A seismic data augmentation method and system based on a generative adversarial network and semi-supervised learning

Abstract

The application belongs to the field of seismic data processing, and discloses a seismic data augmentation method and system based on an adversarial network and semi-supervised learning, which comprises the following steps: collecting seismic signals through a seismic detector to obtain time series data, and obtaining real seismic data after data preprocessing; generating unlabeled seismic data according to the real seismic data and a constructed data generation module; labeling the unlabeled seismic data by using a constructed data labeling module to obtain seismic data with pseudo-labels; forming a new seismic data set by using seismic data with labels and seismic data with pseudo-labels marked by artificial marking, and realizing the augmentation of the seismic data. The application greatly improves the efficiency of constructing a seismic data set under the premise of ensuring data quality, and provides strong data support and technical support for seismic signal processing.

Description

Technical Field

[0001] This invention belongs to the field of earthquake data processing, specifically relating to an earthquake data augmentation method and system based on adversarial networks and semi-supervised learning. Background Technology

[0002] Deep learning technology has been widely used in seismic signal processing, mainly by learning from a large amount of labeled seismic data to improve the accuracy of seismic signal analysis.

[0003] Existing seismic data is primarily acquired through seismic detectors, with manual labeling being a crucial element. This method is not only inefficient but also susceptible to human error, making it difficult to ensure consistent labeling. Analyzing seismic signals using deep learning requires significant manpower and time to create seismic datasets. Furthermore, the high cost of manual labeling makes frequent updates and expansions of the seismic data training set virtually impossible. When new seismic events occur, traditional manual labeling often cannot be completed quickly enough, hindering the rapid analysis of new seismic data by deep learning technologies. Therefore, significantly improving dataset construction efficiency while maintaining labeling quality has become a key challenge for deep learning technologies in seismic signal processing.

[0004] In recent years, generative adversarial networks (GANs) and semi-supervised learning techniques have demonstrated great potential in data generation and automatic annotation. GANs can effectively generate data with realistic features, while semi-supervised learning techniques can achieve automatic annotation of large-scale data using a small number of labeled samples, providing new approaches for seismic data augmentation. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a seismic data augmentation method and system based on adversarial networks and semi-supervised learning. While ensuring data quality, it significantly improves the efficiency of seismic dataset construction, providing strong data support and technical assurance for seismic signal processing.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A seismic data augmentation method based on adversarial networks and semi-supervised learning, the method comprising:

[0008] Seismic signals are collected by seismic detectors to obtain time-series data, and then real seismic data is obtained after data preprocessing.

[0009] Unlabeled earthquake data is generated based on real earthquake data and a constructed data generation module;

[0010] Using the constructed data annotation module, unlabeled seismic data is annotated to obtain seismic data with pseudo-labels;

[0011] A new seismic dataset will be created by combining manually labeled seismic data with pseudo-labeled seismic data, thus augmenting the seismic data.

[0012] Preferably, methods for generating unlabeled seismic data based on real seismic data and a constructed data generation module include:

[0013] During the training phase, a generator is used to generate seismic data x1; a discriminator is used to compare the similarity between real seismic data σ and seismic data x1, and the parameters of the generator are adjusted based on the comparison results.

[0014] Generate unlabeled seismic data x using the trained data generation module. n .

[0015] Preferably, the method for generating seismic data x1 using a generator includes:

[0016] Random noise z is input into a multilayer perceptron, and initial time series data x is obtained through mapping. The initial time series data is then normalized and standardized to obtain normalized time series data x′.

[0017] A lightweight neural network consisting of pooling layers and two perceptron layers is used to perform a global scan of the normalized time series data x′, dividing the sequence into P segments and outputting a weight matrix of a specific size.

[0018] A pre-defined set of N different sizes, F = {F1, F2, ..., F...} N}, where each size corresponds to the length of a partition block, and the size set F is represented as a row vector.

[0019] Multiplying the size weight matrix W by the row vector F yields the dynamic partition size F corresponding to the q-th time series data segment. q ′;

[0020] The dynamic partition size F corresponding to each time series data. q The sequence F' is composed of ''.

[0021] The input sequence x′ is cut by sliding according to sequence F′ to form several time series data x″ of unequal length;

[0022] An attention mechanism is applied to each segment to extract features at different scales.

[0023] The extracted features at different scales are aggregated using a weighted summation method to obtain the aggregated feature Y. fuse ;

[0024] For the aggregated feature Y fuse Seismic data x1, which is the same dimension as the initial time series data x, is recovered using 1D deconvolution. x1 represents the seismic data generated by the generator (G).

[0025] Preferably, the method for comparing the similarity between real seismic data σ and seismic data x1 using a discriminator includes:

[0026] Input the generated seismic data x1 and the actual seismic data σ;

[0027] Use a short time window with a kernel size of 3×3 and a stride of 1 to extract features from local information and obtain local features of the data;

[0028] A medium time window with a kernel size of 9×9 and a stride of 3 is used to extract features from global information to obtain global features of the data.

[0029] Long-term dependencies of data are obtained using a long window with a kernel size of 15×15 and a stride of 5.

[0030] The complete seismic data feature vector is obtained by concatting the local features, global features, and long-term dependencies of the data.

[0031] The complete seismic data feature vector is input into the fully connected layer for feature mapping, and the similarity between the generated seismic data x1 and the real seismic data σ is compared using the sigmoid function.

[0032] Preferably, the method for annotating unlabeled seismic data using the constructed data annotation module to obtain seismic data with pseudo-labels includes:

[0033] During the training phase, in addition to using seismic data σ, seismic data x was also added to the training set. n Partial data x n ′, combine earthquake data σ and data x n The network parameters are updated by backpropagating the weighted sum of the loss values ​​after network training.

[0034] Using the trained data annotation module to annotate unlabeled seismic data x n The data is labeled to obtain seismic data with pseudo-labels.

[0035] Preferably, during the training phase, the processing methods for seismic data σ include:

[0036] Earthquake data σ is manually labeled to obtain true labels δ; the labels of earthquake data σ are predicted using a label prediction network to obtain predicted label results δ1; the predicted label δ1 and the true label δ are then cross-entropy calculated.

[0037] H(δ,δ1)=-∫δ(σ)logδ1(σ)dσ

[0038] Where δ1 represents the prediction result of the earthquake data σ label; δ represents the label of the earthquake data σ that has been manually marked; δ(σ) represents the probability density function of the label δ; and δ1(σ) represents the probability density function of the label δ1.

[0039] Preferably, during the training phase, the seismic data x n Partial data x n The processing methods for ' include:

[0040] For the same seismic data, n random noises are added to perform n different enhancements, resulting in enhanced seismic data x. i ", i∈(1,2,...,n); the earthquake data x is obtained by predicting the labels using a label prediction network. i "Label prediction result λ" i ,i∈(1,2,...,n);

[0041] Using a confidence threshold M on the seismic data x i "Filter the label prediction results:"

[0042]

[0043] Where, λ i Represents earthquake data x i "Predict the label; comparison() represents the comparison function;" This represents the predicted label after comparison with a threshold M; where the confidence threshold M is adaptively and dynamically adjusted.

[0044]

[0045] Among them, M min M represents the initial confidence threshold; max Indicates the final confidence threshold; c represents the current training data volume; M total κ represents the total amount of training data; κ represents the confidence adjustment flexibility factor. The control parameter represents the confidence level smoothing; ε represents the amplification factor for adjusting the training batch data size.

[0046] The same data after filtering Calculate the mean squared error (MSE):

[0047]

[0048] Based on the mean squared error mse, the total mean squared error MSE is obtained:

[0049]

[0050] Where m represents the total number of data samples; n j This represents the number of samples after the j-th data point is randomly augmented.

[0051] Preferably, the seismic data σ and data x n Methods for updating network parameters by backpropagation after weighted summation of loss weights from network training include:

[0052] The weighted sum of the obtained cross-entropy H(δ,δ1) and mean squared error MSE is used as the loss function of the label prediction network:

[0053] L=ρ·H(δ,δ1)+(1-ρ)·MSE

[0054] Where ρ represents the weighting coefficient and L represents the loss function.

[0055] The present invention also provides a seismic data augmentation system based on adversarial networks and semi-supervised learning, the system being used to implement the aforementioned method, the system comprising: an acquisition module, a generation module, an annotation module, and a synthesis module;

[0056] The acquisition module is used to acquire seismic signals through a seismic detector, obtain time series data, and obtain real seismic data after data preprocessing.

[0057] The generation module is used to generate unlabeled earthquake data based on real earthquake data and the constructed data generation module.

[0058] The annotation module is used to annotate unannotated seismic data using the constructed data annotation module, thereby obtaining seismic data with pseudo-labels;

[0059] The synthesis module is used to combine manually labeled seismic data and pseudo-labeled seismic data to form a new seismic dataset, thereby augmenting the seismic data.

[0060] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0061] This invention proposes a seismic data augmentation method and system based on adversarial networks and semi-supervised learning. By constructing a data generation module, automatic generation of seismic data is achieved; by constructing a data annotation module, the efficiency and accuracy of data annotation are improved, significantly reducing annotation costs and reliance on manual annotation. While ensuring data quality, the efficiency of seismic dataset construction is greatly improved, providing strong data support and technical assurance for seismic signal processing. Attached Figure Description

[0062] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0063] Figure 1 This is a schematic diagram of the steps of a seismic data augmentation method based on adversarial networks and semi-supervised learning in an embodiment of the present invention;

[0064] Figure 2 This is a schematic diagram of an adversarial network in an embodiment of the present invention;

[0065] Figure 3 This is a schematic diagram of the generator (G) network in an embodiment of the present invention;

[0066] Figure 4 This is a schematic diagram of the discriminator (D) network in an embodiment of the present invention;

[0067] Figure 5 This is a schematic diagram of the semi-supervised learning training process in an embodiment of the present invention. Detailed Implementation

[0068] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0069] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0070] Example 1

[0071] like Figure 1As shown, this embodiment provides a seismic data augmentation method based on adversarial networks and semi-supervised learning, including: seismic data acquisition, real seismic data, manual labeling, a data generation module, a data annotation module, and a pseudo-label dataset. Specifically:

[0072] Seismic signals are collected using seismic detectors to obtain time-series data, which is then preprocessed to obtain the actual seismic data σ.

[0073] The data generation module built using this invention generates a large amount of new seismic data. During the training phase, a generator produces seismic data x1; a discriminator compares the similarity between seismic data σ and seismic data x1, and adjusts the generator parameters based on the comparison results. After successful training, the data generation module can be used to generate a large amount of new unlabeled seismic data x. n .

[0074] In order to provide the above earthquake data x n For data annotation, this invention employs a data annotation module. During the training phase, in addition to using seismic data σ, seismic data x is also added to the training set to prevent model overfitting. n A small amount of data x n After training, the data annotation module is used to generate a large amount of new, unlabeled seismic data x for the data generation module. n By labeling, a large amount of seismic data with pseudo-labels was obtained.

[0075] This will augment seismic data by combining manually labeled seismic data with a large amount of pseudo-labeled seismic data.

[0076] In this embodiment, data is acquired: seismic data is acquired using a seismic detector, which may contain various noise and outliers.

[0077] Data preprocessing: Cleaning the acquired seismic data to remove noise and outliers, making the data cleaner.

[0078] The preprocessed data is used as the real seismic data σ.

[0079] In this embodiment, the data generation module employs an adversarial network method, such as... Figure 2 As shown, random noise z is input into the generator, which continuously generates seismic data x1. The generated data is then fed into a discriminator and compared with real seismic data σ. The discriminator outputs the similarity between the two, which is fed back to the generator as the discrimination result. Based on this feedback, the generator continuously adjusts its parameters to improve the similarity between the generated data and the real seismic data, thus achieving the training of the adversarial network.

[0080] In the aforementioned adversarial network method, the generator (G) is as follows: Figure 3 As shown, random noise z is mapped to initial time series data through a multilayer perceptron (MLP). Data standardization is then performed using an instance normalization layer. The processed data is then scanned using a lightweight neural network to adaptively divide the time series into multiple scales. Feature extraction is performed through an attention mechanism, and the extracted features are weighted, aggregated, and fused. The fused features are then deconvolutionally used to restore the data dimensionality. Specifically:

[0081] Random noise z is input into a multilayer perceptron (MLP) to obtain initial time series data: x = MLP(z) through mapping, followed by instance normalization.

[0082]

[0083]

[0084] Where x′ represents the normalized time series data; μ(x) represents the mean of the time series data; t is the length of the time series data; σ(x) represents the standard deviation of the time series data; γ represents the learnable scaling factor; and α represents the learnable bias factor.

[0085] Next, a lightweight neural network consisting of pooling layers and two layers of perceptron (MLP) is used to scan the time series x′ as a whole, dividing the sequence into P segments and outputting a weight matrix of a certain size. This is used to adaptively assign the most suitable partition size to each segment of time series data. Each row in matrix W... This indicates the degree of matching between the preset division dimensions corresponding to segment P.

[0086] W=softmax(E2·ReLU(E1·Pool(x′)))

[0087] Where E1 represents the weights of the first fully connected layer; E2 represents the weights of the second fully connected layer; ReLU() represents the activation function; Pool() represents the pooling function.

[0088] The preset N different sizes are set F = {F1, F2, ..., F...} N}, where each size corresponds to the length of a partition block. The set of sizes F is represented as a row vector. Multiplying the size weight matrix W output by the neural network with the row vector F yields the dynamic partition size F corresponding to the q-th segment of the time series data. q ′:

[0089] F q ′=Wq ·F T (q=1,...,P)

[0090] Among them, F q ′: indicates the size of the q-th data segment; W q : represents the weighted matrix of the q-th row; F T : indicates the transpose of row vector F.

[0091] The sequence F′ is composed of the dynamic partition size corresponding to the sequence data for each time period:

[0092] F′={F1′,F2′,...,F P ′}

[0093] The input sequence x′ is cut by sliding according to the sequence F′ to form several time series data x″ of unequal length.

[0094] x″=PatchDivisin(x′,F′)={x1′,x2′,...,x P ′}

[0095] Where x″ represents the time series data set after being partitioned by different sizes; PatchDivisin() represents the partitioning operation; x P ′: Represents time series data after being divided into different sizes.

[0096] After completing the above dynamic multi-scale partitioning of the time series data, in order to capture each segment x P The internal features are processed by applying an attention mechanism to the structure of each segment.

[0097] A novel quadruple attention mechanism (Q, A, K, V) is defined by introducing an additional fitness matrix A into the traditional attention triple (Q, K, V). Here, the fitness matrix A acts as a proxy for the query matrix, aggregating information from K and V and then passing that information back to Q.

[0098] Y i =patchAttention(x″)

[0099]

[0100] Among them, Y i : Represents the features of the i-th time series data after processing; patchAttention(): Represents the attention mechanism function; Q: Represents the query matrix; A: Represents the fitness matrix; K: Represents the key matrix; V: Represents the value matrix; B1, B2: Represent the bias components; d: Represents the dimensions of the query and key; K T: indicates transpose of matrix K; softmax(): indicates normalization operation.

[0101] The extracted features at different scales are aggregated using a weighted summation method:

[0102]

[0103] Among them, w i The weight of the i-th feature is represented by: exp(): the exponential function; g(): the activation function; Y fuse This represents the aggregated features.

[0104] Finally, regarding Y fuse Seismic data x1, which is the same dimension as the initial time series data x, is recovered using 1D deconvolution. x1 represents the seismic data generated by the generator (G).

[0105] In the aforementioned adversarial network methods, the discriminator (D) is as follows: Figure 4 As shown, the generated seismic data x1 and the real seismic data σ are input into the discriminator. A multi-scale CNN method is used, employing a short time window with a kernel size of 3×3 and a stride of 1 to extract local features and learn the local features of the data; a medium time window with a kernel size of 9×9 and a stride of 3 is used to extract global features and learn the global features of the data; and a long time window with a kernel size of 15×15 and a stride of 5 is used to learn the long-term dependencies of the data. The features extracted at different scales by the multi-scale CNN are concatenated using a concat operation to obtain the complete seismic signal data feature vector. The feature vector is input into a fully connected layer for feature mapping, and the similarity between the generated seismic data x1 and the real seismic data σ is compared using a sigmoid function. The result is then fed back to the generator (G). Specifically:

[0106] Input the generated seismic data x1 and the actual seismic data σ.

[0107] We use a short time window with a kernel size of 3×3 and a stride of 1 to extract features from local information and learn the local features of the data.

[0108] h local =Conv1D(η,k=3,s=1),η∈{x1,σ}

[0109] Among them, h local : Represents local information extracted within a short time window; η: Represents seismic data; k: Represents the kernel size; s: Represents the convolution stride.

[0110] We use a medium-sized time window with a kernel size of 9×9 and a stride of 3 to extract features from the global information and learn the global features of the data.

[0111] h global =Conv1D(η,k=9,s=3),η∈{x1,σ}

[0112] Among them, h global : Represents global information extracted within a medium time window.

[0113] Long-term dependencies of the data are learned using a long window with a kernel size of 15×15 and a stride of 5:

[0114] h long-term =Conv1D(η,k=15,s=5),η∈{x1,σ}

[0115] Among them, h long-term : Indicates information about long-term dependency.

[0116] The features extracted from different scales by multi-scale CNN are concatenated using the concat operation to obtain the complete seismic data feature vector:

[0117] h concat =Concat(h local ,h global ,h long-term )

[0118] Among them, h concat : Represents the feature information after the concatenation operation; Concat(): Represents the concatenation operation.

[0119] eigenvector h concat Input to the fully connected layer for feature mapping:

[0120] β=R(ωh concat +b)

[0121] Where β represents the feature map output; R() represents the activation function; ω represents the weight matrix; and b represents the bias vector.

[0122] The similarity between the generated seismic data x1 and the actual seismic data σ is compared using the sigmoid function:

[0123]

[0124] By having the generator (G) and discriminator (D) compete against each other, based on the adversarial loss:

[0125]

[0126] Where, min G max DV(D,G): This represents the discriminator's goal to maximize this value (this value represents a probability value, indicating the degree of difference between generated and real seismic data; the discriminator's goal is to distinguish between generated and real seismic data, and a larger probability value indicates a greater difference, making it easier to distinguish between the two), trying to differentiate between real and generated seismic data (based on the above probability value, the discriminator's goal is to maximize this value to distinguish based on the degree of difference); the generator's goal is to minimize this value (this value is also the aforementioned probability value, indicating the degree of difference between generated and real seismic data; the generator's goal is to minimize the difference between generated and real seismic data), making the generated seismic data closer to the real seismic signal (the generator's goal is to minimize this probability value, which will reduce the difference between generated and real seismic data, making it impossible for the discriminator to distinguish between real and generated seismic data), thus making the discriminator unable to distinguish between them.

[0127] The adversarial network is trained to generate unlabeled seismic data (unlabeled seismic data refers to seismic data generated by the generator through adversarial network training; the generated seismic data is not labeled). The generator's parameters are then adjusted based on the obtained adversarial loss. Specifically, according to the adversarial loss function, the generator aims to reduce the difference between the generated and real seismic data, making it difficult for the discriminator to distinguish between real and generated data. The discriminator's goal is to differentiate between the generated and real seismic data. Based on the loss value, the generator adjusts its parameters to continuously generate seismic data that more closely resembles reality, thus fooling the discriminator; the discriminator also adjusts its parameters to better distinguish between real and generated seismic data. Through this continuous adjustment, the data generated by the generator becomes increasingly similar to real seismic data, and can then be considered and used as correct data.

[0128] A large amount of new seismic signal data was generated using a trained data generation module. n .

[0129] In this embodiment, the data annotation module aims to annotate the large amount of seismic data generated by the data generation module using a semi-supervised learning method. The data annotation module uses two batches of input data: one batch uses real seismic data to ensure the accuracy of the model data annotation; the other batch uses a small amount of generated seismic data to prevent model overfitting. Both batches of data use the same label prediction network. Finally, the two batches of data are weighted and summed with the loss weights after network training, and backpropagated to update the network parameters.

[0130] like Figure 5As shown, firstly, the first batch of data—real earthquake data σ—is manually labeled to obtain true labels. After random augmentation, the earthquake data σ is then used to predict labels using a label prediction network to obtain predicted labels. The predicted labels δ1 and the true labels δ are cross-entropy calculated, and this cross-entropy is used as the loss weight for the first batch of data. Next, to prevent overfitting, a second batch of data—generated earthquake data x—is added. n A small portion of the data x n For the same seismic data, n random noises are added to perform n different enhancements to obtain the enhanced seismic data x. i ", i∈(1,2,...,n); The label prediction result λ is obtained by using a label prediction network to predict the label. i ,i∈(1,2,...,n). To improve the training speed of the label prediction network, an adaptive confidence threshold is used for the seismic data x. i "The label prediction results are filtered, and the mean squared error of the filtered predicted labels is calculated. This total mean squared error is used as the loss weight for the second batch of data. Finally, the obtained cross-entropy and total mean squared error are weighted and summed to obtain the loss value of the label prediction network. Backpropagation is then used to update and optimize the label prediction network parameters. Specifically:"

[0131] The module processes the first batch of data: real earthquake data σ.

[0132] Earthquake data σ is manually labeled to obtain true labels δ. A label prediction network is then used to predict labels (labels are the labels predicted from the actual earthquake data) on the earthquake data σ, resulting in predicted label result δ1. The predicted label δ1 is then cross-entropy calculated with the true label δ.

[0133] H(δ,δ1)=-∫δ(σ)logδ1(σ)dσ

[0134] Where δ1: represents the prediction result of the earthquake data σ label; δ: represents the label of the earthquake data σ that is manually marked; δ(σ): represents the probability density function of the label δ; δ1(σ): represents the probability density function of the label δ1.

[0135] To prevent overfitting, the module includes a second batch of data: generated seismic data x. n A small amount of data x n ′, to be processed.

[0136] In the second batch of data, for the same seismic data, n random noises were added to perform n different enhancements, resulting in enhanced seismic data x. i", i∈(1,2,...,n); The label prediction result λ is obtained by using a label prediction network to predict the labels (labels refer to the labels predicted by the label prediction network on the generated seismic data) of the generated seismic data. i ,i∈(1,2,...,n).

[0137] To improve the training speed of the label prediction network, a confidence threshold M is used for the earthquake data x. i "Filter the label prediction results:"

[0138]

[0139] Where, λ i Represents earthquake data x i "Predict the label; comparison(): represents the comparison function;" This represents the predicted label after comparison with a threshold M; where the confidence threshold M is adaptively and dynamically adjusted.

[0140]

[0141] Among them, M min : Represents the initial confidence threshold; M max : represents the final confidence threshold; c: represents the current amount of training data; M total : Represents the total amount of training data; κ: Represents the flexibility factor for confidence adjustment; ε: represents the control parameter for confidence level smoothing; ε: represents the amplification factor for adjusting the amount of training batch data.

[0142] A label prediction result greater than the label confidence threshold is considered a correct label prediction. Label predictions with a confidence level below the threshold are set to 0. For the same number, if all label predictions from the label prediction network are greater than the confidence threshold, the network is considered strong and only requires fine-tuning. Conversely, if some predictions are greater than the threshold and others are less, the network is considered weak and requires more significant adjustments. This confidence threshold filtering method amplifies the error of low-confidence predictions during error calculation, thereby improving the network training speed and effectiveness.

[0143] The same data after processing Calculate the mean squared error (MSE):

[0144]

[0145] The total mean square error (MSE) after processing the second batch of data is:

[0146]

[0147] Where m represents the total number of data samples; n j : Represents the number of samples after random augmentation of the j-th data point.

[0148] After training the network with the above two batches of data, the weighted sum of the obtained cross-entropy H(δ,δ1) and mean squared error MSE is used as the loss value for semi-supervised learning training:

[0149] L=ρ·H(δ,δ1)+(1-ρ)·MSE

[0150] Where ρ represents the weighting coefficient, and L represents the loss function.

[0151] The loss value measures the gap between the training predicted labels and the actual results. Backpropagation uses the loss value to update and optimize the label prediction network parameters, minimize the prediction error, and improve the performance of the data labeling module.

[0152] In this embodiment, a trained data labeling module is used to label the seismic data generated by the data generation module, resulting in a large amount of seismic data with pseudo-labels.

[0153] In this embodiment, a new seismic dataset is formed by combining manually labeled real seismic data with a large amount of generated pseudo-labeled seismic data, thereby augmenting the seismic data.

[0154] Example 2

[0155] The present invention also provides a seismic data augmentation system based on adversarial networks and semi-supervised learning. The system is used to implement the method of Embodiment 1. The system includes: an acquisition module, a generation module, an annotation module, and a synthesis module.

[0156] The acquisition module is used to acquire seismic signals through a seismic detector, obtain time-series data, and obtain real seismic data after data preprocessing.

[0157] The generation module is used to generate unlabeled earthquake data based on real earthquake data and the constructed data generation module.

[0158] The annotation module is used to annotate unannotated seismic data using the constructed data annotation module, resulting in seismic data with pseudo-labels;

[0159] The synthesis module is used to augment seismic data by combining manually labeled seismic data with pseudo-labeled seismic data into a new seismic dataset.

[0160] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A seismic data augmentation method based on adversarial networks and semi-supervised learning, characterized in that, The method includes: Seismic signals are collected by seismic detectors to obtain time-series data, and then real seismic data is obtained after data preprocessing. Unlabeled earthquake data is generated based on real earthquake data and a constructed data generation module; Using the constructed data annotation module, unlabeled seismic data is annotated to obtain seismic data with pseudo-labels; A new earthquake dataset will be created by combining manually labeled earthquake data with pseudo-labeled earthquake data, thereby augmenting the earthquake data. Methods for generating unlabeled seismic data based on real seismic data and constructed data generation modules include: During the training phase, seismic data is generated using a generator. ; Using a discriminator to compare with real earthquake data and earthquake data The similarity is compared, and the generator parameters are adjusted based on the comparison results; Generate unlabeled seismic data using the trained data generation module. ; Using generators to generate seismic data The methods include: random noise Inputting data into a multilayer perceptron, the initial time series data is obtained through mapping. The initial time series data is normalized using instance normalization. The normalized time series data... ; A lightweight neural network consisting of pooling layers and two perceptron layers is used to process the normalized time series data. Perform a full scan to divide the sequence into The segment outputs a weight matrix of a certain size. ; Preset A collection of different sizes Each size corresponds to the length of a partition block, and the size set is... Represented as row vectors ; Size weight matrix with row vectors Multiply to get the first... Dynamic partition size corresponding to a time series data segment ; The dynamic partition size corresponding to each time series data. Composition sequence ; According to sequence Sliding pair input sequence The data is then segmented to form several time series data of unequal lengths. ; An attention mechanism is applied to each segment to extract features at different scales. The extracted features at different scales are aggregated using a weighted summation method to obtain the aggregated features. ; Features after aggregation Use 1D deconvolution to recover the data from the initial time series. Seismic data of the same dimension , Represents generator The generated seismic data; During the training phase, earthquake data... Partial data The processing methods include: For the same earthquake data, add Random noise Different enhancements were performed to obtain enhanced seismic data. Earthquake data is obtained by predicting labels using a label prediction network. Label prediction results ; Use confidence threshold earthquake data Filter the label prediction results: in, Representing earthquake data Predict labels; This represents a comparison function; Indicates threshold The predicted labels after comparison; where the confidence threshold is... Perform adaptive dynamic adjustments: in, Indicates the initial confidence threshold; Indicates the final confidence threshold; Indicates the current amount of training data; This represents the total amount of training data; A flexibility factor representing confidence level adjustment; Control parameters indicating confidence level smoothing; This represents the amplification factor used to adjust the amount of training batch data.

2. The method according to claim 1, characterized in that, Using a discriminator to compare with real earthquake data and earthquake data Methods for determining similarity include: Input the generated seismic data and real earthquake data ; Use convolution kernel size of A short time window with a step size of 1 is used to extract features from local information to obtain local features of the data; Use convolution kernel size of A medium time window with a step size of 3 is used to extract features from global information to obtain global features of the data. Use convolution kernel size of A long-term window with a step size of 5 is used to obtain the long-term dependencies of the data. Use local features, global features, and long-term dependencies of data. concat The stitching operation yields the complete feature vector of the seismic data; The complete seismic data feature vector is input into the fully connected layer for feature mapping, and then... Seismic data generated by function comparison and real earthquake data The degree of similarity.

3. The method according to claim 1, characterized in that, Methods for labeling unlabeled seismic data using a constructed data labeling module to obtain seismic data with pseudo-labels include: During the training phase, in addition to using seismic data In addition, earthquake data was added to the training set. Partial data Earthquake data and data The network parameters are updated by backpropagating the weighted sum of the loss values ​​after network training. Using the trained data annotation module to label unlabeled seismic data The data is labeled to obtain seismic data with pseudo-labels.

4. The method according to claim 3, characterized in that, During the training phase, earthquake data... The processing methods include: earthquake data Obtaining accurate labels through manual marking Earthquake data A label prediction network is used to predict the labels, and the predicted labels are obtained. Predict labels With real labels Calculate cross-entropy: in, Representing earthquake data The prediction results for the labels; Representing earthquake data Manually marked labels; Indicates label probability density function; Indicates label Probability density function.

5. The method according to claim 4, characterized in that, The same data after filtering Mean square error : According to mean square error The total mean square error is obtained. : in, Indicates the total number of data samples; Indicates the first The number of samples after random augmentation of the data.

6. The method according to claim 5, characterized in that, earthquake data and data Methods for updating network parameters by backpropagation after weighted summation of loss weights following network training include: The obtained cross-entropy and mean square error Weighted summation is used as the loss function for the label prediction network: in, Let L represent the weighting coefficients and L represent the loss function.

7. A seismic data augmentation system based on adversarial networks and semi-supervised learning, said system being used to implement the method according to any one of claims 1-6, characterized in that, The system includes: a data acquisition module, a generation module, a labeling module, and a synthesis module; The acquisition module is used to acquire seismic signals through a seismic detector, obtain time series data, and obtain real seismic data after data preprocessing. The generation module is used to generate unlabeled earthquake data based on real earthquake data and the constructed data generation module. The annotation module is used to annotate unannotated seismic data using the constructed data annotation module, thereby obtaining seismic data with pseudo-labels; The synthesis module is used to combine manually labeled seismic data and pseudo-labeled seismic data to form a new seismic dataset, thereby augmenting the seismic data.

Images