A seismic waveform signal automatic identification method and system based on transfer learning

Abstract

This application relates to an automatic seismic waveform signal recognition method and system based on transfer learning. The method includes: acquiring continuous seismic waveform data of a target area and preprocessing it to obtain target domain waveform samples; inputting the samples into a transfer learning recognition model constructed based on physical consistency constraints, extracting multi-scale spatiotemporal features and globally aggregating them to obtain a global feature vector to be identified; calculating the initial posterior probability of each event type and calculating the physical consistency check value during the inference stage; obtaining the calibrated posterior probability based on the initial posterior probability and the physical consistency check value of each event type, and outputting the recognition result after judging by a confidence threshold. This method can suppress negative transfer phenomena during cross-tectonic region migration without requiring manually labeled data in the target domain, achieving high-precision automatic recognition.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of geophysical exploration and deep learning, and in particular relates to an automatic identification method and system for seismic waveform signals based on transfer learning. Background Technology

[0002] With the development of earthquake monitoring technology and the continuous increase in the density of seismic networks, the deep learning-based automatic seismic waveform recognition technology has made significant progress. This technology can automatically detect earthquake events and identify event types from continuous seismic waveform records, greatly improving the efficiency and automation level of massive earthquake data processing.

[0003] When it is necessary to transfer a deep learning recognition model trained in one region to a new region with significantly different geological structures, existing deep learning-based automatic seismic waveform recognition technologies generally adopt two types of strategies: one is simple fine-tuning, which fixes the weights of some network layers in the pre-trained model and retrains only the last few layers with a small learning rate to adapt the model to the waveform characteristics of the new target region. However, this method requires a certain amount of manually labeled data in the new target region, which is often difficult to meet in sparse seismic network areas or newly deployed seismic networks. The other is unsupervised domain adaptation, which uses adversarial training and other means to put a domain discriminator and a feature encoder in an adversarial game without the need for target domain labels. This minimizes the difference in probability distribution between the source domain and the target domain in the abstract feature space, thereby giving the model cross-domain generalization ability.

[0004] However, current unsupervised domain adaptation methods have a fundamental flaw: these statistically driven domain alignment techniques forcibly match the seismic waveform features of the source and target domains mathematically, completely ignoring the physical laws governing the propagation and attenuation of seismic waves in different geological regions. Since seismic waveforms are formed by the convolution of three complex physical processes—the source, the propagation path, and the site response—the path effects in different geological regions are drastically different. While the "domain-invariant features" learned solely through statistical distribution alignment may be mathematically indistinguishable, they could represent a hybrid representation that violates the wave theory of the target region. This flaw directly leads to a "negative transfer" phenomenon during feature alignment; that is, after cross-domain transfer, although the source domain classification accuracy is maintained, the accuracy in identifying weak signals and complex phases in the target domain significantly decreases. Summary of the Invention

[0005] Therefore, it is necessary to provide a transfer learning-based automatic seismic waveform signal identification method and system that can solve the problems of traditional transfer learning methods relying on manually labeled data of the target domain, purely statistical domain alignment techniques ignoring the physical laws of seismic wave propagation leading to negative migration when migrating across tectonic regions, and a sharp drop in the accuracy of identifying weak seismic signals and complex seismic phases in the target domain.

[0006] In a first aspect, this application provides an automatic identification method for seismic waveform signals based on transfer learning, including:

[0007] S1. Obtain continuous seismic waveform data of the target area to be identified;

[0008] S2. Preprocess the continuous seismic waveform data of the target area to obtain waveform samples of the target domain;

[0009] S3. Input the target domain waveform sample into the transfer learning recognition model constructed based on physical consistency constraints. The feature encoder in the transfer learning recognition model performs multi-scale spatiotemporal feature extraction and global aggregation on the target domain waveform sample to obtain the global feature vector to be recognized.

[0010] S4. Perform classification probability mapping on the global feature vector to be identified to obtain the initial posterior probability of each event type;

[0011] S5. Perform physical consistency constraint calculation on the target domain waveform sample and the global feature vector to be identified to obtain the physical consistency verification value in the inference stage;

[0012] S6. Based on the initial posterior probability and the physical consistency check value, calculate the calibrated posterior probability, and perform confidence threshold discrimination on the calibrated posterior probability to obtain the seismic waveform signal recognition result.

[0013] In one embodiment, the transfer learning recognition model built based on physical consistency constraints is trained in the following manner:

[0014] S31. Obtain the source domain seismic waveform dataset and the target domain unlabeled seismic waveform dataset, and preprocess the source domain seismic waveform dataset and the target domain unlabeled seismic waveform dataset to obtain labeled source domain waveform samples and unlabeled target domain waveform samples.

[0015] S32. Input the source domain waveform sample and the target domain waveform sample into the pre-constructed initial deep neural network to obtain the first global feature vector and the second global feature vector; wherein, the first global feature vector is used to characterize the global feature vector of the source domain waveform sample, and the second global feature vector is used to characterize the global feature vector of the target domain waveform sample.

[0016] S33. Perform classification prediction on the first global feature vector to obtain the source domain classification loss;

[0017] S34. Perform domain adversarial calculations on the first and second global feature vectors to obtain the domain adversarial loss;

[0018] S35. Based on the second global feature vector, a self-supervised physical consistency constraint is applied to the waveform samples in the target domain to obtain the physical consistency loss.

[0019] S36. The source domain classification loss, domain adversarial loss, and physical consistency loss are weighted and summed to obtain the total loss.

[0020] S37. Based on the total loss, backpropagate the parameters of the initial deep neural network to update them until the preset convergence condition is met, and the trained transfer learning recognition model based on physical consistency constraints is obtained.

[0021] In one embodiment, S35 includes:

[0022] S351. The three-channel waveform data contained in the current batch of unlabeled target domain waveform samples are zero-padding and fast Fourier transformed within a preset time window to obtain the complex spectrum of the three channels.

[0023] S352. Extract the amplitude spectrum of the three-channel complex spectrum and calculate the average value to obtain the observed amplitude spectrum;

[0024] S353. Perform a nonlinear mapping transformation on the second global feature vector to obtain the set of prediction parameters;

[0025] S354. Perform forward modeling calculations on the predicted parameter set based on the crustal medium physical forward modeling model to obtain the theoretical amplitude spectrum;

[0026] S355. Map the observed amplitude spectrum and the theoretical amplitude spectrum to the double logarithmic domain to obtain the double logarithmic observed amplitude spectrum and the double logarithmic theoretical amplitude spectrum. Calculate the mean square error between the double logarithmic observed amplitude spectrum and the double logarithmic theoretical amplitude spectrum, and use the mean square error as the physical consistency loss.

[0027] In one embodiment, the expression for the theoretical amplitude spectrum is:

[0028]

[0029] In the formula, Represents the theoretical amplitude spectrum. This represents the frequency value corresponding to a single sampling point on the discrete frequency axis after the Fast Fourier Transform. Represents the set of prediction parameters. The parameter representing the asymptotic level of the low-frequency displacement spectrum. This represents the parameter that controls the corner frequency of the seismic source. Represented by natural constant An exponential function with base 0. This represents the parameter that controls the decay along the entire path. This indicates a high-frequency attenuation trend correction index. It represents pi (π).

[0030] In one embodiment, the expression for the posterior probability is:

[0031]

[0032] In the formula, Represents the posterior probability. Represents the event category variable, Indicates the specific first Event categories, This represents the global feature vector to be identified. This represents a preset coefficient used to balance the weights of the initial posterior probability and the physical consistency penalty term. This represents the global feature vector to be identified, output by the flexible maximum function. Belongs to the The initial posterior probability of the event category. Represented by natural constant An exponential function with base 0. This indicates that the rate at which the physical consistency check value decays with the posterior probability is greater than [a certain value]. The attenuation coefficient, This represents the physical consistency check value.

[0033] In one embodiment, S2 includes:

[0034] S21. Extract the original digital counts of each channel from the continuous seismic waveform data of the target area, perform deconvolution operation on the original digital counts, and obtain the waveform data after instrument response correction.

[0035] S22. Filter the waveform data after instrument response correction to obtain filtered waveform data;

[0036] S23. Perform sliding window truncation on the filtered waveform data according to the preset time window length and sliding step size to generate equal-length three-channel waveform segments.

[0037] S24. Perform mean removal and normalization based on the maximum amplitude of 2 on each three-channel waveform segment to obtain the target domain waveform sample.

[0038] Secondly, this application also provides an automatic seismic waveform signal identification system based on transfer learning, comprising:

[0039] The data acquisition module is used to acquire continuous seismic waveform data of the target area to be identified;

[0040] The preprocessing module is used to preprocess continuous seismic waveform data of the target area to obtain waveform samples of the target domain;

[0041] The feature extraction module is used to input the target domain waveform samples into the transfer learning recognition model built based on physical consistency constraints. The feature encoder in the transfer learning recognition model performs multi-scale spatiotemporal feature extraction and global aggregation on the target domain waveform samples to obtain the global feature vector to be recognized.

[0042] The classification calculation module is used to perform classification calculations on the global feature vector to be identified, and obtain the initial posterior probability of each event type.

[0043] The physical consistency constraint module is used to calculate the physical consistency constraint of the target domain waveform sample and the global feature vector to be identified, and obtain the physical consistency verification value in the inference stage.

[0044] The posterior calibration module is used to calculate the calibrated posterior probability based on the initial posterior probability and the physical consistency check value, and to perform confidence threshold discrimination on the calibrated posterior probability to obtain the seismic waveform signal recognition result.

[0045] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the methods described above.

[0046] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0047] The aforementioned automatic seismic waveform signal identification method and system based on transfer learning, by acquiring continuous seismic waveform data of the target area to be identified, can provide a full-time, full-coverage raw data foundation for cross-regional automatic seismic waveform identification, adapting to the seismic monitoring needs of areas with different network densities. Through standardized preprocessing of the continuous seismic waveform data of the target area, non-seismic interference factors can be effectively eliminated, generating target domain waveform samples with uniform format and amplitude normalization, laying a high-quality data foundation for subsequent feature extraction and calculation in the model. By inputting the target domain waveform samples into a transfer learning identification model constructed based on physical consistency constraints, and using a feature encoder to complete multi-scale spatiotemporal feature extraction and global aggregation, it can simultaneously capture the local phase characteristics and global evolution laws of seismic waveforms at different time scales, generating a globally characteristic vector to be identified that is both discriminative and representative, solving the problem that single-scale features are insufficient to comprehensively represent complex seismic waveforms. The initial and subsequent values ​​of each event type are obtained by classifying and calculating the globally characteristic vector to be identified. The probability of posterior probabilities can leverage the nonlinear fitting capabilities of deep learning to perform preliminary discrimination of earthquake event types, achieving efficient and automated initial screening of massive continuous waveform data. By calculating physical consistency constraints on waveform samples in the target domain and global feature vectors to be identified, a physical consistency verification value is obtained. This organically integrates crustal medium wave theory with deep learning models, constraining the direction of feature learning from the physical laws of seismic wave propagation. This avoids invalid features that violate the geological characteristics of the target area generated by purely statistically driven domain alignment methods, fundamentally suppressing negative migration phenomena in the cross-domain migration process. By combining the initial posterior probability with the physical consistency verification value to perform posterior probability calibration, and obtaining the final identification result based on the confidence threshold, the physical consistency information can be used to scientifically correct the statistical classification results of the model, improving the identification accuracy and robustness of weak seismic signals and complex seismic phases in the target domain. This addresses the core pain points of traditional transfer learning methods, such as the sharp drop in identification accuracy in new areas with huge differences in geological structure and the reliance on a large amount of manually labeled data. This method enables high-precision automatic identification of cross-regional seismic waveform signals without the need for manual annotation of data in the target domain, thereby improving the automation level and identification reliability of seismic monitoring in sparse and newly deployed seismic networks. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1A flowchart illustrating an automatic seismic waveform signal identification method based on transfer learning, provided as an exemplary embodiment of this application;

[0050] Figure 2 A schematic diagram of a process for generating physical consistency loss is provided as an exemplary embodiment of this application;

[0051] Figure 3 This is a schematic diagram of the structure of an automatic seismic waveform signal identification system based on transfer learning, provided as an exemplary embodiment of this application. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0053] In one embodiment, such as Figure 1 As shown, an automatic seismic waveform signal identification method based on transfer learning is provided. This embodiment illustrates the application of this method to an intelligent identification terminal. It is understood that this method can also be applied to an intelligent identification server, and further to a system including both an intelligent identification terminal and an intelligent identification server, and is implemented through the interaction between the intelligent identification terminal and the intelligent identification server. In this embodiment, the method includes the following steps:

[0054] S1. Obtain continuous seismic waveform data of the target area to be identified.

[0055] Specifically, the intelligent identification terminal can obtain uninterrupted raw seismic waveform data streams from the target area through an encrypted real-time data communication link, and connect to an authoritative data platform to import historically archived continuous seismic waveform data in batches.

[0056] Optionally, authoritative data platforms may include, but are not limited to, the National Earthquake Science Data Sharing Center and regional seismic network centers.

[0057] Optionally, the continuous seismic waveform data of the target area can be three-component seismic waveform data containing three vibration components in the vertical, north-south, and east-west directions. The data sampling rate conforms to the general observation standards of the global seismic network, and the data format adopts the standard exchange format commonly used in the field of seismology. It can completely record the vibration characteristics and various natural and anthropogenic background noise signals during the propagation of seismic waves.

[0058] S2. Preprocess the continuous seismic waveform data of the target area to obtain waveform samples of the target domain.

[0059] Specifically, the intelligent identification terminal can extract the original digital counts of each channel in the continuous seismic waveform data of the target area, and eliminate the distortion effects of the seismograph's inherent frequency response, sensitivity coefficient and amplification factor on the waveform through instrument response deconvolution operation, so as to obtain displacement, velocity or acceleration waveform data that truly reflects the ground vibration of the crustal medium.

[0060] Furthermore, the intelligent recognition terminal can filter the corrected waveform data to remove low-frequency baseline drift and high-frequency electromagnetic interference noise that exceed the effective seismic signal frequency band. The intelligent recognition terminal can perform sliding window truncation on the filtered waveform data according to a predefined time window length and sliding step size, generating equal-length three-channel waveform segments. The intelligent recognition terminal can then sequentially perform mean removal and normalization based on the maximum amplitude on each three-channel waveform segment, eliminating model input bias caused by DC offset and amplitude differences, thus obtaining target domain waveform samples.

[0061] Optionally, the target domain waveform samples can be used to characterize the comprehensive characteristics of ground vibration in the target area over a specific time period.

[0062] S3. Input the target domain waveform sample into the transfer learning recognition model constructed based on physical consistency constraints. The feature encoder in the transfer learning recognition model performs multi-scale spatiotemporal feature extraction and global aggregation on the target domain waveform sample to obtain the global feature vector to be recognized.

[0063] Specifically, the intelligent recognition terminal can input pre-processed target domain waveform samples into a pre-trained and converged transfer learning recognition model based on physical consistency constraints. The feature encoder inside the transfer learning recognition model can adopt a deep architecture that combines multi-scale dilated convolution with temporal self-attention mechanism. Through dilated convolutional layers with different dilation coefficients, it can extract local spatiotemporal features of the waveform at different time scales in parallel, extract the morphological features, amplitude changes and arrival time differences of different phases, and perform adaptive weighted aggregation of multi-scale local features through global temporal attention pooling layer, fuse the global evolution context information of the waveform, and generate a global feature vector to be recognized with fixed dimensions.

[0064] Optionally, the transfer learning recognition model built based on physical consistency constraints can be an end-to-end deep neural network model that integrates unsupervised domain adversarial training and forward modeling constraints of crustal medium fluctuation theory, simultaneously optimizing three types of loss functions during the training phase: source domain classification loss, domain adversarial loss, and physical consistency loss.

[0065] Optionally, the global feature vector to be identified can be used to characterize the essential discriminative features of the waveform sample in the target domain. The global feature vector to be identified may include, but is not limited to, statistical separability information and physical interpretability information.

[0066] S4. Perform classification probability mapping on the global feature vector to be identified to obtain the initial posterior probability of each event type.

[0067] Specifically, the intelligent recognition terminal can map the high-dimensional global feature vector to be recognized to a low-dimensional category feature space through cascaded operations of multi-layer linear transformations and non-linear activation functions, thus completing the initial association between features and category labels. The intelligent recognition terminal can normalize the output values ​​of the category feature space using a flexible maximum function, converting the output values ​​into probability values ​​ranging from 0 to 1, with the sum of the probability values ​​corresponding to all preset event types being 1, thus obtaining the initial posterior probability that the waveform sample to be recognized belongs to each preset event type.

[0068] Optionally, the initial posterior probability of each event type can be used to characterize the feature discrimination information obtained by the model based on statistical learning, and to make a preliminary confidence judgment on the event type to which the waveform sample belongs. The preset event types may include, but are not limited to, vibration event types commonly found in the field of earthquake monitoring.

[0069] S5. Perform physical consistency constraint calculation on the target domain waveform sample and the global feature vector to be identified to obtain the physical consistency verification value in the inference stage.

[0070] Specifically, the intelligent recognition terminal can perform zero-padding and Fast Fourier Transform on the three-channel waveform data of the target domain waveform sample within a preset analysis time window, convert the time-domain waveform data into a frequency-domain complex spectrum, extract the amplitude spectrum of each channel's complex spectrum, and calculate the average value of the three channels to obtain the observed amplitude spectrum.

[0071] Furthermore, the intelligent recognition terminal can transform the global feature vector to be identified through two nonlinear mapping layers to obtain a set of predicted parameters. The intelligent recognition terminal can then input the set of predicted parameters into a physical forward model constructed based on the theory of crustal medium fluctuations for forward modeling calculations to obtain the theoretical amplitude spectrum.

[0072] Furthermore, the intelligent recognition terminal can map the observed amplitude spectrum and the theoretical amplitude spectrum to the double logarithmic domain, calculate the mean square error between the observed amplitude spectrum and the theoretical amplitude spectrum, and use the mean square error as the physical consistency verification value in the inference stage.

[0073] Optionally, the three-channel waveform data can be used to characterize the continuous evolution of the ground's vibration state over time in three-dimensional orthogonal space, corresponding to three independent vibration components in the vertical, north-south horizontal, and east-west horizontal directions, respectively. It can completely record all the kinematic characteristics of the compression and stretching vibration of the longitudinal wave, the shear vibration of the transverse wave, and the elliptic polarization vibration of the surface wave during the propagation of the seismic wave.

[0074] Optionally, the physical consistency check value can be used to characterize the degree of difference between the physical parameters inverted from the waveform features and the frequency domain physical features of the actual observed waveform. The smaller the physical consistency check value, the more the learned features conform to the physical laws of seismic wave propagation in the target area.

[0075] S6. Based on the initial posterior probability and the physical consistency check value, calculate the calibrated posterior probability, and perform confidence threshold discrimination on the calibrated posterior probability to obtain the seismic waveform signal recognition result.

[0076] Specifically, the intelligent recognition terminal can use a preset posterior probability calibration formula to calculate the calibrated posterior probability by combining the initial posterior probability and the physical consistency check value. During the calibration process, the influence weights of the statistical classification results and physical consistency constraints are adjusted through preset balancing weight coefficients. An exponentially decaying physical penalty term is used to correct the initial posterior probability that does not conform to the physical laws of the target area, thereby reducing the classification confidence of physically contradictory samples. The intelligent recognition terminal can compare the calibrated posterior probability corresponding to each preset event type with a preset confidence threshold class by class. If the calibrated posterior probability of a certain preset event type is greater than or equal to the confidence threshold, the waveform sample is determined to belong to that preset event type; if the calibrated posterior probabilities of all preset event types are less than the confidence threshold, it is determined to be an unknown vibration event.

[0077] Optionally, the preset event type is a predefined category of common vibration events in the field of earthquake monitoring, which may include, but is not limited to, natural earthquake events, artificial blasting events, mine collapse events, industrial vibration events, and traffic vibration events.

[0078] Optionally, the seismic waveform signal identification results can be used to characterize the types of vibration events detected in the continuous waveforms of the target area and the calibrated confidence levels corresponding to the vibration event types.

[0079] In the aforementioned automatic seismic waveform signal recognition method based on transfer learning, by acquiring continuous seismic waveform data of the target area to be identified, a full-time, full-coverage raw data foundation can be provided for cross-regional automatic seismic waveform recognition, adapting to the seismic monitoring needs of areas with different network densities. By standardizing and preprocessing the continuous seismic waveform data of the target area, non-seismic interference factors can be effectively eliminated, generating target domain waveform samples with uniform format and amplitude normalization, laying a high-quality data foundation for subsequent feature extraction and calculation in the model. By inputting the target domain waveform samples into a transfer learning recognition model constructed based on physical consistency constraints, and using a feature encoder to complete multi-scale spatiotemporal feature extraction and global aggregation, the local phase characteristics and global evolution laws of seismic waveforms at different time scales can be captured simultaneously, generating a global feature vector to be identified that is both discriminative and representative, solving the problem that single-scale features cannot comprehensively represent complex seismic waveforms. Initial posteriors for each event type are obtained by classifying and calculating the global feature vector to be identified. Probability, relying on the nonlinear fitting capabilities of deep learning, can perform preliminary discrimination of earthquake event types, achieving efficient and automated initial screening of massive continuous waveform data. By calculating the physical consistency constraint on the waveform samples of the target domain and the global feature vector to be identified, a physical consistency verification value is obtained. This organically integrates the theory of crustal medium fluctuation with the deep learning model, constraining the feature learning direction from the physical laws of seismic wave propagation. This avoids invalid features that violate the geological characteristics of the target area generated by purely statistical domain alignment methods, fundamentally suppressing the negative migration phenomenon in the cross-domain migration process. By combining the initial posterior probability with the physical consistency verification value to perform posterior probability calibration, and obtaining the final identification result based on the confidence threshold, the physical consistency information can be used to scientifically correct the statistical classification results of the model, improving the identification accuracy and robustness of weak seismic signals and complex seismic phases in the target domain. This can solve the core pain points of traditional transfer learning methods, such as the sharp drop in identification accuracy in new areas with huge differences in geological structure and the reliance on a large amount of manually labeled data. This method enables high-precision automatic identification of cross-regional seismic waveform signals without the need for manual annotation of data in the target domain, thereby improving the automation level and identification reliability of seismic monitoring in sparse and newly deployed seismic networks.

[0080] In one embodiment, the transfer learning recognition model built based on physical consistency constraints is trained in the following ways, including:

[0081] S31. Obtain the source domain seismic waveform dataset and the target domain unlabeled seismic waveform dataset, and preprocess the source domain seismic waveform dataset and the target domain unlabeled seismic waveform dataset to obtain labeled source domain waveform samples and unlabeled target domain waveform samples.

[0082] For example, the intelligent identification terminal can obtain source domain seismic waveform datasets that have been professionally annotated from a globally authoritative earthquake science data sharing platform, and connect to the real-time data interface and historical data archiving platform of the target area seismic network to obtain continuous unlabeled seismic waveform data of the target area throughout the entire time period. The intelligent identification terminal can adopt a standardized preprocessing process that is completely consistent with the identification stage, sequentially performing instrument response deconvolution correction, bandpass filtering, equal-length sliding window truncation, mean removal, and maximum amplitude normalization on the source domain seismic waveform dataset and the target domain unlabeled seismic waveform dataset, and removing invalid waveform segments with discontinuities, saturation, or strong interference, generating labeled source domain waveform samples and unlabeled target domain waveform samples with uniform format.

[0083] Optionally, the source domain seismic waveform dataset can be used to characterize typical waveform features of various vibration events in areas with sufficient annotation resources and clear geological structural features.

[0084] Optionally, the unlabeled seismic waveform dataset of the target domain can be used to characterize the unique crustal medium propagation characteristics, site response features, and background noise distribution patterns of the area to be identified.

[0085] S32. Input the source domain waveform samples and the target domain waveform samples into the pre-constructed initial deep neural network to obtain the first global feature vector and the second global feature vector.

[0086] For example, the intelligent recognition terminal can batch input labeled source domain waveform samples and unlabeled target domain waveform samples from the same training batch into a pre-constructed initial deep neural network. The initial deep neural network can adopt a weight-sharing feature encoder architecture, which is completely consistent with the feature extraction structure in the recognition stage. The feature encoder can extract local seismic phase features at different time scales in parallel through multi-scale dilated convolutional layers, and complete the adaptive weighted aggregation of features through a global temporal attention pooling layer, outputting the first global feature vector corresponding to the source domain waveform sample and the second global feature vector corresponding to the target domain waveform sample, respectively.

[0087] Optionally, the first global feature vector can be used to characterize the global feature vector of the source domain waveform sample, and the second global feature vector can be used to characterize the global feature vector of the target domain waveform sample.

[0088] S33. Perform classification prediction on the first global feature vector to obtain the source domain classification loss.

[0089] For example, the intelligent recognition terminal can input the first global feature vector into the fully connected classification branch of the initial deep neural network, and through the cascade operation of multi-layer linear transformation and nonlinear activation function, map the high-dimensional global features to the low-dimensional category feature space, and then perform normalization processing by the flexible maximum value function to obtain the predicted probability distribution of the source domain samples belonging to each preset event type.

[0090] Furthermore, the intelligent recognition terminal can use the Cross Entropy Loss function to calculate the difference between the predicted probability distribution and the true labeled data of the source domain samples, and use this difference as the source domain classification loss to supervise the initial deep neural network in learning the discrimination patterns of various events in the source domain.

[0091] Preferably, the difference between the predicted probability distribution and the true labeled values ​​of the source domain samples can be calculated using the following formula:

[0092]

[0093] In the formula, Represents the source domain classification loss. This represents the number of source domain samples in each training batch. This indicates the total number of preset event types. Indicates the first The true label of the sample (if the sample belongs to the first sample) (If it is a class, it is 1; otherwise, it is 0). This indicates the initial deep neural network prediction of the first... The sample belongs to the first The initial posterior probability of the class is used to supervise the initial deep neural network learning the discriminative features of various events in the source domain.

[0094] Optionally, the source domain classification loss can be used to characterize the classification prediction accuracy of the transfer learning recognition model on the source domain dataset. The smaller the loss value, the stronger the model's ability to recognize known event types in the source domain.

[0095] S34. Perform domain adversarial calculations on the first and second global feature vectors to obtain the domain adversarial loss.

[0096] For example, the intelligent recognition terminal can connect a gradient reversal layer to the output of the feature encoder, processing the first and second global feature vectors before inputting them into the neighborhood discriminator. During training, the neighborhood discriminator can employ an adversarial game mechanism to learn to distinguish between the source and target domains of the feature vectors. The feature encoder can adjust its parameters using the negative gradient propagated in reverse by the gradient reversal layer to confuse the neighborhood discriminator's judgment, and the classification error of the neighborhood discriminator can be used as the neighborhood adversarial loss.

[0097] Optionally, domain adversarial loss can be used to characterize the degree of difference in the statistical distribution of source domain waveform samples and target domain waveform samples in the abstract feature space. The smaller the loss value, the higher the alignment of the feature distributions of the source domain and the target domain, and the stronger the cross-domain statistical generalization ability of the transfer learning recognition model.

[0098] S35. Based on the second global feature vector, a self-supervised physical consistency constraint is applied to the waveform samples in the target domain to obtain the physical consistency loss.

[0099] For example, the intelligent identification terminal can perform zero-padding and fast Fourier transform on the unlabeled target domain waveform samples of the current batch within a preset analysis window to obtain a three-channel complex spectrum and calculate the average observed amplitude spectrum.

[0100] Furthermore, the intelligent identification terminal can input the second global feature vector into two nonlinear mapping layers to invert and obtain a set of predicted parameters containing source parameters and propagation path parameters. The set of predicted parameters is then substituted into a physical forward model constructed based on crustal medium wave theory to obtain the theoretical amplitude spectrum. The mean square error between the observed amplitude spectrum and the theoretical amplitude spectrum in the double logarithmic domain is calculated, and the mean square error between the observed amplitude spectrum and the theoretical amplitude spectrum in the double logarithmic domain is used as the physical consistency loss.

[0101] Optionally, the physical consistency loss can be used to characterize the degree of matching between the physical parameters inverted from the target domain features by the transfer learning recognition model and the physical features of the actual observed waveform. The smaller the loss value, the more the features learned by the transfer learning recognition model conform to the seismic wave propagation law of the target area.

[0102] S36. The source domain classification loss, domain adversarial loss, and physical consistency loss are weighted and summed to obtain the total loss.

[0103] For example, the intelligent recognition terminal can pre-set weight coefficients for the source domain classification loss, domain adversarial loss, and physical consistency loss, respectively. These weight coefficients can be dynamically adjusted based on the annotation quality of the source domain data and the degree of difference in geological structure between the target domain and the source domain. The intelligent recognition terminal can then perform a weighted summation of the source domain classification loss, domain adversarial loss, and physical consistency loss according to the pre-set weight coefficients to obtain the total loss used for updating the parameters of the transfer learning recognition model.

[0104] Preferably, the source domain classification loss, domain adversarial loss, and physical consistency loss can be weighted and summed using the following formula:

[0105]

[0106] In the formula, Indicates the total loss. The weight coefficients represent the source domain classification loss. The weighting coefficients represent the domain adversarial loss. The weighting coefficients representing the physical consistency loss. Represents the source domain classification loss. Indicates domain-based combat losses. This represents the loss of physical consistency.

[0107] S37. Based on the total loss, backpropagate the parameters of the initial deep neural network to update them until the preset convergence condition is met, and the trained transfer learning recognition model based on physical consistency constraints is obtained.

[0108] For example, the intelligent recognition terminal can employ an improved optimization algorithm of stochastic gradient descent to calculate the gradient of the total loss relative to the trainable parameters of each layer of the initial deep neural network. This gradient is then propagated layer by layer from the output layer to the input layer using the backpropagation algorithm, updating the weight parameters of each layer sequentially. After each training round, the intelligent recognition terminal verifies the model performance. When the total loss decreases to a preset threshold, the accuracy on the validation set shows no significant improvement over multiple consecutive rounds, or the preset maximum number of training rounds is reached, the transfer learning recognition model is deemed to have met the convergence condition. Training is then stopped, and the final model parameters are saved.

[0109] In this embodiment, the intelligent recognition terminal constructs a ternary joint optimization framework that integrates source domain classification loss, domain adversarial loss, and physical consistency loss. Without requiring manually labeled data in the target domain, it collaboratively guides the parameter update direction of the feature encoder from two dimensions: statistical distribution alignment and forward modeling constraints of crustal medium wave theory. This enables the trained transfer learning recognition model to actively learn feature representations that conform to the physical laws of seismic wave propagation in the target area while maintaining the ability to distinguish events in the source domain, thereby fundamentally suppressing the negative migration phenomenon when migrating across tectonic regions.

[0110] In one embodiment, such as Figure 2 As shown, a flowchart illustrating the generation of physical consistency loss is provided, S35, which may include:

[0111] S351. The three-channel waveform data contained in the current batch of unlabeled target domain waveform samples are zero-padding and fast Fourier transformed within a preset time window to obtain the complex spectrum of the three channels.

[0112] For example, the intelligent identification terminal can extract the effective data segment of each unlabeled target domain waveform sample in the current batch within a preset analysis time window. This time window covers the complete vibration process of the seismic wave from its initial wave to its final wave decay, resulting in the extracted time-domain waveform data. The intelligent identification terminal can pad the extracted time-domain waveform data with zeros at the end to ensure that the length of the extracted time-domain waveform data meets the operational requirements of the Fast Fourier Transform (FFT). The intelligent identification terminal can perform FFT operations on the time-domain waveforms of the vertical, north-south, and east-west channels respectively, converting the discrete time-domain signal into a complex representation in the continuous frequency domain, thus obtaining the complex spectrum of the three channels.

[0113] Optionally, the complex spectrum of the three channels can be used to characterize the vibration characteristics of the three orthogonal vibration components at different frequencies. The complex spectrum corresponding to each frequency point may include, but is not limited to, the amplitude and phase information of that frequency component.

[0114] S352. Extract the amplitude spectrum of the three-channel complex spectrum and calculate the average value to obtain the observed amplitude spectrum.

[0115] For example, the intelligent recognition terminal can perform modulo operations on the complex spectrum of each channel at each frequency point to obtain the amplitude spectrum corresponding to the channel. The intelligent recognition terminal can also perform an arithmetic mean operation on the amplitude spectrum values ​​of the three channels at the same frequency point to obtain the average amplitude spectrum that integrates the vibration information of the three channels, i.e., the observed amplitude spectrum.

[0116] Optionally, the numerical value of the amplitude spectrum can be used to represent the vibrational energy intensity of that frequency component.

[0117] Optionally, the observed amplitude spectrum can be used to characterize the overall energy distribution characteristics of seismic events and background noise in the target area in the frequency domain.

[0118] S353. Perform a nonlinear mapping transformation on the second global feature vector to obtain the set of prediction parameters.

[0119] For example, the intelligent recognition terminal can input the second global feature vector corresponding to each target domain waveform sample in the current batch into a two-layer cascaded fully connected nonlinear mapping layer. Each fully connected nonlinear mapping layer is followed by a nonlinear activation function to achieve nonlinear transformation of the features. Through multi-layer nonlinear operations, the intelligent recognition terminal can map the high-dimensional abstract statistical feature vector to a low-dimensional physical parameter space, and output a vector with the same dimension as the preset number of physical parameters, i.e., the set of predicted parameters.

[0120] Optionally, the set of predicted parameters can be used to characterize the key physical parameters inverted from the global characteristics of the target domain waveform, covering the core parameters that control the low-frequency radiation level of the source, the source corner frequency, the full path attenuation coefficient, and the high-frequency attenuation correction coefficient, and can fully describe the main physical influencing factors in the process of seismic waves from source generation to station reception.

[0121] S354. Perform forward modeling calculations on the predicted parameter set based on the crustal medium physical forward modeling model to obtain the theoretical amplitude spectrum.

[0122] For example, the intelligent identification terminal can substitute the set of predicted parameters into a pre-constructed forward model of crustal medium physics. This forward model can be constructed based on elastic wave propagation theory, comprehensively considering the dislocation radiation mode of the seismic source, the geometric diffusion effect during seismic wave propagation, the inelastic absorption and attenuation effect of the crustal medium, and the frequency response characteristics of the local site. The forward model of crustal medium physics can be numerically calculated by substituting the predicted parameters to obtain a theoretical frequency domain energy distribution curve, i.e., a theoretical amplitude spectrum, that is consistent with the frequency range of the observed amplitude spectrum.

[0123] Optionally, the theoretical amplitude spectrum can be used to characterize the ideal frequency domain energy distribution derived from the physical laws of seismic wave propagation under given source parameters and propagation path parameters. The shape of the theoretical amplitude spectrum is completely determined by the set of input prediction parameters.

[0124] S355. Map the observed amplitude spectrum and the theoretical amplitude spectrum to the double logarithmic domain to obtain the double logarithmic observed amplitude spectrum and the double logarithmic theoretical amplitude spectrum. Calculate the mean square error between the double logarithmic observed amplitude spectrum and the double logarithmic theoretical amplitude spectrum, and use the mean square error as the physical consistency loss.

[0125] For example, the intelligent identification terminal can take the natural logarithm of the value of each frequency point in the observed amplitude spectrum and the theoretical amplitude spectrum, and convert the amplitude spectrum in the original linear domain into the amplitude spectrum in the double logarithmic domain, so that the exponential decay characteristic of the seismic wave spectrum is transformed into a linear relationship.

[0126] Furthermore, the intelligent recognition terminal can calculate the sum of squares of the differences between the double logarithmic observed amplitude spectrum and the double logarithmic theoretical amplitude spectrum at all corresponding frequency points, divide it by the total number of frequency points to obtain the mean square error, and use this mean square error as the physical consistency loss during the training phase.

[0127] Preferably, the mean square error between the double logarithmic observed amplitude spectrum and the double logarithmic theoretical amplitude spectrum can be calculated using the following formula:

[0128]

[0129] In the formula, This represents the mean square error between the observed amplitude spectrum and the theoretical amplitude spectrum based on the double logarithmic method. This represents the total number of discrete frequency points within the effective frequency range. Indicates the observed amplitude spectrum at frequency The amplitude at that point, Indicates the theoretical amplitude spectrum at frequency The amplitude at that point, This represents the set of prediction parameters.

[0130] Optionally, the double logarithmic observed amplitude spectrum can be used to characterize the linearized representation of the observed amplitude spectrum in a double logarithmic coordinate system.

[0131] Optionally, the double logarithmic theoretical amplitude spectrum can be used to characterize the linearized representation of the theoretical amplitude spectrum in a double logarithmic coordinate system.

[0132] In this embodiment, the intelligent recognition terminal constructs a physical consistency loss constraint that integrates frequency domain physical observation, feature nonlinear inversion, and differentiable forward modeling spectral shape comparison. This enables the feature encoder to actively learn a feature representation that can self-consistently restore its true physical spectral shape using only unlabeled waveform samples in the target domain. This fundamentally constrains the physical rationality of feature alignment during the transfer learning process and effectively suppresses negative migration phenomena that violate the propagation laws of the crustal medium in the target area.

[0133] In one embodiment, the expression for the theoretical amplitude spectrum can be:

[0134]

[0135] In the formula, Represents the theoretical amplitude spectrum. This represents the frequency value corresponding to a single sampling point on the discrete frequency axis after the Fast Fourier Transform. Represents the set of prediction parameters. The parameter representing the asymptotic level of the low-frequency displacement spectrum. This represents the parameter that controls the corner frequency of the seismic source. Represented by natural constant An exponential function with base 0. This represents the parameter that controls the decay along the entire path. This represents a high-frequency attenuation trend correction index, used to compensate for local site effects and the frequency dependence of the quality factor. It represents pi (π).

[0136] For example, the intelligent recognition terminal can successively substitute the predicted parameter set obtained from the second global feature vector mapping into the theoretical amplitude spectrum expression, traverse all discrete frequency points within the preset effective seismic signal frequency range, calculate the theoretical amplitude value corresponding to each frequency point in turn, and generate a theoretical amplitude spectrum curve that is completely consistent with the frequency resolution of the observed amplitude spectrum.

[0137] In this embodiment, the intelligent recognition terminal can seamlessly integrate the theory of crustal medium fluctuations into the training and inference process of the deep learning model through parameterized differentiable physical forward modeling, so that the physical consistency loss can directly guide the parameter update of the feature encoder through the backpropagation algorithm.

[0138] In one embodiment, the expression for the posterior probability can be:

[0139]

[0140] In the formula, Represents the posterior probability. Represents the event category variable, Indicates the specific first Event categories, This represents the global feature vector to be identified. This represents a preset coefficient used to balance the weights of the initial posterior probability and the physical consistency penalty term. This represents the global feature vector to be identified, output by the flexible maximum function. Belongs to the The initial posterior probability of the event category. Represented by natural constant An exponential function with base 0. This indicates that the rate at which the physical consistency check value decays with the posterior probability is greater than [a certain value]. The attenuation coefficient, This represents the physical consistency check value.

[0141] For example, after obtaining the initial posterior probability and the physical consistency check value of each event type corresponding to a single waveform sample to be identified, the intelligent recognition terminal can substitute the corresponding parameters one by one into the calibration expression to calculate the calibrated posterior probability of each preset event type. During the calibration process, the intelligent recognition terminal can use linear weighted fusion of the basic statistical confidence and physical penalty correction term, and utilize the nonlinear decay characteristics of the exponential function to ensure that the sample with the higher the physical consistency check value has its initial posterior probability decayed by a greater margin.

[0142] In this embodiment, the intelligent identification terminal adopts a posterior probability calibration expression that integrates statistical confidence and physical consistency penalty terms. It can use the physical consistency verification value calculated during the inference stage to perform nonlinear adaptive correction on the initial posterior probability, so that the identification results that violate the physical laws of crustal medium propagation in the target area are effectively suppressed, thereby significantly improving the accuracy and robustness of automatic identification of seismic waveforms across tectonic regions.

[0143] In one embodiment, S2 may include:

[0144] S21. Extract the original digital counts of each channel from the continuous seismic waveform data of the target area, perform deconvolution operation on the original digital counts, and obtain the waveform data after instrument response correction.

[0145] For example, the intelligent identification terminal can parse standard format files of continuous seismic waveform data in the target area and extract the original digital counting sequence corresponding to each channel. The original digital counting sequence can be a discrete integer sequence obtained by analog-to-digital conversion after the seismograph converts ground vibrations into electrical signals. Furthermore, the intelligent identification terminal can load the instrument response file of the corresponding seismograph station, obtain the transfer function parameters including poles, zeros, and sensitivity coefficients, perform frequency domain deconvolution operation on the original digital counting sequence and the instrument response transfer function to eliminate the distortion effect of the seismograph's inherent frequency response and amplification factor on the real ground vibration spectrum, and convert it into a time-domain waveform through inverse Fourier transform to obtain the waveform data after instrument response correction.

[0146] Preferably, the original digital counting sequence and the instrument response transfer function can be deconvolved in the frequency domain using the following formula:

[0147]

[0148] In the formula, Indicates frequency as The actual ground vibration spectrum at that location, This represents the observed spectrum obtained by performing a Fast Fourier Transform on the original digit counting sequence. The instrument response transfer function of the seismograph. This represents the frequency value corresponding to a single sampling point on the discrete frequency axis after the Fast Fourier Transform.

[0149] S22. Filter the waveform data after the instrument response correction to obtain filtered waveform data.

[0150] For example, the intelligent identification terminal can use a Butterworth zero-phase bandpass filter of a preset order to filter the waveform data after instrument response correction. During the filtering process, the intelligent identification terminal can filter out non-target signals outside the passband based on the preset frequency passband range of the effective seismic signal to obtain the filtered waveform data.

[0151] Preferably, the waveform data after instrument response correction can be filtered using the following formula:

[0152]

[0153] In the formula, This represents the amplitude-frequency response of the Butterworth bandpass filter. This represents the frequency value corresponding to a single sampling point on the discrete frequency axis after the Fast Fourier Transform. Indicates the center frequency of the passband. Indicates passband bandwidth. This indicates the preset order of the Butterworth bandpass filter.

[0154] S23. Perform sliding window truncation on the filtered waveform data according to the preset time window length and sliding step size to generate three-channel waveform segments of equal length.

[0155] For example, the intelligent recognition terminal can use the start time of the filtered continuous waveform data as a reference and perform successive sliding window truncation according to a preset time window length and sliding step size. The preset time window length covers the entire process of a complete seismic event from the initial wave to the coda attenuation, and the sliding step size controls the overlap ratio of adjacent windows to ensure that no potential seismic events are missed. During the truncation process, the intelligent recognition terminal can simultaneously truncate waveform data in the vertical, north-south, and east-west directions, ensuring that the time axes of each three-channel waveform segment are completely aligned, generating a set of all three-channel waveform segments of consistent length and continuous time.

[0156] S24. Perform mean removal and normalization based on the maximum amplitude on each three-channel waveform segment to obtain the target domain waveform sample.

[0157] For example, the intelligent recognition terminal can perform mean-reduction processing on each three-channel waveform segment, calculate the arithmetic mean of each channel waveform segment, and subtract it point by point from the original digital counting sequence to eliminate the DC offset component in the waveform. The intelligent recognition terminal can calculate the maximum absolute value of all sampling points of the three-channel waveform segment, divide the values ​​of all sampling points of the three channels by the maximum absolute value, and normalize the waveform amplitude to a preset numerical range to obtain the target domain waveform sample.

[0158] In this embodiment, the intelligent recognition terminal performs instrument response deconvolution correction, zero-phase bandpass filtering, synchronous sliding window truncation, and normalization preprocessing with mean removal and amplitude normalization in sequence. This effectively eliminates the influence of inherent instrument distortion, background noise interference, DC offset, and amplitude differences on waveform data, generating target domain waveform samples that are time-axis aligned, formatted uniformly, have accurate physical meaning, and consistent amplitude distribution. This lays a high-quality data foundation for subsequent multi-scale feature extraction and high-precision earthquake event recognition in the transfer learning recognition model.

[0159] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0160] Based on the same inventive concept, this application also provides a system for implementing the above-mentioned automatic seismic waveform signal identification method based on transfer learning. The solution provided by this system is similar to the implementation scheme described in the above method. Therefore, the specific limitations of one or more embodiments of the automatic seismic waveform signal identification system based on transfer learning provided below can be found in the above-described limitations of the automatic seismic waveform signal identification method based on transfer learning, and will not be repeated here.

[0161] In one exemplary embodiment, such as Figure 3 As shown, an automatic seismic waveform signal identification system 70 based on transfer learning is provided, comprising:

[0162] Data acquisition module 71 can be used to acquire continuous seismic waveform data of the target area to be identified;

[0163] Preprocessing module 72 can be used to preprocess continuous seismic waveform data of the target area to obtain waveform samples of the target domain;

[0164] The feature extraction module 73 can be used to input the target domain waveform sample into the transfer learning recognition model built based on physical consistency constraints, and perform multi-scale spatiotemporal feature extraction and global aggregation on the target domain waveform sample through the feature encoder in the recognition model to obtain the global feature vector to be recognized.

[0165] The classification calculation module 74 can be used to perform classification calculations on the global feature vector to be identified, and obtain the initial posterior probability of each event type.

[0166] The physical consistency constraint module 75 can be used to calculate the physical consistency constraint of the target domain waveform sample and the global feature vector to be identified, and obtain the physical consistency verification value in the inference stage.

[0167] The posterior calibration module 76 can be used to calculate the calibrated posterior probability based on the initial posterior probability and the physical consistency check value, and to perform confidence threshold discrimination on the calibrated posterior probability to obtain the seismic waveform signal recognition result.

[0168] In one embodiment, the feature extraction module includes:

[0169] The data preprocessing unit can be used to acquire source domain seismic waveform datasets and target domain unlabeled seismic waveform datasets, and preprocess the source domain seismic waveform datasets and target domain unlabeled seismic waveform datasets to obtain labeled source domain waveform samples and unlabeled target domain waveform samples.

[0170] The shared feature extraction unit can be used to input source domain waveform samples and target domain waveform samples into a pre-constructed initial deep neural network to obtain a first global feature vector and a second global feature vector; wherein, the first global feature vector is used to characterize the global feature vector of the source domain waveform sample, and the second global feature vector is used to characterize the global feature vector of the target domain waveform sample.

[0171] The source domain classification loss calculation unit can be used to perform classification prediction on the first global feature vector to obtain the source domain classification loss;

[0172] The domain adversarial loss calculation unit can be used to perform domain adversarial calculation on the first global feature vector and the second global feature vector to obtain the domain adversarial loss;

[0173] The physical consistency loss calculation unit can be used to perform self-supervised physical consistency constraints on waveform samples in the target domain based on the second global feature vector, and obtain the physical consistency loss.

[0174] The joint loss weighting unit can be used to perform a weighted summation of the source domain classification loss, domain adversarial loss, and physical consistency loss to obtain the total loss;

[0175] The convergence judgment unit can be used to backpropagate and update the parameters of the initial deep neural network based on the total loss until the preset convergence condition is met, so as to obtain the trained transfer learning recognition model based on physical consistency constraints.

[0176] In one embodiment, the physical consistency loss calculation unit includes:

[0177] The frequency domain transformation subunit can be used to perform zero-padding and fast Fourier transform on the three-channel waveform data contained in the current batch of unlabeled target domain waveform samples within a preset time window to obtain the complex spectrum of the three channels.

[0178] The observation spectrum generation subunit can be used to extract the amplitude spectrum of the three-channel complex spectrum and calculate the average value to obtain the observation amplitude spectrum;

[0179] The physical parameter inversion subunit can be used to perform nonlinear mapping transformation on the second global feature vector to obtain a set of prediction parameters;

[0180] The physical forward modeling sub-unit can be used to perform forward modeling calculations on the predicted parameter set based on the crustal medium physical forward modeling model to obtain the theoretical amplitude spectrum;

[0181] The loss calculation subunit can be used to map the observed amplitude spectrum and the theoretical amplitude spectrum to the double logarithmic domain to obtain the double logarithmic observed amplitude spectrum and the double logarithmic theoretical amplitude spectrum, calculate the mean square error between the double logarithmic observed amplitude spectrum and the double logarithmic theoretical amplitude spectrum, and use the mean square error as the physical consistency loss.

[0182] In one embodiment, the physical forward modeling subunit includes:

[0183] The expression for the theoretical amplitude spectrum is:

[0184]

[0185] In the formula, Represents the theoretical amplitude spectrum. This represents the frequency value corresponding to a single sampling point on the discrete frequency axis after the Fast Fourier Transform. Represents the set of prediction parameters. The parameter representing the asymptotic level of the low-frequency displacement spectrum. This represents the parameter that controls the corner frequency of the seismic source. Represented by natural constant An exponential function with base 0. This represents the parameter that controls the decay along the entire path. This indicates a high-frequency attenuation trend correction index. It represents pi (π).

[0186] In one embodiment, the post-calibration module includes:

[0187] The expression for the posterior probability is:

[0188]

[0189] In the formula, Represents the posterior probability. Represents the event category variable, Indicates the specific first Event categories, This represents the global feature vector to be identified. This represents a preset coefficient used to balance the weights of the initial posterior probability and the physical consistency penalty term. This represents the global feature vector to be identified, output by the flexible maximum function. Belongs to the The initial posterior probability of the event category. Represented by natural constant An exponential function with base 0. This indicates that the rate at which the physical consistency check value decays with the posterior probability is greater than [a certain value]. The attenuation coefficient, This represents the physical consistency check value.

[0190] In one embodiment, the preprocessing module includes:

[0191] The instrument response correction unit can be used to extract the original digital counts of each channel in the continuous seismic waveform data of the target area, perform deconvolution operation on the original digital counts, and obtain the waveform data after instrument response correction.

[0192] The filtering unit can be used to filter the waveform data after the instrument response correction to obtain filtered waveform data.

[0193] The sliding window truncation unit can be used to truncate the filtered waveform data according to the preset time window length and sliding step size, and generate three-channel waveform segments of equal length.

[0194] The normalization processing unit can be used to perform mean removal and normalization based on the maximum amplitude on each three-channel waveform segment to obtain waveform samples in the target domain.

[0195] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the aforementioned automatic identification method for seismic waveform signals based on transfer learning.

[0196] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0197] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0198] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for automatic identification of seismic waveform signals based on transfer learning, characterized in that, The method includes: S1. Obtain continuous seismic waveform data of the target area to be identified; S2. Preprocess the continuous seismic waveform data of the target area to obtain waveform samples of the target domain; S3. Input the target domain waveform sample into the transfer learning recognition model constructed based on physical consistency constraints, and use the feature encoder in the transfer learning recognition model to perform multi-scale spatiotemporal feature extraction and global aggregation on the target domain waveform sample to obtain the global feature vector to be recognized. S4. Perform classification probability mapping on the global feature vector to be identified to obtain the initial posterior probability of each event type; S5. Perform physical consistency constraint calculation on the target domain waveform sample and the global feature vector to be identified to obtain the physical consistency verification value in the inference stage; S6. Based on the initial posterior probability and the physical consistency check value, calculate the calibrated posterior probability, and perform confidence threshold discrimination on the calibrated posterior probability to obtain the seismic waveform signal recognition result.

2. The method of claim 1, wherein, The transfer learning recognition model based on physical consistency constraints is trained in the following ways: S31. Obtain the source domain seismic waveform dataset and the target domain unlabeled seismic waveform dataset, and preprocess the source domain seismic waveform dataset and the target domain unlabeled seismic waveform dataset to obtain labeled source domain waveform samples and unlabeled target domain waveform samples. S32. Input the source domain waveform sample and the target domain waveform sample into a pre-constructed initial deep neural network to obtain a first global feature vector and a second global feature vector; wherein, the first global feature vector is used to characterize the global feature vector of the source domain waveform sample, and the second global feature vector is used to characterize the global feature vector of the target domain waveform sample. S33. Perform classification prediction on the first global feature vector to obtain the source domain classification loss; S34. Perform domain adversarial calculation on the first global feature vector and the second global feature vector to obtain the domain adversarial loss; S35. Based on the second global feature vector, a self-supervised physical consistency constraint is applied to the target domain waveform samples to obtain the physical consistency loss; S36. The source domain classification loss, the domain adversarial loss, and the physical consistency loss are weighted and summed to obtain the total loss; S37. Based on the total loss, backpropagation is performed to update the parameters of the initial deep neural network until the preset convergence condition is met, thereby obtaining the trained transfer learning recognition model constructed based on physical consistency constraints.

3. The method of claim 2, wherein, The S35 includes: S351. The three-channel waveform data contained in the current batch of unlabeled target domain waveform samples are zero-padding and fast Fourier transformed within a preset time window to obtain the complex spectrum of the three channels. S352. Extract the amplitude spectrum of the three-channel complex spectrum and calculate the average value to obtain the observed amplitude spectrum; S353. Perform a nonlinear mapping transformation on the second global feature vector to obtain the prediction parameter set; S354. Perform forward modeling calculations on the predicted parameter set based on the crustal medium physical forward modeling model to obtain the theoretical amplitude spectrum; S355. Map the observed amplitude spectrum and the theoretical amplitude spectrum to the double logarithmic domain to obtain the double logarithmic observed amplitude spectrum and the double logarithmic theoretical amplitude spectrum. Calculate the mean square error between the double logarithmic observed amplitude spectrum and the double logarithmic theoretical amplitude spectrum, and use the mean square error as the physical consistency loss.

4. The method of claim 3, wherein, The expression for the theoretical amplitude spectrum is: In the formula, This represents the theoretical amplitude spectrum. This represents the frequency value corresponding to a single sampling point on the discrete frequency axis after the Fast Fourier Transform. Represents the set of prediction parameters. The parameter representing the asymptotic level of the low-frequency displacement spectrum. This represents the parameter that controls the corner frequency of the seismic source. Represented by natural constant An exponential function with base 0. This represents the parameter that controls the decay along the entire path. This indicates a high-frequency attenuation trend correction index. It represents pi (π).

5. The method of claim 1, wherein, The expression for the posterior probability is: In the formula, This represents the posterior probability. Represents the event category variable, Indicates the specific first Event categories, This represents the global feature vector to be identified. This represents a preset coefficient used to balance the weights of the initial posterior probability and the physical consistency penalty term. This represents the global feature vector to be identified, output by the flexible maximum function. Belongs to the The initial posterior probability of the event category. Represented by natural constant An exponential function with base 0. This indicates that the rate at which the physical consistency check value decays with respect to the posterior probability is greater than [a certain value]. The attenuation coefficient, This represents the physical consistency check value.

6. The method according to claim 1, characterized in that, The S2 includes: S21. Extract the original digital counts of each channel in the continuous seismic waveform data of the target area, perform deconvolution operation on the original digital counts, and obtain the waveform data after instrument response correction. S22. Filter the waveform data after the instrument response correction to obtain filtered waveform data; S23. Perform sliding window truncation on the filtered waveform data according to the preset time window length and sliding step size to generate three-channel waveform segments of equal length. S24. Perform mean removal processing and normalization processing based on the maximum amplitude on each of the three-channel waveform segments to obtain the target domain waveform sample.

7. A seismic waveform signal automatic recognition system based on transfer learning, characterized in that, The system includes: The data acquisition module is used to acquire continuous seismic waveform data of the target area to be identified; The preprocessing module is used to preprocess the continuous seismic waveform data of the target area to obtain waveform samples of the target domain. The feature extraction module is used to input the target domain waveform sample into the transfer learning recognition model constructed based on physical consistency constraints, and to perform multi-scale spatiotemporal feature extraction and global aggregation on the target domain waveform sample through the feature encoder in the transfer learning recognition model to obtain the global feature vector to be recognized. The classification calculation module is used to classify and calculate the global feature vector to be identified, and obtain the initial posterior probability of each event type; The physical consistency constraint module is used to calculate the physical consistency constraint on the target domain waveform sample and the global feature vector to be identified, so as to obtain the physical consistency verification value in the inference stage. The posterior calibration module is used to calculate the calibrated posterior probability based on the initial posterior probability and the physical consistency check value, and to perform confidence threshold discrimination on the calibrated posterior probability to obtain the seismic waveform signal recognition result.

8. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 6.

9. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 6.

Images