Image classification seismic facies division method and system

Abstract

The embodiment of the invention provides an image classification seismic facies division method and system, and belongs to the technical field of seismic data processing. The method comprises the following steps: performing multiple rounds of training on seismic data based on a pre-trained active learning initial model; pushing the annotation sample to a supervision end, and recovering a corresponding annotation result based on the supervision end; constructing an annotation sample set based on all annotation samples with annotation results; performing clustering processing on the historical seismic profile map to obtain clustering results of different seismic facies; based on the labeled sample set and the clustering result, expanding a basic training sample, and based on the expanded training sample and an improved LKA-ConvNext model, performing prediction model training; and predicting the unlabeled pre-stack seismic data based on the prediction model to obtain a seismic facies prediction result. According to the scheme of the invention, the problem that the existing seismic facies division scheme has large demand for labeled samples and strong multiplicity of solutions is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of seismic data processing, and particularly to an image classification seismic facies division method and an image classification seismic facies division system. Background Art

[0002] Currently, there are two main problems in using artificial intelligence for seismic signal reflection pattern analysis:

[0003] 1) Most traditional reflection pattern analyses are supervised, but such methods often rely on rich labeled sample data. However, in actual projects, the acquisition of labeled data is extremely expensive and difficult. Under such small-sample conditions, the trained model is prone to overfitting. Even after expanding the labeled samples by cutting or sliding time windows, it is still difficult to meet the requirements of model training. At the same time, such a sample expansion method does not consider the physical characteristics and geological properties for seismic reflection pattern analysis, and the reliability is relatively low.

[0004] 2) Currently, the multi-angle gathers of pre-stack seismic data contain rich formation reflection information and anisotropic characteristics. However, general supervised models are difficult to capture the global deep features between multi-angle gathers at relatively long distances, and do not strictly follow the physical characteristics and geological properties of geology, and are more susceptible to noise and local changes, resulting in strong multi-solution of the results.

[0005] According to whether well logging labels are used in the classification process, seismic signal reflection pattern analysis methods can be divided into two categories: supervised and unsupervised. Unsupervised methods mainly rely on the automatic analysis of pre-stack seismic data and do not require prior provision of geological labels or supervision information. They mainly identify underground structures and geological features by analyzing the reflected waves in geological data. Unsupervised methods mainly include clustering algorithms and feature mapping methods: Clustering algorithms start from the overall data and automatically classify seismic reflection signals into different clusters according to different difference criteria, and each type of signal belongs to the same seismic reflection pattern. The feature mapping method transforms seismic attributes through a certain transformation to facilitate the display of underlying structural features or reservoir characteristics. Supervised methods are based on existing well logging information or geological labels to establish a seismic pattern classification model. Supervised methods can make good use of the accurate information of geological data, learn the internal features from the labeled data through supervised learning, and thus infer geological structures and geological features, so they have received wide attention in recent years. Supervised seismic reflection pattern analysis mainly uses the convolutional neural network (CNN) as the core architecture, establishes a learning task by constructing a network model, and then trains the model parameters with the help of labels (usually well logging data), and finally enables the model to realize the prediction of seismic facies. Among them, according to the task category of the constructed model, supervised seismic pattern analysis can be roughly divided into two methods: image classification and semantic segmentation.

[0006] However, both image classification and semantic segmentation methods have high requirements for the quantity of well logging sample data and rely on rich well logging data samples. Although the image classification-based method expands the sample quantity to a certain extent by slicing labeled seismic images into patches, the degree of label repetition is high, still unable to meet the needs of model training, and it is necessary to assume that each patch belongs to the same seismic pattern, with low accuracy. In addition, existing supervised pre-stack models are directly transplanted from traditional image processing models. When processing, pre-stack seismic data is regarded as ordinary pictures for processing, and it is impossible to specifically learn the rich unique features of pre-stack seismic data. And the current commonly used seismic sample expansion methods such as structural modeling, data slicing and rotation still start from the data itself, lacking the guidance of prior knowledge, and the data reliability cannot be guaranteed. In terms of model structure, existing network models rely on convolutional neural networks to extract deep features in pre-stack seismic data. However, the multi-angle gathers of pre-stack seismic data are rich in formation reflection information and anisotropic features, and ordinary convolutional neural networks are difficult to accurately capture the correlation information between different angle gathers at a relatively long distance, and are more vulnerable to noise and local changes, resulting in strong multi-solution of the results. Aiming at the defects of large demand for labeled samples and strong multi-solution of the supervised model in the existing solutions, only a new seismic facies division scheme needs to be created. Summary of the Invention

[0007] The purpose of the embodiments of the present invention is to provide an image classification seismic facies division method and system to at least solve the defects of large demand for labeled samples and strong multi-solution of the supervised model in the existing solutions.

[0008] To achieve the above purpose, the first aspect of the present invention provides an image classification seismic facies division method, the method includes: collecting seismic data, and performing multiple rounds of training on the seismic data based on a pre-trained active learning initial model; after each round of training is completed, performing a sample screening once to obtain standard samples, and pushing the labeled samples to the supervision end, and based on the supervision end, retrieving the corresponding labeling results; obtaining the labeled samples with labeling results for each round of training, when the labeled samples with labeling results reach a preset quantity, constructing a labeled sample set based on all the labeled samples with labeling results; collecting historical seismic profiles, and performing clustering processing on the historical seismic profiles to obtain clustering results of different seismic facies; expanding the basic training samples based on the labeled sample set and the clustering results, and performing prediction model training based on the expanded training samples and an improved LKA-ConvNext model; performing prediction on unlabeled pre-stack seismic data based on the prediction model to obtain seismic facies prediction results.

[0009] Optionally, the training rule of the initial model is as follows: identify well logging data with seismic facies labeling results; in a pre-constructed basic upgrade network, use the well logging data with seismic facies labeling results as model input data for model training to obtain a result model, which serves as the initial model.

[0010] Optionally, the seismic data is unlabeled basic data obtained through actual measurement; after each round of training is completed, a sample screening is performed to obtain standard samples, and the labeled samples are pushed to the supervision end, and the corresponding labeling results are retrieved based on the supervision end, including: after each round of training is completed, calculate the uncertainty of each item of seismic data based on the training results; compare the uncertainties of each item of seismic data, and select the seismic data with the greatest uncertainty as the standard sample; push the standard sample to a preset supervision end; in response to the operation instruction of the preset supervision end, parse the labeling result based on the operation instruction.

[0011] Optionally, to obtain labeled samples with labeling results for each round of training, when the number of labeled samples with labeling results reaches a preset quantity, a labeled sample set is constructed based on all the labeled samples with labeling results, including: for each labeled sample with a labeling result obtained, filter out the labeled sample from the seismic data, and perform the next round of training based on the remaining seismic data; perform the training corresponding to the preset quantity for the corresponding number of times, and terminate the model training after the last round of training is completed and the labeled sample corresponding to the last round of training is obtained; summarize the labeled samples with labeling results obtained after each round of training to form a data set, which serves as the labeled sample set.

[0012] Optionally, collect historical seismic profiles and perform clustering processing on the historical seismic profiles to obtain clustering results of different seismic facies, including: perform clustering on the historical seismic profiles, and classify the seismic profiles according to the clustering results; perform the same-color labeling on the parts of the clustering results belonging to the same cluster to obtain the clustering results of different seismic facies.

[0013] Optionally, the improved LKA-ConvNext model is composed of alternating stacks of LKA-ConvNext Blocks and downsampling layers; the improved LKA-ConvNext model includes a constraint layer constructed based on the physical properties of seismic data; the improved LKA-ConvNext model is an inverse bottleneck structure with first upsampling convolution and then downsampling.

[0014] Optionally, the improved LKA-ConvNext model further includes: a LayerNorm normalization layer, a GELU activation function, a Layer Scale parameter scaling layer, and a per-sample inactivation layer; the GELU activation function is:

[0015] GELU(x) = x * Φ(x)

[0016] Where x is the input sample.

[0017] Optionally, expanding the basic training samples based on the labeled sample set and the clustering result, and training the prediction model based on the expanded training samples and the improved LKA-ConvNext model includes: combining the labeled sample set and the clustering result to form training samples, where each data sample in the training samples has a corresponding labeled result; inputting the training samples into the improved LKA-ConvNext model for class prediction; during the prediction process, calculating the loss between the physical attributes of the predicted class of the pre-stack input data and the true physical attributes of the true class, and updating the model parameters based on the calculated loss; re-executing the training based on the updated model parameters, and re-determining the loss between the physical attributes of the predicted class of the pre-stack input data and the true physical attributes of the true class based on the new training results until the loss value reaches the minimum, stopping the training, and obtaining the prediction model.

[0018] Optionally, predicting the unlabeled pre-stack seismic data based on the prediction model to obtain the seismic facies prediction result includes: collecting the unlabeled pre-stack seismic data to be predicted, using the unlabeled pre-stack seismic data as a parameter, and performing the prediction model training; inserting the training result into the seismic profile based on the coordinate range of the unlabeled pre-stack seismic data to obtain a predicted profile diagram as the seismic facies prediction result.

[0019] The second aspect of the present invention provides an image classification seismic facies division system, which includes: a collection unit for collecting seismic data and performing multiple rounds of training on the seismic data based on a pre-trained active learning initial model; a labeling unit for performing a sample screening once after each round of training to obtain standard samples, pushing the labeled samples to the supervision end, and receiving the corresponding labeled results based on the supervision end; a processing unit for obtaining the labeled samples with labeled results for each round of training, and constructing a labeled sample set based on all the labeled samples with labeled results when the number of labeled samples with labeled results reaches a preset quantity; a clustering unit for collecting historical seismic profile diagrams and performing clustering processing on the historical seismic profile diagrams to obtain clustering results of different seismic facies; a training unit for expanding the basic training samples based on the labeled sample set and the clustering result, and training the prediction model based on the expanded training samples and the improved LKA-ConvNext model; a prediction unit for predicting the unlabeled pre-stack seismic data based on the prediction model to obtain the seismic facies prediction result.

[0020] On the other hand, the present invention provides a computer-readable storage medium, on which instructions are stored, and when the instructions are run on a computer, the computer is caused to execute the above-mentioned image classification seismic facies division method.

[0021] Through the above technical solutions, the present invention proposes a label data augmentation processing algorithm, which uses a small number of labeled samples for active learning to preliminarily augment the labeled samples, performs clustering processing on seismic profile data to obtain the overall distribution of different seismic facies, and combines the labeled sample information and expert experience for correction and secondary supplementation, and finally maps it onto the pre-stack data volume to obtain a large number of pre-stack labels. The present invention adopts the LKA-ConvNext model with better feature extraction ability and robustness to extract non-linear features in seismic signals, fully excavates the global features of pre-stack seismic data, and on this basis, introduces physical property multi-modal as a constraint to make the model follow the physical laws of seismic data, limit the model prediction results within a certain space, reduce the multi-solution of the prediction results, and realize intelligent seismic pattern analysis based on pre-stack data.

[0022] Other features and advantages of the embodiments of the present invention will be described in detail in the subsequent specific embodiments. Description of the Drawings

[0023] The drawings are used to provide a further understanding of the embodiments of the present invention, and constitute a part of the specification, and are used together with the following specific embodiments to explain the embodiments of the present invention, but do not constitute a limitation to the embodiments of the present invention. In the drawings:

[0024] Figure 1 is a flowchart of the steps of the image classification seismic facies division method provided by an embodiment of the present invention;

[0025] Figure 2 is a diagram of the clustering result of a seismic profile provided by an embodiment of the present invention;

[0026] Figure 3 is a schematic diagram of the correction of a seismic profile provided by an embodiment of the present invention;

[0027] Figure 4 is a schematic diagram of the synthetic data model parameters provided by an embodiment of the present invention;

[0028] Figure 5 is a comparison diagram of the prediction model data results of different models provided by an embodiment of the present invention;

[0029] Figure 6 is a system structure diagram of the image classification seismic facies division system provided by an embodiment of the present invention. Specific Embodiments

[0030] The following will describe in detail the specific embodiments of the present invention with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for explaining and understanding the present invention, and are not used to limit the present invention.

[0031] According to whether well logging labels are used in the classification process, seismic signal reflection pattern analysis methods can be divided into two categories: supervised and unsupervised. Unsupervised methods mainly rely on the automatic analysis of pre-stack seismic data and do not require the prior provision of geological labels or supervision information. They mainly identify subsurface structures and geological features by analyzing the reflected waves in geological data. Unsupervised methods mainly include clustering algorithms and feature mapping methods: Clustering algorithms start from the overall data and automatically classify seismic reflection signals into different clusters according to different difference criteria, and each class of signals belongs to the same seismic reflection pattern. Feature mapping methods transform seismic attributes through a certain transformation to facilitate the display of underlying structural features or reservoir characteristics. Supervised methods, on the other hand, establish a seismic pattern classification model based on existing well logging information or geological labels. Supervised methods can make good use of the accurate information of geological data, learn the internal features from the labeled data through supervised learning, and thus infer geological structures and geological features, so they have received wide attention in recent years. Supervised seismic reflection pattern analysis mainly takes the convolutional neural network (CNN) as the core architecture, establishes a learning task by constructing a network model, and then trains the model parameters with the help of labels (usually well logging data), and finally enables the model to achieve the prediction of seismic facies. Among them, according to the task categories of the constructed model, supervised seismic pattern analysis can be roughly divided into two methods: image classification and semantic segmentation.

[0032] 1. Seismic reflection pattern analysis based on image classification:

[0033] An image classification method refers to calculating the overall category to which an image belongs through an algorithm model. In 2012, Alex Krizhevsky proposed the AlexNet network, which won the first place in the ISLVRC challenge through a more reasonable network construction and also greatly increased the attention in the field of image classification. In 2014, Simonyan [8] designed the VGG network architecture, achieving a larger receptive field by stacking multiple small convolutional kernels; in 2015, He et al. [9] proposed the residual network architecture and designed the ResNet model, realizing the fusion between features at different levels by designing residual modules; in 2017, the lightweight network MobileNet proposed by Howard

[10] et al. significantly reduced the model parameters while ensuring the training effect and is often used in devices with weak performance such as mobile terminals; in 2021, the Vision Transformer proposed by Alexey et al.

[11] introduced the attention mechanism in natural language processing into image processing, dividing the image into small patches and calculating self-attention respectively to replace the convolutional network, significantly improving the classification accuracy. Subsequently, the image classification model has developed rapidly and has also been widely applied in the field of seismic exploration: in 2016, Alaudah et al. proposed a weakly supervised label mapping algorithm, which can generate a large amount of training data by inputting a small number of pre-stack labels and use it for seismic pattern analysis; in 2018, Chevitarese et al. proposed that by dividing the pre-stack seismic profile image into multiple small image patches along the length and width and assuming that each patch belongs to the same category, inputting these patches into the classification network for training and prediction and then combining them, a complete seismic profile classification result based on pre-stack data can be obtained; Dramsch et al. proposed that the sliding window mechanism can be applied to the pre-stack seismic profile to greatly increase the number of cut patches and improve the model training effect.

[0034] 2. Seismic reflection pattern analysis based on semantic segmentation:

[0035] In addition to image classification, the method of semantic segmentation can also be used to divide the categories to which an image belongs, that is, to distinguish different geological targets on the same seismic profile. Different from classification, semantic segmentation does not output a single category for each image, but outputs the category to which each pixel on the image belongs. In 2015, Long et al. proposed the FCN model (Full Convolutional Neural Network). By migrating the feature extraction part of the classification task to the segmentation task and adding upsampling to restore the image size, the first semantic segmentation model was constructed. Subsequently, the field of semantic segmentation has developed rapidly and achieved good results in the field of geological exploration. In 2018, Dramsch achieved seismic pattern analysis based on patch-based CNN, and the classification results can be obtained by inputting pre-stack seismic data. In 2019, Alaudah et al. proposed a deconvolution network based on small-scale images and seismic profiles. The model can not only perform pre-stack seismic pattern analysis but also complete data augmentation to increase the number of samples. In 2020, Zhang proposed that the Deeplabv3+ network model can be used to extract multi-scale features of pre-stack seismic data through the ASPP pyramid structure, and good results were also obtained.

[0036] However, both image classification and semantic segmentation methods have high requirements for the quantity of well logging sample data and rely on rich well logging data samples. Although the method based on image classification expands the number of samples to a certain extent by cutting the labeled seismic images into patches, the repetition degree of the labels is relatively high, still unable to meet the needs of model training, and it is necessary to assume that each patch belongs to the same seismic pattern, with low accuracy. In addition, existing supervised pre-stack models are directly transplanted from traditional image processing models. When processing, the pre-stack seismic data is regarded as ordinary pictures for processing, and it is impossible to specifically learn the rich unique features of pre-stack seismic data. And today's commonly used seismic sample augmentation methods such as structural modeling, data slicing and rotation still start from the data itself and lack the guidance of prior knowledge, so the data reliability cannot be guaranteed. In terms of the model structure, existing network models rely on convolutional neural networks to extract deep features in pre-stack seismic data. However, the multi-angle gathers of pre-stack seismic data are rich in formation reflection information and anisotropic features, while ordinary convolutional neural networks are difficult to accurately capture the correlation information between different-angle gathers at a relatively long distance and are more vulnerable to noise and local changes, which will result in strong multi-solution of the results.

[0037] In view of the above problems, the solution of the present invention proposes an image classification seismic facies division method. The solution of the present invention proposes a labeled data augmentation processing algorithm, which uses a small number of labeled samples for active learning to initially augment the labeled samples, performs clustering processing on seismic profile data to obtain the overall distribution of different seismic facies, and combines the labeled sample information and expert experience for correction and secondary supplementation, and finally maps it to the pre-stack data volume to obtain a large number of pre-stack labels. The solution of the present invention uses the LKA-ConvNext model with better feature extraction ability and robustness to extract non-linear features in seismic signals, fully excavates the global features of pre-stack seismic data, and on this basis, introduces physical property multi-modal as a constraint to make the model follow the physical laws of seismic data, limits the model prediction results within a certain space, reduces the multi-solution of the prediction results, and realizes intelligent seismic pattern analysis based on pre-stack data.

[0038] Figure 1 It is the flowchart of the image classification seismic facies division method provided by an embodiment of the present invention. As Figure 1 shown, an embodiment of the present invention provides an image classification seismic facies division method, and the method includes:

[0039] Step S10: Collect seismic data and perform multiple rounds of training on the seismic data based on a pre-trained initial model of active learning.

[0040] Specifically, the training rule of the initial model is: identify well logging data with seismic facies label results; in a pre-constructed basic upgrade network, use the well logging data with seismic facies label results as model input data for model training to obtain a result model as the initial model.

[0041] In an embodiment of the present invention, for the processing of labeled data, the solution of the present invention draws on the processing process of handwritten digit recognition in the field of computer vision, and converts the pre-stack seismic data of multi-angle gathers into a picture form for training a prediction initial model. The core idea is to train an initial model of active learning with a small number of labeled well logging data pictures, perform one-time augmentation of labeled samples on the seismic data using the active learning model, and the active learning model will select the samples with the most abundant information, the greatest help for model training, and the greatest uncertainty according to the query strategy for expert annotation, making the annotation more accurate, realizing one-time augmentation of labeled well logging sample data, and finally using a clustering algorithm to obtain the overall distribution of different seismic facies on the seismic profile, and then combining the accurate labels of the labeled sample data augmented by active learning for manual correction and augmentation to obtain a large number of labeled sample data, realizing secondary augmentation of the samples.

[0042] Specifically, the key idea of active learning is to allow a machine learning model to select the data it needs from a data set for training. It uses a specific query function to select the most informative and most helpful samples for model training from an unlabeled sample set, and then labels them, so as to minimize the number of samples required for training the model, thereby solving the problem of difficult acquisition of sample labels.

[0043] In the embodiments of the present invention, there are multiple different query strategies in active learning to measure the information richness of samples. At the same time, active learning mainly has two application scenarios, one is active learning based on a data pool, and the other is active learning based on a data stream. In active learning based on a data pool, the elements in all unlabeled sample sets will be traversed by the query function and the information richness of each sample will be calculated, and then one or more samples will be selected and handed over to an expert for labeling. In active learning based on a data stream, unlabeled data will be sent to the active learning model in sequence, and the model will decide whether to manually label the sample according to the query function. If not, the sample will be discarded. At the same time, the key to active learning lies in the design of the query function. Currently, there are generally 2 most widely used query functions. One is the query function based on uncertainty. This query function uses 1 - max(p i (x)) to measure the uncertainty of a sample. That is to say, the greater the probability that a sample belongs to a certain definite class, the lower its uncertainty. Therefore, the information contained in this sample is less and it is less helpful for the training of the model. The solution of the present invention uses a query method based on uncertainty to select the samples to be marked. In each round, the sample with the greatest uncertainty is selected and handed over to an expert for manual labeling to achieve the expansion of labeled sample data.

[0044] Step S20: After each round of training is completed, perform a sample screening to obtain standard samples, and push the labeled samples to the supervision end, and based on the supervision end, recover the corresponding labeling results.

[0045] Specifically, the seismic data is the measured unlabeled basic data; the step of, after each round of training is completed, performing a sample screening to obtain standard samples, and pushing the labeled samples to the supervision end, and based on the supervision end, recovering the corresponding labeling results includes: after each round of training is completed, calculating the uncertainty of each item of seismic data respectively based on the training results; comparing the uncertainties of each item of seismic data, and selecting the seismic data with the greatest uncertainty as the standard sample; pushing the standard sample to a preset supervision end; in response to the operation instruction of the preset supervision end, parsing and obtaining the labeling result based on the operation instruction.

[0046] Step S30: Obtain the labeled samples with labeling results for each round of training. When the number of labeled samples with labeling results reaches a preset quantity, construct a labeled sample set based on all the labeled samples with labeling results.

[0047] Specifically, for each obtained labeled sample with a labeling result, filter out this labeled sample from the seismic data, and perform the next round of training based on the remaining seismic data; perform the training corresponding to the preset quantity for the corresponding number of times. After completing the last round of training and obtaining the labeled sample corresponding to the last round of training, terminate the model training; summarize the labeled samples with labeling results obtained after each round of training to form a data set as the labeled sample set.

[0048] Step S40: Collect historical seismic profiles and perform clustering processing on the historical seismic profiles to obtain the clustering results of different seismic facies.

[0049] Specifically, as Figure 2 , perform clustering on the post-stack seismic profile, classify the seismic profile according to the clustering results, manually frame out the parts of the clustering results belonging to the same cluster, and distinguish different clusters with different colors. In the clustering results, the specific categories of different clusters are not yet clear. According to the xline and inline positions of the well logging data, project the well logging data onto the post-stack seismic profile, and then combine the accurate seismic facies information and expert experience carried by the well logging interpretation data to expand the clustering results, such as Figure 3 . The specific situation can be divided into the following two types:

[0050] a. When the well logging is distributed in the clustering results (yellow label), it can be determined that all the clustering results in this category belong to the category to which the well logging belongs, thereby realizing the expansion of the labeled data;

[0051] b. When the well logging is distributed outside the background or clustering results (white label), through analysis combining expert experience, it can be judged that the blocks in the vicinity of the well logging belonging to the same category, and the clustering results are corrected and expanded.

[0052] Step S50: Expand the basic training samples based on the labeled sample set and the clustering results, and perform prediction model training based on the expanded training samples and the improved LKA-ConvNext model.

[0053] Specifically, the network model used in the present invention is an improvement based on the LKA-ConvNext network. The solution of the present invention introduces physical property multi-modalities as constraints on the basis of referring to the basic LKA-ConvNext network. The improved network model of the present invention is based on the LKA-ConvNext network architecture and is composed of alternating stacks of LKA-ConvNext Blocks and downsampling layers. Each part mainly uses operations such as large kernel attention and inverted bottleneck structures. Finally, a physical property multi-modal constraint layer composed of stacked multi-layer Conv2d and Linear is used.

[0054] 1) Physical property constraint layer: In the field of computer vision, regularization methods such as L1 regularization, L2 regularization, and Dropout are usually used to reduce the overfitting of the model, thereby improving its generalization ability and reducing the multi-solution problem. However, this method only simply considers reducing the multi-solution problem of the model and does not consider the physical laws of seismic data. Early neural networks mainly focused on simulating the human brain learning process through training and rarely considered the combination with physical models. However, for special physical problems, better combination of their physical properties and deep learning is required to better solve physical problems. Embedding physical information into the deep learning architecture can improve the performance of the model, which requires adding physical features, constraining network weights or activation functions in the network to meet physical constraints. In order to make the trained model more applicable to the field of seismic reflection pattern analysis, the solution of the present invention introduces the physical properties of seismic data (such as multiple attribute parameters such as dip angle) into the network as physical prior knowledge as constraints and uses it as part of the loss function, combining the training of the model with physical properties, which can be better applicable to the field of seismic reflection pattern analysis, can make full use of the characteristics of seismic data, improve the reliability and stability of the model, and reduce the multi-solution problem of the model. This physical property layer is composed of stacked multi-layer Conv, linear, and attention.

[0055] 2) Inverted bottleneck: In ordinary network models, a bottleneck structure that first reduces the dimension to compress data, extracts features, and then increases the dimension to reconstruct the data is usually adopted. However, during the process of compressing data, it is easy to cause information loss of the input data, restricting the expressive ability of the model. Therefore, the solution of the present invention adjusts the traditional bottleneck structure and adopts an inverted bottleneck structure that first increases the dimension by convolution and then reduces the dimension. By constructing cross-layer connections, information is converted between different-dimensional feature spaces, thereby avoiding the information loss caused by reducing the dimension and compression, and improving the model effect. On this basis, the solution of the present invention constructs an inverted bottleneck architecture integrating large kernel attention by cascading an Attention module and two Linear fully connected layers, where the height and width of the input and output remain unchanged, and only the number of channels changes.

[0056] Furthermore, the network module of the solution of the present invention mainly includes the LKA-ConvNextBlock of the LKA-ConvNext model and the downsampling layer (DownSample). By adding LayerNorm normalization, GELU activation function, Layer Scale parameter scaling, and per-sample inactivation layer (Drop Path) on the basis of the inverted bottleneck structure, its structure is made more perfect to achieve the extraction of different-level features of images. In the LKA-ConvNext Block, the dimensions of the input and output remain unchanged. The main function of the downsampling layer (DownSample) is to change the image dimension through a 2×2 convolution with a stride of 2.

[0057] 1) LayerNorm normalization: In a neural network model, it is often necessary to normalize the input or intermediate features to improve the performance, generalization ability, and stability of the model. The commonly used BatchNorm normalization has poor robustness to different input samples, so there may be problems when dealing with non-stationary signals such as seismic signals. Therefore, the solution of the present invention chooses to adopt LayerNorm normalization, which has stronger robustness and is more stable for small-batch training, and normalizes all neurons of an intermediate layer along the channel dimension of the input. The calculation formula is:

[0058]

[0059] For each input sample x, calculate the mean and variance of the output results of all neurons, and use the mean and variance to normalize to obtain the output result y.

[0060] 2) GELU activation function: The role of the activation function is to enhance the non-linear expression ability of the network model and the fitting ability for complex mapping relationships. In traditional network models, the most common activation functions are the RELU and ELU activation functions, etc.

[26] . However, due to the poor stability of the pre-stack seismic data itself, zero values are likely to appear during neuron calculations, making it impossible to continue backpropagation. Moreover, the non-linear expression ability of these activation functions is not flexible enough to capture the complex non-linear relationships in seismic signals. Therefore, the solution of the present invention abandons the most commonly used activation functions such as RELU and adopts a more comprehensive GELU activation function. The calculation formula is as follows:

[0061] GELU(x) = x * Φ(x)

[0062] where represents the cumulative probability distribution of the Gaussian distribution. As a non-monotonic activation function, the GELU activation function adds a Gaussian error linear unit to the input based on RELU, resulting in a smoother distribution, avoiding the appearance of invalid parameter values in neurons, thereby stabilizing the network gradient flow and avoiding phenomena such as gradient explosion.

[0063] In the embodiment of the present invention, first, the pre-stack seismic labels are uniformly scaled to a size of 224×224×3. Then, first, it passes through a 4×4 convolution and normalization layer, and then sequentially passes through multiple cascaded LKA-ConvNext Blocks to expand the number of channels and extract deep features. Three downsampling layers are added in the middle to change the image size. Finally, through the global pooling operation (GlobalAvg Pooling), the mean value of each channel in the input data is obtained, converting the length and width of the input to 1×1. Then, through the fully connected layer, the number of channels is mapped to the number of classes (classes) and physical features to be classified, obtaining the probability that the input is judged as each class and the physical features. For the classes output, the maximum probability is the classification result judged by the network. For the physical attribute output, the output physical features pass through the attribute multi-modal layer to obtain the physical attributes of the data.

[0064] Specifically, the labeled sample set and the clustering result are combined to form a training sample, and each data sample in the training sample has a corresponding labeling result; the training sample is input into the improved LKA-ConvNext model for class prediction; during the prediction process, the loss between the physical attributes of the predicted class of the input pre-stack data and the true physical attributes of the true class is calculated, and the model parameters are updated based on the calculated loss; training is re-executed based on the updated model parameters, and the loss between the physical attributes of the predicted class of the input pre-stack data and the true physical attributes of the true class is re-determined based on the new training results until the loss value reaches the minimum, and training is stopped to obtain the prediction model.

[0065] In the embodiments of the present invention, the calculation rule for the loss of the seismic signal pattern analysis task is as follows:

[0066]

[0067] Among them, L class is the loss of the seismic signal pattern analysis task; c is the total number of categories; y ij is the true label; is the model prediction value. The loss proposed by the solution of the present invention can effectively measure the difference between the category distribution predicted by the model and the true category label, and effectively constrain the prediction of the seismic signal reflection pattern.

[0068] Furthermore, the calculation rule for the loss between the physical property of the predicted category and the true physical property of the true category is as follows:

[0069]

[0070] Among them, n is the total number of samples; y i is the true physical property of the true category; is the physical property of the predicted category; this loss introduces physical property guidance for the model, and the formula of the combined loss function is as follows:

[0071] L total = λ1L class + λ2L phy

[0072] Among them, λ1 and λ2 are weight parameters used to balance the loss calculation

[0073] Step S60: Based on the prediction model, predict the unlabeled prestack seismic data to obtain the seismic facies prediction result.

[0074] Specifically, collect the unlabeled prestack seismic data to be predicted, use the unlabeled prestack seismic data as the input parameter, and execute the training of the prediction model; based on the range coordinates of the unlabeled prestack seismic data, insert the training result into the seismic profile to obtain a prediction profile diagram as the seismic facies prediction result.

[0075] In the embodiments of the present invention, calculate the position ranges of the xline and inline numbers of the corresponding 3D prestack data for the augmented large batch of labeled data, and project them onto the 3D prestack seismic data volume to obtain the prestack multi-trace seismic data at the corresponding positions. Calculate the mean value and difference of the labeled prestack multi-trace seismic data at each xline and inline position, plot the results as n curves, and fill the positive value regions of the curves to generate pictures to obtain a large batch of prestack labeled data.

[0076] Embodiment:

[0077] To quantitatively analyze the effect of the model proposed in the solution of the present invention on prestack seismic reflection pattern analysis, a model as shown in Figure 4 was used for testing. Figure 4 (a), 4(b) and 4(c) are the longitudinal wave velocity, transverse wave velocity and density model parameters respectively. The designed model can be regarded as a 200×700 matrix. There are 200 sampling points longitudinally in the model, and the sampling interval between sampling points is 1 ms; there are 700 columns horizontally, and the data of each column is called a trace. The total number of traces in the designed model is 700, and the trace interval is 1 m. The present invention uses the Zoeppritz equation to calculate the reflection coefficients of the prestack common reflection point (CRP) gather corresponding to each column of data. Each prestack gather has 5 traces, and the incidence angle ranges from 0° to 40°, with an incidence angle interval of 10°. Then, Ricker wavelets with 6 different frequencies, phases and amplitude intensities are convolved with the reflection coefficients to obtain the CRP gather. Finally, the synthetic model has 700 gathers, and every 100 gathers are one kind of seismic signal pattern, with a total of 7 kinds of seismic signal patterns.

[0078] During the implementation process, there are 100 sample pictures for each of the seven seismic reflection pattern types. The common parameters of all network models are: the division ratio of the training set and the validation set of the network is 8:2, the input resolution is 128×128, the loss function is cross entropy, the optimizer is adam, and the initial learning rate is 0.0005. In addition to the improved model proposed in the solution of the present invention, a conventional ConvNext model and a small-scale AlexNet model were also selected for comparative experiments to verify the effect of the improved model proposed in the solution of the present invention. The comparison of the accuracy rates of the training results of different models is shown in Table 1, and the training result diagram is as shown in Figure 5 , where (a) is the real result; (b) is the prediction result of the model of the solution of the present invention; (c) is the prediction result of the traditional ConvNext model; (d)

[0079] is the prediction result of the AlexNet model. It can be seen that each model can accurately predict most of the seismic signals. However, for the two types of signals between 400 and 600 traces, it is relatively difficult for the model to predict. Among them, the AlexNet model completely confuses these two types of seismic signals and lacks the ability to identify; the traditional ConvNext model can distinguish these two types of signals to a certain extent, but there is a large gap from the real result and it is impossible to judge the real distribution; the model of the solution of the present invention has the highest accuracy rate for the prediction results of these two types of signals, and the prediction result is closest to the real result distribution, which shows the effectiveness of the method proposed in the present invention.

[0080]

[0081] Table 1 Statistical table of the prediction results of different models on the model data

[0082] The method of the present invention is applied to the actual pre-stack seismic data in a certain work area to verify the effectiveness of the method. The Crossline range of the field seismic volume data is 7600 - 7970, and the Inline range is 7700 - 8250, including three seismic facies: background, channel 1, and channel 2. After the above-mentioned labeled data processing steps, a total of 78,594 labeled images are finally obtained, including 33,494 background labels, 19,896 channel 1 labels, and 25,204 channel 2 labels.

[0083] In this experiment, the common parameters of all network models are as follows: the division ratio of the training set and the validation set of the network is 8:2, the input resolution is 256×256, the loss function is cross-entropy, the optimizer is adam, and the initial learning rate is 0.001. This experiment selects comparative experiments to verify the effect of the improved model proposed by the present invention. First, the effects of three methods under traditional small-sample conditions are compared. Under traditional small-sample conditions, the effect of AlexNet is the worst. A large amount of background is misjudged as channels, and the contour and continuity of the channels cannot be seen, with very low accuracy, indicating that the effects of ConvNext and LKA are a bit better, but there are still many backgrounds misjudged as channels, with insufficient accuracy. Secondly, active learning models with different numbers of training rounds are trained as comparative experiments to verify the sample augmentation achieved by using the active learning method proposed by the present invention. It can be seen from the effect diagrams that at the 10th round after active learning, the result accuracy is not high, the boundary contour is not obvious, and it is impossible to accurately distinguish what is a channel. However, at the 30th round, the contour of the channel has been significantly improved, and a general contour of the channel can be obtained, but there are still many backgrounds misjudged as channels, with low continuity. After 50 rounds of training, the contour of the channel is clearer, reducing the possibility of background misjudged as channels and improving the accuracy to a certain extent, indicating that the sample augmentation of active learning has a great improvement on seismic reflection pattern analysis, and to a certain extent solves the problems of background misjudged as channels, low channel continuity, and low accuracy in small samples. Later, experiments are carried out using a large attention model after sample augmentation based on active learning, with and without physical attribute layer constraints as comparative experiments. It can be seen from the effect diagrams that after sample augmentation through active learning in the same way, there are still some cases where the background is misjudged as a channel without physical attribute layer constraints, and the multi-solution property is still relatively strong. However, with physical attribute layer constraints, many places where the background is misjudged as a channel are reduced, indicating that the multi-solution property is reduced to a certain extent.

[0084] After the model training is completed, we select five existing well logging data that have not participated in the training for prediction to test the model effect. The names of the five well logging data are QL3, QL16, QL203, QL205, and QL207. The well logging data all have accurate category information. The work area where the QL3 well logging is located belongs to mudstone. Among the remaining four well logging data, QL205 and QL207 belong to River 1; QL16 and QL203 belong to River 2. The prediction results of different models for these five well logging data are shown in Table 2. It can be seen that all models made mistakes in the prediction of QL3 well logging, but the improved model of the present invention can accurately distinguish the categories of the other four well logging, while there are certain errors in the prediction of these four well logging data by the ConvNext model and the AlexNet model. It can be seen that the method proposed by the present invention has a significant improvement in accuracy compared with the traditional model.

[0085]

[0086]

[0087] Table 2 Statistical table of prediction results of different models for well logging data

[0088] Next, the model is applied to the actual unlabeled data to predict the complete seismic profile, and qualitative analysis is carried out according to the prediction results. It can be seen that in the application of unlabeled data, the traditional supervised ConvNext model cannot maintain the high accuracy during training: for the area where river channels overlap (the upper right corner area of the prediction result), the model is difficult to predict. The prediction results of the small-scale AlexNet model are not ideal and cannot accurately predict continuous river channels. The improved model of the present invention and the ConvNext model have better prediction effects in these areas, and the improved method of the present invention has a better prediction effect for River Channel 1 and better continuity of the results.

[0089] Figure 6 It is the system structure diagram of the image classification seismic facies division system provided by an embodiment of the present invention. As Figure 6As shown in the figure, an image classification seismic facies division system is provided in an embodiment of the present invention. The system includes: an acquisition unit, configured to acquire seismic data and perform multiple rounds of training on the seismic data based on a pre-trained active learning initial model; an annotation unit, configured to perform a sample screening once after each round of training to obtain standard samples, and push the annotated samples to a supervision end, and based on the supervision end, retrieve the corresponding annotation results; a processing unit, configured to obtain the annotated samples with annotation results for each round of training, and when the number of annotated samples with annotation results reaches a preset number, construct an annotated sample set based on all the annotated samples with annotation results; a clustering unit, configured to acquire historical seismic profiles and perform clustering processing on the historical seismic profiles to obtain clustering results of different seismic facies; a training unit, configured to expand basic training samples based on the annotated sample set and the clustering results, and perform prediction model training based on the expanded training samples and an improved LKA-ConvNext model; a prediction unit, configured to perform prediction on unannotated prestack seismic data based on the prediction model to obtain seismic facies prediction results.

[0090] An embodiment of the present invention also provides a computer-readable storage medium, on which instructions are stored, and when running on a computer, the instructions cause the computer to execute the above-mentioned image classification seismic facies division method.

[0091] Those skilled in the art can understand that all or part of the steps in the methods of the above embodiments can be completed by instructing relevant hardware through a program. The program is stored in a storage medium, including several instructions to cause a single-chip microcomputer, a chip, or a processor to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical disks and other various media that can store program codes.

[0092] The above has described in detail the optional embodiments of the present invention in conjunction with the accompanying drawings. However, the embodiments of the present invention are not limited to the specific details in the above embodiments. Within the technical concept scope of the embodiments of the present invention, various simple modifications can be made to the technical solutions of the embodiments of the present invention, and these simple modifications all fall within the protection scope of the embodiments of the present invention. In addition, it should be noted that, among the various specific technical features described in the above specific embodiments, they can be combined in any appropriate manner without conflict. To avoid unnecessary repetition, the embodiments of the present invention do not separately describe various possible combination methods.

[0093] In addition, any combination can be made among various different embodiments of the present invention, as long as it does not violate the idea of the embodiments of the present invention, and it should also be regarded as the content disclosed by the embodiments of the present invention.

Claims

1. An image classification method for seismic facies division, characterized in that, The method includes: Collecting seismic data and performing multiple rounds of training on the seismic data based on a pre-trained active learning initial model; After each round of training, performing a sample screening to obtain standard samples, pushing the labeled samples to a supervision end, and recovering corresponding labeling results based on the supervision end; Obtaining labeled samples with labeling results for each round of training, and when the number of labeled samples with labeling results reaches a preset quantity, constructing a labeled sample set based on all the labeled samples with labeling results; Collecting historical seismic profiles and performing clustering processing on the historical seismic profiles to obtain clustering results of different seismic facies; Expanding basic training samples based on the labeled sample set and the clustering results, and performing prediction model training based on the expanded training samples and an improved LKA-ConvNext model; Performing prediction on unlabeled pre-stack seismic data based on the prediction model to obtain seismic facies prediction results.

2. The method according to claim 1, wherein The training rule of the active learning initial model is: Identifying well logging data with seismic facies label results; In a pre-constructed basic upgrade network, using the well logging data with seismic facies label results as model input data for model training to obtain a result model as the active learning initial model.

3. The method according to claim 1, characterized in that, The seismic data is measured unlabeled basic data; The step of, after each round of training, performing a sample screening to obtain standard samples, pushing the labeled samples to a supervision end, and recovering corresponding labeling results based on the supervision end includes: After each round of training, calculating the uncertainty of each item of seismic data respectively based on the training results; Comparing the uncertainties of each item of seismic data and selecting the seismic data with the largest uncertainty as the standard sample; Pushing the standard sample to a preset supervision end; Responding to an operation instruction of the preset supervision end and parsing to obtain a labeling result based on the operation instruction.

4. The method according to claim 1, wherein The step of obtaining labeled samples with labeling results for each round of training, and when the number of labeled samples with labeling results reaches a preset quantity, constructing a labeled sample set based on all the labeled samples with labeling results includes: For each obtained labeled sample with a labeling result, filtering out the labeled sample from the seismic data and performing the next round of training based on the remaining seismic data; Performing training for a corresponding number of times of the preset quantity, terminating the model training after completing the last round of training and obtaining the labeled sample with a labeling result corresponding to the last round of training; Summarizing the labeled samples with labeling results obtained after each round of training to form a data set as the labeled sample set.

5. The method according to claim 1, wherein The step of collecting historical seismic profiles and performing clustering processing on the historical seismic profiles to obtain clustering results of different seismic facies includes: Performing clustering on the historical seismic profiles and classifying the seismic profiles according to the clustering results; Performing the same-color labeling on the parts with clustering results belonging to the same cluster to obtain clustering results of different seismic facies.

6. The method according to claim 1, wherein The improved LKA-ConvNext model is composed of alternating stacks of LKA-ConvNext Blocks and downsampling layers; The improved LKA-ConvNext model includes a constraint layer constructed based on the physical properties of seismic data; The improved LKA-ConvNext model is an inverse bottleneck structure with first upsampling convolution and then downsampling.

7. The method according to claim 6, wherein The improved LKA-ConvNext model further includes: a LayerNorm normalization layer, a GELU activation function, a Layer Scale parameter scaling layer, and a per-sample inactivation layer; The GELU activation function is: GELU(x) = x * Φ(x) where x is the input sample.

8. The method according to claim 1, wherein The method of augmenting the basic training samples based on the labeled sample set and the clustering result, and training the prediction model based on the augmented training samples and the improved LKA-ConvNext model includes: Combining the labeled sample set and the clustering result to form training samples, where each data sample in the training samples has a corresponding labeled result; Inputting the training samples into the improved LKA-ConvNext model for class prediction; During the prediction process, calculating the loss between the physical property of the predicted class of the input pre-stack data and the true physical property of the true class, and updating the model parameters based on the calculated loss; Re-executing the training based on the updated model parameters, and re-determining the loss between the physical property of the predicted class of the input pre-stack data and the true physical property of the true class based on the new training results, until the loss value reaches the minimum, stopping the training, and obtaining the prediction model.

9. The method according to claim 1, characterized in that The method of predicting the unlabeled pre-stack seismic data based on the prediction model to obtain the seismic facies prediction result includes: Collecting the unlabeled pre-stack seismic data to be predicted, using the unlabeled pre-stack seismic data as input parameters, and executing the prediction model training; Inserting the training result into the seismic profile based on the range coordinates of the unlabeled pre-stack seismic data to obtain a predicted profile diagram as the seismic facies prediction result.

10. An image classification seismic facies division system, characterized in that, The system includes: a collection unit for collecting seismic data and performing multiple rounds of training on the seismic data based on a pre-trained active learning initial model; a labeling unit for performing a sample screening once after each round of training to obtain standard samples, pushing the labeled samples to the supervision end, and receiving the corresponding labeled results based on the supervision end; a processing unit for obtaining the labeled samples with labeled results for each round of training, and constructing a labeled sample set based on all the labeled samples with labeled results when the number of labeled samples with labeled results reaches a preset quantity; a clustering unit for collecting historical seismic profile diagrams and performing clustering processing on the historical seismic profile diagrams to obtain clustering results of different seismic facies; a training unit for augmenting the basic training samples based on the labeled sample set and the clustering result, and training the prediction model based on the augmented training samples and the improved LKA-ConvNext model; a prediction unit for predicting the unlabeled pre-stack seismic data based on the prediction model to obtain the seismic facies prediction result.

11. A computer-readable storage medium, characterized in that, Instructions are stored on the computer-readable storage medium, which when running on a computer cause the computer to execute the seismic facies classification method for image classification according to any one of claims 1-9.

Images