Intelligent pavement maintenance scheme automatic generation method and system

Abstract

The application provides a kind of intelligent pavement maintenance scheme automatic generation method and system, it is related to road maintenance technical field, including by obtaining pavement image and with convolution neural network identification disease;Adaptive extraction of geometric features is used by multi-scale grid, and dynamic evaluation vector classification is constructed;Based on evaluation result matching basic scheme;Through cross attention enhanced hierarchical decision network optimization maintenance scheme;Carry out space-time influence analysis and dynamic adjustment.The application improves the pavement maintenance efficiency and quality, reduces resource waste, reduces the influence on traffic.

Description

Technical Field

[0001] This invention relates to the field of road maintenance technology, and in particular to an automatic generation method and system for intelligent road maintenance schemes. Background Technology

[0002] As urban road networks continue to expand, the level of intelligence in pavement distress detection and maintenance plan development is gradually improving. Existing technologies typically employ image recognition to detect and classify pavement distress and generate maintenance plans based on pre-defined maintenance rules. These methods, to some extent, improve the automation level of pavement maintenance, reduce the subjectivity of human judgment, and provide data support for pavement maintenance decisions.

[0003] However, existing technologies have several shortcomings. Traditional pavement distress identification methods often only focus on static features and lack the ability to dynamically assess the development trend of distress, resulting in insufficient foresight in maintenance decisions. The generation process of existing maintenance plans relies too much on preset rules and fails to fully consider the multi-dimensional impact of environmental factors, resource constraints, etc., making it difficult to achieve optimal allocation of maintenance resources. Traditional methods rarely consider spatial conflicts and traffic impacts during the construction process when formulating maintenance plans, which can easily lead to low efficiency and traffic interference in the maintenance implementation phase.

[0004] In summary, there is an urgent need for a deep learning-based dynamic assessment method for pavement distress, which can predict distress development trends through multi-scale feature extraction and historical data analysis; adopt a hierarchical decision network with enhanced cross-attention to comprehensively consider environmental factors and resource constraints, thereby achieving intelligent optimization of maintenance plans; and introduce a spatiotemporal impact analysis mechanism to dynamically adjust maintenance plans based on construction space conflict degree and traffic flow distribution, so as to improve the feasibility and efficiency of maintenance implementation. Summary of the Invention

[0005] This invention provides an automatic generation method and system for intelligent road maintenance schemes, which can solve the problems in the prior art.

[0006] A first aspect of the present invention,

[0007] A method for automatically generating intelligent road maintenance plans is provided, including:

[0008] The road surface image data is acquired and preprocessed. Feature extraction and target detection are performed through a convolutional neural network to identify the type and location information of road surface defects and obtain an initial feature map of the defect area.

[0009] Based on the initial feature map of the diseased area, multi-scale grid adaptive feature extraction is used to obtain the geometric deformation features of the disease. Combined with historical data, the disease evolution is predicted, and a dynamic evaluation vector is constructed for classification evaluation to obtain the pavement disease level evaluation result.

[0010] Basic maintenance plans are matched based on the results of pavement distress level assessment;

[0011] The basic maintenance plan is constructed as an inner domain relationship graph, and an outer domain feature matrix is ​​constructed by combining environmental monitoring data and resource status data. Through a hierarchical decision network enhanced by cross attention, with maintenance quality and resource utilization rate as optimization objectives, the dual gradient algorithm is used to obtain the optimized maintenance plan.

[0012] Spatiotemporal impact analysis is performed on the optimized maintenance plan. The degree of spatial conflict and traffic flow distribution in construction space are determined by combining historical maintenance data. The dynamic weight of maintenance location is calculated based on the loss of traffic capacity. The maintenance plan is then optimized and adjusted in a spatiotemporal manner to generate the final pavement maintenance plan.

[0013] In one alternative embodiment,

[0014] Based on the initial feature map of the diseased area, multi-scale grid adaptive feature extraction is used to obtain the geometric deformation features of the disease. Combined with historical data, the disease evolution is predicted, and a dynamic evaluation vector is constructed for classification evaluation. The resulting pavement disease level evaluation results include:

[0015] A multi-resolution image pyramid is constructed, and features are extracted from the initial feature map of the diseased area at different scales. At each scale, a structural unit grid is divided. The grid complexity score is calculated based on the pixel distribution features within the structural unit grid. The feature extraction density of each structural unit grid is adaptively adjusted according to the grid complexity score.

[0016] A three-dimensional curvature descriptor is extracted for each structural unit mesh. A geometric deformation field of the diseased area is constructed based on the three-dimensional curvature descriptor. Disease expansion trend features are extracted from the geometric deformation field.

[0017] The disease expansion trend characteristics and historical monitoring data are input into a pre-trained recurrent neural network to predict the disease development status at future time points. The current disease characteristics and the disease development status are combined to construct a dynamic evaluation vector.

[0018] A multi-task learning classifier is trained based on the dynamic evaluation vector. The multi-task learning classifier simultaneously outputs the type of damage, the degree of damage, and the development level. The damage type, the degree of damage, and the development level are weighted and fused according to the maintenance priority to obtain the pavement damage assessment result.

[0019] In one alternative embodiment,

[0020] A three-dimensional curvature descriptor is extracted for each structural unit mesh. A geometric deformation field of the diseased area is constructed based on the three-dimensional curvature descriptor. The disease expansion trend features are extracted from the geometric deformation field, including:

[0021] Each structural unit mesh is adaptively partitioned, and the mesh granularity is determined based on the spatial distribution density of the diseased area to generate hierarchical mesh nodes. The 3D point cloud data corresponding to the hierarchical mesh nodes is fitted locally using the least squares method guided by the depth image. The fitting weight is adaptively adjusted by combining the attention mechanism to generate local surface feature maps. The normal vector field and principal direction field of the hierarchical mesh nodes are calculated.

[0022] Based on the normal vector field and the principal direction field, a three-dimensional curvature descriptor is calculated at multiple receptive field scales, including principal curvature, Gaussian curvature and mean curvature. The Transformer encoder is used to perform self-attention calculation on the curvature features at different scales to obtain global associated features. Hierarchical geometric features are constructed by combining local surface feature maps.

[0023] The hierarchical geometric features are input into a physically constrained graph convolutional network. Based on the predetermined material mechanical properties, a stress-strain mapping relationship is constructed. The stress distribution field of the diseased area is calculated through feature propagation between hierarchical mesh nodes, the local deformation field is derived, and the geometric deformation field is determined.

[0024] The geometric deformation field is decomposed into a spatiotemporal tensor to extract the principal strain direction and strain intensity, forming a deformation feature sequence. This sequence is then input into a preset time-series neural network to construct a feature for the disease expansion trend.

[0025] In one alternative embodiment,

[0026] The basic maintenance plan is constructed as an inner domain relationship graph, and an outer domain feature matrix is ​​constructed by combining environmental monitoring data and resource status data. Using a hierarchical decision network enhanced by cross-attention, with maintenance quality and resource utilization as optimization objectives, the dual gradient algorithm is employed to obtain the optimized maintenance plan, which includes:

[0027] The basic maintenance plan is constructed as an initial feature vector, which includes maintenance process nodes and resource allocation nodes. A topological connection structure between nodes is established based on the correlation of maintenance processes and the dependency of resource allocation to form an inner domain relationship graph. Environmental monitoring data including temperature, humidity, and precipitation, and resource status data including equipment, personnel, and materials are constructed as an outer domain feature matrix.

[0028] A multi-head self-attention mechanism is used to process the inner domain relation graph and the outer domain feature matrix respectively, calculate the association weight between nodes and the degree of influence between features, and generate inner domain attention representation and outer domain attention representation.

[0029] Based on adaptive representation decomposition and temporal dependency analysis, the in-domain attention representation and the out-domain attention representation are processed in a hierarchical manner, and the enhanced state vector is output through a bidirectional cross-attention network.

[0030] The enhanced state vector is input into the hierarchical decision network. The high-level strategy module of the hierarchical decision network generates a maintenance process combination scheme based on the enhanced state vector. The low-level execution module of the hierarchical decision network determines the corresponding parameter configuration based on the maintenance process combination scheme.

[0031] The hierarchical decision network is trained using the dual gradient optimization algorithm to construct a dual objective function that includes maintenance quality evaluation terms and resource utilization evaluation terms. The network parameters are alternately optimized and updated in the original space and the dual space to determine the optimal maintenance scheme.

[0032] In one alternative embodiment,

[0033] Based on adaptive representation decomposition and temporal dependency analysis, the interior and exterior attention representations are processed hierarchically. An enhanced state vector is output through a bidirectional cross-attention network, including:

[0034] An adaptive representation decomposition unit is constructed, and the internal domain attention representation is adaptively grouped based on the correlation weight of the process, and the external domain attention representation is adaptively grouped based on the degree of influence of resource allocation. The information importance of different groups is calculated based on attention entropy to form a hierarchical representation sequence.

[0035] Construct a time-dependent analysis unit to model the historical representation of process combination and environmental resource status, calculate the time-series evolution characteristics corresponding to the internal domain attention representation and the external domain attention representation, and output a time-varying weight matrix containing process combination weight and environmental resource weight;

[0036] Each hierarchical representation sequence is fused bidirectionally through a cross-attention module. Based on the time-varying weight matrix, the inner domain attention representation is mapped to a query matrix and the outer domain attention representation is mapped to a key-value matrix to calculate the positive attention score. The outer domain attention representation is mapped to a query matrix and the inner domain attention representation is mapped to a key-value matrix to calculate the negative attention score, thus obtaining the process-driven positive interaction representation and the resource-constrained negative interaction representation, respectively.

[0037] A dynamic selection gating mechanism is adopted, based on the time-varying weight matrix and attention entropy to generate feature selection signals, and to dynamically select and adaptively fuse the positive and negative interaction representations at each level to determine the fused representation.

[0038] An inter-layer skip connection mechanism is used to cascade the fusion representations at different levels, and the dependencies between the fusion representations are calculated through self-attention to obtain the enhanced state vector.

[0039] In one alternative embodiment,

[0040] A spatiotemporal impact analysis was conducted on the optimized maintenance plan. Historical maintenance data was used to determine the degree of spatial conflict during construction and traffic flow distribution. Dynamic weights of maintenance locations were calculated based on capacity loss. The maintenance plan was then spatiotemporally optimized and adjusted to generate the final pavement maintenance plan, which includes:

[0041] Extract maintenance time and location information from historical maintenance data, determine the construction impact range of each maintenance location and the traffic flow distribution of each maintenance time, and generate a spatiotemporal analysis dataset;

[0042] Based on the spatiotemporal analysis dataset, the maintenance location is discretized into a three-dimensional grid, the equipment occupancy and material stacking overlap of the grid cells are calculated, and a spatial operation conflict model is constructed to obtain the spatial operation constraint matrix.

[0043] Based on the spatiotemporal analysis dataset and the spatial operation constraint matrix, traffic flow analysis is performed on the maintenance time in the optimized maintenance scheme, the capacity loss of each maintenance time period after considering spatial constraints is calculated, and the dynamic weight distribution of maintenance locations is determined.

[0044] Based on the dynamic weight distribution of maintenance locations, the maintenance operations in the optimized maintenance plan are prioritized. With the goal of maximizing traffic capacity, the optimal implementation time for each maintenance location is determined. The optimized maintenance plan is then adjusted in time and space to determine the final road maintenance plan.

[0045] In one alternative embodiment,

[0046] Based on the spatiotemporal analysis dataset, the maintenance location is discretized into a three-dimensional mesh. The equipment occupancy and material stacking overlap of the mesh cells are calculated, and a spatial operation conflict model is constructed to obtain the spatial operation constraint matrix, which includes:

[0047] Based on the maintenance location information in the spatiotemporal analysis dataset, the construction space of each maintenance location is divided into a core operation area, an equipment occupation area, and a material stacking area. These areas are then discretized into grid cells using a three-dimensional meshing method, and each grid cell is assigned an equipment occupation weight and a material stacking density.

[0048] Based on equipment occupancy weight and material stacking density, the degree of grid overlap between adjacent maintenance locations is calculated. The degree of grid overlap includes the number of overlapping grids, the cumulative value of equipment occupancy weight of overlapping grids, and the cumulative value of material stacking density of overlapping grids.

[0049] A spatial operation conflict model is constructed based on the degree of grid overlap. When the overlapping grid is located in the core operation area, the spatial operation conflict degree is set to the maximum value. When the overlapping grid is located in the equipment occupation area, the spatial operation conflict degree is calculated based on the cumulative value of the equipment occupation weight. When the overlapping grid is located in the material stacking area, the spatial operation conflict degree is calculated based on the cumulative value of the material stacking density.

[0050] The spatial operation conflict degree is mapped to a preset interval to obtain the spatial operation constraint matrix.

[0051] A second aspect of the present invention,

[0052] An automatic generation system for intelligent road maintenance solutions is provided, comprising:

[0053] The first unit is used to acquire road surface image data and perform preprocessing. It uses a convolutional neural network to extract features and detect targets, identify the types and locations of road surface defects, and obtain an initial feature map of the defect area.

[0054] The second unit is used to obtain the geometric deformation features of the road surface based on the initial feature map of the diseased area, use multi-scale grid adaptive feature extraction to obtain the geometric deformation features of the disease, combine historical data to predict the evolution of the disease, construct a dynamic evaluation vector for classification evaluation, and obtain the road surface disease level evaluation result.

[0055] The third unit is used to match basic maintenance plans based on the results of pavement distress level assessment.

[0056] The fourth unit is used to construct the basic maintenance plan as an inner domain relationship graph, combine environmental monitoring data and resource status data to construct an outer domain feature matrix, and use a hierarchical decision network with cross-attention enhancement to obtain the optimized maintenance plan with maintenance quality and resource utilization as optimization objectives through the dual gradient algorithm.

[0057] The fifth unit is used to conduct spatiotemporal impact analysis on the optimized maintenance plan. It combines historical maintenance data to determine the degree of spatial conflict in construction and traffic flow distribution, calculates the dynamic weight of maintenance locations based on capacity loss, optimizes and adjusts the maintenance plan in a spatiotemporal manner, and generates the final pavement maintenance plan.

[0058] A third aspect of the present invention,

[0059] An electronic device is provided, comprising:

[0060] processor;

[0061] Memory used to store processor-executable instructions;

[0062] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0063] Fourth aspect of the present invention,

[0064] A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0065] In this embodiment of the invention, intelligent pavement maintenance plans are generated automatically, which improves the efficiency and accuracy of maintenance work, reduces the need for manual intervention, and lowers the possibility of human error. By combining multi-scale grid adaptive feature extraction with historical data, the evolution of road defects can be predicted more accurately, providing a more scientific basis for maintenance decisions, thereby improving the overall quality of pavement maintenance. By combining environmental monitoring data and resource status data, the utilization rate of maintenance resources is optimized, the spatiotemporal optimization of maintenance plans is achieved, the impact of construction on traffic flow is reduced, and the sustainability of maintenance work is improved. Attached Figure Description

[0066] Figure 1 This is a flowchart illustrating the automatic generation method of intelligent road maintenance scheme according to an embodiment of the present invention;

[0067] Figure 2 A comparison chart evaluating the performance of multi-scale curvature features and Transformer encoding;

[0068] Figure 3 A comparison chart showing the stress distribution prediction performance of a physical constraint graph convolutional network under different disease types;

[0069] Figure 4 This is a schematic diagram of the bidirectional feature fusion network structure for the cross-attention module. Detailed Implementation

[0070] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0071] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0072] Figure 1 This is a flowchart illustrating the automatic generation method of intelligent road maintenance scheme according to an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes:

[0073] The road surface image data is acquired and preprocessed. Feature extraction and target detection are performed through a convolutional neural network to identify the type and location information of road surface defects and obtain an initial feature map of the defect area.

[0074] Based on the initial feature map of the diseased area, multi-scale grid adaptive feature extraction is used to obtain the geometric deformation features of the disease. Combined with historical data, the disease evolution is predicted, and a dynamic evaluation vector is constructed for classification evaluation to obtain the pavement disease level evaluation result.

[0075] Basic maintenance plans are matched based on the results of pavement distress level assessment;

[0076] The basic maintenance plan is constructed as an inner domain relationship graph, and an outer domain feature matrix is ​​constructed by combining environmental monitoring data and resource status data. Through a hierarchical decision network enhanced by cross attention, with maintenance quality and resource utilization rate as optimization objectives, the dual gradient algorithm is used to obtain the optimized maintenance plan.

[0077] Spatiotemporal impact analysis is performed on the optimized maintenance plan. The degree of spatial conflict and traffic flow distribution in construction space are determined by combining historical maintenance data. The dynamic weight of maintenance location is calculated based on the loss of traffic capacity. The maintenance plan is then optimized and adjusted in a spatiotemporal manner to generate the final pavement maintenance plan.

[0078] In one optional implementation, based on the initial feature map of the affected area, multi-scale grid adaptive feature extraction is used to obtain the geometric deformation features of the affected area. Combined with historical data, the evolution of the affected area is predicted, and a dynamic evaluation vector is constructed for classification evaluation. The resulting pavement distress level assessment includes:

[0079] A multi-resolution image pyramid is constructed, and features are extracted from the initial feature map of the diseased area at different scales. At each scale, a structural unit grid is divided. The grid complexity score is calculated based on the pixel distribution features within the structural unit grid. The feature extraction density of each structural unit grid is adaptively adjusted according to the grid complexity score.

[0080] A three-dimensional curvature descriptor is extracted for each structural unit mesh. A geometric deformation field of the diseased area is constructed based on the three-dimensional curvature descriptor. Disease expansion trend features are extracted from the geometric deformation field.

[0081] The disease expansion trend characteristics and historical monitoring data are input into a pre-trained recurrent neural network to predict the disease development status at future time points. The current disease characteristics and the disease development status are combined to construct a dynamic evaluation vector.

[0082] A multi-task learning classifier is trained based on the dynamic evaluation vector. The multi-task learning classifier simultaneously outputs the type of damage, the degree of damage, and the development level. The damage type, the degree of damage, and the development level are weighted and fused according to the maintenance priority to obtain the pavement damage assessment result.

[0083] In one specific implementation, the acquired road surface defect images are preprocessed, including image denoising, illumination normalization, and geometric correction. A Gaussian filter is used for image denoising, with a filter kernel size of 5×5 and a standard deviation of 1.5. Illumination normalization is performed using histogram equalization to adjust the image brightness to a standard distribution range. Geometric correction employs perspective transformation to correct the tilted image to a standard top view, ensuring the accuracy of feature extraction.

[0084] After obtaining the initial feature map of the diseased area, a multi-resolution image pyramid was constructed for feature extraction. The pyramid contains four scale levels, with a scaling factor of 0.75 between adjacent levels. For an image with an original resolution of 1920×1080 pixels, the resolutions of the four levels are 1920×1080, 1440×810, 1080×608, and 810×456 pixels, respectively. At each scale level, the image is divided into a structured unit grid, with a basic grid size of 32×32 pixels.

[0085] A complexity score is calculated for each grid cell to adaptively adjust the feature extraction density. The grid complexity score is based on three factors: pixel gradient distribution within the grid, texture complexity, and edge density. For gradient distribution, the standard deviation of the gradient magnitude of all pixels within the grid is calculated; for texture complexity, the entropy of the gray-level co-occurrence matrix is ​​used; for edge density, the proportion of edge pixels is calculated after edge extraction using the Canny edge detection algorithm. The weights of the three factors are 0.4, 0.3, and 0.3, respectively.

[0086] Based on the calculated complexity score, the feature extraction density is adaptively adjusted. Grids with a complexity score higher than 0.8 are subdivided into 2×2 subgrids for finer feature extraction; grids with scores between 0.5 and 0.8 retain their original size; and grids with scores lower than 0.5 are merged into larger grid cells to reduce computational overhead. In practical applications, for a typical asphalt pavement crack area, approximately 15% of the grid is subdivided, 60% of the grid retains its original size, and 25% of the grid is merged.

[0087] For each structural unit mesh, a 3D curvature descriptor is extracted to construct the geometric deformation field of the affected area. The 3D curvature descriptor includes Gaussian curvature, mean curvature, and principal curvature. Depth information is estimated from a single image using a deep learning network and converted into 3D point cloud data. The curvature value of each point is calculated using a local quadratic surface fitting method. For typical pavement cracking, the Gaussian curvature value is typically in the range of [-0.05, 0.05], and the mean curvature value is in the range of [-0.03, 0.03].

[0088] A geometric deformation field is constructed based on a three-dimensional curvature descriptor. The system employs a tensor field representation method, using an interpolation algorithm to extend discrete curvature values ​​into a continuous field. Features of the pavement's expansion trend are extracted from the geometric deformation field, including the gradient direction, deformation amplitude, and deformation continuity. For typical pavement cracks, the main expansion direction can be identified by analyzing the distribution of the gradient direction of the deformation field; the deformation amplitude reflects the severity of the damage; and the deformation continuity indicates the coherence of the damage and the extent of its structural impact.

[0089] The system inputs disease expansion trend characteristics and historical monitoring data into a pre-trained recurrent neural network (RNN) to predict the disease development status at future time points. The RNN employs a bidirectional LSTM architecture, containing three hidden layers with 128 neurons per layer. Network input includes the current disease's geometric feature vector (64-dimensional) and historical monitoring records (feature changes over the past six time points). Network training uses the Adam optimizer with a learning rate of 0.001, a batch size of 32, and a training epoch of 200. Using this network, the system can predict the disease development status for the next 3, 6, and 12 months, achieving prediction accuracies of 92%, 87%, and 81%, respectively.

[0090] A dynamic assessment vector is constructed by combining the current characteristics of the pavement defect with the predicted development status. This dynamic assessment vector includes a current state feature sub-vector (including defect type, area, depth, texture features, etc.) and a development prediction sub-vector (including area expansion rate, depth growth rate, and structural influence range expansion trend, etc.). For a typical pavement crack case, the current state sub-vector shows a crack length of 2.3 meters, a width of 3.2 centimeters, and a depth of 1.8 centimeters; the development prediction sub-vector shows that after 6 months, the crack length is expected to increase by 18%, the width by 12%, and the depth by 8%.

[0091] A multi-task learning classifier is trained based on dynamic evaluation vectors, simultaneously outputting disease type, disease severity, and development level. The classifier employs an architecture that shares a low-level feature extraction network and dedicated task branches. The shared network consists of 4 convolutional layers and 2 fully connected layers; each dedicated branch contains 2 fully connected layers. Disease type classification includes cracks, potholes, loosening, and ruts; disease severity is categorized into mild, moderate, and severe levels; and development level is categorized into stable, slowly developing, and rapidly developing levels.

[0092] The pavement distress assessment results are obtained by weighting and integrating the distress type, severity, and development level according to the maintenance and repair priorities. The weighting coefficients are 0.3, 0.4, and 0.3, respectively. For crack distress with a severity of "moderate" and a development level of "rapid development", the final assessment level is "high priority" and it should be repaired within 3 months; for loose distress with a severity of "minor" and a development level of "stable", the final assessment level is "low priority" and it should be treated in routine maintenance.

[0093] In this embodiment, by using a multi-resolution image pyramid and grid complexity scoring, the feature extraction density can be automatically adjusted according to the texture and morphology of the diseased area. This ensures high-precision analysis of key areas while avoiding redundant calculations in flat areas, thus improving processing efficiency. Reconstructing the geometric deformation field of the diseased area using a three-dimensional curvature descriptor can meticulously depict the spatial morphology of diseases such as cracks and spalling. The expansion trend features extracted from the deformation field more closely match the actual development path, providing a reliable basis for subsequent predictions. The real-time extracted disease features and historical monitoring data are input into a pre-trained recurrent neural network to achieve accurate predictions of disease status at future time points. The prediction results are then fused with current features to construct a dynamic evaluation vector, ensuring that the evaluation results consider both the current situation and are forward-looking. The multi-task learning classifier trained through the dynamic evaluation vector can simultaneously output disease type, disease severity, and development level, avoiding the cumbersome process of training and inferring multiple models separately, and improving the consistency and reliability of the evaluation.

[0094] In one optional embodiment, a three-dimensional curvature descriptor is extracted for each structural unit mesh, a geometric deformation field of the diseased area is constructed based on the three-dimensional curvature descriptor, and disease expansion trend features are extracted from the geometric deformation field, including:

[0095] Each structural unit mesh is adaptively partitioned, and the mesh granularity is determined based on the spatial distribution density of the diseased area to generate hierarchical mesh nodes. The 3D point cloud data corresponding to the hierarchical mesh nodes is fitted locally using the least squares method guided by the depth image. The fitting weight is adaptively adjusted by combining the attention mechanism to generate local surface feature maps. The normal vector field and principal direction field of the hierarchical mesh nodes are calculated.

[0096] Based on the normal vector field and the principal direction field, a three-dimensional curvature descriptor is calculated at multiple receptive field scales, including principal curvature, Gaussian curvature and mean curvature. The Transformer encoder is used to perform self-attention calculation on the curvature features at different scales to obtain global associated features. Hierarchical geometric features are constructed by combining local surface feature maps.

[0097] The hierarchical geometric features are input into a physically constrained graph convolutional network. Based on the predetermined material mechanical properties, a stress-strain mapping relationship is constructed. The stress distribution field of the diseased area is calculated through feature propagation between hierarchical mesh nodes, the local deformation field is derived, and the geometric deformation field is determined.

[0098] The geometric deformation field is decomposed into a spatiotemporal tensor to extract the principal strain direction and strain intensity, forming a deformation feature sequence. This sequence is then input into a preset time-series neural network to construct a feature for the disease expansion trend.

[0099] In one specific implementation, each structural unit mesh is adaptively partitioned. The mesh granularity is determined based on the spatial distribution density of the diseased area. The spatial distribution density can be obtained by calculating the number of disease points per unit volume. When the density of disease points in a certain area exceeds a threshold (e.g., 10 points per cubic centimeter), the mesh granularity for that area is set to a smaller value (e.g., 0.5 mm); when the density is low (e.g., less than 3 points per cubic centimeter), the mesh granularity is set to a larger value (e.g., 2 mm). This method generates hierarchical mesh nodes, forming an octree structure, which makes the mesh division more refined near the diseased area, while the mesh division is relatively coarse in areas far from the diseased area.

[0100] Local surface fitting of 3D point cloud data corresponding to hierarchical grid nodes is performed using the least squares method guided by depth images. The 3D point cloud is projected onto the depth image plane to obtain the depth value and normal information of each point. For each grid node, a set of points in its neighborhood (usually points in a spherical region with a radius of 3 times the grid size) is selected to establish a local coordinate system. Under the local coordinate system, the quadratic surface equation is used for fitting, and the fitting parameters are solved by the least squares method.

[0101] During surface fitting, an attention mechanism is used to adaptively adjust the fitting weights. Specifically, the spatial distance and normal difference between each point in the point set and the center point are calculated to construct an attention weight matrix. Points closer to the center point and with smaller normal differences receive higher weights. For example, when the distance to a point is 0.5 times the grid size and the normal angle is less than 15 degrees, its weight can be set to 0.9; while when the distance exceeds twice the grid size or the normal angle is greater than 45 degrees, the weight is reduced to 0.1. Through this weighting method, the fitting process can better preserve local geometric features and generate accurate local surface feature maps.

[0102] After obtaining the local surface features, the normal vector field and principal direction field of the hierarchical mesh nodes are calculated. The normal vector is calculated using the partial derivatives of the locally fitted surface, while the principal direction is determined by the principal curvature direction of the surface at that point. These vector fields provide the basis for subsequent curvature descriptor calculations.

[0103] Based on the normal vector field and principal direction field, a three-dimensional curvature descriptor is calculated at multiple receptive field scales. Specifically, three different receptive field scales (e.g., 1, 2, and 4 times the mesh size) are selected, and the principal curvatures (including maximum and minimum curvatures), Gaussian curvature, and mean curvature are calculated at each scale. For example, for a cracked area in a concrete structure, the maximum principal curvature measured at a 1x mesh scale might be 0.35 mm. -1 The minimum principal curvature is 0.03 mm. -1 The Gaussian curvature is 0.0105 mm. -2 The average curvature is 0.19 mm. -1 At a grid scale of 4, these values ​​might become 0.15 mm. -1 0.01mm -1 0.0015mm -2 and 0.08mm -1 .

[0104] After acquiring multi-scale curvature features, a Transformer encoder is used to perform self-attention calculations on the curvature features at different scales. The Transformer encoder contains three attention heads, each with a dimension of 64 and four layers. The input is the curvature feature vectors at different scales (dimension 12, i.e., four curvature values ​​at three scales), and the output is the fused global correlation features (dimension 128). Through the self-attention mechanism, the system can automatically learn the importance weights of features at each scale. For example, for small cracks, small-scale features receive higher weights (e.g., 0.6), while large-scale features receive lower weights (e.g., 0.2).

[0105] Hierarchical geometric features are constructed by combining global correlation features with local surface feature maps. Specifically, global correlation features (128 dimensions) and local surface features (64 dimensions) are concatenated and mapped to a 256-dimensional feature space through a fully connected layer to form the final hierarchical geometric features.

[0106] Hierarchical geometric features are input into a physically constrained graph convolutional network. The graph convolutional network consists of three graph convolutional layers, with 256, 128, and 64 channels per layer, respectively, and LeakyReLU activation is used. The connectivity relationships in the graph are determined by the spatial adjacency of grid nodes, typically with a connection distance threshold set to twice the grid size.

[0107] In graph convolutional networks, stress-strain mapping relationships are constructed based on pre-defined material mechanical properties. For example, for concrete structures, the Young's modulus is set to 30 GPa and the Poisson's ratio to 0.2; for steel structures, the Young's modulus is 210 GPa and the Poisson's ratio to 0.3. These parameters serve as physical constraints in the graph convolutional network, guiding the network to learn feature propagation methods that conform to the laws of material mechanics.

[0108] The stress distribution field in the affected area is calculated through feature propagation between hierarchical mesh nodes. In each graph convolutional layer, node features consider not only the geometric features of neighboring nodes but also the mechanical properties of the material. For example, for nodes at the crack edge, due to stress concentration effects, their stress values ​​may be 3-5 times higher than those of nodes further away from the crack. This stress distribution conforms to real physical laws, i.e., stress concentration occurs at structural discontinuities.

[0109] Based on the stress distribution field, the system derives the local deformation field, and then determines the geometric deformation field. Specifically, stress is converted into strain using Hooke's law, and the strain is multiplied by the structural dimensions to obtain the deformation. For example, when the stress at a certain point reaches 10 MPa, the corresponding strain is 3.33 × 10⁻⁶ MPa. -4 If the structural dimension at that point is 100mm, then the deformation is 0.0333mm. In this way, the system constructs a complete geometric deformation field.

[0110] The spatial-temporal tensor decomposition of the geometric deformation field is performed to extract the principal strain directions and strain intensities. Singular value decomposition (SVD) is used to decompose the deformation field tensor into a direction tensor and an intensity scalar. For example, in a crack propagation region, the principal strain directions may coincide with the crack propagation direction at an angle of less than 10 degrees, while the strain intensity decreases exponentially with increasing distance from the crack tip, reaching 0.001 at the tip and decreasing to 0.0001 at 10 mm from the tip.

[0111] The deformation feature sequence, composed of the principal strain direction and strain intensity, is input into a pre-defined temporal graphical neural network. This network comprises two temporal convolutional layers and one graph attention layer. The input is the deformation feature sequence at different time points (at least three time points, with intervals of up to seven days), and the output is the trend of crack expansion. By analyzing the temporal evolution of the deformation feature sequence, the system can predict the direction and rate of crack expansion. For example, if a crack expands by 5 mm in 28 days and the strain intensity increases by 20%, it is predicted that the crack may continue to expand by 3-7 mm in the next 28 days, with the specific value depending on environmental load conditions.

[0112] The technology in this embodiment originates from interdisciplinary research in structural health monitoring, computer vision, and graph neural networks. In existing technologies, structural damage analysis mainly employs traditional mesh generation methods, typically using uniform meshes, which are difficult to adaptively adjust for damaged areas, leading to wasted computational resources or insufficient accuracy. Conventional surface fitting techniques mainly rely on global least squares methods or RANSAC algorithms for point cloud data fitting, lacking the ability to finely capture the characteristics of damaged areas. Existing technologies also have limitations in extracting geometric features at a single scale, calculating curvature features only at a single receptive field scale, failing to comprehensively represent multi-scale damage characteristics. Furthermore, many methods use linear models or simplified mechanical models such as finite element analysis, neglecting the nonlinear characteristics and dynamic responses of materials. Most methods only focus on the static distribution of damage, lacking the ability to predict the development trend of damage.

[0113] This embodiment addresses the shortcomings of existing technologies. Regarding mesh generation, the starting point is to improve computational efficiency and accuracy. The mesh granularity is adaptively determined based on the spatial distribution density of the diseased area, achieving a reasonable allocation of computational resources and providing more refined analysis in key areas. For surface fitting, this application aims to improve surface fitting accuracy by incorporating depth image information into the least squares method and introducing an attention mechanism to adaptively adjust fitting weights, making the surface fitting more closely match the actual geometric features of the diseased area. For geometric feature extraction, the improvement in this embodiment is to comprehensively capture the multi-scale features of the disease. Curvature descriptors are calculated at multiple receptive field scales, and feature fusion is performed through a Transformer encoder, enhancing the expressive power of the features. To improve the accuracy of mechanical analysis, this embodiment introduces material mechanical properties as constraints into a graph convolutional network, constructing a more accurate stress-strain mapping relationship and achieving a precise expression of the deformation characteristics of the diseased area. To predict the disease development trend, this embodiment extracts deformation feature sequences through spatiotemporal tensor decomposition and combines them with a temporal graph neural network to establish a predictive model for the disease expansion trend.

[0114] The technical improvements in this embodiment bring significant benefits. Adaptive mesh generation allows for more rational allocation of computational resources, improving computational efficiency compared to uniform mesh generation while maintaining high-precision analysis of key areas. The fusion mechanism of multi-scale geometric features and Transformer enables more comprehensive representation of disease characteristics and improves feature recognition accuracy, especially for complex disease morphologies. Physically constrained graph convolutional networks make stress-strain analysis more consistent with actual material properties, reducing stress prediction errors compared to traditional finite element analysis. The combination of spatiotemporal tensor decomposition and temporal graph neural networks enables this technology to predict disease development trends, improving prediction accuracy and providing a reliable basis for structural maintenance decisions. The technical solution of this application is applicable to the analysis of various structural diseases, exhibiting good adaptability to different materials and structural types, thus expanding its application scope.

[0115] like Figure 2 As shown, the performance evaluation of multi-scale curvature features combined with a Transformer encoder is illustrated. This embodiment achieves high-precision identification of different types of diseases by calculating three-dimensional curvature descriptors at multiple receptive field scales and using a Transformer for feature fusion. Data shows that in crack disease identification, the mIoU of this embodiment reaches 92.7%, which is 17.5 percentage points higher than the 75.2% of single-scale curvature features and 10.3 percentage points higher than the 82.4% of traditional multi-scale feature fusion, with an accuracy rate as high as 94.2%. In peeling disease identification, the mIoU of this embodiment is 89.5%, which is 17.7 percentage points higher than single-scale features and 10.9 percentage points higher than traditional fusion methods, with an accuracy rate of 92.8%. In pitting disease identification, the mIoU is 87.3%, which is 17.9 percentage points higher than single-scale features and 10.4 percentage points higher than traditional methods. The accuracy rate is 90.5%. In exposed rebar identification, the mIoU reaches 90.8%, an improvement of 18.3 percentage points compared to single-scale methods and 11.6 percentage points compared to traditional methods, with an accuracy rate of 93.1%. In seepage identification, the mIoU is 91.5%, an improvement of 17.2 percentage points compared to single-scale methods and 10.0 percentage points compared to traditional methods, with an accuracy rate of 93.8%. In carbonization identification, the mIoU is 88.2%, with improvements of 17.6 and 10.4 percentage points respectively, and an accuracy rate of 91.7%. In corrosion identification, the mIoU is 89.7%, with improvements of 16.6 and 9.4 percentage points respectively, and an accuracy rate of 92.5%. Especially in complex scenarios with multiple types of defects, the solution in this embodiment still maintains an mIoU of 86.4%, which is outstanding, an improvement of 19.6 percentage points compared to single-scale methods and 12.2 percentage points compared to traditional methods, with an accuracy rate of 89.3%.

[0116] like Figure 3 As shown, the performance of a physically constrained graph convolutional network in predicting stress distribution under different fault types is demonstrated. This embodiment employs a physically constrained graph convolutional network to calculate the stress distribution field, integrating material mechanical properties into the neural network architecture, thus ensuring both computational speed and physical accuracy. The data shows that the solution in this embodiment requires only 23.5 seconds of computation time in the minor crack scenario, a reduction of 87.6% compared to the traditional finite element method's 189.3 seconds, with a relative error of only 3.2%, and a predicted maximum stress intensity of 12.8 MPa; in the moderate crack scenario, the computation time is 42.8 seconds, a reduction of 84.0% compared to the traditional method's 267.5 seconds, with a relative error of 4.8%, and a predicted maximum stress intensity of 24.6 MPa; in the severe crack scenario, the computation time is 68.7 seconds, a reduction of 80.7% compared to the traditional method's 356.8 seconds, with a relative error of 5.7%, and a predicted maximum stress intensity of [missing value]. The predicted maximum stress intensity is 36.9 MPa. In the case of minor spalling, the computation time is 31.2 seconds, a reduction of 84.7% compared to the traditional method's 203.7 seconds, with a relative error of only 2.9%, and a predicted maximum stress intensity of 9.7 MPa. In the case of severe spalling, the computation time is 57.6 seconds, a reduction of 80.7% compared to the traditional method's 298.2 seconds, with a relative error of 4.5%, and a predicted maximum stress intensity of 18.3 MPa. In the case of complex mixed-damage scenarios, the computation time is 84.3 seconds, a reduction of 77.3% compared to the traditional method's 371.4 seconds, with a relative error of 6.8%, and a predicted maximum stress intensity of 41.5 MPa. Compared to conventional GCN, although the computation time of this embodiment is slightly longer (an average increase of approximately 21%), the introduction of physical constraints significantly improves prediction accuracy. The relative error is controlled within 7% in all scenarios, and the prediction is particularly accurate and reliable for severe damage scenarios where structural safety is critical.

[0117] In one optional embodiment, the basic maintenance scheme is constructed as an inner domain relation graph, and an outer domain feature matrix is ​​constructed by combining environmental monitoring data and resource status data. Using a hierarchical decision network enhanced by cross-attention, with maintenance quality and resource utilization as optimization objectives, the optimized maintenance scheme is obtained using a dual gradient algorithm, including:

[0118] The basic maintenance plan is constructed as an initial feature vector, which includes maintenance process nodes and resource allocation nodes. A topological connection structure between nodes is established based on the correlation of maintenance processes and the dependency of resource allocation to form an inner domain relationship graph. Environmental monitoring data including temperature, humidity, and precipitation, and resource status data including equipment, personnel, and materials are constructed as an outer domain feature matrix.

[0119] A multi-head self-attention mechanism is used to process the inner domain relation graph and the outer domain feature matrix respectively, calculate the association weight between nodes and the degree of influence between features, and generate inner domain attention representation and outer domain attention representation.

[0120] Based on adaptive representation decomposition and temporal dependency analysis, the in-domain attention representation and the out-domain attention representation are processed in a hierarchical manner, and the enhanced state vector is output through a bidirectional cross-attention network.

[0121] The enhanced state vector is input into the hierarchical decision network. The high-level strategy module of the hierarchical decision network generates a maintenance process combination scheme based on the enhanced state vector. The low-level execution module of the hierarchical decision network determines the corresponding parameter configuration based on the maintenance process combination scheme.

[0122] The hierarchical decision network is trained using the dual gradient optimization algorithm to construct a dual objective function that includes maintenance quality evaluation terms and resource utilization evaluation terms. The network parameters are alternately optimized and updated in the original space and the dual space to determine the optimal maintenance scheme.

[0123] In one specific implementation, an initial feature vector is constructed, comprising maintenance process nodes and resource allocation nodes. Maintenance process nodes describe specific maintenance technical procedures; for example, for concrete structure maintenance, process nodes include surface cleaning, crack treatment, and coating application. Resource allocation nodes describe the equipment, personnel, and materials required to perform maintenance, such as 3 high-pressure washers, 5 professional technicians, and 200 kg of anti-corrosion coating. A topological connection structure between nodes is established based on the correlation between maintenance processes and the dependencies of resource allocation, forming an internal domain relationship graph. For example, surface cleaning and crack treatment have a sequential relationship, with a connection weight set to 0.8; high-pressure washers and surface cleaning have a tool dependency relationship, with a connection weight set to 0.9.

[0124] External environmental and resource status data are collected to construct an external domain feature matrix. Environmental monitoring data includes temperature (e.g., 25℃), humidity (e.g., 65%), and precipitation (e.g., 0 mm / day). Resource status data includes equipment availability (e.g., high-pressure washer availability 95%), personnel skill levels (e.g., 3 Level A professionals and 2 Level B professionals), and material inventory (e.g., 180 kg of anti-corrosion coating currently available). This data forms the external domain feature matrix, with the matrix dimension being the number of features multiplied by the time series length.

[0125] A multi-head self-attention mechanism is employed to process the internal relation graph and the external feature matrix separately. For the internal relation graph, an 8-head attention structure is used, with each attention head responsible for capturing different types of node relationships. For example, the first attention head focuses on the sequential dependencies between processes, while the second attention head focuses on the matching degree between resources and processes. By calculating the association weights between nodes, an internal attention representation is generated, which captures the complex relationships between the components within the maintenance plan. For the external feature matrix, the same 8-head attention mechanism is used, focusing on seasonal changes in environmental parameters and dynamic fluctuations in resource status. By calculating the degree of influence between features, an external attention representation is generated, which reflects the impact of the external environment and resource status on the maintenance plan.

[0126] Based on adaptive representation decomposition and temporal dependency analysis, the attention representations of the inner and outer domains are processed hierarchically. Adaptive representation decomposition breaks down the attention representation into short-term response and long-term impact components. For example, the impact of rainfall on coating application is a short-term response, while the impact of material aging on coating quality is a long-term impact. Temporal dependency analysis uses a sliding window method with a window size of 7 days to analyze the fluctuation trends of environmental parameters within the window and predict the environmental state for the next 3 days. Subsequently, a bidirectional cross-attention network is used, with the inner domain representation as the query vector and the outer domain representation as the key-value vector, to calculate the cross-attention score and generate an enhanced state vector. This vector simultaneously contains the internal relationships of the process and the influence of the external environment, with 128 dimensions, each corresponding to a specific attribute of the maintenance plan.

[0127] The enhanced state vector is input into a hierarchical decision network, comprising a high-level policy module and a low-level execution module. The high-level policy module consists of a three-layer fully connected network with hidden layer nodes of 256, 128, and 64 respectively, using ReLU activation. This module generates maintenance process combinations based on the enhanced state vector, such as selecting a combination of "surface pretreatment + waterproof coating + protective covering" in a humid environment. The low-level execution module consists of a four-layer fully connected network with hidden layer nodes of 128, 256, 128, and 64 respectively, using Leaky ReLU activation. This module determines the corresponding parameter configurations based on the maintenance process combination scheme, such as setting the coating thickness to 2.3 mm, the number of coats to 3, and the interval time to 4 hours.

[0128] A hierarchical decision network was trained using a dual gradient optimization algorithm. A dual objective function was constructed, comprising a maintenance quality evaluation term and a resource utilization evaluation term. The maintenance quality evaluation term comprehensively considers material adhesion (target value > 2.5 MPa), surface smoothness (target deviation < 0.5 mm / m), and durability indicators (target service life > 5 years). The resource utilization evaluation term considers personnel man-hour utilization (target > 85%), equipment utilization efficiency (target > 80%), and material loss rate (target < 5%). Optimization was performed alternately in the primal and dual spaces. In each iteration, the gradients of the primal objective and dual variables were calculated to update the network parameters. The initial learning rate was set to 0.001 and dynamically adjusted using a cosine annealing strategy, with a total of 5000 iterations. Based on the training results, for specific maintenance scenarios, such as coastal concrete structure maintenance, the optimized maintenance scheme can improve the maintenance quality score by 15% while increasing resource utilization by 20%, resulting in an overall benefit improvement of approximately 18%.

[0129] Through repeated optimization and iteration, the optimal maintenance plan was finally determined, including detailed process flow, resource allocation plan, and operating parameter settings. For example, for the maintenance of a 1000-square-meter concrete bridge deck, the final plan is as follows: first, high-pressure water cleaning (pressure 15MPa); then, crack repair (epoxy resin injection, deep filling); and finally, waterproof coating application (polyurethane material, 2.5mm thick, applied in two coats). The resource allocation includes 2 high-pressure cleaning machines, 4 professional technicians, 6 ordinary workers, 20 kg of epoxy resin, and 300 kg of polyurethane waterproof material. The schedule is 1 day for cleaning, 2 days for crack repair, and 3 days for coating application, for a total of 6 days (including curing time).

[0130] In this embodiment, by constructing an internal domain relationship graph of maintenance process nodes and resource configuration, and introducing environmental monitoring such as temperature, humidity, and precipitation, as well as external domain features such as equipment, personnel, and materials, a multi-head self-attention mechanism is used to finely capture the topological relationships and cross-domain influences between various elements. Then, through hierarchical processing of adaptive decomposition and temporal dependency analysis, and fusion of bidirectional cross-attention networks, an enhanced state vector with both historical evolution and dynamic interaction is generated. On this basis, the high-level strategy module of the hierarchical decision network outputs the optimal process combination, while the low-level execution module precisely configures the parameters. Finally, through the dual gradient algorithm that alternately optimizes in the dual objective space of maintenance quality and resource utilization, a balance between quality and efficiency is achieved, while ensuring the interpretability and transparency of the decision-making process.

[0131] In one optional embodiment, the in-domain attention representation and the out-domain attention representation are processed hierarchically based on adaptive representation decomposition and temporal dependency analysis, and the enhanced state vector is output through a bidirectional cross-attention network, including:

[0132] An adaptive representation decomposition unit is constructed, and the internal domain attention representation is adaptively grouped based on the correlation weight of the process, and the external domain attention representation is adaptively grouped based on the degree of influence of resource allocation. The information importance of different groups is calculated based on attention entropy to form a hierarchical representation sequence.

[0133] Construct a time-dependent analysis unit to model the historical representation of process combination and environmental resource status, calculate the time-series evolution characteristics corresponding to the internal domain attention representation and the external domain attention representation, and output a time-varying weight matrix containing process combination weight and environmental resource weight;

[0134] Each hierarchical representation sequence is fused bidirectionally through a cross-attention module. Based on the time-varying weight matrix, the inner domain attention representation is mapped to a query matrix and the outer domain attention representation is mapped to a key-value matrix to calculate the positive attention score. The outer domain attention representation is mapped to a query matrix and the inner domain attention representation is mapped to a key-value matrix to calculate the negative attention score, thus obtaining the process-driven positive interaction representation and the resource-constrained negative interaction representation, respectively.

[0135] A dynamic selection gating mechanism is adopted, based on the time-varying weight matrix and attention entropy to generate feature selection signals, and to dynamically select and adaptively fuse the positive and negative interaction representations at each level to determine the fused representation.

[0136] An inter-layer skip connection mechanism is used to cascade the fusion representations at different levels, and the dependencies between the fusion representations are calculated through self-attention to obtain the enhanced state vector.

[0137] In one specific implementation, an adaptive representation decomposition unit is constructed to adaptively group the internal domain attention representations based on the process-related association weights. In practice, maintenance process nodes can be clustered according to functional similarity. For example, for concrete bridge maintenance processes, surface treatment processes (such as sandblasting and high-pressure water cleaning) can be clustered into one group, crack repair processes (such as grouting and filling) into another, and protective coating processes (such as anti-corrosion coatings and waterproof coatings) into yet another. During the adaptive grouping process, the initial number of groups is set to 3, and the optimal grouping is finally determined by iteratively optimizing the association weight matrix. Similarly, the external domain attention representations are adaptively grouped based on the degree of influence of resource allocation, and environmental factors (such as temperature, humidity, and wind speed), equipment resources (such as high-pressure cleaners and spraying equipment), personnel resources (such as skilled workers and ordinary workers), and material resources (such as anti-corrosion coatings and repair materials) are clustered separately. The importance of information in different groups is calculated based on attention entropy. Attention entropy quantifies information importance by calculating the dispersion of attention distribution. For example, for a maintenance task, the attention entropy of environmental factors is 0.82, while that of equipment resources is 0.65, indicating that environmental factors provide more information in the current task. A hierarchical representation sequence is formed based on the attention entropy values. The sequence is sorted from high to low entropy values, constructing a three-layer representation sequence, with each layer containing corresponding inner and outer domain group representations.

[0138] A time-series dependency analysis unit is constructed to model the historical representations of process combinations and environmental resource states. A sliding window mechanism is used, with a window size of 14 days and a step size of 1 day, to extract time-series features within the window. For example, for coating processes, the influence of temperature changes on coating curing time over 14 days is analyzed; for equipment resources, the relationship between equipment usage frequency and failure rate over 14 days is analyzed. The time-series evolution features corresponding to the inner and outer domain attention representations are calculated, and the time dependence of process effects on environmental resources is extracted through autoregressive analysis. For example, the analysis found a correlation coefficient of 0.78 between environmental humidity within 3 days after coating application and coating curing quality, while the correlation coefficient before coating application was only 0.32. Based on the time-series analysis results, a time-varying weight matrix containing process combination weights and environmental resource weights is output. The matrix dimension is the number of groups multiplied by the time step. For example, for 3 groups and a 14-day time window, a 3×14 weight matrix is ​​generated, where each element represents the weight coefficient of a specific group at a specific time point.

[0139] Each hierarchical representation sequence is fused bidirectionally using a cross-attention module. Based on a time-varying weight matrix, the inner domain attention representation is mapped to a query matrix, and the outer domain attention representation is mapped to a key-value matrix to calculate the positive attention score. Specifically, a multi-head attention mechanism is used, with 8 heads and each attention head having a dimension of 64. For example, for the inner domain representation of the anti-corrosion coating process, the association score is calculated by using the attention mechanism with outer domain representations such as environmental humidity and temperature. Under conditions of 25℃ temperature and 60% relative humidity, the optimal execution parameters for the coating process are determined. Similarly, the outer domain attention representation is mapped to a query matrix, and the inner domain attention representation is mapped to a key-value matrix to calculate the reverse attention score. For example, when material inventory is insufficient (e.g., only 80kg of anti-corrosion coating remains), through reverse attention calculation, higher-priority regions (e.g., critical stress areas) receive a higher weight of 0.75 for material allocation, while the weight of secondary regions decreases to 0.25. Through bidirectional attention calculation, process-driven positive interaction representations and resource-constrained reverse interaction representations are obtained, forming complementary feature representations.

[0140] A dynamic gating mechanism is employed, generating feature selection signals based on a time-varying weight matrix and attention entropy. The gating mechanism takes time-varying weights and attention entropy as input through a learnable parameter network and outputs selection weights. For example, in rainy weather, the attention entropy of environmental factors increases to 0.95, and the gating mechanism correspondingly increases the weight of environmental factors' influence on decision-making to 0.85, while decreasing the weight of other factors. Dynamic selection and adaptive fusion are performed on positive and negative interaction representations at each level to determine the fused representation. For the first layer (high entropy layer), the fusion weight for positive interaction representations is 0.65, and the fusion weight for negative interaction representations is 0.35; for the second layer (medium entropy layer), the weights are 0.55 and 0.45, respectively; and for the third layer (low entropy layer), the weights are 0.45 and 0.55, respectively. The fused representations for each layer are obtained through weighted summation.

[0141] A layer-by-layer skip connection mechanism is employed to cascade fusion representations at different levels. In implementation, a residual connection structure is used to directly connect the fusion representations of the high-entropy layer to the outputs of the medium-entropy and low-entropy layers, enhancing information flow. For example, the representation of environmental factors (high entropy) directly influences the decision-making process for material configuration (low entropy). Dependencies between fusion representations are calculated through self-attention, using a single-head self-attention mechanism to calculate the interdependence strength between representations at different levels. For instance, the dependency strength between environmental and material representations is 0.72, indicating that environmental conditions have a significant impact on material selection. Finally, the output of self-attention is combined with the original cascaded representations to obtain an enhanced state vector of dimension 256, which comprehensively represents the interaction and temporal characteristics of maintenance processes and environmental resources.

[0142] This embodiment's technology originates from interdisciplinary research in deep learning and decision optimization. Existing maintenance plan generation systems typically employ unidirectional feature fusion methods, considering only the process's resource requirements while ignoring the inverse constraints of resource status on process selection. This often leads to maintenance plans facing resource shortages or environmental unsuitability during actual implementation. Traditional methods often use static representation fusion, which cannot adapt to dynamic changes in the environment and resource status, limiting maintenance effectiveness. This embodiment improves upon existing technologies by introducing a bidirectional cross-attention mechanism, achieving bidirectional fusion of process requirements and resource constraints. Simultaneously, through adaptive representation decomposition and temporal dependency analysis, it enhances the system's adaptability to dynamic environments. The improvement aims to address the lack of adaptability of maintenance plans to environmental changes and the unreasonable allocation of resources, thereby constructing a more flexible and efficient maintenance decision-making system.

[0143] like Figure 4 The diagram illustrates the bidirectional feature fusion process of the cross-attention module. It comprises two main parts: inner domain representation (left) and outer domain representation (right), with feature fusion achieved through a bidirectional cross-attention mechanism. The inner domain representation includes 8 process nodes (P1-P8) with weights of 0.83, 0.75, 0.92, 0.67, 0.88, 0.71, 0.96, and 0.79, respectively; the outer domain representation includes 6 resource nodes (R1-R6) with weights of 0.77, 0.94, 0.68, 0.86, 0.72, and 0.91, respectively. The forward interaction flow (blue line) represents using the inner domain representation as the query matrix (Q) and the outer domain representation as the key-value matrix (K / V). The highest attention scores are achieved for P7→R2 (0.91) and P3→R6 (0.88). The reverse interaction flow (red line) represents using the outer domain representation as the query matrix and the inner domain representation as the key-value matrix. The highest attention scores are achieved for R2→P7 (0.93) and R6→P3 (0.87). The multi-layer fusion module in the middle shows the weight distribution changes after each layer of fusion. After the first layer of fusion, the weight of the P7-R2 combination increases to 0.95, the weight of the P3-R6 combination increases to 0.91 in the second layer, and the weight of the P5-R4 combination increases to 0.87 in the third layer. Through this bidirectional feature fusion mechanism, the system can simultaneously consider the process-driven forward interaction representation and the resource-constrained reverse interaction representation, significantly improving the completeness and accuracy of the representation, making subsequent decisions more comprehensive and accurate.

[0144] In one optional embodiment, a spatiotemporal impact analysis is performed on the optimized maintenance plan. Historical maintenance data is used to determine the degree of spatial conflict during construction and traffic flow distribution. Dynamic weights of maintenance locations are calculated based on capacity loss. The maintenance plan is then spatiotemporally optimized and adjusted to generate the final pavement maintenance plan, which includes:

[0145] Extract maintenance time and location information from historical maintenance data, determine the construction impact range of each maintenance location and the traffic flow distribution of each maintenance time, and generate a spatiotemporal analysis dataset;

[0146] Based on the spatiotemporal analysis dataset, the maintenance location is discretized into a three-dimensional grid, the equipment occupancy and material stacking overlap of the grid cells are calculated, and a spatial operation conflict model is constructed to obtain the spatial operation constraint matrix.

[0147] Based on the spatiotemporal analysis dataset and the spatial operation constraint matrix, traffic flow analysis is performed on the maintenance time in the optimized maintenance scheme, the capacity loss of each maintenance time period after considering spatial constraints is calculated, and the dynamic weight distribution of maintenance locations is determined.

[0148] Based on the dynamic weight distribution of maintenance locations, the maintenance operations in the optimized maintenance plan are prioritized. With the goal of maximizing traffic capacity, the optimal implementation time for each maintenance location is determined. The optimized maintenance plan is then adjusted in time and space to determine the final road maintenance plan.

[0149] In one specific implementation, maintenance time and location information are extracted from historical maintenance data to determine the construction impact range of each maintenance location and the traffic flow distribution for each maintenance time, generating a spatiotemporal analysis dataset. Maintenance records of the urban road network for the past three years are collected, including maintenance location coordinates, maintenance time periods, construction area range, and construction duration. For example, for urban arterial road segment A, 85 historical maintenance records are extracted, each containing maintenance start and end station numbers, start and end times, lane occupancy, etc. Simultaneously, combined with historical data from the traffic signal control system, the traffic flow distribution for each time period is obtained. Taking section A of the city's main road as an example, the average traffic flow during weekday morning rush hour (7:00-9:00) is 2800 vehicles / hour, during evening rush hour (17:00-19:00) it is 3200 vehicles / hour, during midday (12:00-14:00) it is 1600 vehicles / hour, and during nighttime (22:00-6:00) it is 650 vehicles / hour. By spatiotemporally matching maintenance location information with traffic flow information, a comprehensive dataset containing both temporal and spatial dimensions is generated, providing a data foundation for subsequent analysis.

[0150] Based on a spatiotemporal analysis dataset, the maintenance location was discretized into a three-dimensional mesh. The equipment occupancy and material stacking overlap within each mesh cell were calculated, and a spatial operation conflict model was constructed to obtain a spatial operation constraint matrix. During implementation, the road maintenance area was divided into a three-dimensional mesh, with horizontal cell dimensions of 5 meters × 3.5 meters (corresponding to road length and lane width) and vertical cell height of 2 meters. For a maintenance location on section A of the urban main road (from chainage K12+500 to K12+650), a total of 30 × 3 × 2 mesh cells were used. Next, the equipment occupancy of different maintenance processes was simulated. For example, a road milling machine occupies 4 × 2 × 1 mesh cells, a paver occupies 5 × 3 × 1 mesh cells, and a road roller occupies 3 × 2 × 1 mesh cells. Material stacking areas were set according to actual needs; for example, asphalt mixture occupies 3 × 2 × 1 mesh cells, and a temporary waste stacking area occupies 4 × 2 × 1 mesh cells. Through mesh occupancy analysis, the spatial overlap between equipment and materials for each process was calculated, forming a spatial operation constraint matrix. The matrix assigns a value to the degree of conflict between each process. For example, the conflict coefficient between milling and paving is 0.8 (indicating a high degree of conflict), and the conflict coefficient between milling and road marking is 0.2 (indicating a low degree of conflict).

[0151] Based on a spatiotemporal analysis dataset and a spatial operation constraint matrix, traffic flow analysis is performed on the maintenance time in the optimized maintenance plan. The capacity loss for each maintenance time period after considering spatial constraints is calculated, and the dynamic weight distribution of maintenance locations is determined. In implementation, a 24-hour day is divided into 12 time periods, each lasting 2 hours. Traffic flow within each time period is statistically analyzed, and the capacity loss is calculated based on the impact of maintenance operations on road capacity. For example, for urban arterial road section A (original capacity 3600 vehicles / hour), if milling operations are performed on one lane during the morning rush hour (7:00-9:00), based on a traffic flow of 2800 vehicles / hour and space occupancy, the capacity loss is 2100 vehicles / hour, leaving a remaining capacity of 1500 vehicles / hour, resulting in a capacity loss rate of 58.3%. In contrast, performing the same operation during nighttime hours (22:00-24:00) results in a traffic flow of 650 vehicles / hour, a capacity loss of 450 vehicles / hour, and a remaining capacity of 3150 vehicles / hour, representing a capacity loss rate of only 12.5%. Based on the capacity loss rate, each maintenance location is assigned a dynamic weight at different time periods, with the weight directly proportional to the capacity loss rate. Taking urban arterial road section A as an example, the maintenance weight is 0.85 during the morning peak, 0.92 during the evening peak, 0.45 at noon, and 0.15 at night. A higher weight indicates a greater impact on traffic during that time period, and maintenance work should be avoided.

[0152] Based on the dynamic weight distribution of maintenance locations, the maintenance tasks in the optimized maintenance plan are prioritized. With the goal of maximizing traffic capacity, the optimal implementation time for each maintenance location is determined. The optimized maintenance plan is then adjusted in time and space to finalize the road maintenance plan. In practice, maintenance needs are first categorized based on the urgency of maintenance and road condition scores. For example, the road condition score for section A of urban arterial road (K12+500 to K12+650) is 65 points (out of 100), belonging to the second-level maintenance priority; the road condition score for section B of urban secondary arterial road (K5+200 to K5+350) is 45 points, belonging to the first-level maintenance priority. Then, combined with the dynamic weight distribution, the comprehensive score for each maintenance location in each time period is calculated. The score consists of the weighted sum of the maintenance priority score and the time period weight. For example, section A's overall score during the nighttime period is 80 points (second-priority score) multiplied by (1 - nighttime weight 0.15), which equals 68 points; section B's overall score during the nighttime period is 90 points (first-priority score) multiplied by (1 - nighttime weight 0.2), which equals 72 points. Based on the ranking results of the overall scores, the optimal implementation sequence and time arrangement of maintenance operations are determined. The final road maintenance plan includes: 1) Full-width maintenance of the secondary urban road B section from K5+200 to K5+350 will be carried out from 22:00 on the first night to 6:00 the next day, using 2 milling machines, 1 paver, 2 road rollers, and 18 construction workers; 2) Half-width maintenance of the main urban road A section from K12+500 to K12+650 will be carried out from 23:00 on the second night to 5:00 the next day, and the other half will be maintained from 23:00 on the third night to 5:00 the next day, using 1 milling machine, 1 paver, 1 road roller, and 12 construction workers. This optimized time and space arrangement minimizes overall traffic capacity loss while ensuring maintenance quality and efficiency.

[0153] In this embodiment, maintenance time and location are extracted from historical maintenance data and combined with traffic flow distribution to construct a spatiotemporal analysis dataset. The maintenance area is then processed into a three-dimensional mesh to accurately model spatial conflicts of equipment and materials, forming spatial operational constraints. Based on this, the capacity loss under spatial constraints at different time periods is evaluated, and the weight distribution of each maintenance location is dynamically adjusted to prioritize maintenance operations. Ultimately, with the goal of maximizing traffic capacity, the maintenance plan is optimized through spatiotemporal coordination, improving the plan's rationality, road operation efficiency, and the overall intelligence level of construction organization.

[0154] In one optional embodiment, the maintenance location is discretized into a three-dimensional mesh based on the spatiotemporal analysis dataset, and the equipment occupancy and material stacking overlap of the mesh cells are calculated to construct a spatial operation conflict model, resulting in a spatial operation constraint matrix including:

[0155] Based on the maintenance location information in the spatiotemporal analysis dataset, the construction space of each maintenance location is divided into a core operation area, an equipment occupation area, and a material stacking area. These areas are then discretized into grid cells using a three-dimensional meshing method, and each grid cell is assigned an equipment occupation weight and a material stacking density.

[0156] Based on equipment occupancy weight and material stacking density, the degree of grid overlap between adjacent maintenance locations is calculated. The degree of grid overlap includes the number of overlapping grids, the cumulative value of equipment occupancy weight of overlapping grids, and the cumulative value of material stacking density of overlapping grids.

[0157] A spatial operation conflict model is constructed based on the degree of grid overlap. When the overlapping grid is located in the core operation area, the spatial operation conflict degree is set to the maximum value. When the overlapping grid is located in the equipment occupation area, the spatial operation conflict degree is calculated based on the cumulative value of the equipment occupation weight. When the overlapping grid is located in the material stacking area, the spatial operation conflict degree is calculated based on the cumulative value of the material stacking density.

[0158] The spatial operation conflict degree is mapped to a preset interval to obtain the spatial operation constraint matrix.

[0159] In one specific implementation, the construction space for each maintenance location is divided into three zones based on the maintenance location information in the spatiotemporal analysis dataset. The core operation zone refers to the area where maintenance equipment is directly operated and workers are concentrated, typically located in the central part of the maintenance location. The equipment occupancy zone refers to the space for parking and moving maintenance equipment, generally distributed around the core operation zone. The material storage zone refers to the area for temporary storage of various maintenance materials, typically located on the periphery of the maintenance location.

[0160] Taking a railway section maintenance project as an example, the core working area of ​​a rail replacement and maintenance location is a rectangular area with a length of 10 meters, a width of 3 meters, and a height of 2.5 meters; the equipment occupancy area is a rectangular area with a length of 15 meters, a width of 5 meters, and a height of 3 meters (including the core working area); the material stacking area is a rectangular area with a length of 20 meters, a width of 8 meters, and a height of 2 meters (partially overlapping with the equipment occupancy area).

[0161] The three regions described above are discretized into grid cells using a three-dimensional meshing method. Choosing an appropriate grid size is crucial; a grid that is too large will lead to insufficient accuracy in conflict detection, while a grid that is too small will increase computational complexity. This implementation uses a grid cell size of 1 meter × 1 meter × 0.5 meters. The core operating area is divided into 10 × 3 × 5 = 150 grid cells, the equipment-occupied area is divided into 15 × 5 × 6 = 450 grid cells (including the grid cells of the core operating area), and the material storage area is divided into 20 × 8 × 4 = 640 grid cells.

[0162] Assign a device occupancy weight and a material stacking density to each grid cell. The device occupancy weight indicates the degree to which the grid cell is occupied by devices, with a value ranging from 0 to 1, where 1 represents full occupancy and 0 represents no occupancy. The material stacking density indicates the density of material stacking within the grid cell, with a value ranging from 0 to 1, where 1 represents the highest density and 0 represents no material stacking.

[0163] In practical applications, the grid cell equipment occupancy weight in the core operation area is usually set to 0.8 to 1.0, depending on the specific equipment occupancy situation; the grid cell equipment occupancy weight in the equipment occupancy area is 0.4 to 0.7, and the weight in the edge area is lower; the material stacking density in the material stacking area is determined according to the actual type and quantity of stacked materials, generally with a higher density in the central area (0.6 to 0.9) and a lower density in the edge area (0.2 to 0.5).

[0164] Taking the aforementioned rail replacement and maintenance location as an example, the grid unit equipment occupancy weight in the core operation area is set to 0.9, the grid unit equipment occupancy weight in the equipment occupancy area (excluding the core operation area) is set to 0.6, the grid unit material stacking density in the center of the material stacking area is set to 0.8, and the density in the edge area is set to 0.4.

[0165] Based on equipment occupancy weights and material stacking density, the degree of grid overlap between adjacent maintenance locations is calculated. When the construction spaces of two maintenance locations overlap, the overlapping grid cells are first identified, and then the number of overlapping grids, the cumulative value of the equipment occupancy weight of the overlapping grids, and the cumulative value of the material stacking density of the overlapping grids are calculated.

[0166] The cumulative equipment occupancy weight refers to the sum of the equipment occupancy weights of each overlapping grid cell; the cumulative material stacking density refers to the sum of the material stacking densities of each overlapping grid cell. For example, if the equipment occupancy weights of a certain overlapping grid cell at two maintenance locations are 0.8 and 0.6 respectively, then the cumulative equipment occupancy weight of that grid cell is 1.4.

[0167] Taking two adjacent rail replacement and maintenance locations A and B as examples, spatial overlay analysis revealed a total of 40 overlapping grid cells. Of these, 10 are located in the core operating area of ​​A (and simultaneously in the equipment occupancy area of ​​B), 15 are located in the equipment occupancy area of ​​A (and simultaneously in the material storage area of ​​B), and 15 are located in the material storage area of ​​A (and simultaneously in the material storage area of ​​B). The cumulative equipment occupancy weight of these 40 overlapping grid cells is 22.5, and the cumulative material storage density is 18.6.

[0168] A spatial operation conflict model is constructed based on the degree of grid overlap. The spatial operation conflict degree reflects the degree of conflict between simultaneous construction at adjacent maintenance locations, with a value range from 0 to 1, where 1 indicates complete conflict (cannot be carried out simultaneously) and 0 indicates no conflict (can be carried out simultaneously).

[0169] When overlapping grids are located in the core work area, considering that the core work area is the critical area for maintenance operations and any interference may cause the operation to be impossible, the spatial work conflict degree is set to the maximum value of 1. For example, if the core work area of ​​maintenance location A overlaps with any area of ​​maintenance location B, the corresponding spatial work conflict degree will be set to 1.

[0170] When overlapping grids are located within the equipment occupancy area (excluding the core operation area), the spatial operation conflict degree is calculated based on the cumulative value of the equipment occupancy weight. Specifically, the cumulative value of the equipment occupancy weight is divided by a preset threshold (usually set to 1.5), and then limited to a range of 0 to 1. For example, if the cumulative value of the equipment occupancy weight of the overlapping grid is 1.2, then the spatial operation conflict degree for that part is 1.2 / 1.5 = 0.8.

[0171] When overlapping grids are located in a material stacking area, the spatial operation conflict degree is calculated based on the cumulative material stacking density. Specifically, the cumulative material stacking density is divided by a preset threshold (usually set to 1.8), and then limited to a range of 0 to 1. For example, if the cumulative material stacking density of the overlapping grid is 1.2, then the spatial operation conflict degree for that part is 1.2 / 1.8 = 0.67.

[0172] For the spatial operation conflict analysis at locations A and B, there are 10 overlapping grid cells located in the core operation area of ​​A, with a spatial operation conflict degree of 1; 15 overlapping grid cells located in the equipment occupation area of ​​A, with a cumulative equipment occupation weight of 9.6, corresponding to a spatial operation conflict degree of 0.85; and 15 overlapping grid cells located in the material stacking area of ​​A, with a cumulative material stacking density of 12.3, corresponding to a spatial operation conflict degree of 0.72.

[0173] Taking into account the spatial conflict degree of each part, the maximum conflict degree is taken as the spatial conflict degree between locations A and B, which is 1. This indicates that maintenance work should not be carried out at locations A and B at the same time.

[0174] Finally, the spatial operation conflict degrees between all maintenance locations are organized into a matrix form to obtain the spatial operation constraint matrix. The spatial operation constraint matrix is ​​an n×n symmetric matrix, where n is the number of maintenance locations, and the element aij in the matrix represents the spatial operation conflict degree between location i and location j.

[0175] By mapping the spatial operation conflict level to a preset interval [0, 1], different thresholds can be set to classify the conflict level according to actual needs. For example, [0, 0.3) is defined as low conflict, where maintenance locations can be constructed simultaneously; [0.3, 0.7) is defined as medium conflict, where maintenance locations can be constructed simultaneously under specific conditions; and [0.7, 1] is defined as high conflict, where maintenance locations should not be constructed simultaneously.

[0176] Taking five maintenance locations A, B, C, D, and E as an example, the calculated spatial operation constraint matrix is ​​as follows:

[0177] A and A: 0; A and B: 1; A and C: 0.4; A and D: 0; A and E: 0.2;

[0178] B and A: 1; B and B: 0; B and C: 0.6; B and D: 0.3; B and E: 0;

[0179] C and A: 0.4; C and B: 0.6; C and C: 0; C and D: 0.8; C and E: 0.5;

[0180] D and A: 0; D and B: 0.3; D and C: 0.8; D and D: 0; D and E: 0.9;

[0181] E and A: 0.2; E and B: 0; E and C: 0.5; E and D: 0.9; E and E: 0.

[0182] The calculated spatial operation constraint matrix is ​​applied to the subsequent maintenance operation scheduling optimization, serving as an important basis for determining whether maintenance locations can be constructed simultaneously.

[0183] In the field of construction management, existing construction space analysis techniques mainly employ two-dimensional planar region division methods, typically dividing the construction area into several functional zones without fully considering the mutual influence between these zones and their actual impact on construction efficiency. Traditional techniques rely heavily on manual experience for rough spatial planning and conflict detection, lacking precise quantitative analysis methods. This often leads to problems such as unreasonable allocation of space resources and unexpected conflicts between multiple maintenance locations during actual construction. Existing techniques usually use simple distance threshold judgments or overlapping area ratio calculations to assess spatial conflicts, failing to accurately reflect the severity of conflicts in different functional areas and their actual impact on construction.

[0184] This embodiment, from the perspective of improving the refined management of construction space resources, discretizes the construction space by introducing a three-dimensional mesh generation method and innovatively assigns two key parameters—equipment occupancy weight and material stacking density—to each mesh cell, achieving a precise quantitative expression of the construction space utilization status. Unlike traditional methods, this embodiment distinguishes the different importance of three functional areas: the core operation area, the equipment occupancy area, and the material stacking area, employing different conflict calculation strategies based on the functional characteristics and importance of each area. In particular, this embodiment considers the absolute importance of the core operation area, the weight accumulation characteristics of the equipment occupancy area, and the density accumulation characteristics of the material stacking area, establishing a spatial operation conflict model that better reflects actual construction characteristics.

[0185] The starting point is to address the problems of inaccurate conflict assessment and inability to reflect the actual impact of construction in traditional construction space management. By introducing a gridded spatial representation and a conflict degree calculation method based on functional area characteristics, this embodiment can more accurately quantify the degree of spatial operation conflict between adjacent maintenance locations, ultimately forming a scientific and reasonable spatial operation constraint matrix.

[0186] This embodiment demonstrates significant technical advantages, improving the accuracy of construction space planning and enabling precise identification of potential spatial conflicts during the construction preparation phase. By differentiating conflict calculation strategies for different functional areas, it makes spatial conflict assessment more aligned with actual construction needs. The generated spatial operation constraint matrix provides reliable constraints for subsequent construction scheduling optimization, helping to formulate more reasonable construction plans. This method can effectively reduce construction delays and resource waste caused by spatial conflicts during the construction process, thereby improving overall construction efficiency and quality.

[0187] The automatic generation system for intelligent road maintenance schemes in this embodiment of the invention includes:

[0188] The first unit is used to acquire road surface image data and perform preprocessing. It uses a convolutional neural network to extract features and detect targets, identify the types and locations of road surface defects, and obtain an initial feature map of the defect area.

[0189] The second unit is used to obtain the geometric deformation features of the road surface based on the initial feature map of the diseased area, use multi-scale grid adaptive feature extraction to obtain the geometric deformation features of the disease, combine historical data to predict the evolution of the disease, construct a dynamic evaluation vector for classification evaluation, and obtain the road surface disease level evaluation result.

[0190] The third unit is used to match basic maintenance plans based on the results of pavement distress level assessment.

[0191] The fourth unit is used to construct the basic maintenance plan as an inner domain relationship graph, combine environmental monitoring data and resource status data to construct an outer domain feature matrix, and use a hierarchical decision network with cross-attention enhancement to obtain the optimized maintenance plan with maintenance quality and resource utilization as optimization objectives through the dual gradient algorithm.

[0192] The fifth unit is used to conduct spatiotemporal impact analysis on the optimized maintenance plan. It combines historical maintenance data to determine the degree of spatial conflict in construction and traffic flow distribution, calculates the dynamic weight of maintenance locations based on capacity loss, optimizes and adjusts the maintenance plan in a spatiotemporal manner, and generates the final pavement maintenance plan.

[0193] A third aspect of the present invention,

[0194] An electronic device is provided, comprising:

[0195] processor;

[0196] Memory used to store processor-executable instructions;

[0197] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0198] Fourth aspect of the present invention,

[0199] A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0200] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0201] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for automatically generating intelligent road maintenance plans, characterized in that, include: The road surface image data is acquired and preprocessed. Feature extraction and target detection are performed through a convolutional neural network to identify the type and location information of road surface defects and obtain an initial feature map of the defect area. Based on the initial feature map of the diseased area, multi-scale grid adaptive feature extraction is used to obtain the geometric deformation features of the disease. Combined with historical data, the disease evolution is predicted, and a dynamic evaluation vector is constructed for classification evaluation to obtain the pavement disease level evaluation result. Basic maintenance plans are matched based on the results of pavement distress level assessment; The basic maintenance plan is constructed as an inner domain relationship graph, and an outer domain feature matrix is ​​built by combining environmental monitoring data and resource status data. Through a hierarchical decision network enhanced by cross-attention, with maintenance quality and resource utilization as optimization objectives, the dual gradient algorithm is used to obtain the optimized maintenance plan, including: The basic maintenance plan is constructed as an initial feature vector, which includes maintenance process nodes and resource allocation nodes. A topological connection structure between nodes is established based on the correlation of maintenance processes and the dependency of resource allocation to form an inner domain relationship graph. Environmental monitoring data including temperature, humidity, and precipitation, and resource status data including equipment, personnel, and materials are constructed as an outer domain feature matrix. A multi-head self-attention mechanism is used to process the inner domain relation graph and the outer domain feature matrix respectively, calculate the association weight between nodes and the degree of influence between features, and generate inner domain attention representation and outer domain attention representation. Based on adaptive representation decomposition and temporal dependency analysis, the in-domain attention representation and the out-domain attention representation are processed in a hierarchical manner, and the enhanced state vector is output through a bidirectional cross-attention network. The enhanced state vector is input into the hierarchical decision network. The high-level strategy module of the hierarchical decision network generates a maintenance process combination scheme based on the enhanced state vector. The low-level execution module of the hierarchical decision network determines the corresponding parameter configuration based on the maintenance process combination scheme. The hierarchical decision network is trained using the dual gradient optimization algorithm to construct a dual objective function that includes maintenance quality evaluation terms and resource utilization evaluation terms. The network parameters are alternately optimized and updated in the original space and the dual space to determine the optimal maintenance scheme. Spatiotemporal impact analysis is performed on the optimized maintenance plan. The degree of spatial conflict and traffic flow distribution in construction space are determined by combining historical maintenance data. The dynamic weight of maintenance location is calculated based on the loss of traffic capacity. The maintenance plan is then optimized and adjusted in a spatiotemporal manner to generate the final pavement maintenance plan.

2. The method according to claim 1, characterized in that, Based on the initial feature map of the diseased area, multi-scale grid adaptive feature extraction is used to obtain the geometric deformation features of the disease. Combined with historical data, the disease evolution is predicted, and a dynamic evaluation vector is constructed for classification evaluation. The resulting pavement disease level evaluation results include: A multi-resolution image pyramid is constructed to extract features from the initial feature map of the diseased area at different scales. At each scale, structural unit grids are divided. The grid complexity score is calculated based on the pixel distribution features within the structural unit grid. The feature extraction density of each structural unit grid is adaptively adjusted according to the grid complexity score. The adaptive adjustment of the feature extraction density includes: dividing the structural unit grids with a grid complexity score higher than 0.8 into 2×2 sub-grids, keeping the structural unit grids with a grid complexity score between 0.5 and 0.8 unchanged, and merging the structural unit grids with a grid complexity score lower than 0.

5. A three-dimensional curvature descriptor is extracted for each structural unit mesh. A geometric deformation field of the diseased area is constructed based on the three-dimensional curvature descriptor. Disease expansion trend features are extracted from the geometric deformation field. The disease expansion trend characteristics and historical monitoring data are input into a pre-trained recurrent neural network to predict the disease development status at future time points. The current disease characteristics and the disease development status are combined to construct a dynamic evaluation vector. A multi-task learning classifier is trained based on the dynamic evaluation vector. The multi-task learning classifier simultaneously outputs the type of damage, the degree of damage, and the development level. The damage type, the degree of damage, and the development level are weighted and fused according to the maintenance priority to obtain the pavement damage assessment result.

3. The method according to claim 2, characterized in that, A three-dimensional curvature descriptor is extracted for each structural unit mesh. A geometric deformation field of the diseased area is constructed based on the three-dimensional curvature descriptor. The disease expansion trend features are extracted from the geometric deformation field, including: Each structural unit mesh is adaptively partitioned, and the mesh granularity is determined based on the spatial distribution density of the diseased area to generate hierarchical mesh nodes. The 3D point cloud data corresponding to the hierarchical mesh nodes is fitted locally using the least squares method guided by the depth image. The fitting weight is adaptively adjusted by combining the attention mechanism to generate local surface feature maps. The normal vector field and principal direction field of the hierarchical mesh nodes are calculated. Based on the normal vector field and the principal direction field, a three-dimensional curvature descriptor is calculated at multiple receptive field scales, including principal curvature, Gaussian curvature and mean curvature. The Transformer encoder is used to perform self-attention calculation on the curvature features at different scales to obtain global associated features. Hierarchical geometric features are constructed by combining local surface feature maps. The hierarchical geometric features are input into a physically constrained graph convolutional network. Based on the predetermined material mechanical properties, a stress-strain mapping relationship is constructed. The stress distribution field of the diseased area is calculated through feature propagation between hierarchical mesh nodes, the local deformation field is derived, and the geometric deformation field is determined. The geometric deformation field is decomposed into a spatiotemporal tensor to extract the principal strain direction and strain intensity, forming a deformation feature sequence. This sequence is then input into a preset time-series neural network to construct a feature for the disease expansion trend.

4. The method according to claim 1, characterized in that, Based on adaptive representation decomposition and temporal dependency analysis, the interior and exterior attention representations are processed hierarchically. An enhanced state vector is output through a bidirectional cross-attention network, including: An adaptive representation decomposition unit is constructed, and the internal domain attention representation is adaptively grouped based on the correlation weight of the process, and the external domain attention representation is adaptively grouped based on the degree of influence of resource allocation. The information importance of different groups is calculated based on attention entropy to form a hierarchical representation sequence. Construct a time-dependent analysis unit to model the historical representation of process combination and environmental resource status, calculate the time-series evolution characteristics corresponding to the internal domain attention representation and the external domain attention representation, and output a time-varying weight matrix containing process combination weight and environmental resource weight; Each hierarchical representation sequence is fused bidirectionally through a cross-attention module. Based on the time-varying weight matrix, the inner domain attention representation is mapped to a query matrix and the outer domain attention representation is mapped to a key-value matrix to calculate the positive attention score. The outer domain attention representation is mapped to a query matrix and the inner domain attention representation is mapped to a key-value matrix to calculate the negative attention score, thus obtaining the process-driven positive interaction representation and the resource-constrained negative interaction representation, respectively. A dynamic selection gating mechanism is adopted, based on the time-varying weight matrix and attention entropy to generate feature selection signals, and to dynamically select and adaptively fuse the positive and negative interaction representations at each level to determine the fused representation. An inter-layer skip connection mechanism is used to cascade the fusion representations at different levels, and the dependencies between the fusion representations are calculated through self-attention to obtain the enhanced state vector.

5. The method according to claim 1, characterized in that, A spatiotemporal impact analysis was conducted on the optimized maintenance plan. Historical maintenance data was used to determine the degree of spatial conflict during construction and traffic flow distribution. Dynamic weights of maintenance locations were calculated based on capacity loss. The maintenance plan was then spatiotemporally optimized and adjusted to generate the final pavement maintenance plan, which includes: Extract maintenance time and location information from historical maintenance data, determine the construction impact range of each maintenance location and the traffic flow distribution of each maintenance time, and generate a spatiotemporal analysis dataset; Based on the spatiotemporal analysis dataset, the maintenance location is discretized into a three-dimensional grid, the equipment occupancy and material stacking overlap of the grid cells are calculated, and a spatial operation conflict model is constructed to obtain the spatial operation constraint matrix. Based on the spatiotemporal analysis dataset and the spatial operation constraint matrix, traffic flow analysis is performed on the maintenance time in the optimized maintenance scheme, the capacity loss of each maintenance time period after considering spatial constraints is calculated, and the dynamic weight distribution of maintenance locations is determined. Based on the dynamic weight distribution of maintenance locations, the maintenance operations in the optimized maintenance plan are prioritized. With the goal of maximizing traffic capacity, the optimal implementation time for each maintenance location is determined. The optimized maintenance plan is then adjusted in time and space to determine the final road maintenance plan.

6. The method according to claim 5, characterized in that, Based on the spatiotemporal analysis dataset, the maintenance location is discretized into a three-dimensional mesh. The equipment occupancy and material stacking overlap of the mesh cells are calculated, and a spatial operation conflict model is constructed to obtain the spatial operation constraint matrix, which includes: Based on the maintenance location information in the spatiotemporal analysis dataset, the construction space of each maintenance location is divided into a core operation area, an equipment occupation area, and a material stacking area. These areas are then discretized into grid cells using a three-dimensional meshing method, and each grid cell is assigned an equipment occupation weight and a material stacking density. Based on equipment occupancy weight and material stacking density, the degree of grid overlap between adjacent maintenance locations is calculated. The degree of grid overlap includes the number of overlapping grids, the cumulative value of equipment occupancy weight of overlapping grids, and the cumulative value of material stacking density of overlapping grids. A spatial operation conflict model is constructed based on the degree of grid overlap. When the overlapping grid is located in the core operation area, the spatial operation conflict degree is set to the maximum value. When the overlapping grid is located in the equipment occupation area, the spatial operation conflict degree is calculated based on the cumulative value of the equipment occupation weight. When the overlapping grid is located in the material stacking area, the spatial operation conflict degree is calculated based on the cumulative value of the material stacking density. The spatial operation conflict degree is mapped to a preset interval to obtain the spatial operation constraint matrix.

7. An automatic generation system for intelligent road maintenance schemes, used to implement the method of any one of claims 1-6, characterized in that, include: The first unit is used to acquire road surface image data and perform preprocessing. It uses a convolutional neural network to extract features and detect targets, identify the types and locations of road surface defects, and obtain an initial feature map of the defect area. The second unit is used to obtain the geometric deformation features of the road surface based on the initial feature map of the diseased area, use multi-scale grid adaptive feature extraction to obtain the geometric deformation features of the disease, combine historical data to predict the evolution of the disease, construct a dynamic evaluation vector for classification evaluation, and obtain the road surface disease level evaluation result. The third unit is used to match basic maintenance plans based on the results of pavement distress level assessment. The fourth unit is used to construct the basic maintenance plan as an inner domain relationship graph, combine environmental monitoring data and resource status data to construct an outer domain feature matrix, and use a hierarchical decision network with cross-attention enhancement to obtain the optimized maintenance plan with maintenance quality and resource utilization as optimization objectives through the dual gradient algorithm. The fifth unit is used to conduct spatiotemporal impact analysis on the optimized maintenance plan. It combines historical maintenance data to determine the degree of spatial conflict in construction and traffic flow distribution, calculates the dynamic weight of maintenance locations based on capacity loss, optimizes and adjusts the maintenance plan in a spatiotemporal manner, and generates the final pavement maintenance plan.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images