Method and system for generating geological interpretation conclusions from acoustic remote detection imaging logging

Abstract

The present invention belongs to the technical field of petroleum exploration and development, and relates to a method and system for generating geological interpretation conclusions for acoustic wave long-range detection imaging logging. By mining the relationship between a label matrix and generated text, the label matrix features are optimized. Feature fusion is simultaneously adopted to achieve the intersection of image and text features, strengthen the fusion of visual semantic information, improve the accuracy of text description of images, quickly generate geological interpretation conclusions for acoustic wave long-range detection imaging logging, reduce the diversity of conclusions, improve the interpretation compliance rate, and enhance the interpretation and analysis capabilities of non-acoustic logging professionals for acoustic wave long-range detection imaging.

Description

Technical Field

[0001] The present invention belongs to the field of geophysical well logging technology, relates to acoustic wave remote detection imaging logging technology, and specifically relates to a method and system for generating geological interpretation conclusions of acoustic wave remote detection imaging logging. Background Art

[0002] As a powerful exploration tool that can see far through the well, acoustic remote sensing imaging logging is widely used in the detection and discovery of concealed oil and gas reservoirs, fractures and caves near wells, and other small geological anomalies. It has demonstrated unique application value in deep-sea tight oil and gas, shale oil and gas, and fracture-cavity oil and gas reservoirs. However, the current challenge of geological interpretation of acoustic remote sensing imaging is the multi-solution nature of the technology, especially the need for professional personnel to interpret and analyze acoustic remote sensing imaging, which limits the deepening and promotion of the application of this technology. To this end, people have compared the results of acoustic remote sensing imaging with typical imaging in the constructed library by constructing a "forward modeling-imaging-interpretation conclusion" library to achieve geological interpretation of acoustic remote sensing imaging logging. However, there are still technical obstacles such as the cumbersome and time-consuming interpretation process and the low interpretation consistency rate.

[0003] In recent years, the rapid development of artificial intelligence has provided a new research approach for image feature recognition. By using neural networks to learn the feature relationships between images and text reports, relatively accurate textual interpretations of given images can be automatically generated. This approach provides assistance to non-specialists and reduces the complexity of interpretation conclusions. This technology has demonstrated its technical superiority in generating diagnostic reports for medical image reports. Text generation methods generally consist of a convolutional neural network and a natural language processing network. First, a convolutional neural network (CNN) or its variants is used to extract features from the input image. CNNs excel at capturing low-level local features such as edges and textures, and gradually extract higher-level semantic information. To better capture complex geological features in images, existing methods also incorporate multi-scale feature extraction modules to ensure a balanced processing of detailed and global information. For text generation, early natural language processing networks typically used LSTMs, BiLSTMs, GRUs, and other techniques, and extracted features through an attention mechanism. With the development of the Transformer, the Transformer and related variants have been used as text feature extraction models and report generation architectures. This basic structure is often more suitable for traditional natural image short sentence description tasks. However, in acoustic wave long-range detection imaging data, normal samples (i.e., samples exhibiting normal geological characteristics) account for a larger proportion of the total samples, while the number of abnormal samples (i.e., samples reflecting abnormal geological characteristics) is relatively small. This data imbalance problem causes the model to focus more on conventional geological structure characteristics and ignore abnormal features such as cracks and caves. As a result, the generated interpretation conclusions are more biased towards the description of non-feature areas and fail to achieve geological interpretation conclusions that are consistent with reality. Summary of the Invention

[0004] In response to the problems existing in the prior art, the present invention provides a method and system for generating geological interpretation conclusions of acoustic wave long-range detection imaging well logging, which mines the relationship between the label matrix and the generated text, optimizes the features of the label matrix, and simultaneously adopts feature fusion to realize the intersection of image and text features, strengthens the fusion of visual semantic information, improves the accuracy of the text description of the image, quickly generates geological interpretation conclusions of acoustic wave long-range detection imaging well logging, reduces the diversity of conclusions, improves the interpretation compliance rate, and thereby improves the interpretation and analysis capabilities of non-acoustic logging professionals for acoustic wave long-range detection imaging.

[0005] In a first aspect, the present invention provides a method for generating geological interpretation conclusions of acoustic wave remote detection imaging logging, the steps of which are as follows:

[0006] Data acquisition step: obtaining acoustic wave remote detection imaging images and text sequences;

[0007] Image feature extraction step: extract local features of the acoustic wave remote detection imaging image, and perform average pooling on the local features to extract global features;

[0008] Label classification step: perform label classification on local features to obtain label vectors;

[0009] One-time encoding step: Encode the global features to obtain the hidden vector;

[0010] Label feature fusion step: construct a label matrix, fuse the label matrix with the hidden vector and label vector to generate label matrix features;

[0011] Text indexing step: index the text sequence through word vectors to obtain text features;

[0012] Vector feature fusion step: Fuse the text features with the hidden vector to generate a fusion vector, and add the text features and the fusion vector to obtain the text vector;

[0013] Explanation conclusion generation step: Decode the label matrix features and text vectors to generate a report and obtain the explanation conclusion;

[0014] Secondary encoding step: Encode text features to obtain enhanced text features;

[0015] Text feature extraction step: Pool the enhanced text features to obtain the global text feature vector;

[0016] Loss function design steps: The similarity between the generated report and the target report is maximized by minimizing the cross entropy loss function as the text loss function, the similarity between the label vector and the true vector is used as the label loss function, and the cosine similarity between the label matrix features and the global text feature vector is used as the text-label loss function;

[0017] Optimization step: The interpretation conclusion is optimized by minimizing the total loss function formed by the text loss function, label loss function and text-label loss function to generate the final interpretation conclusion.

[0018] In some embodiments, in the image feature extraction step, the acoustic wave remote detection imaging image is input into the image feature extractor, the features output by the last convolutional layer of the image feature extractor are extracted as local features, the local features are input into the classifier for average pooling, and all features of each row are spliced ​​into a sequence as global features.

[0019] In some embodiments, in the label classification step, the local features are input into a classifier for classification to generate a label vector.

[0020] In some embodiments, in the label feature fusion step, a specific method for fusing the label matrix with the hidden vector and the label vector to generate label matrix features is:

[0021] Calculate the similarity between the label matrix and the hidden vector Expressed as:

[0022]

[0023] Where, is the label matrix, each row vector of the label matrix , indicating the category The text feature vector of is the number of labels, is the hidden vector, is the hidden vector The transpose of is the dimension of the label matrix;

[0024] Based on similarity Calculate the attention score with the label vector Expressed as:

[0025]

[0026] Where, is the normalized exponential function, is the label vector;

[0027] Leveraging Attention Score , normalization and residual connection fusion label features to obtain the label matrix feature representation as follows:

[0028]

[0029]

[0030] Where, is the label matrix feature, is the normalization function, is the input variable, is the mean, is the standard deviation.

[0031] In some embodiments, in the text feature extraction step, a method for pooling the enhanced text features to obtain a global text feature vector is: pooling the enhanced text features and taking the first vector of the enhanced text features as the global text feature vector.

[0032] In some embodiments, in the loss function design step, the minimized text loss function is expressed as:

[0033]

[0034]

[0035] Where, To minimize the text loss function, The number of words to generate a report for, For the target report generated The probability distribution of words, for Target text at the moment, for The text of the moment, is the label matrix, This is an image of acoustic wave long-range detection imaging.

[0036] In some embodiments, in the loss function design step, the label loss function is expressed as:

[0037]

[0038] Where, is the label loss function, Predicted for the image Category labels, For the real The label of the category, is the number of labels.

[0039] In some embodiments, in the loss function design step, the text-label loss function is:

[0040]

[0041]

[0042] Where, is the text-label loss function, is the global text feature vector, is the label matrix feature, For the real The label of the category, , indicating the category The text feature vector of is the dimension of the label matrix.

[0043] In some embodiments, during the optimization step, the total loss function is expressed as:

[0044]

[0045] Where, is the total loss function, To minimize the text loss function, is the label loss function, is the text-label loss function, is the weighting factor of the label loss function, is the weighting factor of the text-label loss function.

[0046] In a second aspect, the present invention provides a system for generating geological interpretation conclusions from acoustic wave remote detection imaging logging, which is used to implement the method for generating geological interpretation conclusions from acoustic wave remote detection imaging logging described in the first aspect of the present invention, comprising:

[0047] A data acquisition module, used to acquire acoustic wave remote detection imaging images and text sequences;

[0048] Image feature extractor, extracting local features of acoustic wave remote detection imaging images;

[0049] The classifier performs average pooling on local features to extract global features and classifies local features to obtain label vectors;

[0050] Transformer neural network, including:

[0051] The encoder encodes the global features to obtain a hidden vector, indexes the text sequence through the word vector to obtain text features, encodes the text features to obtain enhanced text features, and pools the enhanced text features to obtain a global text feature vector;

[0052] Decoder, built-in:

[0053] The label attention module constructs a label matrix and fuses the label matrix with the hidden vector and label vector to generate label matrix features;

[0054] The feature fusion module fuses the text features with the hidden vector to generate a fusion vector, and adds the text features and the fusion vector to obtain the text vector;

[0055] The interpretation conclusion generation module decodes the label matrix features and text vectors to generate a report and obtain the interpretation conclusion;

[0056] The loss function design module uses the similarity between the generated report and the target report as the text loss function, the similarity between the label vector and the true vector as the label loss function, and the cosine similarity between the label matrix features and the global text feature vector as the text-label loss function;

[0057] The optimization module optimizes the explanation conclusion by minimizing the total loss function formed by the text loss function, label loss function and text-label loss function to generate the final explanation conclusion.

[0058] Compared with the prior art, the advantages and positive effects of the present invention are:

[0059] (1) The method and system for generating geological interpretation conclusions of acoustic wave long-range imaging logging provided by the present invention optimize the features of the label matrix by mining the relationship between the label matrix and the generated text, and at the same time adopt feature fusion to realize the intersection of image and text features, strengthen the fusion of visual semantic information, and improve the accuracy of text description of images. Through the encoding, fusion and optimization mechanism of comprehensive image features and text features, the geological interpretation conclusions of acoustic wave long-range imaging logging are automatically and quickly generated, the diversity of conclusions is reduced, the interpretation compliance rate is improved, and the interpretation and analysis ability of non-acoustic logging professionals on acoustic wave long-range imaging is enhanced.

[0060] (2) The method and system for generating geological interpretation conclusions of acoustic wave remote detection imaging logging provided by the present invention optimize the label matrix by designing a text-label loss function. By constraining feature consistency and accuracy, the deviation in label matrix learning is reduced, thereby further improving the accuracy of text description of images. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 This is a flow chart of the method for generating geological interpretation conclusions of acoustic wave remote detection imaging logging according to an embodiment of the present invention;

[0062] Figure 2 This is a structural block diagram of the system for generating geological interpretation conclusions of acoustic wave remote detection imaging logging according to an embodiment of the present invention;

[0063] Figure 3 This is a schematic diagram of the structure of the classifier according to an embodiment of the present invention;

[0064] Figure 4 This is a schematic diagram of the structure of the label attention module according to an embodiment of the present invention;

[0065] Figure 5This is a schematic diagram of the structure of the feature fusion module according to an embodiment of the present invention;

[0066] Figure 6 Schematic diagram of the loss value change curve of the model described in an embodiment of the present invention on the Cave dataset;

[0067] Figure 7 Schematic diagram of the loss value change curve of the model according to an embodiment of the present invention on the Crack dataset;

[0068] Figure 8 This is a boxplot of the model described in an embodiment of the present invention on the Cave dataset;

[0069] Figure 9 This is a boxplot of the model described in an embodiment of the present invention on the Crack dataset;

[0070] Figure 10 Schematic diagram of two dataset images and interpretation conclusions according to an embodiment of the present invention;

[0071] Figure 11 This is a comparison chart of the qualitative experimental results described in the embodiments of the present invention;

[0072] Figure 12 is the weighting factor of the label loss function described in the embodiment of the present invention Sensitivity diagram of ;

[0073] Figure 13 The weighting factor of the text-label loss function described in the embodiment of the present invention is Sensitivity diagram of ;

[0074] Figure 14 This is a diagram showing the results of acoustic wave remote detection processing for the full-depth MX-1 well in the Tarim Oilfield according to an embodiment of the present invention;

[0075] Figure 15 This is a simulation diagram of the acid fracturing fracture of the MX-1 well in an embodiment of the present invention;

[0076] Figure 16 This is a graph showing the acid fracturing construction of the MX-1 well in an embodiment of the present invention.

[0077] In the figure, 1. Data acquisition module, 2. Image feature extractor, 3. Classifier, 4. Transformer neural network, 41. Encoder, 42. Decoder, 421. Label attention module, 422. Feature fusion module, 423. Explanation conclusion generation module, 424. Loss function design module, 425. Optimization module. DETAILED DESCRIPTION

[0078] The present invention will be described in detail below with reference to exemplary embodiments with reference to the accompanying drawings. However, it should be understood that elements, structures, and features in one embodiment may also be beneficially combined in other embodiments without further description.

[0079] Acoustic remote sensing imaging logging is a key technology used in oil and gas field development and geological structure analysis. It generates images and related geological interpretation information from acoustic detection data. However, existing geological interpretation methods often rely on manual experience or simple statistical models, making it difficult to fully utilize the complex information in the imaging data, resulting in low accuracy and consistency in the interpretation results. The present invention provides a method and system for generating geological interpretation conclusions from acoustic remote sensing imaging logging. By extracting local and global features of acoustic remote sensing imaging images, performing label classification and feature encoding, and combining them with text features indexed from text sequences, a feature fusion strategy and optimization design are used to generate high-quality geological interpretation reports. The core of this method is the use of deep learning methods (such as label classification and encoding-decoding architectures) and loss function-based optimization design to ensure that the generated interpretation conclusions are close to the actual target in terms of content. The following describes the aforementioned method and system for generating geological interpretation conclusions from acoustic remote sensing imaging logging in detail, with reference to the accompanying figures.

[0080] See also Figure 1 The first embodiment of the present invention provides a method for generating geological interpretation conclusions of acoustic wave remote detection imaging logging, the steps of which are as follows:

[0081] S1. Data acquisition step: Acquire acoustic wave remote detection imaging images and text sequences.

[0082] Specifically, the text sequence is , , express The text of the moment, Represents the vocabulary consisting of all characters in the text sequence, Indicates the number of characters in the text sequence.

[0083] S2, image feature extraction step: extract local features of the acoustic wave remote detection imaging image, and perform average pooling on the local features to extract global features.

[0084] Specifically, in some embodiments, the acoustic wave remote detection imaging image is input into an image feature extractor (e.g., a deep convolutional neural network) to extract the features output by the last convolutional layer of the image feature extractor. As local features, 、 and are the height, width and channels of the image respectively.

[0085] Input the local features into the classifier for average pooling, and concatenate all the features of each row into a sequence as the global feature ,in, , the process is as follows:

[0086]

[0087] in, Represented as an image feature extractor, For acoustic wave remote detection imaging images, is the number of image blocks, It is Image block features.

[0088] Average pooling is used to obtain global features and improve the stability and expressiveness of features.

[0089] S3, label classification step: perform label classification on the local features to obtain label vectors.

[0090] Specifically, in some embodiments, the local features Input into the classifier (e.g., multi-classification neural network) to classify the input acoustic wave remote detection image.

[0091] Specifically, in some embodiments, the Figure 3 The classifier structure shown performs label classification on local features. The classifier receives the local features extracted in step S1. , local features are transformed through pooling layer and convolution layer The image is converted into a 1×1 dimension image with 2048 features, and then passes through a fully connected layer with 1024 neurons. To reduce fitting, the dropout technique is applied after the fully connected layer. Finally, it passes through an output layer with 14 neurons, each neuron corresponds to a classification label. The output layer uses Activation function, get the final classification result, generate the label vector of the image , is the number of image labels. The process is as follows:

[0092]

[0093] in, and It is the trainable weight of the fully connected layer in the classifier, which is used to perform linear transformation on the pooled image features and gradually extract deeper feature information. Represents local features An averaging operation is performed to reduce the dimension of image features while retaining the main feature information. It is an activation function, which represents the rectified linear unit and introduces nonlinear features. It is the activation function of the output layer, which obtains the probability of each category.

[0094] S4, one-time encoding step: Encode the global features to obtain the hidden vector.

[0095] Specifically, in some embodiments, a Transformer neural network is used to encode the global features to obtain a hidden vector, and the process is shown in the following formula:

[0096]

[0097] Where, is the hidden vector, is a standard Transformer neural network encoder, is the number of image blocks, It is Image block features, It is the enhanced Image block features.

[0098] S5. Label feature fusion step: construct a label matrix, and fuse the label matrix with the hidden vector and label vector to generate label matrix features.

[0099] In order to solve the data imbalance problem, a label matrix is ​​introduced to enhance text features. This can be achieved through the label attention module of the Transformer neural network. The structure of the label attention module can be found in Figure 4 .

[0100] Specifically, the constructed label matrix is ​​expressed as , each row vector of the label matrix , indicating the category The text feature vector of is the number of labels. It should be noted that the label matrix is ​​a trainable label matrix, and the text information of different labels is learned during the training process. The label matrix is ​​randomly set.

[0101] Specifically, the specific method of fusing the label matrix with the hidden vector and the label vector to generate the label matrix features is:

[0102] Calculate the similarity between the label matrix and the hidden vector Expressed as:

[0103]

[0104] Where, is the hidden vector The transpose of is the dimension of the label matrix;

[0105] Based on similarity Calculate the attention score with the label vector Expressed as:

[0106]

[0107] Where, is the normalized exponential function, is the label vector;

[0108] Leveraging Attention Score , normalization and residual connection fusion label features to obtain the label matrix feature representation as follows:

[0109]

[0110]

[0111] Where, is the label matrix feature, is the normalization function, is the input variable, is the mean, is the standard deviation.

[0112] Using the label matrix , hidden vector and label vector Jointly calculate attention scores and with the label matrix Weighted calculation, build feature mapping for text features of the same label, and suppress text features of other labels.

[0113] S6, text indexing step: index the text sequence through word vectors to obtain text features.

[0114] Specifically, in some embodiments, a Transformer neural network encoder is used as a text feature extractor to extract text sequences. Perform word vector indexing to obtain text features , the process is shown as follows:

[0115]

[0116] in, Represents the word vector indexing process.

[0117] S7, vector feature fusion step: perform feature fusion on the text feature and the hidden vector to generate a fusion vector, and add the text feature and the fusion vector to obtain the text vector.

[0118] Specifically, in some embodiments, a Transformer neural network decoder is used for feature fusion, and a feature fusion module is introduced into the Transformer neural network decoder, such as Figure 5 As shown in the figure, the feature fusion module fuses the text features with the hidden vector to generate a fusion vector. The fusion vector is then added to the text features through the residual layer MLP, which serves as the final output of the encoding stage and is input to the Transformer neural network decoder. The feature fusion method effectively achieves information complementarity between different features and enhances the richness of feature representation. The specific process is as follows:

[0119]

[0120]

[0121] Where, for All text features before time, is a hidden vector, representing the enhanced image features, is the fusion vector, is the number of Chinese characters input to the decoder, is a text vector.

[0122] S8, explanation conclusion generation step: decode the label matrix features and text vectors to generate a report and obtain the explanation conclusion.

[0123] S9, secondary encoding step: encode the text features to obtain enhanced text features.

[0124] Specifically, the Transformer neural network encoder is used as a text feature extractor to encode text features to obtain enhanced text features. The process is as follows:

[0125]

[0126] is the enhanced text feature output by the encoder, To enhance the length of text features, To enhance text features feature vectors.

[0127] S10, text feature extraction step: pooling the enhanced text features to obtain a global text feature vector.

[0128] Specifically, the first feature vector of the text feature is enhanced as the global text feature vector.

[0129] S11. Loss function design steps: The similarity between the generated report and the target report is maximized by minimizing the cross entropy loss function as the minimized text loss function, the similarity between the label vector and the true vector is used as the label loss function, and the cosine similarity between the label matrix feature and the global text feature vector is used as the text-label loss function.

[0130] Specifically, the minimized text loss function is expressed as:

[0131]

[0132]

[0133] Where, To minimize the text loss function, The number of words to generate a report for, For the target report generated The probability distribution of words, for Target text at the moment, for The text of the moment, is the label matrix, This is an image of acoustic wave long-range detection imaging.

[0134] The label loss function is expressed as:

[0135]

[0136] Where, is the label loss function, Predicted for an image Category labels, For the real The label of the category, is the number of labels.

[0137] The text-label loss function is:

[0138]

[0139]

[0140] Where, is the text-label loss function, is the global text feature vector, is the label matrix feature, For the real The label of the category, , indicating the category The text feature vector of is the dimension of the label matrix.

[0141] Minimizing the cross-entropy loss improves the matching of the generated text to the target text. Accurate label assignment is ensured by optimizing the label loss function. Using the text-label loss function, the learning of the label matrix is ​​optimized to strengthen the correlation between text content and label information.

[0142] S12. Optimization step: Optimize the interpretation by minimizing the total loss function formed by the text loss function, the label loss function, and the text-label loss function to generate the final interpretation. The total loss function comprehensively optimizes the text, label, and text-label relationship to improve the quality of the final interpretation.

[0143] Specifically, the total loss function is expressed as:

[0144]

[0145] Where, is the total loss function, To minimize the text loss function, is the label loss function, is the text-label loss function, is the weighting factor of the label loss function, is the weighting factor of the text-label loss function.

[0146] In the above-mentioned method for generating geological interpretation conclusions of acoustic wave long-range detection imaging logging of the present invention, the relationship between the label matrix and the generated text is mined through a model, the label matrix features are optimized, and feature fusion is adopted at the same time to realize the intersection of image and text features, strengthen the fusion of visual semantic information, and improve the accuracy of text description of the image. By integrating the encoding, fusion and optimization mechanism of image features and text features, the geological interpretation conclusions of acoustic wave long-range detection imaging logging are automatically and quickly generated. The interpretation conclusions finally generated can simultaneously meet the optimization goals of text, label and text-label relationship, reduce the diversity of conclusions, and have a high interpretation compliance rate.

[0147] A second embodiment of the present invention provides a system for generating geological interpretation conclusions from acoustic wave remote detection imaging logging, which is used to implement the method for generating geological interpretation conclusions from acoustic wave remote detection imaging logging described in the first aspect of the present invention, comprising:

[0148] Data acquisition module 1, used to acquire acoustic wave remote detection imaging images and text sequences;

[0149] Image feature extractor 2, extracting local features of the acoustic wave remote detection imaging image;

[0150] Classifier 3 performs average pooling on local features to extract global features, and performs label classification on local features to obtain label vectors;

[0151] Transformer neural network 4, including:

[0152] Encoder 41 encodes the global features to obtain a hidden vector, indexes the text sequence through word vectors to obtain text features, encodes the text features to obtain enhanced text features, and pools the enhanced text features to obtain a global text feature vector;

[0153] Decoder 42, including:

[0154] The label attention module 421 constructs a label matrix and performs feature fusion of the label matrix with the hidden vector and the label vector to generate label matrix features;

[0155] A feature fusion module 422 performs feature fusion on the text feature and the latent vector to generate a fusion vector, and adds the text feature and the fusion vector to obtain a text vector;

[0156] The interpretation conclusion generation module 423 decodes the label matrix features and the text vector to generate a report and obtain the interpretation conclusion;

[0157] a loss function design module 424 that uses the similarity between the generated report and the target report as the minimized text loss function by minimizing the cross entropy loss function, uses the similarity between the label vector and the true vector as the label loss function, and uses the cosine similarity between the label matrix feature and the global text feature vector as the text-label loss function;

[0158] The optimization module 425 optimizes the interpretation conclusion by minimizing the total loss function formed by the text loss function, the label loss function and the text-label loss function to generate a final interpretation conclusion.

[0159] In the above-mentioned acoustic wave long-range detection imaging logging geological interpretation conclusion generation system of the present invention, an image feature extractor, a classifier, an encoder and a decoder of a Transformer neural network form a model, and the generation of acoustic wave long-range detection imaging logging geological interpretation conclusions is realized through the model. The model uses a total loss function formed by minimizing a text loss function, a label loss function and a text-label loss function as a loss function. The model mines the relationship between the label matrix and the generated text, optimizes the label matrix features, and simultaneously adopts feature fusion to realize the intersection of image and text features, strengthen the fusion of visual semantic information, and improve the accuracy of the text description of the image. By integrating the encoding, fusion and optimization mechanism of image features and text features, the acoustic wave long-range detection imaging logging geological interpretation conclusion is automatically and quickly generated. The interpretation conclusion finally generated can simultaneously meet the optimization goals of text, label and text-label relationship, reduce the diversity of conclusions, and have a high interpretation compliance rate.

[0160] The method for generating geological interpretation conclusions of acoustic wave remote detection imaging logging and the training method of the system described in the above embodiment of the present invention are as follows:

[0161] The experiment was conducted based on the Ubuntu 20.04 operating system, with a graphics card configuration of 12GB, NVIDIA GeForce RTX3060 GPU, and a parallel computing architecture of CUDA11.0. PyCharm 2021 development tools, python 3.7 programming language, and Pytorch 1.7.1 training framework were used. The experiment used a pre-trained ResNet101 neural network to extract far-field detection image features, and then input the generated 512-dimensional image features into the fully connected layer to obtain a 14-dimensional label vector. Set the number of image feature input layers of the Transformer neural network to 3, the multi-head attention head to 8, the hidden dimension to 512, the input label matrix size to 14×512, and set the balance parameters and 1 and 0.5. In the experiments of training Cave and Crack datasets, Adam optimizer, 1e-4 model learning rate and 5e-5 visual extraction learning rate were used for model optimization. The model loss changes on the two datasets are as follows: Figure 6 、 7 As shown. The Cave dataset was trained for 50 rounds, and after 10 rounds of iteration, the loss curve tended to be stable. The Crack dataset was trained for 30 rounds, and after 10 rounds of iteration, the loss curve tended to be stable, and the model achieved the best results. At the same time, Figure 8 、 Figure 9 The box plots in the figure show the distribution of evaluation indicators during model training. Based on the above analysis, it can be concluded that the proposed method for generating geological interpretation conclusions for acoustic remote detection imaging logging and the Transformer neural network in the system exhibit stable progress during the iterative process.

[0162] Figure 6 is the loss value change curve of the model on Cave. Figure 5 As shown in the figure, the Cave dataset is trained for 50 rounds, and the loss curve tends to be stable after 10 rounds of iteration.

[0163] Figure 7 is the loss value change curve of the model on Crack. Figure 6 As shown in the figure, the number of training rounds for the Crack dataset is 30. After 10 rounds of iteration, the loss curve tends to be stable and the model achieves the best effect.

[0164] Figure 8 This is a boxplot of the model's evaluation indicators on the Cave dataset. Figure 9This is a boxplot of the model's evaluation metrics on the Crack dataset. The figure shows the distribution of the evaluation metrics during model training. It can be seen that the model of the present invention exhibits a stable progress during the iterative process.

[0165] In order to verify the effectiveness of the method and system for generating geological interpretation conclusions of acoustic wave remote detection imaging logging described in the above embodiment of the present invention, the following specific embodiments are used for description.

[0166] Example: Finite difference numerical simulation method is used. According to the characteristics of fracture-cavity carbonate reservoir in Tarim Oilfield, in order to simplify the model complexity, the fracture-cavity aggregate model is split into fracture and cave models, including cave model and crack model.

[0167] Finite-difference numerical simulations were conducted according to the parameters shown in Table 1, and acoustic remote sensing migration imaging was performed on the models. The geological model size was set to 30 m × 20 m to capture and simulate geological structural features over a large area. In the cave model, 25,200 cave models were generated by varying the size and number of caves, their distance from the wellbore axis, and the type of cave filling. The fracture model simulated the morphological variations of fractures under different geological conditions by considering their inclination, width, location, and number, resulting in 2,640 simulation models. Table 1 shows the design details of the two models, and the formation elastic parameters used in the numerical simulations are shown in Table 2. After the geological models were constructed, numerical simulations of the borehole acoustic field were performed to generate dipole array waveform data. Using migration imaging, acoustic remote sensing images of the corresponding models were generated. Each geological model was associated with an interpretation report, resulting in the construction of the cave (Cave) and fracture (Crack) datasets, consisting of 25,200 and 2,640 remote sensing images, respectively, and corresponding interpretation reports. The reports were manually annotated and detailedly describe the geological features and parameters (such as cave location, size, and shape) in the images.

[0168] Table 1

[0169]

[0170] Table 2

[0171]

[0172] Table 3 shows the number of images, reports, and average word count in the dataset. The dataset intuitively displays features such as caves and cracks in geological structures.

[0173] Table 3

[0174]

[0175] The experiment split the data into training set, validation set and test set according to 70%-10%-20%, as shown in Table 4.

[0176] Table 4

[0177]

[0178] Figure 10 This figure shows examples of images and reports from the dataset, showing, from left to right, the model type, remote detection images, and interpretation conclusions. For example, the first row shows an image of a single cave, with the corresponding interpretation being: a single cave exists 10 meters from the wellbore axis. It's important to emphasize that achieving a final, rapid on-site interpretation requires more extensive numerical calculations that align with the geological model.

[0179] Figure 11 The following is a comparison of the qualitative experimental results of the method of the present invention and the system. Figure 11 The figure shows, from top to bottom, simulated images of a single wellbore cave, a single through-well fracture, and a cluster of wellbore caves. The heatmap is the output of the last hidden layer of the image feature extraction module, reflecting the model's ability to capture the input image's focus and features. The left column shows the original image, the middle column shows the heatmap generated by the R2Gen method, and the right column shows the heatmap generated by the proposed method and system. Comparing the heatmap results reveals that the text generated by the proposed method and system accurately captures and describes the location and characteristics of fractures and wellbore caves. For example, in the first row, the high-heat region generated by the proposed method and system fully represents the location and size of a single cave, while the heatmap generated by the R2Gen method fails to fully cover the cave area. In the second row, the proposed method and system relatively fully identify the through-well fracture, while the R2Gen method misidentifies unrelated areas outside the fracture. In the third row, the proposed method and system generate multiple high-heat regions within a range of 5-15 meters from the wellbore axis, clearly representing the cave cluster within this range. However, the heatmap generated by the R2Gen method lacks sufficient coverage and fails to accurately display the location and extent of the cave cluster. Through qualitative analysis of the heat map results, the method and system of the present invention outperform the R2Gen method in capturing geological feature details, accurately describing the location and characteristics of fractures and caves, and maintaining consistency under various complex situations.

[0180] Table 5 shows the comparative experimental results of different models on two datasets. As shown in Table 5, among these comparative methods, the traditional basic image captioning method ATT2IN failed to achieve satisfactory results when directly applied to the generation of interpretation conclusions for remote detection images. The R2Gen and R2GenCMN models using memory networks only considered using memory records to generate information. Although they achieved good results, they did not solve the data bias problem where conventional geological structure data accounted for a large proportion while data on abnormal geological structure features such as caves and cracks accounted for a small proportion. The CMCL model gradually improved model performance by estimating the instance training difficulty and model capability. Prior Guided Trans introduced more complex prior knowledge. Compared with the above work, the model structure proposed in the present invention is simpler and achieved better or the same results. Overall, the method and system of the present invention outperformed most comparative methods in all indicators.

[0181] Table 5

[0182]

[0183] Tables 6 and 7 show the ablation test results of the model proposed by the method and system of the present invention on the Cave and Crack datasets, respectively. The base model is a Transformer-based encoder-decoder model. Experiment (a) shows the result of adding a feature fusion module to the base model, experiment (b) adds a label feature matrix, and experiment (c) optimizes the label matrix using a text-label loss. Experiment (d) shows the model proposed by the method and system of the present invention. Comparing the results of the base model and experiment (a) on the two datasets shows that the Transformer network with feature fusion significantly improves on most metrics. This is because traditional Transformer models only fuse image features in the multi-head attention module, while remote detection images contain a large amount of complex geological information. The Transformer encoder tends to capture coarse-grained features and dependencies in the image, while ignoring fine-grained geological anomaly information. Therefore, adding a feature fusion module integrates visual and textual features through the label matrix, allowing the module to focus on important visual features and suppress irrelevant information, thereby providing the model with a richer feature representation. In experiment (b), the label matrix was first initialized using a normal distribution. The features of different label reports were then learned during subsequent training. Compared to the baseline results, various evaluation metrics showed significant improvements, demonstrating that the label matrix trained through this experiment can capture and memorize the diverse features of geological anomalies, providing a richer feature representation during decoding and improving the quality and reliability of the generated text. Experiment (c) further optimized the label matrix by adding a text-label loss, resulting in significant improvements in experimental metrics compared to experiment (b). Because geological anomalies have complex and diverse characteristics, relying solely on traditional loss functions can lead to errors when learning the label matrix. The text-label loss function, however, further optimizes the label matrix, constraining feature consistency and accuracy, reducing bias in label matrix learning and improving overall model performance.

[0184] Table 6

[0185]

[0186] Table 7

[0187]

[0188] Figure 12 、 Figure 13 These are the experimental results of the sensitivity of two corresponding parameters in the model. and The two parameters balance the text loss, label loss, and text-label loss. In order not to affect the performance of the model in generating text reports, the label loss and text-label loss parameters are set to be smaller than the text loss parameter. The impact of the corresponding parameters on the results is tested on the Cave dataset. Response parameters The sensitivity of Figure 12 As shown on the left. The experimental results show that as Increase, the size of each indicator fluctuates greatly until When it reaches 1, the values ​​of most indicators reach their maximum and tend to be stable. The sensitivity of Figure 13 As shown on the right. When the value increases from 0.1 to 0.5, the values ​​of most evaluation indicators gradually increase to the highest level. When the size of increases to 1 again, the size of the evaluation index gradually decreases. It is a parameter that adjusts the loss of the optimization matrix. If the parameter value is too small, the matrix cannot accurately reflect the label vector. If the parameter value is too large, the model will introduce too much noise information, resulting in performance degradation. Experimental results show that 1. When the value is 0.5, the best result is achieved.

[0189] Figures 14 to 16 The following is a diagram of the analysis results of a field example.

[0190] Figure 14 This is the result of acoustic wave remote detection of MX-1 well in Tarim Oilfield (7650-7740m). From left to right, the figure shows the dipole variable density full-wave waveform, the monopole full-wave waveform, the acoustic wave time difference and the corresponding STC correlation diagram, the horizontal well acoustic wave remote detection imaging diagram (up and down and left and right) and the extracted imaging heat map. Figure 1 As shown in the flowchart of the method, the present invention uses a sliding window to segment the original acoustic wave remote detection imaging image. The starting position of the window is set at the top of the image (usually the starting depth), the window size is 2 meters, the step length is 1 meter, and it slides gradually from top to bottom. Each image slice cut out is treated as an independent image and analyzed by the method and system of the present invention. The thermal map generated in the image feature encoding stage is re-spliced ​​according to its position in the original image to ensure that the geological features within the entire depth range are fully presented. From the results of the thermal map imaging, it can be seen that this layer has developed well fractures (7660-7680m), high-angle fractures beside the well (7680-7710m) and caves beside the well (7680-7730m).

[0191] Figure 15 This is the acid fracturing fracture simulation for Well MX-1. It can be seen that the layer with the longest fracture extension corresponds to the layer with developed fractures near the well.

[0192] Figure 16 This is the acid fracturing construction curve of Well MX-1. It can be found that there are many points where the pressure drops rapidly during the fracturing process. Analysis shows that this is the response of connecting natural fractures.

[0193] The above embodiments are used to explain the present invention rather than to limit the present invention. Any modifications and changes made to the present invention within the spirit of the present invention and the protection scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A method for generating geological interpretation conclusions of acoustic wave remote detection imaging logging, characterized in that: The steps are: Data acquisition step: obtaining acoustic wave remote detection imaging images and text sequences; Image feature extraction step: extract local features of the acoustic wave remote detection imaging image, and perform average pooling on the local features to extract global features; Label classification step: perform label classification on local features to obtain label vectors; One-time encoding step: Encode the global features to obtain the hidden vector; Label feature fusion step: construct a label matrix, fuse the label matrix with the hidden vector and label vector to generate label matrix features; Text indexing step: index the text sequence through word vectors to obtain text features; Vector feature fusion step: Fuse the text features with the hidden vector to generate a fusion vector, and add the text features and the fusion vector to obtain the text vector; Explanation conclusion generation step: Decode the label matrix features and text vectors to generate a report and obtain the explanation conclusion; Secondary encoding step: Encode text features to obtain enhanced text features; Text feature extraction step: Pool the enhanced text features to obtain the global text feature vector; Loss function design steps: The similarity between the generated report and the target report is maximized by minimizing the cross entropy loss function as the text loss function, the similarity between the label vector and the true vector is used as the label loss function, and the cosine similarity between the label matrix features and the global text feature vector is used as the text-label loss function; Optimization step: The interpretation conclusion is optimized by minimizing the total loss function formed by the text loss function, label loss function and text-label loss function to generate the final interpretation conclusion.

2. The method for generating geological interpretation conclusions of acoustic wave remote detection imaging logging according to claim 1, characterized in that: In the image feature extraction step, the acoustic wave remote detection imaging image is input into the image feature extractor, and the features output by the last convolutional layer of the image feature extractor are extracted as local features. The local features are input into the classifier for average pooling, and all features of each row are spliced ​​into a sequence as the global feature.

3. The method for generating geological interpretation conclusions of acoustic wave remote detection imaging logging according to claim 1, characterized in that: In the label classification step, the local features are input into the classifier for classification to generate a label vector.

4. The method for generating geological interpretation conclusions of acoustic wave remote detection imaging logging according to claim 1, characterized in that: In the label feature fusion step, the specific method of fusing the label matrix with the hidden vector and the label vector to generate the label matrix feature is: Calculate the similarity between the label matrix and the hidden vector Expressed as: Where, is the label matrix, each row vector of the label matrix , indicating the category The text feature vector of is the number of labels, is the hidden vector, is the hidden vector The transpose of is the dimension of the label matrix; Based on similarity Calculate the attention score with the label vector Expressed as: Where, is the normalized exponential function, is the label vector; Utilizing Attention Score , normalization and residual connection fusion label features to obtain the label matrix feature representation as follows: Where, is the label matrix feature, is the normalization function, is the input variable, is the mean, is the standard deviation.

5. The method for generating geological interpretation conclusions of acoustic wave remote detection imaging logging according to claim 1, characterized in that: In the text feature extraction step, the method for pooling the enhanced text features to obtain the global text feature vector is: pooling the enhanced text features and taking the first vector of the enhanced text features as the global text feature vector.

6. The method for generating geological interpretation conclusions of acoustic wave remote detection imaging logging according to claim 1, characterized in that: In the loss function design step, the minimized text loss function is expressed as: Where, To minimize the text loss function, The number of words to generate a report for, For the target report generated The probability distribution of words, for Target text at the moment, for The text of the moment, is the label matrix, This is an image of acoustic wave long-range detection imaging.

7. The method for generating geological interpretation conclusions of acoustic wave remote detection imaging logging according to claim 1, characterized in that: In the loss function design step, the label loss function is expressed as: Where, is the label loss function, Predicted for an image Category labels, For the real The label of the category, is the number of labels.

8. The method for generating geological interpretation conclusions of acoustic wave remote detection imaging logging according to claim 1, characterized in that: In the loss function design step, the text-label loss function is: Where, is the text-label loss function, is the global text feature vector, is the label matrix feature, For the real The label of the category, , indicating the category The text feature vector of is the dimension of the label matrix.

9. The method for generating geological interpretation conclusions of acoustic wave remote detection imaging logging according to claim 1, characterized in that: In the optimization step, the total loss function is expressed as: Where, is the total loss function, To minimize the text loss function, is the label loss function, is the text-label loss function, is the weighting factor of the label loss function, is the weighting factor of the text-label loss function.

10. A system for generating geological interpretation conclusions of acoustic wave remote detection imaging logging, used to implement the method for generating geological interpretation conclusions of acoustic wave remote detection imaging logging according to any one of claims 1 to 9, characterized in that: include: A data acquisition module, used to acquire acoustic wave remote detection imaging images and text sequences; Image feature extractor, extracting local features of acoustic wave remote detection imaging images; The classifier performs average pooling on local features to extract global features and classifies local features to obtain label vectors; Transformer neural network, including: The encoder encodes the global features to obtain a hidden vector, indexes the text sequence through the word vector to obtain text features, encodes the text features to obtain enhanced text features, and pools the enhanced text features to obtain a global text feature vector; Decoder, built-in: The label attention module constructs a label matrix and fuses the label matrix with the hidden vector and label vector to generate label matrix features; The feature fusion module fuses the text features with the hidden vector to generate a fusion vector, and adds the text features and the fusion vector to obtain the text vector; The interpretation conclusion generation module decodes the label matrix features and text vectors to generate a report and obtain the interpretation conclusion; The loss function design module uses the similarity between the generated report and the target report as the text loss function, the similarity between the label vector and the true vector as the label loss function, and the cosine similarity between the label matrix features and the global text feature vector as the text-label loss function; The optimization module optimizes the explanation conclusion by minimizing the total loss function formed by the text loss function, label loss function and text-label loss function to generate the final explanation conclusion.

Images