A pre-stack seismic reflection analysis method based on improved focal point loss

Abstract

The application discloses a prestack seismic reflection analysis method based on an improved focal point loss, which is applied to the field of seismic stratigraphy, and aims at the fact that the existing depth clustering method often ignores the representation ability of potential features and always has the sample misclassification problem in the clustering layer. First, prestack seismic data is obtained according to actual work area data and time window data, so as to prepare a network training sample data set. Then, the data set is iteratively trained by using a DCIF network model, the target distribution of the clustering layer is regarded as the real label distribution, the improved focal point loss is introduced into the clustering module, and is used in an unsupervised manner. The method is combined with the two modules to optimize the two modules, and learns the representation for the depth clustering in an end-to-end manner, and directly generates a seismic facies map.

Description

Technical Field

[0001] This invention belongs to the field of seismic stratigraphy, and specifically relates to a seismic reflection pattern analysis technique. Background Technology

[0002] Seismic reflection pattern analysis is one of the main goals of seismic stratigraphy, namely, analyzing seismic facies related to sedimentary environments and lithofacies distribution. In recent years, researchers have demonstrated through practical experiments the superiority of deep learning in the clustering task of seismic reflection pattern analysis. For example, Qian et al. proposed a DCAE network structure that can extract more accurate seismic reflection pattern features than conventional methods and exhibits superior seismic facies identification performance. However, this method requires two steps: deep feature extraction and deep feature clustering. The deep feature clustering step requires using traditional clustering algorithms such as K-Means and SOM to cluster the feature space to complete the seismic reflection pattern analysis.

[0003] Deep clustering is a novel clustering method that can directly cluster raw images. This method combines feature learning and clustering into a unified framework. Xie proposed the Deep Embedded Clustering (DEC) network to improve clustering performance by applying an autoencoder to extract embedded features from the input and designing clustering layers. Duan et al. combined convolutional autoencoders and clustering algorithms to effectively map high-dimensional data to a low-dimensional feature space and perform clustering analysis in this space, proposing the Deep Convolutional Embedded Clustering (DCEC) network to perform seismic reflection pattern analysis. Compared to the DCAE method, the DCEC method focuses more on further training in the feature space to promote more accurate clustering. The DCAE method, on the other hand, emphasizes the data representation capability, aiming to recover the original input seismic signal. Therefore, the DCEC method can be seen as a further development and improvement of the DCAE method.

[0004] Existing deep clustering methods still face challenges in pre-stack reflection pattern analysis, which warrant further investigation. Specifically, DEC and DCEC neglect the network's reconstruction capability, thus affecting the representation of latent features to some extent. Recent studies have shown that enhancing the representation of latent features can significantly improve clustering results. Furthermore, the label assignment mechanisms in existing deep clustering methods are inefficient when handling hard samples. Specifically, due to the uneven distribution of seismic data or the influence of noise interference, some samples in the clustering layer are always misclassified, referred to as hard samples. These may be the reason limiting further improvements in seismic reflection pattern analysis performance. Therefore, how to effectively mine these hard samples in deep clustering methods is a highly anticipated issue. Currently, focus loss has achieved significant success in image classification and object detection. This loss function employs a weight adjustment mechanism, giving greater weight to hard-to-classify samples, thereby focusing on these hard samples. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a pre-stack seismic reflection analysis method based on an improved focus loss function. In deep clustering, an improved focus loss function is used, which makes deep feature clustering pay more attention to hard samples, thereby improving the label allocation mechanism in the existing clustering framework.

[0006] The technical solution adopted in this invention is: a pre-stack seismic reflection analysis method based on improved focal loss, comprising:

[0007] S1. Obtain pre-stack seismic data by adding time windows based on the data's stratigraphic position;

[0008] S2. Construct a deep clustering network based on improved focus loss; the deep clustering network based on improved focus loss includes a convolutional autoencoder network module and a clustering module;

[0009] S3. Input the pre-stack seismic data obtained in step S1 into the deep clustering network based on improved focal loss constructed in step S2 for training, and generate a seismic phase map based on the trained deep clustering network based on improved focal loss.

[0010] The beneficial effects of this invention are as follows: The method of this invention introduces a binary cross-entropy penalty term into the reconstruction term to enhance the network's reconstruction ability, thereby enhancing the representation ability of latent features. Then, an improved focus loss is introduced to improve the label assignment mechanism of the clustering layer. To address the challenge of using focal loss with real labels, a mechanism for applying focal loss in an unsupervised manner is designed, utilizing self-training in deep clustering. Finally, the performance of the model of this invention in pre-stack seismic reflection pattern analysis is evaluated using both artificial model data and actual field data. The comprehensive experimental results show that the method of this invention can provide more geological information and obtain seismic facies maps that are closer to the original data. Finally, the discriminancy index is used to further verify the advantages of the DCIF method in pre-stack seismic reflection pattern analysis. Attached Figure Description

[0011] Figure 1 This is a diagram of the CAE network structure.

[0012] Figure 2 This is a network structure diagram of the clustering module;

[0013] Figure 3 For different modulation factors, the focus loss function is used.

[0014] Figure 4 A comparison of the focus loss function and the improved focus loss function;

[0015] Figure 5 This is a schematic diagram of the overall process of the DCIF method;

[0016] Figure 6 Data for artificial models;

[0017] Wherein, (a) is the model of artificial data; (b) is the root mean square amplitude attribute;

[0018] Figure 7 This is a rendering of the seismic phases of the artificial model work area.

[0019] Wherein, (a) is the root mean square amplitude; (b) is the DCIF effect diagram; (c) is the DCEC effect diagram; (d) is the DCAE+SOM effect diagram; and (e) is the Gabor+SOM effect diagram.

[0020] Figure 8 The root mean square amplitude attribute of seismic data in the ZG work area;

[0021] Figure 9 Seismic phase effect diagram of the ZG work area;

[0022] Wherein, (a) is the root mean square amplitude; (b) is the DCIF effect diagram; (c) is the DCEC effect diagram; (d) is the DCAE+SOM effect diagram; and (e) is the Gabor+SOM effect diagram.

[0023] Figure 10 The root mean square amplitude attribute of seismic data in the Super work area;

[0024] Figure 11 Seismic phase rendering of the Super work area;

[0025] Among them, (a) is the root mean square amplitude; (b) is the DCIF effect diagram; (c) is the DCEC effect diagram; (d) is the DCAE+SOM effect diagram; and (e) is the Gabor+SOM effect diagram. Detailed Implementation

[0026] To facilitate understanding of the technical content of this invention by those skilled in the art, the following technical terms will be explained first:

[0027] 1. Convolutional Autoencoder Network Model

[0028] Autoencoders are unsupervised learning methods that utilize neural networks to automatically learn features from data through bottleneck structures. Compared to traditional supervised learning methods, autoencoders do not require a large number of labeled samples and can automatically learn discriminative features. In recent years, various autoencoders have been widely used in research on clustering and anomaly detection tasks. Convolutional autoencoders (CAEs) are a special type of autoencoder that uses convolutional layers to extract high-level features from data and preserves local relationships through the convolutional kernels in each layer. Therefore, CAEs have become a natural choice for autoencoders processing images and data with local spatial connectivity (such as seismic data). CAEs mainly consist of two parts: the encoder part compresses the input into low-dimensional features, and the decoder takes the latent features as input and reconstructs the original data as accurately as possible.

[0029] Assuming pre-stack seismic images are used as input variables, they can be represented as a matrix. R represents a matrix, with superscripts indicating dimension and size. Here, N1 and N2 represent the number of sampling points and CDP gathers in the pre-stack seismic data along the time direction, respectively, and N is the total number of CDP gathers. The goal of CAE is to automatically learn the data that best represents X. n The deep features, in other words, are obtained by training to establish a mapping H(·), thereby enabling the reconstruction As much as possible with X n Similarity means The most representative of X n The depth features.

[0030] From pre-stack image X n Extract densely overlapping image patches and represent each patch as a feature vector. First, for each image patch X... nIt is convolved with a set of convolutional kernels w, where each filter w is a basis. Here, w is a tensor of size r×l×l×k, where r is the number of channels in the input (and since pre-stack seismic images can be represented as a single channel, r is always equal to 1), l represents the spatial size of the convolutional kernels, and k represents the number of convolutional kernels. Intuitively, this can be understood as the input X... n The model is convolved k times by a kernel of size l×l. Then, a max-pooling layer is added after the convolutional layers to reduce the number of model parameters and help prevent overfitting. These steps constitute the encoding process, where the original X... n After passing through convolutional and pooling layers, i feature maps H of the pre-stack seismic image can be obtained. i .

[0031] In summary, for input X n The representation of the i-th latent feature can be defined by formula (1).

[0032] H i =f(P(X) n *w (i) +B i )) (1)

[0033] The symbol * represents convolution or deconvolution operation, w (i) B represents the i-th convolutional kernel. i Let P(·) denote the bias, P(·) denote the max pooling operation, and f(·) denote the activation function of the nonlinear mapping.

[0034] To reconstruct the input data, a decoding process needs to be designed, which includes unpooling and deconvolution steps. The reconstructed input data can be represented by formula (2).

[0035]

[0036] Among them, H i Let E represent the i-th feature. i Ω represents the bias of each input channel, and Ω represents the combination of latent features. This indicates an unpooling operation. This indicates the deconvolution operation.

[0037] Formulas (1) and (2) are key to establishing the CAE model. The role of the loss function is to ensure that the learned representation retains the important features of the original data. In regression problems, mean square error (MSE) is a fairly common reconstruction loss function, and it is also often used in denoising autoencoders as a distance measurement between the original data and the decoded data.

[0038] However, practical experience has shown that the ability to use only MSE as the reconstruction loss function is limited. Especially in deep clustering, enhancing the reconstruction ability of the network means that the model can better learn and represent the latent features in the input data. This improved representation ability can further enhance the clustering performance. Therefore, enhancing the reconstruction ability of the network is very important in deep clustering. In such a setting, an appropriate loss is Binary CrossEntropy (BCE). Specifically, when the input data is standardized so that each value is within the range [0,1], i.e., the pixel is within [0,1], the degree of reconstruction can be interpreted as an estimate of the probability of a particular pixel being of interest. The main purpose of BCE is to maximize the probability of correct prediction using the maximum likelihood principle. Therefore, the loss function is used as shown in Equation (3), which is a weighted sum of the mean squared error (MSE) and the binary cross-entropy (BCE) after the training dataset is standardized.

[0039]

[0040] The MSE and BCE loss functions are defined as follows:

[0041]

[0042]

[0043] Wherein, λ controls the BCE loss for L MB The loss contributes to the structural differences between the original and reconstructed data, thereby enhancing the data's representational power in the feature space.

[0044] The above explains the working principle of CAE network models. The structure of a CAE network is as follows: Figure 1 As shown, this model can automatically capture the embedding features of the original pre-stack data.

[0045] The following section introduces an improved focus loss embedding mechanism into the clustering module, which further enhances the performance of deep clustering methods in seismic reflection pattern analysis.

[0046] A crucial component of deep clustering methods is the clustering layer. The clustering layer utilizes embedded features learned from feature learning as its input. The structure of the clustering module is as follows: Figure 2 As shown, it will be described in detail below.

[0047] The primary goal of clustering is to accurately divide data points into a specified number of categories. The effectiveness of clustering can be improved by learning features and defining clustering layers. A common approach in clustering layers is to use KL divergence to define the loss, which is a measure of the difference between the soft-label distribution Q and the target distribution P. The soft-label distribution refers to the probability distribution of each data point belonging to each category, while the target distribution refers to the probability distribution of evenly distributing all data points into each category. KL divergence is used to measure the similarity between two probability distributions, as shown in Equation (6).

[0048]

[0049] Define the soft label distribution Q using the student T-distribution, and calculate the representation H in the embedding layer. i With cluster centroid μ j The similarity is shown in formula (7).

[0050]

[0051] H here i That is, the latent features obtained according to formula (1);

[0052] The parameter α controls the degree of freedom of the student's T-distribution. Specifically, α is fixed at 1. Therefore, q ij The probability of assigning instance i to cluster center j is called soft label assignment. The target distribution P is in the form of formula (8).

[0053]

[0054] Where p ij It is q ij As defined, the goal of optimization is to match Q and P, so it can be viewed as a form of self-training.

[0055] However, in practical experience, it has been found that using only formula (6) to define the KL divergence between the soft label distribution and the target distribution as the clustering loss in the clustering module results in some difficult-to-classify misclassified samples. To solve this problem, this invention introduces an improved focus loss, a hard sample mining method designed to improve the label assignment mechanism. Focus loss was originally proposed to address the problem of a severe imbalance between positive and negative samples in single-stage target detection. For the binary classification case, let p and y represent the predicted label and the true label, respectively, and define p... t as follows:

[0056]

[0057] Simultaneously, the class imbalance problem can be solved by introducing weight factors α∈[0,1] for class 1 and 1-α for class -1. For ease of labeling, a similar p... t Define α in this way t , The α-balanced cross-entropy loss (CEL) can then be expressed as formula (10).

[0058] CELoss(p,y)=CELoss(p t )=-α t log(p t (10)

[0059] When there are many easily classifiable samples, the CEL calculation method tends to be ineffective for difficult-to-classify samples. Focal loss (FL) adds a modulation factor to CEL, which reduces the weight of easily classifiable samples. This allows the model to focus more on the difficult-to-classify samples. Therefore, the focal loss function is formalized as follows:

[0060] l FL (p t )=-α t (1-p t ) γ log(p t (11)

[0061] Where (1-p) t ) γ Let γ represent the modulation factor, and γ be an adjustable focusing parameter satisfying γ≥0. Setting γ>0 reduces the number of easily classified examples (p). t With a relative loss greater than 0.5, focus attention on difficult, misclassified examples, such as... Figure 3 As shown. The solid line represents the curve CEL.

[0062] according to Figure 3 The influence of different modulation factors γ on focus loss shows that as the modulation factor γ increases, the loss value of the sample decreases rapidly, especially for easily classified samples, thus increasing the weight of difficult-to-classify samples. However, this invention finds that although the weight of difficult-to-classify samples is relatively increased, the loss of difficult-to-classify samples also decreases rapidly. Therefore, to further improve the weight of difficult-to-classify samples, this invention improves the focus loss function by optimizing the modulation factor. To this end, this invention considers the base of the modulation factor of focus loss as a function, denoted as f(p) = 1 - p, where the range of p is [0,1], and the range of f(p) is also [0,1]. Let the optimized base function be h(p), which needs to satisfy the following conditions simultaneously:

[0063] 1. h(p) passes through (0, 1) and (1, 0) and the variable p is in the range [0, 1]. h(p) is a decreasing function.

[0064] 2. In the range (0, 1), h(p) needs to intersect f(p) at (0.5, 0.5), and in the interval (0, 0.5), h(p) is greater than f(p), and in the interval (0.5, 1), h(p) is less than f(p).

[0065] Therefore, in this invention, by scaling and translating the cosine function, the function h(p) is obtained, and the previously mentioned conditions are satisfied. Then h(p) = 1 / 2(1+cos(π·p)), which is used as the base of the focal loss modulation factor. Thus, the improved focal loss function (IFL) is generated as shown in formula (12).

[0066]

[0067] Figure 4 The diagram illustrates the focus loss function and its optimized counterpart, with γ = 2. The two loss functions intersect at (0.5, 0.5). Within the interval (0, 0.5), the IFL value is greater than the FL value. However, within the interval (0.5, 1), the IFL value is less than the FL value. This indicates that the improved focus loss function further increases the weight of hard-to-classify samples, thus making the model more focused on classifying hard samples.

[0068] In clustering tasks based on pre-stack reflection patterns, multi-class scenarios are often required, while the aforementioned focus loss function is primarily designed for binary classification tasks. Therefore, to accommodate multi-class requirements, the p in formula (12) needs to be modified. t Modify and redefine the calculation method of the loss function. Specifically, change p... t Redefining as follows:

[0069] p t =y pred *y true (13)

[0070] Where y pred and y true These are the predicted labels and true labels in the case of multiple classes.

[0071] Generally, real labels are needed in this case, but this becomes supervised learning, which is inconsistent with the unsupervised clustering task of this invention. Fortunately, as mentioned above, the soft label distribution Q and the target distribution P generated in the clustering module provide a direct solution. The predicted labels are generated by Q, and the real labels are generated by P, so the focus loss can be directly applied to the clustering layer during self-training. In practice, an α-balanced focus loss variant is used, taking this form because it offers a slight improvement in accuracy compared to the non-α-balanced form. Therefore, the objective can be formalized as follows:

[0072]

[0073] According to the definition of formula (14), it can be known that for an easily classifiable sample, i.e., p t Approximately 1, 1 / 2(1+cos(π·p) t The value is very small. For samples that are difficult to classify, i.e., p... t Approaching 0, 1 / 2(1+cos(π·p) t The larger value of L indicates that the improved focus loss can further increase the weight of hard samples in the loss function while decreasing the weight of easy samples, thus focusing more on training hard-to-classify samples. Therefore, the objective function L of the clustering module... CIFL It can be formalized as:

[0074] L CIFL =l CL +l IFL (15)

[0075] Among them, l CL To enable the network to cluster in a self-training manner, l IFL An improved focus loss is employed to address the issue of using only l in the clustering module. CL This addresses the problem of some hard samples being difficult to mine.

[0076] By using the binary cross-entropy enhanced latent feature representation learning module and the improved focus loss embedding clustering module defined above, the objective function of the DCIF network can be formalized as:

[0077] L = L MB +ηL CIFL (16)

[0078] Where η controls the clustering layer loss L CIFL The contribution of η to the total loss L ranges from 0 to 1.

[0079] This invention presents a Deep Clustering (DCIF) method based on improved focus loss, aiming to address the sample misclassification problem in the clustering layers of existing deep clustering methods. To learn more effective features, a binary cross-entropy loss is introduced into the method. Specifically, the loss function is a weighted sum of the mean squared error (MES) of the training dataset and the binary cross-entropy (BCE) of the normalized data. The BCE loss penalizes the structural differences between the original and reconstructed data, thereby learning more features. Secondly, an optimized focus loss is used to improve the label assignment mechanism in the clustering module. However, focus loss is typically applied in supervised learning scenarios because it requires real labels. Therefore, embedding focus loss into an unsupervised clustering framework is the biggest challenge. To address this issue, this invention utilizes the self-training within the network to treat the target distribution in the clustering layers as the real label distribution, thus designing a mechanism to apply focus loss in an unsupervised manner. By introducing this mechanism, the performance of deep clustering methods in seismic reflection pattern analysis is further improved, which is beneficial for generating accurate pre-stack seismic phase maps.

[0080] Table 1 DCIF Algorithm Flow

[0081]

[0082] Table 1 summarizes the algorithm flow for pre-stack seismic reflection mode analysis based on improved focal loss.

[0083] The training strategy for the DCIF network will be introduced below. During training, this invention utilizes stochastic gradient descent and backpropagation to jointly optimize the convolutional autoencoder network and the cluster center points μ. j The parameters are: where ρ is the learning rate, m is the number of samples in the mini-batch, and μ is the cluster center. j The update is as follows:

[0084]

[0085] The weights W of the feature learning module e and W d Updated to:

[0086]

[0087]

[0088] Furthermore, the clustering module of this invention treats the target distribution as true labels and relies on the soft label distribution for updating via equation (8). Therefore, when updating the target distribution, label C i Assigned to sample X according to the following formula n :

[0089]

[0090] This invention is based on an improved focal loss-based deep clustering (DCIF) algorithm. Its goal is to achieve deep clustering in the hidden layers through automatic learning using KL divergence and an improved focal loss, thereby directly generating seismic phase maps. The processing flow based on the DCIF algorithm for pre-stack seismic reflection mode analysis can be summarized as follows:

[0091] (1) Obtaining the dataset. To study the actual work area data, pre-stack seismic data is obtained based on the stratigraphic level and by adding a time window. Since this invention uses a deep learning method, the four-dimensional seismic data needs to be segmented along the timeline to process it into two-dimensional seismic data, which is then used as the training set input. Figure 5 In the network shown.

[0092] (2) Training the DCIF network. First, the DCIF network needs to be constructed, including convolutional layers, pooling layers, and upsampling layers. Parameters λ and α are then set to control the binary cross-entropy loss, the focus loss optimized by the clustering layer, and the contribution of the clustering layer to the total loss function. t And η. In addition, it is necessary to determine important parameters such as the number of pre-training iterations and the input image patch size. To obtain the initial parameters of the DCIF network: encoder weights W... e and decoder weights W d This invention uses a CAE model for pre-training to obtain an initial estimate of the feature mapping, including weights. Specifically, the dataset obtained in (1) is input into the CAE model, and then the initial parameters of the CAE are used as the initial parameters of the DCIF network, such as the encoder weights and decoder weights mentioned above.

[0093] After initializing the CAE network, the seismic data prepared in (1) is used to train the CAE network. The reconstruction loss function of the CAE network is composed of a weighted sum of the MSE loss function and the BCE loss function. The clustering layer loss function is composed of the KL divergence and the improved focal loss function. Then, the CAE network weights are updated through stochastic gradient descent (SGD) and backpropagation (BP). This process is repeated until the loss value decreases to a stable level, thus completing the training of the network.

[0094] (3) Unsupervised pre-stack seismic reflection pattern analysis. By introducing a binary cross-entropy penalty term to learn more effective features, and introducing an improved focus loss to improve the label assignment mechanism of the clustering layer, deep clustering is completed in the hidden layer through automatic learning, and seismic phase maps are directly generated.

[0095] The following section will provide a detailed overview of the algorithm and its application analysis in seismic data.

[0096] The overall process of this invention is as follows: Figure 5 As shown. First, pre-stack seismic data is obtained based on actual work area data and time window data to prepare the training sample dataset for the network. Next, the DCIF network model is used for iterative training on the dataset. The upper part incorporates a binary cross-entropy loss function, with the objective function expressed as L. MB This is represented by a binary cross-entropy penalty term, which penalizes the structural differences between the original and reconstructed data, thereby learning more features. The lower part is a clustering module that generates soft-label distributions and target distributions to measure the target L. CIFL By leveraging the network's self-training, the target distribution of the clustering layer is treated as the true label distribution. An improved focus loss is introduced into the clustering module and used in an unsupervised manner. This invention jointly optimizes the convolutional autoencoder (CAE) module and the clustering layer module, learning a deep clustering-oriented representation in an end-to-end manner to directly generate seismic phase maps.

[0097] To comprehensively verify the performance of the improved focus loss method proposed in this invention in seismic reflection pattern analysis and highlight its superiority, this invention conducts experimental comparisons with three existing algorithms. First, the Gabor feature extraction method proposed by Song Chengyun is introduced, which this invention uses as a representative of waveform feature extraction. The second deep learning method is unsupervised deep convolutional autoencoder feature extraction proposed by Qian, which this invention uses as a representative of deep learning feature extraction. Finally, deep convolutional embedded clustering is introduced, which implements end-to-end deep clustering and serves as a representative of deep clustering before the improvement. The following sections will provide detailed descriptions and experimental results for each of these three existing algorithms.

[0098] (1) Gabor Feature Extraction Algorithm

[0099] Gabor feature extraction-based methods treat target data as two-dimensional images, making them applicable to various forms of pre-stack gathers. These gathers contain numerous features, such as amplitude information, information on the influence of azimuth on amplitude transformation, and information on the influence of incident angle on amplitude variation. This information, after being processed by the Gabor function, is all contained in the Gabor features, thus providing a better description of the details of seismic facies. SOM clustering is used to cluster the features extracted by Gabor, completing the pre-stack seismic reflection mode analysis. The main principles of Gabor feature extraction and the SOM clustering algorithm are existing known technologies and will not be described in detail here.

[0100] (2) Depthwise Convolutional Autoencoder (DCAE)

[0101] DCAE is an unsupervised deep learning method that can learn nonlinear, discriminative, and invariant features from unlabeled data over long periods. Compared to convolutional neural networks, which require large amounts of labeled training data, DCAE has advantages in that it can automatically learn deep features from pre-stack seismic data. This invention uses SOM clustering to cluster the features extracted by DCAE to complete pre-stack seismic reflection pattern analysis. The main principles of DCAE feature extraction and the SOM clustering algorithm are existing known technologies and will not be described in detail here.

[0102] (3) Deep Convolutional Embedded Clustering (DCEC)

[0103] DCEC, a popular deep clustering method, directly constructs a clustering-oriented loss function on embedded features to jointly refine features and assign clusters, thereby better learning clustering information in the feature space and achieving deep clustering. The main principle of the DCEC algorithm applied to pre-stack seismic reflection mode analysis is a known existing technology, which will not be described in detail here.

[0104] Experiment 1

[0105] To verify the effectiveness of the proposed DCIF network in analyzing reflection patterns of pre-stack seismic data and to validate the theoretical model of the proposed DCIF network, this experiment will test its performance by applying the method to pre-stack seismic data from an artificial model.

[0106] The artificial model data contains 500 XLines (connecting seismic lines) and 400 Inline lines (main seismic lines), totaling 549,500 pre-stack seismic traces. This data includes 5 azimuth points (CDP) and 30 time sampling points (Timeline), meaning the physical model data size is (500×400×30×5). If the input network data image size is set to Timeline×CDP, meaning the sample image dimension is 30×5, then the dataset contains 200,000 sample images. The invention is designed as follows... Figure 6 The three models shown in (a) correspond to media 1, 2, and 3 formed by reflection coefficients with different azimuths and incident angles, respectively. To demonstrate the effectiveness of the final seismic phase diagram features, and because the root mean square amplitude attribute of the data is sensitive to amplitude changes and can be used to identify seismic anomalies caused by sedimentation such as channels, faults, and fractures, and to distinguish between continuous deposition and chaotic reflections, the root mean square amplitude attribute of the seismic data was selected for verification. The root mean square amplitude of the artificial model data is calculated as follows: Figure 6 As shown in (b).

[0107] Mainly includes Figure 6 The artificial model data shown. Figure 6(a) illustrates three media, each formed by reflection coefficients at different azimuths and incident angles. Media 3 simulates river channel features, media 2 simulates rock features, and media 1 simulates other features. From Figure 6 (b) The root mean square attribute shows that there are obvious channel features and other lithological features in the artificial data. When comparing the algorithm of this invention with existing algorithms, the main observation is whether the algorithm can correctly identify and classify these two features.

[0108] The following compares the performance of four methods on artificial model data. Among these methods, the traditional feature extraction algorithm employs the SOM unsupervised clustering algorithm, setting the number of classifications for all experiments to 3 (to distinguish between river channels, rocks, and other materials). The four methods are: the DCIF of this invention, whose processed seismic facies results are as follows... Figure 7 (b) shows the seismic phase results obtained from existing DCEC processing. Figure 7 As shown in (c); the existing DCAE feature extraction + SOM clustering method produces the following seismic facies results: Figure 7 As shown in (d); the existing Gabor feature extraction + SOM clustering algorithm produces the following seismic facies results: Figure 7 As shown in (e).

[0109] exist Figure 7 In the seismic phase renderings of the artificial model data, based on the interpretation of the root mean square amplitude, it can be seen that the Gabor+SOM method ( Figure 7 (e) and DCAE+SOM method ( Figure 7 (d) can identify the contour features of the data, but it does not classify channels and other lithologies. The DCEC method ( Figure 7 (c) As a representative of deep clustering, DCEC has significantly outperformed the DCAE+SOM and Gabor+SOM methods, especially in river regions, where it can correctly classify more areas. However, DCEC still has shortcomings, namely, it cannot completely identify river features. In contrast, the DCIF method of this invention (…) Figure 7 (b) It can not only fully identify the contours of river channels and other lithologies, but also classify them correctly.

[0110] In summary, the seismic facies results obtained using the DCIF method are closer to those obtained using artificial model data. Comparative experiments demonstrate that the method proposed in this invention outperforms the other three existing methods.

[0111] Experiment 2

[0112] To further verify the effectiveness of the proposed Deep Clustering with Improved Focal Loss (DCIF) method in practical applications, this experiment conducted pre-stack seismic data tests in actual work areas ZG and Super.

[0113] This experiment studies 4D pre-stack seismic data located in the ZG seismic work area of ​​the Sichuan Basin in southwestern China, covering an area of ​​approximately 120 square kilometers. By vertically selecting a suitable time window centered on the target horizon, the experiment yields data of size 1182×261×15×50, with an Xline size of 1182 and an Inline size of 261. It includes 15 azimuth dipole (CDP) points and 50 timeline sampling points. This step aims to flatten the horizon. The root mean square amplitude calculation for the ZG seismic work area is as follows: Figure 8 As shown.

[0114] This experiment primarily interprets the root mean square amplitude properties of the data from the ZG work area. Figure 8 In this diagram, the elliptical region primarily represents the fault area. When comparing the algorithm of this invention with existing algorithms, for the fault area, the main observations are whether the fault is continuous, clear, and other possible identification errors.

[0115] The following compares the performance of four methods on the ZG work area data. Among these methods, the traditional feature extraction algorithm employs the SOM unsupervised clustering algorithm, and the number of classifications in all experiments is set to 5 (based on experience and the Davidson-Bolding index, a classification fitness metric). The four methods are: using the DCIF of this invention, the seismic facies results obtained are as follows... Figure 9 As shown in (b); the seismic phase results obtained by DCEC processing are as follows: Figure 9 As shown in (c), the seismic facies results obtained by DCAE feature extraction + SOM clustering are as follows: Figure 9 As shown in (d), the seismic facies results obtained by using the Gabor feature extraction + SOM clustering algorithm are as follows. Figure 9 As shown in (e).

[0116] exist Figure 9 In the seismic phase renderings of the ZG work area data, the interpretation of the root mean square amplitude reveals that the Gabor+SOM method ( Figure 9 (e) and DCAE+SOM method ( Figure 9 (d) The faults identified in the fault regions circled by the ellipse are blurry and discontinuous, while the DCEC method ( Figure 9 (c) Some contours were identified more clearly than those of the two traditional algorithms, but the overall tomographic discontinuity remained. In contrast, the DCIF method of this invention ( Figure 9(b) Not only can the fault outline be clearly identified, but the identified faults are also continuous. Therefore, the DCIF method of this invention provides more accurate geological information. A comprehensive comparison of the seismic facies results of the four methods leads to the conclusion that the seismic facies results of the DCIF method of this invention are closer to the actual ZG work area data. Therefore, the method proposed in this invention is superior to the other three existing methods.

[0117] Experiment 3

[0118] This experiment studies 4D pre-stack seismic data located in the Super seismic work area of ​​the Sichuan Basin in southwestern China, covering an area of ​​approximately 80 square kilometers. By vertically selecting a suitable time window centered on the target horizon, a data size of 551×371×41×10 is obtained, with an Xline size of 551 and an Inline size of 371. This includes 10 azimuth dipoles (CDP) and 41 timeline sampling points, a step aimed at flattening the horizon plane. If the input network image size is set to Timeline×CDP, i.e., the sample image dimension is 41×10, the dataset contains 204,421 sample images. To demonstrate the effectiveness of the final seismic phase map features, and because the root mean square amplitude (RMS) attribute of the data is sensitive to amplitude changes and can be used to identify seismic anomalies caused by sedimentation such as channels, faults, and fractures, and to distinguish between continuous deposition and stray reflections, the RMS amplitude attribute of the seismic data is used for verification. The RMS amplitude calculation for the Super work area is as follows: Figure 10 As shown.

[0119] This experiment primarily interprets the root mean square amplitude properties of the Super work area data. Figure 10 In the diagram, the elliptical regions in the upper left and upper right corners represent small fault areas, while the elliptical region at the bottom represents large fault areas. The comparison between the algorithm mentioned in this invention and existing algorithms primarily focuses on whether the fault regions are continuous and clear, and other areas prone to identification errors.

[0120] The following compares the performance of four methods on Superworks data. Among these methods, the traditional feature extraction algorithm employs the SOM unsupervised clustering algorithm, and the number of classifications in all experiments is set to 5 (based on empirical and classification fitness metrics, the Davidson-Bolding index). The four methods are: the DCIF of this invention, whose seismic facies results are as follows... Figure 11 As shown in (b); the existing DCEC, the seismic phase results obtained from its processing are as follows Figure 11 As shown in (c); the existing DCAE feature extraction + SOM clustering method produces the following seismic facies results: Figure 11As shown in (d); the existing Gabor feature extraction + SOM clustering algorithm produces the following seismic facies results: Figure 11 As shown in (e).

[0121] exist Figure 11 In the seismic phase rendering of the Super seismic data, based on the interpretation of root mean square amplitude, we first observe the small fault area, using the DCAE+SOM method ( Figure 11 (d)), Gabor+SOM method ( Figure 11 (e) shows that both the DCAE+SOM and Gabor+SOM methods can clearly identify the contours of small fault regions, but most fault regions (the upper left fault region) are not identified, and the faults are discontinuous. Secondly, observing large fault regions, both the DCAE+SOM and Gabor+SOM methods are unclear about the features of large faults and do not perform good classification. The DCEC method ( Figure 11 (c) As a representative of deep clustering, the DCEC method of this invention is significantly superior to the DCAE+SOM and Gabor+SOM methods, mainly in small fault regions (upper right corner). The DCEC method of this invention identifies faults that are more continuous and clearer. However, it does not clearly identify the same small fault regions (upper left corner) and large fault regions. In contrast, the DCIF method of this invention ( Figure 11 (b) The identified small fault regions are clearer and more continuous, especially the fault region in the upper left corner. Regarding the identification of large fault features, other methods cannot clearly identify them, while the DCIF of this invention can clearly identify large fault features.

[0122] Analysis of the seismic facies in the ZG and Super seismic zones yields the following conclusions: The proposed Deep Clustering with Improved Focal Loss (DCIF) method can be used for pre-stack seismic reflection pattern analysis, providing clearer geological information. The excellent results in both zones also demonstrate the advantages of the LMVAE algorithm.

[0123] In this embodiment Figure 6 , 7 In 8, 9, 10, and 11, the horizontal axis Xline represents the seismic interpretation profile of the connecting line, and the vertical axis Inline represents the seismic interpretation profile of the main survey line. Figure 3 , 4 The vertical axis represents the focal loss function, and the horizontal axis represents the probability of ground truth samples. Figure 5In this context, Improving Focal loss represents the improved focal loss function, Clustering loss represents the clustering loss function, and Total loss represents the total loss function.

[0124] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the scope of the claims of the invention.

Claims

1. A pre-stack seismic reflection analysis method based on improved focal loss, characterized in that, include: S1. Obtain pre-stack seismic data by adding time windows based on the data's stratigraphic position; S2. Construct a deep clustering network based on improved focus loss; The deep clustering network based on improved focus loss includes a convolutional autoencoder network module and a clustering module; S3. Input the pre-stack seismic data obtained in step S1 into the deep clustering network based on improved focal loss constructed in step S2 for training, and generate a seismic phase map based on the trained deep clustering network based on improved focal loss. The loss function expression for the clustering module is: ； in, Denotes KL divergence, This represents the improved focus loss function; The expression is: ； in, , and These are the predicted labels and true labels in the case of multiple classes. As a weighting factor, These are adjustable focusing parameters.

2. The pre-stack seismic reflection analysis method based on improved focal loss according to claim 1, characterized in that, The convolutional autoencoder network module includes an encoding layer, an embedding layer, and a decoding layer. The encoding layer compresses the input data into low-dimensional features. The low-dimensional features are processed by the embedding layer to obtain latent features, which are then used as input to the decoding layer. The decoding layer outputs the reconstructed data.

3. The pre-stack seismic reflection analysis method based on improved focal loss according to claim 2, characterized in that, The clustering module includes a clustering layer, the input of which is the output of the embedding layer of the convolutional autoencoder network module, and the output of the clustering layer is the label of the reconstructed data.

4. The pre-stack seismic reflection analysis method based on improved focal loss according to claim 3, characterized in that, In step S3, during training, the loss function of the convolutional autoencoder network module is composed of a weighted average of the root mean square error loss function and the binary cross-entropy loss function; the loss function of the clustering module is composed of the KL divergence and the improved focal loss function. The weights of the convolutional autoencoder network modules are updated using stochastic gradient descent and backpropagation, thus completing the training of a deep clustering network based on an improved focus loss.

5. The pre-stack seismic reflection analysis method based on improved focal loss according to claim 4, characterized in that, The loss function expression for the convolutional autoencoder network module is: ； in, This represents the root mean square root error loss function. This represents the binary cross-entropy loss function. Controlling the contribution of binary cross-entropy loss to root mean square error loss, This represents the data input to the convolutional autoencoder network module. This represents the reconstructed data output by the convolutional autoencoder network module.

Images