Machine vision-based intelligent planning method and system for cutting track of tunneling equipment

Abstract

The application provides a machine vision-based intelligent planning method and system for a cutting track of a tunneling device, and belongs to the field of machine vision. The method comprises the following steps: acquiring continuous machine vision sensing data of a tunneling face area, and establishing a multi-scale visual feature expression; constructing a spatial semantic field based on the multi-scale visual feature expression; constructing a spatial gradient distribution feature by using the spatial semantic field; extracting a local cutting main direction, projecting the spatial gradient distribution feature along the local cutting main direction, and generating a comprehensive cutting trend evaluation value; taking a current cutting position as a starting point, iteratively expanding the track according to the comprehensive cutting trend evaluation value, and establishing a continuously evolving cutting track. Through multi-scale perception and cuttability quantitative evaluation of the rock mass structure of the working face, the cutting track is generated along the optimal direction of the rock mass cutting, the cutting efficiency is improved, and the stress of the cutter is balanced.

Description

Technical Field

[0001] This invention relates to the field of machine vision, and more particularly to a method and system for intelligent planning of cutting trajectories for tunneling equipment based on machine vision. Background Technology

[0002] Currently, the cutting trajectory planning of tunneling equipment mainly relies on preset programs. That is, the movement path of the cutting head is pre-set according to the cross-sectional shape and design dimensions of the tunnel, and the equipment performs reciprocating cutting according to a fixed trajectory pattern. This method can meet the basic tunneling requirements under homogeneous rock conditions. However, in actual engineering, the geological conditions of the working face rock mass often have significant heterogeneity and uncertainty. Rock masses in different areas have significant differences in hardness, degree of joint development, bedding orientation, and structural integrity, and the cutting difficulty and suitable cutting direction also vary accordingly.

[0003] However, the preset trajectory planning method cannot perceive the actual structural characteristics of the rock mass at the working face in real time, nor can it identify the differences in machinability distribution between different areas. The cutting head can only perform mechanical cutting according to a fixed path and sequence. When the cutting head moves to areas with high hardness or complex structure, the trajectory is not adjusted accordingly, which can easily lead to a sudden increase in cutting resistance and uneven force on the tool. When the cutting direction does not match the orientation of the rock mass joints or bedding, the cutting efficiency is significantly reduced. Summary of the Invention

[0004] This invention addresses the technical problem in existing technologies where the cutting trajectory planning of tunneling equipment mainly relies on preset programs and cannot adaptively adjust according to the actual structural characteristics and machinability differences of the working face rock mass, resulting in low cutting efficiency and uneven force on the cutting tools. It provides a machine vision-based intelligent cutting trajectory planning method and system for tunneling equipment to solve this problem.

[0005] To address the aforementioned technical problems, this invention provides an intelligent planning method for cutting trajectories of tunneling equipment based on machine vision, comprising: acquiring continuous machine vision perception data of the tunneling face area, wherein the machine vision perception data includes surface texture information, structural morphology information, and spatial distribution information of the face; performing multi-scale feature parsing processing based on the machine vision perception data to extract regional-level structural features, mesoscale structural features, and micro-texture features respectively, and performing multi-scale feature spatial alignment and fusion to establish a multi-scale visual feature expression; constructing a spatial semantic field based on the multi-scale visual feature expression, wherein each position in the spatial semantic field is associated with a cutting difficulty index, a structural stability index, and a cutting direction trend index; calculating the rate of change of the indexes of adjacent spatial positions using the cutting difficulty index and the structural stability index to construct a spatial gradient distribution feature; extracting the local main cutting direction based on the cutting direction trend index, and projecting the spatial gradient distribution feature along the local main cutting direction to generate a comprehensive cutting trend evaluation value; and iteratively expanding the trajectory based on the comprehensive cutting trend evaluation value, starting from the current cutting position, to establish a continuously evolving cutting trajectory.

[0006] Furthermore, this invention also provides an intelligent planning system for the cutting trajectory of tunneling equipment based on machine vision, comprising: a visual perception acquisition module for acquiring continuous machine vision perception data of the tunneling working face area, wherein the machine vision perception data includes surface texture information, structural morphology information, and spatial distribution information of the working face; a multi-scale feature parsing module for performing multi-scale feature parsing processing based on the machine vision perception data, extracting regional-level structural features, mesoscale structural features, and micro-texture features respectively, and performing multi-scale feature spatial alignment and fusion to establish a multi-scale visual feature representation; and a spatial semantic field construction module for constructing a spatial semantic field based on the multi-scale visual feature representation. A spatial semantic field is constructed, in which each position is associated with a cutting difficulty index, a structural stability index, and a cutting direction trend index; a gradient distribution construction module is used to calculate the rate of change of the indexes of adjacent spatial positions using the cutting difficulty index and the structural stability index, and construct a spatial gradient distribution feature; a cutting trend evaluation module is used to extract the local main cutting direction based on the cutting direction trend index, and project the spatial gradient distribution feature along the local main cutting direction to generate a comprehensive cutting trend evaluation value; a trajectory iteration planning module is used to iteratively expand the trajectory based on the comprehensive cutting trend evaluation value, starting from the current cutting position, and establish a continuously evolving cutting trajectory.

[0007] The beneficial effects of this invention are:

[0008] Continuous machine vision perception data of the tunneling face area was acquired, including surface texture, structural morphology, and spatial distribution information, providing a foundation of raw visual data for subsequent rock mass feature analysis. Multi-scale feature parsing was performed on the machine vision perception data to extract regional-level structural features, mesoscale structural features, and micro-texture features. Multi-scale feature spatial alignment and fusion were then performed to establish a multi-scale visual feature representation, enabling the complete characterization of rock mass information from macroscopic structure to microscopic texture within a unified spatial coordinate system. Based on this, a spatial semantic field was constructed using the multi-scale visual feature representation. Each location in the spatial semantic field is associated with cutting difficulty indicators, structural stability indicators, and cutting direction trend indicators, transforming visual features into quantitative semantic information directly reflecting the rock mass's machinability and providing a decision-making basis for trajectory planning. The rate of change of indicators at adjacent spatial locations was calculated using the cutting difficulty and structural stability indicators to construct a spatial gradient distribution feature, revealing the spatial variation trend and transition law of rock mass machinability. Based on the cutting direction trend index, the local main cutting direction is extracted, and the spatial gradient distribution characteristics are projected along the local main cutting direction to generate a comprehensive cutting trend evaluation value. The gradient change information and the direction optimization information are coupled into a unified trajectory advancement evaluation criterion. Then, starting from the current cutting position, the trajectory is iteratively expanded according to the comprehensive cutting trend evaluation value to establish a continuously evolving cutting trajectory, which can adaptively extend and generate the cutting trajectory along the optimal direction of rock mass cutting resistance step by step.

[0009] The above technical solution enables multi-scale perception and quantification of machinability of the working face rock structure based on machine vision, and enables the cutting trajectory to adaptively evolve and generate along the optimal direction of rock cutting resistance, thereby improving cutting efficiency, balancing the force on the tool, and achieving trajectory rationalization. Attached Figure Description

[0010] Figure 1 A flowchart illustrating the intelligent planning method for cutting trajectories of tunneling equipment based on machine vision provided by the present invention;

[0011] Figure 2 This is a schematic diagram of the process for generating multi-scale visual feature representations provided by the present invention;

[0012] Figure 3 A schematic diagram of the process for establishing machine vision perception data provided by the present invention;

[0013] Figure 4 This is a schematic diagram of the intelligent planning system for the cutting trajectory of tunneling equipment based on machine vision provided by the present invention.

[0014] In the attached diagram, the components represented by each number are as follows:

[0015] Visual perception acquisition module 11, multi-scale feature parsing module 12, spatial semantic field construction module 13, gradient distribution construction module 14, cutting trend evaluation module 15, trajectory iterative planning module 16. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Example 1, as Figure 1 As shown, this embodiment of the invention provides a machine vision-based intelligent planning method for the cutting trajectory of tunneling equipment, including:

[0018] S1. Acquire continuous machine vision perception data of the tunneling face area, wherein the machine vision perception data includes surface texture information, structural morphology information and spatial distribution information of the face.

[0019] During tunneling operations, the cutting trajectory planning of the tunneling equipment is based on the actual state information of the rock mass at the working face. Therefore, it is necessary to collect machine vision perception data of the tunneling working face area. Machine vision perception data is raw visual data obtained by performing image acquisition and depth perception on the tunneling working face area. It includes image data recording the visual characteristics of the working face surface and depth information reflecting the spatial distances of various areas of the working face. Machine vision perception data contains surface texture information, structural morphology information, and spatial distribution information of the working face.

[0020] Surface texture information of the working face refers to the visual characteristics of the rock mass surface as revealed by the acquired image data. In image data, surface texture information is represented by pixel-level grayscale distribution, color channel response, and local texture repetition characteristics. The grayscale variation patterns, color transition rules, and texture density of different areas in the image reflect the differences in the material composition of the corresponding rock mass surface. For example, dense rock masses with high hardness usually have a fine and uniform surface texture and gentle grayscale changes; while areas containing weak interlayers or with a high degree of weathering show obvious unevenness in surface texture, abrupt color changes, or disordered texture direction.

[0021] Structural morphology information refers to the spatial geometric features of the working surface jointly reflected in image data and depth information. In image data, structural morphology information is reflected through edge features, contour distribution, and gradients of light and shadow, which reflect the exposure traces of structural surfaces such as cracks and joints on the working surface. In depth information, structural morphology information is reflected through local depth undulations, discontinuous depth regions, and abrupt changes in depth gradient, which reflect the degree of crack opening, the attitude of joint surfaces, and the spatial trend of bedding interfaces.

[0022] Spatial distribution information is the positional coordinates of each region of the working surface in three-dimensional space and the spatial geometric relationships between these regions, determined by performing spatial coordinate calculations on depth information. Specifically, based on depth information, the distance measurements of each point on the working surface in the depth direction are obtained, and combined with the pixel position relationships of each point in the image data, spatial geometric calculations are used to determine the three-dimensional spatial coordinates of each point, thereby establishing a three-dimensional spatial position description of the working surface region. Spatial distribution information provides a unified spatial positioning benchmark, enabling all features described in the surface texture and structural morphology information of the working surface to be mapped to a defined three-dimensional spatial position.

[0023] The acquisition of machine vision perception data is continuous, meaning that visual data of the working face area is continuously acquired during the operation of the tunneling equipment. Continuous acquisition ensures that the acquired visual data always reflects the latest state of the working face and provides a data foundation for dynamic response during the cutting process.

[0024] S2. Based on the machine vision perception data, perform multi-scale feature parsing processing to extract regional structural features, mesoscale structural features and micro-texture features respectively, and perform multi-scale feature space alignment and fusion to establish multi-scale visual feature representation.

[0025] Multi-scale feature parsing processing involves extracting and parsing features from machine vision perception data at regional, meso-scale, and micro-scale, generating corresponding regional-level structural features, meso-scale structural features, and micro-texture features.

[0026] After extracting regional-level structural features, mesoscale structural features, and microscale texture features, multi-scale feature spatial alignment and fusion processing is performed. Spatial alignment uses the three-dimensional spatial coordinate description established by the obtained spatial distribution information as a reference, uniformly mapping features at each scale to the same spatial coordinate system, so that features at different scales can correspond to each other in spatial location.

[0027] Based on spatial alignment, a fusion process is performed to integrate and associate features at various scales under a unified spatial coordinate system, establishing a multi-scale visual feature representation. This multi-scale visual feature representation is a feature description system that is spatially continuous and scale-complete. Each spatial location is simultaneously associated with regional-level structural features, mesoscale structural features, and microscale texture features, providing a feature foundation for the subsequent construction of a spatial semantic field.

[0028] S3. Construct a spatial semantic field based on the multi-scale visual feature expression, wherein each position in the spatial semantic field is associated with a cutting difficulty index, a structural stability index, and a cutting direction tendency index.

[0029] After establishing multi-scale visual feature representations, a spatial semantic field is constructed based on these representations. The spatial semantic field is a continuous semantic description field covering the entire working face area, where each spatial location is associated with three semantic indicators characterizing the rock mass's cutability: cutting difficulty index, structural stability index, and cutting direction tendency index.

[0030] The cutting difficulty index, structural stability index, and cutting direction tendency index are all derived from regional-level structural features, mesoscale structural features, and micro-texture features in multi-scale visual feature representation. Features at each scale provide information support for the generation of the indices at different levels. The three indices quantify the cutability of rock masses at each location from three dimensions: cutting resistance, structural safety, and cutting direction suitability. Specifically, the cutting difficulty index represents the magnitude of the cutting resistance that needs to be overcome when the rock mass at the corresponding spatial location is cut. The structural stability index represents the ability of the rock mass at the corresponding spatial location to maintain structural integrity after being cut, i.e., the degree of risk of abnormal collapse or excessive fragmentation of the rock mass at that location during the cutting process. The cutting direction tendency index describes, in vector form, the suitable cutting direction indicated by the rock mass structure at the corresponding spatial location.

[0031] The spatial semantic field transforms the structural and material information of the visual level contained in the multi-scale visual feature expression into a semantic description reflecting the rock mass's machinability. This allows subsequent trajectory planning to no longer rely on indirect inference from the original visual features, but instead make decisions based on explicit cutting performance indicators.

[0032] S4. Calculate the rate of change of indicators of adjacent spatial positions using cutting difficulty indicators and structural stability indicators, and construct spatial gradient distribution characteristics.

[0033] The spatial semantic field covers the entire working surface area, where the spatial coordinates of each location are determined by the spatial distribution information obtained in S1. Each location is associated with a cutting difficulty index value and a structural stability index value, forming a cutting difficulty scalar field and a structural stability scalar field that are continuously distributed in space.

[0034] For the scalar field of cutting difficulty, for each spatial location, determine its adjacent locations in three-dimensional spatial coordinates. For the current location and a certain adjacent location, calculate the difference in the cutting difficulty index value between the two locations, and then divide it by the spatial distance between the two locations to obtain the rate of change of the cutting difficulty index between the current location and that adjacent location. The coordinate relationship between the current location and that adjacent location determines the spatial direction corresponding to this rate of change of the cutting difficulty index.

[0035] A positive rate of change in the cutting difficulty index indicates that the cutting difficulty increases along that spatial direction; a negative rate of change indicates that the cutting difficulty decreases along that spatial direction; the larger the absolute value of the rate of change in the cutting difficulty index, the more drastic the change in cutting difficulty along that spatial direction.

[0036] Perform the above calculations on the current position and all its adjacent positions to obtain the rate of change of the cutting difficulty index between the current position and each adjacent position, as well as their respective spatial directions. Combine the rate of change of the cutting difficulty index corresponding to each adjacent position with the spatial direction of that adjacent position relative to the current position to form a vector of the rate of change of cutting difficulty in that adjacent position direction.

[0037] By combining the cutting difficulty change rate vectors of all adjacent positions at the current position, the cutting difficulty gradient vector at that position is obtained. The direction of this cutting difficulty gradient vector is the direction in which the cutting difficulty changes the fastest in space at that position, and the magnitude is the degree of drastic change in cutting difficulty in that direction.

[0038] By performing the above calculations at all locations in the cutting difficulty scalar field, the cutting difficulty gradient vector distribution covering the entire working surface area can be obtained.

[0039] The structural stability scalar field is processed using the same local gradient calculation method as the cutting difficulty scalar field, resulting in a structural stability gradient vector distribution covering the entire working surface area. The direction of the structural stability gradient vector indicates the direction in which the structural stability changes the fastest in space at that location, and the magnitude indicates the degree of drastic change in that direction.

[0040] After obtaining the cutting difficulty gradient vector and structural stability gradient vector at each location, the cutting difficulty gradient vector and structural stability gradient vector at the same location are correlated and combined to form the spatial gradient distribution feature of that location. That is, the spatial gradient distribution feature of each spatial location is jointly constituted by the cutting difficulty gradient vector and the structural stability gradient vector at that location. By performing the above correlation and combination on all locations within the entire working surface area, a spatial gradient distribution feature covering the entire working surface area can be obtained.

[0041] S5. Extract the local main cutting direction based on the cutting direction trend index, and project the spatial gradient distribution characteristics along the local main cutting direction to generate a comprehensive cutting trend evaluation value.

[0042] The extraction of the dominant cutting direction involves statistically analyzing the directional distribution of cutting direction indices at various locations within the spatial neighborhood of the target location. This analysis identifies the most frequent and representative dominant cutting direction within the area, which serves as the dominant local cutting direction. The dominant local cutting direction represents the overall directional trend of the rock mass structure within that local area, providing a directional reference for subsequent projection calculations.

[0043] After determining the principal direction of local cutting, the spatial gradient distribution characteristics are projected along this direction. For each spatial location, the cutting difficulty gradient vector at that location is projected along the principal direction of local cutting to obtain the component value of the cutting difficulty gradient vector along this direction. This component value reflects the changing trend of cutting difficulty as the cutting difficulty progresses along the principal direction. Similarly, the structural stability gradient vector at that location is projected along the principal direction of local cutting to obtain the component value of the structural stability gradient vector along this direction. This component value reflects the changing trend of structural stability as the cutting difficulty progresses along the principal direction.

[0044] The projection components of the cutting difficulty gradient vector and the structural stability gradient vector onto the local cutting principal direction are combined to generate a comprehensive cutting trend evaluation value for that location. A better comprehensive cutting trend evaluation value indicates that advancing along that direction is more favorable for cutting conditions, providing a decision-making basis for subsequent trajectory iterative expansion starting from the current cutting position.

[0045] S6. Starting from the current cutting position, the trajectory is iteratively expanded based on the comprehensive cutting trend evaluation value to establish a continuously evolving cutting trajectory.

[0046] The cutting trajectory starts from the current cutting position and gradually determines the direction of the trajectory and the next trajectory node based on the comprehensive cutting trend evaluation value. Through continuous iteration, the trajectory is extended forward to establish a continuously evolving cutting trajectory.

[0047] In each step of the trajectory iterative expansion, the foremost position of the current trajectory is taken as the current cutting position, and several candidate expansion points are determined within the spatial neighborhood of this position. Each candidate expansion point corresponds to a comprehensive cutting trend evaluation value for that position. The comprehensive cutting trend evaluation value reflects the changing trend of cutting resistance, the changing trend of surrounding rock stability, and the degree of matching between the advancing direction and the suitable cutting direction of the rock mass structure when advancing from the current cutting position towards the candidate expansion point. Based on the comparison and evaluation of the comprehensive cutting trend evaluation values ​​of each candidate expansion point, the candidate expansion point with the best comprehensive cutting trend evaluation value is selected as the next node of the trajectory, and this node is included in the cutting trajectory. Subsequently, using this new node as the new current cutting position, the process of determining candidate expansion points, comparing evaluation values, and selecting the optimal point is repeated to continue extending the trajectory forward. Through this trajectory iterative expansion method, starting from the initial current cutting position, the trajectory gradually evolves and generates along the optimal direction indicated by the comprehensive cutting trend evaluation value, eventually forming a continuous cutting trajectory.

[0048] As an optional implementation, the spatial gradient distribution features are projected along the local cutting principal direction to generate a comprehensive cutting trend evaluation value, including:

[0049] S51. Based on the spatial gradient distribution characteristics, identify the continuity of gradient change along the local cutting main direction, detect gradient abrupt change positions, and use the gradient abrupt change positions as boundaries to divide the current spatial region into cutting sub-regions.

[0050] S52. Calculate the projection change trend of cutting difficulty index and structural stability index on the local cutting main direction for each cutting sub-region, and determine the corresponding local cutting mode based on the consistency of gradient change in each cutting sub-region.

[0051] S53. Based on the local cutting mode corresponding to each cutting sub-region, perform segmented constraint fusion processing on the trend of cutting difficulty change and the trend of structural stability change to generate a comprehensive cutting trend evaluation value.

[0052] When generating a comprehensive cutting trend evaluation value, firstly, along the main local cutting direction, the spatial gradient distribution characteristics of each spatial location are examined sequentially. The changes in the cutting difficulty gradient vector and structural stability gradient vector between adjacent locations are analyzed to ensure they remain continuous and gradual. When the gradient vector between adjacent locations exhibits a sharp change in magnitude or direction that exceeds the normal gradual change range, that location is identified as a gradient abrupt change location. Using the detected gradient abrupt change locations as boundaries, the current spatial region is divided into several cutting sub-regions, ensuring that the gradient changes within each cutting sub-region remain relatively continuous and consistent. Different cutting sub-regions are separated by gradient abrupt change locations.

[0053] After dividing the cutting sub-regions, the projection trends of the cutting difficulty index and structural stability index on the local main cutting direction are calculated for each sub-region. For a given cutting sub-region, the projection components of the cutting difficulty gradient vector at each location within the sub-region are sequentially taken along the local main cutting direction. The variation trend of these projection components along the local main cutting direction is analyzed, thus obtaining the projection trend of the cutting difficulty in the local main cutting direction within that sub-region. Similarly, the projection components of the structural stability gradient vector at each location within the sub-region are taken along the local main cutting direction to obtain the projection trend of the structural stability in the local main cutting direction within that sub-region.

[0054] Based on this, and considering the consistency between the trends of the projected changes in cutting difficulty and the projected changes in structural stability within each cutting sub-region, the local cutting mode corresponding to that sub-region is determined. The local cutting mode is a summary description of the characteristics of the changing cutting conditions within that sub-region. For example, if the cutting difficulty continuously decreases and the structural stability continuously increases within a certain cutting sub-region, then the local cutting mode for that sub-region is a favorable advancement mode; if the cutting difficulty continuously increases and the structural stability decreases within a certain cutting sub-region, then the local cutting mode is an unfavorable advancement mode; if the cutting difficulty changes gradually within a certain cutting sub-region but the structural stability fluctuates significantly, then the local cutting mode is a stability-sensitive mode.

[0055] Based on the local cutting modes corresponding to each cutting sub-region, a segmented constraint fusion process is performed on the trends of cutting difficulty and structural stability to generate a comprehensive cutting trend evaluation value. The segmented constraint fusion process refers to using different fusion constraint strategies for cutting sub-regions with different local cutting modes to integrate the trends of projected cutting difficulty and projected structural stability into a unified comprehensive cutting trend evaluation value.

[0056] For example, for cutting sub-regions with favorable propulsion modes, the weights of the cutting difficulty projection component and the structural stability projection component are kept similar during fusion, so that the generated comprehensive cutting trend evaluation value reflects the changing trends of both in a balanced way. For cutting sub-regions with unfavorable propulsion modes, the weight of the structural stability projection component is increased during fusion, so that the generated comprehensive cutting trend evaluation value reflects the risks brought about by the decline in structural stability. For cutting sub-regions with stability-sensitive modes, the weight of the structural stability projection component is further increased during fusion, so that the generated comprehensive cutting trend evaluation value responds preferentially to the fluctuations in structural stability.

[0057] Through segmented constraint fusion processing, the comprehensive cutting trend evaluation value generated in different cutting sub-regions has differentiated decision characteristics, and can provide trajectory advancement evaluation for the actual cutting conditions of each sub-region.

[0058] As an optional implementation method, such as Figure 2As shown, multi-scale feature parsing processing is performed based on the machine vision perception data, including:

[0059] S21. Construct the depth information distribution of the working surface area based on the machine vision perception data, and generate a three-dimensional surface morphology expression corresponding to the image data space.

[0060] S22. Based on the three-dimensional surface morphology representation, perform scaled structural interpretation processing on the working surface region, wherein:

[0061] S23. At the regional scale, identify the overall structural partitions of the working face based on depth continuity and morphological undulation features, and extract regional structural features.

[0062] S24. At the mesoscale, based on the depth gradient direction distribution and local geometric change features, information on crack orientation and structural extension is extracted to form mesoscale structural features.

[0063] S25. At the microscale, surface roughness and local heterogeneous features are extracted based on the coupling relationship between image texture intensity distribution and depth perturbation changes to form microtexture features.

[0064] S26. Map the features at each scale to a unified spatial coordinate system, and perform weighted fusion processing based on the structural consistency constraints between features at different scales to generate multi-scale visual feature representations.

[0065] Machine vision perception data includes image data and depth information. Image data records the visual characteristics of the working surface in the form of a two-dimensional pixel array, while depth information records the spatial distances between different areas of the working surface. Depth information distribution is a description of the depth distribution covering the entire working surface area, formed by organizing and arranging depth information according to spatial relationships, where each location corresponds to a depth measurement value. Based on the depth information distribution and combined with the positional relationships of each pixel in the image data, the depth measurements at each location are converted into three-dimensional spatial coordinates, thereby generating a three-dimensional surface morphology representation that corresponds one-to-one with the image data in space.

[0066] After establishing the three-dimensional surface morphology representation, the working surface area is subjected to multi-scale structural interpretation processing, and features are extracted at the regional scale, mesoscale and microscale respectively.

[0067] At the regional scale, the overall structural partitions of the working face are identified based on depth continuity and morphological undulation characteristics, and regional structural features are extracted. In the three-dimensional surface morphology representation, the depth distribution across the entire working face area is analyzed for continuity, identifying regions where depth values ​​maintain a continuous gradient in space and locations where significant abrupt changes or breaks occur. Regions with continuous depth gradients typically correspond to the same rock stratum or rock masses with similar structural properties, while depth abrupt changes or breaks typically correspond to interfaces or structural boundaries between different rock strata. Simultaneously, the large-scale morphological undulation characteristics in the three-dimensional surface morphology representation are analyzed to identify the distribution patterns of uplifts, depressions, and flat areas on the overall surface. Different morphological undulation patterns reflect the spatial occurrence and distribution of different rock strata. Combining the depth continuity analysis results with the morphological undulation characteristic analysis results, the overall structural partitions of the working face are identified, the spatial extent of each partition and the boundary positions between partitions are determined, forming regional structural features that characterize the layered distribution and structural boundaries of the rock mass.

[0068] At the mesoscale, information on fracture orientation and structural extension is extracted based on the distribution of depth gradient directions and local geometric changes, forming mesoscale structural features. Specifically, in the representation of 3D surface morphology, the depth gradient is calculated for each local region of the working surface, obtaining the distribution of depth gradient directions at each location. The depth gradient direction reflects the direction of the most drastic surface morphological change at that location. When the depth gradient directions of multiple adjacent locations show a regular arrangement along a certain direction, it usually indicates the presence of structural surfaces such as fractures or joints extending along that direction in that region. Simultaneously, the geometric change characteristics of each local region are analyzed, including abrupt changes in surface curvature, the distribution of local grooves or ridges, etc. Geometric change characteristics are the geometric representation of fracture opening, joint surface exposure, and structural surface intersections on the 3D surface. By combining the arrangement of depth gradient directions with local geometric change characteristics, the propagation direction of fractures, the extension trend of joint surfaces, and the spatial intersection relationships between different structural surfaces within each local region are extracted, forming mesoscale structural features.

[0069] At the microscale, surface roughness and local heterogeneity features are extracted based on the coupling relationship between image texture intensity distribution and depth perturbation variation, forming microscale texture features. Specifically, the texture intensity distribution of each local region is extracted from the image data. Texture intensity reflects the intensity of grayscale changes in that region; regions with high texture intensity exhibit rich surface visual variations, while regions with low texture intensity show gentler surface visual variations. Simultaneously, depth perturbation variations of each local region are extracted from the 3D surface morphology representation; these are the subtle fluctuations in depth values ​​within a small spatial range, reflecting the geometric undulations of the surface at the microscale. A coupling relationship exists between image texture intensity distribution and depth perturbation variation: regions with high surface roughness typically have high image texture intensity and significant depth perturbation variation; regions with high material heterogeneity exhibit irregular spatial fluctuations in texture intensity distribution, and correspondingly irregular fluctuations in depth perturbation variation. Based on this coupling relationship, image texture intensity distribution and depth perturbation variation are jointly analyzed to evaluate the surface roughness level and material heterogeneity of each local region, forming microscale texture features.

[0070] Subsequently, the features at each scale are uniformly mapped to the same spatial coordinate system. This spatial coordinate system is based on the three-dimensional spatial coordinate description established by the spatial distribution information obtained in S1, so that the features at each scale are aligned in spatial position. At the same spatial position, there can be regional structural features, mesoscale structural features and microscale texture features at that position.

[0071] Based on spatial alignment, a weighted fusion process is performed based on structural consistency constraints between features at different scales. Structural consistency constraints mean that features at different scales should maintain logical consistency in structural information. For example, within a structural zone determined by regional structural features, the fracture orientation reflected by mesoscale structural features should be consistent with the overall structural features of the zone, and the degree of material heterogeneity reflected by microtextural features should also match the lithological type of the zone.

[0072] During fusion, features at each scale are weighted separately. The fusion weight of a particular scale feature is determined based on the degree of structural consistency between that scale feature and other scale features. When the structural information reflected by a particular scale feature is consistent with the structural information reflected by other scale features, that scale feature receives a higher fusion weight. When a particular scale feature contradicts other scale features in terms of structural information, the fusion weight of that scale feature is reduced to ensure the structural rationality of the fusion result.

[0073] For example, within a certain local region, if the three scale features are highly consistent in structural information, the fusion weights of the regional structural features, mesoscale structural features, and microtexture features can be set to 0.3, 0.4, and 0.3, respectively. If the microtexture features in this region contradict the structural information of the other two scale features, the fusion weight of the microtexture features should be reduced, for example, by adjusting the fusion weights of the three to 0.35, 0.5, and 0.15, so that the fusion result relies more on the regional structural features and mesoscale structural features with consistent structural information.

[0074] After the above weighted fusion process, a multi-scale visual feature representation is generated. Each spatial location is associated with a multi-scale visual feature obtained by weighted fusion of the regional-level structural features, mesoscale structural features, and microscale texture features of that location.

[0075] As an optional implementation, constructing a spatial semantic field based on the multi-scale visual feature representation includes:

[0076] S31. Based on the multi-scale visual feature expression, perform structural correlation analysis on the working face area, and aggregate feature points with similar structural attributes and spatial continuity to form multiple semantic interpretation units. Each semantic interpretation unit represents a local rock mass structure area.

[0077] S32. For each semantic interpretation unit, determine the overall structural distribution type and stability benchmark based on regional structural features, extract the crack orientation and structural extension trend based on mesoscale features, evaluate the degree of local material heterogeneity based on microtexture features, and construct the constraint relationship between cross-scale features.

[0078] S33. Based on the constraint relationship between cross-scale features, perform machinability semantic inference on each semantic interpretation unit to generate cutting difficulty index, structural stability index and cutting direction tendency index respectively.

[0079] S34. Taking each semantic interpretation unit as the initial starting point, the cutting difficulty index, structural stability index and cutting direction tendency index are diffused and propagated to adjacent regions in space. The propagation range is constrained according to structural continuity and feature similarity to construct a semantic distribution with spatial continuity and form a spatial semantic field.

[0080] Structural association analysis compares the similarity and spatial continuity of feature points in multi-scale visual feature representations, grouping feature points that are similar in structural attributes and spatially continuous into the same group to form a semantic interpretation unit. The similarity judgment of structural attributes comprehensively considers the regional-level structural features, mesoscale structural features, and micro-texture features of each feature point.

[0081] For two adjacent feature points, feature similarity is calculated based on the multi-scale visual features obtained through weighted fusion. When the feature similarity meets a preset similarity threshold, the two points are determined to have similar structural attributes and are classified into the same semantic interpretation unit. Spatial continuity judgment requires that feature points classified into the same semantic interpretation unit be spatially connected and without spatial discontinuities. The similarity threshold is set according to the actual geological conditions of the working face rock mass and the cutting accuracy requirements.

[0082] Through the above processing, the working face area is divided into multiple semantic interpretation units. Each semantic interpretation unit represents a local rock mass structure area, which is relatively uniform in structural properties and continuous in space, providing an analysis unit for semantic inference of machinability.

[0083] For each semantic interpretation unit, there are several feature points that are similar in structural properties and spatially connected. Each feature point is associated with regional structural features, mesoscale structural features, and microscale texture features. Statistical analysis is performed on the regional structural features of all feature points within the unit, including the distribution of depth continuity feature values ​​and morphological undulation feature values ​​at the location of each feature point.

[0084] The overall level of depth continuity characteristic values ​​indicates the degree of stratification integrity of the rock mass in the region. A high overall depth continuity characteristic value indicates that the rock strata in the region are intact and continuous, while significant fluctuations or local interruptions in the depth continuity characteristic value indicate the presence of interlayer transitions or structural disturbances. The overall amplitude and variation pattern of morphological undulation characteristic values ​​indicate the attitude characteristics of the rock strata in the region. Small undulation amplitude and gentle changes indicate that the rock strata in the region are nearly horizontal, while large undulation amplitude and regular changes indicate that the rock strata are inclined or folded.

[0085] Based on the analysis of depth continuity and morphological undulation, the region is categorized into corresponding overall structural distribution types, such as intact and dense rock strata, interlayer transition zones, fault-affected zones, or folded deformation zones. After determining the overall structural distribution type, the stability benchmark for the region is determined according to the general mechanical characteristics of the rock mass structure corresponding to that type. For example, the stability benchmark for intact and dense rock strata is relatively high, the stability benchmark for interlayer transition zones is intermediate, and the stability benchmark for fault-affected zones is relatively low.

[0086] Based on the mesoscale structural features of each feature point within the unit, fracture orientation and structural extension trends are extracted. Specifically, the mesoscale structural features of all feature points within the unit are summarized and analyzed to extract the fracture orientation and joint extension directions reflected by each feature point. The fracture orientation directions of all feature points within the unit are statistically analyzed to determine the most frequently occurring orientation as the dominant fracture orientation of the region; the joint extension directions are statistically analyzed to determine the dominant extension direction and its spatial extent, thus forming the structural extension trend of the region.

[0087] The degree of local material heterogeneity is assessed based on the microtexture characteristics of each feature point within the unit. Specifically, statistical analysis is performed on the microtexture characteristics of all feature points within the unit, calculating the mean and dispersion of the surface roughness value and material heterogeneity value for each feature point. A higher mean and lower dispersion indicate a higher overall degree of material heterogeneity and a more uniform distribution in the region; a lower mean and lower dispersion indicate better material homogeneity in the region; a higher dispersion indicates significant spatial fluctuations in material properties within the region, with potential significant differences in local hardness and cutting resistance at different locations.

[0088] After interpreting each semantic interpretation unit based on three scale features, constraints between cross-scale features are constructed. Specifically, the overall structural distribution type determined by regional structural features constrains the degree of fracture development reflected by mesoscale structural features. For example, in regions with an overall structural distribution type of intact and dense rock layers, the degree of fracture development is usually low. If the degree of fracture development reflected by the mesoscale structural features in this region is abnormally high, this anomaly may originate from errors in local feature extraction and needs to be marked and corrected in the constraints. The fracture density and structural plane intersection relationship reflected by mesoscale structural features constrain the degree of material heterogeneity reflected by microtexture features. In regions with dense fracture development and multiple sets of structural plane intersections, the degree of material heterogeneity is usually high. If the degree of heterogeneity reflected by microtexture features does not match the fracture development status, it also needs to be marked in the constraints.

[0089] After constructing cross-scale constraint relationships, machinability semantic inference is performed on each semantic interpretation unit based on the constraint relationships between cross-scale features, generating cutting difficulty index, structural stability index and cutting direction tendency index respectively.

[0090] For inferring the cutting difficulty index, the basic cutting difficulty level of the semantic interpretation unit is determined based on the overall structural distribution type identified by regional structural features. For example, the basic cutting difficulty level is higher in areas of intact and dense rock strata, and relatively lower in areas affected by faults. Based on this basic cutting difficulty level, the level is further adjusted according to the degree of fracture development reflected by mesoscale structural features. Higher fracture development indicates lower rock mass structural integrity, making it easier for the rock to break along fracture surfaces during cutting, thus reducing the cutting difficulty and lowering the basic cutting difficulty level. Further adjustments are made based on the degree of material heterogeneity reflected by microscopic texture features. Areas with high material heterogeneity exhibit drastic changes in local rock mass mechanical properties, resulting in greater resistance fluctuations during cutting, and the cutting difficulty is adjusted accordingly. All these corrections and adjustments are conducted under the constraints of cross-scale relationships. When the value of a certain scale feature is inconsistent with the range expected by the cross-scale constraints, the correction range of that scale feature for cutting difficulty is limited.

[0091] For inferring structural stability indices, the stability benchmark determined by regional-level structural features is used as the initial value. Based on the stability benchmark, the initial value is adjusted according to the fracture density and structural plane intersection relationships reflected by mesoscale structural features. Higher fracture density indicates a more severe degree of rock mass truncation by structural planes, resulting in lower structural stability; therefore, the initial stability value is adjusted downwards. Regions with dense intersections of multiple structural planes are prone to forming unstable wedge-shaped blocks or loose blocks, further reducing structural stability. Auxiliary adjustments are made based on the degree of material heterogeneity reflected by microtextural features. Regions with high material heterogeneity have lower local resistance to fracturing, leading to a corresponding downward adjustment of the structural stability value.

[0092] For inferring the cutting direction tendency index, the dominant fracture orientation and structural extension trend extracted from mesoscale structural features are used as the main basis for determining the appropriate cutting direction. Cutting along the fracture orientation and structural extension direction can utilize the structural weaknesses of the rock mass to obtain less cutting resistance. Based on this, the overall trend of the rock strata reflected by regional structural features is used as a directional constraint. When the dominant fracture orientation is basically consistent with the overall trend of the rock strata, the cutting direction tendency index is directly taken from the dominant fracture orientation. When there is a deviation between the dominant fracture orientation and the overall trend of the rock strata, the cutting direction is adjusted under the constraints of cross-scale relationships, so that the cutting direction tendency index reflects the guiding role of local fracture orientation without deviating too much from the overall trend of the rock strata.

[0093] Each semantic interpretation unit is formed by aggregating feature points. There may be spatial gaps between semantic interpretation units, meaning that some spatial locations have not been assigned to any semantic interpretation unit and have not yet obtained corresponding semantic index values. Therefore, it is necessary to use each semantic interpretation unit as an initial starting point and spread the index values ​​to the surrounding spatial locations.

[0094] Specifically, using the spatial location of each semantic interpretation unit and its corresponding index value as the initial information source, the index value is propagated gradually from the boundary position of each semantic interpretation unit to adjacent spatial locations. For a spatial location that has not yet obtained an index value, when there are semantic interpretation units with index values ​​or locations that have completed propagation around it, the existing index values ​​are weighted and interpolated based on the distance relationship and feature similarity between this location and the surrounding locations with existing index values ​​to obtain the index value for this location. The closer the existing index value is, the higher the interpolation weight is given, and the more similar the existing index values ​​are, the higher the interpolation weight is given.

[0095] During the diffusion and propagation process, the propagation range is constrained based on structural continuity and feature similarity. Structural continuity constraint means that the propagation of index values ​​proceeds along the direction of rock mass structural continuity. Specifically, the multi-scale visual feature expression between adjacent locations is checked to ensure continuity along the propagation path. When significant changes occur in the regional structural features between adjacent locations along the propagation path, such as crossing different structural partition boundaries, it is determined that the structure is discontinuous in that direction, and propagation is inhibited or terminated in that direction to avoid erroneously propagating the index values ​​of one structural region to adjacent regions with different structural properties.

[0096] Feature similarity constraint means that during the propagation process, the higher the similarity between the location of the received index value and the location of the propagation source in terms of multi-scale visual features, the more complete the propagation. When the feature difference between the location of the received index value and the location of the propagation source exceeds a preset propagation threshold, the propagation in that direction is suppressed, and the propagated index value is attenuated or the propagation is terminated.

[0097] After diffusion and propagation processing, each spatial location in the entire working face area obtains corresponding cutting difficulty indicators, structural stability indicators, and cutting direction tendency indicators, forming a spatial semantic field covering the entire working face area.

[0098] As an optional implementation, the trajectory is iteratively extended based on the comprehensive cutting trend evaluation value to establish a continuously evolving cutting trajectory, which also includes:

[0099] S61. During the trajectory expansion process, the spatial gradient distribution characteristics in the neighborhood of the current cutting front are subjected to directional consistency screening, and candidate expansion points with deviations from the local cutting main direction exceeding a preset threshold are eliminated.

[0100] S62. Construct joint constraint weights for the remaining candidate extension points based on the cutting difficulty index and the structural stability index, and perform comprehensive cutting trend evaluation value correction and sorting according to the joint constraint weights.

[0101] S63. Select the corrected optimal candidate expansion point as the next trajectory expansion node to perform iterative expansion.

[0102] The current cutting front refers to the foremost position to which the cutting trajectory has currently extended. Within the spatial neighborhood of this position, several candidate extension points exist, each corresponding to a spatial gradient distribution characteristic at that location. For each candidate extension point, the angle between the cutting difficulty gradient vector direction and the local main cutting direction, as well as the angle between the structural stability gradient vector direction and the local main cutting direction, are calculated. When either angle exceeds a preset directional deviation threshold, it indicates a significant deviation between the gradient change direction at the candidate extension point and the suitable cutting direction indicated by the rock mass structure. Advancing along this direction will deviate from the guiding direction of the weak surface of the rock mass structure, resulting in poor cutting performance. Therefore, this candidate extension point is removed from the candidate set. The directional deviation threshold is determined based on the tolerance for trajectory directional deviation in the cutting operation. A smaller threshold requires stricter directional consistency, resulting in a more concentrated directional orientation of the retained candidate extension points. A larger threshold allows for a wider range of directional deviations, leading to a larger number of retained candidate extension points. Through directional consistency screening, candidate extension points with excessive directional deviations are excluded, ensuring that the subsequent selected trajectory advancement direction does not deviate excessively from the local main cutting direction.

[0103] After completing the directional consistency screening, all remaining candidate expansion points meet the directional consistency requirement. However, the cutting difficulty and structural stability conditions at the locations of each candidate expansion point still differ, requiring further comprehensive consideration of these two factors to determine the optimal expansion direction. The joint constraint weight is a weight value calculated based on the cutting difficulty index value and structural stability index value at the location of each candidate expansion point, used to correct the comprehensive cutting trend evaluation value of each candidate expansion point.

[0104] Specifically, for each remaining candidate expansion point, the cutting difficulty index and structural stability index value of that point are obtained. The cutting difficulty index values ​​of all remaining candidate expansion points are normalized to obtain normalized values ​​ranging from 0 to 1; a higher cutting difficulty index value results in a larger normalized value. Similarly, the structural stability index values ​​of all remaining candidate expansion points are normalized to obtain normalized values ​​ranging from 0 to 1; a higher structural stability index value results in a larger normalized value. The normalized cutting difficulty value is then inverted to obtain (1 - normalized cutting difficulty value). This value reflects the favorable cutting resistance of the candidate expansion point; a larger value indicates lower cutting resistance and greater ease of advancement.

[0105] The joint constraint weights are calculated using a basic weight plus an adjustment term: Joint Constraint Weight = Basic Weight + Adjustment Coefficient × [Cutting Difficulty Weight Coefficient × (1 - Normalized Cutting Difficulty Value) + Structural Stability Weight Coefficient × Normalized Structural Stability Value]. The basic weight ensures that all candidate extension points selected through directional screening have a minimum weight for ranking, preventing complete rejection due to extreme cutting conditions. The adjustment coefficient controls the impact of differences in cutting conditions on the weights. The cutting difficulty weight coefficient and structural stability weight coefficient are set according to the specific needs of the actual cutting operation. When the cutting operation prioritizes reducing cutting resistance, the cutting difficulty weight coefficient is set larger; when the cutting operation prioritizes ensuring the safety of the surrounding rock, the structural stability weight coefficient is set larger. By adjusting the joint constraint weights, the ranking results not only reflect the gradient change trend and directional matching degree of each candidate direction but also take into account the actual cutting difficulty and surrounding rock stability conditions of each candidate direction, selecting the optimal trajectory advancement direction while comprehensively considering both cutting resistance and surrounding rock safety.

[0106] After calculating the joint constraint weights of each candidate expansion point, the original comprehensive cutting trend evaluation value of each candidate expansion point is multiplied by the joint constraint weight of that point to obtain the corrected comprehensive cutting trend evaluation value. The candidate expansion points are then ranked from best to worst according to the corrected comprehensive cutting trend evaluation values.

[0107] After completing the corrected sorting, the candidate extension point with the highest corrected comprehensive cutting trend evaluation value is selected as the next trajectory extension node. This node serves as the new current cutting front position, and candidate extension points within its spatial neighborhood are redefined. The complete process of directional consistency screening, joint constraint weight construction, corrected sorting, and optimal point selection is executed again to determine the next trajectory extension node. This process is repeated iteratively, and the cutting trajectory gradually extends forward from the starting position. Each selected extension node undergoes directional consistency screening and comprehensive optimization of cutting conditions, ensuring that the final generated cutting trajectory maintains overall consistency with the main local cutting direction. Simultaneously, it achieves progressive optimization in terms of both cutting resistance and surrounding rock stability, forming a continuously evolving cutting trajectory.

[0108] As an optional implementation, establishing a continuously evolving cut trajectory also includes:

[0109] S64. During the slicing process, the spatial semantic field is dynamically corrected based on continuously updated machine vision perception data, and the corresponding spatial gradient distribution features and candidate direction evaluation results are recalculated to perform real-time adaptive adjustment of the slicing trajectory.

[0110] During the cutting process, machine vision perception data acquisition continues, constantly acquiring the latest image data and depth information of the working face area. When the newly acquired machine vision perception data changes compared to the data previously used to construct the spatial semantic field, it indicates that the state of the rock mass at the working face has been updated, and the spatial semantic field needs to be dynamically corrected.

[0111] Specifically, based on newly acquired machine vision perception data, the surface texture, structural morphology, and spatial distribution information of the changed area are re-extracted. Multi-scale feature analysis is then re-executed on this area to update its regional-level structural features, mesoscale structural features, and micro-texture features. Based on the updated multi-scale visual feature representation, the spatial semantic field of this area is locally corrected, and the cutting difficulty index, structural stability index, and cutting direction tendency index at each location in the area are re-inferred. In the corrected spatial semantic field, the index values ​​of the changed areas are updated to new values ​​corresponding to the current actual state of the working surface, while the index values ​​of the unchanged areas remain unchanged.

[0112] After the spatial semantic field is dynamically corrected, the spatial gradient distribution characteristics of the corrected region are recalculated. The cutting difficulty gradient vector and structural stability gradient vector at each location within the region are also recalculated, as well as the gradient transition relationship between the region and its adjacent regions. Based on the updated spatial gradient distribution characteristics, the local main cutting direction of the corrected region is re-extracted, and the projection processing of the spatial gradient distribution characteristics along the local main cutting direction is re-performed to generate an updated comprehensive cutting trend evaluation value.

[0113] Based on the updated comprehensive cutting trend evaluation value, the evaluation results of candidate directions at the current cutting front position are recalculated, including re-performing direction consistency screening, joint constraint weight construction, and corrected sorting, to redetermine the optimal trajectory advancement direction. If the recalculated optimal advancement direction is consistent with the previously planned trajectory direction, the cutting trajectory continues to execute according to the original path; if the recalculated optimal advancement direction deviates from the previously planned trajectory direction, starting from the current cutting position, the trajectory is iteratively expanded again according to the updated comprehensive cutting trend evaluation value to generate an adjusted cutting trajectory, enabling the cutting trajectory to respond promptly to changes in the working face rock conditions.

[0114] As an optional implementation, the local main cutting direction is extracted based on the cutting direction tendency index, including:

[0115] S54. Statistically analyze the directional distribution of cutting direction tendency index in the target space neighborhood and construct the directional frequency histogram feature.

[0116] S55. Identify the dominant direction interval based on the histogram feature of the direction frequency, and perform stability screening on the dominant direction interval in combination with the structural stability index to establish the constrained local cutting main direction.

[0117] Within the target spatial neighborhood, the directional distribution of cutting direction tendency indices is statistically analyzed, and a histogram of directional frequencies is constructed. The target spatial neighborhood refers to all spatial locations within a certain spatial range, centered on the location where the local main cutting direction needs to be determined. Each spatial location within this neighborhood is associated with a corresponding cutting direction tendency index in the spatial semantic field. The cutting direction tendency index describes, in vector form, the suitable cutting direction indicated by the rock mass structure at that location.

[0118] The directional range in three-dimensional space is divided into several directional intervals according to preset angular intervals, such as 10 degrees or 15 degrees per interval, thus dividing the complete directional space into multiple non-overlapping directional intervals. The cutting direction tendency indexes at all locations within the target space neighborhood are individually assigned to their respective directional intervals, and the number of cutting direction tendency indexes falling into each directional interval is counted. The number of indexes within each directional interval constitutes the directional frequency histogram feature, reflecting the degree of concentration of the cutting direction tendency indexes in each direction within the target space neighborhood.

[0119] The identification of dominant direction intervals is based on the number of indicators in each direction interval from the histogram of direction frequencies. The direction interval with the highest number of indicators, or several direction intervals with an indicator number exceeding a preset frequency threshold, are selected as candidate dominant direction intervals. When the number of indicators in one direction interval is significantly higher than other direction intervals, that direction interval is the only candidate dominant direction interval; when multiple direction intervals have high and similar indicator numbers, all of these direction intervals are considered candidate dominant direction intervals. The frequency threshold is determined based on the concentration of the cutting direction tendency indicator distribution within the neighborhood. Higher frequency thresholds can be set for regions with more concentrated direction distribution to filter out the most concentrated directions, while lower frequency thresholds are appropriately used for regions with more dispersed direction distribution to retain sufficient candidate directions.

[0120] After identifying candidate dominant direction intervals, stability screening is performed on these intervals using structural stability indices. Specifically, for each candidate dominant direction interval, the structural stability index values ​​corresponding to each location within that interval are obtained, and the average value of the structural stability index values ​​within that interval is calculated. If the average structural stability index value within a candidate dominant direction interval is lower than a preset stability threshold, that candidate dominant direction interval is eliminated. The stability threshold is set based on the safety requirements of the cutting operation for the surrounding rock; the higher the safety requirements, the higher the threshold is set, and the more stringent the screening.

[0121] After stability screening, if only one candidate dominant direction interval remains, the center direction of that interval is selected as the local cutting main direction; if multiple candidate dominant direction intervals pass the stability screening, the center direction of the interval with the most indicators is selected as the local cutting main direction.

[0122] As an optional implementation method, such as Figure 3As shown, a multi-view imaging unit and a depth sensing unit are deployed on the tunneling equipment. The multi-view imaging unit is used to acquire a continuous image sequence of the tunneling working face area from different observation angles. The depth sensing unit is used to synchronously acquire depth information corresponding to the image data space. After performing time synchronization and spatial registration on the continuous image sequence and depth information, machine vision perception data is established.

[0123] A multi-view imaging unit refers to multiple imaging devices installed on a tunneling machine. Each imaging device is positioned at a different location on the machine, acquiring images of the working face area from various observation angles. By deploying multiple imaging devices at different locations on the tunneling machine, each device acquires images of the working face area from its own observation angle. The images from multiple perspectives complement each other, covering all areas of the working face and eliminating blind spots and occlusion problems associated with single-view observations. Each imaging device continuously acquires images during the tunneling machine's operation, forming a continuous image sequence. Each frame in the continuous image sequence records the working face image information acquired by that imaging device from its observation angle at a specific moment.

[0124] A depth sensing unit is a sensing device installed on tunneling equipment to acquire spatial distance information of different areas of the working face. Examples include structured light sensors, time-of-flight sensors, or lidar. The depth sensing unit works synchronously with the multi-view imaging unit. While the multi-view imaging unit acquires images, the depth sensing unit simultaneously acquires the spatial depth information corresponding to the image data.

[0125] After the multi-view imaging unit and the depth sensing unit acquire continuous image sequences and depth information respectively, time synchronization and spatial registration are performed on them to establish machine vision perception data. Time synchronization refers to aligning the continuous image sequences acquired by the multi-view imaging unit and the depth information acquired by the depth sensing unit on the time axis, so that the image frames acquired by each imaging device at the same time correspond to the depth information acquired by the depth sensing unit. Time synchronization eliminates time deviations between acquisition devices through unified timestamp calibration, ensuring that the image data and depth information at the same time are strictly corresponding in time.

[0126] Spatial registration refers to the unified alignment of image data acquired by imaging devices from multiple perspectives and depth information obtained by depth sensing units in spatial coordinates. Spatial registration uses the spatial positional relationships and coordinate transformation parameters between the devices to uniformly transform image data and depth information from different perspectives into the same spatial coordinate system, enabling data from different devices to correspond and correlate spatially. The coordinate transformation parameters required for spatial registration can be obtained by pre-calibrating each device. The calibration process determines the relative positional relationships and coordinate transformation matrices between each imaging device and depth sensing unit.

[0127] After time synchronization and spatial registration, continuous image sequences and depth information from multiple perspectives are integrated into a dataset that is uniformly aligned in time and space, establishing machine vision perception data. This provides a spatiotemporally processed data foundation for the subsequent extraction of surface texture information, structural morphology information, and spatial distribution information from the working surface.

[0128] As an optional implementation, after time synchronization and spatial registration, abnormal frame filtering is performed based on image sharpness and depth continuity, and continuity compensation processing is performed to establish machine vision perception data.

[0129] Abnormal frame filtering is based on image sharpness and depth continuity. For image sharpness, a sharpness evaluation value is calculated for each frame in the continuous image sequence. The sharpness evaluation value can be measured by indicators such as the gradient intensity of edge regions and the energy proportion of high-frequency components in the image. The sharpness evaluation value of each frame is compared with a preset sharpness threshold. When the sharpness evaluation value of a frame is lower than the sharpness threshold, the frame is determined to be an image sharpness abnormal frame.

[0130] For depth continuity, the depth information of each frame is checked to see if the changes in depth measurement values ​​between adjacent spatial locations remain within a reasonable range. When there is a sharp jump in the depth measurement values ​​between adjacent locations that exceeds the normal fluctuation range or when there is a region with missing depth values, it is determined that there is a depth continuity anomaly in that frame.

[0131] When a frame is simultaneously or separately identified as an image sharpness abnormality frame or a depth continuity abnormality frame, the frame is removed from the dataset and not included in subsequent machine vision perception data.

[0132] After the abnormal frames are removed, a continuity compensation process is performed to compensate for and fill the data gaps caused by the removed frames. For each time corresponding to a removed frame, the image data and depth information of the normal frames before and after that time are acquired, and interpolation estimation is performed on the image data and depth information of the removed time based on the data of the adjacent normal frames.

[0133] At the image data level, based on the image content and pixel change trends of adjacent frames, an estimated image for that moment is generated through temporal interpolation, so that the estimated image maintains a smooth transition with the preceding and following frames in terms of visual content.

[0134] At the depth information level, based on the depth measurements and depth change trends of adjacent frames, estimated depth information for that moment is generated through temporal interpolation, so that the estimated depth values ​​remain continuous with the spatial distribution of the preceding and following frames.

[0135] When the number of frames removed is small, the data difference between adjacent normal frames is small, and the interpolation estimation accuracy is high. When the number of frames removed is large, the data difference between adjacent normal frames may be large, and the interpolation estimation accuracy decreases. In this case, quadratic interpolation or cubic spline interpolation can be used, or weighted interpolation can be performed by combining data from multiple adjacent frames to improve the estimation accuracy.

[0136] After abnormal frame filtering and continuity compensation processing, the image data and depth information of each frame in the machine vision perception data have a reliable quality level, while maintaining continuity in the time series. This eliminates the adverse effects of abnormal data caused by interference from the acquisition environment on subsequent processing, providing a high-quality data foundation for the future.

[0137] Example 2, as Figure 4 As shown, based on the same inventive concept as the machine vision-based intelligent planning method for tunneling equipment cutting trajectories provided in Embodiment 1, this embodiment of the invention also provides a machine vision-based intelligent planning system for tunneling equipment cutting trajectories, including:

[0138] The visual perception acquisition module 11 is used to acquire continuous machine vision perception data of the tunneling face area, including surface texture information, structural morphology information and spatial distribution information of the face.

[0139] The multi-scale feature parsing module 12 is used to perform multi-scale feature parsing processing based on the machine vision perception data, extract regional structural features, mesoscale structural features and micro-texture features respectively, and perform multi-scale feature space alignment and fusion to establish multi-scale visual feature representation.

[0140] The spatial semantic field construction module 13 is used to construct a spatial semantic field based on the multi-scale visual feature expression. Each position in the spatial semantic field is associated with a cutting difficulty index, a structural stability index, and a cutting direction tendency index.

[0141] The gradient distribution construction module 14 is used to calculate the rate of change of indicators of adjacent spatial positions using cutting difficulty indicators and structural stability indicators, and to construct spatial gradient distribution features.

[0142] The cutting trend evaluation module 15 is used to extract the local main cutting direction based on the cutting direction trend index, and project the spatial gradient distribution characteristics along the local main cutting direction to generate a comprehensive cutting trend evaluation value.

[0143] The trajectory iteration planning module 16 is used to expand the trajectory iteratively based on the comprehensive cutting trend evaluation value, starting from the current cutting position, and to establish a continuously evolving cutting trajectory.

[0144] Furthermore, the cutting trend evaluation module 15 is also used for:

[0145] Based on the spatial gradient distribution characteristics, the continuity of gradient change along the local cutting main direction is identified, and the gradient abrupt change position is detected. The current spatial region is divided into cutting sub-regions using the gradient abrupt change position as the boundary.

[0146] The projection trends of cutting difficulty index and structural stability index on the main local cutting direction are calculated for each cutting sub-region, and the corresponding local cutting mode is determined based on the consistency of gradient changes within each cutting sub-region.

[0147] Based on the local cutting mode corresponding to each cutting sub-region, the cutting difficulty change trend and the structural stability change trend are subjected to segmented constraint fusion processing to generate a comprehensive cutting trend evaluation value.

[0148] Furthermore, the multi-scale feature parsing module 12 is also used for:

[0149] Based on the machine vision perception data, the depth information distribution of the working surface area is constructed, and a three-dimensional surface morphology expression corresponding to the image data space is generated.

[0150] Based on the aforementioned three-dimensional surface morphology representation, a multi-scale structural interpretation process is performed on the working surface region, wherein:

[0151] At the regional scale, the overall structural partitioning of the working face is identified based on depth continuity and morphological undulation features, and regional structural features are extracted.

[0152] At the mesoscale, information on fracture orientation and structural extension is extracted based on the depth gradient direction distribution and local geometric change features to form mesoscale structural features.

[0153] At the microscale, surface roughness and local heterogeneous features are extracted based on the coupling relationship between image texture intensity distribution and depth perturbation changes, forming microtexture features;

[0154] Features at each scale are mapped to a unified spatial coordinate system, and weighted fusion processing is performed based on structural consistency constraints between features at different scales to generate multi-scale visual feature representations.

[0155] Furthermore, the spatial semantic field construction module 13 is also used for:

[0156] Based on the multi-scale visual feature expression, structural correlation analysis is performed on the working face area, and feature points with similar structural attributes and spatial continuity are aggregated to form multiple semantic interpretation units. Each semantic interpretation unit represents a local rock mass structure area.

[0157] For each semantic interpretation unit, the overall structural distribution type and stability benchmark are determined based on regional structural features, the crack orientation and structural extension trend are extracted based on mesoscale features, the degree of local material heterogeneity is evaluated based on microtexture features, and the constraint relationship between cross-scale features is constructed.

[0158] Based on the constraint relationship between cross-scale features, machinability semantic inference is performed on each semantic interpretation unit to generate cutting difficulty index, structural stability index and cutting direction tendency index respectively.

[0159] Starting from each semantic interpretation unit, the cutting difficulty index, structural stability index, and cutting direction tendency index are diffused and propagated to adjacent regions in space. The propagation range is constrained according to structural continuity and feature similarity to construct a spatially continuous semantic distribution and form a spatial semantic field.

[0160] Furthermore, the trajectory iterative planning module 16 is also used for:

[0161] During the trajectory expansion process, the spatial gradient distribution characteristics in the neighborhood of the current cutting front are subjected to directional consistency screening, and candidate expansion points whose deviation from the local cutting main direction exceeds a preset threshold are eliminated.

[0162] The remaining candidate extension points are constructed with joint constraint weights based on the cutting difficulty index and the structural stability index, and then sorted according to the comprehensive cutting trend evaluation value based on the joint constraint weights.

[0163] The corrected optimal candidate expansion point is selected as the next trajectory expansion node to perform iterative expansion.

[0164] Furthermore, the trajectory iterative planning module 16 is also used for:

[0165] During the truncation process, the spatial semantic field is dynamically corrected based on continuously updated machine vision perception data, and the corresponding spatial gradient distribution features and candidate direction evaluation results are recalculated to perform real-time adaptive adjustment of the truncation trajectory.

[0166] Furthermore, the cutting trend evaluation module 15 is also used for:

[0167] Within the target space neighborhood, the directional distribution of the cutting direction tendency index is statistically analyzed, and a directional frequency histogram feature is constructed.

[0168] Based on the histogram feature of directional frequency, the dominant direction interval is identified, and the stability of the dominant direction interval is screened by combining the structural stability index to establish the constrained local cutting main direction.

[0169] Furthermore, a multi-view imaging unit and a depth sensing unit are deployed on the tunneling equipment. The multi-view imaging unit is used to acquire a continuous image sequence of the tunneling working face area from different observation angles. The depth sensing unit is used to synchronously acquire depth information corresponding to the image data space. After performing time synchronization and spatial registration on the continuous image sequence and depth information, machine vision perception data is established.

[0170] Furthermore, after time synchronization and spatial registration, abnormal frames are filtered out based on image sharpness and depth continuity, and continuity compensation processing is performed to establish machine vision perception data.

[0171] Although preferred embodiments of the invention have been described, those skilled in the art, once they have learned the basic inventive concept, can make other changes and modifications to these embodiments.

[0172] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A machine vision-based intelligent planning method for cutting trajectories of tunneling equipment, characterized in that, The method includes: Acquire continuous machine vision perception data of the tunneling face area, wherein the machine vision perception data includes surface texture information, structural morphology information and spatial distribution information of the face. Based on the machine vision perception data, multi-scale feature parsing processing is performed to extract regional-level structural features, mesoscale structural features, and micro-texture features, respectively. Multi-scale feature space alignment and fusion are then performed to establish a multi-scale visual feature representation. A spatial semantic field is constructed based on the multi-scale visual feature representation, and each position in the spatial semantic field is associated with a cutting difficulty index, a structural stability index, and a cutting direction tendency index. By using cutting difficulty index and structural stability index, the rate of change of index of adjacent spatial positions is calculated, and spatial gradient distribution characteristics are constructed. Based on the cutting direction trend index, the local main cutting direction is extracted, and the spatial gradient distribution characteristics are projected along the local main cutting direction to generate a comprehensive cutting trend evaluation value. Starting from the current cutting position, the trajectory is iteratively expanded based on the comprehensive cutting trend evaluation value to establish a continuously evolving cutting trajectory.

2. The intelligent planning method for cutting trajectory of tunneling equipment based on machine vision as described in claim 1, characterized in that, The spatial gradient distribution features are projected along the local cutting principal direction to generate a comprehensive cutting trend evaluation value, including: Based on the spatial gradient distribution characteristics, the continuity of gradient change along the local cutting main direction is identified, and the gradient abrupt change position is detected. The current spatial region is divided into cutting sub-regions using the gradient abrupt change position as the boundary. The projection trends of cutting difficulty index and structural stability index on the main local cutting direction are calculated for each cutting sub-region, and the corresponding local cutting mode is determined based on the consistency of gradient changes within each cutting sub-region. Based on the local cutting mode corresponding to each cutting sub-region, the cutting difficulty change trend and the structural stability change trend are subjected to segmented constraint fusion processing to generate a comprehensive cutting trend evaluation value.

3. The intelligent planning method for cutting trajectory of tunneling equipment based on machine vision as described in claim 1, characterized in that, Multi-scale feature parsing processing is performed based on the machine vision perception data, including: Based on the machine vision perception data, the depth information distribution of the working surface area is constructed, and a three-dimensional surface morphology expression corresponding to the image data space is generated. Based on the aforementioned three-dimensional surface morphology representation, a multi-scale structural interpretation process is performed on the working surface region, wherein: At the regional scale, the overall structural partitioning of the working face is identified based on depth continuity and morphological undulation features, and regional structural features are extracted. At the mesoscale, information on fracture orientation and structural extension is extracted based on the depth gradient direction distribution and local geometric change features to form mesoscale structural features. At the microscale, surface roughness and local heterogeneous features are extracted based on the coupling relationship between image texture intensity distribution and depth perturbation changes, forming microtexture features; Features at each scale are mapped to a unified spatial coordinate system, and weighted fusion processing is performed based on structural consistency constraints between features at different scales to generate multi-scale visual feature representations.

4. The intelligent planning method for cutting trajectory of tunneling equipment based on machine vision as described in claim 1, characterized in that, Constructing a spatial semantic field based on the multi-scale visual feature representation includes: Based on the multi-scale visual feature expression, structural correlation analysis is performed on the working face area, and feature points with similar structural attributes and spatial continuity are aggregated to form multiple semantic interpretation units. Each semantic interpretation unit represents a local rock mass structure area. For each semantic interpretation unit, the overall structural distribution type and stability benchmark are determined based on regional structural features, the crack orientation and structural extension trend are extracted based on mesoscale features, the degree of local material heterogeneity is evaluated based on microtexture features, and the constraint relationship between cross-scale features is constructed. Based on the constraint relationship between cross-scale features, machinability semantic inference is performed on each semantic interpretation unit to generate cutting difficulty index, structural stability index and cutting direction tendency index respectively. Starting from each semantic interpretation unit, the cutting difficulty index, structural stability index, and cutting direction tendency index are diffused and propagated to adjacent regions in space. The propagation range is constrained according to structural continuity and feature similarity to construct a spatially continuous semantic distribution and form a spatial semantic field.

5. The intelligent planning method for cutting trajectory of tunneling equipment based on machine vision as described in claim 1, characterized in that, Based on the comprehensive cutting trend evaluation value, the trajectory is iteratively extended to establish a continuously evolving cutting trajectory, which also includes: During the trajectory expansion process, the spatial gradient distribution characteristics in the neighborhood of the current cutting front are subjected to directional consistency screening, and candidate expansion points whose deviation from the local cutting main direction exceeds a preset threshold are eliminated. The remaining candidate extension points are constructed with joint constraint weights based on the cutting difficulty index and the structural stability index, and then sorted according to the comprehensive cutting trend evaluation value based on the joint constraint weights. The corrected optimal candidate expansion point is selected as the next trajectory expansion node to perform iterative expansion.

6. The intelligent planning method for cutting trajectory of tunneling equipment based on machine vision as described in claim 1, characterized in that, Establishing a continuously evolving cut-off trajectory also includes: During the truncation process, the spatial semantic field is dynamically corrected based on continuously updated machine vision perception data, and the corresponding spatial gradient distribution features and candidate direction evaluation results are recalculated to perform real-time adaptive adjustment of the truncation trajectory.

7. The intelligent planning method for cutting trajectory of tunneling equipment based on machine vision as described in claim 1, characterized in that, Based on the cutting direction tendency index, the local main cutting direction is extracted, including: Within the target space neighborhood, the directional distribution of the cutting direction tendency index is statistically analyzed, and a directional frequency histogram feature is constructed. Based on the histogram feature of directional frequency, the dominant direction interval is identified, and the stability of the dominant direction interval is screened by combining the structural stability index to establish the constrained local cutting main direction.

8. The intelligent planning method for cutting trajectory of tunneling equipment based on machine vision as described in claim 1, characterized in that, A multi-view imaging unit and a depth sensing unit are deployed on the tunneling equipment. The multi-view imaging unit is used to acquire a continuous image sequence of the tunneling working face area from different observation angles. The depth sensing unit is used to synchronously acquire depth information corresponding to the image data space. After performing time synchronization and spatial registration on the continuous image sequence and depth information, machine vision perception data is established.

9. The intelligent planning method for cutting trajectory of tunneling equipment based on machine vision as described in claim 8, characterized in that, After time synchronization and spatial registration, abnormal frames are filtered out based on image sharpness and depth continuity, and continuity compensation processing is performed to establish machine vision perception data.

10. A machine vision-based intelligent planning system for the cutting trajectory of tunneling equipment, characterized in that, The system is used to implement the machine vision-based intelligent planning method for cutting trajectories of tunneling equipment as described in any one of claims 1 to 9, the system comprising: The visual perception acquisition module is used to acquire continuous machine vision perception data of the tunneling face area. The machine vision perception data includes surface texture information, structural morphology information and spatial distribution information of the face. The multi-scale feature parsing module is used to perform multi-scale feature parsing processing based on the machine vision perception data, extracting regional-level structural features, mesoscale structural features and micro-texture features respectively, and performing multi-scale feature space alignment and fusion to establish a multi-scale visual feature representation. The spatial semantic field construction module is used to construct a spatial semantic field based on the multi-scale visual feature expression. Each position in the spatial semantic field is associated with a cutting difficulty index, a structural stability index, and a cutting direction tendency index. The gradient distribution construction module is used to calculate the rate of change of indicators of adjacent spatial positions using cutting difficulty indicators and structural stability indicators, and to construct spatial gradient distribution characteristics. The cutting trend evaluation module is used to extract the local main cutting direction based on the cutting direction trend index, and project the spatial gradient distribution characteristics along the local main cutting direction to generate a comprehensive cutting trend evaluation value. The trajectory iteration planning module is used to expand the trajectory iteratively based on the comprehensive cutting trend evaluation value, starting from the current cutting position, and to establish a continuously evolving cutting trajectory.

Images