A coal mine drawing intelligent recognition method based on multi-source automation pre-training

Abstract

The application provides a coal mine drawing intelligent recognition method based on multi-source automation pre-training, steps comprising: receiving a coal mine vector drawing package, automatically distinguishing mining engineering plan and various professional layout drawings; automatically positioning the legend area, analyzing the symbol graphics and corresponding text instructions, automatically constructing a safety measurement point sample library with text as a label and graphics as a sample; taking a pre-trained visual detection model as a base, and obtaining a safety measurement point recognition model through sample library field adaptation fine tuning; performing first-level recognition on the professional layout drawing with clean layers, extracting spatial structure elements and generating standardized training samples; superimposing non-target layer interference elements on the samples to simulate a high-noise environment for enhanced training, and obtaining a spatial structure recognition model; using the two types of models to recognize the elements and measurement points of the mining engineering plan respectively, and outputting the corresponding results. Advantageous effects: the double-source architecture is used to realize zero manual annotation to construct training samples, and mixed noise enhanced training is used to improve the drawing recognition accuracy.

Description

Technical Field

[0001] This invention relates to the field of coal mine engineering drawing recognition technology, specifically to an intelligent recognition method for coal mine drawings based on multi-source automated pre-training. Background Technology

[0002] In the field of information technology for mine safety production, intelligent recognition and analysis technology for coal mine drawings has become a core technical means supporting various business operations. In coal mine safety supervision scenarios, the supervision platform relies on this technology to conduct automatic compliance reviews of the mining engineering plan and various professional layout drawings submitted by the mine, quickly identify safety hazards, and generate standardized review reports. In the mine geological survey information system, core elements such as roadways, working faces, goaf areas, and safety monitoring points in the mine map can be automatically extracted and converted into structured GIS data for direct system access. In the internal management of mining enterprises, updated mine maps can be analyzed automatically and regularly to track the progress of hazard rectification and promote the upgrading of safety management from post-inspection to pre-prediction.

[0003] The current mainstream technical approach in this field is based on a multimodal vision-language large model, which continuously improves recognition capabilities through an iterative closed loop of sample library construction, manual annotation, model pre-training, fine-tuning, deployment, and feedback optimization. However, existing technical solutions face insurmountable technical bottlenecks when applied to coal mine scenarios. Mining engineering plans typically overlay nearly twenty layers of all elements, including roadways, geological structures, electromechanical equipment, and pipeline systems. The mixed layer information results in an accuracy rate of only about 40%-50% when directly using AI models for recognition, which is insufficient to meet the accuracy threshold for practical engineering applications. At the same time, the coal mining industry lacks a unified mapping standard. Different mines have significant differences in legend symbols, color line types, and annotation formats. Models trained for a single mine suffer from significant time-domain offset issues when applied across mines, resulting in extremely poor versatility. Furthermore, mine map annotation is highly dependent on manual work, with annotation time for a single map reaching several hours. The massive amount of maps from thousands of coal mines nationwide makes the construction of a universal sample library infeasible, creating a vicious cycle of "annotation-training" and failing to meet the real-time requirements for map submission and review. Summary of the Invention

[0004] To address the problems in the existing technology, this invention provides a method for intelligent recognition of coal mine drawings based on multi-source automated pre-training.

[0005] This invention provides a method for intelligent recognition of coal mine drawings based on multi-source automated pre-training, comprising the following steps: Step S1: Receive a vector file package containing multiple coal mine drawings and automatically identify and distinguish between mining engineering plan and various professional layout drawings; Step S2: Automatically locate the legend area in the coal mine drawing vector file, parse the symbol graphic unit and text description unit corresponding to each legend, use the text description as a label and the symbol graphic unit as a sample to automatically construct a safety measurement point symbol sample library; use the pre-trained visual detection model as a base and use the safety measurement point symbol sample library to perform domain adaptation fine-tuning to obtain a safety measurement point recognition model. Step S3: Perform primitive recognition on the clean professional layout diagram of the classified layers, extract spatial structure primitives and crop them to generate standardized training samples; Step S4: Add non-target layer interference elements of the mining engineering plan to the standardized training samples to simulate a high-noise environment with multiple layers and perform enhancement training to obtain a spatial structure recognition model. Step S5: Use the spatial structure recognition model to identify spatial structure elements in the mining project plan and output the spatial structure class recognition result; at the same time, use the safety measuring point recognition model to identify safety measuring points in the mining project plan and output the safety measuring point class recognition result.

[0006] The present invention is further improved in that, in step S2, the legend area is located by keyword matching, and the keywords include legend, symbol, sensor, and annotation description; The parsed symbolic graphic units are exported as image formats with a resolution of at least 600 DPI for use in building a sample library.

[0007] The present invention is further improved in that, in step S2, when processing drawings of different coal mines, the degree of difference between the legend of the current drawing and the existing sample library is determined by the pixel geometric shape similarity. If the similarity meets the preset threshold, the existing security measurement point recognition model can be directly reused; If the conditions are not met, the process reverts to the pre-trained visual detection model base and re-executes fine-tuning, forming an adaptive fine-tuning pipeline.

[0008] The present invention is further improved in that, in step S5, after obtaining the identification results of the safety measurement point class, the GIS spatial proximity analysis algorithm is used to automatically match and associate each safety measurement point with the corresponding text label.

[0009] The present invention is further improved, and the professional layout diagram includes one or more of the following: personnel positioning layout diagram, electromechanical equipment layout diagram, ventilation system diagram, safety monitoring system layout diagram, dust prevention pipeline layout diagram, compressed air self-rescue system layout diagram, underground drainage system layout diagram, and disaster avoidance route diagram.

[0010] In a further improvement to the present invention, in step S4, the non-target layer interference elements include one or more of contour lines, borehole symbols, and text annotations, which are used to simulate the high-noise interference environment of multiple layers superimposed on the mining engineering plan.

[0011] The present invention is further improved so that the spatial structure identification results include one or more of the following: roadway system and topological connection relationship, mining face, goaf, water accumulation area, protective coal pillar, and airtight wall.

[0012] The present invention is further improved by including, after step S5: performing coordinate matching and data fusion of the spatial structure class identification results and the safety measurement point class identification results in a unified GIS coordinate system to generate a digital mining map model containing topological attributes and feature attributes.

[0013] The present invention is further improved by inputting the digital mining map model into a preset safety hazard rule knowledge base, performing spatial geometric calculations and compliance verification through an inference engine, and outputting a structured safety hazard analysis report.

[0014] The present invention is further improved in that the rules of the safety hazard rule knowledge base are derived from industry standards and cover five major types of hazards: insufficient roadway spacing, disordered mining continuity, abnormal roadway structure, unsealed abandoned roadways, and non-compliant safety measuring point layout; the safety hazard analysis report includes hazard type, hazard level, location information, procedural basis, and rectification suggestions.

[0015] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention provides an intelligent recognition method for coal mine drawings based on multi-source automated pre-training, which solves the problem that the accuracy of direct AI recognition is low due to the overlay of all elements in existing mining engineering plan drawings, and cannot meet the practical requirements of engineering. At the same time, it addresses the problems of the lack of unified drafting standards in the industry, high cost of manual annotation, and serious cross-mine area offset. The specific beneficial effects are as follows: 1. Completely eliminates the bottleneck of manual annotation and achieves zero-manual sample construction. This invention uses a legend-driven vector automatic parsing mechanism to directly extract symbols and corresponding labels from CAD vector drawings, automatically constructing a safety measurement point sample library. The entire process requires no manual annotation, breaking the vicious cycle of "high annotation costs and inability to establish a general sample library" in the coal mining industry, and significantly reducing the manpower and time costs of model implementation.

[0016] 2. A significant leap in recognition accuracy, reaching practical engineering standards. This invention utilizes the inherent layered characteristics of coal mine drawings—"clean professional drawings, complex comprehensive drawings"—and employs a "professional" approach... Figure 1 The hierarchical strategy of "secondary identification sampling - mixed noise enhancement training - comprehensive map secondary identification" improves the recognition accuracy of spatial structure elements in mining engineering plan from about 45% to over 95%, and the recognition accuracy of safety measurement points is no less than 98%, meeting the engineering practical threshold for safety production review.

[0017] 3. Adaptive solution to cross-mine offset problem, with strong versatility. This invention has a built-in legend difference adaptive mechanism, which uses geometric similarity to determine the reuse or fine-tuning of the dynamic decision model, eliminating the need to train a separate model for each coal mine; in batch testing, it can automatically cluster coal mines with different mapping styles, covering the needs of multi-mine identification with very few fine-tunings, adapting to the current situation where there is no unified mapping standard in the industry.

[0018] 4. Fully automated operation to meet real-time review requirements. This invention automates the entire process from drawing input to recognition result output. The entire process for a single drawing takes only 3-4 minutes. When combined with a rule knowledge base, it can directly generate a hazard analysis report, which is about 100 times more efficient than traditional manual review and can support the regulatory scenario requirement of "drawings submitted for review immediately".

[0019] 5. High architectural compatibility, supporting deployment in multiple scenarios. This method supports a degraded recognition mode, achieving an accuracy rate of 85%-90% even without a corresponding professional layout diagram. The recognition results can be output in a standardized GIS data format, directly compatible with various business systems such as grounding information systems and security supervision platforms. It also supports predictive security analysis, making it flexible in application scenarios. Attached Figure Description

[0020] To more clearly illustrate the solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 This is a flowchart of a coal mine drawing intelligent recognition method based on multi-source automated pre-training, according to the present invention. Detailed Implementation

[0022] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs; the terminology used herein in the specification of the application is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having” and any variations thereof in the specification, claims and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0023] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0024] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0025] like Figure 1 As shown, this invention provides an intelligent recognition method for coal mine maps based on multi-source automated pre-training. The core concept is a two-stage hybrid architecture of "visual coarse screening + vector semantic fine judgment". The AI ​​visual model quickly completes the coarse positioning of global difference areas, and then the comparison range is locked within a local grid by combining GIS spatial index. Finally, high-precision vector primitive semantic comparison is completed within the local range. This not only solves the core problems of high complexity and long time consumption of full map primitive comparison, but also overcomes the defect of high false alarm rate of pixel-level comparison, and achieves a dual improvement in efficiency and accuracy of mine map difference detection.

[0026] The steps of the intelligent recognition method for coal mine drawings based on multi-source automated pre-training in this example are as follows: Step S1: Receive a vector file package containing multiple coal mine drawings and automatically identify and distinguish between mining engineering plan and various professional layout drawings.

[0027] After receiving the DWG / DXF format vector drawings uploaded by the coal mine, the system first reads the drawing name attributes, number of layers, total number of elements, and element density characteristics of each drawing through the CAD secondary development interface. A preset keyword library is used to match the drawing name field, initially distinguishing between mining engineering plan drawings and professional layout drawings. A second verification is then performed based on the number of layers threshold: drawings with more than 15 layers are classified as mining engineering plan drawings (comprehensive drawings), while those with fewer than 8 layers are classified as professional layout drawings. This completes the drawing classification and archiving. Professional layout drawings include categories such as personnel positioning layout drawings, electromechanical equipment layout drawings, ventilation system diagrams, safety monitoring system layout drawings, dust suppression sprinkler pipeline layout drawings, compressed air self-rescue system layout drawings, underground drainage system layout drawings, and disaster evacuation route maps.

[0028] Functions and benefits: By leveraging the inherent layered characteristics of coal mine drawings, which are characterized by "complex comprehensive drawing layers and simplified specialized drawing layers," the complexity of the drawings can be classified in advance, providing a preliminary judgment basis for subsequent classification and recognition strategies. This avoids redundant enhanced recognition processing on low-complexity specialized layout drawings, while ensuring that high-complexity comprehensive drawings are matched with corresponding enhanced recognition models, thereby improving the overall recognition efficiency and accuracy from the perspective of process architecture.

[0029] Step S2: Automatically locate the legend area in the coal mine drawing vector file, parse the symbol graphic unit and text description unit corresponding to each legend, use the text description as a label and the symbol graphic unit as a sample to automatically construct a safety measurement point symbol sample library; use the pre-trained visual detection model as a base and use the safety measurement point symbol sample library to perform domain adaptation fine-tuning to obtain a safety measurement point recognition model.

[0030] (1) Automatic positioning and parsing of legend area: Call the CAD vector programming interface, traverse the area around the drawing frame, match text entities through keywords such as "legend", "symbol", "sensor", "annotation", etc., and lock the legend table range; parse legend entries line by line, separate vector graphic entities and corresponding text annotation entities in each entry, export symbol graphic entities as PNG format images at a resolution of not less than 600 DPI, use the corresponding text annotation as classification labels, form a one-to-one corresponding sample data, and automatically build a safety measurement point symbol sample library.

[0031] (2) Model adaptive fine-tuning: Based on the pre-trained multimodal visual detection model, the automatically constructed sample library is input into the model to perform domain adaptation pre-training and fine-tuning. During the training process, the small target detection enhancement strategy is enabled to optimize the recognition recall rate of small-sized measurement point symbols and obtain a safety measurement point recognition model that is adapted to the current mine mapping style.

[0032] (3) Cross-mine adaptive mechanism: When processing new mine drawings, the geometric contour features of the newly analyzed legend symbols are extracted and pixel-level shape similarity is calculated with the symbol features of the existing sample library. If the similarity is higher than the preset threshold, the existing safety measurement point recognition model is directly reused. If the similarity is lower than the threshold, it reverts to the pre-trained base model and performs fine-tuning again based on the new sample library, forming an automatically iterative adaptive pipeline. In batch testing of 15 different coal mines, this mechanism can automatically cluster the drawings into 7 style clusters, and only 6 adaptive fine-tunings are needed to complete the recognition and adaptation of all mines.

[0033] Functions and beneficial effects: (1) By directly extracting labeled samples from the legends provided in the drawings through vector analysis, the work of manually marking measurement point symbols is completely replaced. The time required to construct a sample on a single drawing is reduced from 2-8 hours to a few seconds, thus breaking the vicious cycle of "high labeling costs and inability to establish a general sample library" in the industry. (2) To address the differences in mapping styles among different mines, the model can be dynamically reused and adaptively fine-tuned by judging geometric similarity. There is no need to train the model separately for each mine, which effectively solves the domain offset problem caused by the lack of a unified mapping standard in the industry and greatly improves the cross-mine versatility of the model. (3) Small target detection enhancement strategy adapts to the characteristics of small size and low proportion of measuring point symbols in coal mine drawings, improves the recognition accuracy of small-sized safety measuring points, and avoids missed detection.

[0034] Step S3: Perform primitive recognition on the clean professional layout diagram obtained from the classification layer, extract spatial structure primitives and crop them to generate standardized training samples.

[0035] For the various professional layout maps selected by classification, the basic visual segmentation model is called to perform first-level recognition. Since the professional layout maps only retain the core elements of the corresponding system, the layer structure is clean, and there are very few interfering elements, the model can directly identify the vector outlines of spatial structure elements such as roadways, working faces, goaf areas, protective coal pillars, sealed walls, chambers, and water accumulation areas. The recognition results are cropped and exported according to the element type to generate standardized training samples with classification labels. The accuracy of first-level recognition for clean professional layout maps can reach more than 98%.

[0036] Functions and benefits: By fully utilizing the natural advantages of professional layout maps, which have low information density and few interferences, labeled samples of spatial structure primitives can be obtained with nearly 100% accuracy, replacing the work of manually labeling spatial primitives and constructing a high-quality training dataset at low cost; at the same time, it avoids the high error caused by directly labeling on complex composite maps, providing a reliable data foundation for subsequent augmentation training.

[0037] Step S4: Add non-target layer interference elements of the mining engineering plan to the standardized training samples to simulate a high-noise environment with multiple layers and perform enhancement training to obtain a spatial structure recognition model.

[0038] Based on the standardized clean sample generated in step S3, interference elements such as contour lines, borehole symbols, and text annotations are extracted from the non-target layer of the mining engineering plan. By randomly positioning, randomly adjusting the transparency, and randomly scaling the interference elements, they are superimposed onto the clean sample to simulate the high-noise background of multiple layers superimposed on the comprehensive map. The enhanced dataset is used to train the visual recognition model. During the training process, the superposition ratio of interference elements is dynamically adjusted to enhance the model's ability to resist background noise, and finally, a spatial structure recognition model is obtained.

[0039] Functions and benefits: By artificially synthesizing interference, the complex scene of all elements superimposed on the mining engineering plan can be simulated at low cost, allowing the model to learn to distinguish target primitives from background interference during the training phase; without the need for manual annotation of complex composite maps, the robustness of the model in high-noise environments can be greatly improved, providing core support for high-precision recognition of subsequent composite maps.

[0040] Step S5: Use the spatial structure recognition model to identify spatial structure elements in the mining project plan and output the spatial structure class recognition result; at the same time, use the safety measuring point recognition model to identify safety measuring points in the mining project plan and output the safety measuring point class recognition result.

[0041] (1) Spatial structure primitive recognition: Call the spatial structure recognition model trained in step S4, perform secondary enhanced recognition on the mining engineering plan, extract all spatial structure primitives such as roadway system and topological connection relationship, mining face, goaf (distinguishing between waterless / waterlogged), waterlogged area, protective coal pillar, and sealed wall, and output vector recognition results with spatial coordinates. The comprehensive recognition accuracy can reach more than 95%.

[0042] (2) Safety measurement point identification and label association: Call the safety measurement point identification model trained in step S2 to simultaneously detect safety measurement points such as methane sensor, CO sensor, dust sensor, wind speed sensor, temperature sensor, damper switch sensor, and personnel positioning substation in the whole map; calculate the spatial distance between each measurement point symbol and the surrounding text labels through the GIS spatial proximity analysis algorithm, associate the closest and semantically matched text labels with the measurement point symbols, and complete the attribute assignment of the measurement points.

[0043] Functions and beneficial effects: (1) By adopting a two-level recognition strategy of "clean-up drawing training - enhanced training - comprehensive drawing recognition", the recognition accuracy of spatial structure elements of mining engineering plan map is increased from 40%-50% to more than 95%, breaking through the recognition accuracy bottleneck of full-element superimposed drawings and reaching the engineering practical standard; (2) The spatial proximity algorithm automatically matches the measurement points and text labels, replacing the manual verification of measurement point information, and realizes the structured output of the recognition results, which can be directly connected to the subsequent GIS system and analysis module.

[0044] Extended Steps: Specific Implementation Methods for Dual-Source Data Fusion and Automated Security Risk Analysis: (1) Unified coordinate system fusion: The spatial structure identification results and the safety measurement point identification results are mapped to the unified GIS coordinate system. The spatial alignment of the two types of elements is completed through coordinate matching, and a complete digital mine map model including roadway topology, working face boundary, goaf attributes, measurement point location and parameters is constructed.

[0045] (2) Rule-based hazard reasoning: The digital mine map model is input into the preset safety hazard rule knowledge base. The rules in the knowledge base come from industry standards such as the "Coal Mine Safety Regulations", "Coal Mine Water Prevention and Control Rules", "Coal Mine Rockburst Prevention and Control Rules", and "Coal Mine Safety Monitoring System and Detection Instrument Use and Management Specifications". They cover more than 20 judgment rules in five categories: insufficient roadway spacing, disordered mining and excavation continuity, abnormal roadway bends and intersections, unsealed abandoned roadways, and non-compliant safety measuring point layout. The reasoning engine automatically performs spatial geometric calculations such as roadway spacing verification, mining stress field superposition analysis, intersection detection, acute angle bend identification, sealed wall matching, measuring point compliance comparison, and prediction of the breakthrough distance between opposing tunnels. The hazards are divided into three levels according to their severity: major, general, and warning. Finally, a structured analysis report containing hazard type, location, regulatory basis, and rectification suggestions is output, which supports the display of visual heat map superposition.

[0046] Functions and benefits: It automates the entire process from drawing input to hazard report output, reducing the processing time for a single drawing to less than 5 minutes, which is about 100 times more efficient than traditional manual review; rule-based reasoning based on unified industry standards avoids the experience bias and inconsistent standards of manual review, supporting the real-time supervision needs of "drawings submitted for review".

[0047] This embodiment uses the scenario of submitting and reviewing drawings of a local state-owned high-gas mine as an example to verify the complete implementation process of this method and further illustrate the implementation process and effects of the present invention: Scenario Setting: The mine has a certified production capacity of 1.2 million tons / year, mining the No. 3 coal seam of the Permian Shanxi Formation, with an average thickness of 5.2m. It employs fully mechanized, one-pass, full-height mining technology (high-height fully mechanized mining), with a mining depth of approximately 650m. The hydrogeological conditions are moderate, and the spontaneous combustion tendency is classified as Class II. The submitted drawing package includes one mining engineering plan (DWG format, scale 1:2000, CGCS2000 coordinate system, approximately 82,000 elements, covering 18 layers including development, preparation, mining, ventilation, safety monitoring, electromechanical, transportation, drainage, dust control, compressed air, grouting, and gas extraction) and one drawing for each of the eight professional layout categories. All are DWG files of the same scale and coordinate system. The regulatory platform must immediately complete the automatic safety compliance review after the drawings are submitted.

[0048] Implementation steps and corresponding effects: 1. Receiving and classifying drawing packages After receiving the drawing package, the platform reads the layer information and drawing name field of each drawing: one drawing has 18 layers and the drawing name contains the words "Mining Engineering Plan", which is identified as a comprehensive drawing; the remaining 8 drawings have 3-7 layers each, and the drawing names correspond to personnel positioning system layout, electromechanical equipment layout, ventilation system layout, safety monitoring system layout, dust suppression sprinkler pipeline layout, compressed air self-rescue system layout, underground drainage system layout, and disaster avoidance route map, respectively, and complete the classification and archiving.

[0049] This step enables preliminary classification of drawing complexity, providing a basis for subsequent differential identification, avoiding redundant enhancement processing on low-complexity drawings, and improving overall processing efficiency.

[0050] 2. Construction of safety measurement point samples and model training (Mechanism 1) First, automatic vector parsing of the legend area is performed: using the CAD vector programming interface, the legend table in the lower right corner of the mining engineering plan frame is automatically located using keywords such as "legend," "symbol," "sensor," and "annotation," identifying 28 legend entries. The symbol graphic unit and text description unit of each legend are separated line by line, and the symbol graphic unit is exported as a 600DPI PNG image. Using the corresponding text description as a label, a structured safety measuring point symbol sample library is automatically constructed. This step takes approximately 2.1 seconds.

[0051] This process requires no manual intervention, improving efficiency by hundreds of times compared to the 2-8 hours required for manual annotation of a single drawing. It completely solves the industry vicious cycle of high cost of coal mine drawing annotation and the inability to establish a general sample library. Automated pre-training and fine-tuning are then performed: using the CLIP-ViT-L pre-trained visual model as a base, domain adaptation training is performed using the aforementioned 28 samples, undergoing 50 multi-scale enhancement iterations, totaling 1400 training steps. Then, 200 fine-tuning steps are performed, with small target detection enhancement optimization enabled during training to obtain a safety measurement point recognition model adapted to the mine's mapping style. This step takes approximately 45 seconds (hardware environment: NVIDIA RTX 3060).

[0052] Because the similarity between the mine's legend style and the system's existing sample library was below a preset threshold, the system automatically triggered readjustment. If the mine belonged to the same style cluster, the existing model could be reused directly without repeated training. In batch testing of 15 different coal mines, this mechanism could automatically cluster into 7 style clusters, requiring only 6 adaptive fine-tunings to complete the identification and adaptation of all mines, effectively solving the problem of cross-mine offset caused by the lack of a unified mapping standard in the industry. Finally, the safety monitoring point identification and labeling association were performed: using the fine-tuned model, safety monitoring points were detected across the entire map, identifying a total of 18 methane sensors (mean confidence level 0.97), 5 CO sensors (mean confidence level 0.96), 5 dust sensors (mean confidence level 0.94), 8 wind speed sensors (mean confidence level 0.98), 8 temperature sensors (mean confidence level 0.95), 12 damper switch sensors (mean confidence level 0.99), and 22 personnel positioning substations (mean confidence level 0.99). Each monitoring point was automatically matched with associated text labels using a GIS spatial proximity analysis algorithm, achieving a labeling association matching success rate of 96.3%. This step took approximately 6.8 seconds.

[0053] 3. Professional setup Figure 1 Level Identification (Mechanism Two, Step One) The eight categories of specialized layout maps obtained from the classification were subjected to first-level recognition one by one. Since the specialized layout maps only contain the core elements of the corresponding system, the layer structure is clean and there is very little interference, so the recognition accuracy can generally reach over 98%. This recognition generated a total of 1570 high-quality standardized training samples, including 1539 tunnel samples, 31 chamber samples, 22 sealed wall samples, and 9 waterlogged area samples. This step took about 40 seconds.

[0054] The recognition accuracy and number of generated samples for various professional layout diagrams are as follows: (1) Personnel positioning system layout diagram: recognition accuracy rate 99.1%, 186 samples generated; (2) Layout diagram of electromechanical equipment: recognition accuracy rate 98.7%, 235 samples generated; (3) Ventilation system diagram: recognition accuracy rate 98.5%, 220 samples generated; (4) Security monitoring system layout diagram: recognition accuracy rate 99.2%, 189 samples generated; (5) Dust control sprinkler pipeline layout diagram: recognition accuracy rate 98.9%, 191 samples generated; (6) Layout diagram of compressed air self-rescue system: recognition accuracy rate 99.0%, 183 samples generated; (7) Layout diagram of underground drainage system: recognition accuracy rate 98.2%, 196 samples generated; (8) Disaster evacuation route map: the identification accuracy rate was 99.3%, and 201 samples were generated.

[0055] This step obtains high-precision labeled data at extremely low computational cost, replacing the work of manually labeling spatial primitives and providing a reliable data foundation for subsequent mixed noise enhancement training.

[0056] 4. Hybrid noise enhancement pre-training (Mechanism 2, step 2) Based on the 1570 standardized clean samples obtained in the previous step, non-target layer interference elements such as contour lines, borehole symbols, and text annotations from the mining engineering plan were randomly superimposed. By using random positions, scaling, and transparency, a high-noise interference environment of multiple layers superimposed on the comprehensive map was simulated, and mixed noise enhancement training was performed. This training consisted of 3000 basic training steps and 500 fine-tuning steps, resulting in a spatial structure recognition model with strong anti-interference capabilities. This step took approximately 95 seconds.

[0057] Through this enhancement strategy, the model learns the ability to extract targets in complex backgrounds during the training phase, significantly improving its anti-interference capabilities and providing core technical support for the high-precision identification of subsequent mining engineering plan maps.

[0058] 5. Secondary Enhanced Recognition of Mining Engineering Plan View (Mechanism 2, Step 3) The trained spatial structure recognition model was invoked to perform secondary enhanced recognition on the mining engineering plan, extracting all spatial structure elements, including the roadway system and topological connections, mining faces, goaf (distinguishing between waterless and waterlogged areas), waterlogged areas, protective coal pillars, and sealed walls. Actual measurements showed that the comprehensive recognition accuracy of spatial structure elements reached 96.0%, while the baseline accuracy using a general model was only 43.2%. This solution improved the recognition accuracy by 52.8 percentage points, and this step took approximately 18.3 seconds.

[0059] The results validate the effectiveness of the hierarchical recognition + hybrid noise enhancement strategy, successfully breaking through the industry bottleneck of low recognition accuracy of full-element overlay comprehensive map, and enabling the recognition results to reach the engineering practical threshold for safety production review.

[0060] 6. Integration of dual-source results and comprehensive analysis of safety hazards.

[0061] The results of safety measurement point identification and spatial structure identification are matched and fused under the unified CGCS2000 Gauss-Kruger projection 3-degree zone coordinate system to form a complete digital mine map model that includes the centerline network and topological adjacency relationship of all roadways, working face outline, goaf boundary (distinguishing between waterless and waterlogged), location and type of sealed wall, location of all safety measurement points and associated annotation attributes.

[0062] The digital mine map model is input into a reasoning engine based on five categories of rule knowledge bases built on industry standards. This engine automatically performs spatial geometric calculations, including roadway spacing verification, mining-induced stress field superposition analysis, three-way intersection detection, acute-angle bend identification, sealed wall matching, measurement point compliance comparison, and prediction of the breakthrough distance between opposing tunnels. Ultimately, seven safety hazards were identified, as follows: 1. Category A - Insufficient spacing between tunnels (2 locations in total) (1) The actual width of the protective coal pillar between the transport roadway and the return air roadway is 28m, while the design requires a width of ≥35m. The actual width is only 80% of the design value. The calculated stress concentration factor is 2.1, which is considered a major hidden danger. It is recommended to expand the support to relieve pressure or reinforce it. (2) The actual width of the coal pillar (narrow coal pillar along the goaf) in the 105 and 106 longwall mining face sections is 6.8m, while the design requires a width of ≥8m. The actual width is 85% of the design value, which is judged as a general hidden danger. It is recommended to strengthen mine pressure monitoring.

[0063] 2. Category B - Disordered Mining Continuity (1 location in total) The tunneling directions of the four tunneling faces in Mining Area No. 3 (301 return air roadway tunneling, 301 transport roadway tunneling, 302 return air roadway tunneling, and mining area water tank tunneling) form a three-dimensional intersection in the middle of the mining area. It is expected to intersect at an elevation of -520m in three months. This is considered a general hidden danger. It is recommended to optimize the tunneling sequence and avoid multiple roadways crossing the same area at the same time.

[0064] 3. Category C - Abandoned roadways not sealed (1 location in total) The abandoned return airway (old roadway, about 180m long, and has been out of use for about 3 years) in the north wing of mining area No. 3 has no airtight wall markings at both ends of its boundary, and is only 65m away from the main transport roadway in use. It is determined to be a major hidden danger. It is recommended to immediately construct fireproof / waterproof airtight walls and test the gas and water accumulation in the old roadway.

[0065] 4. Class E safety monitoring points were not installed in compliance with regulations (3 locations in total). (1) No methane sensor was installed at the upper corner (U-shaped ventilation return air corner) of the 105 longwall mining face, which was determined to be a major hidden danger; the absolute gas emission rate of this working face is 3.8 m³ / h. 3 / min, the historical maximum gas concentration in the upper corner has reached 1.42%, which has exceeded the alarm concentration and must be installed immediately; (2) The ventilation from the 301 tunneling face to the 302 tunneling face was connected in series. No methane sensor was added to the air intake of the 302 tunneling face, which violated Article 172 of the Coal Mine Safety Regulations. It was determined to be a general hidden danger. It is recommended to stop tunneling the 302 tunneling face or add a methane sensor immediately. (3) The dust sensor at the 105 working face is installed at a distance of 28m from the working face, which exceeds the 10-15m range required by the regulations. It is judged as a general hidden danger. It is recommended to move the sensor forward and recalibrate it.

[0066] The system automatically outputs a structured hazard analysis report, which includes hazard type, hazard level, location information, regulatory basis, and rectification suggestions. It also supports the overlay output of visual heat maps.

[0067] The entire process in this embodiment takes approximately 208 seconds (3.5 minutes), while the traditional manual review of drawings of the same scale (including distance measurement, comparison of regulations and clauses, and writing review comments) is expected to take 6-8 hours. This solution is about 100 times more efficient than the traditional manual review, fully meeting the real-time supervision requirements of "reviewing drawings upon submission", and verifying the engineering practical value and technological advancement of this method.

[0068] As can be seen from the above, this invention provides an intelligent recognition method for coal mine drawings based on multi-source automated pre-training, which solves the problem that the accuracy of direct AI recognition is low due to the overlay of all elements in existing mining engineering plan drawings, and cannot meet the practical requirements of engineering. At the same time, it addresses the problems of the lack of unified drafting standards in the industry, high cost of manual annotation, and serious cross-mine area offset. The specific beneficial effects are as follows: 1. Completely eliminates the bottleneck of manual annotation and achieves zero-manual sample construction. This invention uses a legend-driven vector automatic parsing mechanism to directly extract symbols and corresponding labels from CAD vector drawings, automatically constructing a safety measurement point sample library. The entire process requires no manual annotation, breaking the vicious cycle of "high annotation costs and inability to establish a general sample library" in the coal mining industry, and significantly reducing the manpower and time costs of model implementation.

[0069] 2. A significant leap in recognition accuracy, reaching practical engineering standards. This invention utilizes the inherent layered characteristics of coal mine drawings—"clean professional drawings, complex comprehensive drawings"—and employs a "professional" approach... Figure 1 The hierarchical strategy of "secondary identification sampling - mixed noise enhancement training - comprehensive map secondary identification" improves the recognition accuracy of spatial structure elements in mining engineering plan from about 45% to over 95%, and the recognition accuracy of safety measurement points is no less than 98%, meeting the engineering practical threshold for safety production review.

[0070] 3. Adaptive solution to cross-mine offset problem, with strong versatility. This invention has a built-in legend difference adaptive mechanism, which uses geometric similarity to determine the reuse or fine-tuning of the dynamic decision model, eliminating the need to train a separate model for each coal mine; in batch testing, it can automatically cluster coal mines with different mapping styles, covering the needs of multi-mine identification with very few fine-tunings, adapting to the current situation where there is no unified mapping standard in the industry.

[0071] 4. Fully automated operation to meet real-time review requirements. This invention automates the entire process from drawing input to recognition result output. The entire process for a single drawing takes only 3-4 minutes. When combined with a rule knowledge base, it can directly generate a hazard analysis report, which is about 100 times more efficient than traditional manual review and can support the regulatory scenario requirement of "drawings submitted for review immediately".

[0072] 5. High architectural compatibility, supporting deployment in multiple scenarios. This method supports a degraded recognition mode, achieving an accuracy rate of 85%-90% even without a corresponding professional layout diagram. The recognition results can be output in a standardized GIS data format, directly compatible with various business systems such as grounding information systems and security supervision platforms. It also supports predictive security analysis, making it flexible in application scenarios.

[0073] The specific embodiments described above are preferred embodiments of the present invention and are not intended to limit the specific scope of the present invention. The scope of the present invention includes, but is not limited to, these specific embodiments. All equivalent changes made in accordance with the present invention are within the protection scope of the present invention.

Claims

1. A method for intelligent recognition of coal mine drawings based on multi-source automated pre-training, characterized in that, Includes the following steps: Step S1: Receive a vector file package containing multiple coal mine drawings and automatically identify and distinguish between mining engineering plan and various professional layout drawings; Step S2: Automatically locate the legend area in the coal mine drawing vector file, parse the symbol graphic unit and text description unit corresponding to each legend, use the text description as a label and the symbol graphic unit as a sample to automatically construct a safety measurement point symbol sample library; use the pre-trained visual detection model as a base and use the safety measurement point symbol sample library to perform domain adaptation fine-tuning to obtain a safety measurement point recognition model. Step S3: Perform primitive recognition on the clean professional layout diagram of the classified layers, extract spatial structure primitives and crop them to generate standardized training samples; Step S4: Add non-target layer interference elements of the mining engineering plan to the standardized training samples to simulate a high-noise environment with multiple layers and perform enhancement training to obtain a spatial structure recognition model. Step S5: Use the spatial structure recognition model to identify spatial structure elements in the mining project plan and output the spatial structure class recognition result; at the same time, use the safety measuring point recognition model to identify safety measuring points in the mining project plan and output the safety measuring point class recognition result.

2. The intelligent recognition method for coal mine drawings based on multi-source automated pre-training as described in claim 1, characterized in that: In step S2, the legend area is located by keyword matching, and the keywords include legend, symbol, sensor, and annotation description; The parsed symbolic graphic units are exported as image formats with a resolution of at least 600 DPI for use in building a sample library.

3. The intelligent recognition method for coal mine drawings based on multi-source automated pre-training as described in claim 1, characterized in that: In step S2, when processing drawings of different coal mines, the degree of difference between the legend of the current drawing and the existing sample library is determined by the pixel geometric similarity. If the similarity meets the preset threshold, the existing security measurement point recognition model can be directly reused; If the conditions are not met, the process reverts to the pre-trained visual detection model base and re-executes fine-tuning, forming an adaptive fine-tuning pipeline.

4. The intelligent recognition method for coal mine drawings based on multi-source automated pre-training as described in claim 1, characterized in that: In step S5, after obtaining the identification results of the safety measurement points, the GIS spatial proximity analysis algorithm is used to automatically match and associate each safety measurement point with its corresponding text label.

5. The intelligent recognition method for coal mine drawings based on multi-source automated pre-training as described in claim 1, characterized in that: The professional layout diagrams include one or more of the following: personnel positioning layout diagram, electromechanical equipment layout diagram, ventilation system diagram, safety monitoring system layout diagram, dust control pipeline layout diagram, compressed air self-rescue system layout diagram, underground drainage system layout diagram, and disaster evacuation route diagram.

6. The intelligent recognition method for coal mine drawings based on multi-source automated pre-training as described in claim 1, characterized in that: In step S4, the non-target layer interference elements include one or more of contour lines, borehole symbols, and text annotations, used to simulate a high-noise interference environment of multiple layers superimposed on the mining engineering plan.

7. The intelligent recognition method for coal mine drawings based on multi-source automated pre-training as described in claim 1, characterized in that: The spatial structure identification results include one or more of the following: roadway system and topological connection relationship, mining face, goaf, water accumulation area, protective coal pillar, and sealed wall.

8. The intelligent recognition method for coal mine drawings based on multi-source automated pre-training according to claim 1, characterized in that, Following step S5, the method further includes: performing coordinate matching and data fusion of the spatial structure class identification results and the safety measurement point class identification results under a unified GIS coordinate system to generate a digital mining map model containing topological attributes and feature attributes.

9. The intelligent recognition method for coal mine drawings based on multi-source automated pre-training as described in claim 1, characterized in that: The digital mining map model is input into a preset safety hazard rule knowledge base, and spatial geometric calculations and compliance checks are performed through an inference engine to output a structured safety hazard analysis report.

10. The intelligent recognition method for coal mine drawings based on multi-source automated pre-training according to any one of claims 1-9, characterized in that: The rules in the safety hazard rule knowledge base are derived from industry standards and cover five major types of hazards: insufficient roadway spacing, disordered mining continuity, abnormal roadway structure, unsealed abandoned roadways, and non-compliant safety monitoring point layout. The safety hazard analysis report includes the hazard type, hazard level, location information, procedural basis, and rectification suggestions.

Images