Speed spectrum intelligent picking method based on improved YOLOv8-pose neural network, medium and equipment

Abstract

The present application relates to a kind of speed spectrum intelligent pickup method based on improved YOLOv8-pose neural network, medium and equipment, belong to geophysical exploration technical field.The method converts traditional velocity analysis problem into image target and key point detection, and the energy group in the speed spectrum is regarded as detection target, and the corresponding pickup point is regarded as key point.YOLOv8-pose network is improved on the basis of maintaining original efficient detection and key point regression capability, first, the SPPF module of main network part is replaced by SPPELAN module;Second, using Pose detection head in detection head part simultaneously outputs the boundary box of energy group and the coordinates of corresponding pickup point.In addition, confidence loss, boundary box regression and key point regression joint loss are used during network training.The present application also proposes the medium and equipment for running the method, and high-precision, high-efficiency speed spectrum pickup can be realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of exploration geophysics, and particularly relates to a velocity spectrum intelligent picking method based on an improved YOLOv8-pose neural network, a medium and equipment. BACKGROUND

[0002] As a key link in the seismic data processing flow, the accuracy of velocity analysis directly determines the reliability of subsequent seismic imaging quality and geological interpretation results. With the continuous progress of seismic exploration technology, seismic data has achieved significant improvement in acquisition quality, spatial resolution and signal-to-noise ratio, and the data scale has also shown exponential growth. The emergence of large-scale and high-precision data has greatly increased the complexity and challenge of each link of seismic data processing and interpretation. The traditional velocity spectrum picking method mainly relies on manual intervention and experience, and has problems such as low efficiency, strong subjectivity and poor repeatability, which has been difficult to meet the requirements of modern seismic exploration for processing automation and refinement. In view of this situation, in recent years, researchers have proposed a variety of new methods for efficient picking, which can be divided into two categories in general: one is based on optimization search strategy, which improves the accuracy and stability of velocity spectrum picking by improving the traditional algorithm model; the other relies on deep learning theory, and uses large-scale data to train automatic model to realize intelligent extraction of velocity spectrum features. Especially the method based on neural network, with its excellent feature expression ability and high automation characteristics, has shown broad application prospects in the field of intelligent seismic data processing, and is expected to become the main direction of the development of future velocity spectrum picking technology.

[0003] The speed spectrum picking method based on the optimization search strategy mainly introduces mathematical optimization technology and search algorithm to efficiently locate the best speed solution in the speed spectrum space, thereby improving the accuracy and calculation efficiency of picking. Typical methods include global search, genetic algorithm, simulated annealing, and particle swarm optimization. This kind of method usually models the speed spectrum picking problem as an optimization problem, taking some form of energy function, cost function or likelihood function as the target, and finding the speed parameter that optimizes the target function through systematic search. The main advantages of the optimization search strategy are: on the one hand, they can effectively avoid falling into local optimum in complex, multi-peak speed spectrum space, improving the globality and stability of the picking result; on the other hand, compared with traditional manual picking method, these algorithms have higher automation and repeatability when dealing with large-scale data, significantly reducing the influence of human factors on the results. However, this kind of method also has certain limitations: first, the calculation cost is relatively high, especially in wide range speed search, optimization algorithm may cause the processing time to increase significantly; second, part of the heuristic algorithm (such as genetic algorithm, particle swarm optimization) is sensitive to the actual application environment in terms of parameter setting and convergence, which needs to be adjusted and optimized for different data types. In addition, the optimization search strategy may still have picking errors or stability problems under low signal-to-noise ratio or high signal complexity, limiting its wide application in extreme complex geological conditions. SUMMARY

[0004] In view of the shortcomings of the prior art, the present application provides a new speed spectrum intelligent picking technology based on an improved YOLOv8-pose neural network. The present application applies the key point detection theory to intelligent speed spectrum picking, further improving the accuracy of intelligent speed spectrum picking.

[0005] The present application is implemented as follows:

[0006] A speed spectrum intelligent picking method based on an improved YOLOv8-pose neural network, the method comprising:

[0007] (1) An improved YOLOv8-pose network model is constructed; the network is composed of a backbone network, a neck network and a detection head, the speed spectrum energy group is regarded as a detection target, and the speed picking point is regarded as the corresponding detection key point, and the SPPF module in the network structure is replaced with the SPPELAN module;

[0008] Further, the backbone network part is responsible for extracting features from the input velocity spectrum image, converting the image into a feature representation with rich semantic information; the neck network part is responsible for fusing and enhancing the features extracted by the backbone network at different levels, providing multi-scale information representation for subsequent prediction; the detection head part uses a Pose detection head to output the bounding box, confidence and corresponding velocity picking point coordinates (key points) of each velocity spectrum energy group in a joint detection form, and through optimizing the bounding box regression and key point regression, the model learns the judgment standard of professional personnel for picking points while positioning the energy group, so as to realize more accurate and reasonable joint prediction of the energy group and the picking point.

[0009] In addition, the SPPF module in the network structure is replaced by the SPPELAN module, which adds multi-scale receptive field design, branch structure and hierarchical aggregation, so that the feature map can capture more rich information at different spatial scales. This design takes into account the network performance and computational efficiency, and shows lower delay in the prediction process of the velocity picking point, and better adapts to velocity spectrum energy groups of various forms.

[0010] (2) Making a velocity spectrum data set;

[0011] Further, the data in the velocity spectrum data set is seismic data with different signal-to-noise ratios. After obtaining the velocity spectrum data in the seismic data, it is converted into an image, the data set converted into an image is preprocessed, and the data set is divided into a training set, a validation set and a test set.

[0012] (3) Network model training and loss function design; the velocity spectrum data set in (2) is used to train the network model constructed in (1) to obtain the best model training weight. After adding key point detection, the loss function in the network model is composed of confidence loss, bounding box loss and key point loss, as shown in formula (1), and the weight of each part needs to be determined through multiple experiments to determine the best weight coefficient.

[0013] ;

[0014] In the formula, , , , total loss, confidence loss, bounding box loss and key point loss, , , corresponding to the weight of the confidence loss, the bounding box loss and the key point loss.

[0015] Further, the specific form of the confidence loss is shown in formula (2):

[0016] ;

[0017] N is the target number matched, ∈{0,1} is the true class label, where 1 represents a target, and 0 represents no target, ∈{0,1} is the model predicted probability.

[0018] Further, the bounding box loss of the network model An IoU type loss is used, as shown in formula (3), which comprehensively considers the area overlap, center point distance and aspect ratio:

[0019] ;

[0020] In the formula is the intersection ratio of the predicted box and the real box; is the Euclidean distance between the centers of the predicted box and the real box; c is the diagonal length of the minimum circumscribed rectangle of the predicted box and the real box; is the weight coefficient; is the aspect ratio consistency index, as shown in formula (4):

[0021] ;

[0022] In the formula is the width of the real box, is the height of the real box, w is the width of the predicted box, and h is the height of the real box.

[0023] Further, the key point loss in the network model L1 Loss is adopted, and the regression error of the key points is directly summed, as shown in formula (5):

[0024] ;

[0025] In the formula, N is the number of energy groups detected; K is the number of key points of each energy group, and the value is 1; , is the real coordinate of the i-th target k-th key point; , is the predicted coordinate of the i-th target k-th key point.

[0026] (4) Intelligent picking of speed spectrum; using the trained network model for intelligent picking.

[0027] The application also provides a computer readable storage medium, the computer readable storage medium stores a computer program, the computer program is suitable for being loaded and executed by a processor, and the speed spectrum intelligent picking method based on the improved YOLOv8-pose neural network is executed.

[0028] The application also provides a computer device, the device comprises a memory and a processor, the memory stores a computer program, and the computer program is executed by the processor to make the processor execute the speed spectrum intelligent picking method based on the improved YOLOv8-pose neural network.

[0029] Compared with the prior art, the application has the following beneficial effects:

[0030] The application proposes an improved YOLOv8-pose network architecture with innovation, replaces the SPPF module in the original network structure with an SPPELAN module, the new module can capture more rich information under different spatial scales, the design takes into account the network performance and computing efficiency, and lower delay is shown in the prediction process of the speed picking point, and the method can better adapt to the detection and identification of speed spectrum energy groups of various morphologies.

[0031] The application innovatively proposes an intelligent speed spectrum picking method based on deep learning, the method applies the key point detection theory to intelligent speed spectrum picking, regards the energy group as a detection target, regards the picking point as a corresponding key point, makes the network learn the judgment basis and rules of the picking point position of professional processors in the actual processing process, and thus realizes more accurate and reasonable joint prediction of the energy group and the picking point.

[0032] The application researches and develops an improved YOLOv8-pose neural network-based speed spectrum intelligent picking method, which can realize high-precision and high-efficiency speed spectrum picking, and the picking accuracy can reach more than 95% compared with manual picking.

[0033] The application has good picking results for seismic data speed spectrum under different geological conditions, the position of the picked energy group and the picking point are high in accuracy, and the efficiency is greatly improved, which provides effective help for speed analysis of large-area and high-density three-dimensional seismic exploration data. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 It is a flowchart of the improved YOLOv8-pose neural network-based speed spectrum intelligent picking method;

[0035] Figure 2 It is a schematic view of the improved YOLOv8-pose network structure;

[0036] Figure 3is a schematic diagram of the SPPF module structure;

[0037] Figure 4 is a schematic diagram of the SPPELAN module structure;

[0038] Figure 5 is a part of the original velocity spectrum diagram showing two work areas; (a) is the velocity spectrum diagram at low coverage times of work area 1, (b) is the velocity spectrum diagram at low coverage times of work area 2, (c) is the velocity spectrum diagram at high coverage times of work area 1, and (d) is the velocity spectrum diagram at high coverage times of work area 2;

[0039] Figure 6 is a schematic diagram of the velocity spectrum label under the key point detection task; (a) is a velocity spectrum diagram; (b) is a local enlarged area diagram in the red dashed box in (a); (c) is a local enlarged area diagram in the red dashed box in (a);

[0040] Figure 7 is a curve diagram of the change of the key point loss in network training;

[0041] Figure 8 is a comparison diagram of the professional and the network model pickup results; (a) is a professional pickup result diagram; (b) is a network model pickup result diagram;

[0042] Figure 9 is a velocity field diagram picked up by a professional;

[0043] Figure 10 is a velocity field diagram picked up by a network model;

[0044] Figure 11 is a difference diagram of the professional pickup and the network model pickup velocity field. DETAILED DESCRIPTION

[0045] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application.

[0046] The application provides a new speed spectrum intelligent picking method based on an improved YOLOv8-pose neural network, which successfully realizes efficient speed spectrum intelligent picking. The method converts the traditional speed analysis work into an image recognition process, applies the key point detection theory to the intelligent speed spectrum picking, regards the speed spectrum energy group as a detection target, and regards the speed picking point as the corresponding key point. First, when making the speed spectrum image label, the accurate position of the speed spectrum energy group is labeled, and the corresponding speed picking point is also labeled. In this way, the network can not only learn the spatial and energy distribution characteristics in the speed spectrum image, but also learn the judgment basis and rules of the picking point position in the actual processing process of professionals, so that more accurate and reasonable joint prediction of the energy group and the picking point is realized. In addition, the SPPF module in the network structure is improved into an SPPELAN module, so that the feature map can capture more rich speed information under different spatial scales, and better adapt to the effective picking of various morphological speed spectrum energy groups. The application realizes efficient picking of multiple actual work area data, and the application will be described in detail below with reference to the drawings.

[0047] As shown in Figure 1 The speed spectrum intelligent picking method based on the improved YOLOv8-pose neural network has the following specific steps:

[0048] Step S10: Constructing an improved YOLOv8-pose network model. Figure 2 The network model constructed by the application is shown, which is composed of a backbone network, a neck network and a detection head. The speed spectrum energy group is regarded as a detection target, and the speed picking point is regarded as the corresponding detection key point. The backbone network part is responsible for extracting features from the input speed spectrum image, and converting the image into a feature representation with rich semantic information; the neck network part is responsible for fusing the features extracted by the backbone network; the detection head part uses a Pose detection head, and outputs the boundary box position, confidence and corresponding picking point coordinates of the energy group after prediction. The network can not only learn the spatial and energy distribution characteristics in the speed spectrum image in the training process, but also learn the judgment and selection criteria of the picking point position in the actual processing process of professionals, so that more accurate and reasonable joint prediction of the energy group and the picking point is realized. In addition, the SPPF module in the network structure is replaced by an SPPELAN module, Figure 3 、 Figure 4The specific architecture of the SPFF module and the SPPELAN module is respectively shown. The SPPELAN module adds a multi-scale receptive field design, a branch structure and hierarchical aggregation, so that the feature map can capture more rich information at different spatial scales. The design takes into account the network performance and computing efficiency, and shows lower delay in the speed picking point prediction process, and better adapts to various forms of speed spectrum energy group;

[0049] Step S20: Making a speed spectrum dataset. Two groups of seismic data with different signal-to-noise ratios from Jiangsu oilfield in eastern China are selected for processing, and the speed spectrum data are converted into images for dataset making, Figure 5 Part of the speed spectrum images of two work areas are shown. The speed spectrum of work area 1 has a medium-high signal-to-noise ratio feature, the shallow-middle layer energy group focuses well, the event continuity is good, but there is obvious multiple wave interference; the speed spectrum energy group focusing degree of the seismic data of work area 2 is poor due to the influence of complex fault zone structure, and the event continuity is obviously reduced. The dataset is preprocessed using preprocessing methods including bad channel editing and cutting, noise suppression, amplitude compensation, and a label dataset is made according to the picking results of professional processors;

[0050] Figure 6 The difference between the traditional target detection task and the key point detection task label of the application is shown. Wherein, (a) is the original speed spectrum without picking; (b) is the local enlarged area graph in the red dashed line box in the figure, which is the label format of the traditional target detection, and the boundary box of the energy group is labeled, only the position of the energy group is labeled, and the required picking point cannot be directly obtained; (c) is the local enlarged area graph in the red dashed line box in the figure (a), which is the key point positioning label format, directly marking the "time-velocity" key points corresponding to the effective energy group, so that the network can learn the spatial position relationship of the picking points; and the dataset is divided into a training set, a validation set and a test set according to a ratio of 6:2:2;

[0051] Step S30: Network model training and loss function design. The network model is trained using the speed spectrum dataset, and the best model training weight is obtained by adjusting the network training hyperparameters for multiple times.

[0052] The training environment is:

[0053] (1) Processor: 13th Gen Intel(R) Core(TM) i9-13700KF;

[0054] (2) Memory: 256GB;

[0055] (3) System type: 64-bit operating system;

[0056] (4) Graphics card: NVIDIA GeForce RTX4090, 24GB;

[0057] (5) Platform: PyTorch neural network framework based on PyCharm compiler.

[0058] Based on the above training environment, the improved YOLOv8-pose network constructed in step S10 of the embodiment is trained using the training set. After adding key point detection, the loss function of this network is composed of confidence loss, bounding box loss and key point loss, as shown in formula (1). The weights of each part need to be determined through multiple experiments to determine the best weight coefficient.

[0059] ;

[0060] In the formula , , , are the total loss, confidence loss, bounding box loss and key point loss, respectively. , , correspond to the weights of the classification loss, bounding box loss and key point loss, respectively.

[0061] The specific form of the confidence loss is shown in formula (2):

[0062] ;

[0063] In the formula, N is the number of matched targets, ∈{0,1} is the true class label (1 represents a target, and 0 represents no target), ∈{0,1} is the probability predicted by the model.

[0064] The bounding box loss of the network model uses an IoU type loss, as shown in formula (3), which considers the area overlap, center point distance and aspect ratio:

[0065] ;

[0066] In the formula is the intersection ratio of the predicted box and the true box; is the Euclidean distance between the centers of the predicted box and the true box; c is the diagonal length of the minimum bounding rectangle of the predicted box and the true box; is the weight coefficient; is the aspect ratio consistency index, as shown in formula (4):

[0067] ;

[0068] In the formula The width of the actual bounding box. is the height of the ground truth bounding box, w is the width of the predicted bounding box, and h is the height of the ground truth bounding box.

[0069] The network's critical point loss Using L1 Loss, the regression errors of the key points are directly summed, as shown in formula (5):

[0070] ;

[0071] In the formula, N is the number of detected energy clusters; K is the number of key points for each energy cluster, which is 1 in this invention; ( , ) is the true coordinate of the k-th key point of the i-th target; , () are the predicted key point coordinates.

[0072] After multiple trials, the network hyperparameters during training were finally set as follows: 400 training epochs with an early stopping strategy; batch size of 32; and learning rate of 0.01.

[0073] Keypoint training loss curve as shown Figure 7 As shown in the figure, this graph illustrates the trend of the loss function as the model progresses through training epochs. Specifically, the loss value rapidly decreases from approximately 0.16 to approximately 0.04 within the first 20 epochs, indicating that the model quickly and effectively fits the training data in the initial stage. After the 20th epoch, the loss value remains relatively stable within a small fluctuation range between 0.03 and 0.04, demonstrating the stability and good convergence of the training process.

[0074] Step S40: Intelligent velocity spectrum picking and picking accuracy analysis. The trained network model is used for intelligent picking, and the picking accuracy is analyzed, along with generalization experiments.

[0075] Figure 8 (a) shows the picking results from professional processors, while (b) shows the intelligent picking results from the improved YOLOv8-pose neural network proposed in this invention. A comparison of the two shows that the results of energy cluster location identification and corresponding velocity curve interpolation are highly consistent with those of manual picking. Furthermore, thanks to the algorithm's high efficiency, the overall processing efficiency of automated picking is significantly improved compared to manual methods, substantially enhancing the automation level and timeliness of data processing, thus verifying the superior performance of the proposed method in terms of accuracy and efficiency.

[0076] Figure 9To show the velocity field generated by the picking result of professional processing personnel, Figure 10 To show the velocity field corresponding to the picking result of the YOLOv8-pose network model proposed in the present study, Figure 11 To show the velocity field corresponding to the picking result of the YOLOv8-pose network model proposed in the present study, Figure 9 , Figure 10 The difference between the two figures is the velocity field. The target horizon is mainly distributed between 1.5s and 4s. From the comparison results, it can be seen that the velocity values and overall velocity change trend obtained by the method in the shallow layer, middle layer and target layer are highly consistent with the artificial picking results, and the overall picking accuracy can reach more than 95%, indicating that the proposed model has good accuracy and reliability in the intelligent picking task of velocity spectrum. The data of two work areas are used in the present embodiment, and the method has good performance in different work areas, which can indicate the effectiveness of the method in different geological environments.

[0077] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any modification, equivalent replacement and improvement made by those skilled in the art within the technical range disclosed in the present application, as long as it is within the spirit and principles of the present application, should be covered in the protection scope of the present application.

Claims

1. A velocity spectrum intelligent picking method based on an improved YOLOv8-pose neural network, characterized in that, The method includes: (1) Construct an improved YOLOv8-pose network model; the network consists of a backbone network, a neck network and a detection head. The velocity spectrum energy cluster is regarded as the detection target, the velocity pickup point is regarded as its corresponding detection key point, and the SPPF module in the network structure is replaced with the SPPELAN module. (2) Create a velocity spectrum dataset; (3) Network model training and loss function design; The network model constructed in (1) is trained using the velocity spectrum dataset in (2) to obtain the optimal model training weights; The loss function in the network model consists of confidence loss, bounding box loss and key point loss; Keypoint loss in the network model Using L1 Loss, the regression errors of the key points are directly summed, as shown in formula (5): ； In the formula, N is the number of detected energy clusters; K is the number of keypoints for each energy cluster, with a value of 1; , ) is the true coordinate of the k-th key point of the i-th target; , () represents the coordinates of the k-th predicted keypoint of the i-th target; (4) Intelligent picking of velocity spectrum; intelligent picking is performed using the trained network model.

2. The method according to claim 1, characterized in that, The backbone network is responsible for extracting features from the input velocity spectrum image and transforming the image into a feature representation with rich semantic information; the neck network is responsible for fusing and enhancing the features extracted from different levels by the backbone network; the detection head uses the Pose detection head to output the bounding box, confidence score and corresponding velocity pickup point coordinates, i.e., key points, for each velocity spectrum energy cluster in a joint detection form.

3. The method according to claim 1, characterized in that, The data in the velocity spectrum dataset consists of seismic data with different signal-to-noise ratios. After obtaining the velocity spectrum data from the seismic data, it is converted into images. The converted image dataset is preprocessed and divided into training set, validation set and test set.

4. The method according to claim 1, characterized in that, The optimal weighting coefficients for the confidence loss, bounding box loss, and keypoint loss in (3) were determined through multiple experiments: ； In the formula , , , These are the total loss, confidence loss, bounding box loss, and keypoint loss, respectively. , , These correspond to the weights of confidence loss, bounding box loss, and keypoint loss, respectively.

5. The method according to claim 4, characterized in that, Confidence loss The specific form is shown in formula (2): ； In the formula, N is the number of matched targets. ∈{0,1} are the true category labels, where 1 represents a target and 0 represents no target. ∈{0,1} is the probability predicted by the model.

6. The method according to claim 5, characterized in that, Boundary loss of network model The loss function used is an IoU-type loss, as shown in formula (3), which comprehensively considers area overlap, center point distance, and aspect ratio: ； In the formula This represents the interaction ratio between the predicted bounding box and the ground truth bounding box. is the Euclidean distance between the centers of the predicted bounding box and the ground truth bounding box; c is the diagonal length of the minimum bounding rectangle between the predicted bounding box and the ground truth bounding box. These are the weighting coefficients; The aspect ratio consistency index is shown in formula (4): ； In the formula The width of the actual bounding box. is the height of the ground truth bounding box, w is the width of the predicted bounding box, and h is the height of the ground truth bounding box.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that is adapted to be loaded by a processor and executed by the velocity spectrum intelligent picking method based on the improved YOLOv8-pose neural network according to any one of claims 1-6.

8. A computer device, characterized in that, The device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the velocity spectrum intelligent picking method based on the improved YOLOv8-pose neural network according to any one of claims 1-6.

Images