A seismic phase prediction method based on Transformer structure self-supervision

Abstract

The present invention provides a self-supervised seismic phase prediction method based on a Transformer structure. The method performs preliminary identification of seismic volume data based on well logging information and a clustering algorithm, selects well-connected profiles as unlabeled data, performs masking processing, and obtains masked unlabeled data. An auxiliary network is used to reconstruct a large amount of masked unlabeled data, trains the auxiliary network parameters, and obtains pre-trained weights. The pre-trained weights of the auxiliary network are transferred to a downstream network. The downstream network is further trained using labeled data, and the auxiliary network weights are loaded into the downstream network. On this basis, the downstream network is further trained in a semi-supervised manner to achieve seismic phase prediction and obtain a predicted profile. The predicted profiles are spliced ​​to obtain a seismic profile. The present invention enables the downstream seismic phase identification network to have a better initial solution by migrating the training weight parameters of the auxiliary network, thereby realizing intelligent prediction of small-sample seismic phases based on the concept of self-supervision.

Description

Technical Field

[0001] The present invention relates to the field of intelligent analysis of earthquake phases, and in particular to an earthquake phase prediction method based on Transformer structure self-supervision. Background Art

[0002] Seismic facies are the sum of sedimentary facies as expressed on a seismic profile. They refer to the distribution of seismic reflection units within a given space. Their distribution within different spatial ranges represents different seismic attributes. Seismic facies identification plays a crucial role in reservoir prediction. It is the primary method for analyzing geological conditions and predicting oil and gas reservoirs, and is also a key factor in the successful development of oil and gas accumulation areas.

[0003] In the early days, seismic phase prediction was typically performed manually, relying on the expertise and comprehensive analytical skills of seismologists to analyze and interpret seismic profiles and slices. However, this approach presented several challenges. First, it was inefficient. The number of seismic interpreters was significantly imbalanced with the vast volume of seismic data. Analyzing regional seismic data required a significant amount of time, and training experienced interpreters required significant resources. Second, it was highly subjective. Different interpreters possessed varying levels of theoretical and practical knowledge, leading to differing perspectives on the interpretation of the same detection data.

[0004] In actual seismic phase identification projects, a large amount of unlabeled data is often available, yet this data is often underutilized. With the continuous development of artificial intelligence and computer application technology, further exploration is needed to combine the advantages of supervised and unsupervised methods, incorporate the concept of self-supervision to fully utilize this unlabeled data, and design auxiliary tasks to mine the data's inherent characteristics as supervisory signals, thereby improving the model's feature extraction capabilities. Summary of the Invention

[0005] To address the above-mentioned problems in the prior art, this application proposes a seismic phase prediction method based on Transformer structure self-supervision, which includes the following steps:

[0006] S1. Preliminary identification of seismic volume data is performed based on well logging information and clustering algorithms. The well-connected profile is selected as unlabeled data and masked to obtain masked unlabeled data.

[0007] S2. Use the auxiliary network to reconstruct a large amount of masked unlabeled data, train the auxiliary network parameters, and obtain pre-trained weights;

[0008] S3, transfer the pre-trained weights of the auxiliary network to the downstream network;

[0009] S4. Further train the downstream network using the labeled data, load the auxiliary network weights obtained in step S2 into the downstream network, and continue to train the downstream network in a semi-supervised manner on this basis to achieve earthquake phase prediction and obtain a predicted profile;

[0010] S5. Splice the predicted sections to obtain the seismic section.

[0011] Preferably, the auxiliary network is used to reconstruct a large amount of masked unlabeled data. Unlabeled samples are input to the auxiliary network and mask reconstruction is performed in a context-based generation manner, so that the network model obtains the initial distribution characteristics of the data set.

[0012] Preferably, step S2 specifically includes:

[0013] S21. Obtain unlabeled samples: Obtain seismic profiles along the inline and crossline directions of the seismic volume data; then use a normalization method to scale the seismic profile data to a predetermined interval, and process the seismic profile data into a single-channel grayscale image format;

[0014] S22. Obtain training data: Randomly mask a portion of the seismic profile. The masked image serves as the network input, while the original seismic profile grayscale image serves as the true result. Specifically, a sub-image of a predetermined size is randomly obtained from the large seismic grayscale image. This sub-image block serves as the label. This sub-image is then randomly masked to remove some information. The masked sub-image block serves as the input to the auxiliary network.

[0015] S23. Training the image reconstruction capability of the network: using the image obtained in step 22 to train the auxiliary network to reconstruct the missing part of the seismic profile, and finally obtaining the weight parameter file of the auxiliary network.

[0016] Preferably, step S4 specifically includes:

[0017] S41. Obtaining data labels: Connect multiple wells on a plane, then vertically downwardly obtain seismic profile data for each well based on the connection position; then use a normalization method to scale the seismic profile data to a predetermined interval, and process the seismic profile data into a single-channel grayscale image; then define data labels based on the boundary lines on the well log;

[0018] S42. Obtain training data: Use a sliding window to crop the grayscale image to obtain a series of sub-images with label information. The labels of the sub-images are the same as the labels of the original grayscale images in which they are located.

[0019] S43, training the seismic phase network recognition model of the downstream task: the seismic phase network recognition model first loads the training weight parameters obtained by the auxiliary task, and then uses the sub-image obtained in step 42 to train the seismic phase network recognition model;

[0020] S44. Earthquake phase prediction: Use the trained downstream network to identify earthquake phase data.

[0021] Preferably, the auxiliary network is an asymmetric network, including an encoding part and a decoding part, both of which are stacks of multiple Transformer layers. The encoding part is composed of multiple Transformer layers and is responsible for the unmasked image part, which is needed in downstream tasks. The decoding part is composed of multiple Transformer layers and is responsible for restoring the masked part of the image. The input of the encoding part is the unmasked image sub-block, and the output of the unmasked image sub-block needs to be unshuffled with the unmasked image part to restore the complete image sub-block sequence, which is finally input to the decoding part.

[0022] Preferably, there is a preprocessing module before the encoding part. In the preprocessing module, the image needs to be divided into multiple non-overlapping small blocks first, and the input size of the image is set. After the block transformation, it is input through the linear mapping head and the input of the encoding part is obtained by adding it to the position embedding of the image.

[0023] Preferably, the encoding part of the downstream network is consistent with the encoding part of the auxiliary network, the encoding input is all sub-blocks of the image, and the output is the output of the 4th, 6th, 8th, and 12th layers in the Transformer.

[0024] Preferably, the decoding part of the downstream network has a total of three steps; the first step: the four outputs of the encoding part are first linearly transformed to change the image size to 16 height × 16 width × the number of channels, and then half of the channels are reduced through 3 × 3 convolution; the second step: the four outputs of the first step are spliced ​​with the next layer from top to bottom, and after splicing, the number of channels is reduced by half through two 3 × 3 convolutions, and finally upsampled by 4 times to the output size of 4 height × 4 width × 4 channels; the third step: the four outputs of the second step are spliced ​​and upsampled by 4 times to restore to the original input size.

[0025] Preferably, the step of using an auxiliary network to reconstruct a large amount of masked unlabeled data includes: obtaining the overall distribution of different seismic phases on the seismic profile based on machine learning clustering and expert experience, and then combining the accurate labels of the well logging data to select structural well-connected profiles with the same or similar geological structures and split them along this direction to obtain seismic profile data.

[0026] Preferably, the normalization method includes: scaling the obtained seismic profile data to a range of 0-255, and processing the seismic profile data into a single-channel grayscale image form; the normalization calculation is as follows:

[0027]

[0028] Where x represents the specific value of a point in the seismic profile, represents the maximum value in the seismic profile, represents the minimum value in the seismic profile, Represents the pixel value of a point in the grayscale image after normalization.

[0029] The above technical features can be combined in various suitable ways or replaced by equivalent technical features, as long as the purpose of the present invention can be achieved.

[0030] The earthquake phase prediction method based on Transformer structure self-supervision provided by the present invention has at least the following beneficial effects compared with the existing technology:

[0031] This method uses expert experience and clustering algorithms to perform a preliminary analysis of seismic data, and combines well logging labels to obtain connected well seismic profiles as unlabeled input. Subsequently, based on the concept of self-supervised learning, a context-generated self-supervised learning approach enables an auxiliary network to reconstruct missing seismic images. By migrating the training weight parameters of the auxiliary network, the downstream seismic phase recognition network has a better initial solution, thereby achieving intelligent prediction of small-sample seismic phases based on self-supervision. Experiments have shown that:

[0032] (1) Under the premise of having a good initial solution, the downstream network can identify the distribution location of the seismic phase in the study area with limited data labels.

[0033] (2) Compared with the initial weights obtained on conventional datasets such as ImageNet, the self-supervised weights obtained from seismic profile reconstruction have better prediction effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The present invention will be described in more detail below based on embodiments and with reference to the accompanying drawings, wherein:

[0035] Figure 1 The seismic phase identification framework diagram based on the self-supervision concept of the present invention is shown;

[0036] Figure 2 The internal structure diagram of the Transformer of the present invention is shown;

[0037] Figure 3 The internal structure diagram of the scaled dot product attention mechanism of the present invention is shown;

[0038] Figure 4 Shows the Q, K, V matrix calculation process diagram of the present invention;

[0039] Figure 5 shows a schematic diagram of the image processing method of the present invention;

[0040] Figure 6 Shows the multi-head attention structure diagram of the present invention;

[0041] Figure 7 Shows the auxiliary network structure diagram of the present invention;

[0042] Figure 8 shows the downstream network structure diagram of the present invention;

[0043] Figure 9 Shows a schematic diagram of seismic well-connecting profile selection according to the present invention;

[0044] Figure 10 The normalized grayscale image of the seismic section of the present invention is shown;

[0045] Figure 11 A schematic diagram showing a random acquisition method of sub-images in a seismic image according to the present invention is shown;

[0046] Figure 12 A schematic diagram of the random masking method of seismic images of the present invention is shown, with the randomly obtained image blocks on the left and the grayscale image after masking on the right;

[0047] Figure 13 shows a schematic diagram of reconstruction of the masked part under different masking ratios of the present invention;

[0048] Figure 14 shows the loss variation graph of the present invention;

[0049] FIG15( a ) shows a cross-sectional reconstruction schematic diagram of the present invention with Inline being 3338;

[0050] FIG15( b ) shows a cross-sectional reconstruction schematic diagram of the present invention with an Inline of 3565;

[0051] FIG15( c ) shows a cross-sectional reconstruction schematic diagram of the present invention with Inline 3706;

[0052] Figure 16(a) shows the prediction results with an Inline of 4100 obtained using the ImageNet pre-trained weights;

[0053] Figure 16(b) shows the prediction results of the self-supervised training weights obtained by the present invention with an Inline of 4100;

[0054] Figure 16(c) shows the prediction results of the ImageNet pre-training weights obtained with an Inline of 4200;

[0055] FIG16( d ) shows the prediction result with an Inline of 4200 obtained by the self-supervised training weights of the present invention. DETAILED DESCRIPTION

[0056] The present invention will be further described below with reference to the accompanying drawings.

[0057] In response to the defects of the above models, the present invention proposes a seismic phase identification method based on the idea of ​​self-supervision. In terms of the model, the present invention adopts the TransFormer architecture and improves the design to form an auxiliary network and a downstream network. The auxiliary network uses a large amount of unlabeled seismic data to perform self-supervised learning to restore the missing parts of the image, and then transfers the parameter weights of the encoded part of the trained network model to the downstream seismic phase identification task. In data processing, for unlabeled data, the seismic data body is preliminarily identified by combining expert experience and machine learning clustering, and the well section is selected as the unlabeled input based on the existing well logging, and a masking operation is performed for auxiliary network training and reconstruction; for labeled data, sub-image training samples are obtained by segmentation through sliding windows.

[0058] like Figure 1 As shown, the present invention provides an earthquake phase prediction method based on Transformer structure self-supervision, comprising the following steps:

[0059] S1. Preliminary identification of seismic volume data is performed based on well logging information and clustering algorithms. The well-connected profile is selected as unlabeled data and masked to obtain masked unlabeled data.

[0060] S2. Use the auxiliary network to reconstruct a large amount of masked unlabeled data, train the auxiliary network parameters, and obtain pre-trained weights;

[0061] S3, transfer the pre-trained weights of the auxiliary network to the downstream network;

[0062] S4. Further train the downstream network using the labeled data, load the auxiliary network weights obtained in step S2 into the downstream network, and continue to train the downstream network in a semi-supervised manner on this basis to achieve earthquake phase prediction and obtain a predicted profile;

[0063] S5. Splice the predicted sections to obtain the seismic section.

[0064] In one embodiment, Figure 2As shown in the figure, the Transformer architecture used in this paper consists of two main components. From bottom to top, the first component is a multi-head attention mechanism, consisting of multiple scaled dot-product attention mechanisms, and the second component is a fully connected feedforward neural network. Each component undergoes a normalization operation beforehand, and the outer layers use residual connections similar to ResNet. The following is a detailed description of the involved structures:

[0065] The scaled dot product attention mechanism is the core part of Transformer. Figure 3 As shown, Self-Attention needs to calculate three matrices Q, K, and V. These matrices are calculated by inputting the matrix X and the linear transformation matrices WQ, WK, and WV. The calculation method is shown in Figure 4 In the field of natural language, each row of X represents the vector of a word, while in computer vision, the image is divided into multiple 16×16 small blocks ( Figure 5 ), a small block is equivalent to a word, that is, a line in X.

[0066] After obtaining the three matrices Q, K, and V, the calculation of the final output Z is shown in formula (1):

[0067] (1)

[0068] In the above formula, in order to prevent The calculated value of is too large, so it is divided by . The resulting matrix is ​​an N×N square matrix, where N is the number of image patches, representing the attention between each patch in the image. Softmax is then used to calculate the attention coefficient between each patch and the other patches, summing each row of the matrix to 1. Finally, the matrix is ​​multiplied by V to produce the output Z.

[0069] Multi-head attention is composed of multiple attentions in the previous paragraph, see Figure 6 The multi-head attention will concatenate multiple outputs Z together, and then linearly transform them through the linear layer to obtain the final output , The same dimension as the input X.

[0070] The fully connected feedforward neural network of the present invention is composed of two fully connected neural networks. The first one has a ReLU activation function at the end, and the second one has no activation function at the end. The calculation is shown in formula (2):

[0071] (2)

[0072] In one embodiment, the auxiliary network is a network model used in the auxiliary task, and its purpose is to perform self-supervised training based on context generation on unlabeled seismic data. The auxiliary network used in the present invention is an asymmetric network, such as Figure 7 As shown, both the encoding and decoding parts are stacked with multiple Transformer layers. The encoding part consists of 12 Transformer layers and is responsible for the unmasked portion of the image, which is required for downstream tasks. The decoding part consists of 4 Transformer layers and is responsible for restoring the masked portion of the image, which is not required for downstream tasks. In addition, the encoding part also needs to go through a preprocessing module. The specific operations of each part are as follows:

[0073] 1. Preprocessing module: In the preprocessing module, the image first needs to be divided into multiple 16×16 non-overlapping small blocks. In the present invention, the image input size is set to 192×192×3, so after the block transformation, the 192×192×3 image will become 12×12×16×16×3 in size. The 16×16×3 sub-block is equivalent to a word in the natural language field. Then, after the linear mapping head input, it will become 12×12×768 in size. Finally, the input of the encoding part is obtained by adding it to the position embedding of the image. Because the Transformer layer needs to use global information, position embedding is required to obtain the relative position relationship of each image sub-block. The calculation of position embedding is shown in Equations (3) and (4):

[0074] (3)

[0075] (4)

[0076] Where pos represents the position of the image sub-block, i represents the dimension, and any Can be express.

[0077] 2. Encoding part: The input of the encoding part is the unmasked image sub-block. The output of this part needs to be unshuffled with the masked image part to restore the complete image sub-block sequence and finally input to the decoding part.

[0078] 3. Decoding part: This part only calculates the loss of the masked image part to restore the original image.

[0079] In one embodiment, the downstream network is a network model used in the downstream seismic phase identification task. Figure 8 As shown in the figure, the encoding part of the downstream network is consistent with the encoding part of the auxiliary network. In the decoding part, in order to compare with the network in the previous section, this part is a pyramid-like upsampling module, which is described in detail as follows:

[0080] 1. Encoding: In the downstream network, seismic images no longer require masking. Therefore, the encoding input is all sub-blocks of the image, and the output is the output of the 4th, 6th, 8th, and 12th layers of the Transformer. In addition, before the downstream network training begins, the encoding needs to load the trained weight parameters of the auxiliary network.

[0081] 2. Decoding part: The decoding part is a pyramid-like structure. Figure 8 As shown in the middle right half, the decoding process consists of three steps. Step 1: The four outputs of the encoding process are first linearly transformed to H / 16×W / 16×C (H and W are the height and width of the original input image, and C is the number of channels). Then, a 3×3 convolution is performed to reduce the number of channels by half. Step 2: The four outputs of the first step are concatenated with the next layer from top to bottom. After concatenation, two 3×3 convolutions are performed to reduce the number of channels by half, and finally the output size is upsampled by 4 to H / 4×W / 4×C / 4. Step 3: The four outputs of the second step are concatenated and upsampled by 4 to restore the original input size.

[0082] In one embodiment, the data processing of the present invention mainly includes two parts: unlabeled data and labeled data. The processing of these two parts of data will be explained in detail below.

[0083] 1. Unlabeled data processing: This part is mainly about the data processing of auxiliary tasks. In the auxiliary tasks, the present invention does not need data containing seismic phase label information, but only seismic profile information data. In the field of geophysical exploration, unlabeled data is very abundant. Among them, in order to make more reasonable use of existing unlabeled data, the constraints of prior knowledge are introduced. The present invention obtains the overall distribution of different seismic phases on the seismic profile based on machine learning clustering and expert experience, and then combines the accurate labels of the well logging data to select structural well sections with the same or similar geological structures and cut them along this direction to obtain seismic profile data. Figure 9 As shown in the figure, for wells initially identified by experts as sharing the same lithology, a well-connected profile can be constructed between them. Wells with the same lithology are more similar, and the resulting seismic profiles have a more regular distribution and spatial continuity, which is more conducive to network model training.

[0084] Next, the normalization method is used to scale the obtained seismic profile data to the range of 0-255, and the seismic profile data is processed into a single-channel grayscale image. The normalization calculation is shown in formula (5), and the obtained grayscale image is as follows: Figure 10 shown.

[0085] (5)

[0086] Where x represents the specific value of a point in the seismic profile, represents the maximum value in the seismic profile, represents the minimum value in the seismic profile, Represents the pixel value of a point in the grayscale image after normalization.

[0087] Next, we need to obtain input images and labels to complete the training. In this experiment, the label of the auxiliary task is the grayscale image itself, and the input is the masked earthquake grayscale image. The specific operation is as follows: Figure 11 As shown in the figure, first, a 192×192 size sub-image (rectangular area) is randomly obtained from the large-scale seismic grayscale image. The sub-image block is the label; then a random masking operation is performed on the sub-image, so that the sub-image loses some information. The masked sub-image block is the input of the auxiliary network ( Figure 12 ).

[0088] 2. Label Data Processing: This section primarily processes data for downstream seismic phase identification tasks, including labeling, sliding window segmentation, and overlapping window prediction. Labeling involves obtaining profile images containing seismic phase information from well logging; sliding window segmentation involves performing overlapping sliding window segmentation on the original labeled data to obtain training samples; and overlapping window prediction involves using sub-image blocks to extract cores and perform overlapping splicing when predicting profiles.

[0089] To validate the effectiveness of the proposed method, an experimental validation study was conducted in the Gaoshiti work area. Located in central Sichuan, the Dengying Formation deposits date back to the formation of the Deyang-Anyue rift, a period that likely occurred during the late Dengying Formation. High-quality reservoirs developed along the platform margins on both sides of the rift. The primary objective of this experiment was to identify high-quality reservoirs within the Dengying Formation. The seismic volume data used in the experiment had an inline range of 3700-4900, a crossline range of 2300-2800, a time depth of 2100-2500 ms, and a sampling interval of 2 ms.

[0090] To demonstrate the effectiveness of the auxiliary network, this invention will be experimentally verified in a real-world work area. The network model used in this experiment is an asymmetric Transformer-based network model. The experimental parameters are: input size of 192×192, random masking, masked image sub-block size of 16×16, loss function of MSE, batch size of 256, optimizer of adamw, and experimental epoch number of 1600. The model's initial learning rate is 3e-4, the learning strategy is warmup and cosine decay, the warmup epoch number is 80, and the total epoch number is 1600. The experimental code uses the frameworks of TensorFlow and Keras, and the training graphics card is a Tesla P100.

[0091] In order to test the reconstruction ability of the auxiliary network, the present invention uses 50%, 75% and 90% masking ratios to conduct experiments. In addition, the present invention also uses the MSE quantitative evaluation index to measure the reconstruction ability of the auxiliary network under different masking ratios. Figure 13 As shown in Table 1, when the masking ratio is smaller, the reconstruction ability of the auxiliary network is stronger, but the training time will also increase dramatically.

[0092] Table 1 Comparison of average MSE and training time under different masking ratios

[0093] Different masking ratios MSE Training time 50% 0.0003 8h27m 75% 0.0008 4h30m 90% 0.0045 3h11m

[0094] Under the premise of selecting 75% masking ratio, the present invention also analyzes the convergence of the auxiliary network loss. Figure 14 As shown in FIG, when the epoch is around 1200, the loss of the auxiliary network reaches the convergence level. Therefore, based on the above experiments, the present invention will select a masking ratio of 75% and a training epoch number of 1200.

[0095] Based on the above parameter settings, we selected several sections to examine the reconstruction capabilities of the auxiliary network. As shown in Figures 15(a)-15(c), we observed that the auxiliary network was able to reliably recover the missing seismic images, with only a small portion of detail not being fully restored. This demonstrates that the auxiliary network possesses preliminary inference capabilities for this dataset. We then migrated the parameters of the auxiliary network to the downstream network to verify the effectiveness of our proposed method.

[0096] In one embodiment, the present invention loads the ImageNet dataset pre-trained weights and the self-supervised pre-trained weights of this dataset in the downstream network respectively and compares the prediction results to verify the superiority of the initial weights provided by the self-supervised learning method of the present invention. The number of original labeled seismic profiles is 30, and the training set and test set are divided according to a ratio of 9:1. The data enhancement method is the sliding window segmentation method in Chapter 3, with a sliding window size of 192×192 and a horizontal and vertical step size of 20 pixels. A total of 1020 samples were obtained in the training set according to the above-mentioned data enhancement method, and the ratio of training and verification sets for all subsequent experiments was set to 9:1. The comparison of the prediction results of the two weights is shown in Table 2. It can be seen that the prediction accuracy of the self-supervised training weights used in the present invention is significantly better than that of the ImageNet pre-trained weights.

[0097] Table 2 Comparison of prediction accuracy and training time of two weights

[0098] Different initial weights Accuracy (MIOU) Training takes time Loading ImageNet pre-trained weights Background: 89.29 Reef slope: 56.75 Reef flat: 63.93 Intra-reef back sand facies and reef flat: 52.82 Total accuracy: 65.7 58m Loading self-supervised training weights Background: 97.06 Reef slope: 69.32 Reef flat: 82.38 Intra-reef back sand facies and reef flat: 88.56 Total accuracy: 84.33 58m

[0099] Next, we will use two profile prediction results to qualitatively analyze the advantages of using self-supervised training weights as initial weights. In the prediction results of this work area, category 0 represents background, category 1 represents reef slope, category 2 represents reef flat, and category 3 represents intra-platform reef back sand facies and reef flat. The prediction results are as follows: Figure 16a -16d. It can be seen that when the ImageNet pre-training weights are used as the initial weights, as shown in Figure 16a As shown by the red solid line, the intra-platform back-reef sand facies and reef flat are clearly separated from other categories, such as Figure 16a As shown by the red dotted line, the reef slope is distributed above the reef back sand facies and reef flat within the platform; Figure 16c As shown in Figure 2, there are obvious misalignments in the prediction results, which are inconsistent with the geological distribution of the Gaoshiti work area. However, when the self-supervised training weights are used as the initial weights, these phenomena are significantly reduced ( Figure 16b , Figure 16d ). Therefore, from the above analysis, it can be seen that the initial weights provided by the self-supervised learning method of the present invention have greater advantages.

[0100] The present invention processes unlabeled data based on expert experience and clustering algorithms, selects seismic profiles, and constructs auxiliary networks and downstream networks in a self-supervisory manner to implement a seismic phase identification method based on the idea of ​​self-supervision.

[0101] Although the present invention is described herein with reference to specific embodiments, it should be understood that these embodiments are merely illustrative of the principles and applications of the invention. It should be understood that many modifications may be made to the illustrative embodiments, and that other arrangements may be devised, without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that the various dependent claims and features described herein may be combined in ways other than those described in the original claims. It should also be understood that features described in conjunction with individual embodiments may be employed in conjunction with other described embodiments.

Claims

1. A self-supervised earthquake phase prediction method based on Transformer structure, characterized in that: The following steps are involved: S1. Preliminary identification of seismic volume data is performed based on well logging information and clustering algorithms. The well-connected profile is selected as unlabeled data and masked to obtain masked unlabeled data. S2. Use the auxiliary network to reconstruct a large amount of masked unlabeled data, train the auxiliary network parameters, and obtain pre-trained weights; The auxiliary network is an asymmetric network, consisting of an encoding part and a decoding part. Both the encoding part and the decoding part are stacked with multiple Transformer layers. The encoding part is composed of multiple Transformer layers and is responsible for the unmasked image part, which is needed for downstream tasks. The decoding part is composed of multiple Transformer layers and is responsible for restoring the masked part of the image. The input of the encoding part is the unmasked image sub-block. The output of this part needs to be unshuffled with the unmasked image part to restore the complete image sub-block sequence, which is finally input to the decoding part. S3, transfer the pre-trained weights of the auxiliary network to the downstream network; S4, further training the downstream network using the labeled data, loading the auxiliary network weights obtained in step S2 into the downstream network, and continuing to train the downstream network in a semi-supervised manner on this basis to achieve earthquake phase prediction and obtain a predicted profile; step S4 specifically includes: S41. Obtaining data labels: Connect multiple wells on a plane, then vertically downwardly obtain seismic profile data for each well based on the connection position; then use a normalization method to scale the seismic profile data to a predetermined interval, and process the seismic profile data into a single-channel grayscale image; then define data labels based on the boundary lines on the well log; S42. Obtain training data: Use a sliding window to crop the grayscale image to obtain a series of sub-images with label information. The labels of the sub-images are the same as the labels of the original grayscale images in which they are located. S43, training the seismic phase network recognition model of the downstream task: the seismic phase network recognition model first loads the training weight parameters obtained by the auxiliary task, and then uses the sub-image obtained in step 42 to train the seismic phase network recognition model; S44. Earthquake phase prediction: Identify earthquake phase data using the trained downstream network; S5. Splice the predicted sections to obtain the seismic section.

2. The earthquake phase prediction method based on Transformer structure self-supervision according to claim 1 is characterized in that: The process of using auxiliary networks to reconstruct a large amount of masked unlabeled data. Through context-based generation, unlabeled samples are input to the auxiliary network and mask reconstruction is performed, so that the network model obtains the initial distribution characteristics of the data set.

3. The earthquake phase prediction method based on Transformer structure self-supervision according to claim 1, characterized in that: Step S2 specifically includes: S21. Obtain unlabeled samples: Obtain seismic profiles along the inline and crossline directions of the seismic volume data; then use a normalization method to scale the seismic profile data to a predetermined interval, and process the seismic profile data into a single-channel grayscale image format; S22. Obtain training data: Randomly mask a portion of the seismic profile, use the masked image as the network input, and use the original seismic profile grayscale image as the true result. Specifically, obtain a sub-image of a predetermined size from the large-scale seismic grayscale image. This sub-image block serves as the label. Then, perform a random masking operation on this sub-image, causing the sub-image to lose some information. The masked sub-image block serves as the input to the auxiliary network. S23. Training the network's image reconstruction capability: using the image obtained in step 22 to train the auxiliary network to reconstruct the missing portion of the seismic profile, and ultimately obtaining a weight parameter file for the auxiliary network.

4. The earthquake phase prediction method based on Transformer structure self-supervision according to claim 1, characterized in that: Before the encoding part, there is also a preprocessing module. In the preprocessing module, the image first needs to be divided into multiple non-overlapping small blocks, the input size of the image is set, and after the block transformation, it is input through the linear mapping head and added to the position embedding of the image to obtain the input of the encoding part.

5. The earthquake phase prediction method based on Transformer structure self-supervision according to claim 1, characterized in that: The encoding part of the downstream network is consistent with the encoding part of the auxiliary network. The encoding input is all sub-blocks of the image, and the output is the output of the 4th, 6th, 8th, and 12th layers in the Transformer.

6. The earthquake phase prediction method based on Transformer structure self-supervision according to claim 1, characterized in that: The decoding part of the downstream network has a total of three steps. The first step: the four outputs of the encoding part are first linearly transformed to the size of 16 height × 16 width × number of channels, and then the number of channels is reduced by half through 3×3 convolution. The second step: the four outputs of the first step are spliced ​​from top to bottom with the next layer. After splicing, the number of channels is reduced by half through two 3×3 convolutions, and finally upsampled by 4 times to the output size of 4 height × 4 width × 4 number of channels. The third step: the four outputs of the second step are spliced ​​and upsampled by 4 times to restore to the original input size.

7. The earthquake phase prediction method based on Transformer structure self-supervision according to claim 1, characterized in that: The steps of using the auxiliary network to reconstruct a large amount of masked unlabeled data include: based on machine learning clustering and expert experience, obtaining the overall distribution of different seismic phases on the seismic profile, and then combining the accurate labels of the well logging data to select structural well sections with the same or similar geological structures and split them along this direction to obtain seismic profile data.

8. The earthquake phase prediction method based on Transformer structure self-supervision according to claim 2, characterized in that: The normalization method includes: scaling the obtained seismic profile data to the range of 0-255, and processing the seismic profile data into a single-channel grayscale image format; the normalization calculation is as follows: Where x represents the specific value of a point in the seismic profile, represents the maximum value in the seismic profile, represents the minimum value in the seismic profile, Represents the pixel value of a point in the grayscale image after normalization.

Images