furthermore, the time-domain signal of the original geoscientific signal is transformed into the frequency domain using Hilbert transform. The saddle point of the augmented Lagrange function is then solved in the frequency domain to obtain the frequency-domain IMF component, center frequency, and the iterative update relationship of the frequency-domain Lagrange multiplication operator, expressed as:

[0061] ;

[0062] ;

[0063] ;

[0064] in, Indicates the first The number obtained in the iteration is... Frequency domain of each IMF component This represents the frequency domain signal obtained through fitting. Represents the frequency domain of other IMF components, subscript This represents the index of other IMF components besides the current one. Represents frequency variables. Indicates the first The center frequency of the IMF component obtained in the next iteration. Indicates the first Frequency-domain Lagrange multiplication operator for the next iteration Indicates the first Frequency-domain Lagrange multiplication operator for the next iteration Indicates the iteration count index. Indicates the first Frequency domain of each IMF component This indicates the iteration step size.

[0065] For each IMF component in the frequency domain, calculate the ratio of the L2 norm of the difference between two consecutive iterations to the L2 norm of the previous iteration, and sum these ratios for all IMF components. The iteration ends when the sum of these ratios is less than the convergence threshold, as shown below:

[0066] ;

[0067] in, Indicates the first The number obtained in the iteration is... Frequency domain of each IMF component This represents the convergence threshold, which takes the value of... .

[0068] Furthermore, each geoscientific signal sample and each set of VMD candidate parameters are combined to perform VMD decomposition and obtain the IMF component set.

[0069] The true fitness value is calculated based on the IMF component set and according to the entropy class index.

[0070] It should be noted that entropy metrics include, but are not limited to, minimum envelope entropy, minimum sample entropy, and maximum energy entropy.

[0071] It should be noted that the envelope entropy is used as the true fitness value. The envelope entropy can effectively reflect the sparsity characteristics of geoscientific signals. A small envelope entropy value indicates that the geoscientific signal sequence is ordered and has few noise components.

[0072] Specifically, for each IMF component, the envelope signal is calculated using Hilbert transform, the probability distribution of the envelope signal is calculated to obtain the probability of each geoscientific signal sampling point, and the envelope entropy of the IMF component is calculated, expressed as:

[0073] ;

[0074] ;

[0075] in, The envelope entropy of the IMF component is represented by... Indicates the number of sampling points in a geoscientific signal sample. The envelope signal after normalization is represented as the first... The probability of a geoscientific signal sampling point This represents the index of the geoscientific signal sampling points after the envelope signal has been normalized. This represents the amplitude of the envelope signal.

[0076] Furthermore, the envelope entropy of all IMF components is averaged or the minimum value is selected as the true fitness value under the current VMD candidate parameter combination.

[0077] It should be noted that, in the embodiments of this application, the minimum envelope entropy is preferred as the true fitness value. The minimum envelope entropy directly corresponds to the purest and most periodic IMF component. The minimum envelope entropy is used to characterize the VMD decomposition effect and to judge the merits of VMD candidate parameter combinations.

[0078] Furthermore, using geoscientific signal samples, VMD candidate parameter combinations, and corresponding true fitness values ​​as elements, training samples are constructed from all collected geoscientific signal samples and all VMD candidate parameter combinations to generate a training dataset.

[0079] It should be noted that the training dataset is constructed by combining the true fitness values ​​generated from each geoscientific signal sample and each set of VMD candidate parameter combinations.

[0080] It should be noted that each training data point in the training dataset is in the form of {geological signal sequence, (K, α), true fitness value}.

[0081] S2, a CNN-LSTM deep learning agent model is constructed, and the deep learning agent model is trained offline using the training data set to establish a mapping relationship between the geoscience signals, the VMD candidate parameter combination, and the true fitness value.

[0082] It should be noted that the CNN-LSTM deep learning agent model refers to a CNN-LSTM deep learning agent model composed of CNN and LSTM.

[0083] Further, the geoscience signal sequence and the VMD candidate parameter combination are spliced or fused to generate a feature vector as the input of the CNN-LSTM deep learning agent model.

[0084] Specifically, the CNN feature extraction refers to constructing a one-dimensional convolutional neural network, setting 1-2 one-dimensional convolutional layers, using 64 convolutional kernels with a size of 3 in each layer, using a rectified linear unit activation function, performing convolution operation on the input geoscience signal sequence once, extracting local spatial features and local patterns of the geoscience signal, and connecting a max-pooling layer after each one-dimensional convolutional layer.

[0085] The LSTM captures the time sequence information, which includes inputting the feature sequence extracted by the CNN feature extraction module into the long short-term memory network, setting 1-2 long short-term memory network layers, configuring 100 neural units in each layer, learning the time sequence dependence and long-term trend between features, and outputting the time sequence feature representation.

[0086] In the full connection output stage, the time sequence feature representation is input into one or more full connection layers in sequence, and the full connection layer maps the high-dimensional feature representation to a low-dimensional feature representation. In order to reduce the risk of overfitting, a dropout layer is inserted between adjacent full connection layers.

[0087] A full connection layer containing only a single neuron is set as the output layer, a linear activation function is used, and a single scalar is output as the predicted fitness value, which represents the predicted minimum envelope entropy.

[0088] It should be noted that preprocessing refers to arranging the geoscience signal samples into fixed-length geoscience signal sequences, normalizing the VMD candidate parameter combination within the value range, and combining the geoscience signal sequence and the normalized VMD candidate parameter combination according to the input order specified by the CNN-LSTM deep learning agent model.

[0089] The constructed training data set is divided into a training set and a validation set in a ratio of 8:2, the mean square error is used as the loss function, and the CNN-LSTM deep learning agent model is trained through the Adam optimizer.

[0090] Further, in the CNN-LSTM deep learning agent model, the local features and local patterns of the geoscience signal are extracted by a one-dimensional convolutional layer, the local features and local patterns of the geoscience signal are obtained through a long short-term memory network layer to obtain a time sequence feature representation, and the time sequence feature representation is output through a fully connected layer to output a predicted fitness value.

[0091] Further, taking the real fitness value as a reference, an error between the predicted fitness value and the real fitness value is calculated through a loss function, parameters of the CNN-LSTM deep learning agent model are adjusted, and repeated training is performed until the error is stable, and a trained CNN-LSTM deep learning agent model is obtained.

[0092] Further, the trained CNN-LSTM deep learning agent model is saved, and a mapping relationship between the geoscience signal, the VMD candidate parameter combination, and the real fitness value is established.

[0093] It should be noted that the trained CNN-LSTM deep learning agent model outputs a predicted fitness value for any given geoscience signal sample and a set of VMD parameter combinations.

[0094] It should be noted that the CNN-LSTM deep learning agent model extracts local features and local patterns of the geoscience signal through CNN, obtains a time sequence feature representation of the geoscience signal through LSTM, quickly outputs a predicted fitness value, replaces a large number of real evaluations to form a closed-loop feedback, reduces the number of calls of VMD decomposition, and improves processing timeliness.

[0095] S3, the collected geoscience signal sample is input to the CNN-LSTM deep learning agent model through the iterative optimization algorithm to generate a VMD candidate parameter combination, and a predicted fitness value is obtained, and the predicted fitness value is fed back to the iterative optimization algorithm to obtain an optimal VMD parameter combination.

[0096] Further, the geoscience signal sample is collected, the population and search boundary of the iterative optimization algorithm are initialized, and the geoscience signal sample is generated through the internal mechanism of the iterative optimization algorithm to generate a VMD candidate parameter combination.

[0097] It should be noted that the iterative optimization algorithm includes but is not limited to sparrow search algorithm, particle swarm optimization algorithm, and genetic algorithm.

[0098] In the embodiments of the present application, the iterative optimization algorithm uses the sparrow search algorithm, specifically, the sparrow search algorithm is initialized, the population size is set to 30, and the maximum number of iterations is set to 100; the sparrow search algorithm is started, and in the iteration, the sparrow search algorithm generates 30 sets of VMD candidate parameter combinations according to the internal mechanism.

[0099] It should be noted that by updating the position update formulas of discoverers, joiners, and watchers, the sparrow search algorithm updates the position of all individuals in the population in each iteration, generating a new set of VMD candidate parameter combinations.

[0100] Specifically, the sparrow search algorithm updates the position by simulating the behavior of finders, joiners, and watchers in a sparrow population, where position refers to the combination of VMD candidate parameters.

[0101] It should be noted that updating the location of the discoverer involves selecting individuals with better fitness values ​​as discoverers, who are responsible for directional searching towards the best location globally and providing search direction for the population. The remaining individuals act as participants, following the discoverer in the search. A certain proportion (e.g., 10%-20%) of the individuals in the population are selected as watchdogs, who will engage in anti-predation behavior when they are aware of danger.

[0102] Specifically, the discoverer's location update is represented as:

[0103] ;

[0104] in, Indicates the first During the nth iteration The first sparrow discoverer in the first A new position in each dimension, This indicates the individual sparrow's ID number. Indicates the discoverer, This represents the dimension number of an individual sparrow in the VMD parameter space. Indicates the iteration round number. Indicates the first During the nth iteration The discoverer sparrow was in the first Position in each dimension Indicates the maximum number of iterations. For random perturbation parameters, This represents the warning value, with a range of [0, 1]. This represents the safety threshold, with a value range of [0.5, 1]. Represents a random number that follows a normal distribution. This represents a vector of all 1s with the same dimensions as the VMD parameter space.

[0105] Specifically, the update of the joiner's position is represented as:

[0106] ;

[0107] in, Indicates the first During the nth iteration The first sparrow joined in the... new position of the i-th dimension, representing the joiner, representing the i-th iteration, position of the i-th dimension, representing the i-th iteration, representing the optimal position occupied by the current discoverer at the i-th iteration, representing the i-th sparrow individual with the optimal fitness value in the discoverer at the current iteration, representing the position of the individual with the worst fitness value in the current population, representing that the joiner with a lower fitness value needs to fly to other places to forage, representing a random weight factor generated by a normal distribution, representing the sparrow population size.

[0108] Specifically, the position update of the alarm is represented as:

[0109] ;

[0110] wherein, new position of the i-th dimension, representing the i-th iteration, position of the i-th dimension, representing the alarm, representing the i-th iteration, position of the i-th dimension, representing the position of the optimal individual in the current population, representing the step control parameter, representing a random number with a value range of [-1, 1], representing the fitness value of the current alarm individual, representing the current global optimal fitness value, representing the current global worst fitness value, representing the minimum constant to avoid zero denominator. The collected geoscience signal samples and VMD candidate parameter combinations are input into the trained CNN-LSTM deep learning agent model, and the predicted fitness value is output.

[0111] Further, the predicted fitness value is fed back to the sparrow search algorithm, the population individual position is updated, new VMD candidate parameter combinations are generated, and the next iteration is entered.

[0112]

[0113] ​Furthermore, the trained CNN-LSTM deep learning agent model is driven by the new VMD candidate parameter combination to repeatedly predict and feedback until the sparrow search algorithm converges, thus obtaining the optimal VMD parameter combination.

[0114] Specifically, 100 iterations are set as the upper limit for convergence. After 100 iterations, the VMD candidate parameter combination is generated, input into the CNN-LSTM deep learning surrogate model to obtain the predicted fitness value, and the position of individuals in the population of the sparrow search algorithm is updated. The sparrow search algorithm converges and outputs the VMD candidate parameter combination with the smallest predicted fitness value as the optimal VMD parameter combination.

[0115] It should be noted that the Sparrow Search algorithm, combined with the predicted fitness value output by the CNN-LSTM deep learning surrogate model, continuously generates and filters VMD candidate parameter combinations, quickly approximating the optimal VMD parameter combination, reducing computational load and time cost, and improving the efficiency and accuracy of VMD candidate parameter optimization.

[0116] S4. Based on the optimal VMD parameter combination, the original geoscientific signal to be processed is decomposed by VMD to obtain IMF components. The IMF components are then filtered and reconstructed to obtain the denoised geoscientific signal.

[0117] Furthermore, the optimal VMD parameter combination is used as the VMD decomposition parameter to perform a VMD decomposition on the original geoscientific signal to be processed, and the IMF component set is obtained.

[0118] Furthermore, based on the IMF component set, the Pearson correlation coefficient between each IMF component and the original geoscientific signal is calculated, and effective IMF components are selected according to the Pearson correlation coefficient threshold.

[0119] It should be noted that the raw geoscientific signal refers to the raw, noisy transient electromagnetic signal to be processed.

[0120] Specifically, the Pearson correlation coefficient between each IMF component and the original geoscientific signal is calculated and expressed as:

[0121] ;

[0122] in: Represents the original geoscientific signal With IMF components The Pearson correlation coefficient between them The first one representing the original geoscientific signal The value of each sampling point, The first component of the IMF The value of each sampling point, This represents the average value of all sampling points of the original geoscientific signal. denotes the average value of all sampling points of the IMF component, denotes the total number of sampling points of the original geoscience signal.

[0123] It should be noted that the Pearson correlation coefficient has a value range of [-1, 1], and the closer the Pearson correlation coefficient is to 1, the stronger the positive correlation, and the closer the Pearson correlation coefficient is to -1, the stronger the negative correlation.

[0124] Further, the Pearson correlation coefficient threshold is set, and the value range of the Pearson correlation coefficient threshold is (0.1, 0.4), and the specific value of the Pearson correlation coefficient threshold can be determined according to prior knowledge of geoscience signal processing or experimental test.

[0125] Further, when the absolute value of the Pearson correlation coefficient of a certain IMF component is greater than the Pearson correlation coefficient threshold, it is determined to be an effective IMF component; when the absolute value of the Pearson correlation coefficient of a certain IMF component is less than or equal to the Pearson correlation coefficient threshold, it is determined to be a noise or pseudo IMF component, and the effective IMF component is determined according to the Pearson correlation coefficient.

[0126] Further, the effective IMF component is superimposed and reconstructed to obtain a denoised geoscience signal.

[0127] Specifically, all effective IMF components are linearly superimposed to obtain a denoised geoscience signal.

[0128] wherein the denoised geoscience signal is represented as:

[0129] ;

[0130] wherein, denotes the denoised geoscience signal, denotes the index set of all effective IMF components, denotes the effective IMF component index.

[0131] It should be noted that by superimposing and reconstructing the effective IMF component, the noise or pseudo IMF component is eliminated, and effective denoising of the original geoscience signal is achieved.

[0132] Specifically, as shown in Figure 4 , a series of IMF components and a residual component obtained by VMD decomposition are displayed in the form of a three-dimensional waterfall chart, and by screening and reconstructing the effective IMF component, a pure denoised geoscience signal in Figure 5 is obtained.

[0133] As shown in Figure 5As shown, the change of the original geoscience signal is intuitively presented in the form of a two-dimensional waveform graph, the blue curve represents the original noisy transient electromagnetic signal, and the orange curve represents the denoised geoscience signal after processing by the method of the present application. The orange curve superimposes obvious high-frequency noise and spike pulses on the basis of the original periodic geoscience signal, indicating that the background noise is significantly suppressed, and the geoscience signal is smoother, while effectively retaining the main geoscience signal characteristics.

[0134] In summary, the present application constructs a CNN-LSTM deep learning agent model to realize fast prediction of fitness values; generates VMD candidate parameter combinations through an iterative optimization algorithm to realize fast determination of the globally optimal VMD parameter combination and improve signal processing quality; and through the collection of different geoscience signal samples to construct a training data set, the deep learning model and the iterative optimization algorithm are flexibly replaced to realize the generality and scalability of the method.

[0135] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit it. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application, which should be covered in the scope of the claims of the present application.

Claims

1. A geoscience signal processing method based on deep learning acceleration variational modal decomposition optimization, characterized in that: The application relates to a method for processing geoscience signals. The method comprises the following steps: collecting geoscience signal samples, generating VMD candidate parameter combinations in a preset VMD parameter space, calculating real fitness values through VMD decomposition, and constructing a training data set; a CNN-LSTM deep learning agent model is constructed, the deep learning agent model is trained offline by using the training data set, and a mapping relationship among the geoscience signals, the VMD candidate parameter combinations and the real fitness values is established; the geoscience signal samples collected are used to generate VMD candidate parameter combinations through an iterative optimization algorithm, the VMD candidate parameter combinations are input into the CNN-LSTM deep learning agent model, a predicted fitness value is obtained, the predicted fitness value is fed back to the iterative optimization algorithm, and an optimal VMD parameter combination is obtained; based on the optimal VMD parameter combination, the original geoscience signals to be processed are subjected to VMD decomposition, IMF components are obtained, the IMF components are screened and reconstructed, and denoised geoscience signals are obtained. The training data set is constructed in the following specific steps: geoscience signal samples are collected, sampling information and numbers are registered, a mode number and a range of a quadratic penalty factor are set according to the preset VMD parameter space, and a VMD candidate parameter combination list is generated; the geoscience signal samples and the VMD candidate parameter combinations are subjected to VMD decomposition, and an IMF component set is obtained; real fitness values are calculated according to the IMF component set and an entropy index; training samples are constructed by taking the geoscience signal samples, the VMD candidate parameter combinations and the corresponding real fitness values as elements, all the collected geoscience signal samples and all the VMD candidate parameter combinations are used to generate the training data set; The CNN-LSTM deep learning agent model is constructed in the following specific steps: the geoscience signal samples and the VMD candidate parameter combinations in the training data set are preprocessed and input into the CNN-LSTM deep learning agent model; in the CNN-LSTM deep learning agent model, local features of the geoscience signals are extracted through a one-dimensional convolution layer, the local features of the geoscience signals are input into a long short-term memory network layer to obtain time sequence dependency of the geoscience signals, and the time sequence dependency of the geoscience signals is output as a predicted fitness value through a full connection layer; The CNN-LSTM deep learning agent model is trained offline by using the training data set, and a mapping relationship among the geoscience signals, the VMD candidate parameter combinations and the real fitness values is established in the following specific steps: the error between the predicted fitness value and the real fitness value is calculated by using a loss function with the real fitness value as a reference, the parameters of the CNN-LSTM deep learning agent model are adjusted, and the training is repeated until the error is stable, and a trained CNN-LSTM deep learning agent model is obtained; the trained CNN-LSTM deep learning agent model is saved, and a mapping relationship among the geoscience signals, the VMD candidate parameter combinations and the real fitness values is established. 2.The geoscience signal processing method based on deep learning acceleration variational modal decomposition optimization of claim 1, wherein: The preset VMD parameter space refers to preset upper and lower limits and sampling ranges of two key parameters, a mode number and a quadratic penalty factor, of variational mode decomposition. 3.The geoscience signal processing method based on deep learning acceleration variational modal decomposition optimization of claim 2, wherein: The CNN-LSTM deep learning agent model refers to a CNN-LSTM deep learning agent model composed of a CNN and an LSTM. 4.The geoscience signal processing method based on deep learning acceleration variational modal decomposition optimization of claim 3, wherein: The preprocessing refers to arranging the geoscience signal sample / geoscience signal into a geoscience signal sequence of fixed length, normalizing the VMD candidate parameter combination in the value range, and combining the geoscience signal sequence and the normalized VMD candidate parameter combination according to the input order specified by the CNN-LSTM deep learning agent model.

5. The geoscience signal processing method based on deep learning acceleration variational modal decomposition optimization of claim 4, wherein: The collected geoscience signal sample is input into the CNN-LSTM deep learning agent model through the iterative optimization algorithm to generate a VMD candidate parameter combination, and a predicted fitness value is obtained, and the specific steps are as follows, Initialize the population and search boundary of the iterative optimization algorithm, and generate a VMD candidate parameter combination through the internal mechanism of the iterative optimization algorithm. The collected geoscience signal sample and the VMD candidate parameter combination are input into the trained CNN-LSTM deep learning agent model to output a predicted fitness value. 6.The geoscience signal processing method based on deep learning acceleration variational modal decomposition optimization of claim 5, wherein: The predicted fitness value is fed back to the iterative optimization algorithm to obtain an optimal VMD parameter combination, and the specific steps are as follows, The predicted fitness value is fed back to the iterative optimization algorithm to update the individual position of the population and generate a new VMD candidate parameter combination. The trained CNN-LSTM deep learning agent model is driven by the new VMD candidate parameter combination to repeatedly predict and feedback until the iterative optimization algorithm converges, and the optimal VMD parameter combination is obtained.

7. The geoscience signal processing method based on deep learning acceleration variational modal decomposition optimization of claim 6, wherein: Based on the optimal VMD parameter combination, the original geoscience signal to be processed is decomposed by VMD to obtain IMF components, and the IMF components are screened and reconstructed to obtain a denoised geoscience signal, and the specific steps are as follows, The optimal VMD parameter combination is used as the VMD decomposition parameter to perform VMD decomposition on the original geoscience signal to be processed to obtain an IMF component set. Based on the IMF component set, the Pearson correlation coefficient of each IMF component and the original geoscience signal is calculated, and effective IMF components are selected according to the Pearson correlation coefficient threshold. The effective IMF components are superimposed and reconstructed to obtain a denoised geoscience signal. The original geoscience signal refers to an original noisy transient electromagnetic signal to be processed.

Images