Geological iterative matching technology based on machine learning assistance
By extracting geological image features using CNN and combining them with SVM for data matching and classification, the geological model is optimized, solving the data integration problem in geological exploration and achieving efficient and accurate visualization and interpretation of exploration results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- YANCHENG INST OF TECH
- Filing Date
- 2023-11-21
- Publication Date
- 2026-04-17
AI Technical Summary
Existing technologies struggle to effectively integrate and analyze geological exploration data, especially in complex geological scenarios, leading to low exploration efficiency and insufficient accuracy.
Convolutional Neural Networks (CNNs) are used to extract geological image features, combined with Support Vector Machines (SVMs) for data matching and classification, gradient descent is used to optimize the model, and the results are presented using neural network visualization algorithms to achieve iterative geological matching.
It improves the efficiency and accuracy of geological data analysis, enables accurate identification of geological patterns in complex geological scenarios, reduces human intervention, and enhances the predictive performance and visualization capabilities of exploration models.
Smart Images
Figure CN121880946A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to applications in geological exploration, oil and coal mine detection, and machine learning. It is applicable to various geological iterations and the exploration of underground oil resources, and is particularly suitable for deep-earth engineering data collection below 5,000 meters. Specifically, it is a machine learning-assisted geological iteration matching technology, which has a significant cost-performance advantage in large-area and geologically complex scenarios. Background Technology
[0002] In recent years, with the development of computer science and geological exploration, a large amount of data has emerged in fields such as geophysics and geological exploration technology. This data includes seismic data, geophysical exploration data, and rock chemical analysis results, providing more information to reveal geological iterations. However, the resolution and spatial distribution of this data are uneven, making the effective integration and analysis of this data a major challenge for geological research. Geological iteration matching technology stems from the geological need to understand geological iterations and internal structures, as well as the continuous development of computer learning technology. By applying geological iteration matching technology, we can better understand the Earth's evolutionary processes and resource distribution, providing more information and methods for research and applications in geology and petroleum exploration.
[0003] Geological iterative techniques are based on machine learning, which enables automatic learning from data and experience, allowing performance to improve automatically without explicit programming. A model is built based on the relationship between input features and corresponding labels, and then used to predict new unlabeled data. Large datasets are divided into training and test sets. The training set is used for algorithm learning and parameter tuning, while the test set is used to evaluate the model's performance and generalization ability. Training the model using the training set involves matching input features with corresponding labels or target values to adjust the model's parameters or weights, enabling it to accurately predict targets.
[0004] Seismic data is frequently used in oil and gas exploration to detect potential oil and gas reservoirs. Neural network visualization algorithms can analyze seismic data to help identify seismic facies, structural interpretations, and sedimentary characteristics, thereby determining potential oil and gas exploration targets. These algorithms can also help identify potential oil and gas reservoirs, including lithological boundaries, variations in oil saturation, and fracture systems. Based on the visualization results from neural networks, explorers can develop more accurate exploration strategies and select areas most likely to contain oil and gas for further exploration. Furthermore, neural network visualization algorithms can monitor and analyze geological and production data, aiding in emergency response and risk assessment. Real-time data visualization allows for the timely detection and response to potential geological hazards, environmental problems, or equipment malfunctions, reducing safety risks and production losses.
[0005] Support Vector Machines (SVMs) can be used for target identification and classification in oil and gas exploration. By training the model and utilizing the classification capabilities of SVMs, the ability of explorers to identify subsurface structures and targets can be improved. SVMs can learn feature weights from data, helping to determine the importance of key features in oil and gas exploration. This is very helpful for optimizing exploration models, reducing redundant information, and improving prediction performance. The feature selection capabilities of SVMs can improve the efficiency of data analysis and model interpretation. SVMs can be applied to seismic inversion, enabling the inversion and prediction of subsurface attributes by learning the relationship between seismic data and subsurface models. This helps to more accurately understand subsurface oil and gas reservoir conditions. SVMs can be used to build predictive models in oil and gas exploration. By training the SVM model and utilizing its learning capabilities, prediction accuracy can be improved, providing a reliable basis for decision-making. SVM matching algorithms are well-suited for multi-classification problems in complex geological and geomorphological applications, accurately achieving intelligent matching in geological exploration. Summary of the Invention
[0006] The purpose of this invention is to provide a machine learning-assisted geological iterative matching technology, which is particularly suitable for various geological iterations and the exploration of underground oil resources.
[0007] This invention is achieved through the following technical solution:
[0008] This paper presents a machine learning-assisted iterative geological matching technique, characterized by the use of convolutional neural networks to capture feature data extraction from geological images. The extracted data is then cleaned and standardized through normalization to ensure data quality and consistency, thereby improving the accuracy of subsequent machine learning models. A vector machine (SVM) learning algorithm is used to establish a correlation between geological data and the model to match the geological model with actual data. Gradient descent is used for iterative optimization, gradually improving the model's accuracy and performance. Finally, a neural network visualization algorithm is used to present the results of the geological data and model in an intuitive and understandable way. This iterative geological matching technique, based on a combination of machine learning-assisted algorithms, helps geological experts better understand geological data, identify geological patterns, and optimize geological models to match actual data.
[0009] The convolutional neural network (CNN) used in geological image acquisition first learns feature patterns from the images to provide automated interpretation and classification. By training the model, it automatically identifies and locates these targets, reducing manual intervention and improving detection accuracy and efficiency. Then, it learns feature representations of the geological images, identifying key features such as texture, structure, and lithology to help optimize the performance of exploration and prediction models. Next, it learns spatial and contextual information from the geological images to recover missing data or remove noise, improving image quality and enabling applications in geological image reconstruction tasks. This supports exploration personnel in better utilizing geological image data and making accurate decisions.
[0010] The method utilizes singular value decomposition and curvelet transform to separate and detect feature vectors of weak signals over a wide range. Machine learning algorithms are used to learn features and classify and predict three geological patterns: overlying strata, reservoir, and underlying strata. Then, the feature vectors of the overlying strata, reservoir, and underlying strata are used as inputs to train an SVM classifier. The trained SVM classifier is then used to classify geological patterns, and the three geological patterns of overlying strata, reservoir, and underlying strata are matched using a geological model with varying weak signal amplitudes.
[0011] The key feature is the intuitive presentation of geological data using a neural visualization algorithm. By visualizing the feature maps of the intermediate or convolutional layers of the network, abstract features extracted from geological images at different levels can be observed. This helps understand how the network learns geological features and patterns. Drawing network structure diagrams provides an intuitive understanding of the network's hierarchical organization and information transmission paths, thus revealing the network's complexity and function. The network's prediction results are visualized, explaining the basis of the predictions. Geological attributes or subsurface structures predicted by the network are displayed visually, providing an intuitive understanding of the network's predictive power and results. Finally, the uncertainty of the model is estimated, as geological data often contains noise and uncertainty. By visualizing the model's uncertainty, a more comprehensive evaluation and interpretation of the prediction results can be achieved.
[0012] Compared with the prior art, the present invention has the following advantages:
[0013] (1) This invention is based on the application of Convolutional Neural Networks (CNNs) for geological image data acquisition. CNNs can automatically learn and extract features from images without the need for manual feature extraction. Through multi-layer convolution and pooling operations, CNNs extract highly abstract and representative features from the original images. By employing local connectivity and weight sharing, they gain the ability to perceive local regions of the image. Convolutional neural networks have a certain degree of scale invariance, enabling them to obtain useful features for images of different scales. Through large-scale training and parameter optimization, they can still accurately process and classify images even in complex situations, exhibiting a certain degree of robustness and generalization ability. By learning the structured information, hierarchical structure, and multi-layer convolution operations of the image, higher-level structural information in the image is gradually combined and learned, thereby better understanding the semantic content of the image.
[0014] (2) This invention utilizes the Support Vector Machine (SVM) algorithm for geological iteration identification and classification. SVM can effectively process data in high-dimensional feature spaces. It maps low-dimensional data to a high-dimensional feature space, thereby better processing and separating complex geological features. By mapping data to a nonlinear feature space and constructing a nonlinear decision boundary for matching, it better adapts to the complexity of geological data. By introducing regularization parameters, it effectively controls the complexity of the model, thereby reducing the risk of overfitting. SVM possesses good mathematical properties and stability. In geological matching, stability is crucial for ensuring the reliability of the results. This ensures consistent performance under different datasets and parameter selections. Attached Figure Description
[0015] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0016] Figure 1 Machine learning-assisted geological matching flowchart
[0017] Figure 2 Schematic diagram of a convolutional neural network model
[0018] Figure 3 Data Iterative Optimization Process Diagram
[0019] Figure 4 SVM Overall Flowchart
[0020] Figure 5 Visualization of the deconvolution process Detailed Implementation
[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments. The accompanying drawings are provided to provide a further understanding of the present invention and constitute a part of this application, but do not constitute a limitation on the present invention.
[0022] Machine learning-assisted geological iterative matching technology takes into account the incompleteness of the measured data during oil exploration, the limited amount of data, and the inconsistency of different data sources. This can cause important geological features to be easily overlooked, and the subsequent data processing is greatly affected by complex geological conditions, resulting in misjudgment or misinterpretation of the acquired data.
[0023] First, a convolutional neural network is used to extract relevant features from the geological data. For example... Figure 2 The model shown helps identify geological features and patterns. The convolutional layer, a feature extraction layer, is used to extract features from geological data and is the core of the convolutional neural network. The characteristics of local connectivity and weight sharing in convolutional neural networks are fully reflected in the computation process of the convolutional layer. Given input geological image data, the convolution operation obtains target features by performing local multiplication and accumulation on the image through a sliding convolution kernel. The calculation process of the convolutional layer output value is as follows:
[0024]
[0025] The pixel value at each corresponding position can be obtained:
[0026] Where xi,j represents the element in the i-th row and j-th column of the feature image; ω m，n ω represents the weight of the m-th row and n-th column of the feature image; b a represents the bias term of the convolution kernel; i，j represents the element in the i-th row and j-th column of the feature image; f represents the activation function.
[0027] The width and height of the final feature map satisfy the following relationship:
[0028] W2=(W1-F+2P) / S+1
[0029] H2=(H1-F+2P) / S+1
[0030] Where W2 and W1 represent the widths of the feature images before and after convolution, respectively; F represents the width of the convolution kernel; P represents the number; S represents the amplitude; and H2 and H1 represent the heights of the feature images before and after convolution, respectively.
[0031] After convolution, a nonlinear transformation is typically applied to the geological feature map, introducing an activation function. The activation function introduces nonlinearity into the geological model, improving its expressive power and representational ability. The Sigmoid function, with an output range of 0 to 1, can be used in models where the output is a predicted probability. The formula for the Sigmoid function is:
[0032]
[0033] Compared to the Sigmoid function, the ReLU function is faster to compute and does not suffer from gradient saturation when the input is positive, making it a commonly used activation function in geomatching techniques. The formula for the ReLU function is:
[0034]
[0035] Pooling layers are used for geological data compression, reducing data dimensionality and effectively decreasing the number of model parameters, thus mitigating overfitting during geotechnical data training. They are also used to extract more abstract features. A typical pooling operation is max pooling, which selects the maximum value from a specific region as the pooled value. The expression for max pooling is:
[0036] f(x) = max(x) [i，i+k][j，j+k] )
[0037] The fully connected layer follows the single convolutional layer block. The output of the single convolutional layer block serves as the input to the fully connected layer. It first flattens the input mini-batch of samples, then performs data classification to complete the classification task. The Softmax function expresses the results of multi-class classification in probabilities, and its formula is:
[0038]
[0039] In this system, the numerator is an exponential function, which, based on its properties, maps the real-valued outputs of the geological model to a range from zero to positive infinity. The denominator sums all results, normalizes the data, and transforms multi-class outputs into probabilistic outputs. This effectively captures local features, textures, shapes, and other information from geological images. Figure 3 The stacking of multiple convolutional and pooling layers shown allows the network to gradually learn higher-level geological features, thereby achieving more accurate geological image analysis and classification.
[0040] Feature extraction and selection were performed on the collected geological data to obtain feature vectors for three geological models: overlying strata, reservoir, and underlying strata. Feature learning and classification prediction were then performed using a convolutional neural network. An SVM algorithm was selected to build the classifier, and optimization methods were used to improve the classifier's performance and generalization ability. Figure 4As shown, the feature vectors of the overlying strata, reservoir, and underlying strata are used as input to an SVM classifier to train the classifier for classification and prediction of the input data. The trained SVM classifier is then used to classify and predict the acquired geological data, categorizing it into three geological models: overlying strata, reservoir, and underlying strata. A geological model based on the amplitude variation of fiber optic signals is used to match these three geological models. This geological model searches for reflected waves from the acquired seismic data and extracts their amplitudes. The amplitude variation of the extracted reflected waves is related to the properties of the layer in which they reside. The degree of matching between the amplitude variation of the reflected waves and the geological model is calculated, thus matching the geological model.
[0041] Finally, a neural network visualization algorithm is used to present the geological data and model results in an intuitive and understandable way. First, the feature maps obtained from each layer are used as input, and deconvolution and unpooling are performed to obtain the deconvolution results, which are then used to verify and display the feature maps extracted from each layer. The entire process is illustrated in the diagram below. Figure 5 As shown, the entire process of the network begins with normal image prediction from the top left. Then, the feature map is processed in reverse: unpooling, activation functions, and deconvolution. A deconvolutional network is used to complete the mapping. A deconvolutional network can be viewed as a convolutional model, using the same components but in reverse. Therefore, it maps features to pixels using gradient ascent, adjusting the input data of selected layers to maximize feature activation values. The geological data and model results are presented in an intuitive and understandable way.
Claims
1. A machine learning-assisted geological iterative matching technique, characterized by: Feature data extraction from geological images is captured using convolutional neural networks. The extracted data is then cleaned and standardized through normalization to ensure quality and consistency, improving the accuracy of subsequent machine learning models. A vector machine (SVM) learning algorithm is used to establish a correlation between geological data and the model to match the geological model with actual data. Gradient descent is used for iterative optimization, gradually improving the model's accuracy and performance. Finally, a neural network visualization algorithm is used to present the results of the geological data and model in an intuitive and understandable way. This geological iterative matching technique, based on a combination of machine learning-assisted algorithms, helps geological experts better understand geological data, identify geological patterns, and optimize geological models to match actual data.
2. As described in claim 1, characterized in that Convolutional neural networks (CNNs) are used to capture features from geological images. First, the geological image is preprocessed to reduce interference and standardize it. Local geological features are extracted by sliding convolutional kernels across the image, outputting specimen data. Fully connected layers are used to process two-dimensional tensors, establishing a relationship between features and the target output. Overfitting is avoided by applying layers, and a loss function and optimizer (stochastic gradient descent) are selected during training to minimize discrepancies and approximate true values. This allows for the extraction of features at various levels from geological images, aiding in tasks such as geological image classification, segmentation, and recognition.
3. As described in claim 1, characterized in that The Super Vector Machine (SVM) learning algorithm establishes the correlation between geological data and the model. First, the feature vectors of the overlying strata, reservoir, and underlying strata are used as input to the SVM classifier, and the classifier is trained. The SVM separates data samples of different classes by finding an optimal hyperplane. Using the training dataset, the SVM model is trained to learn the best decision boundary and separating hyperplane. Finally, the performance of the trained SVM model is evaluated using the test dataset. The trained SVM classifier is then used for geological pattern classification, matching the input overlying strata, reservoir, and underlying strata patterns using a geological model with weak signal amplitude variations.
4. As described in claim 1, characterized in that This paper utilizes neural network visualization algorithms to present geological data and model results in an intuitive and understandable way. First, the feature maps obtained from each layer are used as input, and deconvolution and unpooling are performed to obtain the deconvolution results, which are then used to verify and display the feature maps extracted from each layer. Gradient ascent technique is used to adjust the input data of selected strata to maximize the activation values of features. Heatmaps are used to visualize the feature maps of specific layers, representing the importance of features as color intensity in the image. This process presents the geological data and model results in an intuitive and understandable manner.