A drilling camera intelligent interpretation method based on a geological visual large model

Abstract

The application discloses a kind of based on geological visual big model's borehole camera intelligent interpretation method, comprising: 1) collection borehole image and make pretreatment, construct sample library containing geological label;2) based on existing general visual big model, embed geological feature attention module and lithology classification adapter, combined with deep coal mine geological features are optimized, construct geological visual big model;3) construct geological algorithm dictionary and convert geological knowledge experience into algorithm module capable of calculation, and embed geological visual big model;4) through AI module realizes sample self-adapting diagnosis and repeated learning.The method of the present application can significantly improve the accuracy of geological feature recognition in borehole camera image by constructing geological visual big model, embedding geological feature attention module and geological algorithm dictionary in the model and fusing sample self-adapting diagnosis and repeated learning mechanism, realize the efficient and accurate interpretation of borehole image, and provide accurate geological data support for underground operation.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of coal mine borehole camera image processing and artificial intelligence, specifically a method for intelligent interpretation of borehole camera images based on a large geological visual model. Background Technology

[0002] In deep coal mining, borehole camera technology is a core means of directly acquiring information about the underground rock mass structure, and the image interpretation results directly determine the accuracy of geological understanding and the safety of mining. Traditional coal mine borehole camera interpretation methods mainly rely on manual identification. These methods have drawbacks such as low efficiency, strong subjectivity, and unstable identification accuracy in data image processing and analysis, making it impossible to accurately and efficiently provide complex geological conditions in a short time, thus affecting the efficiency of coal mining.

[0003] Especially in the post-processing and interpretation of borehole camera images, current research still relies heavily on manual, experience-based recognition. This approach fails to efficiently and accurately extract crucial information such as coal seam geology from massive images and complex data. The limitations of existing image recognition-based intelligent interpretation methods are mainly reflected in the following aspects:

[0004] 1. Low image recognition efficiency: Traditional image processing methods are based on manual recognition, which is time-consuming and difficult to invest in for long periods of time, resulting in low efficiency when dealing with a large number of image processing tasks.

[0005] 2. High sample dependence: Current image recognition and analysis mainly rely on the experience of technical personnel, which makes it difficult to quickly identify complex geological structures and results in insufficient accuracy in image interpretation.

[0006] 3. Inapplicability of existing AI models: Existing intelligent vision models are not embedded in deep coal mine geological scenes, making it difficult to accurately analyze rock strata, fissures, etc. in coal mines.

[0007] These shortcomings result in insufficient extraction of geological information from borehole camera data, making it difficult to meet the needs of real-time and accurate analysis of coal seam geology for safe mining in deep coal mines. Summary of the Invention

[0008] The purpose of this invention is to address the shortcomings of existing technologies by providing an intelligent interpretation method for borehole images based on a large geological visual model. By constructing a large visual model adapted to the geology of deep coal mines, embedding a geological algorithm dictionary, and integrating sample adaptive diagnosis and iterative learning mechanisms, this method achieves efficient and accurate interpretation of borehole images, providing technical support for geological analysis of deep coal mines.

[0009] To achieve the above objectives, the present invention adopts the following technical solution.

[0010] A method for intelligent interpretation of borehole camera images based on a large geological visual model includes the following steps:

[0011] Step S1: Acquire borehole images and perform preprocessing to construct a sample library with geological labels;

[0012] Borehole images were acquired and subjected to denoising and enhancement processes to obtain a geologically labeled sample library. The sample library was divided into a training set, a validation set, and a test set according to a certain ratio, wherein:

[0013] The number of borehole image samples is less than 5,000, and the sample library is divided into training set, validation set and test set in a 7:2:1 ratio;

[0014] The sample library contains 5,000 to 50,000 borehole images, divided into training, validation, and test sets in a 6:3:1 ratio.

[0015] The number of borehole image samples exceeds 50,000, and the sample library is divided into training set, validation set, and test set in a 5:4:1 ratio;

[0016] Step S2: Based on the existing general-purpose visual model, embed the geological feature attention module and lithology classification adapter, combine them with the geological features of deep coal mines for special optimization, construct a geological visual model, train and validate the constructed geological visual model, and evaluate the accuracy of the geological visual model.

[0017] Step S3: Construct a geological algorithm dictionary to transform geological knowledge and experience into computable algorithm modules and embed them into a large geological visual model;

[0018] Step S4: Implement adaptive diagnosis and iterative learning of samples through the AI ​​module.

[0019] Specifically, in step S1, the image is denoised and enhanced to obtain a sample library containing geological labels, as detailed below:

[0020] Denoising: Gaussian filtering is used to smooth the noise in the borehole image, and median filtering is used to remove noise caused by particles inside the borehole;

[0021] Enhancement processing: The adaptive histogram equalization CLAHE algorithm is used to perform block-based local contrast adjustment on the borehole image, enhance the feature difference between dark and bright areas, and improve the identification of rock strata boundaries;

[0022] The rock strata interfaces, fractures, and faults in the denoised and enhanced borehole images are labeled to obtain a sample library with geological labels:

[0023] For rock strata interfaces, a combined algorithm of gray-level gradient and edge tracking is used. The interface position is located by calculating the abrupt change point of the first derivative of pixel gray-level. Then, Hough transform is used to fit the interface line to mark the boundary position of different lithologies. The interface burial depth, dip angle, and coverage of horizontal bedding and wavy bedding are recorded. For fractures, the edges are extracted by the Canny edge detection operator, and noise is removed by morphological operations. Then, the fracture length, strike, and density are calculated by connected component analysis and classified and labeled. For faults, a gray-level variance identification method based on fracture zone is used. When the gray-level variance of a local area is greater than the preset threshold of gray-level variance and the gray-level continuous length of a local area is greater than the preset threshold of gray-level continuous length, it is determined to be a fault and the top and bottom boundaries, drop, and dip angle of the fault fracture zone are marked.

[0024] Specifically, the process of constructing the large-scale geological visual model in step S2 is as follows:

[0025] Step S21: Loading and initializing the basic model;

[0026] Based on the Swing Transformer model, the parameters of the first 6 layers of the Swing Transformer model are frozen to retain the general feature extraction capability, and the parameters of the last 4 layers and the classification head are unfrozen to adjust for the coal mine scenario; the parameters of the newly added modules are initialized: the branch weights of the geological feature attention module are initialized, and the fully connected layer of the lithology classification adapter is initialized with He.

[0027] The core constraint of the branch weights [a,b,c] of the geological feature attention module is a+b+c=1, where a is the texture feature branch, b is the edge feature branch, and c is the semantic feature branch. When the edge feature is the core of the geological visual model (such as when the cracks and borehole wall scratches are separated by the edge linearity zone), the edge feature has the greatest impact on the accuracy of the geological visual model, so it is given a high weight of 0.4. At this time, the texture features (the bedding texture of the rock strata, the disordered texture of the fault fracture zone) and the semantic features (multi-feature association logic) are auxiliary supplements, and their importance is relatively balanced. Therefore, they are given weights of 0.3 respectively, that is, the branch weights of the geological feature attention module are initialized as [0.3,0.4,0.3]. Similarly, if it is necessary to improve the crack recognition accuracy (such as when the accuracy of micro-crack interpretation of the geological visual model is low), the weight of the edge feature branch can be further increased, that is, the branch weights of the geological feature attention module are initialized as [0.2,0.5,0.3], to enhance the geological visual model's ability to capture crack edges.

[0028] Step S22: Embedding the geological feature attention module;

[0029] A geological feature attention module is embedded between the 4th and 5th layers of the Swin Transformer model. The feature map output from the 4th layer is processed by the geological feature attention module and outputs an enhanced feature map with the same dimension. This enhanced feature map is then passed to the 5th layer for further feature extraction. Subsequently, the classification head parameters of the original Swin Transformer model are replaced, and a lithology classification adapter is connected to receive the feature vector output from the last layer of the model. After dimensionality reduction by a fully connected layer, the lithology probability distribution is output through the classification head.

[0030] Step S23: Model training and optimization;

[0031] The preprocessed training set from step S1 is input into the model. Data augmentation is enabled during model training, and the model undergoes iterative training. The iterative changes follow the following pattern: learning rate , t Indicates the first t wheel;

[0032] After each training round, the feature recognition accuracy and lithology classification accuracy are calculated on the validation set. An early stopping strategy is adopted. If the validation set accuracy improves by less than 0.5% for 10 consecutive rounds, the training is terminated early. Finally, the model weight with the highest validation set accuracy is saved, including the basic model parameters, the parameters of the newly added modules and the optimizer state, to construct a large geological vision model.

[0033] Step S24: Model evaluation and optimization;

[0034] The model performance was evaluated on the test set, and the accuracy of the constructed geological visual model was assessed based on feature recognition accuracy, lithology classification accuracy, and single-image inference time.

[0035] Specifically, the evaluation criteria for the accuracy of the geological visual large model in step S24 are as follows:

[0036] Feature recognition accuracy: rock interface ≥90%, fissure ≥85%, fault ≥80%; lithology classification accuracy ≥88%; single image inference time ≤0.5s.

[0037] Specifically, the embedding process of the geological algorithm dictionary in step S3 is as follows:

[0038] Step S31: Construction and encoding of the geological algorithm dictionary;

[0039] The geological algorithm dictionary is constructed by a team of 3-5 senior coal mine geology experts, who combine their own experience in deep coal mining with coal mine geological work standards to sort out key geological laws and feature recognition standards, forming a structured list of geological rules.

[0040] The geological rule list covers three core categories of rules: feature recognition rules, association rules, and anomaly exclusion rules. The feature recognition rules define feature-category judgment criteria based on the image representation and spatial attributes of geological features, used to accurately identify basic geological objects, including rock strata, fissures, and faults. The association rules reveal the inherent coupling relationship between different geological features, used to supplement the geological significance of features. The anomaly exclusion rules define invalid features that do not conform to geological laws, used to purify the interpretation results.

[0041] The encoding process transforms the organized geological rules into production rules that can be executed by a computer and stores them in a database. The algorithm module is developed using the computer programming language Python and includes three core functions: feature parameter extraction function, rule matching function, and result correction function.

[0042] Step S32: Interface development and integration testing;

[0043] Based on a RESTful architecture for representational state transition, two core interfaces are designed: one is the validation interface / api / validate for model calls to the dictionary, which receives feature data in JSON format, including feature type, quantization parameters, and model interpretation confidence; the other is the correction interface / api / correct for model feedback to the dictionary, which returns the correction type, corrected feature parameters, and corrected confidence. The interfaces are developed using the FastAPI framework, support batch data transmission, and generate interface documentation using the Swagger tool to facilitate callers in viewing parameter formats and return value descriptions.

[0044] Construct a simulation test set, which includes model output data covering common geological features and marginal cases. Use the Postman tool to simulate the model calling the / api / validate interface; ensure that the processing time for a single data point is ≤0.5s and the batch processing time is ≤5s to meet the efficiency requirements of real-time interpretation.

[0045] The sample size of the simulated test set is determined as follows: Based on feature type, at least 20-30 sample data points are required for each key feature category; based on sample library size, 3%-5% of the total test set size is used if the sample library contains less than 5000 images, and 0.5%-1% of the total test set size is used if the sample library contains more than 5000 images; based on accuracy requirements, high-precision requires ≥100 sample data points, while rapid verification only requires 50-80 sample data points.

[0046] Step S33: Embedding and integration, and setting trigger conditions;

[0047] The geological visual big data model and the geological algorithm dictionary are deployed on the same GPU server, and a low-latency communication link is established through the internal LAN. After the geological visual big data model completes the initial interpretation of a single borehole image, the system automatically triggers data encapsulation, packages the interpretation results in a preset JSON format, and sends them to the geological algorithm dictionary through the / api / validate interface. After the geological algorithm dictionary completes rule verification and correction, it returns the correction results through the / api / correct interface. After receiving the feedback, the geological visual big data model updates the final interpretation and outputs it in combination with the correction results, while recording correction logs for subsequent iterative optimization of the geological visual big data model.

[0048] The triggering conditions for geological algorithm dictionary verification adopt a hierarchical triggering mechanism to ensure the targeted and efficient use of the dictionary: For low-confidence interpretation results with a model interpretation confidence level below 0.75, geological algorithm dictionary verification is forcibly triggered to avoid the omission of key information due to model misjudgment; For high-risk features, verification is triggered regardless of confidence level, with a focus on checking their consistency with regional geological patterns; For borehole images interpreted for the first time in a new mining area, a full verification mode is initiated, that is, the geological algorithm dictionary is called for all feature types, and the region-specific rules in the geological algorithm dictionary are used to quickly adapt to the new scenario; In addition, trigger exemption conditions are set: For regular features that have been historically verified and whose model interpretation accuracy is stable at over 95%, geological algorithm dictionary verification is skipped to reduce redundant calculations and improve overall interpretation efficiency.

[0049] Furthermore, the formulas for the feature parameter extraction function, rule matching function, and result correction function are expressed as follows:

[0050] Feature parameter extraction functions include:

[0051] Crack length extraction function;

[0052] Let the pixel coordinates of the starting point of the crack in the image be... The endpoint pixel coordinates are The image resolution is r Unit: pixels / m, then the crack length ;

[0053] Rock strata interface grayscale difference extraction function;

[0054] Let the average gray values ​​on the left and right sides of the rock stratum interface be respectively , Then the grayscale difference ;

[0055] Crack density extraction function;

[0056] Crack density 1m represents the unit length, and N is the number of cracks;

[0057] Rule matching function:

[0058] Let the threshold vector of a certain rule in the rule base be... The feature parameter vector output by the model is Then the rule matching degree ,in, To avoid the denominator being 0; A match is considered successful when S≥0.8;

[0059] Result correction function:

[0060] Let the initial confidence level of the model be... If the rule matching degree is S, the feature type correction coefficient is K, the correct category K=1, and the incorrect category K=0.3, then the corrected confidence score is... .

[0061] Specifically, in step S4, the AI ​​module enables adaptive diagnosis and iterative learning of samples, as follows:

[0062] Step 41: Sample adaptive diagnosis;

[0063] First, the training set samples preprocessed in step S1 are feature-quantized and input into the geological vision big data model. The feature vectors output from the penultimate layer of the geological vision big data model are extracted as the digital representation of the samples. Based on the geological feature attention module embedded in the geological vision big data model, an anomaly detection model is constructed using the DeepMineVision analysis algorithm. By calculating the outlier degree of the sample feature vectors, suspected anomaly samples are marked. Then, an anomaly sample list containing sample ID, anomaly type, and feature visualization results is generated. Next, the anomaly samples are corrected using an automatic correction module: multiple lightweight models are called to perform secondary interpretation of the samples. If the annotations of two or more lightweight models are consistent, the majority result is used as the correction. Based on the rules in the geological algorithm dictionary, the interpretation results of multiple lightweight models are verified twice to avoid incorrect corrections due to feature extraction errors. Then, the quality filtering module performs quantitative evaluation on the feature-blurred samples in the abnormal samples and automatically removes invalid data: the gradient entropy algorithm is used to evaluate the degree of image detail retention. The lower the gradient entropy, the more blurred the image. If the signal-to-noise ratio (SNR) is <15dB, it is judged as a completely invalid sample. Samples that meet both gradient entropy <0.3 and SNR <15dB are directly removed from the sample library without entering the correction process. Finally, the corrected sample library is re-divided into training and validation sets according to the original ratio, so that the sample purity is improved to over 95%, laying a high-quality data foundation for subsequent model training.

[0064] Step 42: Study repeatedly;

[0065] Borehole image samples corresponding to low-confidence interpretation results are collected. When the geological vision model interprets new borehole images, interpretation results with a confidence level below 0.7 are selected, and the corresponding image regions are extracted and a pseudo-label generation module is activated. Combining the rules of the geological algorithm dictionary and the feature distribution of high-confidence samples, feature matching is performed on low-confidence regions to generate pseudo-labels, forming an incremental sample set. Subsequently, an incremental training strategy is configured: mini-batch gradient descent is adopted, with the initial learning rate set to 1 / 10 of that in the pre-training stage, decaying by 50% every 5 rounds. At the same time, incremental samples are given a loss weight of 2 times to enhance the model's attention to new features. Finally, the learning process is controlled through iterative evaluation: after each round of learning, the feature recognition accuracy IOU is calculated on the validation set. If the accuracy improvement is less than 0.5% for 3 consecutive rounds or the cumulative learning reaches 20 rounds, the current round of learning is terminated and the optimal weight is saved.

[0066] Compared with the prior art, the present invention has the following beneficial effects:

[0067] 1. Completely eliminate reliance on manual interpretation, achieve efficient processing of borehole images from acquisition to interpretation, and meet the needs of deep coal mines for "rapid detection and timely decision-making".

[0068] 2. Significantly reduces reliance on large-scale labeled samples. By dynamically optimizing sample quality and model parameters, it achieves "rapid iteration with small samples," significantly improving the model's adaptability to geological differences in different mining areas.

[0069] 3. Break through the bottlenecks of general AI models in terms of geological expertise and adaptability to complex scenarios, and achieve accurate identification of features such as rock strata interfaces, fissures, and faults. The interpretation results are more in line with the professional needs of coal mine geological work, and provide reliable data support for roof stability assessment and gas risk early warning. Attached Figure Description

[0070] Figure 1 This is a flowchart of a borehole camera intelligent interpretation method based on a large geological visual model according to the present invention;

[0071] Figure 2 This is a comparison diagram of the interpretation results of fault fracture zones by the method of the present invention and the manual interpretation method in an embodiment of the present invention. Detailed Implementation

[0072] To facilitate understanding and implementation of the present invention by those skilled in the art, the various steps of the method proposed in this invention are described in detail below. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various modifications or alterations to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0073] Example

[0074] like Figure 1 As shown, this invention discloses an intelligent interpretation method for borehole camera images based on a large geological visual model, comprising the following steps:

[0075] Step S1: Acquire borehole images and perform preprocessing to construct a sample library with geological labels;

[0076] Borehole images were acquired and denoised and enhanced to obtain a sample library with geological labels. The sample library was divided into training set, validation set and test set according to a certain ratio.

[0077] Step S2: Based on the existing general-purpose visual model, embed the geological feature attention module and lithology classification adapter, combine them with the geological features of deep coal mines for special optimization, construct a geological visual model, train and validate the constructed geological visual model, and evaluate the accuracy of the geological visual model.

[0078] Step S3: Transform geological knowledge and experience into computable algorithm modules, construct a geological algorithm dictionary and embed it into a large geological visual model;

[0079] Step S4: Implement adaptive diagnosis and iterative learning of samples through the AI ​​module.

[0080] Specifically, in step S1, the image is denoised and enhanced to obtain a sample library containing geological labels, as detailed below:

[0081] In this embodiment, a GD3Q-GCX type digital camera mining borehole imager is selected to perform borehole imaging on the deep coal seam working face to obtain high-definition borehole images. Drilling parameters (such as borehole diameter, construction date, and water inflow within the borehole) and environmental data (such as underground temperature and humidity) are recorded simultaneously as auxiliary information for subsequent interpretation. A sufficient number of borehole wall images (1920×1080 resolution) are acquired and denoised. This involves using Gaussian filtering to smooth image noise and median filtering to remove noise caused by particles within the borehole. Enhancement processing utilizes adaptive histogram equalization (CLAHE). The algorithm performs local contrast adjustment on image blocks to enhance the feature differences between dark areas (such as shadows on rock strata interfaces) and bright areas (such as the walls of reflective holes), improving the recognizability of rock strata boundaries by 40%. Next, it labels rock strata interfaces, fissures, and faults in the image. For rock strata interfaces, a combination of gray-level gradient and edge tracking algorithms is used. Interface locations are determined by calculating the abrupt changes in the first derivative of pixel gray levels, and then Hough transform is used to fit straight lines on the interfaces to mark the boundaries of different lithologies. The algorithm records the interface depth and dip angle, covering different morphologies such as horizontal bedding and wavy bedding. For fissures, edge detection is used... After edge extraction using the Canny operator, noise is removed using morphological operations (erosion, dilation). Then, connected component analysis is used to calculate fracture length, orientation, and density, and classification is performed. For faults, a gray-scale variance-based fault identification method is used. When the gray-scale variance of a local area exceeds a preset threshold and the continuous gray-scale length of a local area exceeds a preset threshold, it is identified as a fault, and the top and bottom boundaries, displacement, and dip angle of the fault fracture zone are marked. This results in a coal mine geological sample library with geological labels. The sample library is divided into training, validation, and test sets according to a certain ratio for subsequent model testing and optimization.

[0082] The number of borehole images is less than 5000, and the sample library is divided into training set, validation set and test set in a 7:2:1 ratio;

[0083] The sample library contains 5,000 to 50,000 borehole images, divided into training, validation, and test sets in a 6:3:1 ratio.

[0084] The sample library contains more than 50,000 borehole images, divided into training, validation, and test sets in a 5:4:1 ratio.

[0085] Specifically, the process of constructing the large-scale geological visual model in step S2 is as follows:

[0086] Step S21: Loading and initializing the basic model;

[0087] Based on the Swin Transformer model, the swin-base version was selected. The parameters of the first 6 layers of the Swin Transformer model were frozen to retain the general feature extraction capability. The parameters of the last 4 layers and the classification head were unfrozen to adjust for the coal mine scenario. The parameters of the newly added modules were initialized: the branch weights of the geological feature attention module were initialized, and the fully connected layer of the lithology classification adapter was initialized with He.

[0088] Step S22: Embedding the geological feature attention module;

[0089] A geological feature attention module is embedded between the 4th and 5th layers of the Swin Transformer model. The feature map output from the 4th layer is processed by the geological feature attention module and outputs an enhanced feature map with the same dimension. This enhanced feature map is then passed to the 5th layer for further feature extraction. Subsequently, the classification head parameters of the original Swin Transformer model are replaced, and a lithology classification adapter is connected to receive the feature vector output from the last layer of the model. After dimensionality reduction by a fully connected layer, the lithology probability distribution is output through the classification head.

[0090] The geological feature attention module strengthens the weight of key features such as rock texture and fracture edges by weighting coefficients, adds a lithology classification adapter before the output layer, and designs a dedicated classification head for common coal mine lithologies including sandstone, mudstone and coal seam. It improves classification sensitivity by increasing the training data of lithology samples by 10%, and achieves specific optimization of geological features.

[0091] Step S23: Model training and optimization;

[0092] The preprocessed training set from step S1 is input into the model. Data augmentation is enabled during model training, and the model undergoes iterative training. The iterative changes follow the following pattern: learning rate , t Indicates the first t wheel;

[0093] After each training round, the feature recognition accuracy and lithology classification accuracy are calculated on the validation set (IOU ≥ 0.5 is considered correct). An early stopping strategy is adopted. If the validation set accuracy improves by < 0.5% for 10 consecutive rounds, the training is terminated early. Finally, the model weight with the highest validation set accuracy is saved, including the basic model parameters, the parameters of the newly added modules and the optimizer state, to construct a large geological vision model.

[0094] Step S24: Model evaluation and optimization;

[0095] The model performance was evaluated on the test set, and the accuracy of the constructed geological visual model was assessed based on feature recognition accuracy, lithology classification accuracy, and single-image inference time.

[0096] Specifically, the evaluation criteria for the accuracy of the geological visual large model in step S24 are as follows:

[0097] Feature recognition accuracy: rock interface ≥90%, fissure ≥85%, fault ≥80%; lithology classification accuracy ≥88%; single image inference time ≤0.5s.

[0098] Specifically, the embedding process of the geological algorithm dictionary in step S3 is as follows:

[0099] Step S31: Construction and encoding of the geological algorithm dictionary;

[0100] The geological algorithm dictionary is constructed by a team of 3-5 senior coal mine geology experts, who combine their own experience in deep coal mining with coal mine geological work standards to sort out key geological laws and feature recognition standards, forming a structured list of geological rules.

[0101] The geological rule list covers three core categories of rules: feature recognition rules, association rules, and anomaly exclusion rules. Feature recognition rules define feature-category judgment criteria based on the image representation and spatial attributes of geological features, used to accurately identify basic geological objects such as strata, fissures, and faults (e.g., if a continuous black band appears within 3m of the coal seam roof and the grayscale value is below 50, it is judged as pseudo-roof mudstone). Association rules reveal the inherent coupling relationship between different geological features (e.g., the correlation between structure and stress, fissures and gas), used to supplement the geological significance of features (e.g., when the fissure density exceeds 8 fissures / m and the distance from the coal seam is less than 5m, the gas outburst risk level increases). The anomaly exclusion rules define invalid features that do not conform to geological laws (such as labeling errors or image noise) to purify the interpretation results (e.g., a single fracture exceeding 5m in length but without branching extension and without abrupt changes in lithology around it is judged as a misidentification caused by image noise). Simultaneously, this embodiment also clarifies the quantitative parameter standards for geological features. For example, fracture density is divided into: micro-fractures (<3 fractures / m), medium-sized fractures (3-8 fractures / m), and dense fractures (>8 fractures / m); rock layer interfaces are defined with a grayscale variance ≥30 and a continuous grayscale length ≥0.3m as the effective judgment threshold; faults are classified according to elevation difference into: small faults (<5m), interrupted faults (5-10m), and large faults (>10m).

[0102] The encoding process transforms the organized geological rules into production rules (IF-THEN format) that can be executed by a computer. For example: IF lithology = fine-grained sandstone AND fracture strike concentrated at 30°-60° AND dip angle > 70° THEN associated structure is a normal fault. These rules are stored in the rules table of a MySQL database. Each rule includes fields such as rule ID, applicable lithology, triggering conditions, and conclusion. The parameter library is stored in the parameters table of the database in key-value pairs, supporting fast querying by feature type. The algorithm module is developed in Python and includes three core functions: feature parameter extraction function, rule matching function, and result correction function.

[0103] Step S32: Interface development and integration testing;

[0104] Based on a RESTful architecture for representational state transition, two core interfaces are designed: First, the model calls the dictionary for validation via the interface ` / api / validate`, which receives feature data in JSON format, including feature type (fracture, fault), quantization parameters (length, orientation, density, etc.), and model interpretation confidence. Second, the model corrects the dictionary feedback via the interface ` / api / correct`, returning results including correction type (category correction, parameter adjustment), corrected feature parameters, and corrected confidence. The interfaces are developed using the FastAPI framework, supporting batch data transmission, and interface documentation is generated using Swagger to facilitate users' understanding of parameter formats and return value descriptions.

[0105] A simulation test set was constructed, containing model output data covering common geological features (such as medium-sized fractures in sandstone and faults in coal seam roofs) and edge cases (such as suspected fractures in blurred images). The model calls the / api / validate interface were simulated using the Postman tool. The testing focused on three aspects: First, functional verification, checking whether the dictionary could correctly parse the input data and match the corresponding rules (such as calling the medium-sized fracture classification rule for a fracture with a length of 1.2m and a strike of 30°); second, logical verification, verifying the rationality of the correction results (such as correcting a dense fracture zone that the model misclassified as a fault to a fracture development zone); and third, performance testing, recording the interface response time to ensure that the processing time for a single data item is ≤0.5s and the batch processing time is ≤5s, meeting the efficiency requirements of real-time interpretation.

[0106] The sample size of the simulated test set is determined as follows: Based on feature type, at least 20-30 sample data points are required for each key feature category; based on sample library size, 3%-5% of the total test set size is used if the sample library contains less than 5000 images, and 0.5%-1% of the total test set size is used if the sample library contains more than 5000 images; based on accuracy requirements, high-precision requires ≥100 sample data points, while rapid verification only requires 50-80 sample data points.

[0107] Step S33: Embedding and integration, and setting trigger conditions;

[0108] The geological visual big data model and the geological algorithm dictionary are deployed on the same high-performance GPU server (such as a device equipped with an NVIDIA A100 graphics card), and a low-latency communication link is established through an internal LAN (network latency controlled within 10ms). The geological algorithm dictionary call step is embedded in the inference process of the geological visual big data model to form a closed-loop workflow: after the geological visual big data model completes the initial interpretation of a single borehole image (outputting feature type, parameters, and confidence level), the system automatically triggers data encapsulation, packages the interpretation results in a preset JSON format, and sends them to the geological algorithm dictionary through the / api / validate interface; after the geological algorithm dictionary completes rule validation and correction, it returns the correction results (including the corrected feature category, parameter adjustment value, and correction basis) through the / api / correct interface; after receiving the feedback, the geological visual big data model updates the final interpretation output based on the correction results, and records the correction log (including the original result, correction content, and rule matching ID) for subsequent model iteration and optimization.

[0109] In this embodiment, a hierarchical triggering mechanism is adopted for the geological algorithm dictionary verification to ensure the targeted and efficient use of the geological algorithm dictionary: For low-confidence interpretation results with a model interpretation confidence level below 0.75, geological algorithm dictionary verification is forcibly triggered to avoid the omission of key information due to model misjudgment; For high-risk features (such as faults, dense fractures, and coal seam-roof contact zones), verification is triggered regardless of the confidence level, with a focus on checking their consistency with regional geological patterns (such as whether the fault strike matches the main structural line of the mine); For borehole images interpreted for the first time in a new mining area, a full verification mode is initiated, that is, the geological algorithm dictionary is called for all feature types, and the regional specific rules in the geological algorithm dictionary (such as the fact that sandstone fractures in a certain mining area are mostly oriented northeast) are used to quickly adapt to the new scenario.

[0110] In addition, trigger exemption conditions are set: for routine features (such as the texture features of typical mudstone) that have been historically verified and whose model interpretation accuracy is stable at over 95%, geological algorithm dictionary verification can be skipped to reduce redundant calculations and improve overall interpretation efficiency.

[0111] Furthermore, the formulas for the feature parameter extraction function, rule matching function, and result correction function are expressed as follows:

[0112] Feature parameter extraction functions include:

[0113] Crack length extraction function;

[0114] Let the pixel coordinates of the starting point of the crack in the image be... The endpoint pixel coordinates are The image resolution isr Unit: pixels / m, then the crack length ;

[0115] Rock strata interface grayscale difference extraction function;

[0116] Let the average gray values ​​on the left and right sides of the rock stratum interface be respectively , Then the grayscale difference ;

[0117] Crack density extraction function;

[0118] Crack density 1m represents the unit length, and N is the number of cracks;

[0119] Rule matching function:

[0120] To determine the degree of matching between the model's output features and rules in the rule base, a similarity calculation is used: feature parameter - rule threshold. Let the threshold vector of a rule in the rule base be... The feature parameter vector output by the model is Then the rule matching degree ,in, To avoid the denominator being 0; A match is considered successful when S≥0.8;

[0121] Result correction function:

[0122] Let the initial confidence level of the model be... If the rule matching degree is S, the feature type correction coefficient is K, the correct category K=1, and the incorrect category K=0.3, then the corrected confidence score is... .

[0123] Specifically, in step S4, the AI ​​module enables adaptive diagnosis and iterative learning of samples, as follows:

[0124] Step 41: Sample adaptive diagnosis;

[0125] First, the training set samples preprocessed in step S1 are quantized and input into the geological vision big data model. The feature vectors output from the penultimate layer of the geological vision big data model are extracted as digital representations of the samples. Based on the geological feature attention module embedded in the geological vision big data model, an anomaly detection model is constructed using the DeepMineVision analysis algorithm (anomaly ratio threshold is set to 6%). By calculating the outlier degree of the sample feature vectors, suspected anomalous samples are marked—including blurred images caused by water accumulation in the borehole and equipment vibration (high feature vector dispersion), conflicting samples misjudged by annotators (such as simultaneously annotating fractures and rock strata interfaces in the same area), and incomplete samples missing key geological features (such as faults). Then, an anomalous sample list containing sample IDs, anomaly types, and feature visualization results is generated. Next, the anomalous samples are corrected using an automatic correction module: multiple lightweight models (obtained based on the parameters of the big data model, each focusing on different aspects) are called. (Lithology, fracture, and fault identification) The samples undergo secondary interpretation. If the annotations of two or more lightweight models are consistent, the majority result is used as the correction basis. Combined with the rules in the geological algorithm dictionary, the interpretation results of multiple lightweight models are verified a second time to avoid incorrect correction due to feature extraction errors. Then, the quality filtering module performs quantitative evaluation on the feature-blurred samples in the abnormal samples and automatically removes invalid data: The gradient entropy algorithm (gradient entropy < 0.3 is considered blurry) is used to evaluate the degree of image detail retention. The lower the gradient entropy, the more blurry the image. If the signal-to-noise ratio (SNR) is < 15dB, it is judged as a completely invalid sample. Samples that meet both gradient entropy < 0.3 and SNR < 15dB are directly removed from the sample library without entering the correction process, ensuring that the retained samples truly reflect the geological features. Finally, the corrected sample library is re-divided into training and validation sets according to the original ratio, so that the sample purity is improved to over 95%, laying a high-quality data foundation for subsequent model training.

[0126] Step 42: Study repeatedly;

[0127] Collect borehole image samples corresponding to low-confidence interpretation results. When the geological vision model interprets new borehole images, it filters out interpretation results with a confidence level below 0.7 (such as suspected micro-fractures or areas with uncertain lithology), extracts the corresponding image regions, and starts the pseudo-label generation module: combining the rules of the geological algorithm dictionary (such as the correlation between fracture length and strike) and the feature distribution of high-confidence samples, it performs feature matching on low-confidence areas to generate pseudo-labels, forming an incremental sample set (the number of new samples in each batch is controlled at 5%-10% of the original training set to avoid drastic fluctuations in data distribution); Next, an incremental training strategy was configured: the parameters of the first 6 layers of the model were frozen to retain the general feature extraction capability, while only the parameters of the last 4 layers and the classification head were unfrozen; mini-batch gradient descent was adopted, with the initial learning rate set to 1 / 10 of that in the pre-training stage, decaying by 50% every 5 rounds, and the incremental samples were given double the loss weight to enhance the model's attention to new features; finally, the learning process was controlled through iterative evaluation: after each round of learning, the feature recognition accuracy IOU was calculated on the validation set, and if the accuracy improvement was less than 0.5% for 3 consecutive rounds or the cumulative learning reached 20 rounds, the current round of learning was terminated and the optimal weights were saved.

[0128] Through the above process, the geological visual big model can quickly absorb new geological feature information while retaining existing knowledge, achieve dynamic adaptation to different mining area scenarios, and shorten the debugging cycle of new mining area models from several months to 1-2 weeks.

[0129] The following experiment uses the 12212 and 12213 working faces of Xinyuan Coal Mine in a county of Shanxi Province as examples to verify the feasibility and accuracy of the method of the present invention.

[0130] This mine is a high-gas mine with a complex roof stratum structure and localized areas with fractures and faults, which significantly impacts drilling operations and gas extraction.

[0131] In this example, seven directional boreholes were selected as test objects, with a borehole depth ranging from 70m. The borehole walls were imaged using a GD3Q-GCX mining digital borehole imager, resulting in 9860 borehole images with a resolution of 1920×1080. After filtering, noise reduction, and contrast enhancement of the original borehole images, they were manually annotated by geologists. The annotations included rock strata interfaces, fractures, and fault features, ultimately forming 9200 borehole image annotation samples, which were divided into training, validation, and test sets in a 6:3:1 ratio.

[0132] The segmented sample data were input into the geological visual big data model proposed in this invention for interpretation, and the interpretation results were compared with the traditional manual interpretation results. The results are shown in Table 1 below.

[0133] Table 1. Comparison of interpretation results between the geological visual large-scale model of the present invention and the manual interpretation method.

[0134] ;

[0135] As shown in Table 1 above, on the test set, the geological visual large model proposed by the method of this invention has an accuracy rate of 92.7% for rock layer interface identification, 86.4% for fracture identification, 82.1% for fault identification, and 89.3% for lithology classification. The average processing time per image is 0.41s. Compared with the results of manual interpretation, the method of this invention reduces the processing time by about 87.5% under the same data conditions, and improves the accuracy of identification of major geological features by 8% to 12%.

[0136] like Figure 2 As shown, taking borehole #5 out of seven directional boreholes as an example, the manual interpretation result indicates the presence of a fault fracture zone at a depth of 21.2-21.3m; the interpretation result of the method of this invention shows a fault at a depth of 21.1-21.3m. The positional errors of the two methods are 0.1m, which are within the allowable engineering error range, indicating that both the method of this invention and the manual interpretation method can obtain accurate interpretation results of the fault fracture zone.

[0137] To further illustrate the application effect of the method of the present invention, a comparison of the construction cycle of working face 12213 and control working face 12212 was made: Without the method of the present invention, the high-level directional drilling layout of the roof lacked precise stratigraphic basis, resulting in a long construction cycle. The average monthly footage per drilling rig was approximately 362.4 m, and the drilling site construction cycle exceeded 4 months. Furthermore, the gas extraction effect was poor, with the cumulative extraction volume of 11 boreholes being only 0.01–0.33 m³ / min, and the gas concentration in the upper corner remaining at 0.62%–0.74%. However, after optimizing the borehole layout using the method of the present invention on working face 12213, 8 ineffective boreholes were reduced, decreasing the drilling workload by approximately 3012 m. Drilling efficiency was improved, with the average monthly footage per drilling rig reaching 795 m. The m value was approximately 119% higher than that of the control face 12212; the average borehole extraction concentration increased by 12.5%, and the gas concentration in the upper corner remained stable at 0.37%~0.45% during the mining period, which was significantly lower than that of the control face.

[0138] In summary, this example verifies the applicability of the method of the present invention in intelligent interpretation of borehole images through actual mine test data; the comparison results with manual interpretation methods show that the method of the present invention can effectively improve the interpretation accuracy and efficiency of coal mine borehole images; the comparison results with control working faces show that the method of the present invention has significant advantages in optimizing borehole layout, shortening construction cycle, and improving gas extraction efficiency.

[0139] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for intelligent interpretation of borehole camera images based on a large geological visual model, characterized in that, Includes the following steps: Step S1: Acquire borehole images and perform preprocessing to construct a sample library with geological labels; Borehole images were acquired and denoised and enhanced to obtain a sample library with geological labels. The sample library was divided into training set, validation set and test set according to a certain ratio. Step S2: Based on the existing general-purpose visual model, embed the geological feature attention module and lithology classification adapter, combine them with the geological features of deep coal mines for special optimization, construct a geological visual model, train and validate the constructed geological visual model, and evaluate the accuracy of the geological visual model. The construction of the large-scale geological visual model includes: Geological feature attention module embedding; A geological feature attention module is embedded between the 4th and 5th layers of the Swin Transformer model. The feature map output from the 4th layer is processed by the geological feature attention module and outputs an enhanced feature map with the same dimension. This enhanced feature map is then passed to the 5th layer for further feature extraction. Subsequently, the classification head parameters of the original Swin Transformer model are replaced, and a lithology classification adapter is connected to receive the feature vector output from the last layer of the model. After dimensionality reduction by a fully connected layer, the lithology probability distribution is output through the classification head. Model training and optimization; Input the preprocessed training set from step S1 into the model, enable data augmentation during model training, and iterate the training of the model. Step S3: Construct a geological algorithm dictionary to transform geological knowledge and experience into computational algorithm modules and embed them into a large geological visual model; The embedding of the geological algorithm dictionary includes: Embedding integration and trigger condition settings; The geological visual big data model and the geological algorithm dictionary are deployed on the same GPU server, and a low-latency communication link is established through the internal LAN. After the geological visual big data model completes the initial interpretation of a single borehole image, the system automatically triggers data encapsulation, packages the interpretation results in a preset JSON format, and sends them to the geological algorithm dictionary through the / api / validate interface. After the geological algorithm dictionary completes rule verification and correction, it returns the correction results through the / api / correct interface. After receiving the feedback, the geological visual big data model updates the final interpretation and outputs it in combination with the correction results, while recording correction logs for subsequent iterative optimization of the geological visual big data model. Step S4: Implement sample adaptive diagnosis and iterative learning through the AI ​​module, including: Sample adaptive diagnosis; First, the training set samples preprocessed in step S1 are quantized for features and input into the geological vision big model. The feature vectors output from the penultimate layer of the geological vision big model are extracted as digital representations of the samples. Based on the geological feature attention module embedded in the geological vision big model, an anomaly detection model is constructed using the DeepMineVision analysis algorithm. By calculating the outlier degree of the sample feature vectors, suspected anomalous samples are marked. Then, an anomalous sample list containing sample ID, anomaly type, and feature visualization results is generated. Finally, the anomalous samples are corrected using an automatic correction module.

2. The intelligent interpretation method for borehole camera images based on a large geological visual model according to claim 1, characterized in that, In step S1, the images are denoised and enhanced to obtain a sample library containing geological labels, as detailed below: Denoising: Gaussian filtering is used to smooth the noise in the borehole image, and median filtering is used to remove noise caused by particles inside the borehole; Enhancement processing: The adaptive histogram equalization CLAHE algorithm is used to perform block-based local contrast adjustment on the borehole image, enhance the feature difference between dark and bright areas, and improve the identification of rock strata boundaries; The rock strata interfaces, fractures, and faults in the denoised and enhanced borehole images are labeled to obtain a sample library with geological labels: For rock strata interfaces, a combined algorithm of gray-level gradient and edge tracking is used. The interface position is located by calculating the abrupt change point of the first derivative of pixel gray-level. Then, Hough transform is used to fit the interface line to mark the boundary position of different lithologies. The interface burial depth, dip angle, and coverage of horizontal bedding and wavy bedding are recorded. For fractures, the edges are extracted by the Canny edge detection operator, and noise is removed by morphological operations. Then, the fracture length, strike, and density are calculated by connected component analysis and classified and labeled. For faults, a gray-level variance identification method based on fracture zone is used. When the gray-level variance of a local area is greater than the preset threshold of gray-level variance and the gray-level continuous length of a local area is greater than the preset threshold of gray-level continuous length, it is determined to be a fault and the top and bottom boundaries, drop, and dip angle of the fault fracture zone are marked.

3. The intelligent interpretation method for borehole camera images based on a large geological visual model according to claim 1, characterized in that, The process of constructing the large-scale geological visual model in step S2 is as follows: Step S21: Loading and initializing the basic model; Based on the Swing Transformer model, the parameters of the first 6 layers of the Swing Transformer model are frozen to retain the general feature extraction capability, and the parameters of the last 4 layers and the classification head are unfrozen to adjust for the coal mine scenario; the parameters of the newly added modules are initialized: the branch weights of the geological feature attention module are initialized, and the fully connected layer of the lithology classification adapter is initialized with He. Step S22: Embedding the geological feature attention module; A geological feature attention module is embedded between the 4th and 5th layers of the Swin Transformer model. The feature map output from the 4th layer is processed by the geological feature attention module and outputs an enhanced feature map with the same dimension. This enhanced feature map is then passed to the 5th layer for further feature extraction. Subsequently, the classification head parameters of the original Swin Transformer model are replaced, and a lithology classification adapter is connected to receive the feature vector output from the last layer of the model. After dimensionality reduction by a fully connected layer, the lithology probability distribution is output through the classification head. Step S23: Model training and optimization; The preprocessed training set from step S1 is input into the model. Data augmentation is enabled during model training, and the model undergoes iterative training. The iterative changes follow the following pattern: learning rate , t Indicates the first t wheel; After each training round, the feature recognition accuracy and lithology classification accuracy are calculated on the validation set. An early stopping strategy is adopted. If the validation set accuracy improves by less than 0.5% for 10 consecutive rounds, the training is terminated early. Finally, the model weight with the highest validation set accuracy is saved, including the basic model parameters, the parameters of the newly added modules and the optimizer state, to construct a large geological vision model. Step S24: Model evaluation and optimization; The model performance was evaluated on the test set, and the accuracy of the constructed geological visual model was assessed based on feature recognition accuracy, lithology classification accuracy, and single-image inference time.

4. The intelligent interpretation method for borehole camera images based on a large geological visual model according to claim 3, characterized in that, The criteria for evaluating the accuracy of the geological visual large model in step S24 are as follows: Feature recognition accuracy: rock interface ≥90%, fissure ≥85%, fault ≥80%; lithology classification accuracy ≥88%; single image inference time ≤0.5s.

5. The intelligent interpretation method for borehole camera images based on a large geological visual model according to claim 1, characterized in that, The embedding process of the geological algorithm dictionary in step S3 is as follows: Step S31: Construction and encoding of the geological algorithm dictionary; The geological algorithm dictionary is constructed by a team of 3-5 senior coal mine geology experts, who combine their own experience in deep coal mining with coal mine geological work standards to sort out key geological laws and feature recognition standards, forming a structured list of geological rules. The encoding process involves converting the organized geological rules into production rules that can be executed by a computer and storing them in a database. The algorithm module is developed using the Python programming language and contains three core functions: feature parameter extraction function, rule matching function, and result correction function. Step S32: Interface development and integration testing; Based on the RESTful architecture of representational state transition, two core interfaces are designed: one is the validation interface / api / validate for model calls to the dictionary, which receives feature data in JSON format, including feature type, quantization parameters and model interpretation confidence; the other is the correction interface / api / correct for model feedback to the dictionary, which returns the correction type, corrected feature parameters and corrected confidence. A simulation test set was constructed, which included model output data covering common geological features and marginal cases. The Postman tool was used to simulate the model calling the / api / validate interface to ensure that the data processing time met the efficiency requirements of real-time interpretation. Step S33: Embedding and integration, and setting trigger conditions; The geological visual big model and the geological algorithm dictionary are deployed on the same GPU server and a low-latency communication link is established through the internal local area network. After the geological visual big model completes the initial interpretation of a single borehole image, the system automatically triggers data encapsulation, packages the interpretation results in a preset JSON format, and sends them to the geological algorithm dictionary through the / api / validate interface. After the geological algorithm dictionary completes rule verification and correction, it returns the correction results through the / api / correct interface; after receiving the feedback, the geological vision big model updates the final interpretation and outputs it in combination with the correction results, and records the correction log for subsequent iterative optimization of the geological vision big model; The triggering conditions for geological algorithm dictionary verification adopt a hierarchical triggering mechanism. For low-confidence interpretation results with a model interpretation confidence level below 0.75, geological algorithm dictionary verification is forcibly triggered to avoid the omission of key information due to model misjudgment. For high-risk features, verification is triggered regardless of confidence level, with a focus on checking their consistency with regional geological patterns. For borehole images interpreted for the first time in a new mining area, a full verification mode is initiated, that is, the geological algorithm dictionary is called for all feature types, and the region-specific rules in the geological algorithm dictionary are used to quickly adapt to the new scenario. In addition, trigger exemption conditions are set: for conventional features that have been historically verified and whose model interpretation accuracy is stable at over 95%, geological algorithm dictionary verification is skipped.

6. The intelligent interpretation method for borehole camera images based on a large geological visual model according to claim 5, characterized in that, The formulas for the feature parameter extraction function, rule matching function, and result correction function are expressed as follows: Feature parameter extraction functions include: Crack length extraction function; Let the pixel coordinates of the starting point of the crack in the image be... The endpoint pixel coordinates are The image resolution is r Unit: pixels / m, then the crack length ; Rock strata interface grayscale difference extraction function; Let the average gray values ​​on the left and right sides of the rock stratum interface be respectively , Then the grayscale difference ; Crack density extraction function; Crack density 1m represents the unit length, and N is the number of cracks; Rule matching function: Let the threshold vector of a certain rule in the rule base be... The feature parameter vector output by the model is Then the rule matching degree ,in, To avoid the denominator being 0; A match is considered successful when S≥0.8; Result correction function: Let the initial confidence level of the model be... If the rule matching degree is S, the feature type correction coefficient is K, the correct category K=1, and the incorrect category K=0.3, then the corrected confidence score is... .

7. The intelligent interpretation method for borehole camera images based on a large geological visual model according to claim 1, characterized in that, In step S4, the AI ​​module enables adaptive diagnosis and iterative learning of samples, as follows: Step 41: Sample adaptive diagnosis; First, the training set samples preprocessed in step S1 are feature-quantized and input into the geological vision big data model. The feature vectors output from the penultimate layer of the geological vision big data model are extracted as the digital representation of the samples. Based on the geological feature attention module embedded in the geological vision big data model, an anomaly detection model is constructed using the DeepMineVision analysis algorithm. By calculating the outlier degree of the sample feature vectors, suspected anomaly samples are marked. Then, an anomaly sample list containing sample ID, anomaly type, and feature visualization results is generated. Next, the anomaly samples are corrected through an automatic correction module: multiple lightweight models are called to perform secondary interpretation of the samples. If two or more lightweight models are used, the automatic correction module will correct the samples. If the annotations of the models are consistent, the majority results are used as the basis for correction. Combined with the rules in the geological algorithm dictionary, the interpretation results of multiple lightweight models are verified a second time to avoid incorrect corrections due to feature extraction errors. Then, the quality filtering module is used to quantitatively evaluate the feature blurry samples in the abnormal samples and automatically remove invalid data. The gradient entropy algorithm is used to evaluate the degree of image detail preservation. The lower the gradient entropy, the blurrier the image. If the signal-to-noise ratio (SNR) is <15dB, it is determined to be a completely invalid sample. Samples that meet both gradient entropy <0.3 and SNR <15dB are directly removed from the sample library without entering the correction process. Finally, the corrected sample library is re-divided into training and validation sets according to the original proportions. Step 42: Study repeatedly; Borehole image samples corresponding to low-confidence interpretation results are collected. When the geological vision big model interprets new borehole images, interpretation results with a confidence level below 0.7 are selected, and the corresponding image regions are extracted and the pseudo-label generation module is activated. Combining the rules of the geological algorithm dictionary and the feature distribution of high-confidence samples, feature matching is performed on the low-confidence regions to generate pseudo-labels and form an incremental sample set. Subsequently, an incremental training strategy is configured: the mini-batch gradient descent method is adopted, with the initial learning rate set to 1 / 10 of that in the pre-training stage, decaying by 50% every 5 rounds, while the incremental samples are given a loss weight of 2 times to enhance the model's attention to new features. Finally, the learning process is controlled through iterative evaluation: after each round of learning, the feature recognition accuracy IOU is calculated on the validation set. If the accuracy improvement is less than 0.5% for three consecutive rounds or the cumulative learning reaches 20 rounds, the current round of learning is terminated and the optimal weights are saved.

Images