Seismic horizon tracing method based on deep convolutional neural network and random forest

Abstract

This invention discloses a seismic horizon tracking method based on deep convolutional neural networks and random forests, belonging to the field of automatic horizon tracking technology for seismic data in oil and gas exploration. The seismic horizon tracking method based on deep convolutional neural networks and random forests includes the following steps: Step 1, seismic data enhancement processing; Step 2, horizon sample label construction and optimization, establishing a sample library; Step 3, deep convolutional neural network training and parameter optimization to obtain high-level features of seismic data for target geological horizons; Step 4, based on the high-level features trained by the deep convolutional neural network and the professional feature vector dataset of seismic waveforms, adjusting the output of the deep convolutional network, and connecting it to a random forest classifier for training to construct a classification tree; Step 5, automatic identification of seismic reflection co-phase axes, identifying and automatically tracking co-phase axes of the same horizon based on the results of the classification tree. This seismic horizon tracking method based on deep convolutional neural networks and random forests not only further improves the accuracy of automatic tracking of seismic reflection co-phase axes but also greatly improves the efficiency of horizon tracking, showing great application and promotion prospects in the automatic tracking of geological target horizons in the field of oil and gas exploration.

Description

Technical Field

[0001] The present invention relates to a method for automatically tracking seismic reflection events of a target geological layer in the field of oil and gas exploration, and specifically provides a seismic layer tracking method based on a deep convolutional neural network and a random forest. Background Art

[0002] As geological targets for oil and gas exploration become increasingly complex, the requirements for exploration work become more sophisticated, and the workload and difficulty of interpretation for researchers have greatly increased. Artificial intelligence technologies with core algorithms such as deep learning have brought revolutionary breakthroughs in many fields such as face and voice recognition, and have also received widespread attention in the field of oil and gas exploration as one of the most promising technologies. However, the problems faced by oil and gas exploration are more complex, and intelligent recognition technology is still in the stage of theoretical research and model testing.

[0003] Stratum identification and interpretation are the basis of oil and gas exploration work such as sedimentary phase analysis and reservoir prediction. The method usually adopted by manual event tracking is to identify the extreme points or zero-crossing points of the seismic event in the seismic waveform based on the dynamic and kinematic characteristics of seismic waves. The problem it faces is low efficiency, unavoidable subjective influence, and poor verifiability. Later, new automatic stratigraphic tracking technologies such as waveform feature tracking, correlation tracking, and ant tracking, which find similar seed points in adjacent channels based on the "seed point" characteristics, have emerged. Their advantages are directionality and high efficiency, but the problem is that the existing technology can only meet the requirements of structural interpretation when the stratigraphic structure is relatively simple, and its accuracy cannot meet the exploration needs when the geological structure is complex.

[0004] From the perspective of seismic layer interpretation needs, seismic data, as a time series signal, has great similarities with speech recognition. Seismic layer tracing can essentially be mathematically defined and classified, converting the complex layer tracing problem into a seismic waveform unit classification problem. Therefore, big data technology with deep learning methods as the core can provide a more effective tool for automatic layer tracing. This technology has stronger nonlinear representation capabilities and can generate intelligent models with end-to-end functions, which can greatly reduce the workload and ensure the accuracy of interpretation.

[0005] Among the various deep learning network models, the deep convolutional neural network (DCNN) has a strong advantage in extracting spatial structural features because it can learn higher-level abstract features. It is a model with the best application effect in high-level feature extraction. Random forest, as an integrated learning algorithm based on decision trees, balances errors through multiple numbers to produce a high-accuracy classifier. It is one of the algorithms with better performance in classification algorithms. Therefore, this patent chooses the most basic layer coaxial automatic tracking problem in oil and gas exploration as the application scenario, integrates the respective technical advantages of deep convolutional neural networks and random forest algorithms, reshapes the professional process of seismic layer interpretation, uses deep convolutional neural networks as seismic waveform high-level feature extractors, and random forests as classification identifiers, realizing an integrated method based on deep convolutional neural networks and random forests to solve the problem of automatic tracking of seismic reflection wave coaxials in the field of oil and gas exploration, greatly improving the work efficiency and accuracy of geological target layer tracking, and realizing the technical application of deep learning in the field of oil and gas exploration. Summary of the invention

[0006] The purpose of the present invention is to provide a seismic layer tracking method based on deep convolutional neural network and random forest to solve the current seismic layer tracking accuracy and efficiency problems. The method uses deep convolutional neural network as a seismic reflection wave high-level feature extractor of the geological target layer, multi-dimensionally characterizes the high-level features of the seismic waveform unit, expands the seismic waveform professional feature vector data set, inherits the advantages of the traditional layer waveform tracking technology based on the professional features of the seismic waveform, and then uses random forest to classify the seismic data of the geological target layer, further improves the automatic tracking accuracy of the seismic iso-axis, and also greatly improves the efficiency of layer tracking.

[0007] To achieve the above-mentioned purpose, the present invention provides the following technical scheme, and the seismic layer tracking method based on deep convolutional neural network and random forest comprises the following steps: step 1, seismic data enhancement processing; step 2, layer sample label construction and optimization, and establishment of sample library; step 3, deep convolutional neural network training and parameter optimization, and acquisition of high-level features of seismic data of target geological layer; step 4, based on the high-level features of seismic waveforms trained by deep convolutional neural network and the professional feature vector data set of seismic waveforms, the output of deep convolutional network is adjusted, connected to random forest classifier training, and a classification tree is constructed; step 5, seismic reflection isotropic axis is automatically identified, and the isotropic axis of the same layer is identified and automatically tracked according to the results of the classification tree.

[0008] Furthermore, in step 1, seismic data enhancement processing generally adopts conventional filtering, frequency extension and other means to effectively enhance the energy of weak reflection waves and improve the lateral continuity of seismic waveforms, providing a high signal-to-noise ratio and high-resolution seismic data volume for further construction and optimization of layer sample labels. The purpose of seismic data enhancement processing is to prepare data input for obtaining key sample labels of seismic layer data and training target layer high-level features.

[0009] Furthermore, in step 2, the layer sample labels are constructed and optimized to establish a sample library; the top and bottom layer times are selected as the upper and lower limits, which contain multiple seismic waveforms in the layer segment to be identified, and the sparse grid line layer seismic layers based on expert experience interpretation are used as sample labels, and different layers in the layer segment are assigned different numerical labels;

[0010] Furthermore, in step 3, a deep convolutional neural network is trained, and the seismic waveform data in the layer segment is trained using the layer sample labels constructed based on expert experience. The last layer uses a fully connected network, and the output is a high-level feature vector set of the seismic waveform of the target layer. The Center Loss loss function based on cosine distance is used as the error loss function, and the parameters of each layer in the network are optimized through back propagation and stochastic gradient descent algorithms, and the deep grid is trained until the classification accuracy requirements are met. The purpose of network training is to extract high-level features of seismic waveforms through the deep convolutional neural network recognition process.

[0011] A deep convolutional neural network includes an input layer, a convolutional layer, a linear rectifier function (Relu) activation layer, a pooling layer, a fully connected layer, and an output layer. The input layer is a seismic reflection waveform unit within a certain time window above and below the target layer; the alternating appearance of the convolutional layer, the linear rectifier function (Relu) activation layer, and the pooling layer realizes the automatic extraction of high-level feature information of the seismic waveform; the fully connected layer realizes the output of high-level features of the seismic waveform; the fully connected layer realizes the output of high-level features of the seismic waveform, and the fully connected layer is placed after the last pooling layer. The size, number, and step length of the convolutional filter in the convolutional layer are set respectively; the pooling area and the pooling step length in the pooling layer are set; the nodes of the fully connected layer are set, and the number of nodes in the output layer is determined according to the high-level features of the target layer to be identified, that is, according to the experience or feature data of the actual situation of the application scenario.

[0012] The calculation of the loss function includes: calculating the cosine distance between each sample and the corresponding class center as the error loss function, that is, Center Loss based on cosine distance, and calculating the partial derivatives of the loss function with respect to the central features of each class and the partial derivatives of the parameters in the network, which mainly considers the spatial anisotropy information of the input seismic waveform unit, and is used to optimize the parameters in the network during the back propagation process.

[0013] Further, in step 4, based on the high-level feature vector data set trained by the deep convolutional neural network and the professional features of the seismic waveform of the target layer, adjustments are made at the output end of the deep convolutional network, and the high-level feature results of the seismic waveform obtained by the last fully connected activation layer of the deep convolutional network and the professional features of the seismic waveform of the target layer are used as the input of the random forest classifier and connected to the random forest classifier. Among them, the high-level features of the seismic data of the target layer are obtained by training the deep convolutional neural network for each new input of the seismic data of the layer, and the professional features of the seismic waveform of the target layer are extracted by conventional methods, including the maximum peak time, waveform category, wave length, maximum peak value, maximum trough value, number of peaks, number of troughs, peak length, trough length, inclination and other features.

[0014] In addition, the random forest classifier is trained and several classification trees are constructed using the Gini index as the criterion. When constructing each classification tree, samples are extracted with replacement from the high-level features of each layer of seismic waves and the professional feature vector data set as the training set. The classification tree is split using the feature with the smallest Gini index until the Gini index is less than the threshold. Several classification trees are established to form a random forest.

[0015] Furthermore, in step 5, the seismic layer automatic tracking system module is used to input the high-level features and professional features of the acquired seismic data of each layer into the trained random forest, and each classification tree in the random forest generates a classification result based on the high-level features and professional features of the seismic data of the target layer, and the seismic reflection phase axis tracking result of the entire target layer is obtained based on the classification result.

[0016] Compared with the prior art, the seismic layer automatic tracking method based on deep neural network and random forest of the present invention has the following outstanding beneficial effects: (a) seismic layer tracking method based on deep convolutional neural network and random forest, deep neural network can extract more discriminative high-level features, simple, efficient, and robust, even in the area with complex seismic reflection phase axis, the high-level feature extraction effect is also good; (b) the seismic waveform professional feature vector data set is expanded, inheriting the advantages of traditional layer waveform tracking technology based on seismic waveform professional features; (c) using random forest as a classifier, giving full play to the advantages of ensemble learning, fast classification of high-dimensional features, high accuracy and generalization ability. At the same time, when each tree of the random forest selects the features to be used, they are randomly generated from all high-level features and professional features, reducing the risk of overfitting. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a technical roadmap for automatic tracking of seismic horizons based on deep convolutional neural networks and random forests.

[0018] Figure 2This is a flowchart of the automatic tracking technology of seismic layers based on deep convolutional neural networks and random forests.

[0019] Figure 3 This is a schematic diagram of the section construction of the layer sample label in this case.

[0020] Figure 4 This is a schematic diagram of the construction plane of the layer sample labels in this case.

[0021] Figure 5 This is a schematic diagram of the automatic layer tracking section in this case.

[0022] Figure 6 This is a schematic diagram of the automatic layer tracking plan in this case. DETAILED DESCRIPTION

[0023] The present invention will be further explained below in conjunction with specific implementation schemes. The detailed description and technical contents of the present invention are described as follows in conjunction with the accompanying drawings. However, the drawings are only provided for reference and illustration and are not intended to limit the present invention.

[0024] Figure 1 This is the technical roadmap for automatic tracking of seismic horizons based on deep convolutional neural networks and random forests in the present invention, such as Figure 1 As shown, the method comprises the following steps:

[0025] The first step is to enhance the processing of seismic data, which effectively enhances the energy of weak reflection waves and improves the lateral continuity of seismic waveforms, providing a high signal-to-noise ratio and high-resolution seismic data volume for the construction and optimization of further layer sample labels. In this case, a three-dimensional seismic data volume in the east was used as an example, and the seismic data volume was subjected to frequency extension processing, and the lateral continuity of the processed seismic waveform was enhanced.

[0026] The second step is to construct and optimize the layer sample labels and establish a sample library. First, we carried out manual tracking of the target layer based on expert experience. The three-dimensional seismic data selected in this case contains 160 lines, 160 traces per line, and 25m trace spacing. We selected a target test layer, manually tracked the target layer based on a 10×10 grid, and obtained the coarse grid layer interpretation results. On this basis, we constructed the target layer seismic waveform label, and selected the top and bottom layer times as the upper and lower limits, which included multiple seismic waveforms in the layer segment to be identified, and different layers in the layer segment were assigned different numerical labels. Figure 3 This is a schematic diagram of the section construction of the layer sample labels in this case. Figure 4 This is a schematic diagram of the plane construction of the layer sample labels in this case.

[0027] The third step is to train the deep convolutional neural network and optimize the parameters to obtain the high-level features of the seismic reflection waves of the target geological layer. The training samples are the characteristic values ​​of each seismic reflection waveform explained based on expert experience. For each seismic waveform feature of the layer in the training sample, a deep convolutional neural model is built separately. Considering the scale of the input data, the number of internal parameters of the deep convolutional neural network is minimized to avoid serious overfitting. The deep convolutional neural network includes an input layer, a convolutional layer, a fully connected layer, and a loss function layer. Each convolutional layer contains convolution, activation, and pooling operations, and the fully connected layer is placed after the last pooling layer. In this case, during the training process of the deep convolutional neural network, Figure 2 Taking the deep convolutional neural network with three convolutional layers as an example, the input of the deep convolutional neural network is single-channel seismic waveform data with 1 channel, and each channel includes 128 data points. The size of the convolutional filter is 3×1, and the number of convolutional filters in each convolutional layer is 32, 64, and 32 respectively. The convolution step is 1. In the pooling layer, the activation function is Relu, the pooling area is 3×1, and the pooling step is 1. The maximum pooling is used, and the nodes of the fully connected layer are all 64. The output of the last convolutional layer is combined into a one-dimensional vector and then input into the fully connected layer.

[0028] The fourth step is to adjust the output of the deep convolutional network based on the high-level features and professional feature vector data sets of the stratigraphic seismic waveforms trained by the deep convolutional neural network, connect to the random forest classifier training, and build a classification tree. Use the Gini index as the criterion to build k classification trees. When building each classification tree, n samples are extracted with replacement from the high-level features and professional feature vector data sets based on the deep convolutional neural network training as the training set. Use the feature with the smallest Gini index to split the binary tree, and repeat this process until the Gini index is less than a certain threshold. The standard for finding the best classification features and thresholds is: find the best features and thresholds so that the Gini value of the current node minus the Gini value of the left child node and the Gini value of the right child node are the largest. In this case, 0.45 is selected as the judgment criterion based on the test results.

[0029] The fifth step is to automatically identify the seismic reflection event axis. According to the results of the classification tree, the event axis of the same layer is identified and automatically tracked. The seismic reflection waveforms of the layers to be detected and identified in the work area are automatically identified, classified and tracked. Figure 5 This is a schematic diagram of the automatic tracing profile of the layers in this case. Figure 6 This is a schematic diagram of the automatic layer tracking in this case. From the automatic identification classification and tracking results, the automatic layer interpretation (lower tracking line) effectively avoids the tracking method based only on the single attribute of the amplitude value (upper tracking line), fully utilizes the high-level characteristics of the seismic waveform, effectively avoids the string axis phenomenon, and greatly reduces the automatic tracking time, which greatly improves the accuracy and efficiency of seismic interpretation.

[0030] The present invention can be applied to structural interpretation in the field of oil and gas exploration. The method integrates the respective technical advantages of deep convolutional neural networks and random forest algorithms, reshapes the professional process of seismic structural interpretation, uses deep convolutional neural networks as high-level feature extractors and random forests as classification identifiers, and realizes an integrated method based on deep convolutional neural networks and random forests to solve the problem of automatic identification and tracking of seismic reflection wave event axes in the field of oil and gas exploration, greatly improving the work efficiency and accuracy of geological target layer tracking, and realizing the technical application of deep learning in the field of oil and gas exploration.

[0031] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A seismic horizon tracking method based on deep convolutional neural network and random forest, characterized by: The method comprises the following steps: Step 1, seismic data enhancement processing; Step 2: construct and optimize layer sample labels and establish a sample library; Step 3: deep convolutional neural network training and parameter optimization to obtain high-level characteristics of seismic reflection waves of target geological layers; Step 4: Based on the high-level features of the deep convolutional neural network training and the seismic waveform professional feature vector dataset, the deep convolutional network output is adjusted, connected to the random forest classifier training, and a classification tree is constructed; Step 5: Automatically identify the seismic reflection event axis, and identify the event axis in the same layer according to the results of the classification tree and automatically track it; In step 1, seismic data enhancement processing is performed to obtain the reflection characteristics of the reflection waves of the geological target layer after the seismic data is input; seismic data enhancement processing generally adopts conventional filtering, frequency extension and other means, and its purpose is to effectively enhance the energy of weak reflection waves and improve the lateral continuity of seismic waveforms, and provide high signal-to-noise ratio and high-resolution seismic data bodies for further construction and optimization of layer sample labels.

2. The seismic horizon tracing method based on deep convolutional neural network and random forest according to claim 1, characterized in that: In step 2, the layer sample labels are constructed and optimized to establish a sample library; the top and bottom layer times are selected as the upper and lower limits, which contain multiple seismic waveforms in the layer segment to be identified, and the coarse grid line layer seismic layers based on expert experience interpretation are used as sample labels, and different layers in the layer segment are assigned different numerical labels.

3. The seismic horizon tracing method based on deep convolutional neural network and random forest according to claim 1, characterized in that: In step 3, the deep convolutional neural network is trained, and the seismic waveform data in the layer segment is trained using the constructed layer sample labels. The last layer adopts a fully connected network, and the output is a high-level feature vector set of the seismic waveform of the target layer. The Center Loss loss function based on cosine distance is selected as the error loss function, and the parameters of each layer in the network are optimized through back propagation and stochastic gradient descent algorithms. The deep network is trained until the classification accuracy requirements are met. The purpose of deep convolutional network training is to extract high-level features of seismic waveforms.

4. The seismic horizon tracing method based on deep convolutional neural network and random forest according to claim 1, characterized in that: In step 4, the output layer of the last layer of the deep convolutional network structure is replaced with a random forest classifier, and the high-level features of the seismic waveform obtained by the last fully connected activation layer of the deep convolutional network and the professional feature vector data set of the seismic waveform of the target layer are used as the input of the random forest classifier, and connected to the random forest classifier training to build a classification tree; wherein, the high-level features of the seismic data of the target layer are obtained by training the convolutional neural network after each new seismic data is input, and the professional features of the seismic waveform of the target layer are mainly extracted by conventional methods, including maximum peak time, waveform category, wave length, maximum peak value, maximum trough value, number of peaks, number of troughs, peak length, trough length, inclination and other features.

5. The seismic horizon tracing method based on deep convolutional neural network and random forest according to claim 3, characterized in that: A deep convolutional neural network includes an input layer, a convolutional layer, a linear rectifier function (Relu) activation layer, a pooling layer, a fully connected layer and an output layer; the input layer is a seismic reflection waveform unit within a certain time window above and below the target layer; the pooling layer is set between adjacent convolutional layers, and the alternating appearance of the convolutional layer and the pooling layer realizes the automatic extraction of high-level feature information of the seismic waveform; the fully connected layer realizes the output of high-level features of the seismic waveform; the size, number and step size of the convolutional filter in the convolutional layer are set respectively; the pooling area and the pooling step size in the pooling layer are set; the nodes of the fully connected layer are set, and the number of nodes in the output layer is determined according to the high-level features of the seismic waveform of the identified layer, that is, according to the experience or feature data of the actual situation of the application scenario.

6. The seismic horizon tracing method based on deep convolutional neural network and random forest according to claim 4, characterized in that: The random forest classifier is trained and several classification trees are constructed using the Gini index as the criterion. When constructing each classification tree, samples are extracted with replacement from the high-level features of each layer seismic waveform and the professional feature vector set as the training set. The classification tree is split using the feature with the smallest Gini index until the Gini index is less than the threshold. Several classification trees are established to form a random forest.

7. The seismic horizon tracing method based on deep convolutional neural network and random forest according to claim 5, characterized in that: The module of automatic seismic layer tracking system is used to input the high-level features and professional features of the seismic waveform of each layer into the trained random forest. Each classification tree in the random forest generates a classification result according to the high-level features and professional features of the seismic waveform of the target layer, and the seismic reflection event axis tracking result of the entire target layer is obtained according to the classification result.

Images