Slope rock mass structure plane generalization identification method and system based on multi-level domain adaptation

Abstract

This invention discloses a generalized identification method and system for slope rock mass structural surfaces based on multi-level domain adaptation. This invention relates to the field of slope rock mass identification technology. By constructing four levels—global domain adaptation, regional domain adaptation, local domain adaptation, and sample-level adaptive adaptation—this invention achieves hierarchical and progressive evolution of model capabilities. A variational autoencoder is used to explicitly decouple features into domain-invariant and domain-specific parts, enabling the model to distinguish between the essential geometric features of rock mass structural surfaces and the weathering and unloading features of specific regions. Through regional meta-learning training, it can quickly adapt to new regions with only a small number of labeled samples.

Description

Technical Field

[0001] This invention relates to the field of slope rock mass identification technology, specifically to a method and system for generalized identification of slope rock mass structural surfaces based on multi-level domain adaptation. Background Technology

[0002] Currently, in the field of rock slope structural surface identification, automated methods integrating UAV remote sensing and computer vision have become an important development direction. A representative technical approach is to acquire high-precision, high-resolution 3D point cloud models of slopes using close-up UAV photogrammetry, and then utilize deep neural networks for feature learning and identification. For example, using improved ResNet-18 convolutional neural network architectures can effectively extract structural surface features from point cloud data. This type of method demonstrates significant potential in engineering applications: on the one hand, close-up UAV photogrammetry provides a rich and spatially continuous data foundation that is difficult to achieve with traditional methods; on the other hand, for large slopes with tens or even hundreds of millions of point clouds, the trained neural network model can quickly predict using a small number of new samples by adjusting hyperparameters. Compared to processing methods relying on manual interaction or traditional clustering algorithms (such as those used in some commercial software), this greatly improves identification efficiency and reduces computational costs when processing massive amounts of data, providing a feasible tool for intelligent engineering geological exploration.

[0003] However, the aforementioned existing technologies suffer from a fundamental limitation that severely restricts their universality and widespread application in practical engineering. This limitation stems from the complex and variable geological properties of rock mass structural surfaces: due to the vastly different geological histories, tectonic movements, and lithological combinations of different slopes, the dominant orientation (strike, dip), spacing, extensibility, number of groups, and spatial distribution patterns of their structural surfaces exhibit strong regional specificity and non-stationarity. This leads to a situation where a recognition model trained and optimized to perfection on a specific slope dataset often suffers a significant drop in recognition accuracy or even fails when directly applied to another slope with significantly different geological conditions; that is, the model lacks sufficient cross-regional and cross-scenario generalization ability. Although, in the long run, constructing a standardized structural surface sample library covering various rock mass conditions globally may be a solution to this problem, at present, it faces significant challenges due to multiple practical constraints such as data acquisition costs, uniformity of annotation standards, geological privacy, and intellectual property rights. Therefore, existing technologies are essentially still "customized" solutions for specific scenarios, making it difficult to form a universal and transferable intelligent recognition capability.

[0004] Based on the above analysis, the key drawback of existing technologies lies in the fact that the strong regional heterogeneity of geological conditions makes it difficult to widely cover training data. This leads to the overfitting of purely data-driven deep learning models to local features, lacking the constraints and guidance of geological principles, and ultimately resulting in a severe deficiency in cross-regional generalization ability. As a result, current methods remain at the rudimentary stage of customizing solutions for individual slopes, far from reaching the level of intelligent and universal engineering tools that can be "trained once and used in multiple locations." Summary of the Invention

[0005] Based on the problems raised in the background technology above, the purpose of this invention is to provide a generalized identification method and system for slope rock mass structure surfaces based on multi-level domain adaptation. By improving model learning methods, data utilization strategies or system architecture, the adaptability and inference ability of the model to the rock mass structure characteristics of unknown areas are enhanced, thus promoting the slope engineering geological survey towards a truly intelligent and universal direction.

[0006] This invention is achieved through the following technical solution:

[0007] The first aspect of this invention provides a generalized identification method for slope rock mass structural surfaces based on multi-level domain adaptation, comprising the following steps:

[0008] Step S1: Measure the outcrop rock mass to obtain slope point cloud data, and perform multi-level feature decomposition and representation learning on the slope point cloud data to obtain a multi-scale feature pyramid.

[0009] Step S2: Perform progressive domain adaptation training on the multi-scale feature pyramid to generate an optimized site model and a domain-aligned feature representation;

[0010] Step S3: Based on the multi-task learning framework with geological knowledge embedding, perform multi-task joint optimization on the domain-aligned feature representation to obtain the structural surface identification result, attitude prediction result, and joint group division result; wherein, the multi-task joint optimization includes: extracting basic features from the domain-aligned feature representation using a shared feature encoder, inputting the basic features into three task branches—structural surface segmentation head, attitude regression head, and structural surface group division head—for processing, and performing geological knowledge constraint verification and multi-task joint optimization on the processed data;

[0011] Step S4: Analyze the structural surface identification results, the attitude prediction results, and the joint group division results through uncertainty quantification and active learning mechanisms to generate slope rock mass identification results, and feed the slope rock mass identification results back to step S2 for data supplementation.

[0012] In the above technical solution, firstly, a multi-level feature decomposition and representation learning network is constructed to establish a feature extraction system with four levels: global geological general, regional semantic, site specific, and instance details. This system accurately captures structural surface features at different levels of abstraction. At the same time, a feature decoupling network is used to explicitly separate domain-invariant and domain-specific features. Combined with a multi-scale feature pyramid, features of different resolutions are fused. This fundamentally solves the core problem of existing models having one-sided feature extraction and difficulty in adapting to different geological scenarios, laying a high-quality feature foundation for subsequent domain adaptation training.

[0013] Then, a progressive domain adaptation training strategy is adopted, which optimizes the model's adaptability in four stages. The first stage is based on large-scale pre-training on a global multi-regional mixed dataset and learning common knowledge across geological regions by combining multi-task objective functions. The second stage trains the model to quickly adapt to the initial parameters of new regions through regional meta-learning. The third stage adopts adversarial domain adaptation combined with prototype alignment to achieve site adaptation and generate domain-indistinguishable but task-related features. This step effectively makes up for the shortcoming that the training data cannot cover a wide range of geological conditions, avoids the model overfitting to local features, and achieves accurate adaptation of the model from cross-regional to single-site.

[0014] Next, a multi-task learning framework is constructed based on embedded geological knowledge. Basic features are extracted through a shared feature encoder and then processed collaboratively by a structural surface segmentation head, an attitude regression head, and a structural surface group segmentation head. At the same time, geological constraints such as attitude consistency and topological rationality are introduced to solve the problems of existing models lacking geological principle guidance and prediction results not conforming to actual geological laws, thus balancing recognition accuracy and geological rationality.

[0015] Finally, the model's generalization ability is further improved through uncertainty quantification and active learning feedback optimization.

[0016] In one optional embodiment, the slope point cloud data undergoes multi-level feature decomposition and representation learning, including:

[0017] Rock mass structural features at different levels of abstraction were extracted from the slope point cloud data;

[0018] The extracted rock mass structural surface features are explicitly decomposed into domain-invariant features and domain-specific features using a feature decoupling network;

[0019] A multi-scale feature pyramid is used to hierarchically fuse domain-invariant features and domain-specific features, outputting a multi-scale pyramid; the hierarchical fusion includes:

[0020] The first level is used to learn the basic geometric morphology, topological relationship and mechanical genesis of the structural surfaces in domain-invariant features and domain-specific features, and to generate the overall structural features of the slope.

[0021] The second level is used to learn discriminative features from domain-invariant features and domain-specific features, and to generate construct domain features;

[0022] The third level is used to learn the lithology, weathering, and unloading characteristics at the slope scale in both domain-invariant and domain-specific features, and to generate specific slope segment features.

[0023] The fourth level is used to learn instance detail features from domain-invariant and domain-specific features to generate point cloud detail features.

[0024] In one optional embodiment, the feature decoupling network includes: a shared geometric coding module, an attitude constraint coding module, a topological graph neural network, a mechanical origin discrimination branch, and an adversarial decoupling dual-branch encoder;

[0025] Specifically, a feature decoupling network is used to explicitly decompose the extracted rock mass structural surface features into domain-invariant features and domain-specific features, including:

[0026] The shared geometric coding module is used to extract the basic geometric features from the rock mass structural surface features; wherein, the shared geometric coding module includes a first-level geometric abstraction layer, a second-level geometric abstraction layer and a third-level geometric abstraction layer, and the first-level geometric abstraction layer, the second-level geometric abstraction layer and the third-level geometric abstraction layer are respectively used to extract local features of different dimensions and sizes from the rock mass structural surface features;

[0027] The basic geometric features are input in parallel into the attitude constraint encoding module and the topology graph neural network to obtain attitude perception features and topology features;

[0028] The basic geometric features, the attitude perception features, and the topological relationship features are concatenated and then input in parallel to the mechanical origin discrimination branch and the adversarial decoupling dual-branch encoder. The mechanical origin discrimination branch outputs mechanical pattern features, and the adversarial decoupling dual-branch encoder outputs domain-invariant features and domain-specific features. The mechanical pattern features serve as a supervision signal to guide the feature decoupling process of the adversarial decoupling dual-branch encoder.

[0029] In one optional embodiment, the orientation constraint encoding module includes: a normal vector-orientation converter, a spherical position encoder, and an orientation consistency constraint layer;

[0030] Inputting the basic geometric features into the attitude constraint encoding module includes:

[0031] Extract local features of the second-level geometric abstraction layer and corresponding point normal vector estimates from the basic geometric features;

[0032] The normal vector estimate is converted into geological attitude parameters using the normal vector-attitude converter; wherein, the dip angle and strike of the normal vector estimate are calculated, the normal vector-attitude converter determines two candidate attitudes using the dip angle and strike, determines the final attitude from the two candidate attitudes through local neighborhood consensus voting, and converts the final attitude into geological attitude parameters;

[0033] The geological attitude parameters are mapped to a high-dimensional feature space by a spherical position encoder, and the attitude consistency constraint layer is used to perform feature constraints in the high-dimensional feature space to generate attitude perception features.

[0034] The attitude-aware features are concatenated with the global features of the basic geometric features to generate an attitude-enhanced set of features. In an optional embodiment, the multi-scale feature pyramid undergoes progressive domain adaptation training, including:

[0035] A multi-task objective function, including structural surface segmentation loss, geological rationality constraint loss, and comparative learning loss, is used to pre-train the overall structural features of the slope to generate a general geological basic model.

[0036] Regional meta-learning is performed on the general geological model and the tectonic domain features to generate regional adaptive initialization parameters;

[0037] Unlabeled data are extracted from the slope point cloud data. Adversarial training is performed using the region adaptation initialization parameters and the construction domain features. Domain alignment is performed during adversarial training to generate a basic site model and a domain-aligned feature representation.

[0038] The basic site model is optimized using the point cloud detail features to generate an optimized site model.

[0039] In one optional embodiment, performing regional meta-learning on the general geological model and the tectonic domain features includes: using the parameters of the general geological model as initial parameters, generating parameter modulation vectors from the tectonic domain features through a tectonic domain encoder, and performing an affine transformation on the initial parameters using the parameter modulation vectors to obtain tectonic domain-aware parameters; wherein generating parameter modulation vectors from the tectonic domain features through the tectonic domain encoder includes: randomly sampling from the current tectonic domain features as support set samples, performing dimensionality reduction processing on the support set samples through two fully connected layers in the tectonic domain encoder to generate embedding vectors; aggregating the embedding vectors, and performing dimensionality increase processing on the aggregated embedding vectors through two fully connected layers in the tectonic domain encoder to generate parameter modulation vectors;

[0040] The gradients of the constructed domain perception parameters are updated sequentially in the inner and outer loops to obtain the meta-learning adaptation parameters.

[0041] The region prototype vector of the constructed domain is constructed using the meta-learning adaptation parameters;

[0042] The meta-learning adaptation parameters and the region prototype vector are iterated to generate region adaptation initialization parameters.

[0043] In one optional embodiment, the base features are input to a structural surface segmentation head for processing, including:

[0044] The structural surface segmentation head receives the base features and processes them through a point-by-point multilayer perceptron to output a point-by-point structural surface mask. The base features are sequentially processed by the point-by-point multilayer perceptron through linear transformation, batch normalization, activation function and random deactivation operations, and finally the probability of each point belonging to the structural surface is output through the Sigmoid function.

[0045] In an optional embodiment, the basic features are input into the morphology regression head for processing, including:

[0046] After receiving the basic features, the attitude regression head performs local geometric enhancement on the basic features;

[0047] For each point in the base features of local geometric enhancement, search for neighboring points to construct a local point set, and calculate the covariance matrix of the local point set;

[0048] Normal vector features are extracted through the covariance matrix. The normal vector features are concatenated with the basic features and then input into the attitude regression multilayer perceptron. The attitude regression multilayer perceptron receives the concatenated normal vector features and basic features through the input layer, and then performs stepwise dimensionality reduction through multiple fully connected layers. A ReLU activation function is connected after each fully connected layer. Finally, the output layer outputs the predicted values ​​of dip and tilt angle for each point.

[0049] The predicted dip and tilt angle values ​​are mapped to the ranges of 0 to 360 degrees and 0 to 90 degrees respectively using activation functions, and then optimized using the L2 loss function to obtain the attitude prediction results.

[0050] In an optional embodiment, the base features are input to the structural surface group partitioning head for processing, including:

[0051] The orientation prediction results are converted into three-dimensional unit vectors;

[0052] The three-dimensional unit vector is concatenated with the base features to obtain joint features;

[0053] The optimized site model is used to learn the prototype vectors of each joint group, and the distance between the joint feature and the prototype vector at each point is calculated;

[0054] Based on the distance, the probability of each point belonging to each joint group is output through the Softmax function to complete point-by-point grouping;

[0055] The point-by-point grouping is optimized by using cross-entropy loss and prototype separation regularization term to obtain the joint group division results.

[0056] A second aspect of the present invention provides a generalized identification system for slope rock mass structural surfaces based on multi-level domain adaptation, comprising:

[0057] The multi-scale feature module is used to measure the outcrop rock mass, acquire slope point cloud data, and perform multi-level feature decomposition and representation learning on the slope point cloud data to obtain a multi-scale feature pyramid.

[0058] The progressive domain adaptation module is used to progressively adapt the multi-scale feature pyramid to generate an optimized site model and a domain-aligned feature representation.

[0059] The multi-task joint optimization module is used to perform multi-task joint optimization on the domain-aligned feature representation based on a multi-task learning framework with geological knowledge embedding, to obtain structural surface identification results, attitude prediction results, and joint group division results. The multi-task joint optimization includes: extracting basic features from the domain-aligned feature representation using a shared feature encoder, inputting the basic features into three task branches—structural surface segmentation head, attitude regression head, and structural surface group division head—for processing, and performing geological knowledge constraint verification and multi-task joint optimization on the processed data.

[0060] The feedback module is used to analyze the structural surface identification results, the attitude prediction results, and the joint group division results through uncertainty quantification and active learning mechanisms, generate slope rock mass identification results, and feed the slope rock mass identification results back to the progressive domain adaptation module for data supplementation.

[0061] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0062] The core of this invention lies in its ability to precisely overcome the inherent limitations of existing models that are purely data-driven through a collaborative design of three core processes: multi-level feature decomposition and representation learning, providing full-scale, high-quality feature support for generalized recognition; progressive domain adaptation training, effectively addressing the pain points of insufficient training data coverage and the model's tendency to overfit local features; and embedding geological knowledge into multi-task learning, compensating for the shortcomings of existing models that lack geological principle guidance, achieving full-scale feature adaptation and deep coupling with geological knowledge, ultimately breaking free from the bottleneck of single-site customization, achieving the intelligent goal of "one-time training, multiple applications," and significantly improving the generalization and engineering practicality of slope rock mass structure surface recognition. Attached Figure Description

[0063] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0064] Figure 1 This is a flowchart illustrating the generalized identification method for slope rock mass structure surfaces based on multi-level domain adaptation provided in Embodiment 1 of the present invention.

[0065] Figure 2 This is a schematic diagram of the generalized identification system for slope rock mass structure surfaces based on multi-level domain adaptation provided in Embodiment 2 of the present invention. Detailed Implementation

[0066] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0067] Embodiment 1 of this invention provides a generalized identification method for slope rock mass structural surfaces based on multi-level domain adaptation, such as... Figure 1 As shown, the generalized identification method for slope rock mass structural surfaces based on multi-level domain adaptation includes the following steps:

[0068] Step S1: Measure the outcrop rock mass to obtain slope point cloud data, and perform multi-level feature decomposition and representation learning on the slope point cloud data to obtain a multi-scale feature pyramid.

[0069] Step S2: Perform progressive domain adaptation training on the multi-scale feature pyramid to generate an optimized site model and a domain-aligned feature representation;

[0070] Step S3: Based on the multi-task learning framework with geological knowledge embedding, perform multi-task joint optimization on the domain-aligned feature representation to obtain the structural surface identification result, attitude prediction result, and joint group division result; wherein, the multi-task joint optimization includes: extracting basic features from the domain-aligned feature representation using a shared feature encoder, inputting the basic features into three task branches—structural surface segmentation head, attitude regression head, and structural surface group division head—for processing, and performing geological knowledge constraint verification and multi-task joint optimization on the processed data;

[0071] Step S4: Analyze the structural surface identification results, the attitude prediction results, and the joint group division results through uncertainty quantification and active learning mechanisms to generate slope rock mass identification results, and feed the slope rock mass identification results back to step S2 for data supplementation.

[0072] It should be noted that, firstly, a multi-level feature decomposition and representation learning network is constructed to establish a feature extraction system with four levels: global geological general, regional semantic, site-specific, and instance detail. This system accurately captures structural surface features at different levels of abstraction. At the same time, a feature decoupling network is used to explicitly separate domain-invariant and domain-specific features. Combined with a multi-scale feature pyramid, features of different resolutions are fused. This fundamentally solves the core problem of existing models having one-sided feature extraction and being difficult to adapt to different geological scenarios, laying a high-quality feature foundation for subsequent domain adaptation training.

[0073] Then, a progressive domain adaptation training strategy is adopted, which optimizes the model's adaptability in four stages. The first stage is based on large-scale pre-training on a global multi-regional mixed dataset and learning common knowledge across geological regions by combining multi-task objective functions. The second stage trains the model to quickly adapt to the initial parameters of new regions through regional meta-learning. The third stage adopts adversarial domain adaptation combined with prototype alignment to achieve site adaptation and generate domain-indistinguishable but task-related features. This step effectively makes up for the shortcoming that the training data cannot cover a wide range of geological conditions, avoids the model overfitting to local features, and achieves accurate adaptation of the model from cross-regional to single-site.

[0074] Next, a multi-task learning framework is constructed based on embedded geological knowledge. Basic features are extracted through a shared feature encoder and then processed collaboratively by a structural surface segmentation head, an attitude regression head, and a structural surface group segmentation head. At the same time, geological constraints such as attitude consistency and topological rationality are introduced to solve the problems of existing models lacking geological principle guidance and prediction results not conforming to actual geological laws, thus balancing recognition accuracy and geological rationality.

[0075] Finally, the model's generalization ability is further improved through uncertainty quantification and active learning feedback optimization.

[0076] Compared to existing technologies, the core inventiveness of this invention lies in its precise breakthrough of the inherent limitations of purely data-driven existing models through the collaborative design of three core processes: multi-level feature decomposition and representation learning, providing full-scale, high-quality feature support for generalized recognition; progressive domain adaptation training, effectively addressing the pain points of insufficient training data coverage and the model's tendency to overfit local features; and embedding geological knowledge into multi-task learning, compensating for the shortcomings of existing models lacking geological principle guidance, achieving full-scale feature adaptation and deep coupling with geological knowledge, ultimately truly breaking free from the bottleneck of single-site customization, achieving the intelligent goal of "one-time training, multiple applications," and significantly improving the generalization and engineering practicality of slope rock mass structure surface recognition.

[0077] In one optional embodiment, the slope point cloud data undergoes multi-level feature decomposition and representation learning, including:

[0078] Rock mass structural features at different levels of abstraction were extracted from the slope point cloud data;

[0079] The extracted rock mass structural surface features are explicitly decomposed into domain-invariant features and domain-specific features using a feature decoupling network;

[0080] A multi-scale feature pyramid is used to hierarchically fuse domain-invariant features and domain-specific features, outputting a multi-scale pyramid; the hierarchical fusion includes:

[0081] The first level is used to learn the basic geometric morphology, topological relationship and mechanical genesis of the structural surfaces in domain-invariant features and domain-specific features, and to generate the overall structural features of the slope.

[0082] The second level is used to learn discriminative features from domain-invariant features and domain-specific features, and to generate construct domain features;

[0083] The third level is used to learn the lithology, weathering, and unloading characteristics at the slope scale in both domain-invariant and domain-specific features, and to generate specific slope segment features.

[0084] The fourth level is used to learn instance detail features from domain-invariant and domain-specific features to generate point cloud detail features.

[0085] The process begins with preprocessing the slope point cloud data: the 3D space is divided into cubic meshes based on the slope scene, and a representative point is retained within each voxel of the cubic mesh; in this embodiment, the center point is selected. For each representative point, the average distance of its k-nearest neighbors is calculated, and points whose average distance exceeds the global mean ± 3 standard deviations are removed, thereby eliminating noise points (such as birds and dust reflections) collected by the UAV.

[0086] Then, geometric features are extracted from the cleaned slope point cloud data to form a low-level feature representation: First, the normal vector of each point is estimated. This estimation is based on principal component analysis, searching for the 30 nearest neighbors of each point, constructing a local covariance matrix, and obtaining the eigenvector corresponding to the minimum eigenvalue through eigenvalue decomposition. This vector is the direction of the local surface normal vector at that point. Second, the curvature feature of each point is calculated. Curvature reflects the degree of bending of the local surface and is obtained by calculating the deviation of the normal vector of the point from the normal vectors of its nearest neighbors. The larger the curvature value, the more uneven the surface at that point, possibly located at the edge of a structural plane or in a fractured rock mass area. Third, the local density feature of each point is calculated. Local density is defined as the reciprocal of the average distance between the point and its nearest neighbors. The density feature can distinguish between dense and loose areas on the rock mass surface. Finally, a fast point feature histogram feature is extracted. This feature forms a 33-dimensional histogram by statistically analyzing the angle and distance relationship between the normal vectors of a point and its nearest neighbors, which can effectively characterize the local geometric pattern. The combination of the above geometric features forms a 38-dimensional geometric feature vector, including a 3-dimensional normal vector, a 1-dimensional curvature, a 1-dimensional density, and a 33-dimensional fast point feature histogram. This feature vector constitutes a low-level geometric representation of the point.

[0087] Furthermore, for slope structural surface identification, this embodiment proposes to extract semantic features at the structural surface level based on geometric feature extraction, forming a mid-level feature representation, thereby supporting the subsequent generalized identification of structural surfaces. In this embodiment, the semantic extraction at the structural surface level is implemented based on a region growing segmentation and random sampling consistency plane fitting algorithm. Specifically, it includes: selecting points with curvature less than 0.1 as candidate seed points from all points in the fast point feature histogram feature, and sorting them in ascending order of curvature. Low curvature points are usually located inside the structural surface rather than at the edge, making them suitable as the starting point for region growing; starting from the seed point, iteratively searching for its nearest neighbors; if the angle between the normal vector of a nearest neighbor point and the normal vector of the seed point is less than 5 degrees, then the point is assigned to the same region. This process continues until no new points are added, ultimately forming an initial structural surface region. For each structural surface region, a random sampling consensus algorithm is used to fit a planar model. Specifically, three points are randomly selected from the structural surface region to define a plane. The distances from all points within the region to this plane are calculated, and the number of interior points with a distance less than 0.05 meters is counted. After multiple iterations, the planar model with the most interior points is selected as the fitting result for that structural surface. Finally, based on the fitted planar model and the region's point set, the structural surface attribute features are calculated. These features include: the structural surface's normal vector (used to determine attitude), dip and dip angle (geological attitude parameters), structural surface area, surface roughness (standard deviation of the distance from a point to the fitted plane), extension dimensions (lengths in three directions based on principal component analysis), continuity (ratio of maximum to minimum extension dimensions), centroid coordinates, and fit confidence.

[0088] The above-mentioned structural surface attribute features are combined to form a 16-dimensional structural surface semantic feature vector. This feature vector is indexed and assigned to all points belonging to the structural surface, so that each point obtains its intermediate structure representation.

[0089] Furthermore, based on the semantic features of structural surfaces, the structural surfaces are treated as nodes in a graph, and the spatial relationships between structural surfaces are treated as edges. A graph neural network is used to extract contextual relationship features between structural surfaces, forming a high-level feature representation. Specifically, this includes: using the centroid (geometric center) of each structural surface as a node, calculating the Euclidean distance between the centroids of all structural surfaces to form a spatial distance matrix. Simultaneously, the angle between the normal vectors of all structural surfaces is calculated to form a directional similarity matrix. Combining spatial distance and directional relationships, a weighted adjacency matrix is ​​constructed (structural surfaces that are spatially adjacent and orthogonal in direction (angle close to 90 degrees) are assigned higher adjacency weights, because orthogonal tangency relationships typically represent X-type joint systems in geology). For each pair of structural surfaces, the centroid distance, the angle between their normal vectors, the direction of their intersection (obtained through the cross product of normal vectors), and the type of mechanical relationship are calculated. The mechanical relationship type is determined based on prior geological knowledge: if the angle between the normal vectors of two structural surfaces is less than 15 degrees, they are considered coplanar (belonging to the same set of joints); if the angle is between 75 and 105 degrees, they are considered orthogonal and tangent; if the centroid of one structural surface is located within the extension range of another structural surface, it is considered a restrictive relationship. The semantic features of the structural surfaces are used as the initial features of the nodes, and information is propagated through a three-layer graph attention network. Each layer of the graph attention network calculates the attention coefficients between nodes, aggregates the features of neighboring nodes, and introduces edge features for relationship weighting. The first layer maps 16-dimensional features to 256 dimensions (4-head attention, 64 dimensions per head), the second layer maintains 256 dimensions, and the third layer maps to a 128-dimensional output. The final obtained contextual relationship feature is a 128-dimensional vector, which characterizes the semantic role and relationship attributes of each structural surface in the entire slope rock mass system.

[0090] After multi-scale feature extraction, a hierarchical feature abstraction module fuses features from different scales and maps them to a unified feature space, forming three abstract levels of feature representation. The fused features are then output for use by the subsequent decoupling network. The hierarchical feature abstraction module consists of four sub-encoders and one feature fusion network. The first sub-encoder is a low-level geometric feature encoder, taking 38-dimensional geometric features as input. It is mapped to 128-dimensional low-level geometric features through two fully connected layers (hidden layer dimensions of 64 and 128, respectively). This encoder uses batch normalization and ReLU activation functions to learn point-level local geometric patterns. The second sub-encoder is a mid-level structural feature encoder, taking 16-dimensional structural surface semantic features as input. It is mapped to 64-dimensional mid-level structural features through two fully connected layers (hidden layer dimensions of 32 and 64, respectively). This encoder learns structural surface-level attribute representations. The third sub-encoder is a high-level semantic feature encoder, taking 128-dimensional contextual relationship features as input. It is mapped to 512-dimensional high-level semantic features through two fully connected layers (hidden layer dimensions of 256 and 512, respectively). This encoder learns scene-level global semantics. The fourth sub-encoder is a feature fusion network, which concatenates the features from the above three layers (total dimensions are 128 + 64 + 512 = 704 dimensions) and maps them to a 512-dimensional fused feature through two fully connected layers (hidden layer dimension is 512). This fused feature integrates point-level geometric details, structural surface-level attribute information, and scene-level relational semantics, serving as the input to the subsequent feature decoupling network.

[0091] The hierarchical feature abstraction module processes the following steps: For each point in the input point cloud, a low-level encoder first obtains low-level geometric features, which mainly include local geometric attributes such as normal vectors and curvature; a mid-level encoder then obtains mid-level structural features, which mainly include structural surface attributes such as attitude and roughness; a high-level encoder then obtains high-level semantic features, which mainly include global semantics such as structural surface groupings and tangent relationships; finally, a fusion network obtains fusion features, which will be used as input to a feature decoupling network to separate domain-invariant features from domain-specific features.

[0092] In one optional embodiment, the feature decoupling network includes: a shared geometric coding module, an attitude constraint coding module, a topological graph neural network, a mechanical origin discrimination branch, and an adversarial decoupling dual-branch encoder;

[0093] Specifically, a feature decoupling network is used to explicitly decompose the extracted rock mass structural surface features into domain-invariant features and domain-specific features, including:

[0094] The basic geometric features in the rock mass structural surface features are extracted using the shared geometric coding module.

[0095] The basic geometric features are input in parallel into the attitude constraint encoding module and the topology graph neural network to obtain attitude perception features and topology features;

[0096] The basic geometric features, the attitude perception features, and the topological relationship features are concatenated and then input in parallel to the mechanical origin discrimination branch and the adversarial decoupling dual-branch encoder. The mechanical origin discrimination branch outputs mechanical pattern features, and the adversarial decoupling dual-branch encoder outputs domain-invariant features and domain-specific features. The mechanical pattern features serve as a supervision signal to guide the feature decoupling process of the adversarial decoupling dual-branch encoder.

[0097] It should be noted that the characteristics of rock mass structural planes are significantly different from those of data features in the following ways: First, structural planes have clear geological occurrence attributes, including three interrelated parameters: strike, dip angle, and dip direction. These parameters determine the orientation of the structural planes in three-dimensional space. Second, there are complex spatial topological relationships between structural planes, including parallel assemblages, intersecting relationships, and constraint relationships. These relationships reflect the tectonic evolution history of the rock mass. Third, the geometric morphology of structural planes is controlled by mechanical genesis. Tensile, shear, and compressive structural planes have different morphological characteristics and distribution patterns. Fourth, the development degree of structural planes is closely related to geological factors such as lithology, weathering, and unloading, exhibiting strong domain specificity.

[0098] Therefore, in this embodiment, the structure of the feature decoupling network is improved to address the aforementioned specificities. First, an attitude constraint encoding module is introduced to embed the spherical geometric characteristics of geological attitudes into the feature learning process. Second, a topological graph neural network is constructed to explicitly model the high-order relationships between structural surfaces. Finally, a mechanical genesis discrimination branch is designed to achieve joint decoupling of mechanical patterns and geometric morphology. This enables the feature decoupling network to more effectively process the features of rock mass structural surfaces and extract truly geologically significant domain-invariant properties.

[0099] Specifically, the shared geometry encoding module is constructed using an improved PointNet++ network. Existing PointNet++ networks use sphere queries for local neighborhood grouping, but this isotropic grouping method is inefficient for structural surfaces with significant planar characteristics. In this embodiment, at each sampling layer, the local principal curvature of candidate center points is first estimated. If the principal curvature ratio is less than a threshold (indicating a local approximation of a plane), a flattened ellipsoid query is used instead of the standard sphere query. The major axis of the ellipsoid extends along the estimated plane, the minor axis along the normal direction, and the ratio of the major semi-axis to the minor semi-axis is 3:1. This anisotropic grouping strategy can more efficiently capture the local geometric context of the structural surface. Specifically, the shared geometry encoding module includes three levels of geometric abstraction layers. The first, second, and third levels of geometric abstraction layers are used to extract local features of different dimensions from the rock mass structural surface features. These layers sample the rock mass structural surface features through adaptive flattened grouping with different radii, thereby obtaining local features of different sizes and mapping these local features to different dimensions.

[0100] The first-level geometric abstraction layer uses adaptive flattened grouping with a radius of 0.2 meters, sampling 512 center points, and uses a multilayer perceptron to map local features to 128 dimensions; the second-level geometric abstraction layer uses grouping with a radius of 0.4 meters, sampling 256 center points, and mapping features to 256 dimensions; the third-level geometric abstraction layer uses grouping with a radius of 0.8 meters, sampling 128 center points, and mapping features to 512 dimensions; finally, a 512-dimensional global geometric feature vector is obtained through global max pooling; the global geometric feature vector represents the global shape attributes of the point cloud, but it does not yet show the encoding of geological semantic information such as attitude and topology.

[0101] The attitude constraint coding module is one of the core concepts of this method. Its function is to explicitly embed the spherical geometric characteristics of geological attitudes into the feature learning process, enabling the network to respect the spatial distribution patterns of geological attitudes. Specifically, geological attitudes exhibit periodicity in strike (0 degrees are equivalent to 360 degrees) and boundedness in dip (0 to 90 degrees). Feature learning in standard Euclidean space struggles to handle these periodic and bounded parameters; therefore, this embodiment improves the attitude constraint coding structure.

[0102] In one optional embodiment, the orientation constraint encoding module includes: a normal vector-orientation converter, a spherical position encoder, and an orientation consistency constraint layer;

[0103] Inputting the basic geometric features into the attitude constraint encoding module includes:

[0104] Extract local features of the second-level geometric abstraction layer and corresponding point normal vector estimates from the basic geometric features;

[0105] The normal vector estimate is converted into geological attitude parameters using the normal vector-attitude converter; wherein, the dip angle and strike of the normal vector estimate are calculated, the normal vector-attitude converter determines two candidate attitudes using the dip angle and strike, determines the final attitude from the two candidate attitudes through local neighborhood consensus voting, and converts the final attitude into geological attitude parameters;

[0106] The geological attitude parameters are mapped to a high-dimensional feature space by a spherical position encoder, and the attitude consistency constraint layer is used to perform feature constraints in the high-dimensional feature space to generate attitude perception features.

[0107] The attitude-aware features are concatenated with the global features of the basic geometric features to generate an attitude-enhanced set of features.

[0108] It should be noted that the local features of the second-level geometric abstraction layer and the normal vector estimation of the corresponding points are extracted from the basic geometric features. The local features of the second-level geometric abstraction layer are the 256-dimensional local features of the second-level geometric abstraction layer.

[0109] The dip angle estimated by the normal vector is the angle between the normal vector and the horizontal plane, and the strike is the azimuth angle of the horizontal projection of the normal vector. Since the normal vector has bidirectional ambiguity, the normal vector-attitude converter will output two candidate attitudes. It is necessary to determine one of the two candidate attitudes and convert it into geological attitude parameters. In this embodiment, this is achieved through local neighborhood consensus voting.

[0110] In this embodiment, spherical harmonic basis function encoding is used to map geological attitude parameters into a high-dimensional feature space. Specifically, for dip angle, associated Legendre polynomials are used; for strike, Fourier series are used; the combination of the two forms a 32-dimensional attitude-aware feature. This encoding method naturally satisfies the geometric characteristics of spherical measurement, enabling similar attitudes to have similar representations in the feature space.

[0111] The attitude consistency constraint layer enforces geological consistency during feature learning. It defines an attitude smoothing loss to penalize abrupt changes in attitude features between adjacent points, and an attitude clustering loss to encourage points with the same attitude to cluster in the feature space. These constraints are added to the total loss in the form of an auxiliary loss function to guide the network in learning geologically reasonable attitude representations.

[0112] Finally, the attitude-aware features are concatenated with the 512-dimensional global features to form a 544-dimensional attitude-enhanced geometric feature. Through the attitude-constrained encoding module, the spherical geometric characteristics of geological attitudes are explicitly embedded into feature learning, so that the feature representation output by the network respects the spatial distribution law of geological attitudes. The structural surfaces of similar attitudes have inherent similarity in the feature space, which improves the geographical consistency of domain-invariant features.

[0113] Furthermore, topological graph neural networks explicitly model higher-order topological relationships between structural surfaces, associating the individual properties of structural surfaces with their roles in the rock mass structure. The topological relationships of rock mass structural surfaces include several types: parallel relationships indicate that structural surfaces belong to the same set of joint systems; tangential relationships indicate that two sets of structural surfaces cut each other, usually forming X-shaped joints; restrictive relationships indicate that the extension of one set of structural surfaces is terminated by another set; and inclusion relationships indicate that smaller structural surfaces develop on top of larger ones. These relationships contain rich geological information, but standard feature encoding treats them as independent samples, losing the context of these relationships.

[0114] The 512-dimensional global geometric features from the shared geometric encoding module and the structural surface instance mask are processed by a topological graph neural network. The structural surface instance mask divides the point cloud into several connected regions, and each region corresponds to a potential structural surface instance.

[0115] The processing steps of the topological graph neural network include: extracting representative features of each structural surface based on instance masks (including average pooling of geometric features of all points on the structural surface, centroid coordinates, and principal and secondary extension directions obtained through principal component analysis). Then, a structural surface adjacency graph is constructed, with each structural surface as a node. If the centroid distance between two structural surfaces is less than a preset threshold (usually 20% of the slope scale), an edge is established between the nodes. The initial weight of the edge is determined by both spatial proximity and directional relationship; structural surfaces with similar directions (angle less than 15 degrees) or orthogonal directions (angle 75 to 105 degrees) are assigned higher weights.

[0116] For each edge, multidimensional relationship features are calculated (centroid Euclidean distance, angle between normal vectors, intersection direction vector (obtained from the cross product of normal vectors and normalized), and mechanical relationship type encoding). The mechanical relationship type is inferred based on geometric relationships: parallel relationships correspond to an angle between normal vectors less than 15 degrees; orthogonal and tangent relationships correspond to an angle between normal vectors between 75 and 105 degrees and a relatively close centroid distance; constraint relationships correspond to the centroid of one structural surface being located within the projection range of the principal extension axis of another structural surface. These relationship features are mapped to a 16-dimensional edge embedding vector using a multilayer perceptron.

[0117] In the propagation phase of the graph attention network, a relation-aware graph attention mechanism is used for message passing. Standard graph attention calculates attention coefficients based solely on node features. This embodiment employs relation-enhanced attention calculation, where the attention coefficient is determined by both node feature similarity and edge embedding; specifically, it is the result of modulating the dot product of node features through edge embedding. This design ensures that message passing strength depends not only on node attribute similarity but also on the relation type. Specifically, it includes three graph attention layers with a hidden layer dimension of 128. Each layer is followed by a normalization and ReLU activation layer, ultimately outputting a 128-dimensional topological relation-enhanced feature for each structural surface.

[0118] The 128-dimensional topological relationship enhancement features of each structural facet are concatenated with the attitude enhancement geometric features to form a 672-dimensional comprehensive shared feature representation, which is then input into the subsequent adversarial decoupled dual-branch encoder. Through a topological relationship graph neural network, higher-order relationships between structural faces are explicitly modeled, enabling the network to identify the role of structural faces in the rock mass structure (such as dominant structural faces and associated structural faces). These relationship features, as domain-invariant properties, enhance the model's ability to understand complex rock mass structures.

[0119] The internal structure of the mechanical origin discrimination branch is a three-layer classification network: the first fully connected layer maps 672 dimensions to 256 dimensions, followed by batch normalization and ReLU activation; the second fully connected layer maps 256 dimensions to 64 dimensions; and the third fully connected layer maps 64 dimensions to 3 dimensions, corresponding to the logits of the three categories of tensile, shear, and compressive stresses, which are then processed by Softmax to output the mechanical origin probability distribution. Through the mechanical origin discrimination branch, geomechanical knowledge is introduced as a supervisory signal to guide the network in learning essential properties related to mechanical patterns. These properties have clear physical meaning and cross-slope universality, improving the geological interpretability of domain-invariant features.

[0120] Furthermore, the adversarial decoupled bi-branch encoder processes the comprehensive shared features after attitude constraint, topology enhancement, and mechanical supervision. The dimensionality is 672 dimensions multiplied by the batch size, significantly higher than the 512-dimensional input of the standard decoupled network, containing richer geological semantic information. Based on this, the adversarial decoupled bi-branch encoder in this embodiment introduces an attitude-topology joint attention mechanism and designs a hierarchical domain-specific feature decomposition. Specifically, the attitude-topology joint attention mechanism is inserted between the first and second fully connected layers, decomposing the 672-dimensional features into an attitude-related subspace (256 dimensions), a topology-related subspace (256 dimensions), and a general subspace (160 dimensions), calculating self-attention separately before merging. This decomposed attention allows the network to explicitly focus on different types of geological attributes, avoiding information mixing. Attention calculation employs a multi-head mechanism: four attention heads are used each for the attitude and topology subspaces, and two attention heads are used for the general subspace. The outputs are concatenated to restore the 672-dimensional feature. By employing the attitude-topology joint attention mechanism, the network can explicitly focus on different types of geological attributes when processing high-dimensional integrated features, avoiding information mixing and improving the purity and quality of feature decoupling.

[0121] Standard decoupled networks output a single domain-specific feature vector, but the domain-specific attributes of rock mass surfaces contain multiple levels: lithological attributes (color, texture, mineral composition), weathering attributes (surface alteration degree), and acquisition attributes (illuminance, angle, resolution). This embodiment expands the output layer of the domain-specific encoder into three parallel sub-branches, each outputting 64-dimensional features, corresponding to the lithological subspace, weathering subspace, and acquisition subspace, respectively. The total dimension remains 192 dimensions (different from the standard 128 dimensions; this is an enhancement design and can be adjusted in actual implementation). Each sub-branch connects to an independent domain classifier subheader, forming a multi-task domain classification framework, enabling the network to explicitly learn different types of domain-specific attributes.

[0122] The domain-invariant feature encoder maintains the standard architecture, but expands the input dimension to 672 dimensions while keeping the output dimension at 128 dimensions. This encoder also incorporates an attitude-topology joint attention mechanism, sharing attention computation with the domain-specific branch but with independent parameters.

[0123] The encoder's adversarial training mechanism is consistent with that of standard networks, achieving adversarial learning of domain-invariant features through a gradient inversion layer. The difference lies in that, since the domain-specific branch contains three subspaces, the domain classification loss is expanded to a weighted sum of the three sub-losses, with the weights adaptively adjusted according to the classification difficulty during training.

[0124] Furthermore, the multi-scale feature pyramid adopts a four-level progressive fusion structure, with each level corresponding to different spatial resolutions and abstraction levels, and achieves dynamic fusion of domain-invariant features and domain-specific features. Through a hierarchical dynamic fusion strategy of domain-invariant and domain-specific features, the network can adaptively balance general knowledge across slopes with slope-specific information at different abstraction levels. By integrating multi-view 3D augmentation at the site-specific feature level, the spatial consistency of features is enhanced by utilizing the geometric constraints of photogrammetry, thereby solving the technical problems of insufficient feature representation from a single viewpoint and inconsistent prediction results from different viewpoints.

[0125] Specifically, firstly, domain-invariant features and domain-specific features are concatenated along the channel dimension to form a 256-dimensional concatenated feature. This concatenated feature is then mapped to a 512-dimensional high-dimensional feature space through a linear projection layer, yielding the basic feature representation. Secondly, the basic feature representation undergoes multi-scale decomposition to generate multi-scale feature sets corresponding to different spatial resolutions, including feature maps at four levels: P2, P3, P4, and P5, corresponding to scales of 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the original resolution, respectively. This multi-scale decomposition is achieved through stride convolution or pooling operations to ensure semantic consistency across all feature levels.

[0126] The first-level feature fusion employs a strong-than-invariant constraint strategy, fusing domain-invariant features and domain-specific features at a 9:1 ratio to ensure that this level of features primarily encodes general geological attributes across slopes, rather than personalized information specific to a particular slope. The P5 feature map is input into the first level. Firstly, the first level adaptively weights the feature channels using a channel attention mechanism with a compression ratio of 16, enhancing the response to key geological attributes. Secondly, it weights the spatial location of the feature map using a spatial attention mechanism, focusing on key areas of structural surface development. Finally, a global average pooling operation compresses the two-dimensional feature map into a one-dimensional vector, resulting in a globally geologically universal feature tensor with dimensions equal to the batch size multiplied by 512, representing the overall structural features of the slope (this tensor characterizes the overall structural pattern of the slope and the global distribution pattern of rock mass structural surfaces, possessing the strongest cross-slope migration capability). The learning focus at this level includes three aspects of general knowledge across geological zones: the basic geometric morphology of structural surfaces, including planar, curved, or broken-line structural surfaces; the topological relationships between structural surfaces, including continuity, tangency, and constraint relationships; and the mechanical origin of structural surfaces, including different genetic types such as tension joints, shear joints, and bedding planes.

[0127] The second-level strategy employs a moderate domain adaptation approach, fusing domain-invariant features with domain-specific features in a 7:3 ratio. This allows the level to simultaneously utilize general knowledge across slopes and region-specific information, adapting to the characteristic differences of different geological provinces or tectonic units. Specifically, the second level first uses 1×1 convolution for lateral connection dimensionality reduction, aligning the number of feature channels after upsampling from the first level with the P4 features. Secondly, it fuses the upsampling features of the first level with the P4 features element-wise, combining deep semantic information and shallow spatial details. Finally, it learns the discriminative feature tensor of different geological provinces (this tensor represents the distribution characteristics of rock mass structural surfaces at the regional scale and has strong regional adaptability) through a regional attention mechanism, i.e., tectonic domain features, such as the specific geological background of the western margin tectonic domain of the Hengduan Mountains. The second level focuses on learning the tectonic domain features of the regional geological background, including different tectonic stress fields such as compressional, extensional, and strike-slip environments; lithological assemblage patterns, including the regional distribution patterns of igneous, sedimentary, and metamorphic rocks; and regional weathering patterns, including the weathering characteristics of rock masses in different climatic zones.

[0128] The third-level strategy employs a balanced fusion approach, integrating domain-invariant features and domain-specific features in a 5:5 ratio. This makes this level the optimal combination of domain-invariant knowledge and domain-specific information, maintaining the ability to identify the essential attributes of structural surfaces while adapting to specific site conditions. The P3 feature map and the upsampling results of the second-level features are used as input to the third level, injecting multi-view 3D enhanced features. A 3D-2D feature projection module integrates multi-view geometric consistency constraints into the feature representation. Secondly, a deformable convolutional network processes the features, enabling the sampling position of the convolutional kernel to adapt to the irregular shape of the slope, enhancing adaptability to complex slope geometry. Finally, a site encoder learns the lithology, weathering, and unloading feature tensors of a specific slope segment (these tensors represent detailed features of the rock mass structure at the site scale and are the feature level directly affected by multi-view 3D enhancement). These features are implemented through lightweight fully connected networks or convolutional layers. This level focuses on slope-scale characteristics, including the rock mass structure at the excavation face, such as the stepped outcrops and excavation damage zones of highway slopes; the characteristics of local unloading fracture development, such as stress release and fracture propagation patterns caused by slope excavation; and the relationship between groundwater outcrops and structural surfaces, such as the spatial correlation between seepage points, drip points and structural surfaces.

[0129] The fourth-level strategy emphasizes domain-specific details, fusing domain-invariant features with domain-specific features in a 3:7 ratio. This allows the level to fully capture the unique attributes of individual samples, such as local illumination variations and point cloud density differences. The upsampling results of the P2 feature map and the third-level features are used as input to the third level, preserving high-resolution feature details and avoiding the loss of fine structural information due to excessive downsampling. Secondly, instance normalization is used to optimize features in real time, adaptively adjusting the data distribution for a single image or frame of point cloud. Finally, 3×3 depthwise separable convolutions are used to capture fine textures, forming a point cloud detail feature tensor (this tensor represents the fine features of rock mass structure at the instance scale, possessing the highest spatial resolution and strongest sample specificity), enhancing the perception of local details while maintaining computational efficiency. This level focuses on learning real-time features of single samples, including local lighting variations in a single 5-meter close-up image, such as shadows, reflections, and exposure differences; detail differences caused by uneven point cloud density, such as fine structure in high-density areas and interpolation uncertainty in low-density areas; and surface roughness and infill features of structural surfaces, such as specular reflection, vegetation cover, and calcareous infill.

[0130] The multi-scale feature pyramid adopts an adaptive fusion strategy based on an attention mechanism to replace the fixed feature fusion ratio, enabling the network to dynamically adjust the contribution weights of domain-invariant features and domain-specific features according to the characteristics of the input data.

[0131] In one optional embodiment, progressive domain adaptation training is performed on the multi-scale feature pyramid, including:

[0132] A multi-task objective function, including structural surface segmentation loss, geological rationality constraint loss, and comparative learning loss, is used to pre-train the overall structural features of the slope to generate a general geological basic model.

[0133] Regional meta-learning is performed on the general geological model and the tectonic domain features to generate regional adaptive initialization parameters;

[0134] Unlabeled data are extracted from the slope point cloud data. Adversarial training is performed using the region adaptation initialization parameters and the construction domain features. Domain alignment is performed during adversarial training to generate a basic site model and a domain-aligned feature representation.

[0135] The basic site model is optimized using the point cloud detail features to generate an optimized site model.

[0136] Through a hierarchical training strategy, the model first learns common knowledge across geological regions at a macroscopic level, then learns discriminative features of specific tectonic domains at the regional level, achieves domain alignment through adversarial training at the site level, and finally performs continuous optimization at the instance level. This progressive training process corresponds to the four-level structure of the multi-scale feature pyramid, ensuring that the feature representations at each level are fully optimized and utilized.

[0137] First, a general geological foundation model is generated during the global pre-training phase. The goal is to learn common knowledge across geological regions and establish a general geological foundation model, providing good parameter initialization for subsequent regional adaptation and site adaptation. This phase employs a multi-task objective function for training, comprising three components: structural surface segmentation loss, geological rationality constraint loss, and contrastive learning loss. The structural surface segmentation loss uses standard cross-entropy loss or Dice loss, while the geological rationality constraint loss is designed based on prior geological knowledge to constrain the geological rationality of the model's output. The contrastive learning loss enhances the discriminability and transferability of features.

[0138] Furthermore, the geological rationality constraint loss specifically includes: attitude continuity constraint, which penalizes abrupt changes in attitude parameters between adjacent regions to ensure the spatial continuity of structural plane attitudes; topological consistency constraint, which encourages parallel structural planes to cluster in the feature space and intersecting structural planes to form specific angular relationships; and mechanical mode constraint, which guides the model to learn the morphological characteristics of structural planes with different mechanical origins. These constraints are added to the total loss in the form of an auxiliary loss function to ensure that the model output conforms to geological laws.

[0139] Furthermore, the contrastive learning loss specifically includes: using different perspectives or different augmented samples from the same structural surface instance as positive sample pairs, and using samples from different structural surface instances or different slopes as negative sample pairs, using InfoNCE loss to bring the feature distance of positive sample pairs closer and push the feature distance of negative sample pairs further apart.

[0140] In one optional embodiment, regional meta-learning is performed on the general geological foundation model and the tectonic domain features, including:

[0141] The parameters of the general geological model are used as initial parameters. The tectonic domain features are used to generate parameter modulation vectors through a tectonic domain encoder. The initial parameters are then subjected to affine transformation through the parameter modulation vectors to obtain the tectonic domain sensing parameters.

[0142] The gradients of the constructed domain perception parameters are updated sequentially in the inner and outer loops to obtain the meta-learning adaptation parameters.

[0143] The region prototype vector of the constructed domain is constructed using the meta-learning adaptation parameters;

[0144] The meta-learning adaptation parameters and the region prototype vector are iterated to generate region adaptation initialization parameters.

[0145] The parameters extracted from the general geological model include: in the early stage of meta-training, freezing the bottom-level feature extraction parameters (shared geometric coding module) and training only the top-level classification head and domain adaptation-related parameters to ensure that the basic geometric features are not destroyed; as training progresses, gradually unfreezing the middle-level parameters (attitude constraint coding module, topological relationship graph neural network) to allow them to adapt to region-specific features; and never unfreezing the bottom-level parameters to maintain the general geometric perception capability across regions.

[0146] The constructive domain encoder is a neural network used to extract task-level modulation information from support set samples. Its output is used to perform affine transformation on the basic model parameters to achieve fast domain adaptation. In this embodiment, the constructive domain encoder includes an input layer for receiving data, two fully connected layers for mapping input features to low-dimensional embeddings, an aggregation module for performing aggregation operations to capture common features within the domain, and two fully connected layers for mapping the aggregated embedding vector to the final parameter modulation vector.

[0147] The acquisition of support set samples includes: randomly sampling from the current tectonic domain features as support set samples, which are used to quickly adapt to task-specific parameters in the inner loop. The support set samples cover typical geological features of the tectonic domain, including representative lithology, dominant structural surface assemblages, and typical weathering patterns. Specifically, generating parameter modulation vectors from the tectonic domain features using a tectonic domain encoder includes: randomly sampling from the current tectonic domain features as support set samples; performing dimensionality reduction on the support set samples through two fully connected layers in the tectonic domain encoder to generate embedding vectors; aggregating the embedding vectors, which enables the tectonic domain encoder to capture common features within the support set; and finally, performing dimensionality increase on the aggregated embedding vectors through two fully connected layers in the tectonic domain encoder to generate parameter modulation vectors.

[0148] Samples that do not overlap with the support set are sampled from the same construction domain as query set samples. These samples are used to evaluate the performance of the adapted parameters in the outer loop and to calculate the meta-loss.

[0149] The inner loop gradient update includes: starting with the meta-parameters, the model updates rapidly along the structural surface segmentation loss gradient direction of the support set samples with the learning rate as the step size, guided by the geological prior of the initial parameters modulated by the tectonic domain features. This shifts general knowledge towards the unique structural surface distribution characteristics of the current slope. After one or a few steps of gradient descent, task-specific parameters are generated, enabling rapid customization of the ability to identify the structural surface of the current slope rock mass, while avoiding overfitting the support set and preserving cross-regional generalization potential.

[0150] The outer loop gradient update includes: the outer loop uses the comprehensive performance of the inner loop adaptation parameters of all tasks on their respective query sets as the optimization objective, and quantifies the generalization ability of the current meta-parameters by calculating the meta-loss, which includes structural surface segmentation loss and geological rationality constraint loss; the gradient of the meta-loss with respect to the initial meta-parameters needs to be backpropagated along the inner loop gradient chain, treating the inner loop gradient as a constant and updating it directly through the query set gradient, which greatly reduces the computational complexity; the meta-parameters are iteratively updated along the descent direction of the meta-loss under the control of the outer loop learning rate, gradually strengthening the model's ability to quickly extract tectonic domain-specific features from a small number of support set samples, and finally achieving small-sample adaptive identification of slope structural surfaces with different geological backgrounds.

[0151] The region prototype is a feature vector that represents the representativeness of the constructed domain and is used for subsequent domain alignment, similarity measurement, and classification decisions. For each constructed domain, the meta-learning adaptation parameters of the domain are used to extract features of all samples. Then, the sample features are aggregated by weighted averaging, with the weights determined by the quality and representativeness of the samples. Finally, the aggregated features are normalized to obtain the region prototype vector of the constructed domain.

[0152] After multiple rounds of meta-training, the final region adaptation initialization parameters are generated. These parameters consist of two parts: the model parameters optimized by meta-learning, and the completed region prototype library.

[0153] Furthermore, the site adaptation stage corresponds to the third level of the multi-scale feature pyramid, namely the site-specific feature level. The goal of this stage is to achieve domain alignment between the source and target domains through adversarial training, generating a site-customized model and domain-aligned feature representation suitable for a specific slope.

[0154] This phase employs an adversarial training framework, comprising three core components: a feature extractor, a structural surface segmentation predictor, and a domain classifier. The feature extractor initializes parameters based on region adaptation and outputs site-specific features; the structural surface segmentation predictor predicts structural surface segmentation results based on the output of the feature extractor; and the domain classifier attempts to distinguish whether features originate from the source domain or the target domain.

[0155] The objective function of adversarial training consists of two parts: a structural surface segmentation loss, computed only on source domain data to ensure the model retains its recognition ability; and a domain adversarial loss, which the feature extractor attempts to minimize (making the domain classifier unable to distinguish the source), while the domain classifier attempts to maximize (accurately distinguishing the source). This adversarial game is achieved through a gradient reversal layer, where the feature extractor learns domain-invariant feature representations.

[0156] Domain alignment is a key technology in this stage. Specifically, it involves minimizing the difference in feature distribution between the source and target domains using maximum mean difference or adversarial domain adaptation methods. Simultaneously, semantic alignment is performed using a regional prototype library, associating the features of the target domain samples with the most similar regional prototypes to maintain semantic consistency. The domain-aligned feature representation maintains structural surface recognition capabilities while exhibiting better adaptability to the data distribution of the target slope.

[0157] The final output is a site-customized model adapted to the specific characteristics of the target slope, as well as a domain-aligned feature representation that is aligned with the source domain in the feature space and can be directly used for structural surface identification.

[0158] The online incremental learning phase corresponds to the fourth level of the multi-scale feature pyramid, namely the instance detail feature level. The goal of this phase is to continuously optimize the model using newly acquired instance data after deployment, preventing performance degradation over time and generating a continuously optimized site model. This phase employs the Elastic Weight Integration algorithm or its variants for training. The Elastic Weight Integration algorithm prevents catastrophic forgetting by estimating the importance of model parameters to the old task and protecting important parameters from excessive modification during parameter updates. First, an importance weight is calculated for each parameter based on the squared gradient estimate of the parameter on the old data; then, during training on new data, parameter updates are restricted to less important directions, penalizing large changes in important parameters through a regularization term.

[0159] In one optional embodiment, the base features are input to a structural surface segmentation head for processing, including:

[0160] The structural surface segmentation head receives the base features and processes them through a point-by-point multilayer perceptron to output a point-by-point structural surface mask. The base features are sequentially processed by the point-by-point multilayer perceptron through linear transformation, batch normalization, activation function and random deactivation operations, and finally the probability of each point belonging to the structural surface is output through the Sigmoid function.

[0161] It should be noted that the above probability values ​​are used to generate a binary structural surface mask after thresholding, which is used to accurately identify the spatial distribution of structural surfaces in the point cloud. This mask, together with the original point cloud, participates in subsequent geological modeling, significantly improving the geometric continuity and topological consistency of tectonic interpretation.

[0162] In an optional embodiment, the basic features are input into the morphology regression head for processing, including:

[0163] After receiving the basic features, the attitude regression head performs local geometric enhancement on the basic features;

[0164] For each point in the base features of local geometric enhancement, search for neighboring points to construct a local point set, and calculate the covariance matrix of the local point set;

[0165] The normal vector features are extracted through the covariance matrix, and the normal vector features are concatenated with the basic features and then input into the attitude regression multilayer perceptron to output the predicted values ​​of dip and tilt angle for each point.

[0166] The predicted dip and tilt angle values ​​are mapped to the ranges of 0 to 360 degrees and 0 to 90 degrees respectively using activation functions, and then optimized using the L2 loss function to obtain the attitude prediction results.

[0167] It should be noted that the attitude regression head achieves accurate prediction of the attitude of structural surfaces through local geometric enhancement and covariance analysis. First, local geometric enhancement is performed on the basic features. For each point, a local point set is constructed by searching for neighboring points and calculating the covariance matrix, thereby extracting the normal vector features. Then, the normal vector features are concatenated with the basic features and input into the attitude regression multilayer perceptron. The attitude regression multilayer perceptron is a fully connected neural network used to regress geological attitude (dip and dip angle) from the concatenated features. It includes an input layer, multiple fully connected layers, and an output layer. Specifically, in this embodiment, the attitude regression multilayer perceptron receives the concatenated normal vector features and basic features through the input layer, then performs progressive dimensionality reduction using multiple fully connected layers. A ReLU activation function is applied after each fully connected layer. Finally, the output layer outputs the predicted dip and dip angle values ​​for each point, which are mapped to effective ranges of 0 to 360 degrees and 0 to 90 degrees respectively through activation functions. Finally, an L2 loss function is used for optimization to obtain attitude prediction results that conform to geological specifications.

[0168] This method makes full use of the local geometric structure information of the point cloud, obtains reliable normal vector estimates through covariance matrix decomposition, and performs regression prediction by combining deep learning features. This not only ensures the geometric rigor of the attitude calculation, but also realizes the direct mapping from the original point cloud to the attitude parameters through end-to-end training, effectively improving the accuracy and robustness of the attitude estimation of the structural surface.

[0169] In an optional embodiment, the base features are input to the structural surface group partitioning head for processing, including:

[0170] The orientation prediction results are converted into three-dimensional unit vectors;

[0171] The three-dimensional unit vector is concatenated with the base features to obtain joint features;

[0172] The optimized site model is used to learn the prototype vectors of each joint group, and the distance between the joint feature and the prototype vector at each point is calculated;

[0173] Based on the distance, the probability of each point belonging to each joint group is output through the Softmax function to complete point-by-point grouping;

[0174] The point-by-point grouping is optimized by using cross-entropy loss and prototype separation regularization term to obtain the joint group division results.

[0175] It should be noted that the joint group classification head achieves intelligent joint group classification through joint modeling of attitude vectors and depth features. First, the attitude prediction results are converted into three-dimensional unit vectors and concatenated with the basic features to form joint features. Then, the prototype vectors of each joint group are learned using an optimized site model. The distance between the joint features and the prototype vectors at each point is calculated. The probability distribution of each point belonging to each joint group is output through the Softmax function. Cross-entropy loss and prototype separation regularization terms are used for optimization to ensure intra-group clustering and inter-group separability, ultimately obtaining joint group classification results with clear geological significance.

[0176] This method deeply integrates geometric orientation information with deep semantic features, and explicitly models the distribution characteristics of joint groups through a prototype learning mechanism. It not only respects the classic concept of "dominant orientation" in geology, but also utilizes the powerful feature learning capabilities of neural networks to achieve automatic grouping in complex scenarios. It effectively solves the instability problem of traditional clustering methods under noise interference and cross-group conditions, and provides a reliable quantitative basis for the systematic analysis of rock mass structural surfaces.

[0177] Furthermore, based on the structural surface identification results, attitude prediction results, joint group division results, and multi-scale feature pyramid, three types of uncertainty are calculated, including:

[0178] Cognitive uncertainty calculation: Using the Monte Carlo Dropout method or the deep ensemble model method, multiple model instances are obtained through multiple forward propagations or parallel runs to obtain multiple prediction results for the same input. The variance of the prediction results is calculated as cognitive uncertainty, which reflects the degree of unfamiliarity of the model with the new geological model.

[0179] Random uncertainty calculation: In step S3, each task head output layer constructs a data noise estimation branch in parallel, and outputs the prediction mean and prediction variance for each prediction value. The prediction variance is used as random uncertainty, which reflects the degree of noise interference in the data acquisition process.

[0180] Domain uncertainty calculation: Using the output of the domain discriminator in step S2, the entropy value of the domain classification probability is calculated as the domain uncertainty, which reflects the degree of deviation of the input sample from the training distribution.

[0181] A weighted fusion mechanism is established to calculate the comprehensive uncertainty. Cognitive uncertainty, accidental uncertainty, and domain uncertainty are linearly weighted and summed to obtain the comprehensive uncertainty index.

[0182] When the overall uncertainty is lower than the preset threshold, the prediction result is determined to have high confidence, and the structural surface identification result, attitude prediction result, and joint group division result output in step S3 are directly applied to engineering practice; when the overall uncertainty is higher than or equal to the preset threshold, the prediction result is determined to have low confidence, triggering the active learning mechanism, and pushing the sample to the expert annotation interface to request manual review.

[0183] Under the active learning trigger condition, a multi-indicator fusion strategy is adopted to screen high-value samples: the comprehensive prediction entropy value is used to screen the most confusing samples of the model, the peak distance of the feature space density is used to screen the representative samples, and the domain uncertainty is used to screen the samples on the edge of the distribution. The screened high-value samples are pushed to experts for fine annotation.

[0184] High-confidence prediction results are directly fed into the engineering output process for slope stability analysis and support design. Low-confidence samples are labeled by experts to generate corrected data, which flows to the fourth stage of the online incremental learning module in step S2. The site-customized model is optimized through real-time gradient updates, and an elastic weight consolidation method is used to prevent catastrophic forgetting of learned knowledge. At the same time, high-value samples selected by active learning are labeled and flow to the first stage of the pre-training dataset in step S2 to supplement and enhance the training data of the global basic model, so as to realize the continuous evolution of the model and knowledge accumulation.

[0185] This invention is achieved through the above method:

[0186] 1. Cross-regional generalization ability: Through four-level progressive domain adaptation and feature decoupling, the model does not need to be reconstructed for each new region and can quickly adapt to different geological environments with only a small number of samples.

[0187] 2. Significantly reduced data costs: Relying on meta-learning to achieve rapid learning capabilities, the requirement for new slope annotation samples has been reduced from hundreds to a dozen, reducing field operation costs by more than 90%.

[0188] 3. The prediction results are physically interpretable: Geomechanical principles are explicitly embedded into the loss function, so that the identification results automatically meet the physical rationality constraints such as attitude distribution, conjugate angle theory and topological relationship.

[0189] 4. Quantifiable reliability assessment: By establishing a three-dimensional uncertainty system encompassing cognition, chance, and domain, the model can self-know its confidence level, providing a reliable basis for engineering decisions.

[0190] 5. Continuous knowledge evolution and accumulation: Through federated learning, secure sharing of multi-engineering experience is achieved, and combined with a lifelong learning mechanism, the model becomes more accurate with use, gradually building an industry-level intelligent recognition system.

[0191] This embodiment uses a steep rock slope project in Nujiang as an example to illustrate the improvement of this method. The slope is located in a high mountain canyon area on the western edge of the Hengduan Mountains. The slope lithology is Yanshanian granite, with an altitude of 3505 meters, an overall slope of 52 degrees, and is in a slightly weathered state. The main challenges faced by the project include: the harsh working environment of the plateau makes on-site geological investigation difficult; the slope has a local slope of more than 80 degrees, forming unstable rock masses, which are difficult to implement with traditional manual measurement; and the need to quickly assess the risk of instability, slippage, and collapse of the slope surface rock mass.

[0192] The input data includes a drone-based close-up photogrammetric point cloud containing over 6 million points, as well as metadata describing the geological background of the slope. The metadata explicitly identifies the lithology as Yanshanian granite, the altitude as 3505 meters, the slope as 52 degrees, and the degree of weak weathering. This information serves as prior knowledge of domain-specific characteristics input into the network.

[0193] Feature decoupling and multi-scale feature extraction:

[0194] The feature decoupling network explicitly separates the input features into two subspaces: domain-invariant features and domain-specific features. Domain-invariant features encode the essential geological properties of the structural surface, including its planar geometry, statistical regularities of attitude, and fracture network topology. These features are universal across slopes in different geological regions. Domain-specific features encode the slope's unique properties, including plateau frost heave weathering characteristics, unloading zone development patterns, and noise characteristics from close-up photography by the DJI Air3 drone. These features are closely related to specific sites and data acquisition conditions.

[0195] The multi-scale feature pyramid constructs four levels of feature representation. The globally applicable geological feature level characterizes the overall structural pattern of the slope, corresponding to the macroscopic geomorphological feature of "towering mountains on both sides." The regional semantic feature level characterizes the geological background of the western edge tectonic domain of the Hengduan Mountains, covering the tectonic stress field and lithological combination patterns of the high mountain and canyon areas of the Three Rivers Basin. The site-specific feature level characterizes the local features of specific slope sections beside excavated roads, focusing on the development of unstable rock masses with local slopes exceeding 80 degrees. The instance detail feature level characterizes the point cloud details of a single 5-meter close-up photograph, achieving precise identification of cracks down to the 5-centimeter level.

[0196] Progressive domain adaptation training:

[0197] The global pre-training phase utilizes a multi-regional hybrid dataset, covering various geological types including granite slopes in the Hengduan Mountains, limestone slopes in the Southwest Fold Belt, loess slopes in Northwest China, and red bed slopes in South China. This phase learns the essential geological laws governing structural surfaces, including planar geometric properties, mechanical genetic classification, and statistical laws governing attitude distribution. During training, Fisher distribution constraints are embedded to ensure that the clustering of joint attitudes conforms to spherical statistical laws; power-law distribution constraints are also embedded to ensure that the relationship between structural surface trace length and spacing conforms to the scale law. Through this phase, the model has pre-learned the general laws of the three joint systems of the granite body, such as the theoretical intersection angle of X-type conjugate joints, rather than simply memorizing specific values ​​such as the 169-degree dip angle and 75-degree dip angle in this specific case.

[0198] The regional meta-learning stage employs a variant of the model-independent meta-learning algorithm to establish a rapid adaptation mechanism. A regional prototype library is constructed, containing representative features from regions such as the Hengduan Mountains, the Qinghai-Tibet Plateau, and the Sichuan Basin. When facing the new slope of the Nujiang River, the model automatically matches the "high mountain and canyon of the Hengduan Mountains" prototype, using similarity metrics to accelerate the convergence process. This stage requires only 10 to 20 labeled photos to complete regional adaptation, significantly reducing the workload of on-site geologists in the harsh environment of the plateau and adapting to actual engineering needs.

[0199] In the site adaptation phase, an adversarial training framework is used to align the source and target domains. The source domain data consists of labeled historical slopes from multiple regions, covering different lithologies and climate zones; the target domain is the new slope at the Nujiang Bridge site, which is unlabeled or weakly labeled. The adversarial training involves a game-theoretic optimization between the feature extractor and the domain discriminator: the feature extractor extracts domain-invariant features, including the planarity and attitude parameters of the structural surfaces; the domain discriminator attempts to distinguish the feature source domains; the optimization objective is to minimize the task loss while maximizing the domain discrimination loss. This mechanism forces the feature extractor to ignore domain-specific interferences such as surface roughness caused by plateau physical weathering, focusing instead on the essential geometric properties of the structural surfaces.

[0200] The online incremental learning phase continuously optimizes the model after deployment, uses an elastic weight integration algorithm to prevent catastrophic forgetting, and adapts to newly collected data through timely gradient updates to ensure long-term stable operation of the model.

[0201] Occurrence prediction and geological constraint verification:

[0202] The predicted attitude of the structural planes outputs three dominant joint systems: Group G1 with an attitude of 169 degrees and a dip of 75 degrees, Group G2 with an attitude of 285 degrees and a dip of 43 degrees, and Group G3 with an attitude of 82 degrees and a dip of 62 degrees. The prediction results are validated by three geological constraints: the attitude consistency constraint uses Fisher distribution to measure the KL divergence between the predicted attitude and the theoretical distribution; the conjugate joint constraint verifies whether the joint angles of the joint groups conform to the theoretical values ​​based on the Anderson mechanical model; and the scale law constraint verifies the power-law relationship between the structural plane trace length and spacing. These constraints ensure that the prediction results not only conform to data-driven principles but also satisfy geomechanical principles.

[0203] The final output enhances the traditional structural surface pole isodensity map into an overlay visualization of an uncertainty heatmap. Green areas represent high-confidence model predictions, from which structural surface information can be automatically extracted; yellow areas represent medium uncertainty, requiring random checks for verification; red areas represent high uncertainty or out-of-domain samples, necessitating detailed manual investigation. In engineering decision-making, red areas directly correspond to high-risk locations where "slope surface rock mass instability, slippage, and collapse may occur," providing crucial information for slope stability assessment and engineering protection measure design. This method achieves a deep integration of artificial intelligence prediction and geological expert experience, improving both efficiency and safety.

[0204] Embodiment 2 of the present invention provides a generalized identification system for slope rock mass structural surfaces based on multi-level domain adaptation, such as... Figure 2 As shown, the slope rock mass structure surface generalization identification system based on multi-level domain adaptation includes:

[0205] The multi-scale feature module is used to measure the outcrop rock mass, acquire slope point cloud data, and perform multi-level feature decomposition and representation learning on the slope point cloud data to obtain a multi-scale feature pyramid.

[0206] The progressive domain adaptation module is used to progressively adapt the multi-scale feature pyramid to generate an optimized site model and a domain-aligned feature representation.

[0207] The multi-task joint optimization module is used to perform multi-task joint optimization on the domain-aligned feature representation based on a multi-task learning framework with geological knowledge embedding, to obtain structural surface identification results, attitude prediction results, and joint group division results. The multi-task joint optimization includes: extracting basic features from the domain-aligned feature representation using a shared feature encoder, inputting the basic features into three task branches—structural surface segmentation head, attitude regression head, and structural surface group division head—for processing, and performing geological knowledge constraint verification and multi-task joint optimization on the processed data.

[0208] The feedback module is used to analyze the structural surface identification results, the attitude prediction results, and the joint group division results through uncertainty quantification and active learning mechanisms, generate slope rock mass identification results, and feed the slope rock mass identification results back to the progressive domain adaptation module for data supplementation.

[0209] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A generalized identification method for slope rock mass structural surfaces based on multi-level domain adaptation, characterized in that, Includes the following steps: Step S1: Measure the outcrop rock mass to obtain slope point cloud data, and perform multi-level feature decomposition and representation learning on the slope point cloud data to obtain a multi-scale feature pyramid; wherein, the multi-level feature decomposition and representation learning on the slope point cloud data includes: Rock mass structural features at different levels of abstraction were extracted from the slope point cloud data; The extracted rock mass structural surface features are explicitly decomposed into domain-invariant features and domain-specific features using a feature decoupling network; A multi-scale feature pyramid is used to hierarchically fuse domain-invariant features and domain-specific features, outputting a multi-scale pyramid; the hierarchical fusion includes: The first level is used to learn the basic geometric morphology, topological relationship and mechanical genesis of the structural surfaces in domain-invariant features and domain-specific features, and to generate the overall structural features of the slope. The second level is used to learn discriminative features from domain-invariant features and domain-specific features, and to generate construct domain features; The third level is used to learn the lithology, weathering, and unloading characteristics at the slope scale in both domain-invariant and domain-specific features, and to generate specific slope segment features. The fourth level is used to learn instance detail features from domain-invariant and domain-specific features to generate point cloud detail features; Step S2: Perform progressive domain adaptation training on the multi-scale feature pyramid to generate an optimized site model and domain-aligned feature representations; wherein, performing progressive domain adaptation training on the multi-scale feature pyramid includes: A multi-task objective function, including structural surface segmentation loss, geological rationality constraint loss, and comparative learning loss, is used to pre-train the overall structural features of the slope to generate a general geological basic model. Regional meta-learning is performed on the general geological model and the tectonic domain features to generate regional adaptive initialization parameters; Unlabeled data are extracted from the slope point cloud data. Adversarial training is performed using the region adaptation initialization parameters and the construction domain features. Domain alignment is performed during adversarial training to generate a basic site model and a domain-aligned feature representation. The basic site model is optimized using the point cloud detail features to generate an optimized site model; Through a hierarchical training strategy, the general geological basic model first learns common knowledge across geological regions at the macro level, then learns discriminative features of specific tectonic domains at the regional level, then achieves domain alignment through adversarial training at the site level, and finally performs continuous optimization at the instance level. The progressive training process corresponds to the four-level structure of the multi-scale feature pyramid, ensuring that the feature representation at each level is fully optimized and utilized. Step S3: Based on the multi-task learning framework with geological knowledge embedding, perform multi-task joint optimization on the domain-aligned feature representation to obtain the structural surface identification result, attitude prediction result, and joint group division result; wherein, the multi-task joint optimization includes: extracting basic features from the domain-aligned feature representation using a shared feature encoder, inputting the basic features into three task branches—structural surface segmentation head, attitude regression head, and structural surface group division head—for processing, and performing geological knowledge constraint verification and multi-task joint optimization on the processed data; Step S4: Analyze the structural surface identification results, the attitude prediction results, and the joint group division results through uncertainty quantification and active learning mechanisms to generate slope rock mass identification results, and feed the slope rock mass identification results back to step S2 for data supplementation.

2. The method for generalized identification of slope rock mass structural surfaces based on multi-level domain adaptation according to claim 1, characterized in that, The feature decoupling network includes: a shared geometric coding module, an attitude constraint coding module, a topological graph neural network, a mechanical origin discrimination branch, and an adversarial decoupling dual-branch encoder; Specifically, a feature decoupling network is used to explicitly decompose the extracted rock mass structural surface features into domain-invariant features and domain-specific features, including: The shared geometric coding module is used to extract the basic geometric features from the rock mass structural surface features; wherein, the shared geometric coding module includes a first-level geometric abstraction layer, a second-level geometric abstraction layer and a third-level geometric abstraction layer, and the first-level geometric abstraction layer, the second-level geometric abstraction layer and the third-level geometric abstraction layer are respectively used to extract local features of different dimensions and sizes from the rock mass structural surface features; The basic geometric features are input in parallel into the attitude constraint encoding module and the topology graph neural network to obtain attitude perception features and topology features; The basic geometric features, the attitude perception features, and the topological relationship features are concatenated and then input in parallel to the mechanical origin discrimination branch and the adversarial decoupling dual-branch encoder. The mechanical origin discrimination branch outputs mechanical pattern features, and the adversarial decoupling dual-branch encoder outputs domain-invariant features and domain-specific features. The mechanical pattern features serve as a supervision signal to guide the feature decoupling process of the adversarial decoupling dual-branch encoder.

3. The method for generalized identification of slope rock mass structural surfaces based on multi-level domain adaptation according to claim 2, characterized in that, The orientation constraint encoding module includes: a normal vector-orientation converter, a spherical position encoder, and an orientation consistency constraint layer; Inputting the basic geometric features into the attitude constraint encoding module includes: Extract local features of the second-level geometric abstraction layer and corresponding point normal vector estimates from the basic geometric features; The normal vector estimate is converted into geological attitude parameters using the normal vector-attitude converter; wherein, the dip angle and strike of the normal vector estimate are calculated, the normal vector-attitude converter determines two candidate attitudes using the dip angle and strike, determines the final attitude from the two candidate attitudes through local neighborhood consensus voting, and converts the final attitude into geological attitude parameters; The geological attitude parameters are mapped to a high-dimensional feature space by a spherical position encoder, and the attitude consistency constraint layer is used to perform feature constraints in the high-dimensional feature space to generate attitude perception features. The attitude-aware features are concatenated with the global features of the basic geometric features to generate an attitude-enhanced set of features.

4. The method for generalized identification of slope rock mass structural surfaces based on multi-level domain adaptation according to claim 1, characterized in that, Regional meta-learning is performed on the aforementioned general geological model and the aforementioned tectonic domain features, including: Using the parameters of the general geological model as initial parameters, the tectonic domain features are processed by a tectonic domain encoder to generate parameter modulation vectors. The initial parameters are then subjected to an affine transformation using these parameter modulation vectors to obtain tectonic domain-aware parameters. Generating parameter modulation vectors from the tectonic domain features using the tectonic domain encoder includes: randomly sampling from the current tectonic domain features as support set samples; performing dimensionality reduction on the support set samples using two fully connected layers in the tectonic domain encoder to generate embedding vectors; aggregating the embedding vectors; and then performing dimensionality increase on the aggregated embedding vectors using two fully connected layers in the tectonic domain encoder to generate parameter modulation vectors. The gradients of the constructed domain perception parameters are updated sequentially in the inner and outer loops to obtain the meta-learning adaptation parameters. The region prototype vector of the constructed domain is constructed using the meta-learning adaptation parameters; The meta-learning adaptation parameters and the region prototype vector are iterated to generate region adaptation initialization parameters.

5. The method for generalized identification of slope rock mass structural surfaces based on multi-level domain adaptation according to claim 1, characterized in that, The base features are input into the structural surface segmentation head for processing, including: The structural surface segmentation head receives the base features and processes them through a point-by-point multilayer perceptron to output a point-by-point structural surface mask. The base features are sequentially processed by the point-by-point multilayer perceptron through linear transformation, batch normalization, activation function and random deactivation operations, and finally the probability of each point belonging to the structural surface is output through the Sigmoid function.

6. The method for generalized identification of slope rock mass structural surfaces based on multi-level domain adaptation according to claim 5, characterized in that, The basic features are input into the attitude regression head for processing, including: After receiving the basic features, the attitude regression head performs local geometric enhancement on the basic features; For each point in the base features of local geometric enhancement, search for neighboring points to construct a local point set, and calculate the covariance matrix of the local point set; Normal vector features are extracted through the covariance matrix. The normal vector features are concatenated with the basic features and then input into the attitude regression multilayer perceptron. The attitude regression multilayer perceptron receives the concatenated normal vector features and basic features through the input layer, and then performs stepwise dimensionality reduction through multiple fully connected layers. A ReLU activation function is connected after each fully connected layer. Finally, the output layer outputs the predicted values ​​of dip and tilt angle for each point. The predicted dip and tilt angle values ​​are mapped to the ranges of 0 to 360 degrees and 0 to 90 degrees respectively using activation functions, and then optimized using the L2 loss function to obtain the attitude prediction results.

7. The method for generalized identification of slope rock mass structural surfaces based on multi-level domain adaptation according to claim 6, characterized in that, The base features are input into the structure surface group partitioning head for processing, including: The orientation prediction results are converted into three-dimensional unit vectors; The three-dimensional unit vector is concatenated with the base features to obtain joint features; The optimized site model is used to learn the prototype vectors of each joint group, and the distance between the joint feature and the prototype vector at each point is calculated; Based on the distance, the probability of each point belonging to each joint group is output through the Softmax function to complete point-by-point grouping; The point-by-point grouping is optimized by using cross-entropy loss and prototype separation regularization term to obtain the joint group division results.

8. A slope rock mass structural surface generalization identification system based on multi-level domain adaptation, used to implement the slope rock mass structural surface generalization identification method based on multi-level domain adaptation as described in any one of claims 1 to 7, characterized in that, A generalized identification system for slope rock mass structure surfaces, including: A multi-scale feature module is used to measure the outcrop rock mass, acquire slope point cloud data, and perform multi-level feature decomposition and representation learning on the slope point cloud data to obtain a multi-scale feature pyramid; wherein, the multi-level feature decomposition and representation learning on the slope point cloud data includes: Rock mass structural features at different levels of abstraction were extracted from the slope point cloud data; The extracted rock mass structural surface features are explicitly decomposed into domain-invariant features and domain-specific features using a feature decoupling network; A multi-scale feature pyramid is used to hierarchically fuse domain-invariant features and domain-specific features, outputting a multi-scale pyramid; the hierarchical fusion includes: The first level is used to learn the basic geometric morphology, topological relationship and mechanical genesis of the structural surfaces in domain-invariant features and domain-specific features, and to generate the overall structural features of the slope. The second level is used to learn discriminative features from domain-invariant features and domain-specific features, and to generate construct domain features; The third level is used to learn the lithology, weathering, and unloading characteristics at the slope scale in both domain-invariant and domain-specific features, and to generate specific slope segment features. The fourth level is used to learn instance detail features from domain-invariant and domain-specific features to generate point cloud detail features; A progressive domain adaptation module is used to progressively adapt the multi-scale feature pyramid to generate an optimized site model and a domain-aligned feature representation; wherein, progressively adapting the multi-scale feature pyramid to the multi-scale feature pyramid includes: A multi-task objective function, including structural surface segmentation loss, geological rationality constraint loss, and comparative learning loss, is used to pre-train the overall structural features of the slope to generate a general geological basic model. Regional meta-learning is performed on the general geological model and the tectonic domain features to generate regional adaptive initialization parameters; Unlabeled data are extracted from the slope point cloud data. Adversarial training is performed using the region adaptation initialization parameters and the construction domain features. Domain alignment is performed during adversarial training to generate a basic site model and a domain-aligned feature representation. The basic site model is optimized using the point cloud detail features to generate an optimized site model; Through a hierarchical training strategy, the general geological basic model first learns common knowledge across geological regions at the macro level, then learns discriminative features of specific tectonic domains at the regional level, then achieves domain alignment through adversarial training at the site level, and finally performs continuous optimization at the instance level. The progressive training process corresponds to the four-level structure of the multi-scale feature pyramid, ensuring that the feature representation at each level is fully optimized and utilized. The multi-task joint optimization module is used to perform multi-task joint optimization on the domain-aligned feature representation based on a multi-task learning framework with geological knowledge embedding, to obtain structural surface identification results, attitude prediction results, and joint group division results. The multi-task joint optimization includes: extracting basic features from the domain-aligned feature representation using a shared feature encoder, inputting the basic features into three task branches—structural surface segmentation head, attitude regression head, and structural surface group division head—for processing, and performing geological knowledge constraint verification and multi-task joint optimization on the processed data. The feedback module is used to analyze the structural surface identification results, the attitude prediction results, and the joint group division results through uncertainty quantification and active learning mechanisms, generate slope rock mass identification results, and feed the slope rock mass identification results back to the progressive domain adaptation module for data supplementation.

Images