Drilling and blasting method rock mass structure identification method and system based on large model agent

Abstract

The invention provides a drilling and blasting method rock mass structure identification method and system based on a large model agent, and relates to the technical field of thinking chains and agents. The method comprises the following steps: preliminarily identifying geologic features on a tunnel face by using an image segmentation small model, and optimizing a preliminary segmentation result of the geologic features and judging a specific category of a structural surface by using a multi-modal large model based on a corrected thinking chain; further guiding the multi-mode large model to automatically generate a geologic feature occurrence information calculation code through a prompt project, and calculating the geologic feature occurrence information calculation code; based on a fault judgment thinking chain, the multi-modal large model performs fault reasoning judgment by using the optimized segmentation result and occurrence parameters; and generating feature descriptions of different geological features according to the domain knowledge graph, and generating a rock mass structure information text description by optimizing a segmentation result and a fault inference result according to a prompt prompt template. The invention aims to realize full-process intelligent processing from geologic feature recognition, occurrence information calculation, fault judgment and reasoning and professional description generation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of thought chain and intelligent agent, and particularly relates to a drilling and blasting method rock mass structure identification method and system based on a large model intelligent agent. BACKGROUND

[0002] The geological conditions of deep buried tunnel construction are complex, and disasters such as rock burst frequently occur during construction and cause great damage. For the drilling and blasting method, determining whether a fault occurs can provide supporting information for the occurrence of some special geological disasters. The appearance of a small fault usually means that the surrounding rock is unstable, and is prone to induce fracture-type rock burst in high ground stress tunnels.

[0003] A fault is a planar damage or planar rheological zone of rock strata or rock mass in the crust under the action of stress, and the rock blocks on both sides of the fault have obvious displacement. The existing methods mostly simulate and restore the overall geological conditions of the rock mass through geological exploration and numerical simulation, and find faults and other geological features accordingly. These methods help construction personnel to understand the overall geological distribution, but in the construction process, the actual exposed faults on the working face still need to be observed manually, which has certain safety risks.

[0004] However, there is currently no method for intelligent fault identification through working face images. This is mainly because on the working face, the visual features of faults and general structural planes are almost indistinguishable, and the "obvious displacement of rock blocks on both sides of the fault" is an important feature that distinguishes faults from structural planes. This is a reasoning task that needs to find the positional relationship between geological features, and conventional target detection visual models cannot achieve it, with the following two limitations: (1) existing methods cannot distinguish between structural planes and faults on the working face: for faults on the working face, existing methods can only identify them as structural planes, and cannot use the important feature of distinguishing structural planes to make effective judgments. (2) Existing methods cannot give complete and professional geological feature descriptions and criteria: current general large models lack deep engineering professional knowledge and cannot meet the standards of field experts, and cannot give professional feature descriptions for preliminary geological feature recognition results, and also cannot give judgment basis in accordance with the physical mechanism of the faults on the working face.

[0005] In view of this, in order to reduce the dependence on manual work, an automatic and intelligent identification method for rock mass structure information on the working face for drilling and blasting construction is needed. SUMMARY

[0006] In view of the deficiencies of the prior art, the present application observes geological features through an image segmentation small model, introduces a thinking chain to enhance the task processing capability of a multi-modal large model, and further uses a domain knowledge graph to empower the multi-modal large model with deep engineering knowledge, so as to automatically generate a professional description of geological information and assist in identifying the stability of rock mass, thereby proposing a drilling and blasting method and system for rock mass structure identification based on a large model agent, aiming to realize intelligent processing of the whole process from geological feature identification, occurrence information calculation, to fault determination reasoning and professional description generation, to make up for the lack of intelligent identification method for rock mass structure information on the working face of the drilling and blasting method, reduce the dependence of the geological information identification and fault reasoning process on artificial expert experience, improve the construction efficiency, and reduce the labor cost.

[0007] In one aspect, the present application proposes a drilling and blasting method for rock mass structure identification based on a large model agent, which comprises the following processes:

[0008] Obtain the working face image of the engineering site, and use an image segmentation small model to preliminarily identify the geological features of the working face image to obtain the preliminary segmentation result of the working face image;

[0009] Based on the modified thinking chain, input the working face image and the preliminary segmentation result of the working face image into the multi-modal large model, and modify the preliminary segmentation result through iterative optimization to obtain the optimized segmentation result of the working face image;

[0010] Based on the optimized segmentation result of the working face image, prompt the engineering to guide the multi-modal large model to automatically generate the occurrence parameters of the working face image according to the structural surface morphological characteristics;

[0011] Based on the fault determination thinking chain, input the optimized segmentation result of the working face image and the occurrence parameters of the working face image into the multi-modal large model for fault reasoning to obtain the fault inference result of the working face image;

[0012] Based on the optimized segmentation result of the working face image, the multi-modal large model searches the pre-constructed domain knowledge graph to generate feature descriptions of different geological features in the working face image;

[0013] Based on the optimized segmentation result of the working face image, the fault inference result and the feature descriptions of different geological features, the multi-modal large model generates a professional text description of the rock mass structure information on the working face according to a preset prompt template.

[0014] Further, the construction method of the image segmentation small model is:

[0015] Collect a plurality of working face images of the engineering site, and label the geological features of each working face image respectively; wherein the geological features include structural surfaces and dikes;

[0016] Each face image and labeled geological features are taken as a sample, a geological feature image segmentation dataset is constructed, and the geological feature image segmentation dataset is divided into a training set, a test set and a validation set according to a predetermined proportion;

[0017] A visual model is selected as an image segmentation small model, the image segmentation small model is iteratively trained using the training set, and periodic testing is performed on the test set during the training process. After training, the trained image segmentation small model is verified using the validation set, and the final image segmentation small model is obtained.

[0018] For any face image, the face image is input into the finally obtained image segmentation small model for preliminary identification, and the preliminary segmentation result of the face image is output. The preliminary segmentation result includes: a target detection frame, a segmentation result image and a classification probability of each pixel.

[0019] Further, based on the modified thinking chain, the face image and the preliminary segmentation result of the face image are input into the multi-modal large model, and the preliminary segmentation result is iteratively optimized and corrected to obtain the optimized segmentation result of the face image. The specific content is:

[0020] Based on the preliminary segmentation result of the face image, the segmentation result image is cropped according to the target detection frame to obtain local images of different geological features in the face image. At the same time, the original images corresponding to each local image are cropped in the face image.

[0021] The pixel-level recognition results of each local image are extracted respectively, including: pixel point category, category probability and binary mask of each pixel point; wherein the pixel point category includes: structural plane, dike and background;

[0022] A modified thinking chain is constructed to guide the multi-modal large model to compare and observe the local image and the original image corresponding to the local image, and to correct the pixel point category of the pixel-level recognition result of the local image;

[0023] For each local image, the local image, the original image corresponding to the local image and the pixel-level recognition result of the local image are input into the multi-modal large model, and the pixel point category of the pixel-level recognition result of the local image is corrected based on the modified thinking chain to obtain the corrected segmentation result; wherein the corrected segmentation result includes: segmentation result visualization image and final mask.

[0024] extract the structural surface from the corrected segmentation result, and input the original image corresponding to the structural surface into the multi-modal large model, prompt the multi-modal large model to identify the category of the structural surface by giving the geological feature description of different structural surface categories, and obtain the corrected structural surface category; wherein the structural surface category includes: hard structural surface and weak structural surface;

[0025] Splice the corrected segmentation result and the corrected structural surface category corresponding to all local images respectively to generate the optimized segmentation result of the working face image, and write the corrected structural surface category and the head and tail pixel point coordinate values of each structural surface into a JSON file.

[0026] Further, the specific content of the correction thought chain is:

[0027] Set a probability threshold and a maximum number of iterations, and initialize the current iteration round to 0. For any group of input local images, the original image corresponding to the local image, and the pixel-level recognition result of the local image, execute the following steps:

[0028] Step A1: Determine whether the current mask M has redundancy by comparing the local image and the original image. If not, the current mask M is taken as the final mask and step A4 is executed; if yes, step A2 is executed;

[0029] Step A2: Extract all boundary pixels of the target region from the current mask , and respectively correct the pixel point category of each boundary pixel to obtain the updated mask ; determine whether the updated mask is completely the same as the current mask , if yes, the updated mask is taken as the final mask and step A4 is executed; if not, step A3 is executed;

[0030] The target region refers to the set of connected regions formed by all pixel points with a pixel value of 1 in the current mask .

[0031] For any boundary pixel , if the category probability of the boundary pixel does not exceed the probability threshold, the pixel point category of the boundary pixel is modified to background; if the category probability of the boundary pixel exceeds the probability threshold, the pixel point category of the boundary pixel is not changed.

[0032] Step A3: superimpose the updated mask on the original image to generate a revised visualization, increment the current iteration round by 1, and determine whether the current iteration round reaches the maximum iteration number. If yes, the updated mask is taken as the final mask , and step A4 is executed; if no, the revised visualization is taken as the new local image, the updated mask is taken as the new current mask , and step A1 is returned.

[0033] Step 4: superimpose the final mask on the original image to obtain a segmentation result visualization, and take the segmentation result visualization and the final mask as the final output.

[0034] Further, the specific content of prompting the engineering to guide the multi-modal large model to automatically generate the occurrence parameters of the discontinuity surface image based on the optimized segmentation result of the discontinuity surface image is:

[0035] For any discontinuity surface contained in the optimized segmentation result, the morphological features of the discontinuity surface are extracted.

[0036] According to a preset discontinuity surface morphological feature prompt word template, the morphological features of the discontinuity surface are used to generate a discontinuity surface morphological feature prompt word of the discontinuity surface, and the optimized segmentation result is input into a multi-modal large model.

[0037] The multi-modal large model simultaneously processes the discontinuity surface morphological feature prompt word of the discontinuity surface and the optimized segmentation result to generate a discontinuity surface occurrence calculation code of the discontinuity surface.

[0038] The occurrence parameters of the discontinuity surface are automatically generated by executing the discontinuity surface occurrence calculation code.

[0039] All discontinuity surfaces contained in the optimized segmentation result are traversed to generate the occurrence parameters of each discontinuity surface and are summarized to obtain the occurrence parameters of the discontinuity surface image.

[0040] Further, the specific content of the fault determination thought chain is:

[0041] Step B1: call a JSON file, filter out discontinuity surfaces with hard discontinuity surface and weak discontinuity surface categories, and extract the start and end pixel point coordinate values of all filtered discontinuity surfaces, and start traversing each filtered discontinuity surface in turn.

[0042] Step B2: determine the spatial range of the current discontinuity surface according to the start and end pixel point coordinate values of the current discontinuity surface . In addition, all geological features are extracted from the optimized segmentation result, and each geological feature is divided into a left feature or a right feature of the current structural surface .

[0043] The specific content of the division of each geological feature into a left feature or a right feature of the current structural surface is as follows: for each geological feature , the first and last pixel points of the geological feature are obtained, and one pixel point is randomly sampled from each side of the midpoint of the geological feature , and it is judged whether all of the first and last pixel points of the geological feature and the two randomly sampled pixel points are located on the left side and the right side of the current structural surface , if all are located on the left side, the geological feature is taken as a left feature of the current structural surface ; if all are located on the right side, the geological feature is taken as a right feature of the current structural surface .

[0044] Step B3: if there are geological features of the same category in the left feature and the right feature of the current structural surface , the current structural surface is marked as a fault candidate, and the left feature and the right feature of the same category are taken as a geological feature pair of the current structural surface .

[0045] Step B4: according to the occurrence parameters of the working face image, for each geological feature pair of the current structural surface , the slope angle between the current structural surface and the left geological feature in the geological feature pair and the slope angle between the current structural surface and the right geological feature in the geological feature pair are calculated, if there is at least one geological feature pair satisfying and , the difference is within a set threshold, the fault candidate mark of the current structural surface is kept, and step B5 is performed; otherwise, the fault candidate mark of the current structural surface is cancelled, and step B2 is returned until the traversal is completed.

[0046] Step B5: for each geological feature pair of the current structural surface , the slope angle ​If the included angle of the slope of at least one pair of geological features is not more than a preset included angle value, the current structural surface is marked as a fault, and step B2 is returned to until the traversal is completed.

[0047] In another aspect, the present application provides a drilling and blasting rock mass structure identification system based on a large model agent, which comprises:

[0048] A preliminary segmentation module is configured to obtain a working face image, and use an image segmentation small model to preliminarily identify geological features of the working face image to obtain a preliminary segmentation result of the working face image.

[0049] A segmentation correction module is configured to input the working face image and the preliminary segmentation result of the working face image into a multi-modal large model based on a correction thought chain, and correct the preliminary segmentation result through iterative optimization to obtain an optimized segmentation result of the working face image.

[0050] An occurrence calculation module is configured to prompt an engineering to guide the multi-modal large model to automatically generate occurrence parameters of the working face image according to structural surface morphological features based on the optimized segmentation result of the working face image.

[0051] A fault inference module is configured to input the optimized segmentation result of the working face image and the occurrence parameters of the working face image into the multi-modal large model for fault reasoning based on a fault determination thought chain to obtain a fault inference result of the working face image.

[0052] A knowledge graph search module is configured to search a pre-constructed domain knowledge graph by the multi-modal large model to generate feature descriptions of different geological features in the working face image based on the optimized segmentation result of the working face image.

[0053] A rock mass structure information generation module is configured to generate a professional text description of rock mass structure information on the working face by the multi-modal large model according to a preset prompt template based on the optimized segmentation result of the working face image, the fault inference result, and the feature descriptions of different geological features.

[0054] In a third aspect, the present application provides an electronic device, comprising one or more processors and a memory for storing instructions that, when executed by the one or more processors, cause the one or more processors to perform the drilling and blasting rock mass structure identification method based on a large model agent.

[0055] In a fourth aspect, the present application provides a computer-readable storage medium storing executable instructions that, when executed, cause a processor to perform the drilling and blasting rock mass structure identification method based on a large model agent.

[0056] In a fifth aspect, the present application provides a computer program product comprising computer programs or instructions, which, when executed by a processor, implement the method for identifying rock mass structure of drill-and-blast method based on large model agent.

[0057] The beneficial effects of the above technical solution are as follows:

[0058] The method of the present application takes the identification of rock mass structure information on the working face as the main line, constructs an agent system integrating "structure surface identification, automatic extraction of occurrence, spatial reasoning, fault determination, and professional description generation", has the ability of geological feature identification and automatic calculation of occurrence information, fault determination, and automatic generation of professional description of geological information, effectively realizes the automatic process of geological feature identification and automatic calculation of occurrence information, intelligent fault determination, and automatic generation of professional description of geological information, not only significantly improves the efficiency and accuracy of geological information identification and fault determination, improves the construction efficiency, but also greatly reduces the dependence on manual operation, reduces the time and cost. The main effects are as follows:

[0059] 1. Correcting the segmentation result of the slender structure surface: the method of the present application constructs a geological feature image segmentation dataset by manual annotation, and obtains a preliminary segmentation result of the geological feature by training a Mask-RCNN traditional vision model, which is low-cost and lightweight, to help the multi-modal large model locate the local area of the geological feature. Based on the preliminary segmentation result, the pixel-level recognition result is extracted as an intermediate medium, and a "correction thinking chain" is constructed to guide the multi-modal large model to gradually iterate and correct the preliminary segmentation result, and finally the large model is guided to classify the specific structure surface category through the prompt prompt, which reduces the model training cost and the dependence on manual expert experience.

[0060] 2、Geological feature recognition and occurrence automatic extraction required for fault determination: In view of the dependence on structural surface distribution and occurrence information in the fault determination process, the method of the present application uses a lightweight visual model to preliminarily segment the structural surface of the working face image, providing a basic input for large model reasoning. The method of the present application combines a multi-modal large model and a "modified thinking chain" to modify the segmentation result of the small model, gradually refining the structural surface recognition result, improving the recognition accuracy, and solving the redundancy phenomenon of the segmentation result of the traditional small model in identifying geological features. At the same time, the method of the present application uses the natural language ability of the multi-modal large model to input the unique characteristics of different structural surfaces, which can distinguish hard structural surfaces and weak structural surfaces in detail. In addition, the method of the present application prompts the large model geological feature occurrence information calculation method, lets the large model automatically write python code and calculates the occurrence information through the running code to decouple the natural language ability and numerical calculation ability of the large model, so as to improve the calculation accuracy and calculation method of the occurrence index, and provide quantifiable basis for subsequent judgment of whether the geological features on both sides of the structural surface have spatial displacement, and form the basis support for intelligent fault determination.

[0061] 3、Enhance the explainability and generalizability of occurrence information calculation: The method of the present application decouples the natural language understanding ability and numerical calculation ability of the large model, and first lets the large model generate the corresponding python code based on the artificially prompted occurrence information calculation method, and automatically calculates the occurrence information through the code running.

[0062] 4、Intelligent fault determination: The method of the present application has the ability to infer whether a fault occurs, and guides the large model to gradually understand the physical mechanism behind the geological feature of the fault through the construction of "fault determination thinking chain", and determines the fault based on the geological feature segmentation result, solves the problem that the traditional small model cannot determine the fault and the low reasoning ability of the general large model, improves the accuracy of the large model fault determination, and assists the engineering site to determine the stability of the current rock mass. Combining the fault determination result with the characteristics of the current tunnel is helpful for immediate warning of extreme geological disasters such as fractured rock burst and gushing water.

[0063] 5、Geological information professional description automatic generation: the method has the ability of comprehensive analysis text generation of geological information, the large model is used for comprehensively identifying the results of geological features and the results of fault determination, and the deep engineering field knowledge of the large model is enhanced by constructing a field knowledge graph. According to the geological feature categories of the identification results, detailed geological feature descriptions are obtained, finally, a prompt template is designed to guide the large model to integrate various results and standardize the output answer format, realize the comprehensive summary of geological information, and generate a complete summary text. The method solves the problem of high cost of fine-tuning the large model with field knowledge, and the generated comprehensive text description breaks the limitation of single output mode of traditional methods, enhances the explainability, realizes richer interactive experience, and assists in completing the information summary work such as engineering daily report.

[0064] In summary, the design of the whole method emphasizes the accuracy of the geological feature segmentation result, the generalization of the occurrence information calculation, the logic of the fault intelligent determination and the professionalism of the geological information summary. The method can realize geological feature recognition, occurrence information calculation and fault determination with low labor cost by constructing an intelligent agent, connecting traditional visual models, multi-modal large models and field knowledge graphs, and then replacing manual generation of professional description of geological information on the current working face to assist in rock mass stability identification. The method not only improves the efficiency and accuracy of geological information identification and fault determination, but also provides strong support for identifying the current rock mass stability on the construction site, replacing the traditional manual analysis and summary method, assisting the construction personnel to quickly master the geological information on the construction site, and improving the engineering quality and safety level. The method promotes the modernization and intelligent development of the construction industry, provides a new solution for geological information comprehensive analysis and identification based on the working face image on the construction site, and leads the industry to a more efficient, safer and more intelligent direction. BRIEF DESCRIPTION OF DRAWINGS

[0065] Figure 1 The flowchart of the drilling and blasting method rock mass structure identification method based on the large model intelligent agent in this embodiment;

[0066] Figure 2 The scheme framework diagram of the drilling and blasting method rock mass structure identification method based on the large model intelligent agent in this embodiment;

[0067] Figure 3 The correction thought chain flowchart in this embodiment;

[0068] Figure 4 The fault determination thought chain flowchart in this embodiment;

[0069] Figure 5A structural diagram of a drill-and-blast rock mass structure identification system based on a large model agent in the embodiment. DETAILED DESCRIPTION

[0070] In order to facilitate the understanding of the present application, the specific embodiments of the present application are further described in detail below in combination with the drawings and embodiments. The following embodiments are used to illustrate the present application, but are not used to limit the scope of the present application. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present application more thorough and comprehensive.

[0071] The embodiment proposes a drill-and-blast rock mass structure identification method based on a large model agent. First, a traditional visual small model is used to preliminarily locate each geological feature. Then, a thinking chain is used to endow a multi-modal large model with thinking and reasoning ability, to gradually guide the correction of the segmentation result, and to automatically calculate the occurrence information. Next, the large model is guided to analyze the relationship between each geological feature on the working face to determine whether a fault exists on the current working face. Finally, in order to improve the professionalism of the automatically generated professional geological information, the large model is enhanced by an external domain knowledge graph to enhance the deep engineering field knowledge. When the method is used for geological information identification and fault determination tasks, only the working face image is needed to quickly obtain the recognition result of the geological feature and automatically calculate the occurrence information of the geological feature, further determine whether a fault exists, and finally automatically generate a summary text containing the distribution position of the geological features such as faults on the working face, and the professional description of the features.

[0072] Embodiment 1

[0073] A drill-and-blast rock mass structure identification method based on a large model agent in the embodiment, as shown in Figure 1 and Figure 2 , the method includes the following processes:

[0074] Obtain the working face image of the construction site, and use the image segmentation small model to preliminarily identify the geological features of the working face image to obtain the preliminary segmentation result of the working face image.

[0075] The construction method of the image segmentation small model is as follows:

[0076] Collect several working face images of the construction site, and label the geological features of each working face image respectively; wherein the geological features include: structural surfaces and dikes.

[0077] In the embodiment, the structural plane is an important geological feature for assisting in identifying the stability degree of the rock mass, and the fine classification thereof includes hard structural plane, weak structural plane, fault, and the like. The structural plane appears as a narrow crack on the working face, and different types of structural planes have unique characteristics. For example, the hard structural plane appears as a closed, slightly filled, or unfilled structural plane; and the weak structural plane appears as a structural plane generally filled with a certain thickness of weak material, such as a mudified, softened, or broken thin interlayer. However, in the background of the working face, the traditional visual model cannot well capture the unique characteristics of different structural planes, and forcibly distinguishing different types of structural planes during training may reduce the performance of the model. Therefore, the embodiment preliminarily identifies the geological features of the working face image by using the visual model, and unifies different types of structural planes as structural planes, and then differentiates and classifies the structural planes in detail according to the unique characteristics of different types of structural planes by using the multi-modal large model, so as to realize fine classification.

[0078] Each working face image and the labeled geological features are taken as a sample to construct a geological feature image segmentation dataset, and the geological feature image segmentation dataset is divided into a training set, a test set, and a validation set according to a predetermined ratio.

[0079] The visual model is selected as the image segmentation small model, the image segmentation small model is iteratively trained by using the training set, and the test set is periodically tested during the training process. After the training is completed, the trained image segmentation small model is verified by using the validation set, and the final image segmentation small model is obtained.

[0080] In the embodiment, the Mask-RCNN traditional visual model is selected as the image segmentation small model. The geological feature image segmentation dataset is divided into a training set, a test set, and a validation set, and the training and verification are performed based on the divided dataset. The total number of training rounds is set to 80 rounds, in which the test set is tested once every 5 training rounds, and the validation set is verified after the training is completed, so as to obtain the trained image segmentation small model. It should be noted that, in addition to the Mask-RCNN traditional visual model, other visual models applicable to the image segmentation task can also be used as the image segmentation small model, and different models may have different accuracies.

[0081] For any working face image, the working face image is input into the finally obtained image segmentation small model for preliminary identification, and the preliminary segmentation result of the working face image is output.

[0082] In the embodiment, the original working face image is input into the trained image segmentation small model to preliminarily identify the geological features on the working face and obtain a segmentation result image. This step aims to realize the positioning of the geological features, help the multi-modal large model lock the local range of the geological features, and improve the identification and analysis capabilities of the multi-modal large model.

[0083] The preliminary segmentation result includes: a target detection frame, a segmentation result image, and a classification probability of each pixel point.

[0084] Based on the modified thought chain, the working face image and the preliminary segmentation result of the working face image are input into the multi-modal large model, and the preliminary segmentation result is modified through iterative optimization to obtain the optimized segmentation result of the working face image.

[0085] The specific content of the step of inputting the working face image and the preliminary segmentation result of the working face image into the multi-modal large model and modifying the preliminary segmentation result through iterative optimization to obtain the optimized segmentation result of the working face image is:

[0086] Based on the preliminary segmentation result of the working face image, the segmentation result image is cropped according to the target detection frame to obtain a local image of different geological features in the working face image; meanwhile, the original image corresponding to each local image is cropped in the working face image.

[0087] The pixel-level recognition result of each local image is extracted, including: pixel point category, category probability and binary mask of each pixel point; wherein the pixel point category includes: structural plane, dike and background.

[0088] In this embodiment, the multi-modal large model is enhanced to modify the segmentation result by constructing a modification thought chain. Specifically: first, in order to make the multi-modal large model focus on the geological feature area on the working face, based on the preliminary segmentation result, the local image of each geological feature is cropped according to the target detection frame. Next, the pixel-level recognition result of each geological feature is extracted, which represents which pixel point category in the current region belongs to the geological feature and which pixel point category belongs to the background, and also contains the classification probability of each pixel point. Since the multi-modal large model cannot directly modify the segmentation result image, the pixel-level recognition result is intended to serve as an intermediate carrier for modifying the segmentation result. Then, for each geological feature, the original image of the geological feature region, the recognition result image, i.e. the cropped local image, and the pixel-level recognition result are input into the multi-modal large model. Next, by constructing a modification thought chain, the large model is guided to iteratively compare and observe the recognition result image and the original image, and to modify the pixel point category on the pixel-level recognition result.

[0089] The modification thought chain is constructed to guide the multi-modal large model to compare and observe the local image and the original image corresponding to the local image, and to modify the pixel point category on the pixel-level recognition result of the local image.

[0090] The specific content of the modification thought chain is:

[0091] Set a probability threshold and a maximum number of iterations, and initialize the current iteration round to 0. For any set of input local images, the original images corresponding to the local images, and the pixel-level recognition results of the local images, perform the following steps:

[0092] Step A1: Determine whether the current mask M has redundancy by comparing the local image and the original image. If not, the current mask M is taken as the final mask and step A4 is performed; if so, step A2 is performed.

[0093] Step A2: Extract all boundary pixels of the target region from the current mask and respectively correct the pixel point category of each boundary pixel to obtain an updated mask . Determine whether the updated mask is identical to the current mask . If so, the updated mask is taken as the final mask and step A4 is performed; if not, step A3 is performed.

[0094] The target region refers to the set of connected regions formed by all pixel points with a pixel value of 1 in the current mask .

[0095] For any boundary pixel , if the category probability of the boundary pixel does not exceed the probability threshold, the pixel point category of the boundary pixel is modified to background; if the category probability of the boundary pixel exceeds the probability threshold, the pixel point category of the boundary pixel is not changed.

[0096] Step A3: Superimpose the updated mask and the original image to generate a corrected visualization image, increment the current iteration round by 1, and determine whether the current iteration round reaches the maximum number of iterations. If so, the updated mask is taken as the final mask and step A4 is performed; if not, the corrected visualization image is taken as a new local image, the updated mask is taken as a new current mask , and step A1 is returned.

[0097] Step A4: Superimpose the final mask and the original image to obtain a segmentation result visualization image, and take the segmentation result visualization image and the final mask as the final output.

[0098] In this embodiment, as shown in Figure 3 , the task definition is: given a set of inputs <original image, segmentation result image, pixel-level recognition result>, please automatically and iteratively eliminate the redundant areas in the segmentation result of the geological feature according to the following rules, only correct the boundary pixels, the probability threshold is 0.5, and the maximum number of iterations is limited to 10 times. The specific execution process is as follows:

[0099] Input: (1) original image: <ORIG_IMG>; for example, the original image is an RGB image;

[0100] (2) segmentation result image: <PRED_VIS>;

[0101] (3) pixel-level recognition result: <PIXEL_RESULT>, including: predicted geological feature class probability and binary mask . Wherein if the pixel point is a geological feature, the pixel point corresponds to 1; otherwise, 0.

[0102] Step 1: observation of recognition result: compare <ORIG_IMG> with <PRED_VIS> to determine whether the current mask has redundancy; the redundancy refers to overflow and burr.

[0103] If no redundancy is found: go to "termination condition and output".

[0104] If redundancy is found: go to Step 2.

[0105] Step 2: screening class fuzzy points: extract "outer layer pixels" in the current mask . For each boundary pixel .

[0106] If , change its class to background, i.e. .

[0107] If , keep it unchanged.

[0108] After completion, the updated mask is obtained.

[0109] Step 3: display the corrected segmentation result: superimpose the updated mask on <ORIG_IMG> to generate the corrected visualization image , and go back to Step 1 for the next round of determination.

[0110] Termination condition and output: stop when Step 1 determines that no redundancy appears, or no pixel changes in this round, or reaches MAX_ITERS, and output FINAL_MASK, i.e., the final mask and FINAL_OVERLAY, i.e. the segmented result after overlaying with <ORIG_IMG>.

[0111] For each local image, input the local image, the original image corresponding to the local image, and the pixel-level recognition result of the local image into the multi-modal large model, and correct the pixel-level recognition result of the local image based on the revised thought chain to obtain a revised segmentation result; wherein the revised segmentation result includes: a segmented result visualization and a final mask.

[0112] Extract the structural surface from the revised segmentation result, and input the original image corresponding to the structural surface into the multi-modal large model, prompt the multi-modal large model to identify the class of the structural surface by giving geological feature descriptions of different structural surface classes, and obtain a revised structural surface class; wherein the structural surface class includes: hard structural surface and weak structural surface.

[0113] Splice all the revised segmentation results and revised structural surface classes corresponding to the respective local images to generate an optimized segmentation result of the working face image, and write the revised structural surface class and the head and tail pixel point coordinates of each structural surface into a JSON file.

[0114] In this embodiment, the original local image corresponding to the structural surface is input into the multi-modal large model, and the multi-modal large model is prompted to classify each structural surface by giving the textual description of the hard structural surface features and the textual description of the weak structural surface features. The revised segmentation result and the revised structural surface class are spliced back to the original working face image size, and finally the geological feature class label recognized and corrected on the working face image, and the coordinates of the two endpoints, i.e., the first pixel points in the geological feature segmentation result, are written into a JSON file for subsequent occurrence information calculation and fault inference.

[0115] Based on the optimized segmentation result of the working face image, the multi-modal large model is prompted to automatically generate the occurrence parameters of the working face image according to the structural surface morphological features.

[0116] The specific content of the optimized segmentation result of the working face image based on the optimized segmentation result of the working face image, prompting the multi-modal large model to automatically generate the occurrence parameters of the working face image according to the structural surface morphological features is:

[0117] For any structural surface included in the optimized segmentation result, extract the morphological features of the structural surface.

[0118] In this embodiment, based on the optimized segmentation result, the occurrence parameters of each structure surface are generated by traversing the final mask All structure surfaces are determined, and the morphological features of each structure surface are extracted respectively.

[0119] According to the preset structure surface morphological feature prompt word template, the structure surface morphological feature prompt word of the structure surface is generated using the morphological feature of the structure surface, and the optimized segmentation result is input into the multi-modal large model.

[0120] The multi-modal large model simultaneously processes the structure surface morphological feature prompt word of the structure surface and the optimized segmentation result to generate the occurrence calculation code of the structure surface.

[0121] In this embodiment, in order to improve the accuracy and adaptability of the intelligent agent in the calculation of the occurrence parameters of the geological structure surface, the natural language understanding ability of the multi-modal large model is decoupled from the numerical calculation process by using code generation. Specifically, the model automatically generates the corresponding occurrence calculation code according to the structure surface morphological feature by prompting the engineering guide model. The data processing process of the model is as follows: the large model processes the text prompt word and the visual image of the structure surface in parallel, understands the text prompt word and selects the appropriate calculation strategy according to the preset rules, and confirms that the structure surface morphological feature described by the text prompt word is consistent with the image by analyzing the visual image, so as to ensure that the text prompt word and the visual image are mutually confirmed and avoid making wrong decisions due to the deviation of a single data source. Based on the selected calculation strategy, executable Python code containing necessary mathematical formulas and library function calls is dynamically assembled and generated. For example: when the structure surface is relatively flat, the model generates a simplified trace length calculation logic based on the first and last pixel points of the structure surface; when the structure surface is relatively tortuous, a segmented addition logic combining the first and last points and multiple discrete sampling points in the middle is generated to adapt to more complex curve shapes. The intelligent agent executes the code internally to automatically complete the calculation of the occurrence parameters of the structure surface based on the segmentation result. This mechanism not only improves the calculation accuracy, but also enhances the adaptability of the occurrence calculation method to various geological shapes, and has good controllability, interpretability and scalability.

[0122] The occurrence parameters of the structure surface are automatically generated by executing the occurrence calculation code of the structure surface.

[0123] All structure surfaces contained in the optimized segmentation result are traversed to generate the occurrence parameters of each structure surface and to aggregate the occurrence parameters to obtain the occurrence parameters of the tunnel face image.

[0124] Based on the fault determination thought chain, the optimized segmentation result of the tunnel face image and the occurrence parameters of the tunnel face image are input into the multi-modal large model for fault reasoning to obtain the fault inference result of the tunnel face image.

[0125] Faults are a type of structural plane. Their unique characteristic lies in the fact that what was originally a single geological feature is divided into two parts by the presence of a fault, and these two parts undergo spatial displacement. Based on the unique characteristics of faults, a fault identification framework is constructed to progressively guide the observation and identification of geological features in large-scale models.

[0126] The specific content of the fault diagnosis thought chain is as follows:

[0127] Step B1: Call the JSON file to filter out all structural surfaces classified as hard structural surfaces and weak structural surfaces, and extract the coordinates of the first and last pixels of all filtered structural surfaces. Then, start traversing each filtered structural surface in turn.

[0128] Step B2: Based on the current structural plane The coordinates of the first and last pixels determine the spatial range of the current structural plane, excluding the current structural plane. In addition, all geological features are extracted from the optimized segmentation results, and each geological feature is divided into current structural planes. Left-side or right-side features.

[0129] The process involves dividing each geological feature into current structural planes. Left-side features or current structural surface The specific content of the right-hand features is as follows: for each geological feature To obtain geological features The first and last pixels, and respectively from the geological features Randomly sample one pixel on each side of the midpoint to determine geological features. Does the first and last pixels and two randomly sampled pixels contain three sampling points that are all located on the current structural surface? If all of them are located on the left side, then the geological features will be... As the current structural plane The features on the left side; if all are located on the right side, then the geological features will be... As the current structural plane The right-side features.

[0130] It should be noted that if a geological feature is interrupted by a structural plane at a certain location, the geological feature will appear as two independent segments on the facet image, typically not on the same straight line. Therefore, this embodiment uses this as a preliminary criterion for determining whether a structural plane is a fault. That is, for any structural plane, other geological features need to be first classified as the left and right sides of the structural plane. Then, by judging whether the left and right sides of the structural plane contain similar geological features, it is determined whether the structural plane is a fault candidate; that is, whether there are independent and discontinuous geological feature entities on the left and right sides of the structural plane.

[0131] Since the geological features themselves may be winding on the actual tunnel face image, the coordinates of the first and last pixel points of the geological features may not be completely located on the same side of the structural plane in most cases, but only a small part exceeds the structural plane, and the main part of the geological features is still located on the same side of the structural plane, i.e., the coordinates of the first and last pixel points are prone to be located at the intersection with other geological features. Considering the above situation, in order to accurately determine whether the geological features constitute a fault candidate, the spatial position of the geological features is determined by judging whether all the three sampling points of the first and last pixel points of the geological features and the two random sampling points located on both sides of the midpoint of the geological features are located on the same side of the structural plane, i.e., the determination standard of the present embodiment does not strictly require that the coordinates of the first and last pixel points must be on the same side, so as to avoid misjudgment due to the intersection of the first and last pixel points of the geological features with other geological features.

[0132] Step B3: If there are geological features of the same category in the left feature and the right feature of the current structural plane , the current structural plane is marked as a fault candidate, and the left feature and the right feature of the same category are taken as a geological feature pair of the current structural plane .

[0133] Step B4: According to the occurrence parameters of the tunnel face image, for each geological feature pair of the current structural plane , the slope angle between the current structural plane and the left geological feature in the geological feature pair and the slope angle between the current structural plane and the right geological feature in the geological feature pair are calculated, respectively. If there is at least one geological feature pair that satisfies and , the difference between which is within a set threshold, the fault candidate mark of the current structural plane is retained, and step B5 is performed; otherwise, the fault candidate mark of the current structural plane is cancelled, and step B2 is returned until the traversal is completed. Step B5: For each geological feature pair of the current structural plane

[0134] , the slope angle between the left geological feature and the right geological feature in the geological feature pair is calculated. If there is at least one geological feature pair whose slope angle does not exceed a preset angle value, the current structural plane is marked as a fault, and step B2 is returned until the traversal is completed. In the present embodiment, as

[0135] Figure 4The execution process of the fault determination thought chain is as follows:

[0136] Step 1: Identify structural planes: Faults belong to structural planes, so both hard structural planes and weak structural planes may be faults. Extract the geological coordinate information of structural planes with class labels of hard structural planes and weak structural planes from the JSON file of the segmentation result.

[0137] Step 2: Analyze geological features on both sides of the structural plane: Traverse each structural plane coordinate information extracted in the previous step to obtain the first and last pixel coordinates to determine the spatial range of the structural plane. For other geological features, extract the first and last coordinates, and randomly sample two pixel coordinates from the midpoint of the geological feature, a total of four representative points. Determine whether all three points are located on one side of the structural plane: if all points are located on the left side of the structural plane, the geological feature belongs to the left side; if all points are located on the right side, it belongs to the right side. Realize it by generating Python code that implements the above logic.

[0138] Step 3: Preliminary screening of symmetrical geological features: Read the geological features contained on the left and right sides of each structural plane. Structural planes with the same geological features on the left and right sides may be faults.

[0139] Step 4: Determine the difference between the structural plane and the lateral feature inclination: Read the results of the previous step, calculate the slope angle between the structural plane and the left geological feature, and the slope angle between the structural plane and the right geological feature. If the difference between the two angle ranges is within a certain threshold, the structural plane may be a fault.

[0140] Step 5: Determine the integrity of the left and right geological features: Read the results of the previous step, calculate the slope angle between the left and right geological features. When the angle is within 10°, it means that the left and right geological features have approximately the same slope, and the original structure is cut by the middle structural plane, so the structural plane is a fault.

[0141] The fault determination thought chain takes the structural plane as the core of analysis, and gradually guides the large model to complete the fault determination by combining the spatial relationship between the structural plane and the surrounding geological features. Starting from identifying the structural plane, it successively judges the consistency of the geological features on both sides of the structural plane, the inclination difference, and the overall continuity, and filters out the structural plane with all the characteristics of the fault.

[0142] Based on the optimized segmentation results of the working face image, the multi-modal large model searches the pre-constructed domain knowledge graph to generate feature descriptions of different geological features in the working face image.

[0143] In this embodiment, the deep engineering field knowledge of the multi-modal large model is enhanced in the manner of domain knowledge graph enhancement. Specifically, first, according to the knowledge points in the monograph in the deep engineering field, a domain knowledge graph is constructed by artificial construction, and stored in the form of a triple of <entity-relation-tail entity>, and then after obtaining the JSON file of the optimized segmentation result, the large model searches in the knowledge graph according to the geological feature category, and returns the detailed feature description corresponding to the geological feature, for example, "hard structural plane-feature is-closed and no filling or micro filling"; "hard structural plane-physical and mechanical properties-higher shear strength and compressive strength". This step aims to solve the problem of high-cost data preparation and model training process for fine-tuning a large model in a field.

[0144] Based on the optimized segmentation result of the working face image, the fault inference result and the feature description of different geological features, the multi-modal large model generates a professional text description of the rock mass structure information on the working face according to the preset prompt template.

[0145] In this embodiment, the multi-modal large model generates a complete summary text by integrating the hard structural plane, weak structural plane and dike in the optimized segmentation result, the fault inference result and the detailed geological feature description information obtained from the domain knowledge graph. Specifically, the prompt template needs to be designed in detail to prompt the large model to integrate the obtained result information, and the output content format of the large model also needs to be standardized. This step aims to replace the traditional manual summary process and automatically generate a professional rock mass structure information text description.

[0146] In summary, geological information identification and fault determination is a comprehensive task that requires accurate identification of geological features, detailed observation of the relationship between features and fault determination. The present application combines the respective advantages of traditional visual models and large models, and integrates thinking chain technology to enhance the reasoning ability of the large model, and constructs a complete intelligent agent to provide an intelligent solution for on-site geological feature analysis and improve the efficiency of on-site construction.

[0147] Embodiment 2:

[0148] A drilling and blasting method rock mass structure identification system based on a large model intelligent agent according to this embodiment, as shown in Figure 5 The system comprises:

[0149] A preliminary segmentation module is configured to obtain a working face image of a construction site, and use an image segmentation small model to preliminarily identify the geological features of the working face image, and obtain a preliminary segmentation result of the working face image.

[0150] The segmentation correction module is configured to input the tunnel face image and a preliminary segmentation result of the tunnel face image into a multi-modal large model based on a correction thinking chain, and correct the preliminary segmentation result through iterative optimization to obtain an optimized segmentation result of the tunnel face image.

[0151] The occurrence calculation module is configured to prompt the engineering to guide the multi-modal large model to automatically generate occurrence parameters of the tunnel face image according to structural surface morphological characteristics based on the optimized segmentation result of the tunnel face image.

[0152] The fault inference module is configured to input the optimized segmentation result of the tunnel face image and the occurrence parameters of the tunnel face image into the multi-modal large model for fault reasoning based on a fault determination thinking chain to obtain a fault inference result of the tunnel face image.

[0153] The knowledge graph search module is configured to search a pre-constructed domain knowledge graph based on the optimized segmentation result of the tunnel face image to generate feature descriptions of different geological features in the tunnel face image.

[0154] The rock mass structure information generation module is configured to generate a professional text description of rock mass structure information on the tunnel face based on the optimized segmentation result of the tunnel face image, the fault inference result and the feature descriptions of different geological features according to a preset prompt template.

[0155] Embodiment 3:

[0156] The embodiment provides an electronic device, which includes one or more processors and a memory. The memory is configured to store instructions, and when the instructions are executed by the one or more processors, the one or more processors are caused to execute the method for identifying rock mass structure of drill-and-blast method based on large model agent.

[0157] The electronic device can be a mobile phone, a computer, a tablet computer or the like, and includes a memory and a processor. The memory stores a computer program, and the computer program is executed by the processor to implement the method for identifying rock mass structure of drill-and-blast method based on large model agent as described in the embodiments. It can be understood that the electronic device can further include an input / output (I / O) interface and a communication component.

[0158] The processor is configured to execute all or part of the steps of the method for identifying rock mass structure of drill-and-blast method based on large model agent as described in the above embodiments. The memory is configured to store various types of data, which can include, for example, instructions of any application program or method in the electronic device, and application program related data.

[0159] The processor can be an Application Specific Integrated Cricuit (ASIC), a Digital Signal Processor (DSP), a Programmable Logic Device (PLD), a Field Programmable Gate Array (FPGA), a controller, a microcontroller, a microprocessor, or other electronic elements, which are used to execute the large model-based agent drilling and blasting method for rock mass structure identification described in the above embodiments.

[0160] Embodiment 3:

[0161] The embodiment provides a computer readable storage medium storing executable instructions, which, when executed, can be stored in one computer readable storage medium if implemented in the form of a software function unit and sold or used as an independent product.

[0162] The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the large model-based agent drilling and blasting method for rock mass structure identification described in various embodiments of the present application.

[0163] The aforementioned storage medium includes a flash memory, a hard disk, a multimedia card, a card-type memory (for example, an SD (Secure Digital Memory Card) or a DX (an abbreviation of Memory Data Register, MDR) memory, etc.), a random access memory (RAM), a static random access memory (SRAM), a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), a programmable read-only memory (PROM), a magnetic memory, a magnetic disk, an optical disk, a server, an APP (an abbreviation of Application) application store, and various media capable of storing a program check code, on which a computer program is stored, which, when executed by a processor, can implement each step of the large model-based agent drilling and blasting method for rock mass structure identification described above.

[0164] Embodiment 4:

[0165] The embodiment provides a computer program product comprising computer programs or instructions, which, when executed by a processor, implement the drilling and blasting rock mass structure identification method based on a large model intelligent agent.

[0166] Based on such understanding, the technical solution of the present application or the part of the technical solution that essentially contributes to the prior art can be embodied in the form of a computer program product.

[0167] Each of the embodiments in the present application is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other. Each embodiment mainly describes the difference from other embodiments.

[0168] The protection scope of the present application is not limited to the above-mentioned embodiments. Obviously, those skilled in the art can make various modifications and changes to the present disclosure without departing from the scope and spirit of the present disclosure. If these modifications and changes belong to the scope of the present disclosure and its equivalent technology, the present disclosure also includes these modifications and changes.

Claims

1. A method for identifying rock mass structure by drilling and blasting based on a large model agent, characterized in that, The method comprises the following steps: An image of a working face is acquired, and a preliminary identification of geological features of the image of the working face is performed using an image segmentation small model to obtain a preliminary segmentation result of the image of the working face; Based on a correction thought chain, the image of the working face and the preliminary segmentation result of the image of the working face are input into a multimodal large model, and the preliminary segmentation result is corrected through iterative optimization to obtain an optimized segmentation result of the image of the working face; Based on the optimized segmentation result of the image of the working face, a structure parameter of the image of the working face is automatically generated according to a structural surface morphological feature by prompting the engineering to guide the multimodal large model; Based on a fault judgment thought chain, the optimized segmentation result of the image of the working face and the structure parameter of the image of the working face are input into the multimodal large model for fault reasoning to obtain a fault inference result of the image of the working face; Based on the optimized segmentation result of the image of the working face, the multimodal large model searches a pre-constructed domain knowledge graph to generate feature descriptions of different geological features in the image of the working face; Based on the optimized segmentation result of the image of the working face, the fault inference result and the feature descriptions of different geological features, a professional text description of rock mass structure information on the working face is generated according to a preset prompt template by the multimodal large model.

2. The method according to claim 1, wherein, The construction method of the image segmentation small model is as follows: A plurality of images of working faces are collected, and geological features of each image of the working face are labeled respectively; The geological features include structural surfaces and dikes. Each image of the working face and the labeled geological features are taken as a sample to construct a geological feature image segmentation data set, and the geological feature image segmentation data set is divided into a training set, a test set and a validation set according to a predetermined proportion; A visual model is selected as the image segmentation small model, the image segmentation small model is iteratively trained using the training set, and periodic testing is performed on the training set during the training process. After the training is completed, the trained image segmentation small model is verified using the validation set to obtain the final image segmentation small model. For any image of the working face, the image of the working face is input into the final image segmentation small model for preliminary identification, and a preliminary segmentation result of the image of the working face is output. The preliminary segmentation result includes a target detection frame, a segmentation result image and a classification probability of each pixel.

3. The method according to claim 2, wherein, Based on the correction thought chain, the image of the working face and the preliminary segmentation result of the image of the working face are input into the multimodal large model, and the preliminary segmentation result is corrected through iterative optimization to obtain the optimized segmentation result of the image of the working face. The specific content is as follows: Based on the preliminary segmentation result of the image of the working face, the segmentation result image is cropped according to the target detection frame to obtain local images of different geological features in the image of the working face. Meanwhile, the original images corresponding to the local images are cropped in the image of the working face; Pixel-level recognition results of the local images are extracted respectively, including pixel categories, classification probabilities and binary masks of each pixel. The pixel categories include structural surfaces, dikes and backgrounds. Construct a correction thought chain for guiding the multimodal large model to compare and observe the local image and the original image corresponding to the local image, and correct the pixel class of the pixel-level recognition result of the local image; For each local image, input the local image, the original image corresponding to the local image, and the pixel-level recognition result of the local image into the multimodal large model, and correct the pixel class of the pixel-level recognition result of the local image based on the correction thought chain to obtain a corrected segmentation result; wherein the corrected segmentation result includes: a segmentation result visualization graph and a final mask; Extract the structural surface from the corrected segmentation result, and input the original image corresponding to the structural surface into the multimodal large model to prompt the multimodal large model to identify the class of the structural surface by giving geological feature descriptions of different structural surface classes, and obtain a corrected structural surface class; wherein the structural surface class includes: hard structural surface and weak structural surface; Splice the corrected segmentation result and the corrected structural surface class corresponding to all local images to generate an optimized segmentation result of the working face image, and write the corrected structural surface class and the head and tail pixel point coordinate values of each structural surface into a JSON file.

4. The method according to claim 3, wherein, The specific content of the correction thought chain is: Set a probability threshold and a maximum number of iterations, and initialize the current iteration round to 0. For any group of input local images, original images corresponding to the local images, and pixel-level recognition results of the local images, perform the following steps: Step Al: judging whether the current mask M has redundancy by comparing the local image and the original image, if not, taking the current mask M as the final mask and performing step A4; if yes, performing step A2; Step A2: Extract all boundary pixels of the target region from the current mask and correct the pixel class of each boundary pixel respectively to get the updated mask ; determine whether the updated mask is the same as the current mask , if so, take the updated mask as the final mask and execute Step A4; If not, perform step A3; The target region refers to a connected region consisting of all pixel points with pixel values of 1 in the current mask The target region refers to a connected region consisting of all pixel points with pixel values of 1 in the current mask For any boundary pixel , if the class probability of the boundary pixel does not exceed the probability threshold, the pixel class of the boundary pixel is modified to background; if the class probability of the boundary pixel exceeds the probability threshold, the pixel class of the boundary pixel is not changed. Step A3: superimpose the updated mask with the original image to generate a revised visualization, increment the current iteration round by 1, and determine whether the current iteration round has reached the maximum number of iterations. If yes, the updated mask is taken as the final mask and Step A4 is executed; if no, the revised visualization is taken as the new local image, the updated mask is taken as the new current mask and Step A1 is returned to. Step 4: Overlaying the final mask with the original image to get the segmentation result visualization, and the segmentation result visualization and the final mask as the final output.

5. The method according to claim 4, wherein, The specific content of the optimized segmentation result based on the working face image for prompting the engineering to guide the multimodal large model to automatically generate the occurrence parameters of the working face image according to the structural surface morphological features is: For any structural surface contained in the optimized segmentation result, extract the morphological features of the structural surface; According to a preset structural surface morphological feature prompt word template, generate a structural surface morphological feature prompt word of the structural surface using the morphological features of the structural surface, and input the structural surface morphological feature prompt word and the optimized segmentation result into the multimodal large model; The multimodal large model simultaneously processes the structural surface morphological feature prompt word of the structural surface and the optimized segmentation result to generate occurrence calculation code of the structural surface; Automatically generate the occurrence parameters of the structural surface by executing the occurrence calculation code of the structural surface; Iterate through all structural surfaces contained in the optimized segmentation result to generate the occurrence parameters of each structural surface and aggregate them to obtain the occurrence parameters of the working face image.

6. The method according to claim 5, wherein, The specific content of the fault determination thought chain is: Step B1: Call the JSON file, filter out all structural surfaces with hard structural surface and weak structural surface classes, and extract the head and tail pixel point coordinate values of all filtered structural surfaces, and start iterating through each filtered structural surface in turn; Step B2: Based on the current structural plane The coordinates of the first and last pixels determine the spatial range of the current structural plane, excluding the current structural plane. In addition, all geological features are extracted from the optimized segmentation results, and each geological feature is divided into current structural planes. Left-side or right-side features; The process involves dividing each geological feature into current structural planes. Left-side features or current structural plane The specific content of the right-hand features is as follows: for each geological feature To obtain geological features The first and last pixels, and respectively from the geological features Randomly sample one pixel on each side of the midpoint to determine geological features. Does the first and last pixels and two randomly sampled pixels contain three sampling points that are all located on the current structural surface? If all of them are located on the left side, then the geological features will be... As the current structural plane The left-side features; If all are on the right side, the geological features are right side features of the current structural plane . Step B3: If there is a geological feature of the same category in the left feature and the right feature of the current structural plane , mark the current structural plane as a fault candidate, and take the left feature and the right feature of the same category as a geological feature pair of the current structural plane ; Step B4: According to the occurrence parameters of the planar image, for each geological feature pair of the current structural plane , the slope angle between the current structural plane and the left geological feature in the geological feature pair and the slope angle between the current structural plane and the right geological feature in the geological feature pair are calculated respectively, if there is at least one geological feature pair satisfying and , the difference of which is within a set threshold, the fault candidate marker of the current structural plane is retained, and step B5 is executed; otherwise, the fault candidate marker of the current structural plane is cancelled, and step B2 is returned until the traversal is completed; Step B5: For each pair of geological features of the current structural surface , calculate the slope angle between the left geological feature and the right geological feature in the pair of geological features , if the slope angle of at least one pair of geological features does not exceed the preset angle value, mark the current structural surface as a fault, and return to step B2 until the traversal is completed.

7. A drilling and blasting rock mass structure identification system based on a large model agent, used to implement the drilling and blasting rock mass structure identification method based on a large model agent in any one of claims 1-6, characterized in that, The system comprises: A preliminary segmentation module for obtaining a working face image of a construction site, and using an image segmentation small model to preliminarily identify the geological features of the working face image to obtain a preliminary segmentation result of the working face image; The segmentation correction module is configured to input the working face image and a preliminary segmentation result of the working face image into a multi-modal large model based on a correction thinking chain, and correct the preliminary segmentation result through iterative optimization to obtain an optimized segmentation result of the working face image. The occurrence calculation module is configured to prompt the engineering to guide the multi-modal large model to automatically generate an occurrence parameter of the working face image based on the optimized segmentation result of the working face image. The fault inference module is configured to input the optimized segmentation result of the working face image and the occurrence parameter of the working face image into the multi-modal large model based on a fault determination thinking chain to perform fault reasoning and obtain a fault inference result of the working face image. The knowledge graph search module is configured to search a pre-constructed domain knowledge graph based on the optimized segmentation result of the working face image to generate feature descriptions of different geological features in the working face image. The rock mass structure information generation module is configured to generate a professional text description of rock mass structure information on the working face based on the optimized segmentation result of the working face image, the fault inference result and the feature descriptions of different geological features.

8. An electronic device, comprising: One or more processors, and a memory for storing instructions that, when executed by the one or more processors, cause the one or more processors to perform the method of any one of claims 1-6. The computer-readable storage medium stores executable instructions that, when executed, cause a processor to perform the method of any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, A computer program or instructions that, when executed by a processor, implement the method of any one of claims 1-6.

10. A computer program product, characterised in that, ​

Images