Method and system for automatically generating intelligent pavement maintenance scheme

Abstract

The invention provides an automatic generation method and system for an intelligent pavement maintenance scheme, and relates to the technical field of road maintenance, and the method comprises the steps: obtaining a pavement image, and recognizing a disease through a convolutional neural network; a multi-scale grid is utilized to adaptively extract geometric features, and dynamic evaluation vector classification is constructed; matching a basic scheme based on the evaluation result; a maintenance scheme is optimized through a cross attention enhanced hierarchical decision network; and space-time influence analysis and dynamic adjustment are carried out. The pavement maintenance efficiency and quality are improved, the resource waste is reduced, and the influence on traffic is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of road maintenance, and particularly to a method and system for automatically generating an intelligent road maintenance plan. Background Art

[0002] With the continuous expansion of the urban road network scale, the intelligent level of road surface disease detection and maintenance plan formulation has been gradually improved. In the prior art, image recognition technology is usually used to detect and classify road surface diseases, and a maintenance plan is generated based on preset maintenance rules. These methods have improved the automation degree of road surface maintenance to a certain extent, reduced the subjectivity of manual judgment, and provided data support for road surface maintenance decision-making.

[0003] However, there are several deficiencies in the prior art. Traditional road surface disease recognition methods often only focus on static features and lack the ability to dynamically evaluate the development trend of diseases, resulting in insufficient foresight in maintenance decision-making; the generation process of existing maintenance plans is too dependent on preset rules, and multi-dimensional influences such as environmental factors and resource constraints are not fully considered, making it difficult to achieve the optimal allocation of maintenance resources; traditional methods rarely consider spatial conflicts and traffic impacts during the construction process when formulating maintenance plans, which easily causes low efficiency and traffic interference during the maintenance implementation stage.

[0004] In summary, there is an urgent need for a dynamic evaluation method for road surface diseases based on deep learning, which can predict the development trend of diseases through multi-scale feature extraction and historical data analysis; adopt a hierarchical decision-making network enhanced by cross-attention to comprehensively consider environmental factors and resource constraints to achieve intelligent optimization of the maintenance plan; at the same time, introduce a spatio-temporal impact analysis mechanism to dynamically adjust the maintenance plan based on the construction space conflict degree and traffic flow distribution to improve the feasibility and efficiency of maintenance implementation. Summary of the Invention

[0005] The embodiments of the present invention provide a method and system for automatically generating an intelligent road maintenance plan, which can solve the problems in the prior art.

[0006] In the first aspect of the embodiments of the present invention,

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

[0008] Obtain road surface image data and perform preprocessing, extract features and detect targets through a convolutional neural network, identify the type and location information of road surface diseases, and obtain an initial feature map of the disease area;

[0009] Based on the initial feature map of the disease area, adopt multi-scale grid adaptive feature extraction to obtain the geometric deformation features of the disease, combine historical data to predict the disease evolution, and construct a dynamic evaluation vector for classification evaluation to obtain the evaluation result of the road surface disease level;

[0010] Match the basic maintenance plan based on the evaluation results of pavement disease levels;

[0011] Construct the basic maintenance plan into an internal domain relationship graph, construct an external domain feature matrix by combining environmental monitoring data and resource status data, and through a hierarchical decision-making network enhanced by cross-attention, with the maintenance quality and resource utilization rate as the optimization objectives, use the dual gradient algorithm to obtain the optimized maintenance plan;

[0012] Conduct spatio-temporal impact analysis on the optimized maintenance plan, determine the construction space conflict degree and traffic flow distribution by combining historical maintenance data, calculate the dynamic weight of the maintenance location based on the loss of traffic capacity, and perform spatio-temporal optimization adjustment on the maintenance plan to generate the final pavement maintenance plan.

[0013] In an alternative embodiment,

[0014] Based on the initial feature map of the disease area, use multi-scale grid adaptive feature extraction to obtain the geometric deformation features of the disease, combine historical data to predict the disease evolution, and construct a dynamic evaluation vector for classification evaluation. The evaluation results of pavement disease levels include:

[0015] Construct a multi-resolution image pyramid, extract features of different proportional scales from the initial feature map of the disease area, divide the structural unit grid at each scale, calculate the grid complexity score based on the pixel distribution features within the structural unit grid, and adaptively adjust the feature extraction density of each structural unit grid according to the grid complexity score;

[0016] Extract the three-dimensional curvature descriptor for each structural unit grid, construct the geometric deformation field of the disease area based on the three-dimensional curvature descriptor, and extract the disease expansion trend features from the geometric deformation field;

[0017] Input the disease expansion trend features and historical monitoring data into a pre-trained recurrent neural network to predict the disease development state at future time nodes, and combine the current disease features with the disease development state to construct a dynamic evaluation vector;

[0018] Train a multi-task learning classifier based on the dynamic evaluation vector. The multi-task learning classifier outputs the disease type, disease degree, and development level at the same time, and perform weighted fusion on the disease type, disease degree, and development level according to the maintenance and repair priority to obtain the pavement disease evaluation results.

[0019] In an alternative embodiment,

[0020] Extract the three-dimensional curvature descriptor for each structural unit grid, construct the geometric deformation field of the disease area based on the three-dimensional curvature descriptor, and the disease expansion trend features extracted from the geometric deformation field include:

[0021] Adaptive meshing is performed on each structural unit grid, the meshing granularity is determined based on the spatial distribution density of the disease area, hierarchical grid nodes are generated, local surface fitting is performed on the three-dimensional point cloud data corresponding to the hierarchical grid nodes by the least squares method guided by depth images, the fitting weights are adaptively adjusted in combination with the attention mechanism, a local surface feature map is generated, and the normal vector field and the principal direction field of the hierarchical grid nodes are calculated;

[0022] Based on the normal vector field and the principal direction field, three-dimensional curvature descriptors are calculated at multiple receptive field scales, including principal curvature, Gaussian curvature, and mean curvature. The self-attention calculation of the curvature features at different scales is performed using a Transformer encoder to obtain global correlation features, and hierarchical geometric features are constructed in combination with the local surface feature map;

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

[0024] The geometric deformation field is subjected to spatio-temporal tensor decomposition, the principal strain direction and the strain intensity are extracted, a deformation feature sequence is formed, and it is input into a preset temporal graph neural network to construct a disease spread trend feature.

[0025] In an alternative embodiment,

[0026] The basic maintenance plan is constructed as an inner domain relationship graph, an outer domain feature matrix is constructed by combining environmental monitoring data and resource status data, and through a hierarchical decision network enhanced by cross-attention, with the maintenance quality and resource utilization rate as the optimization objectives, an optimized maintenance plan is obtained using the dual gradient algorithm, including:

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

[0028] The multi-head self-attention mechanism is used to process the inner domain relationship graph and the outer domain feature matrix respectively, calculate the correlation weights between nodes and the influence degree between features, and generate an inner domain attention representation and an outer domain attention representation;

[0029] Based on adaptive representation decomposition and temporal dependence analysis, the inner domain attention representation and the outer domain attention representation are hierarchically processed, and an enhanced state vector is output through a bidirectional cross-attention network;

[0030] Input the enhanced state vector into a hierarchical decision network. The high-level policy module of the hierarchical decision network generates a maintenance process combination plan based on the enhanced state vector, and the low-level execution module of the hierarchical decision network determines the corresponding parameter configuration based on the maintenance process combination plan;

[0031] Train the hierarchical decision network using the dual gradient optimization algorithm, construct a dual objective function including a maintenance quality evaluation item and a resource utilization rate evaluation item, alternately optimize and update the network parameters in the original space and the dual space, and determine the optimized maintenance plan.

[0032] In an alternative embodiment,

[0033] Perform hierarchical processing on the inner-domain attention representation and the outer-domain attention representation based on adaptive representation decomposition and temporal dependence analysis. The enhanced state vector output by the bidirectional cross-attention network includes:

[0034] Construct an adaptive representation decomposition unit, perform adaptive grouping on the inner-domain attention representation based on the process-related weights, perform adaptive grouping on the outer-domain attention representation based on the influence degree of resource allocation, calculate the information importance of different groups based on attention entropy, and form a hierarchical representation sequence;

[0035] Construct a temporal dependence analysis unit, model the historical representations of the process combination and the environmental resource state, calculate the temporal evolution features corresponding to the inner-domain attention representation and the outer-domain attention representation, and output a time-varying weight matrix including the process combination weight and the environmental resource weight;

[0036] Perform bidirectional feature fusion on each hierarchical representation sequence through a cross-attention module. Based on the time-varying weight matrix, map the inner-domain attention representation to a query matrix, map the outer-domain attention representation to a key-value matrix to calculate the forward attention score, map the outer-domain attention representation to a query matrix, map the inner-domain attention representation to a key-value matrix to calculate the reverse attention score, and respectively obtain the forward interaction representation dominated by the process and the reverse interaction representation constrained by the resources;

[0037] Adopt a dynamic selection gating mechanism, generate a feature selection signal based on the time-varying weight matrix and attention entropy, perform dynamic selection and adaptive fusion on the forward interaction representation and the reverse interaction representation at each level, and determine the fused representation;

[0038] Adopt an inter-layer skip connection mechanism to cascade the fused representations at different levels, and calculate the dependence relationship between the fused representations through self-attention to obtain the enhanced state vector.

[0039] In an alternative embodiment,

[0040] Conduct a spatio-temporal impact analysis on the optimized maintenance plan, determine the construction space conflict degree and traffic flow distribution in combination with historical maintenance data, calculate the dynamic weight of the maintenance location based on the loss of traffic capacity, and conduct spatio-temporal optimization and adjustment on the maintenance plan to generate the final pavement maintenance plan, including:

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

[0042] Discretize the maintenance locations into three-dimensional grids based on the spatio-temporal analysis data set, calculate the degree of equipment occupancy and material stacking overlap of the grid cells, and construct a spatial operation conflict model to obtain a spatial operation constraint matrix;

[0043] Conduct traffic flow analysis on the maintenance time in the optimized maintenance plan based on the spatio-temporal analysis data set and the spatial operation constraint matrix, calculate the loss of traffic capacity in each maintenance time period considering spatial constraints, and determine the dynamic weight distribution of the maintenance locations;

[0044] Rank the maintenance operations in the optimized maintenance plan according to the dynamic weight distribution of the maintenance locations, aim at maximizing the traffic capacity, determine the optimal implementation period of each maintenance location, conduct spatio-temporal adjustment on the optimized maintenance plan, and determine the final pavement maintenance plan.

[0045] In an alternative embodiment,

[0046] Discretize the maintenance locations into three-dimensional grids based on the spatio-temporal analysis data set, calculate the degree of equipment occupancy and material stacking overlap of the grid cells, and construct a spatial operation conflict model to obtain a spatial operation constraint matrix, including:

[0047] Divide the construction space of each maintenance location into a core operation area, an equipment occupancy area, and a material stacking area according to the maintenance location information in the spatio-temporal analysis data set, discretize it into grid cells by a three-dimensional grid division method, and assign an equipment occupancy weight and a material stacking density to each grid cell;

[0048] Based on the equipment occupancy weight and the material stacking density, calculate the grid overlap degree between adjacent maintenance locations, and the grid overlap degree includes the number of overlapping grids, the cumulative value of the equipment occupancy weights of the overlapping grids, and the cumulative value of the material stacking densities of the overlapping grids;

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

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

[0051] In the second aspect of the embodiments of the present invention,

[0052] Provide an automatic generation system for an intelligent road maintenance plan, including:

[0053] The first unit is used to obtain road surface image data and perform preprocessing, extract features and detect targets through a convolutional neural network, identify the types and location information of road surface diseases, and obtain an initial feature map of the disease area;

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

[0055] The third unit is used to match the basic maintenance plan based on the evaluation result of the road surface disease level;

[0056] The fourth unit is used to construct the basic maintenance plan into an internal domain relationship graph, construct an external domain feature matrix in combination with environmental monitoring data and resource status data, and obtain an optimized maintenance plan by using a dual gradient algorithm with the maintenance quality and resource utilization rate as the optimization objectives through a hierarchical decision network enhanced by cross attention;

[0057] The fifth unit is used to perform spatio-temporal impact analysis on the optimized maintenance plan, determine the construction space conflict degree and traffic flow distribution in combination with historical maintenance data, calculate the dynamic weight of the maintenance location based on the loss of traffic capacity, and perform spatio-temporal optimization adjustment on the maintenance plan to generate the final road surface maintenance plan.

[0058] In the third aspect of the embodiments of the present invention,

[0059] Provide an electronic device, including:

[0060] A processor;

[0061] A memory for storing instructions executable by the processor;

[0062] Wherein, the processor is configured to call the instructions stored in the memory to execute the foregoing method.

[0063] In the fourth aspect of the embodiments of the present invention,

[0064] a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0065] In the embodiments of the present invention, an intelligent road maintenance plan is generated in an automated manner, improving the efficiency and accuracy of maintenance work, reducing the need for manual intervention, and lowering the possibility of human errors; by combining multi-scale grid adaptive feature extraction with historical data, the evolution of diseases can be predicted more accurately, providing a more scientific basis for maintenance decision-making, thereby improving the overall quality of road maintenance; by combining environmental monitoring data and resource status data, the utilization rate of maintenance resources is optimized, the spatio-temporal optimization of the maintenance plan is achieved, the impact of construction on traffic flow is reduced, and the sustainability of maintenance work is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 It is a schematic flow chart of the method for automatically generating an intelligent road maintenance plan according to the embodiments of the present invention;

[0067] Figure 2 It is a comparison chart of the evaluation of the effects of multi-scale curvature features and Transformer encoding;

[0068] Figure 3 It is a comparison chart of the stress distribution prediction performance of the physical constraint graph convolutional network for different disease types;

[0069] Figure 4 It is a schematic diagram of the cross-attention module bidirectional feature fusion network structure. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0070] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

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

[0072] Figure 1 It is a schematic flow chart of the method for automatically generating an intelligent road maintenance plan according to the embodiments of the present invention, as Figure 1 shown, the method includes:

[0073] Obtain road surface image data and perform preprocessing, extract features and detect targets through a convolutional neural network, identify the types and location information of road surface diseases, and obtain the initial feature map of the disease area;

[0074] Based on the initial feature map of the disease area, adopt multi-scale grid adaptive feature extraction to obtain the geometric deformation features of the disease, combine historical data to predict the disease evolution, construct a dynamic evaluation vector for classification evaluation, and obtain the evaluation result of the road surface disease level;

[0075] Match the basic maintenance plan based on the evaluation result of the road surface disease level;

[0076] Construct the basic maintenance plan into an internal domain relationship graph, combine environmental monitoring data and resource status data to construct an external domain feature matrix, and through a hierarchical decision network enhanced by cross attention, with the maintenance quality and resource utilization rate as the optimization objectives, adopt the dual gradient algorithm to obtain the optimized maintenance plan;

[0077] Conduct spatio-temporal impact analysis on the optimized maintenance plan, combine historical maintenance data to determine the construction space conflict degree and traffic flow distribution, calculate the dynamic weight of the maintenance location based on the loss of traffic capacity, and perform spatio-temporal optimization adjustment on the maintenance plan to generate the final road surface maintenance plan.

[0078] In an optional implementation manner, based on the initial feature map of the disease area, adopt multi-scale grid adaptive feature extraction to obtain the geometric deformation features of the disease, combine historical data to predict the disease evolution, construct a dynamic evaluation vector for classification evaluation, and the evaluation result of the road surface disease level includes:

[0079] Construct a multi-resolution image pyramid, perform feature extraction on the initial feature map of the disease area at different proportional scales, divide structural unit grids at each scale, calculate the grid complexity score based on the pixel distribution characteristics within the structural unit grids, and adaptively adjust the feature extraction density of each structural unit grid according to the grid complexity score;

[0080] Extract three-dimensional curvature descriptors for each structural unit grid, construct a geometric deformation field of the disease area based on the three-dimensional curvature descriptors, and extract the disease expansion trend features from the geometric deformation field;

[0081] Input the disease expansion trend features and historical monitoring data into a pre-trained recurrent neural network to predict the disease development state at future time nodes, and combine the current disease features with the disease development state to construct a dynamic evaluation vector;

[0082] Train a multi-task learning classifier based on the dynamic evaluation vector. The multi-task learning classifier simultaneously outputs the disease type, disease degree, and development level. The pavement disease evaluation result is obtained by weighted fusion of the disease type, disease degree, and development level according to the maintenance priority.

[0083] In a specific implementation, preprocess the collected pavement disease images. The preprocessing includes image denoising, illumination normalization, and geometric correction. Use a Gaussian filter to denoise the image, with the filter kernel size set to 5×5 and the standard deviation to 1.5. Perform illumination normalization through histogram equalization to adjust the image brightness within the standard distribution range. Geometric correction uses perspective transformation to correct the obliquely captured image into a standard top view to ensure the accuracy of feature extraction.

[0084] After obtaining the initial feature map of the disease area, construct a multi-resolution image pyramid for feature extraction. The pyramid contains 4 scale levels, and the scaling ratio between adjacent levels is 0.75. 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, divide the image into a grid of structural units, with the basic grid size being 32×32 pixels.

[0085] Calculate the complexity score of each grid for adaptively adjusting the feature extraction density. The grid complexity score is comprehensively calculated based on three factors: pixel gradient distribution, texture complexity, and edge density within the grid. For the gradient distribution, calculate the standard deviation of the gradient magnitudes of all pixel points within the grid; for texture complexity, use the entropy value of the gray-level co-occurrence matrix; for edge density, calculate the proportion of edge pixels after extracting edges using the Canny edge detection algorithm. The weights of the three factors are 0.4, 0.3, and 0.3 respectively.

[0086] According to the calculated complexity score, adaptively adjust the feature extraction density. For grids with a complexity score higher than 0.8, subdivide them into 2×2 sub-grids for more refined feature extraction; grids with scores between 0.5 and 0.8 remain the original size; grids with scores lower than 0.5 can be merged into larger grid units to reduce the computational overhead. In practical applications, for a typical asphalt pavement crack area, approximately 15% of the grids are subdivided, 60% of the grids remain the original size, and 25% of the grids are merged.

[0087] For each structural unit grid, three-dimensional curvature descriptors are extracted to construct the geometric deformation field of the disease area. The three-dimensional curvature descriptors include Gaussian curvature, mean curvature, and principal curvature. Depth information is estimated from a single image based on a deep learning network and converted into three-dimensional point cloud data. The local quadratic surface fitting method is used to calculate the curvature value of each point. For typical pavement cracking diseases, the Gaussian curvature value is usually in the range of [-0.05, 0.05], and the mean curvature value is in the range of [-0.03, 0.03].

[0088] Based on the three-dimensional curvature descriptors, a geometric deformation field is constructed. The system uses a tensor field representation method and expands the discrete curvature values into a continuous field through an interpolation algorithm. The disease expansion trend features are extracted from the geometric deformation field, including the gradient direction, deformation amplitude, and deformation continuity of the deformation field. For typical pavement cracks, by analyzing the distribution of the gradient direction of the deformation field, the main expansion direction of the cracks can be identified; the deformation amplitude reflects the severity of the disease; and the deformation continuity indicates the coherence of the disease and the scope of structural influence.

[0089] The disease expansion trend features and historical monitoring data are input into a pre-trained recurrent neural network to predict the disease development state at future time nodes. The recurrent neural network adopts a bidirectional LSTM architecture, including 3 hidden layers with 128 neurons in each layer. The network input includes the geometric feature vector (64-dimensional) of the current disease and the historical monitoring records (feature changes at the past 6 time points). The network is trained using the Adam optimizer with a learning rate of 0.001, a batch size of 32, and 200 training epochs. Through this network, the system can predict the disease development states at 3 months, 6 months, and 12 months in the future, and the prediction accuracies reach 92%, 87%, and 81% respectively.

[0090] The current disease characteristics and the predicted disease development states are combined to construct a dynamic evaluation vector. The dynamic evaluation vector includes a current state feature sub-vector (including disease type, area, depth, texture features, etc.) and a development prediction sub-vector (including area expansion rate, depth growth rate, structural influence range expansion trend, etc.). For a typical pavement crack case, the current state sub-vector shows that the crack length is 2.3 meters, the width is 3.2 centimeters, and the depth is 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] Train a multi-task learning classifier based on a dynamic evaluation vector, and simultaneously output the disease type, disease severity, and development level. The classifier adopts an architecture with a shared underlying feature extraction network and dedicated task branches. The shared network includes 4 convolutional layers and 2 fully connected layers; each dedicated branch includes 2 fully connected layers. The disease type classification includes categories such as cracks, potholes, looseness, and ruts; the disease severity is divided into three levels: minor, moderate, and severe; the development level is divided into three levels: stable type, slow development type, and rapid development type.

[0092] Perform weighted fusion on the disease type, disease severity, and development level according to the maintenance and repair priority to obtain the pavement disease evaluation result. The weighting coefficients are 0.3, 0.4, and 0.3 respectively. For the crack disease with a severity of "moderate" and a development level of "rapid development type", its final evaluation level is "high priority" and it will be repaired within 3 months; for the looseness disease with a severity of "minor" and a development level of "stable type", its final evaluation level is "low priority" and it will be handled during routine maintenance.

[0093] In this embodiment, through the multi-resolution image pyramid and grid complexity scoring, the feature extraction density can be automatically adjusted according to the texture and morphology of the disease area, which not only ensures the high-precision analysis of key areas but also avoids redundant calculations for flat areas, improving the processing efficiency; using the three-dimensional curvature descriptor to reconstruct the geometric deformation field of the disease area can finely depict the spatial morphology of diseases such as cracks and spalling, and the expansion trend features extracted from the deformation field are more consistent with the real development path, providing a reliable basis for subsequent prediction; inputting the disease features extracted in real time and the historical monitoring data into a pre-trained recurrent neural network to achieve accurate prediction of the disease state at each future time node; and fusing the prediction results with the current features to construct a dynamic evaluation vector, making the evaluation result consider both the current situation and be forward-looking; the multi-task learning classifier trained by the dynamic evaluation vector can synchronously output the disease type, disease severity, and development level, avoiding the cumbersome process of separate training and inference of multiple models and improving the consistency and reliability of the evaluation.

[0094] In an alternative embodiment, extract the three-dimensional curvature descriptor for each structural unit grid, construct the geometric deformation field of the disease area based on the three-dimensional curvature descriptor, and the extraction of the disease expansion trend features from the geometric deformation field includes:

[0095] Perform adaptive meshing on each structural unit grid, determine the grid division granularity based on the spatial distribution density of the disease area, generate hierarchical grid nodes, perform local surface fitting on the three-dimensional point cloud data corresponding to the hierarchical grid nodes by the least squares method guided by the depth image, adaptively adjust the fitting weights in combination with the attention mechanism, generate a local surface feature map, and calculate the normal vector field and the principal direction field of the hierarchical grid nodes;

[0096] Based on the normal vector field and the principal direction field, calculate three-dimensional curvature descriptors at multiple receptive field scales, including principal curvature, Gaussian curvature, and mean curvature. Use a Transformer encoder to perform self-attention calculations on the curvature features at different scales to obtain global correlation features, and combine them with local surface feature maps to construct hierarchical geometric features;

[0097] Input the hierarchical geometric features into a physically constrained graph convolutional network. Based on the pre-determined material mechanics properties, construct a stress-strain mapping relationship. Calculate the stress distribution field of the disease area through feature propagation between hierarchical grid nodes, deduce the local deformation field, and determine the geometric deformation field;

[0098] Perform spatio-temporal tensor decomposition on the geometric deformation field, extract the principal strain direction and strain intensity, form a deformation feature sequence, and input it into a preset temporal graph neural network to construct the disease expansion trend feature.

[0099] In a specific implementation, perform adaptive meshing on each structural unit grid. Determine the grid division granularity based on the spatial distribution density of the disease area. The spatial distribution density can be obtained by calculating the number of disease points per unit volume. When the disease point density in a certain area exceeds a threshold (such as 10 points per cubic centimeter), set the grid granularity of this area to a smaller value (such as 0.5 mm); when the density is low (such as less than 3 points per cubic centimeter), set the grid granularity to a larger value (such as 2 mm). Generate hierarchical grid nodes in this way to form an octree structure, making the grid division more refined near the disease area and relatively rough far from the disease area.

[0100] Perform local surface fitting on the three-dimensional point cloud data corresponding to the hierarchical grid nodes through depth image-guided least squares method. Project the three-dimensional point cloud onto the depth image plane to obtain the depth value and normal information of each point; for each grid node, select the point set within its neighborhood (usually the points within a spherical area with a radius three times the grid size), establish a local coordinate system, and use a quadratic surface equation for fitting in the local coordinate system. The fitting parameters are solved by the least squares method.

[0101] During the surface fitting process, adaptively adjust the fitting weights in combination with the attention mechanism. Specifically, calculate the spatial distance and normal difference between each point in the point set and the center point, and construct an attention weight matrix. Points closer to the center point and with smaller normal differences obtain higher weights. Exemplarily, when the distance of 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 2 times the grid size or the normal angle is greater than 45 degrees, the weight is reduced to 0.1. Through this weighted method, the fitting process can better retain local geometric features and generate accurate local surface feature maps.

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

[0103] Based on the normal vector field and the principal direction field, three-dimensional curvature descriptors are calculated at multiple receptive field scales. Specifically, three different receptive field scales are selected (such as 1 times, 2 times, and 4 times the grid size), and the principal curvatures (including the maximum curvature and the minimum curvature), the Gaussian curvature, and the mean curvature are calculated at each scale. For example, for the crack area of a certain concrete structure, the maximum principal curvature measured at the 1-time grid scale may be 0.35 mm -1 , the minimum principal curvature is 0.03 mm -1 , the Gaussian curvature is 0.0105 mm -2 , and the mean curvature is 0.19 mm -1 ; while at the 4-time grid scale, these values may become 0.15 mm -1 , 0.01 mm -1 , 0.0015 mm -2 , and 0.08 mm -1 .

[0104] After obtaining the multi-scale curvature features, the self-attention calculation of the curvature features at different scales is performed using a Transformer encoder. The Transformer encoder contains 3 attention heads, each with a dimension of 64 and 4 layers. The input is the curvature feature vectors at different scales (with a dimension of 12, that is, 4 curvature values at 3 scales), and the output is the fused global correlation features (with a dimension of 128). Through the self-attention mechanism, the system can automatically learn the importance weights of the features at each scale. For example, for small cracks, the small-scale features obtain a higher weight (such as 0.6), while the large-scale features have a lower weight (such as 0.2).

[0105] The global correlation features are combined with the local surface feature map to construct hierarchical geometric features. Specifically, the global correlation features (128-dimensional) are concatenated with the local surface features (64-dimensional) and mapped to a 256-dimensional feature space through a fully connected layer to form the final hierarchical geometric features.

[0106] The hierarchical geometric features are input into a graph convolutional network with physical constraints. The graph convolutional network consists of 3 graph convolutional layers, and the number of channels in each layer is 256, 128, and 64 respectively. The LeakyReLU activation function is used. The connection relationship of the graph is determined by the spatial adjacency relationship of the grid nodes, and usually the connection distance threshold is set to 2 times the grid size.

[0107] In the graph convolutional network, a stress-strain mapping relationship is constructed based on predetermined material mechanical properties. For example, for a concrete structure, its Young's modulus is set to 30 GPa and the Poisson's ratio is 0.2; for a steel structure, the Young's modulus is 210 GPa and the Poisson's ratio is 0.3. These parameters serve as physical constraint conditions in the graph convolutional network, guiding the network to learn the characteristic propagation method that conforms to the laws of material mechanics.

[0108] The stress distribution field of the disease area is calculated through the characteristic propagation between hierarchical grid nodes. In each graph convolutional layer, the node features consider not only the geometric features of neighboring nodes but also the mechanical properties of the material. For example, for the nodes at the crack edge, due to the stress concentration effect, their stress values may be 3 - 5 times that of the nodes far from the crack area. This stress distribution conforms to the real physical law, that is, 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, the stress is converted into strain through Hooke's law, and the strain is multiplied by the structural size to obtain the deformation amount. For example, when the stress at a certain point reaches 10 MPa, the corresponding strain is 3.33×10 -4 , if the structural size where this point is located is 100 mm, then the deformation amount is 0.0333 mm. In this way, the system constructs a complete geometric deformation field.

[0110] Perform spatio-temporal tensor decomposition on the geometric deformation field to extract the principal strain direction and strain intensity. Using the singular value decomposition method, the deformation field tensor is decomposed into a direction tensor and an intensity scalar. For example, for a certain crack propagation area, the principal strain direction may be consistent with the crack propagation direction, with an included angle less than 10 degrees, and the strain intensity decays exponentially with the increase in the distance from the crack tip, reaching 0.001 at the tip and dropping to 0.0001 at a distance of 10 mm from the tip.

[0111] The principal strain direction and strain intensity are combined into a deformation feature sequence and input into a preset temporal graph neural network. This network contains 2 temporal convolutional layers and 1 graph attention layer. The input is the deformation feature sequence at different time points (at least 3 time points, with an interval of 7 days), and the output is the disease expansion trend feature. By analyzing the time evolution law of the deformation feature sequence, the system can predict the expansion direction and rate of the crack. For example, if a certain crack has expanded by 5 mm within 28 days and the strain intensity has increased by 20%, it is predicted that the crack may continue to expand by 3 - 7 mm within the next 28 days, and the specific value depends on the environmental load conditions.

[0112] The technology in this embodiment stems from cross - research in the fields of structural health monitoring, computer vision, and graph neural networks. In the prior art, structural disease analysis mainly uses traditional grid meshing methods, usually with uniform grid meshing, which is difficult to adaptively adjust for disease areas, resulting in waste of computing resources or insufficient accuracy. Conventional surface fitting techniques mainly perform point cloud data fitting based on algorithms such as global least squares or RANSAC, lacking the ability to finely capture the characteristics of disease areas. The prior art also has limitations in single - scale geometric feature extraction, calculating curvature features only at a single receptive field scale and unable to comprehensively express multi - scale disease features. In addition, simplified mechanical models such as linear models or simplified finite element analysis are mostly used, ignoring the non - linear characteristics and dynamic responses of materials. Most methods only focus on the static distribution of diseases and lack the ability to predict the development trend of diseases.

[0113] This embodiment has improved on the deficiencies of the prior art. In terms of grid meshing, the starting point of this embodiment is to improve computational efficiency and accuracy. The grid meshing granularity is adaptively determined based on the spatial distribution density of the disease area, achieving reasonable allocation of computing resources and providing more refined analysis in key areas. In terms of surface fitting, this application aims to improve the surface fitting accuracy, integrates depth image information into the least squares method, and introduces an attention mechanism to adaptively adjust the fitting weights, making the surface fitting more conform to the actual geometric characteristics of the disease area. For geometric feature extraction, the improvement starting point of this embodiment is to comprehensively capture the multi - scale features of diseases, calculate curvature descriptors at multiple receptive field scales, and perform feature fusion through a Transformer encoder, enhancing the feature expression ability. To improve the accuracy of mechanical analysis, this embodiment introduces the mechanical properties of materials as constraint conditions into the graph convolutional network, constructs a more accurate stress - strain mapping relationship, and achieves precise expression of the deformation characteristics of the disease area. To predict the development trend of diseases, this embodiment extracts the deformation feature sequence through spatio - temporal tensor decomposition and combines it with a temporal graph neural network to establish a prediction model for the disease expansion trend.

[0114] The technical improvements in this embodiment have brought remarkable effects. Adaptive mesh refinement makes the allocation of computing resources more reasonable. Compared with uniform mesh refinement, the computing efficiency is improved while maintaining high-precision analysis in key areas. The multi-scale geometric feature and Transformer fusion mechanism enables a more comprehensive expression of disease features, improving the feature recognition accuracy, especially for the recognition of complex disease patterns. The physically constrained graph convolutional network makes the stress-strain analysis more in line with the actual material properties. Compared with traditional finite element analysis, the stress prediction error is reduced. The combination of spatio-temporal tensor decomposition and temporal graph neural network endows this technology with the ability to predict the development trend of diseases, improving the prediction accuracy and providing a reliable basis for structural maintenance decisions. The technical solution of this application is applicable to various structural disease analyses, has good adaptability to different materials and different structural types, and expands the application scope.

[0115] As Figure 2 shown, it demonstrates the effect evaluation of the combination of multi-scale curvature features and Transformer encoder. The solution of this embodiment realizes high-precision recognition of different types of diseases by calculating three-dimensional curvature descriptors at multiple receptive field scales and using Transformer for feature fusion. From the data, in the recognition of crack diseases, the mIoU of the solution of this embodiment reaches 92.7%, an increase of 17.5 percentage points compared with 75.2% of single-scale curvature features and 10.3 percentage points compared with 82.4% of traditional multi-scale feature fusion, and the recognition accuracy is as high as 94.2%; in the recognition of spalling diseases, the mIoU of the solution of this embodiment is 89.5%, an increase of 17.7 percentage points compared with single-scale features and 10.9 percentage points compared with traditional fusion methods, and the accuracy reaches 92.8%; in the recognition of pitted surface diseases, the mIoU is 87.3%, an increase of 17.9 percentage points compared with single-scale features and 10.4 percentage points compared with traditional methods, and the accuracy is 90.5%; in the recognition of exposed reinforcement diseases, the mIoU reaches 90.8%, an increase of 18.3 percentage points compared with single-scale features and 11.6 percentage points compared with traditional methods, and the accuracy is 93.1%; in the recognition of water seepage diseases, the mIoU is 91.5%, an increase of 17.2 percentage points compared with single-scale features and 10.0 percentage points compared with traditional methods, and the accuracy is 93.8%; in the recognition of carbonation diseases, the mIoU is 88.2%, with increase amplitudes of 17.6 and 10.4 percentage points respectively, and the accuracy is 91.7%; in the recognition of corrosion diseases, the mIoU is 89.7%, with increase amplitudes of 16.6 and 9.4 percentage points respectively, and the accuracy is 92.5%. Especially in the complex scenario of mixed multi-type diseases, the solution of this embodiment still maintains an mIoU of 86.4%, showing outstanding performance, an increase of 19.6 percentage points compared with single-scale features and 12.2 percentage points compared with traditional methods, and the accuracy reaches 89.3%.

[0116] AsFigure 3 As shown, it demonstrates the stress distribution prediction performance of the physically constrained graph convolutional network for different disease types. The solution of this embodiment uses a physically constrained graph convolutional network to calculate the stress distribution field, integrating material mechanical properties into the neural network architecture, which not only ensures the calculation speed but also guarantees physical accuracy. As can be seen from the data, the solution of this embodiment only takes 23.5 seconds for the calculation time in the slight crack scenario, a reduction of 87.6% compared with 189.3 seconds of the traditional finite element method, and the relative error is only 3.2%, with the predicted maximum stress intensity value being 12.8 MPa; in the moderate crack scenario, the calculation time is 42.8 seconds, a reduction of 84.0% compared with 267.5 seconds of the traditional method, the relative error is 4.8%, and the predicted maximum stress intensity value reaches 24.6 MPa; in the severe crack scenario, the calculation time is 68.7 seconds, a reduction of 80.7% compared with 356.8 seconds of the traditional method, the relative error is 5.7%, and the predicted maximum stress intensity value is as high as 36.9 MPa; in the slight spalling scenario, the calculation time is 31.2 seconds, a reduction of 84.7% compared with 203.7 seconds of the traditional method, the relative error is only 2.9%, and the predicted maximum stress intensity value is 9.7 MPa; in the severe spalling scenario, the calculation time is 57.6 seconds, a reduction of 80.7% compared with 298.2 seconds of the traditional method, the relative error is 4.5%, and the predicted maximum stress intensity value is 18.3 MPa; in the complex mixed disease scenario, the calculation time is 84.3 seconds, a reduction of 77.3% compared with 371.4 seconds of the traditional method, the relative error is 6.8%, and the predicted maximum stress intensity value reaches 41.5 MPa. Compared with the conventional GCN, although the calculation time of the solution of this embodiment is slightly longer (an average increase of about 21%), due to the introduction of physical constraints, the prediction accuracy is significantly improved, and the relative error is controlled within 7% in all scenarios. Especially for the severe disease scenarios that are critical to structural safety, the prediction is more accurate and reliable.

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

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

[0119] The multi-head self-attention mechanism is used to process the inner domain relationship graph and the outer domain feature matrix respectively, calculate the association weights between nodes and the influence degree between features, and generate the inner domain attention representation and the outer domain attention representation;

[0120] Based on adaptive representation decomposition and temporal dependence analysis, the inner domain attention representation and the outer domain attention representation are hierarchically processed, and an enhanced state vector is output through a bidirectional cross-attention network;

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

[0122] The hierarchical decision network is trained using the dual gradient optimization algorithm, a dual objective function including a maintenance quality evaluation item and a resource utilization rate evaluation item is constructed, and the network parameters are alternately optimized and updated in the original space and the dual space to determine the optimized maintenance plan.

[0123] In a specific implementation, an initial feature vector is constructed, which includes maintenance process nodes and resource configuration nodes. The maintenance process nodes describe the specific maintenance technical processes. For example, for the maintenance of concrete structures, the process nodes include surface cleaning, crack treatment, coating construction, etc.; the resource configuration nodes describe the equipment, personnel, and materials required for maintenance. For example, there are 3 high-pressure cleaning machines, 5 professional technicians, 200 kilograms of anti-corrosion coating, etc. A topological connection structure between nodes is established according to the association degree of maintenance processes and the dependence relationship of resource configuration to form an inner domain relationship graph. For example, there is a sequential relationship between surface cleaning and crack treatment, and the connection weight is set to 0.8; there is a tool dependence relationship between the high-pressure cleaning machine and surface cleaning, and the connection weight is set to 0.9.

[0124] External environment and resource status data are collected to construct an outer domain feature matrix. The environmental monitoring data includes temperature (such as 25°C), humidity (such as 65%), precipitation (such as 0 mm / day), etc.; the resource status data includes the availability of equipment (such as the availability rate of high-pressure cleaning machines is 95%), the skill levels of personnel (such as 3 professional personnel at level A and 2 at level B), the inventory of materials (such as there are currently 180 kilograms of anti-corrosion coating), etc. These data form an outer domain feature matrix, and the matrix dimension is the number of features multiplied by the time series length.

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

[0126] Based on adaptive representation decomposition and time - series dependence analysis, the in - domain and out - domain attention representations are processed in layers. Adaptive representation decomposition decomposes the attention representation into a short - term response part and a long - term influence part. For example, the impact of rainfall on coating construction belongs to the short - term response, and the impact of material aging on coating quality belongs to the long - term influence. Time - series dependence analysis uses the sliding window method, sets the window size to 7 days, analyzes the fluctuation trend of environmental parameters within the window, and predicts the environmental state in the next 3 days. Subsequently, through a bidirectional cross - attention network, the in - domain representation is used as the query vector, and the out - domain representation is used as the key - value vector to calculate the cross - attention score and generate an enhanced state vector. This vector contains both the internal relationships of the process and the influence of the external environment, with a dimension of 128, and each dimension corresponds to a specific attribute of the maintenance plan.

[0127] The enhanced state vector is input into a hierarchical decision - making network, which includes a high - level policy module and a low - level execution module. The high - level policy module consists of three fully - connected networks, and the number of hidden - layer nodes is 256, 128, and 64 respectively, and the activation function uses ReLU. This module generates a maintenance process combination plan based on the enhanced state vector. For example, in a humid environment, a process combination of "surface pretreatment + waterproof coating + protective covering" is selected. The low - level execution module consists of four fully - connected networks, and the number of hidden - layer nodes is 128, 256, 128, and 64 respectively, and the activation function uses Leaky ReLU. This module determines the corresponding parameter configuration based on the maintenance process combination plan, such as setting the coating thickness to 2.3 mm, the number of brushing times to 3 times, and the interval time to 4 hours, etc.

[0128] The hierarchical decision-making network is trained using the dual gradient optimization algorithm. A dual objective function is constructed, which includes a maintenance quality evaluation term and a resource utilization rate evaluation term. The maintenance quality evaluation term comprehensively considers factors such as material adhesion (target value greater than 2.5 MPa), surface flatness (target deviation less than 0.5 mm / m), durability index (target service life greater than 5 years), etc. The resource utilization rate evaluation term considers factors such as personnel man-hour utilization rate (target greater than 85%), equipment usage efficiency (target greater than 80%), material loss rate (target less than 5%), etc. Alternate optimization is carried out in the original space and the dual space. In each round of iteration, the original objective gradient and the dual variable gradient are calculated, and the network parameters are updated. The initial learning rate is set to 0.001 and dynamically adjusted using the cosine annealing strategy. The total number of iterations is 5000 rounds. According to the training results, for specific maintenance scenarios, such as the maintenance of coastal concrete structures, the optimized maintenance plan can improve the resource utilization rate by 20% while ensuring a 15% increase in the maintenance quality score, and the overall benefit is increased by approximately 18%.

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

[0130] In this embodiment, by constructing the maintenance process nodes and resource configuration into an inner domain relationship graph, and introducing environmental monitoring such as temperature, humidity, precipitation, and outer domain features such as equipment, personnel, and materials, the multi-head self-attention mechanism is used to finely capture the topological correlation and cross-domain influence between various elements. Then, through hierarchical processing of adaptive decomposition and time-series dependence analysis, and fusion of the bidirectional cross-attention network, an enhanced state vector with both historical evolution and dynamic interaction is generated; on this basis, the high-level policy module of the hierarchical decision-making network outputs the optimal process combination, while the low-level execution module accurately configures the parameters. Finally, through the dual gradient algorithm that alternately optimizes in the dual objective space of maintenance quality and resource utilization rate, both the balance between quality and efficiency is achieved, and the interpretability and transparency of the decision-making process are ensured.

[0131] In an alternative embodiment, based on adaptive representation decomposition and time-series dependence analysis, hierarchical processing of the inner domain attention representation and the outer domain attention representation is performed, and the enhanced state vector output by the bidirectional cross-attention network includes:

[0132] Construct an adaptive representation decomposition unit to adaptively group the process-based correlation weights of the inner-domain attention representation and the impact degree of resource allocation of the outer-domain attention representation, calculate the information importance of different groups based on attention entropy, and form a hierarchical representation sequence;

[0133] Construct a temporal dependence analysis unit to model the historical representations of process combinations and environmental resource states, calculate the temporal evolution features corresponding to the inner-domain attention representation and the outer-domain attention representation, and output a time-varying weight matrix containing process combination weights and environmental resource weights;

[0134] Pass each hierarchical representation sequence through a cross-attention module for two-way feature fusion. Based on the time-varying weight matrix, map the inner-domain attention representation to a query matrix and the outer-domain attention representation to a key-value matrix to calculate the forward attention score, and map the outer-domain attention representation to a query matrix and the inner-domain attention representation to a key-value matrix to calculate the reverse attention score, respectively obtaining a process-dominated forward interaction representation and a resource-constrained reverse interaction representation;

[0135] Adopt a dynamic selection gating mechanism to generate feature selection signals based on the time-varying weight matrix and attention entropy, dynamically select and adaptively fuse the forward interaction representation and the reverse interaction representation at each level to determine the fused representation;

[0136] Adopt an inter-layer skip connection mechanism to cascade the fused representations at different levels, and calculate the dependence relationship between the fused representations through self-attention to obtain an enhanced state vector.

[0137] In a specific embodiment, an adaptive representation decomposition unit is constructed to adaptively group the correlation weights of the inner-domain attention representation based on the process. In actual implementation, the maintenance process nodes can be clustered according to functional similarity. For example, for the concrete bridge maintenance process, the surface treatment processes (such as sandblasting, high-pressure water cleaning) can be grouped into one group, the crack repair processes (such as grouting, caulking) can be grouped into one group, and the protective coating processes (such as anti-corrosion coating, waterproof coating) can be grouped into one group. During the adaptive grouping process, the initial number of groups is set to 3, and the correlation weight matrix is iteratively optimized to finally determine the optimal grouping. Similarly, the outer-domain attention representation is adaptively grouped based on the influence degree of resource allocation, and the environmental factors (such as temperature, humidity, wind speed), equipment resources (such as high-pressure cleaning machines, spraying equipment), personnel resources (such as technical workers, ordinary workers), and material resources (such as anti-corrosion coatings, repair materials) are clustered respectively. The information importance of different groups is calculated based on the attention entropy. The attention entropy quantifies the information importance by calculating the dispersion degree of the attention distribution. For example, for a certain maintenance task, the attention entropy of environmental factors is 0.82, and the attention entropy 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, sorted from high to low entropy values, and a three-layer representation sequence is constructed, with each layer containing the corresponding inner-domain and outer-domain grouped representations.

[0138] A temporal dependence analysis unit is constructed to model the historical representations of the process combination and the environmental resource state. The sliding window mechanism is used, with the window size set to 14 days and the step size to 1 day, and temporal features are extracted within the window. For example, for the coating process, analyze the influence mode of temperature change on the coating curing time within 14 days; for equipment resources, analyze the relationship between the equipment usage frequency and the failure rate within 14 days. Calculate the temporal evolution features corresponding to the inner-domain attention representation and the outer-domain attention representation, and extract the time dependence of the process effect and the environmental resources through autoregressive analysis. For example, it is found through analysis that there is a correlation coefficient of 0.78 between the environmental humidity within 3 days after coating construction and the coating curing quality, while the humidity correlation before coating construction is only 0.32. Based on the temporal analysis results, a time-varying weight matrix containing the process combination weights and the 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, and each element in the matrix represents the weight coefficient of a specific group at a specific time point.

[0139] Each hierarchical representation sequence undergoes two-way feature fusion through a cross-attention module. Based on a time-varying weight matrix, the in-domain attention representation is mapped to a query matrix, and the out-of-domain attention representation is mapped to a key-value matrix to calculate the forward attention score. In a specific implementation, a multi-head attention mechanism is used, with the number of heads set to 8 and the dimension of each attention head being 64. For example, for the in-domain representation of the anti-corrosion coating process, the correlation score is calculated with out-of-domain representations such as environmental humidity and temperature through the attention mechanism. Under the conditions of a temperature of 25°C and a relative humidity of 60%, the optimal execution parameters of the coating process are determined. Similarly, the out-of-domain attention representation is mapped to a query matrix, and the in-domain attention representation is mapped to a key-value matrix to calculate the reverse attention score. For example, when the material inventory is insufficient (such as only 80 kg of anti-corrosion paint remains), through reverse attention calculation, the areas with higher priority (such as the key stress-bearing parts) obtain a higher weight of 0.75 for material allocation, while the weight of the secondary areas drops to 0.25. Through two-way attention calculation, the forward interaction representation dominated by the process and the reverse interaction representation constrained by resources are obtained respectively, forming complementary feature representations.

[0140] A dynamic selection gating mechanism is adopted to generate a feature selection signal based on a time-varying weight matrix and attention entropy. The gating mechanism inputs the time-varying weight and attention entropy through a learnable parameter network and outputs a selection weight. For example, in rainy weather, the attention entropy of environmental factors increases to 0.95, and the gating mechanism correspondingly increases the influence weight of environmental factors on the decision-making to 0.85 and reduces the influence weights of other factors. The forward and reverse interaction representations at each level are dynamically selected and adaptively fused to determine the fused representation. For the first-level representation (high entropy value layer), the fusion weight of the forward interaction representation is 0.65, and the fusion weight of the reverse interaction representation is 0.35; for the second-level representation (medium entropy value layer), the weights are 0.55 and 0.45 respectively; for the third-level representation (low entropy value layer), the weights are 0.45 and 0.55 respectively. The fused representations of each layer are obtained through weighted summation.

[0141] An inter-layer skip connection mechanism is used to cascade the fused representations at different levels. In the implementation, a residual connection structure is used to directly connect the fused representation of the high entropy value layer to the outputs of the medium entropy value layer and the low entropy value layer to enhance information flow. For example, the representation of environmental factors (high entropy value) directly affects the decision-making process of material configuration (low entropy value). The dependence relationship between the fused representations is calculated through self-attention. A single-head self-attention mechanism is used to calculate the mutual dependence strength between representations at different levels. For example, the dependence strength between the environmental representation and the material representation 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 representation to obtain an enhanced state vector with a dimension of 256, which comprehensively represents the interaction relationship and temporal characteristics between the maintenance process and environmental resources.

[0142] The technology of this embodiment stems from the cross - research in the fields of deep learning and decision optimization. The maintenance plan generation systems in the prior art usually adopt one - way feature fusion methods, only considering the process's demand for resources and ignoring the reverse constraints of resource status on process selection, resulting in the generated maintenance plans often facing problems such as insufficient resources or unsuitable environments during actual execution. Traditional methods mostly use static representation fusion and cannot adapt to the dynamic changes of the environment and resource status, so the maintenance effect is limited. The improvement of this embodiment over the prior art lies in introducing a bidirectional cross - attention mechanism to achieve the bidirectional fusion of process requirements and resource constraints. At the same time, through adaptive representation decomposition and temporal - dependence analysis, the system's adaptability to the dynamic environment is enhanced. The starting point of the improvement is to solve the problems of the lack of adaptability of maintenance plans to environmental changes and unreasonable resource allocation, and to build a more flexible and efficient maintenance decision - making system.

[0143] As Figure 4 shown, it demonstrates the bidirectional feature fusion process of the cross - attention module. The figure includes two main parts: the inner - domain representation (left side) and the outer - domain representation (right side), and feature fusion is achieved through the bidirectional cross - attention mechanism. The inner - domain representation contains 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 contains 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 process (blue connection) means taking 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 P7→R2 (0.91) and P3→R6 (0.88); the reverse interaction process (red connection) means taking the outer - domain representation as the query matrix and the inner - domain representation as the key - value matrix. The highest attention scores are R2→P7 (0.93) and R6→P3 (0.87). The middle multi - layer fusion module shows the change in the weight distribution after each layer of fusion. After the first - layer fusion, the weight of the P7 - R2 combination is increased to 0.95, after the second layer, the weight of the P3 - R6 combination is increased to 0.91, and after the third layer, the weight of the P5 - R4 combination is increased to 0.87. Through this bidirectional feature fusion mechanism, the system can simultaneously consider the forward interaction representation dominated by the process and the reverse interaction representation of resource constraints, greatly improving the integrity and accuracy of the representation, making the subsequent decision - making more comprehensive and accurate.

[0144] In an optional embodiment, a spatio - temporal impact analysis is performed on the optimized maintenance plan. By combining historical maintenance data, the construction space conflict degree and traffic flow distribution are determined, and the dynamic weight of the maintenance location is calculated based on the loss of traffic capacity. The spatio - temporal optimization adjustment of the maintenance plan is carried out, and the final pavement maintenance plan generated includes:

[0145] Extract the maintenance time and maintenance location information from the historical maintenance data, determine the construction influence range of each maintenance location and the traffic flow distribution of each maintenance time, and generate a spatio-temporal analysis data set;

[0146] Based on the spatio-temporal analysis data set, discretize the maintenance locations into three-dimensional grids, calculate the degree of equipment occupancy and material stacking overlap of the grid cells, and construct a spatial operation conflict model to obtain a spatial operation constraint matrix;

[0147] Based on the spatio-temporal analysis data set and the spatial operation constraint matrix, conduct a traffic flow analysis on the maintenance time in the optimized maintenance plan, calculate the loss of traffic capacity in each maintenance time period considering spatial constraints, and determine the dynamic weight distribution of the maintenance locations;

[0148] According to the dynamic weight distribution of the maintenance locations, rank the maintenance operations in the optimized maintenance plan, aiming at maximizing the traffic capacity, determine the optimal implementation time period of each maintenance location, make spatio-temporal adjustments to the optimized maintenance plan, and determine the final road surface maintenance plan.

[0149] In a specific implementation manner, extract the maintenance time and maintenance location information from the historical maintenance data, determine the construction influence range of each maintenance location and the traffic flow distribution of each maintenance time, and generate a spatio-temporal analysis data set. Collect the maintenance record data of the urban road network in the past 3 years, including information such as maintenance location coordinates, maintenance time periods, construction area ranges, construction durations, etc. Exemplarily, for section A of the urban arterial road, 85 historical maintenance records are extracted, and each record includes the starting and ending stake numbers of the maintenance, start and end times, lane occupancy conditions, etc. At the same time, combined with the historical data of the traffic signal control system, obtain the traffic flow distribution of each time period on the road. Taking section A of the urban arterial road as an example, the average traffic volume during the morning rush hour (7:00 - 9:00) on weekdays is 2,800 vehicles per hour, the average traffic volume during the evening rush hour (17:00 - 19:00) is 3,200 vehicles per hour, the average traffic volume during the lunchtime period (12:00 - 14:00) is 1,600 vehicles per hour, and the average traffic volume during the night time period (22:00 - 6:00) is 650 vehicles per hour. Match the maintenance location information with the traffic flow information in space and time to generate a comprehensive data set including time and space dimensions, providing a data basis for subsequent analysis.

[0150] Perform three-dimensional grid discretization on the maintenance locations based on the spatio-temporal analysis dataset, calculate the equipment occupancy and material stacking overlap degree of the grid cells, and construct a spatial operation conflict model to obtain a spatial operation constraint matrix. During the implementation process, divide the road maintenance area into three-dimensional grids, with a horizontal unit size of 5 m × 3.5 m (corresponding to the road length and lane width), and a vertical unit height of 2 m. For a maintenance location (from pile number K12+500 to K12+650) on section A of the urban arterial road, a total of 30×3×2 grid cells are divided. Next, simulate the equipment occupancy of different maintenance processes. Exemplarily, a road milling machine occupies 4×2×1 grid cells, a paver occupies 5×3×1 grid cells, and a roller occupies 3×2×1 grid cells. The material stacking area is set according to actual needs. For example, asphalt mixture occupies 3×2×1 grid cells, and the temporary waste stacking area occupies 4×2×1 grid cells. Through grid occupancy analysis, calculate the spatial overlap degree of equipment and materials in each process, and form a spatial operation constraint matrix. This matrix assigns a value to the conflict degree between each process. For example, the conflict coefficient between milling and paving is 0.8 (indicating a high conflict), and the conflict coefficient between milling and marking construction is 0.2 (indicating a low conflict).

[0151] Based on the spatio-temporal analysis dataset and the spatial operation constraint matrix, conduct traffic flow analysis on the maintenance time in the optimized maintenance plan, calculate the traffic capacity loss in each maintenance time period considering spatial constraints, and determine the dynamic weight distribution of the maintenance locations. In the specific implementation, divide 24 hours a day into 12 time periods, each time period being 2 hours. Conduct statistical analysis on the traffic flow in each time period, and combine the impact of maintenance operations on the road traffic capacity to calculate the traffic capacity loss. For example, for section A of the urban arterial road (with an original traffic capacity of 3,600 vehicles / hour), if a lane is occupied for milling operation during the morning peak period (7:00 - 9:00), according to the traffic flow of 2,800 vehicles / hour and the spatial occupancy situation, the traffic capacity loss is 2,100 vehicles / hour, the remaining traffic capacity is 1,500 vehicles / hour, and the traffic capacity loss rate is 58.3%. In contrast, when the same operation is carried out during the night period (22:00 - 24:00), the traffic flow is 650 vehicles / hour, the traffic capacity loss is 450 vehicles / hour, the remaining traffic capacity is 3,150 vehicles / hour, and the traffic capacity loss rate is only 12.5%. Based on the traffic capacity loss rate, assign dynamic weights to each maintenance location in different time periods. The weight is in a direct proportion relationship with the traffic capacity loss rate. Taking section A of the urban arterial road as an example, the maintenance weight during the morning peak period is 0.85, the maintenance weight during the evening peak period is 0.92, the maintenance weight during the noon period is 0.45, and the maintenance weight during the night period is 0.15. The higher the weight value, the greater the impact of maintenance on traffic in this time period, and the more maintenance operations should be avoided.

[0152] Prioritize the maintenance operations in the optimized maintenance plan according to the dynamic weight distribution of the maintenance locations. With the goal of maximizing the traffic capacity, determine the optimal implementation time period for each maintenance location, adjust the optimized maintenance plan in terms of space and time, and determine the final pavement maintenance plan. In actual operation, first classify the maintenance requirements based on the maintenance urgency and pavement condition score. Exemplarily, the pavement condition score of the section from K12+500 to K12+650 of Road A, the main urban road, is 65 points (out of 100), belonging to the secondary maintenance priority level; the pavement condition score of the section from K5+200 to K5+350 of Road B, the secondary urban road, is 45 points, belonging to the primary maintenance priority level. Then, combined with the dynamic weight distribution, calculate the comprehensive score of each maintenance location in each time period. The score is composed of the weighted sum of the maintenance priority score and the time period weight. For example, the comprehensive score of Road A during the night time period is 80 points of the secondary priority score multiplied by (1 - night weight 0.15), which is equal to 68 points; the comprehensive score of Road B during the night time period is 90 points of the primary priority score multiplied by (1 - night weight 0.2), which is equal to 72 points. Based on the sorting result of the comprehensive scores, determine the optimal implementation order and time arrangement of the maintenance operations. The final pavement maintenance plan includes: 1) The section from K5+200 to K5+350 of Road B, the secondary urban road, is arranged for full-width maintenance from 22:00 to 6:00 the next day on the first night, using 2 milling machines, 1 paver, 2 rollers, and 18 construction workers; 2) The section from K12+500 to K12+650 of Road A, the main urban road, is arranged for half-width maintenance from 23:00 to 5:00 the next day on the second night, and the other half-width is maintained from 23:00 to 5:00 the next day on the third night, using 1 milling machine, 1 paver, 1 roller, and 12 construction workers. Through this space-time optimization arrangement, the overall loss of traffic capacity can be minimized while ensuring the maintenance quality and efficiency.

[0153] In this embodiment, by extracting the maintenance time and location from the historical maintenance data, combining with the traffic flow distribution, constructing a space-time analysis data set, and performing three-dimensional grid processing on the maintenance area, accurately modeling the spatial conflicts of equipment and materials, a spatial operation constraint is formed; on this basis, evaluate the loss of traffic capacity under spatial constraints in different time periods, dynamically adjust the weight distribution of each maintenance location, and achieve the prioritization of maintenance operations. Finally, with the goal of maximizing the traffic capacity, perform space-time coordination and optimization on the maintenance plan, improving the rationality of the plan, the road operation efficiency, and the intelligent level of the overall construction organization.

[0154] In an alternative embodiment, based on the space-time analysis data set, discretize the maintenance locations into three-dimensional grids, calculate the overlapping degree of equipment occupancy and material stacking in the grid cells, and construct a spatial operation conflict model to obtain a spatial operation constraint matrix including:

[0155] According to the maintenance location information in the spatio-temporal analysis dataset, the construction space of each maintenance location is divided into a core operation area, an equipment occupancy area, and a material stacking area, which are discretized into grid cells by a three-dimensional grid division method, and an equipment occupancy weight and a material stacking density are assigned to each of the grid cells;

[0156] Based on the equipment occupancy weight and the material stacking density, calculate the degree of grid overlap between adjacent maintenance locations, where the degree of grid overlap includes the number of overlapping grids, the cumulative value of the equipment occupancy weights of the overlapping grids, and the cumulative value of the material stacking densities of the overlapping grids;

[0157] Construct a spatial operation conflict model according to the degree of grid overlap. When the overlapping grid is located in the core operation area, set the spatial operation conflict degree to the maximum value. When the overlapping grid is located in the equipment occupancy area, calculate the spatial operation conflict degree based on the cumulative value of the equipment occupancy weights. When the overlapping grid is located in the material stacking area, calculate the spatial operation conflict degree based on the cumulative value of the material stacking densities;

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

[0159] In a specific embodiment, the construction space of each maintenance location is divided into three areas according to the maintenance location information in the spatio-temporal analysis dataset. The core operation area refers to the area where maintenance equipment is directly operated and staff are concentrated, usually located in the central part of the maintenance location. The equipment occupancy area refers to the space range where maintenance equipment is parked and moved, generally distributed around the core operation area. The material stacking area refers to the area where various maintenance materials are temporarily stored, usually located in the outer part of the maintenance location.

[0160] Taking the maintenance project of a certain railway section as an example, the core operation area of a rail replacement 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 operation 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] Discretize the above three areas into grid cells by a three-dimensional grid division method. Selecting an appropriate grid size is the key. Too large a grid will result in insufficient conflict judgment accuracy, and too small a grid will increase the computational complexity. In this embodiment, a grid cell size of 1 meter × 1 meter × 0.5 meter is adopted. The core operation area is divided into 10×3×5 = 150 grid cells, the equipment occupancy area is divided into 15×5×6 = 450 grid cells (including the grid cells of the core operation area), and the material stacking area is divided into 20×8×4 = 640 grid cells.

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

[0163] In practical applications, the equipment occupancy weights of grid cells in the core operation area are usually set to 0.8 to 1.0, determined according to the occupancy of specific equipment; the equipment occupancy weights of grid cells in the equipment occupancy area are 0.4 to 0.7, with lower weights in the edge areas; the material stacking densities in the material stacking areas are determined according to the types and quantities of the actually stacked materials. Generally, the density is higher in the central area (0.6 to 0.9) and lower in the edge area (0.2 to 0.5).

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

[0165] Based on the equipment occupancy weight and the material stacking density, calculate the degree of grid overlap between adjacent maintenance positions. When there is an overlap in the construction space of two maintenance positions, first identify the overlapping grid cells, and then calculate the number of overlapping grids, the cumulative value of the equipment occupancy weights of the overlapping grids, and the cumulative value of the material stacking densities of the overlapping grids.

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

[0167] Taking two adjacent rail replacement and maintenance positions A and B as an example, through spatial position overlay analysis, it is found that they have a total of 40 overlapping grid cells, of which 10 are located in the core operation area of A (and also in the equipment occupancy area of B), 15 are located in the equipment occupancy area of A (and also in the material stacking area of B), and 15 are located in the material stacking area of A (and also in the material stacking area of B). The cumulative value of the equipment occupancy weights of these 40 overlapping grid cells is 22.5, and the cumulative value of the material stacking densities is 18.6.

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

[0169] When the overlapping grid is located in the core operation area, considering that the core operation area is the key area for maintenance operations and any interference may cause the operation to be unable to proceed, the spatial operation conflict degree is set to the maximum value of 1. For example, if the core operation area of maintenance position A overlaps with any area of maintenance position B, the corresponding spatial operation conflict degree is set to 1.

[0170] When the overlapping grid is located in the equipment occupancy area (excluding the core operation area), calculate the spatial operation conflict degree based on the cumulative value of equipment occupancy weights. The specific calculation method is to divide the cumulative value of equipment occupancy weights by a preset threshold (usually set to 1.5), and then limit it within the range of 0 to 1. Exemplarily, if the cumulative value of equipment occupancy weights of the overlapping grid is 1.2, the spatial operation conflict degree of this part is 1.2 / 1.5 = 0.8.

[0171] When the overlapping grid is located in the material stacking area, calculate the spatial operation conflict degree based on the cumulative value of material stacking density. The specific calculation method is to divide the cumulative value of material stacking density by a preset threshold (usually set to 1.8), and then limit it within the range of 0 to 1. Exemplarily, if the cumulative value of material stacking density of the overlapping grid is 1.2, the spatial operation conflict degree of this part is 1.2 / 1.8 = 0.67.

[0172] For the spatial operation conflict analysis between positions A and B, since there are 10 overlapping grid cells in the core operation area of A, the spatial operation conflict degree of this part is 1; 15 overlapping grid cells are in the equipment occupancy area of A, the cumulative value of equipment occupancy weights of this part is 9.6, and the corresponding spatial operation conflict degree is 0.85; 15 overlapping grid cells are in the material stacking area of A, the cumulative value of material stacking density of this part is 12.3, and the corresponding spatial operation conflict degree is 0.72.

[0173] Taking into account the spatial operation conflict degrees of all parts, the maximum conflict degree is used as the spatial operation conflict degree between positions A and B, which is 1. This indicates that maintenance operations at positions A and B should not be carried out simultaneously.

[0174] Finally, organize the spatial operation conflict degrees between all maintenance positions 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 positions, and the element aij in the matrix represents the spatial operation conflict degree between positions i and j.

[0175] Map the spatial operation conflict degree to the preset interval [0, 1], and different threshold values can be set according to actual requirements to divide the conflict levels. For example, [0, 0.3) is defined as low conflict, and construction can be carried out simultaneously at the maintenance positions; [0.3, 0.7) is defined as medium conflict, and construction can be carried out simultaneously at the maintenance positions under specific conditions; [0.7, 1] is defined as high conflict, and construction should not be carried out simultaneously at the maintenance positions.

[0176] Taking five maintenance positions 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 optimization of the maintenance operation schedule as an important basis for judging whether construction can be carried out simultaneously at the maintenance positions.

[0183] In the field of engineering construction management, the existing construction space analysis technologies mainly adopt the regional division method based on the two-dimensional plane. Usually, the construction area is simply divided into several functional areas, without fully considering the mutual influence between the functional areas and its actual impact on the construction efficiency. The traditional technologies mainly rely on manual experience for rough space planning and conflict detection, lacking accurate quantitative analysis means, resulting in problems such as unreasonable allocation of space resources and unexpected conflicts between multiple maintenance positions during the actual construction process. The existing technologies usually use simple distance threshold judgment or overlapping area ratio calculation to evaluate the spatial conflict, which cannot accurately reflect the severity of the conflict between different functional areas and its actual impact on the construction.

[0184] From the perspective of improving the refined management of construction space resources, this embodiment discretizes the construction space by introducing a three-dimensional grid division method, and innovatively assigns two key parameters, namely equipment occupancy weight and material stacking density, to each grid unit, realizing an accurate quantitative expression of the utilization status of the construction space. Different from the traditional method, this embodiment distinguishes the different importance of three functional areas: the core operation area, the equipment occupancy area, and the material stacking area, and adopts different conflict calculation strategies according to the functional characteristics and importance of different areas. In particular, considering 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, this embodiment establishes a spatial operation conflict model that better conforms to the actual construction characteristics.

[0185] The starting point is to solve the problems of inaccurate conflict assessment and inability to reflect the actual construction impact degree in traditional construction space management. By introducing a grid-based spatial expression method and a conflict degree calculation method based on the characteristics of functional areas, this embodiment can more accurately quantify the spatial operation conflict degree between adjacent maintenance positions, and finally form a scientific and reasonable spatial operation constraint matrix.

[0186] This embodiment has remarkable technical effects, improving the accuracy of construction space planning, and being able to accurately identify potential spatial conflict problems at the construction preparation stage; by distinguishing the conflict calculation strategies of different functional areas, the spatial conflict assessment is more in line with the actual construction needs; the generated spatial operation constraint matrix provides reliable constraint conditions for subsequent construction scheduling optimization, helping to formulate a more reasonable construction plan; this method can effectively reduce the construction period delay and resource waste caused by spatial conflicts during the construction process, and improve the overall construction efficiency and quality.

[0187] The automatic generation system of the intelligent road surface maintenance plan in the embodiment of the present invention includes:

[0188] The first unit is used to obtain road surface image data and perform preprocessing, extract features and detect targets through a convolutional neural network, identify the type and location information of road surface diseases, and obtain the initial feature map of the disease area;

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

[0190] The third unit is used to match the basic maintenance plan based on the evaluation result of the road surface disease level;

[0191] The fourth unit is used to construct an internal domain relationship diagram for the basic maintenance plan, construct an external domain feature matrix by combining environmental monitoring data and resource status data, and through a hierarchical decision-making network enhanced by cross-attention, with the maintenance quality and resource utilization rate as the optimization objectives, and use the dual gradient algorithm to obtain an optimized maintenance plan;

[0192] The fifth unit is used to conduct spatio-temporal impact analysis on the optimized maintenance plan, determine the construction space conflict degree and traffic flow distribution by combining historical maintenance data, calculate the dynamic weight of the maintenance location based on the loss of traffic capacity, and conduct spatio-temporal optimization adjustment on the maintenance plan to generate the final pavement maintenance plan.

[0193] In the third aspect of the embodiments of the present invention,

[0194] There is provided an electronic device, comprising:

[0195] A processor;

[0196] A memory for storing instructions executable by the processor;

[0197] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0198] In the fourth aspect of the embodiments of the present invention,

[0199] There is provided a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0200] The present invention can be a method, device, system, and / or computer program product. The computer program product can include a computer-readable storage medium, on which computer-readable program instructions for executing various aspects of the present invention are uploaded.

[0201] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for automatically generating an intelligent pavement maintenance plan, characterized in that: include: Obtain pavement image data and perform preprocessing, perform feature extraction and target detection through convolutional neural networks, identify the type and location information of pavement damage, and obtain the initial feature map of the damage area; Based on the initial feature map of the damaged area, multi-scale grid adaptive feature extraction is used to obtain the geometric deformation characteristics of the damage, and the damage evolution is predicted by combining historical data. A dynamic evaluation vector is constructed for classification evaluation to obtain the pavement damage grade evaluation result. Match basic maintenance plans based on pavement damage grade assessment results; The basic maintenance plan is constructed as an internal domain relationship diagram, and the external domain feature matrix is ​​constructed by combining environmental monitoring data and resource status data. Through the cross-attention enhanced hierarchical decision network, the dual gradient algorithm is used to obtain the optimized maintenance plan with maintenance quality and resource utilization as the optimization goals. Conduct spatiotemporal impact analysis on the optimized maintenance plan, determine the construction space conflict degree and traffic flow distribution based on historical maintenance data, calculate the dynamic weight of the maintenance location based on capacity loss, perform spatiotemporal optimization and adjustment on the maintenance plan, and 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 characteristics of the disease, and the disease evolution is predicted in combination with historical data. A dynamic evaluation vector is constructed for classification evaluation, and the pavement disease grade evaluation results are obtained, including: Construct a multi-resolution image pyramid, extract features at different scales from the initial feature map of the diseased area, divide the structural unit grid at each scale, calculate the grid complexity score based on the pixel distribution characteristics within the structural unit grid, and adaptively adjust the feature extraction density of each structural unit grid according to the grid complexity score; Extracting a three-dimensional curvature descriptor for each structural unit grid, constructing a geometric deformation field of the diseased area based on the three-dimensional curvature descriptor, and extracting disease extension trend characteristics from the geometric deformation field; Input the disease expansion trend characteristics and historical monitoring data into the pre-trained recursive neural network to predict the disease development status at future time nodes, and combine the current disease characteristics with the disease development status to construct a dynamic evaluation vector; A multi-task learning classifier is trained based on the dynamic assessment vector, and the multi-task learning classifier simultaneously outputs the type of disease, the degree of disease and the development level. The type of disease, the degree of disease and the development level are weighted and fused according to the maintenance priority to obtain the pavement disease assessment result.

3. The method according to claim 2, characterized in that Extracting a three-dimensional curvature descriptor for each structural unit grid, constructing a geometric deformation field of the diseased area based on the three-dimensional curvature descriptor, and extracting disease extension trend features from the geometric deformation field include: Adaptively divide each structural unit grid, determine the grid division granularity based on the spatial distribution density of the diseased area, generate hierarchical grid nodes, perform local surface fitting on the three-dimensional point cloud data corresponding to the hierarchical grid nodes through the least squares method guided by the deep image, adaptively adjust the fitting weights in combination with the attention mechanism, generate local surface feature maps, and calculate the normal vector field and main direction field of the hierarchical grid nodes; Based on the normal vector field and the main direction field, the three-dimensional curvature descriptor is calculated at multiple receptive field scales, including the principal curvature, Gaussian curvature and mean curvature. The Transformer encoder is used to perform self-attention calculations on the curvature features at different scales to obtain global correlation features, and the hierarchical geometric features are constructed in combination with the local surface feature map. The hierarchical geometric features are input into a physically constrained graph convolutional network, a stress-strain mapping relationship is constructed based on predetermined material mechanical properties, the stress distribution field of the defect area is calculated through feature propagation between hierarchical grid nodes, a local deformation field is derived, and a geometric deformation field is determined; The geometric deformation field is subjected to space-time tensor decomposition, the principal strain direction and strain intensity are extracted, a deformation feature sequence is formed, and a preset time-series graph neural network is input to construct the disease extension trend feature.

4. The method according to claim 1, characterized in that The basic maintenance plan is constructed as an internal domain relationship diagram, and the external domain feature matrix is ​​constructed by combining environmental monitoring data and resource status data. Through the cross-attention enhanced hierarchical decision network, with maintenance quality and resource utilization as optimization goals, 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 configuration nodes. A topological connection structure between nodes is established according to the correlation degree of the maintenance process and the dependency relationship of resource configuration to form an internal domain relationship diagram; environmental monitoring data including temperature, humidity, and precipitation and resource status data including equipment, personnel, and materials are constructed as an external domain feature matrix; A multi-head self-attention mechanism is used to process the inner domain relationship graph and the outer domain feature matrix respectively, calculate the association weights between nodes and the influence degree between features, and generate inner domain attention representation and outer domain attention representation; Based on adaptive representation decomposition and temporal dependency analysis, the internal domain attention representation and the external domain attention representation are hierarchically processed, and the enhanced state vector is output through a bidirectional cross attention network. The enhanced state vector is input into a hierarchical decision network, a high-level strategy module of the hierarchical decision network generates a maintenance process combination scheme based on the enhanced state vector, and a low-level execution module of the hierarchical decision network determines a corresponding parameter configuration based on the maintenance process combination scheme; The dual gradient optimization algorithm is used to train the hierarchical decision network, a dual objective function including maintenance quality evaluation items and resource utilization evaluation items is constructed, the network parameters are optimized and updated alternately in the original space and the dual space, and the optimized maintenance plan is determined.

5. The method according to claim 4, characterized in that Based on adaptive representation decomposition and temporal dependency analysis, the internal domain attention representation and the external domain attention representation are hierarchically processed, and the enhanced state vector is output through the bidirectional cross attention network, including: Construct an adaptive representation decomposition unit to adaptively group the internal domain attention representation based on the process-related weights, adaptively group the external domain attention representation based on the degree of influence of resource allocation, calculate the information importance of different groups based on attention entropy, and form a hierarchical representation sequence; Construct a temporal dependency analysis unit, model the historical representation of process combinations and environmental resource states, calculate the temporal evolution characteristics corresponding to the internal and external domain attention representations, and output a time-varying weight matrix containing process combination weights and environmental resource weights; Each hierarchical representation sequence is bidirectionally fused through the cross attention module. Based on the time-varying weight matrix, the inner domain attention representation is mapped to the query matrix, and the outer domain attention representation is mapped to the key value matrix to calculate the positive attention score. The outer domain attention representation is mapped to the query matrix, and the inner domain attention representation is mapped to the key value matrix to calculate the reverse attention score, and the process-dominated positive interaction representation and resource-constrained reverse interaction representation are obtained respectively. A dynamic selection gating mechanism is used to generate feature selection signals based on the time-varying weight matrix and attention entropy, and the forward interaction representations and reverse interaction representations at each level are dynamically selected and adaptively fused to determine the fused representation. The inter-layer skip connection mechanism is used to cascade the fused representations at different levels, and the dependencies between the fused representations are calculated through self-attention to obtain the enhanced state vector.

6. The method according to claim 1, characterized in that The spatiotemporal impact analysis of the optimized maintenance plan is carried out. The construction space conflict degree and traffic flow distribution are determined in combination with historical maintenance data. The dynamic weight of the maintenance location is calculated based on the capacity loss. The maintenance plan is adjusted spatiotemporally to generate the final pavement maintenance plan, including: Extract maintenance time and location information from historical maintenance data, determine the construction impact range of each maintenance location and the traffic flow distribution at each maintenance time, and generate a spatiotemporal analysis data set; Based on the spatiotemporal analysis data set, the maintenance location is discretized into three-dimensional grids, the equipment occupancy and material stacking overlap of the grid units are calculated, and a spatial operation conflict model is constructed to obtain a spatial operation constraint matrix; Based on the spatiotemporal analysis data set and the spatial operation constraint matrix, traffic flow analysis is performed on the maintenance time in the optimized maintenance plan, the capacity loss of each maintenance time period is calculated after considering the spatial constraints, and the dynamic weight distribution of the maintenance location is determined; The maintenance operations in the optimized maintenance plan are prioritized according to the dynamic weight distribution of the maintenance locations. With the goal of maximizing traffic capacity, the optimal implementation period for each maintenance location is determined, the optimized maintenance plan is adjusted in time and space, and the final pavement maintenance plan is determined.

7. The method according to claim 6, characterized in that Based on the spatiotemporal analysis data set, the maintenance location is discretized into three-dimensional grids, the equipment occupancy and material stacking overlap of the grid unit are calculated, and the spatial operation conflict model is constructed to obtain the spatial operation constraint matrix including: According to the maintenance location information in the spatiotemporal analysis data set, the construction space of each maintenance location is divided into a core operation area, an equipment occupation area, and a material stacking area, and discretized into grid units through a three-dimensional grid division method, and an equipment occupancy weight and a material stacking density are assigned to each of the grid units; Based on the equipment occupancy weight and the material stacking density, the grid overlap degree between adjacent maintenance positions is calculated, wherein the grid overlap degree includes 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; A spatial operation conflict model is constructed according to the degree of grid overlap, and when the overlapping grids are located in the core operation area, the spatial operation conflict degree is set to the maximum value, when the overlapping grids are located in the equipment occupation area, the spatial operation conflict degree is calculated based on the cumulative value of the equipment occupation weight, and when the overlapping grids are 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 the preset interval to obtain the spatial operation constraint matrix.

8. An automatic generation system for intelligent pavement maintenance plans, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain and preprocess the pavement image data, perform feature extraction and target detection through a convolutional neural network, identify the type and location information of pavement damage, and obtain an initial feature map of the damage area; The second unit is used to obtain the geometric deformation characteristics of the disease based on the initial feature map of the disease area by using multi-scale grid adaptive feature extraction, predict the evolution of the disease in combination with historical data, construct a dynamic evaluation vector for classification evaluation, and obtain the pavement disease grade evaluation result; The third unit is used to match the basic maintenance plan based on the pavement disease grade assessment results; The fourth unit is used to construct the basic maintenance plan as an internal domain relationship diagram, combine environmental monitoring data and resource status data to construct an external domain feature matrix, and use a cross-attention enhanced hierarchical decision network to optimize the maintenance quality and resource utilization rate, and use a dual gradient algorithm to obtain the optimized maintenance plan; The fifth unit is used to conduct spatiotemporal impact analysis on the optimized maintenance plan, determine the construction space conflict degree and traffic flow distribution based on historical maintenance data, calculate the dynamic weight of the maintenance location based on the capacity loss, perform spatiotemporal optimization and adjustment on the maintenance plan, and generate the final pavement maintenance plan.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images