Lithology identification method based on lightweight multi-scale adaptive fusion network model

Abstract

This invention discloses a lithology identification method based on a lightweight multi-scale adaptive fusion network model, belonging to the field of geological exploration. The method comprises: constructing a lightweight multi-scale adaptive fusion network model consisting of a data multi-scale preprocessing module, a multi-scale adaptive weighted convolutional neural network module, a channel attention module, and a fine-tuning module; using the data multi-scale preprocessing module to perform multi-scale data preprocessing on the collected data; using the multi-scale adaptive weighted convolutional neural network module to perform multi-scale adaptive weighted processing on the preprocessed lithology data; using the channel attention module to dynamically combine feature maps generated at multiple scales; using the fine-tuning module to further optimize the model output; using a dynamic focal loss function to train the model; and using the trained model to perform lithology identification. This invention effectively alleviates the class imbalance problem of lithology data and improves the accuracy of lithology identification.

Description

Technical Field

[0001] The present invention relates to the field of geological exploration, and in particular to a lithology identification method based on a lightweight multi-scale adaptive fusion network model. Background Art

[0002] Lithology identification and reservoir delineation are processes that utilize specific methods to identify and differentiate lithologies. They are fundamental tasks in formation evaluation and reservoir description, and crucial in various fields, including geology, geological resource discovery, and underground engineering. Traditional lithology identification methods rely heavily on the experience of geologists and physical logging data. These methods are often time-consuming and susceptible to subjective factors, making them difficult to meet the demands of modern oil exploration and development. Although intelligent lithology identification based on deep learning is becoming a trend, with an increasing number of lithology identification tasks being completed quickly and cost-effectively by artificial intelligence, the complexity of lithology identification prevents the full integration of logging information through any mathematical methods, and constructing actual underground lithology sequences presents numerous challenges.

[0003] The three main problems of intelligent lithology identification methods based on deep learning are as follows:

[0004] 1. Ignoring the spatial relationships of lithologic data. Most existing data-driven methods simply aggregate geological data into a single vector and input it into a neural network for prediction. They fail to consider the spatial connections between geological data points or the geological context of neighboring data points. As a result, these methods often produce inaccurate predictions in areas with complex geological structures.

[0005] 2. Data imbalance bias exists. Geological data indicates that some lithology types are more common than others, leading data-driven prediction methods to favor more common lithology types. Traditional oversampling methods, due to their blindness and marginal sampling, struggle to generate highly realistic samples. Consequently, they are limited in improving lithology identification and are unable to meet practical application requirements.

[0006] 3. Insufficient integration of geological knowledge. Vertical stacking patterns of different rock types are crucial for reducing uncertainty in lithologic sequence predictions. These stacking patterns contribute to more accurate predictions of rock layer distribution and properties in geological modeling and resource assessment. However, current methods often overlook this aspect.

[0007] In view of this, it is necessary to develop a lithology identification method based on a lightweight multi-scale adaptive fusion network model to solve the above problems. Summary of the Invention

[0008] The technical problem to be solved by the present invention is to provide a lithology identification method based on a lightweight multi-scale adaptive fusion network model, aiming to alleviate the class imbalance problem of lithology data and improve the accuracy of lithology identification.

[0009] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0010] A lithology identification method based on a lightweight multi-scale adaptive fusion network model includes the following steps:

[0011] Step 1: Build a lightweight multi-scale adaptive fusion network model;

[0012] The lightweight multi-scale adaptive fusion network model includes a data multi-scale preprocessing module, a multi-scale adaptive weighted convolutional neural network module, a channel attention module and a fine-tuning module;

[0013] Step 2: Use a data multi-scale preprocessing module to perform multi-scale data preprocessing on the collected data;

[0014] Step 3: Use a multi-scale adaptive weighted convolutional neural network module to perform multi-scale adaptive weighted processing on the pre-processed lithologic data;

[0015] Step 4: Use the channel attention module to dynamically combine the feature maps generated at multiple scales;

[0016] Step 5: Use the fine-tuning module to further optimize the output of the lightweight multi-scale adaptive fusion network model;

[0017] Step 6: Use dynamic focal loss function to train the lightweight multi-scale adaptive fusion network model;

[0018] Step 7: Use the trained lightweight multi-scale adaptive fusion network model to perform lithology identification.

[0019] A further improvement of the technical solution of the present invention is that: the data multi-scale preprocessing module is responsible for data standardization and feature engineering construction, and the preprocessed data serves as the input of the multi-scale adaptive weighted convolutional neural network module; the multi-scale adaptive weighted convolutional neural network module effectively captures and enhances the key features and subtle differences in the lithological data by analyzing data features at different scales; the channel attention module receives the output of the multi-scale adaptive weighted convolutional neural network module, emphasizes key feature channels to improve the sensitivity and processing capability of the lightweight multi-scale adaptive fusion network model to important information; the fine-tuning module further optimizes the output of the lightweight multi-scale adaptive fusion network model, and enhances the adaptability of the lightweight multi-scale adaptive fusion network model to complex data.

[0020] A further improvement of the technical solution of the present invention is that in step 2, the data multi-scale preprocessing module performs normalization processing on the logging data to ensure that all data are evenly distributed in the interval [0, 1] and enhance the comparability of the data, as shown below:

[0021]

[0022] Where, X norm is the normalized attribute value; X represents the normalized initial attribute value, X min and X max Corresponding to the upper and lower boundaries of a specific attribute within a given well log;

[0023] When constructing the data, single-scale, three-scale, and five-scale methods are used. Contextual information from the lithologic data is taken into account, and the focus is expanded from a single data point to include information from adjacent data points. This fully captures the spatial information of the lithologic data. Taking the three-scale approach as an example, the data construction process is as follows:

[0024] S3={Previous,Current,Next} (2)

[0025] Where S3 represents the three-scale data, Previous represents the previous data, Current represents the current data, and Next represents the next data.

[0026] A further improvement of the technical solution of the present invention is that in step 3, the multi-scale adaptive weighted convolutional neural network module is composed of three shallow convolutional neural network streams, namely MA-CNN-1, MA-CNN-2 and MA-CNN-3; each network stream processes a specific input data segment, namely Input-1, Input-2 and Input-3; as shown below:

[0027] F i =MA-CNN i (Input i θ i ),i=1,2,3 (3)

[0028] Where, F i Represents the output features of the i-th convolutional neural network stream, MA-CNN i Represents the i-th multi-scale adaptive convolutional neural network flow, Input i is the input data processed by the i-th network flow, θ i Represents the parameters of the i-th network flow;

[0029] By dynamically combining the three-scale feature maps (F1, F2, F3) generated by the multi-scale adaptive weighted convolutional neural network module with adaptive weights, we obtain an optimized feature map, which accurately analyzes adjacent segments of different sizes and enhances the recognition and utilization of key features, as shown below:

[0030] Feature=Fusion(F1,F2,F3) (4)

[0031] Where Feature is the optimized feature map.

[0032] A further improvement of the technical solution of the present invention is that: in step 4, the channel attention module includes three basic components: a "squeeze block", an "excitation block" and a "scaling operation"; the channel attention module performs compression-excitation processing on the feature map Feature output by MACNN to achieve effective optimization of feature mapping, strengthen the focus on important channel information and reduce the influence of secondary channels, and enhance the ability of the lightweight multi-scale adaptive fusion network model in multi-scale information processing;

[0033] The extrusion block fsq(.) is as follows:

[0034]

[0035] Where z i Indicates the compression value of the i-th channel, F ijk is the eigenvalue of position (j, k) in the i-th channel, H and W are the height and width of the feature map respectively;

[0036] The stimulus block fex(.) is as follows:

[0037] s i =σ(W ex2 δ(W ex1 z i )) (6)

[0038] Where s i represents the stimulus value of the i-th channel, z i is the value obtained from the compression operation, σ and δ are activation functions respectively;

[0039] The scaling operation fscal(.) is as follows:

[0040] F i scaled =s i ·F i (7)

[0041] In, F i scaled is the scaled feature map of the i-th channel, s i represents the weight obtained through the incentive process, and F i is the initial feature representation of the i-th channel.

[0042] A further improvement of the technical solution of the present invention is that in step 5, the fine-tuning module further enhances the processing capability of lithologic data through feature extraction and dimensionality reduction combined with Dropout technology, while reducing overfitting and improving the robustness of the lightweight multi-scale adaptive fusion network model in practical applications.

[0043] A further improvement of the technical solution of the present invention is that in step 6, the lightweight multi-scale adaptive fusion network model uses a dynamic focus loss function during training to adjust the degree of attention paid to different lithology categories and strengthen the emphasis on minority class samples; the recognition performance of the lightweight multi-scale adaptive fusion network model when dealing with class imbalance problems is improved, and its formula is as follows:

[0044] FL(p t )=-α t ·(1-p t ) γ ·log(p t ) (8)

[0045] Where p t Represents the probability that the model accurately identifies the target category; α t is a weighting factor used to adjust the importance of different categories; γ is an adjustable exponent that is adjusted through the dynamic adaptive mechanism of the neural network.

[0046] Due to the adoption of the above technical solution, the technical advancements achieved by the present invention are:

[0047] 1. By constructing a lightweight multi-scale adaptive fusion network model, the present invention can recognize that the stacking of rock layers is not random, but follows specific geological processes and histories. By identifying these patterns, the uncertainty in lithologic sequence prediction can be reduced; the ability to capture lithologic variability can be improved, especially when processing rock data with complex geological structures. It can significantly improve the accuracy and efficiency of lithologic identification, while increasing the adaptability, flexibility and explanatory power of the model; and effectively improve the overall accuracy and reliability of lithologic identification tasks.

[0048] 2. In the present invention, through normalization processing, data of different units and magnitudes can be converted to the same scale, which not only eliminates the deviation between data of different magnitudes, but also improves the convergence speed and performance of the lightweight multi-scale adaptive fusion network model; multi-scale processing of the normalized data can extract important features from different scales, which helps to capture subtle changes in the data and enhance the lightweight multi-scale adaptive fusion network model's ability to understand complex lithological data; using the lightweight multi-scale adaptive fusion network model to automatically learn the multi-scale representation of data in the preprocessing stage can effectively solve the multi-scale problems that are difficult to handle with manual feature engineering and improve the generalization ability of the model; in the multi-scale data preprocessing, the use of heterogeneous data fusion technology can effectively integrate data from different sources and scales, and improve the lightweight multi-scale adaptive fusion network model's ability to recognize complex geological structures; it can more accurately identify and predict rock sequences based on the vertical stacking patterns of different rock types.

[0049] 3. Traditional multi-scale processing methods often use fixed weight distribution schemes, which may not fully consider the importance and contribution of data at different scales. In the present invention, by introducing an adaptive weight distribution mechanism, the weights of data at each scale can be dynamically adjusted according to the characteristics and context of the data, thereby more effectively utilizing multi-scale information for lithologic identification; by adaptively adjusting the weights of data at each scale, the contribution of information at different scales to lithologic identification can be more accurately reflected, thereby optimizing information utilization in the identification process and improving the accuracy and reliability of lithologic identification; in multi-scale adaptive weighted processing, not only single-scale data is considered, but also more comprehensive lithologic characteristic information is obtained by fusing data at different scales.

[0050] 4. The channel attention module in the present invention can dynamically combine feature maps generated at multiple scales according to the importance of the features. This means that the channel attention module does not simply merge feature maps of all scales together, but adjusts the weight of each feature in the final combination based on its contribution and relevance; this enables the lightweight multi-scale adaptive fusion network model to not only improve processing efficiency, but also improve recognition accuracy by reducing the interference of irrelevant features.

[0051] 5. In the present invention, the dynamic focus loss function is used to enable the lightweight multi-scale adaptive fusion network model to pay more attention to those difficult-to-identify minority class samples during the training process. This attention not only improves the lightweight multi-scale adaptive fusion network model's ability to recognize minority classes, but also helps to improve the overall classification accuracy; the adaptive characteristics of the dynamic focus loss function enable the lightweight multi-scale adaptive fusion network model to better adapt to the complexity and imbalance of lithological data; in traditional focus loss methods, the γ value is usually fixed and needs to be adjusted through experience. The dynamic focus loss method in the present invention reduces the dependence on manual experience by automatically adjusting the γ value, making the lightweight multi-scale adaptive fusion network model training process more efficient and automated; by assigning higher loss weights to challenging samples (especially minority class samples), the lightweight multi-scale adaptive fusion network model can be more focused on correctly identifying these samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. Those skilled in the art can derive other drawings based on these drawings without inventive effort.

[0053] Figure 1 is an architectural diagram of a lightweight multi-scale adaptive fusion network model in an embodiment of the present invention;

[0054] Figure 2 This is a parameter configuration diagram of a lightweight multi-scale adaptive fusion network model in an embodiment of the present invention. DETAILED DESCRIPTION

[0055] It should be noted that the terms "including" and "having" and any variations thereof in the specification and claims of the present invention and the above-mentioned drawings are intended to cover non-exclusive inclusions. For example, a process, method, system, product or apparatus comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or are inherent to these processes, methods, products or apparatuses.

[0056] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments:

[0057] like Figure 1 As shown in the figure, a lithology identification method based on a lightweight multi-scale adaptive fusion network model uses vertical reservoir information for lithology identification, including the following contents:

[0058] Step 1: Build a lightweight multi-scale adaptive fusion network (LMAFNet) model;

[0059] The lightweight multi-scale adaptive fusion network (LMAFNet) model includes a data multi-scale preprocessing module, a multi-scale adaptive weighted convolutional neural network module, a channel attention module, and a fine-tuning module;

[0060] The multi-scale data preprocessing module is responsible for data standardization and feature engineering. The preprocessed data serves as the input of the multi-scale adaptive weighted convolutional neural network module. It is committed to exploring the characteristics of lithologic changes and integrating the adjacent segment information, spatial relationships, and continuity of lithologic data at different scales.

[0061] The multi-scale adaptive weighted convolutional neural network module effectively captures and enhances the key features and subtle differences in lithologic data by analyzing data features at different scales; it uses multi-scale adaptive weighting technology to capture subtle features in lithologic data.

[0062] The channel attention module receives the output of the multi-scale adaptive weighted convolutional neural network module and emphasizes key feature channels to improve the model's sensitivity and processing capabilities for important information. It implements a channel attention mechanism to adaptively integrate feature maps from the multi-scale adaptive weighted convolutional neural network module's output. This enables the lightweight multi-scale adaptive fusion network model to selectively focus on and utilize different features, thereby improving the accuracy of lithology identification.

[0063] The fine-tuning module further optimizes the output of the lightweight multi-scale adaptive fusion network model, enhancing its adaptability to complex data. Following the channel attention module, the fine-tuning module and fully connected layers are used to complete lithology classification. Through feature extraction and dimensionality reduction, combined with dropout technology, the fine-tuning module further enhances its ability to process lithology data, while mitigating overfitting and improving the robustness of the lightweight multi-scale adaptive fusion network model in practical applications.

[0064] Step 2: Use the data multi-scale preprocessing module to perform multi-scale data preprocessing on the collected data; the specific process is as follows:

[0065] In a given dataset, attributes measured in different units can lead to discrepancies because attributes with smaller values can be masked by attributes with larger values in a machine learning model. Therefore, normalizing attribute values is crucial to ensure consistent scaling, as shown below:

[0066]

[0067] Where, X norm is the standardized attribute value; X represents the initial attribute value before normalization, X min and Xmax Correspondence represents the upper and lower boundaries of a specific attribute within a given well log curve.

[0068] There is significant variability in resistivity between different rock types. Directly normalizing this data can mask the unique resistivity characteristics inherent to a particular rock type. To address this issue, the resistivity data are converted to a base-10 logarithmic scale before normalization. This approach effectively reduces the differences in resistivity ranges between rock types. When analyzing well log data from different wells, inaccuracies in the logging system can affect the comparability of curves between wells. This variability affects the comparability of data between wells because they are no longer operated on the same scale and accuracy, complicating the data comparison process. To improve the accuracy of well-to-well comparisons, the data from each well are normalized before integration. This process involves adjusting the values of each attribute in the well log data to ensure a uniform distribution within the range [0,1].

[0069] Normalized data is processed at multiple scales, based on the vertical stacking patterns of different rock types, to help reduce uncertainty in lithologic sequence predictions. When constructing the dataset, contextual information from the lithologic data is incorporated, expanding the focus from a single data point to include information from adjacent data points. For example, with three-scale input, each data point needs to be analyzed in relation to the previous, current, and next data points.

[0070] S3={Previous,Current,Next} (2)

[0071] Where S3 represents the three-scale data, Previous represents the previous data, Current represents the current data, and Next represents the next data.

[0072] By integrating and analyzing vertical reservoir data, the accuracy in delineating the continuity and periodicity of lithologic features can be improved.

[0073] Step 3: Use a multi-scale adaptive weighted convolutional neural network module to perform multi-scale adaptive weighted processing on the pre-processed lithologic data;

[0074] Specifically, lithology identification is a complex, dynamically evolving, nonlinear process that is influenced by gradual changes in physical properties such as the porosity and permeability of underground rocks. LMAFNet uses a specialized convolutional neural network to analyze adjacent segments of different dimensions. This enhances its ability to detect complex patterns and improves its accuracy and reliability in lithology identification. The multi-scale adaptive weighted convolutional neural network module consists of three shallow convolutional neural network streams, namely MA-CNN1, MA-CNN2, and MA-CNN3. Each network stream processes a specific input data segment, namely Input-1, Input-2, and Input-3, as shown below:

[0075] F i =MA-CNN i (Input i θ i ),i=1,2,3 (3)

[0076] Where, F i Represents the output features of the i-th convolutional neural network stream, MA-CNN i Represents the i-th multi-scale adaptive convolutional neural network flow (Multi-Scale Adaption Convolutional Neural Network), Input i is the input data processed by the i-th network flow, θ i Represents the parameters of the i-th network flow.

[0077] By dynamically combining the weights with the three-scale feature maps (F1, F2, F3) generated by the multi-scale adaptive convolutional neural network module, we obtain an optimized feature map, which accurately analyzes adjacent segments of different sizes and enhances the recognition and utilization of key features, as shown below:

[0078] Feature=Fusion(F1,F2,F3) (4)

[0079] These network flows implement adaptive weighting, selectively amplifying or minimizing the contribution of various features, and enhancing the recognition and utilization of key features. LMAFNet's input X is divided into three segments: X1, X2, and X3. Each segment is processed based on the raw well logging data.

[0080] Step 4: Use the channel attention module to dynamically combine the feature maps generated at multiple scales. The specific process is as follows:

[0081] The three-scale feature maps generated by the multi-scale MCNN module are dynamically combined using a channel attention module based on SENet (Squeeze and Excitation Block). The input of the channel attention module consists of feature maps F1, F2, and F3 from the three feature extraction networks MA-CNN1, MA-CNN2, and MA-CNN3. These feature maps are combined using adaptive weights to form the input feature map F.

[0082] The channel attention module consists of three basic components: a "squeeze block," an "excitation block," and a "scaling operation." This module applies a compression-and-excitation (Squeeze-and-Excitation Net) process to the feature maps output by the MACNN. This effectively optimizes the feature mapping, focusing on important channel information while reducing the influence of less important channels, thereby improving the model's ability to process multi-scale information.

[0083] The squeeze block applies global average pooling to compress the feature maps within each channel into a single scalar value that encapsulates the overall importance of each channel. Conversely, the excitation block utilizes a fully connected network to determine the weights for each channel, generating a set of weights that subsequently reweight the feature maps for each channel. Finally, the scaling operation applies these weights obtained from the excitation block to the original feature maps, amplifying the information of meaningful channels while reducing the information of less important channels.

[0084] The extrusion block fsq(.) is shown as follows:

[0085]

[0086] Where z i is the compression value of the i-th channel, F ijk is the eigenvalue at the i-th channel (j, k), H and W are the height and width of the feature map.

[0087] The excitation block fex(.) is shown as follows:

[0088] s i =σ(W ex2 δ(W ex1 z i )) (6)

[0089] Where s i is the excitation value of channel i, z i is the value obtained by the extrusion operation, W ex1 and W ex2 is the value obtained by the squeezing operation, σ and δ are activation functions.

[0090] The scaling operation fscal(.) is as follows:

[0091] F i scaled =s i ·F i (7)

[0092] Where, F i scaled is the scaled feature map of the i-th channel, s i represents the weight obtained in the incentive process, F i is the initial feature representation of the i-th channel.

[0093] Step 5: Use the fine-tuning module to further optimize the output of the lightweight multi-scale adaptive fusion network model;

[0094] Specifically, the fine-tuning module further enhances the processing capability of lithologic data through feature extraction and dimensionality reduction combined with Dropout technology, while alleviating overfitting and improving the robustness of the model in practical applications.

[0095] Step 6: Use the dynamic focus loss function to train the lightweight multi-scale adaptive fusion network model; the specific process is as follows:

[0096] Class imbalance learning methods aim to alleviate the tendency of lightweight multi-scale adaptive fusion network models to favor more frequently occurring classes by increasing the importance of less representative classes. Focal Loss is a loss function introduced to address the class imbalance problem. Its main goal is to reduce the importance of easily classified samples and shift the focus of the lightweight multi-scale adaptive fusion network model to more challenging classification samples, as shown in the following formula:

[0097] FL(p t )=-α t ·(1-p t ) γ ·log(p t ) (8)

[0098] Where p t Indicates the possibility of the model accurately identifying the target class; α t is a weighting factor used to adjust the importance of different categories; the modulation factor is expressed as (1-p t ) γ , where γ is an adjustable exponent; this factor reduces the weight given to samples that are easy to classify and increases the weight given to samples that are difficult to classify.

[0099] In traditional focal loss, the value of γ is usually a predetermined hyperparameter that requires empirical fine-tuning. The present invention proposes a dynamic focal loss method that enables a lightweight multi-scale adaptive fusion network model to dynamically modify the degree of concentration of weights by adjusting the value at different training stages. When the lightweight multi-scale adaptive fusion network model encounters samples from the minority class, its γ value automatically increases, increasing the loss weight of the challenging samples, ensuring that the lightweight multi-scale adaptive fusion network model focuses more on learning these critical but challenging instances. In the field of lithology recognition, faced with obvious class imbalance, traditional loss functions may inadvertently cause the model to overfit to the more dominant class. The dynamic focal loss function can alleviate this class imbalance problem.

[0100] During the training of the lightweight multi-scale adaptive fusion network model, grid search was used to optimize multiple hyperparameters. These hyperparameters, including batch size (batch = [16, 32, 64, 128]), number of training epochs (epochs) = [20, 40, 60, 80, 100], and learning rate (learning rate = [0.01, 0.001, 0.0005, 0.00025], were initially selected based on empirical information. Through grid search and hyperparameter tuning techniques, a batch size of 32 and a number of training epochs of 80 were determined for training the lightweight multi-scale adaptive fusion network model. Training of the lightweight multi-scale adaptive fusion network model began with an initial learning rate of 0.01 and was optimized using the Adam optimizer. Furthermore, to further optimize the performance of the lightweight multi-scale adaptive fusion network model during training, a learning rate scheduler (StepLR) was implemented, reducing the learning rate to one-tenth of its initial value after every 20 training iterations. This method ensures the dynamic adjustment of the learning rate and improves the effectiveness of the lightweight multi-scale adaptive fusion network model training scheme.

[0101] Specific lightweight multi-scale adaptive fusion network model parameters are as follows: Figure 2 As shown, its configuration is as follows:

[0102] First, feature extraction is performed simultaneously through three parallel one-dimensional convolutional layers. The first convolutional layer has 1 input channel, 16 output channels, a kernel size of 1, a stride of 1, and no padding. The second convolutional layer has 3 input channels, 16 output channels, a kernel size of 3, a stride of 1, and padding of 1. The third convolutional layer has 5 input channels, 16 output channels, a kernel size of 5, a stride of 1, and padding of 1. Each convolutional layer is followed by a batch normalization layer and a LeakyReLU activation function to optimize learning and increase nonlinearity. After the three convolutional layers, a lightweight multi-scale adaptive fusion network model introduces an adaptive data fusion mechanism, effectively fusing the outputs of the three convolutional layers through dynamic weights and a softmax function. The fine-tuning module further processes the fused 48-channel output. A one-dimensional convolutional layer is first used to reduce the number of channels to 16, followed by batch normalization and a LeakyReLU activation to enhance feature representation. Subsequently, a 1D transposed convolutional layer halved the number of channels to 8, followed by another convolutional layer to restore the number of channels to 16. Finally, a fully connected layer mapped the features to 128 dimensions. After a batch normalization layer and LeakyReLU activation, a second fully connected layer mapped the 128-dimensional features to the final number of categories.

[0103] Step 7: Use the trained lightweight multi-scale adaptive fusion network model to perform lithology identification.

[0104] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A lithology identification method based on a lightweight multi-scale adaptive fusion network model, characterized in that: The following steps are involved: Step 1: Build a lightweight multi-scale adaptive fusion network model; The lightweight multi-scale adaptive fusion network model includes a data multi-scale preprocessing module, a multi-scale adaptive weighted convolutional neural network module, a channel attention module and a fine-tuning module; Step 2: Use a data multi-scale preprocessing module to perform multi-scale data preprocessing on the collected data; Step 3: Use a multi-scale adaptive weighted convolutional neural network module to perform multi-scale adaptive weighted processing on the pre-processed lithologic data; The multi-scale adaptive weighted convolutional neural network module consists of three shallow convolutional neural network streams, namely MA-CNN-1, MA-CNN-2 and MA-CNN-3; each network stream processes a specific input data segment, namely Input-1, Input-2 and Input-3, as shown below: F i =MA-CNN i (Input i ;θ i ),i=1,2,3 (3) Where, F i Represents the output features of the i-th convolutional neural network stream, MA-CNN i Represents the i-th multi-scale adaptive convolutional neural network flow, Input i is the input data processed by the i-th network flow, θ i Represents the parameters of the i-th network flow; By dynamically combining the three-scale feature maps (F1, F2, F3) generated by the multi-scale adaptive weighted convolutional neural network module with adaptive weights, we obtain an optimized feature map, which accurately analyzes adjacent segments of different sizes and enhances the recognition and utilization of key features, as shown below: Feature=Fusion(F1,F2,F3) (4) In the formula, Feature is the optimized feature map; Step 4: Use the channel attention module to dynamically combine the feature maps generated at multiple scales; Step 5: Use the fine-tuning module to further optimize the output of the lightweight multi-scale adaptive fusion network model; Step 6: Use dynamic focal loss function to train the lightweight multi-scale adaptive fusion network model; Step 7: Use the trained lightweight multi-scale adaptive fusion network model to perform lithology identification.

2. The lithology identification method based on a lightweight multi-scale adaptive fusion network model according to claim 1 is characterized in that: The data multi-scale preprocessing module is responsible for data standardization and feature engineering construction, and the preprocessed data serves as the input of the multi-scale adaptive weighted convolutional neural network module; the multi-scale adaptive weighted convolutional neural network module effectively captures and enhances the key features and subtle differences in lithologic data by analyzing data features at different scales; the channel attention module receives the output of the multi-scale adaptive weighted convolutional neural network module, emphasizes key feature channels to improve the sensitivity and processing capability of the lightweight multi-scale adaptive fusion network model to important information; the fine-tuning module further optimizes the output of the lightweight multi-scale adaptive fusion network model, and enhances the adaptability of the lightweight multi-scale adaptive fusion network model to complex data.

3. The lithology identification method based on a lightweight multi-scale adaptive fusion network model according to claim 1 is characterized in that: In step 2, the data multi-scale preprocessing module performs normalization on the logging data to ensure that all data are evenly distributed in the interval [0, 1] and enhance the comparability of the data, as shown below: Where, X norm is the normalized attribute value; X represents the normalized initial attribute value, X min and X max Corresponding to the upper and lower boundaries of a specific attribute within a given well log; When constructing the data, single-scale, three-scale, and five-scale methods are used. Contextual information from the lithologic data is taken into account, and the focus is expanded from a single data point to include information from adjacent data points. This fully captures the spatial information of the lithologic data. Taking the three-scale approach as an example, the data construction process is as follows: S3={Previous,Current,Next} (2) Where S3 represents the three-scale data, Previous represents the previous data, Current represents the current data, and Next represents the next data.

4. The lithology identification method based on a lightweight multi-scale adaptive fusion network model according to claim 1 is characterized in that: In step 4, the channel attention module includes three basic components: a "squeeze block," an "excitation block," and a "scaling operation." The channel attention module performs compression-excitation processing on the feature map output by MACNN to achieve effective optimization of feature mapping, strengthen the focus on important channel information, and reduce the influence of minor channels, thereby improving the lightweight multi-scale adaptive fusion network model's ability to process multi-scale information. The extrusion block fsq(.) is as follows: Where z i Indicates the compression value of the i-th channel, F ijk is the eigenvalue of position (j, k) in the i-th channel, H and W are the height and width of the feature map respectively; The stimulus block fex(.) is as follows: s i =σ(W ex2 δ(W ex1 z i )) (6) Where s i represents the stimulus value of the i-th channel, z i is the value obtained from the compression operation, σ and δ are activation functions respectively; The scaling operation fscal(.) is as follows: F i scaled =s i ·F i (7) In, F i scaled is the scaled feature map of the i-th channel, s i represents the weight obtained through the incentive process, and F i is the initial feature representation of the i-th channel.

5. The lithology identification method based on a lightweight multi-scale adaptive fusion network model according to claim 1 is characterized in that: In step 5, the fine-tuning module further enhances the processing capability of lithologic data through feature extraction and dimensionality reduction combined with Dropout technology, while alleviating overfitting and improving the robustness of the lightweight multi-scale adaptive fusion network model in practical applications.

6. The lithology identification method based on a lightweight multi-scale adaptive fusion network model according to claim 1, characterized in that: In step 6, during the training process, the lightweight multi-scale adaptive fusion network model uses a dynamic focus loss function to adjust the degree of attention paid to different lithology categories and strengthen the emphasis on minority class samples; the recognition performance of the lightweight multi-scale adaptive fusion network model when dealing with class imbalance problems is improved. The formula is as follows: FL(p t )=-a t ·(1-p t ) γ ·log(p t ) (8) Where p t Represents the probability that the model accurately identifies the target category; α t is a weighting factor used to adjust the importance of different categories; γ is an adjustable exponent that is adjusted through the dynamic adaptive mechanism of the neural network.

Images