Semi-supervised microseismic first arrival intelligent picking method combined with SimMatch and improved TransUGA
By combining SimMatch with an improved TransUGA semi-supervised learning method, and utilizing generative adversarial networks and the Transformer architecture, high-confidence pseudo-labels are generated, solving the problems of low efficiency and poor accuracy of manual annotation in microseismic signal picking, and achieving efficient and accurate first arrival signal picking.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA THREE GORGES UNIV
- Filing Date
- 2023-06-25
- Publication Date
- 2026-05-29
AI Technical Summary
Existing technologies for microseismic signal acquisition suffer from problems such as low efficiency due to manual intervention, high subjectivity, laborious and time-consuming data annotation, and poor generalization. In particular, the accuracy of first arrival signal acquisition is insufficient under conditions of low signal-to-noise ratio and small sample size.
A semi-supervised learning method combining SimMatch and an improved TransUGA is adopted, which integrates generative adversarial networks and the Transformer architecture. By generating pseudo-labels and high-confidence pseudo-labels, semi-supervised training is performed using a small amount of labeled data and a large amount of unlabeled data to extract microseismic signal features.
It improves the accuracy and efficiency of first arrival signal acquisition for microseismic events, reduces the cost of manually labeled data, adapts to different data distributions and acquisition scenarios, and has universal applicability.
Smart Images
Figure CN116990860B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of first-arrival acquisition technology, and more specifically to a semi-supervised microseismic first-arrival intelligent acquisition method that combines SimMatch and an improved TransUGA. Background Technology
[0002] In the field of petroleum exploration, microseismic monitoring technology is widely used to determine the formation and development trends of hydraulic fracturing fractures. Microseismic monitoring mainly consists of three stages: data acquisition and preprocessing, first-arrival information acquisition, and seismic source inversion and location. However, due to the special nature of microseismic signal acquisition methods, the acquired samples often contain a large amount of noise, and the effective signal energy of microseismic signals is relatively weak, resulting in the original effective signal being submerged in noise, making first-arrival information acquisition extremely difficult.
[0003] Currently, traditional data acquisition methods still require significant manual intervention. However, manual acquisition is inefficient and suffers from issues such as subjectivity and the risk of missed detections. Especially in large-scale practical applications, where the amount of data to be processed is enormous, manual acquisition cannot meet the demands for rapid and efficient processing, nor can it guarantee the consistency and accuracy of the results. Traditional algorithms, on the other hand, rely on sufficient manually labeled samples to train the model to achieve high accuracy. However, data labeling is laborious and time-consuming, and is subject to subjectivity and errors, easily leading to poor model generalization. Therefore, how to efficiently, accurately, and stably acquire effective signals in the context of small samples is one of the urgent problems to be solved in microseismic monitoring technology.
[0004] This invention proposes a semi-supervised first-arrival extraction method based on TransUGA and SimMatch, which can automatically and accurately extract first-arrival information from noisy microseismic signals. This method utilizes a semi-supervised learning framework, combining generative adversarial networks and the Transformer architecture, to automatically learn and extract features from a large amount of unlabeled data even with limited labeled data, thereby improving the accuracy of first-arrival extraction. Simultaneously, it obtains high-confidence pseudo-labels, solving the problem of low efficiency associated with manual data annotation. Summary of the Invention
[0005] The technical problem to be solved by this invention is to provide a semi-supervised intelligent first arrival picking method for microseismic data by combining SimMatch and improved TransUGA. This method can effectively solve the problems of low efficiency of manual annotation data, poor accuracy of first arrival signal picking under low signal-to-noise ratio, and difficulty in learning signal features from small sample data. It improves the picking accuracy of effective microseismic signals and greatly reduces the cost of manual annotation data.
[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:
[0007] The semi-supervised microseismic first arrival intelligent picking method combining SimMatch and improved TransUGA has the following steps: Step 1: Obtain the real banded signal in the microseismic data and generate forward modeling simulation signals at different dominant frequencies and velocities using the finite difference wave equation forward modeling method;
[0008] Step 2: Use real signals and generated analog signals to create an original dataset. Input the created original dataset into the constructed UGATIT generative adversarial network. Through the interaction between the generator and the discriminator, pseudo signal samples are generated, and the generated pseudo real signal samples are filtered out.
[0009] Step 3: Use the original labeled data and the pseudo-real signal samples from Step 2 as a new dataset to perform semi-supervised training on the improved TransUNet network;
[0010] Step 4: Obtain the trained TransUNet network and high-confidence pseudo-labels;
[0011] Step 5: Input the microseismic data to be measured into the trained TransUNet network to pick up the first arrival signal.
[0012] The specific process of the UGATIT network generating pseudo-real signal samples in Step 2 above is as follows:
[0013] The UGATIT network consists of two generators G s→t G t→s and two discriminators D s and D t composition;
[0014] Generator: Downsamples the input signal (source domain) and extracts signal features using residual blocks. The generator domain is obtained by guiding the downsampling blocks with AdaLIN. s→t A generator represents a transformation from a source domain to a target domain;
[0015] AdaLIN is defined as follows:
[0016]
[0017] Where a is the attention feature map tensor, μ I μ L and γ and β represent the mean and variance of the channel and layer, respectively; γ and β are the parameters generated by the fully connected layer, ρ is between [0, 1], Δρ represents the parameter update vector determined by the optimizer, and lr is the learning rate.
[0018] The generator dynamically adjusts the value of ρ so that ρ is close to 0 in tasks important for LN and close to 1 in tasks important for instance normalization; it is initialized to 1 and 0 in the residual block and upsampling block of the decoder, respectively.
[0019] UGATIT's total loss function is defined as follows:
[0020]
[0021] In this invention, λ1 = 1, λ2 = 10, λ3 = 10, and λ4 = 1000;
[0022] The four loss functions in the total loss function are defined as follows:
[0023] Combat loss L gan The data distribution used to match the generation domain and the target domain is defined as follows:
[0024]
[0025] in This indicates the expectation that the sample belongs to the source domain. G represents the expectation that a sample belongs to the target domain. s→t (x) represents the microseismic signal sample obtained from the source domain to the target domain transformation, D t (x) represents the discrimination result of the target discriminator on sample x;
[0026] Cyclic consistency loss L cycle The following is a definition used to mitigate pattern collapse:
[0027]
[0028] Identity loss L identity To ensure that the generator input and the distribution of the inputs are similar, it is defined as follows:
[0029]
[0030] CAM classification loss L cam To distinguish the differences between the source and target domains, an auxiliary classifier η is used. s and The classification loss is defined as follows:
[0031]
[0032] Where η s (x) represents the probability that the sample comes from the source domain.
[0033] The specific process of semi-supervised training of the improved TransUNet network in Step 3 above is as follows:
[0034] The improved TransUNet network structure is as follows:
[0035] The improved TransUNet modifies the original encoder to a combination of ResNet50 and Transformer, and connects the intermediate features output from the ResNet50 network through skip connections, and performs feature aggregation with the upsampling module in the decoder, fusing signal features at different resolutions to accelerate the convergence speed of the network.
[0036] The ResNet50 network uses Group Normalization to process the data, making the network model more suitable for picking up the first arrival of signals;
[0037] Before upsampling, ConvReLU is replaced with the SSPCAB module to maximize the features of the encoder output, suppress the influence of noise, and thus improve the network's noise resistance and the accuracy of initial arrival picking.
[0038] The SSPCAB module consists of a masked convolutional layer and a Squeeze-and-Excitation module; the size of the masked convolutional kernel is k = 2k′ + 2d + 1, where k′ is a hyperparameter defining the sub-kernel size, and d is the dilation rate, used to control the local or global nature of the visible information; the receptive field of the visible region of the convolutional kernel is... The loss function of the SSPCAB module is defined as follows:
[0039]
[0040] Where X is the input of this module, For output, G is the SSPCAB function;
[0041] The improved TransUNet loss function is defined as follows:
[0042]
[0043] (8) In the formula The cross-entropy loss function for multi-class classification is calculated using the following formula:
[0044]
[0045] (9) p ic Let y be the predicted probability that the i-th sample belongs to class c, and M be the number of predicted classes. If the true class of sample i is c, then y ic Take 1 otherwise take 0; λ in equation (8) s It controls SSPCAB. Hyperparameters indicating the degree of importance.
[0046] The specific process of Step 4 above is as follows:
[0047] Use the SimMatch structure to improve the confidence of the pseudo labels of unlabeled samples during training;
[0048] For unlabeled samples, randomly apply a weak augmentation (T w (·)) or a strong augmentation (T s (·)), and extract feature information with an encoder F(·), that is Adopt a fully connected class prediction head φ(·) to map h b to semantic similarity, denoted as p = φ(h);
[0049] For a cluster of different samples z k : (k ∈ (1,..., K)), the instance similarity distribution between z w and the i-th instance z i is calculated as follows: The calculation formula is as follows:
[0050]
[0051] where is the temperature coefficient, used to control the smoothness of the data distribution; z b = g(h b ), and g(·) is a non-linear mapping head that maps the feature h to a low-dimensional embedding vector;
[0052] is to use the semantic similarity p w to calibrate the instance similarity q w , expand p w to the K-dimensional space, and the expansion formula is as follows:
[0053]
[0054] where cls(·) is a function that returns the true arrival time;
[0055] is to use the instance similarity q w to calibrate the semantic similarity p w , here aggregate q w >to the L-dimensional space (L << K), and the aggregation formula is as follows:
[0056]
[0057] where 1(·) is an indicator function, if p i w and q j w belong to the same class, the value is 1, otherwise 0;
[0058] Calibrated semantic similarity Similarity to instances The calculation formula is as follows:
[0059]
[0060] Here, α is a hyperparameter used to control the weight relationship between semantic information and instances; when semantic similarity and instance similarity are close, the generated pseudo-labels will have higher confidence.
[0061] The present invention provides a semi-supervised intelligent microseismic first arrival acquisition method combining SimMatch and improved TransUGA, which has the following beneficial effects:
[0062] 1. By inputting a small number of labeled real signal samples and generated simulated signals into the generator of the generative adversarial network UGATIT, a large number of signals close to real samples are generated through self-learning. This expands the dataset and improves the model's stability against signal noise.
[0063] 2. By improving the TransUNet network and combining it with the Transformer architecture and layer skip connections, the temporal information and deep features in the microseismic signal are effectively extracted, which greatly improves the accuracy of first arrival picking.
[0064] 3. The model was semi-supervised by using a small amount of labeled data and a large amount of generated unlabeled data, and high-confidence labels were obtained, which greatly reduced the workload of manually labeling data and improved the efficiency of initial picking.
[0065] 4. Using SimMatch's instance matching and semantic matching to improve the confidence of pseudo-labels can adapt to different data distributions and initial picking scenarios, and has universality. Attached Figure Description
[0066] The present invention will be further described below with reference to the accompanying drawings and embodiments:
[0067] Figure 1 This is a flowchart of the process of this invention;
[0068] Figure 2 It is an improved TransUNet network structure;
[0069] Figure 3 This is the UGATIT architecture diagram;
[0070] Figure 4 This is a diagram illustrating the process of generating unlabeled samples using the UGATIT network. Detailed Implementation
[0071] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and embodiments.
[0072] like Figure 1-4 As shown, the semi-supervised microseismic first-arrival intelligent picking method combining SimMatch and improved TransUGA has the following steps:
[0073] Step 1: Obtain the real labeled signals from the microseismic data and generate forward modeling signals at different dominant frequencies and velocities using the finite difference wave equation forward modeling method;
[0074] Step 2: Use real signals and generated analog signals to create an original dataset. Input the created original dataset into the constructed UGATIT generative adversarial network. Through the interaction between the generator and the discriminator, pseudo signal samples are generated, and the generated pseudo real signal samples are filtered out.
[0075] Step 3: Use the original labeled data and the pseudo-real signal samples from Step 2 as a new dataset to perform semi-supervised training on the improved TransUNet network;
[0076] Step 4: Obtain the trained TransUNet network and high-confidence pseudo-labels;
[0077] Step 5: Input the microseismic data to be measured into the trained TransUNet network to pick up the first arrival signal.
[0078] The specific process of the UGATIT network generating pseudo-real signal samples in Step 2 above is as follows:
[0079] The UGATIT network consists of two generators G s→t G t→s and two discriminators D s and D t composition;
[0080] Generator: Downsamples the input signal (source domain) and extracts signal features using residual blocks. The generator domain is obtained by guiding the downsampling blocks with AdaLIN. s→t A generator represents a transformation from a source domain to a target domain;
[0081] AdaLIN is defined as follows:
[0082]
[0083] Where a is the attention feature map tensor, μ I μ L and ρ and β represent the mean and variance of the channel and layer, respectively; γ and β are the parameters generated by the fully connected layer, ρ is between [0, 1], Δρ represents the parameter update vector determined by the optimizer, and l r The learning rate;
[0084] The generator dynamically adjusts the value of ρ so that ρ is close to 0 in tasks important for LN and close to 1 in tasks important for instance normalization; it is initialized to 1 and 0 in the residual block and upsampling block of the decoder, respectively.
[0085] UGATIT's total loss function is defined as follows:
[0086]
[0087] In this invention, λ1 = 1, λ2 = 10, λ3 = 10, and λ4 = 1000;
[0088] The four loss functions in the total loss function are defined as follows:
[0089] Combat loss L gan The data distribution used to match the generation domain and the target domain is defined as follows:
[0090]
[0091] in This indicates the expectation that the sample belongs to the source domain. G represents the expectation that a sample belongs to the target domain. s→t (x) represents the microseismic signal sample obtained from the source domain to the target domain transformation, D t (x) represents the discrimination result of the target discriminator on sample x;
[0092] Cyclic consistency loss L cycle The following is a definition used to mitigate pattern collapse:
[0093]
[0094] Identity loss L identity To ensure that the generator input and the distribution of the inputs are similar, it is defined as follows:
[0095]
[0096] CAM classification loss L cam To distinguish the differences between the source and target domains, an auxiliary classifier η is used. s and The classification loss is defined as follows:
[0097]
[0098] Where η s (x) represents the probability that the sample comes from the source domain.
[0099] The specific process of semi-supervised training of the improved TransUNet network in Step 3 above is as follows:
[0100] The improved TransUNet network structure is as follows:
[0101] The improved TransUNet modifies the original encoder to a combination of ResNet50 and Transformer, and connects the intermediate features output from the ResNet50 network through skip connections, and performs feature aggregation with the upsampling module in the decoder, fusing signal features at different resolutions to accelerate the convergence speed of the network.
[0102] The ResNet50 network uses Group Normalization to process the data, making the network model more suitable for picking up the first arrival of signals;
[0103] Before upsampling, ConvReLU is replaced with the SSPCAB module to maximize the features of the encoder output, suppress the influence of noise, and thus improve the network's noise resistance and the accuracy of initial arrival picking.
[0104] The SSPCAB module consists of a masked convolutional layer and a Squeeze-and-Excitation module; the size of the masked convolutional kernel is k = 2k′ + 2d + 1, where k′ is a hyperparameter defining the sub-kernel size, and d is the dilation rate, used to control the local or global nature of the visible information; the receptive field of the visible region of the convolutional kernel is... The loss function of the SSPCAB module is defined as follows:
[0105]
[0106] Where X is the input of this module, For output, G is the SSPCAB function;
[0107] The improved TransUNet loss function is defined as follows:
[0108]
[0109] (8) In the formula The cross-entropy loss function for multi-class classification is calculated using the following formula:
[0110]
[0111] (9) p icLet y be the predicted probability that the i-th sample belongs to class c, and M be the number of predicted classes. If the true class of sample i is c, then y ic Take 1 otherwise take 0; λ in equation (8) s It controls SSPCAB. Hyperparameters indicating the degree of importance.
[0112] The specific process of Step 4 above is as follows:
[0113] The SimMatch structure is used to improve the confidence of pseudo-labels for unlabeled samples during training.
[0114] For unlabeled samples, a weak enhancement (T) is randomly applied. w (·)) or strong enhancement (T) s (·)), using an encoder F(·) to extract feature information, i.e. Using a fully connected class prediction head φ(·) to predict h b The mapping is expressed as semantic similarity, denoted as p = φ(h);
[0115] For a cluster of different samples z k :(k∈(1,...,K)), z w and the i-th instance z i Instance similarity distribution between The calculation formula is as follows:
[0116]
[0117] in, This is a temperature coefficient used to control the smoothness of the data distribution; z b =g(h b g(·) is a non-linear mapping head that maps the feature h into a low-dimensional embedding vector;
[0118] To use semantic similarity p w To calibrate instance similarity q w , will p w Expanding to K-dimensional space, the expansion formula is as follows:
[0119]
[0120] Where cls(·) is a function that returns the true first arrival time;
[0121] To use instance similarity q w To calibrate semantic similarity p w , here q w Aggregating to L-dimensional space (L << K), the aggregation formula is as follows:
[0122]
[0123] Where 1(·) is the indicator function, if p i w and q j w If they belong to the same category, the value is 1; otherwise, it is 0.
[0124] Calibrated semantic similarity Similarity to instances The calculation formula is as follows:
[0125]
[0126] Here, α is a hyperparameter used to control the weight relationship between semantic information and instances; when semantic similarity and instance similarity are close, the generated pseudo-labels will have higher confidence.
[0127] This invention combines the Transformer architecture and the UNet network to make feature extraction more efficient. Utilizing the self-attention mechanism of the Transformer module, local features are effectively extracted while preserving global information, resulting in more discriminative feature representations. The TransUNet network employs an end-to-end training approach, directly learning the feature representations of first-arrival signal points from the raw data, avoiding the process of designing complex feature engineering. Furthermore, it uses a semi-supervised learning method, leveraging the UGATIT generative adversarial network to generate a large number of unlabeled samples, and combining this with the SimMatch method to fully utilize known labeled data and generated unlabeled samples for joint training, generating more reliable pseudo-labels and thus improving the first-arrival picking effect. A large number of high-confidence labels are also obtained.
[0128] like Figure 4 As shown in the figure, (a) real signal, (b) fake_s2s feature, (c) fake_s2t2s feature, (d) generated real signal, (e) analog signal, (f) fake_t2t feature, (g) fake_t2s2t feature, and (h) generated analog signal;
[0129] Figure 4 The process of generating a large number of unlabeled real and simulated signals for the UGATIT generative adversarial network is visualized. By comparing the input signals and the generated signals, it is found that the unlabeled samples generated by UGATIT are almost identical to the input samples.
Claims
1. A semi-supervised intelligent microseismic first-arrival acquisition method combining SimMatch and improved TransUGA, characterized by the following steps: for: Step 1: Obtain the real labeled signals from the microseismic data and generate forward modeling signals at different dominant frequencies and velocities using the finite difference wave equation forward modeling method; Step 2: Use real signals and generated analog signals to create an original dataset. Input the created original dataset into the constructed UGATIT generative adversarial network. Through the interaction between the generator and the discriminator, pseudo signal samples are generated, and the generated pseudo real signal samples are filtered out. Step 3: Use the original labeled data and the pseudo-real signal samples from Step 2 as a new dataset to perform semi-supervised training on the improved TransUNet network; Step 4: Obtain the trained TransUNet network and high-confidence pseudo-labels; Step 5: Input the microseismic data to be measured into the trained TransUNet network to pick up the first arrival signal.
2. The semi-supervised microseismic first arrival intelligent pickup method combining SimMatch and improved TransUGA according to claim 1, characterized in that, The specific process of generating pseudo-real signal samples using the UGATIT network in Step 2 is as follows: The UGATIT network consists of two generators G s→t G t→s and two discriminators D s and D t composition; Generator: Downsamples the input signal (source domain) and extracts signal features using residual blocks. The generator domain is obtained by guiding the downsampling blocks with AdaLIN. s→t A generator represents a transformation from a source domain to a target domain; AdaLIN is defined as follows: Where a is the attention feature map tensor, μ I μ L and ρ and β represent the mean and variance of the channel and layer, respectively; γ and β are the parameters generated by the fully connected layer, ρ is between [0, 1], Δρ represents the parameter update vector determined by the optimizer, and l r The learning rate; The generator dynamically adjusts the value of ρ, which is initialized to 1 and 0 in the residual block and upsampling block of the decoder, respectively. UGATIT's total loss function is defined as follows: In this invention, λ1 = 1, λ2 = 10, λ3 = 10, and λ4 = 1000; The four loss functions in the total loss function are defined as follows: Combat loss L gan The data distribution used to match the generation domain and the target domain is defined as follows: in This indicates the expectation that the sample belongs to the source domain. G represents the expectation that a sample belongs to the target domain. s→t (x) represents the microseismic signal sample obtained from the source domain to the target domain transformation, D t (x) represents the discrimination result of the target discriminator on sample x; Cyclic consistency loss L cycle The following is a definition used to mitigate pattern collapse: Identity loss L identity To ensure that the generator input and the distribution of the inputs are similar, it is defined as follows: CAM classification loss L cam To distinguish the differences between the source and target domains, an auxiliary classifier η is used. s and The classification loss is defined as follows: Where η s (x) represents the probability that the sample comes from the source domain.
3. The semi-supervised microseismic first-arrival intelligent pickup method combining SimMatch and improved TransUGA according to claim 2, characterized in that, The specific process of semi-supervised training of the improved TransUNet network in Step 3 is as follows: The improved TransUNet network structure is as follows: The improved TransUNet modifies the original encoder to a combination of ResNet50 and Transformer, and connects the intermediate features output from the ResNet50 network through skip connections, and performs feature aggregation with the upsampling module in the decoder, fusing signal features at different resolutions to accelerate the convergence speed of the network. The ResNet50 network uses Group Normalization to process the data, making the network model more suitable for picking up the first arrival of signals; Before upsampling, ConvReLU is replaced with the SSPCAB module to maximize the features of the encoder output, suppress the influence of noise, and thus improve the network's noise resistance and the accuracy of initial arrival picking. The SSPCAB module consists of a masked convolutional layer and a Squeeze-and-Excitation module. The size of the masked convolutional kernel is k = 2k′ + 2d + 1, where k′ is a hyperparameter defining the kernel size and d is the dilation rate, used to control the local or global nature of the visible information. The receptive field of the visible region of the convolutional kernel is K. i , The loss function of the SSPCAB module is defined as follows: Where X is the input of this module, For output, G is the SSPCAB function; The improved TransUNet loss function is defined as follows: (8) In the formula The cross-entropy loss function for multi-class classification is calculated using the following formula: (9) p ic Let y be the predicted probability that the i-th sample belongs to class c, and M be the number of predicted classes. If the true class of sample i is c, then y ic Take 1 otherwise take 0; λ in equation (8) s It controls SSPCAB. Hyperparameters indicating the degree of importance.
4. The semi-supervised microseismic first-arrival intelligent pickup method combining SimMatch and improved TransUGA according to claim 3, characterized in that, The specific process of Step 4 is as follows: The SimMatch structure is used to improve the confidence of pseudo-labels for unlabeled samples during training. For unlabeled samples, a weak enhancement (T) is randomly applied. w (·)) or strong enhancement (T) s (·)), using an encoder F(·) to extract feature information, i.e. Using a fully connected class prediction head φ(·) to predict h b The mapping is expressed as semantic similarity, denoted as p = φ(h); For a cluster of different samples z k :(k∈(1,...,K)), z w and the i-th instance z i Instance similarity distribution between The calculation formula is as follows: in, t is a temperature coefficient used to control the smoothness of the data distribution; z b =g(h b g(·) is a non-linear mapping head that maps the feature h into a low-dimensional embedding vector; To use semantic similarity p w To calibrate instance similarity q w , will p w Expanding to K-dimensional space, the expansion formula is as follows: Where cls(·) is a function that returns the true first arrival time; To use instance similarity q w To calibrate semantic similarity p w , here q w Aggregating to L-dimensional space (L << K), the aggregation formula is as follows: Where 1(·) is the indicator function, if p i w and q j w If they belong to the same category, the value is 1; otherwise, it is 0. Calibrated semantic similarity Similarity to instances The calculation formula is as follows: Here, α is a hyperparameter used to control the weight relationship between semantic information and instances; when semantic similarity and instance similarity are close, the generated pseudo-labels will have higher confidence.