A method and system for predicting in-floor stress

Abstract

The application provides a floor internal stress prediction method and system, specifically, a floor image is acquired, finite element analysis is performed in combination with macroscopic load and boundary conditions to generate an initial stress field; visual key points in the image are extracted, key points in a high gradient area are selected as sparse key points according to a gradient of the initial stress field, and multi-modal feature representation is generated by fusing visual features and local stress features of the sparse key points; sparse key points are taken as nodes, a stress correlation graph with a weighted edge is constructed according to a spatial distance, the stress correlation graph and the multi-modal feature representation are input into a graph neural network, and stress values of each node are predicted; and a continuous internal stress distribution graph covering the entire floor is generated through a spatial interpolation algorithm.

Description

Technical Field

[0001] This application belongs to the field of prediction, and in particular relates to a method and system for predicting floor internal stress. Background Technology

[0002] The finite element method (FEM) discretizes a continuous structure into a finite number of elements and, based on continuum mechanics, establishes and solves a system of algebraic equations to obtain the stress and strain quantities of each part of the structure. However, for floor structures with geometric details or minor defects, obtaining the stress field requires dense meshing, consuming computational resources and time, making it difficult to meet the needs of rapid assessment. Data-driven methods, such as using convolutional neural network models, learn and predict stress distribution from images or geometric information of the structure. Purely data-driven models lack physical mechanism support, making it difficult to guarantee the reliability of prediction results. Moreover, the models require labeled data for training, and obtaining the training samples themselves requires costly simulations or experiments. Furthermore, existing methods fail to integrate macroscopic information and microscopic features of the structure, making it difficult for the models to detect stress concentration phenomena caused by local visual features. Therefore, how to combine the advantages of physical simulation and data-driven methods to achieve rapid prediction of floor internal stress is a technical challenge that urgently needs to be solved in the current field. Summary of the Invention

[0003] This invention proposes a method for predicting floor internal stress, addressing the problem that existing methods fail to combine the advantages of physical simulation and data-driven approaches to achieve rapid prediction of floor internal stress. The method includes: An image of the floor to be predicted is acquired, and finite element analysis is performed based on the image, macroscopic load, and boundary conditions to generate an initial stress scalar field. A set of initial visual keypoints carrying visual representations are extracted from the image; a gradient threshold is determined based on the gradient distribution of the initial stress scalar field, and initial visual keypoints with gradient magnitudes greater than the gradient threshold at corresponding positions in the initial stress scalar field are selected to form a set of sparse keypoints; for each sparse keypoint, the visual representation of the keypoint is fused with the stress features extracted from the initial stress scalar field at the corresponding position to generate a multimodal feature representation; The sparse keypoints are used as graph nodes. For each graph node, other graph nodes are connected within a preset spatial distance threshold to form graph edges. The stress correlation graph is constructed by using a function based on the spatial distance between the graph nodes as the weight of the connecting edges. The stress correlation graph and the multimodal feature representations corresponding to each graph node are input into the graph neural network model to obtain the stress prediction values ​​of each sparse keypoint. A spatial interpolation algorithm is used to generate a continuous internal stress distribution map covering the entire floor based on the location of the sparse key points and the corresponding stress prediction values. The stress distribution map is then corrected using the initial stress scalar of the non-sparse key points.

[0004] Furthermore, the present invention also relates to a floor internal stress prediction system, comprising the following modules: The first generation module is used to acquire an image of the floor to be predicted, and to perform finite element analysis based on the image, macroscopic load and boundary conditions to generate an initial stress scalar field. A fusion module is used to extract a set of initial visual key points carrying visual representations from the image; determine a gradient threshold based on the gradient distribution of the initial stress scalar field, and filter initial visual key points whose gradient magnitude at the corresponding position in the initial stress scalar field is greater than the gradient threshold to form a set of sparse key points; for each sparse key point, fuse the visual representation of the key point with the stress features extracted from the initial stress scalar field at the corresponding position to generate a multimodal feature representation; The input module is used to treat the sparse key points as graph nodes. For each graph node, it connects other graph nodes within a preset spatial distance threshold to form graph edges, and uses a function based on the spatial distance between the graph nodes as the weight of the connecting edges to construct a stress correlation graph. The stress correlation graph and the multimodal feature representations corresponding to each graph node are input into the graph neural network model to obtain the stress prediction values ​​of each of the sparse key points. The second generation module is used to generate a continuous internal stress distribution map covering the entire floor using a spatial interpolation algorithm based on the location of the sparse key points and the corresponding stress prediction values, and to correct the stress distribution map using the initial stress scalar of the non-sparse key points.

[0005] This invention utilizes finite element analysis to generate an initial stress field, providing a physical basis for prediction. By selecting key points in high stress gradient regions and constructing a stress correlation map, computational resources are concentrated on critical areas of the structure, improving computational efficiency. Visual representations extracted from images are fused with stress features extracted from the initial stress field to generate multimodal feature representations. This enables graph neural network models to detect localized stress concentrations caused by microscopic defects or texture visual features on the floor surface, while avoiding the influence of finite element analysis or a single GNN mode on the results. This invention can reduce computational costs while maintaining prediction accuracy, achieving rapid and reliable prediction of stress distribution within the floor. Attached Figure Description

[0006] Figure 1 A flowchart of the first embodiment; Figure 2This is a schematic diagram of multimodal feature representation sub-fusion. Figure 3 A schematic diagram is constructed for the stress correlation diagram; Figure 4 This is for graph neural network model processing. Detailed Implementation

[0007] The features and exemplary embodiments of various aspects of this application will be described in detail below. To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain this application and not to limit it. For those skilled in the art, this application can be implemented without some of these specific details. The following description of the embodiments is merely to provide a better understanding of this application by illustrating examples.

[0008] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0009] In the first embodiment, the present invention proposes a method for predicting floor internal stress, such as... Figure 1 ,include: S1. Obtain an image of the floor to be predicted, and perform finite element analysis based on the image, macroscopic load, and boundary conditions to generate an initial stress scalar field; A high-resolution industrial camera is used to capture a high-resolution digital image of the floor surface to be predicted from directly above. Image processing techniques are used to extract the floor's planar contour and dimensions from this image. Combined with the floor's thickness and material properties, this image is imported into finite element analysis software such as ANSYS or ABAQUS to create a three-dimensional geometric model. A coarse mesh is created on this model, for example, using large tetrahedral elements. Macroscopic loads are set according to actual working conditions, such as applying a uniformly distributed pressure of 5000 Newtons per square meter to simulate household or pedestrian loads. Boundary conditions are also set, such as setting fixed constraints on the edges of the model. A finite element solver is run to perform structural analysis, calculating the initial stress distribution of the entire model. The von Mises equivalent stress values ​​of all nodes on the upper surface of the model are extracted, and a two-dimensional initial stress scalar field is generated based on its planar coordinate mapping. After acquiring the floor image, a transformation relationship between the image pixel coordinates and the actual physical coordinates of the floor needs to be established using a calibration plate or known floor physical dimensions. The finite element analysis is performed in the same physical coordinate system, ensuring that the position of each data point in the generated initial stress scalar field corresponds to the position of the image pixel.

[0010] S2, extract a set of initial visual key points carrying visual representations from the image; determine a gradient threshold based on the gradient distribution of the initial stress scalar field, and filter initial visual key points whose gradient magnitude at the corresponding position in the initial stress scalar field is greater than the gradient threshold to form a set of sparse key points; for each sparse key point, fuse the visual representation of the key point with the stress features extracted from the initial stress scalar field at the corresponding position to generate a multimodal feature representation; For the acquired high-resolution floor image, the Scale Invariant Feature Transform (SIFT) algorithm is used to detect stable feature points such as corners and edges, generating thousands of initial visual keypoints. A 128-dimensional visual representation sub-vector is calculated for each keypoint. Simultaneously, the gradient of the initial stress scalar field generated in the previous step is calculated using operators such as Sobel in the x and y directions, thus obtaining the gradient magnitude at each location. The gradient magnitudes at all locations are statistically analyzed, sorted in ascending order, and the gradient magnitude at the 90th percentile is selected as the gradient threshold. All initial visual keypoints are iterated through, mapping their coordinates in the image to their corresponding positions in the initial stress scalar field. If the stress gradient magnitude at a given location is greater than the determined gradient threshold, the keypoint is retained; otherwise, it is discarded. The resulting set of fewer keypoints is the sparse keypoint set.

[0011] The visual representation incorporates information about the material inhomogeneity of the floor, such as wood knots, grain direction, or surface microcracks. This information affects stress, and fusing the two helps improve prediction accuracy. For each selected sparse keypoint, the 128-dimensional SIFT visual representation generated in the previous step is obtained. Based on the keypoint's location coordinates, the corresponding stress feature is searched and extracted from the initial stress scalar field data. This stress feature is a vector containing three elements, representing the von Mises equivalent stress value, the stress gradient in the x-direction, and the stress gradient in the y-direction at that point. The 128-dimensional visual representation is concatenated with the 3-dimensional stress feature vector to form a new 131-dimensional feature vector, which is the multimodal feature representation of the sparse keypoint. Optionally, the initial stress scalar of the sparse keypoint can be directly used as the multimodal feature representation.

[0012] In an optional embodiment, determining the gradient threshold based on the gradient distribution of the initial stress scalar field includes: Calculate the gradient magnitudes at all locations in the initial stress scalar field, sort all gradient magnitudes in ascending order, and select the gradient magnitude at the first preset quantile as the gradient threshold.

[0013] Specifically, the initial stress scalar field is considered as a two-dimensional matrix, where each element represents the stress value at a location. To calculate the gradient magnitude, the numerical differentiation method of the Sobel operator is applied at each pixel location to calculate the gradient components in the horizontal direction (X-direction). and the gradient components in the vertical direction, i.e., the Y-direction. The gradient magnitude at each location is calculated using the two components. This process generates a gradient magnitude map with the same dimensions as the original stress field.

[0014] The gradient magnitude values ​​of all pixels in the gradient magnitude map are collected to obtain a one-dimensional numerical list. For example, for a stress field of size 200 pixels × 200 pixels, a list containing 40,000 gradient magnitude values ​​will be obtained. All values ​​in this list are sorted in ascending order. A threshold is selected based on a preset first quantile. Assuming the first preset quantile is the 90th percentile, the value at position 36,000 in the sorted list will be selected as the gradient threshold. If this value is 15.3, then all pixels with a gradient magnitude greater than 15.3 will be initially identified as candidate regions for key points with drastic stress changes.

[0015] To create a comprehensive feature representation that integrates physical and visual information for each selected sparse keypoint, in an optional embodiment, for each sparse keypoint, the visual representation of the keypoint is fused with the corresponding stress features extracted from the initial stress scalar field to generate a multimodal feature representation, including: The stress scalar value, stress gradient X component, and stress gradient Y component at the corresponding positions of the sparse key points are extracted to form a stress feature vector, and the visual representation of the sparse key points is concatenated with the stress feature vector.

[0016] For any sparse keypoint, record its x and y coordinates in the image. Using these coordinates, retrieve the stress scalar value of that point from the initial stress scalar field data, and retrieve the stress gradient X and Y components of that point from the pre-calculated stress gradient field. These three values, for example, stress value 25.8, gradient X component 3.1, and gradient Y component -1.7, are combined to form a three-dimensional stress feature vector [25.8, 3.1, -1.7].

[0017] Simultaneously, visual images of the same region are processed using the SIFT feature extraction algorithm to extract visual features at the same locations as sparse keypoints, generating a high-dimensional visual representation. For example, a SIFT visual representation is typically a 128-dimensional vector. The feature vectors of the two different modalities are concatenated. A three-dimensional stress feature vector is appended to the end of the 128-dimensional visual representation vector, resulting in a 131-dimensional multimodal feature representation. This new representation contains both the object's appearance texture information and its internal mechanical state information. See [link to documentation]. Figure 2 .

[0018] S3, the sparse key points are used as graph nodes. For each graph node, other graph nodes are connected within a preset spatial distance threshold to form graph edges. The stress correlation graph is constructed by using a function based on the spatial distance between the graph nodes as the weight of the connected edges. The stress correlation graph and the multimodal feature representations corresponding to each graph node are input into the graph neural network model to obtain the stress prediction values ​​of each sparse key point. Each sparse keypoint is abstracted as a node in a graph. A spatial distance threshold is set, for example, 50mm. All node pairs are traversed, and the Euclidean distance between any two nodes in real physical space is calculated. If the distance between two nodes is less than 50mm, an edge is created between them. For each created edge, the edge weight is calculated using a Gaussian kernel function. Through these operations, all sparse keypoints and their connections together constitute a weighted undirected graph, i.e., a stress correlation graph. (See [link to relevant documentation]). Figure 3 .

[0019] The adjacency matrix or edge list of the constructed stress correlation graph, along with the 131-dimensional multimodal feature representations corresponding to all graph nodes, are used as input to a pre-trained Graph Convolutional Network (GCN) model. This model contains multiple graph convolutional layers. In each layer, each node updates its own feature representation by aggregating the feature information of its neighbors, thereby learning the stress transmission and influence patterns between nodes. After multiple layers of propagation and transformation, the model's output layer processes the feature representation of each node, outputting a single scalar value for each node. This value is the predicted stress value at the corresponding sparse keypoint location.

[0020] In an optional embodiment, the step of connecting other graph nodes within a preset spatial distance threshold to form graph edges includes: For each graph node, calculate the Euclidean distance between the graph node and all other graph nodes. When the Euclidean distance is less than the preset spatial distance threshold, establish a connection edge between the two graph nodes.

[0021] Specifically, all sparse keypoints are abstracted as nodes in a graph structure, and the position of each node is determined by its physical coordinates (x, y) in two-dimensional space. The graph construction process starts with one node and iterates through all other nodes, calculating the straight-line distances between it and all other nodes. These distances are calculated using the Euclidean distance formula.

[0022] A spatial distance threshold, such as 50mm, is preset. After calculating the Euclidean distance between any two nodes, this distance value is compared with the preset threshold. If the calculated distance is less than 50, a connection edge is created between the two nodes, indicating that they are spatially adjacent. Conversely, if the distance is greater than or equal to 50, no connection is established. For example, node A has coordinates (10,20), node B has coordinates (30,50), and node C has coordinates (80,90). The distance between A and B is approximately 36, which is less than 50, so A and B are connected. The distance between A and C is approximately 99, which is greater than 50, so no connection is established. This process is repeated for all node pairs in the graph, forming an undirected graph where edges represent the spatial proximity relationships between key points.

[0023] In an optional embodiment, using a function based on the spatial distance between the graph nodes as the weight of the connecting edges includes: The Gaussian kernel function is used to calculate the edge weights. ,in Let be the weight of the edge connecting graph node i and graph node j. Let be the Euclidean distance between two graph nodes. These are preset scale parameters.

[0024] After establishing the edge structure of the graph, a weight is assigned to each edge. This is done using a Gaussian kernel function. For any edge connecting node i and node j in the graph, the Euclidean distance between them is obtained. The distance value is the core input for calculating the weights.

[0025] Use preset scale parameters and distance Substitute these values ​​into the Gaussian kernel function formula for calculation. Parameters The rate at which the weights decay with distance is controlled. For example, set... The value is 20. If the distance between two nodes... If the distance between two nodes is 10, the edge weight is approximately 0.88. If the distance between the other pair of nodes is 40, the weight is approximately 0.135. The results indicate that the closer the node pairs are, the higher the edge weight, and the greater their mutual influence. After calculating the weights of all edges, a weighted graph is obtained.

[0026] In an optional embodiment, the graph neural network model is a graph convolutional network model, and the model network structure includes: at least one graph convolutional layer for aggregating neighborhood node features, and at least one fully connected layer for mapping the aggregated features to stress prediction values.

[0027] Specifically, the graph convolutional network model consists of two graph convolutional layers and a multilayer perceptron regression head composed of two fully connected layers. Each graph convolutional layer and fully connected layer is followed by a ReLU activation function, except for the last layer. The model input is graph data, including a node feature matrix and a weighted adjacency matrix. Each row of the node feature matrix is ​​a multimodal feature representation of the corresponding keypoint, and the adjacency matrix defines the connection relationships between nodes and the weights of edges. The model output is a stress prediction vector, with the same dimension as the number of nodes in the input graph. Each element represents the predicted stress value of the corresponding keypoint. The training set of this model consists of a large number of samples generated through finite element simulation. Each sample contains an initial stress field and a corresponding true stress field, which are used to construct the graph data and labels required for training. The training process adopts supervised learning, optimizing the model parameters by minimizing the mean squared error between the model's predicted values ​​and the true stress labels. The mean squared error loss function is... ,in This represents the actual stress value. To predict stress values, the optimizer uses the Adam algorithm for parameter updates.

[0028] The adopted graph convolutional network model takes a previously constructed weighted graph as input, where each node carries a multimodal feature representation. The core of the model is the graph convolutional layer. In the graph convolutional layer, each node aggregates information from its neighbors. For a central node, the feature vectors of all its neighbors are weighted and summed or averaged according to the weights of the connecting edges. The aggregated neighborhood features are combined with the central node's own features, and through a learnable linear transformation matrix and activation function, such as the ReLU function, a new feature representation for that node in the next layer is generated. By stacking multiple graph convolutional layers, each node can perceive information from nodes further away in the graph.

[0029] After information is propagated and updated through one or more graph convolutional layers, each node obtains a high-dimensional feature vector containing rich contextual information. To obtain the stress prediction value, this high-dimensional feature vector is input into at least one fully connected layer at the end of the network. This fully connected layer linearly maps the input feature vector to a single numerical output. The output value is the model's prediction of the stress magnitude at the node's location, such as... Figure 4 The entire model is trained end-to-end by minimizing the error between the predicted stress and the actual stress.

[0030] S4. Using a spatial interpolation algorithm, a continuous internal stress distribution map covering the entire floor is generated based on the location of the sparse key points and the corresponding stress prediction values. The stress distribution map is then corrected using the initial stress scalar of the non-sparse key points.

[0031] The two-dimensional spatial coordinates of all sparse keypoints and their corresponding stress prediction values ​​output by a graph neural network model are used as the input data set. Kriging interpolation is employed to analyze the input data set, calculate and fit a semivariance function, and establish a spatial correlation model for stress values. Based on this model, the stress estimate for each grid point is calculated on a predefined grid across the entire floor area. The estimate for each grid point is a weighted average of the stress prediction values ​​of the surrounding sparse keypoints, with the weights determined by the spatial correlation model. The stress estimates of all grid points collectively constitute a high-resolution, continuous, and smooth internal stress distribution map, which can be visualized using pseudo-color rendering. In one embodiment, the method further includes correcting the stress distribution map using the initial stress scalar of non-sparse key points. Specifically, the coordinate positions of the non-sparse key points and their corresponding initial stress scalars are obtained; the interpolated stress values ​​of the generated continuous internal stress distribution map at the coordinate positions are read, and the difference between the initial stress scalar and the interpolated stress values ​​is calculated to obtain the correction residual; based on the correction residuals of all non-sparse key points, a global residual compensation field covering the entire floor is constructed using an interpolation algorithm; the global residual compensation field is superimposed on the continuous internal stress distribution map to obtain the corrected internal stress distribution map. In another embodiment, an initial background stress map covering the entire floor is generated based on the initial stress scalar field. A distance field matrix covering the entire floor is calculated, centered on the location of the sparse keypoints. Each element in the distance field matrix represents the Euclidean distance from that location to the nearest sparse keypoint. A fusion weight map is constructed based on the distance field matrix, where locations with smaller distances correspond to larger fusion weights. Using the fusion weight map, the continuous internal stress distribution map and the initial background stress map are weighted and summed, where a larger fusion weight corresponds to a higher proportion of the continuous internal stress distribution map, resulting in a corrected final internal stress distribution map. These two methods can integrate finite element analysis and GNN prediction results, reducing errors caused by a single method.

[0032] In an optional embodiment, the spatial interpolation algorithm is the Kriging interpolation algorithm. The algorithm takes the position of the sparse keypoint and the corresponding stress prediction value as input, calculates the spatial autocorrelation between the point to be interpolated and all sparse keypoints, and performs a weighted average of all pixels in the floor area to be predicted to obtain the stress value of each pixel.

[0033] To generate a continuous stress distribution map covering the entire floor area from the stress values ​​at sparse keypoints output by a graph neural network, Kriging interpolation is employed. The algorithm's input consists of two parts: a set of two-dimensional coordinates for all sparse keypoints, and the corresponding stress values ​​predicted by the graph neural network at those coordinates. Kriging interpolation utilizes the spatial correlation of known data points to predict the values ​​at unknown points.

[0034] The spatial structure of known data points is analyzed by constructing a mathematical model representing how stress values ​​vary with distance through a semivariogram. This model represents the spatial autocorrelation between data points. For each pixel within the region to be predicted, the Kriging algorithm calculates an optimal weighted average to estimate the stress value. The weights are determined by solving a system of equations related to the semivariogram model. These weights utilize the distances between the interpolation point and each known keypoint, as well as the spatial relationships between all known keypoints, to ensure that the interpolation result is unbiased and has minimal estimation variance. By repeating this process for all pixels within the region, a complete and smooth high-resolution stress contour map is generated.

[0035] In a second embodiment, the present invention also proposes a floor internal stress prediction system, comprising the following modules: The first generation module is used to acquire an image of the floor to be predicted, and to perform finite element analysis based on the image, macroscopic load and boundary conditions to generate an initial stress scalar field. A fusion module is used to extract a set of initial visual key points carrying visual representations from the image; determine a gradient threshold based on the gradient distribution of the initial stress scalar field, and filter initial visual key points whose gradient magnitude at the corresponding position in the initial stress scalar field is greater than the gradient threshold to form a set of sparse key points; for each sparse key point, fuse the visual representation of the key point with the stress features extracted from the initial stress scalar field at the corresponding position to generate a multimodal feature representation; The input module is used to treat the sparse key points as graph nodes. For each graph node, it connects other graph nodes within a preset spatial distance threshold to form graph edges, and uses a function based on the spatial distance between the graph nodes as the weight of the connecting edges to construct a stress correlation graph. The stress correlation graph and the multimodal feature representations corresponding to each graph node are input into the graph neural network model to obtain the stress prediction values ​​of each of the sparse key points. The second generation module is used to generate a continuous internal stress distribution map covering the entire floor using a spatial interpolation algorithm based on the location of the sparse key points and the corresponding stress prediction values, and to correct the stress distribution map using the initial stress scalar of the non-sparse key points.

[0036] In an optional embodiment, determining the gradient threshold based on the gradient distribution of the initial stress scalar field includes: Calculate the gradient magnitudes at all locations in the initial stress scalar field, sort all gradient magnitudes in ascending order, and select the gradient magnitude at the first preset quantile as the gradient threshold.

[0037] In an optional embodiment, for each of the sparse keypoints, fusing the visual representation of the keypoint with the corresponding stress features extracted from the initial stress scalar field to generate a multimodal feature representation includes: The stress scalar value, stress gradient X component, and stress gradient Y component at the corresponding positions of the sparse key points are extracted to form a stress feature vector, and the visual representation of the sparse key points is concatenated with the stress feature vector.

[0038] In an optional embodiment, the step of connecting other graph nodes within a preset spatial distance threshold to form graph edges includes: For each graph node, calculate the Euclidean distance between the graph node and all other graph nodes. When the Euclidean distance is less than the preset spatial distance threshold, establish a connection edge between the two graph nodes.

[0039] In an optional embodiment, using a function based on the spatial distance between the graph nodes as the weight of the connecting edges includes: The Gaussian kernel function is used to calculate the edge weights. ,in Let be the weight of the edge connecting graph node i and graph node j. Let be the Euclidean distance between two graph nodes. These are preset scale parameters.

[0040] In an optional embodiment, the graph neural network model is a graph convolutional network model, and the model network structure includes: at least one graph convolutional layer for aggregating neighborhood node features, and at least one fully connected layer for mapping the aggregated features to stress prediction values.

[0041] In an optional embodiment, the spatial interpolation algorithm is the Kriging interpolation algorithm. The algorithm takes the position of the sparse keypoint and the corresponding stress prediction value as input, calculates the spatial autocorrelation between the point to be interpolated and all sparse keypoints, and performs a weighted average of all pixels in the floor area to be predicted to obtain the stress value of each pixel.

[0042] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0043] The functional modules shown in the above-described block diagram can be implemented as hardware, software, firmware, or a combination thereof. When implemented in hardware, they can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. Programs or code segments can be stored on a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried on a carrier wave. "Machine-readable medium" can include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, fiber optic media, radio frequency (RF) links, etc. Code segments can be downloaded via computer networks such as the Internet, intranets, etc.

[0044] It should also be noted that the exemplary embodiments mentioned in this application describe methods or systems based on a series of steps or apparatus. However, this application is not limited to the order of the above steps; that is, the steps can be performed in the order mentioned in the embodiments, or in a different order, or several steps can be performed simultaneously.

[0045] The aspects of this application have been described above with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It should be understood that each block in the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that these instructions, executable via the processor of the computer or other programmable data processing apparatus, enable the implementation of the functions / actions specified in one or more blocks of the flowchart illustrations and / or block diagrams. Such a processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor, or a field-programmable logic circuit. It is also understood that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can also be implemented by dedicated hardware performing the specified functions or actions, or can be implemented by a combination of dedicated hardware and computer instructions.

[0046] The above description is merely a specific implementation of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the protection scope of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the protection scope of this application.

Claims

1. A method for predicting floor internal stress, characterized in that, Includes the following steps: An image of the floor to be predicted is acquired, and finite element analysis is performed based on the image, macroscopic load, and boundary conditions to generate an initial stress scalar field. Extract a set of initial visual key points carrying visual representations from the image; A gradient threshold is determined based on the gradient distribution of the initial stress scalar field, and initial visual key points with gradient magnitudes greater than the gradient threshold at corresponding positions in the initial stress scalar field are selected to form a set of sparse key points. For each of the sparse keypoints, the visual representation of the keypoint is fused with the stress features at the corresponding position extracted from the initial stress scalar field to generate a multimodal feature representation. The sparse keypoints are used as graph nodes. For each graph node, other graph nodes are connected within a preset spatial distance threshold to form graph edges. The stress correlation graph is constructed by using a function based on the spatial distance between the graph nodes as the weight of the connecting edges. The stress correlation graph and the multimodal feature representations corresponding to each graph node are input into the graph neural network model to obtain the stress prediction values ​​of each sparse keypoint. A spatial interpolation algorithm is used to generate a continuous internal stress distribution map covering the entire floor based on the location of the sparse key points and the corresponding stress prediction values. The stress distribution map is then corrected using the initial stress scalar of the non-sparse key points.

2. The method according to claim 1, characterized in that, Determining the gradient threshold based on the gradient distribution of the initial stress scalar field includes: Calculate the gradient magnitudes at all locations in the initial stress scalar field, sort all gradient magnitudes in ascending order, and select the gradient magnitude at the first preset quantile as the gradient threshold.

3. The method according to claim 1, characterized in that, For each of the sparse keypoints, the visual representation of the keypoint is fused with the stress features at the corresponding position extracted from the initial stress scalar field to generate a multimodal feature representation, including: The stress scalar value, stress gradient X component, and stress gradient Y component at the corresponding positions of the sparse key points are extracted to form a stress feature vector, and the visual representation of the sparse key points is concatenated with the stress feature vector.

4. The method according to claim 1, characterized in that, The step of connecting other graph nodes within a preset spatial distance threshold to form graph edges includes: For each graph node, calculate the Euclidean distance between the graph node and all other graph nodes. When the Euclidean distance is less than the preset spatial distance threshold, establish a connection edge between the two graph nodes.

5. The method according to claim 1, characterized in that, The step of using a function based on the spatial distance between the graph nodes as the weight of the connecting edges includes: The Gaussian kernel function is used to calculate the edge weights. ,in Let be the weight of the edge connecting graph node i and graph node j. Let be the Euclidean distance between two graph nodes. These are preset scale parameters.

6. The method according to claim 1, characterized in that, The graph neural network model is a graph convolutional network model. The network structure of the model includes: at least one graph convolutional layer for aggregating features of neighboring nodes, and at least one fully connected layer for mapping the aggregated features to stress prediction values.

7. The method according to claim 1, characterized in that, The spatial interpolation algorithm is the Kriging interpolation algorithm. The algorithm takes the position of the sparse keypoint and the corresponding stress prediction value as input, calculates the spatial autocorrelation between the point to be interpolated and all sparse keypoints, and performs a weighted average of all pixels in the floor area to be predicted to obtain the stress value of each pixel.

8. A floor internal stress prediction system, characterized in that, Includes the following modules: The first generation module is used to acquire an image of the floor to be predicted, and to perform finite element analysis based on the image, macroscopic load and boundary conditions to generate an initial stress scalar field. The fusion module is used to extract a set of initial visual key points carrying visual representations on the image; A gradient threshold is determined based on the gradient distribution of the initial stress scalar field, and initial visual key points with gradient magnitudes greater than the gradient threshold at corresponding positions in the initial stress scalar field are selected to form a set of sparse key points. For each of the sparse keypoints, the visual representation of the keypoint is fused with the stress features at the corresponding position extracted from the initial stress scalar field to generate a multimodal feature representation. The input module is used to treat the sparse key points as graph nodes. For each graph node, it connects other graph nodes within a preset spatial distance threshold to form graph edges, and uses a function based on the spatial distance between the graph nodes as the weight of the connecting edges to construct a stress correlation graph. The stress correlation graph and the multimodal feature representations corresponding to each graph node are input into the graph neural network model to obtain the stress prediction values ​​of each of the sparse key points. The second generation module is used to generate a continuous internal stress distribution map covering the entire floor using a spatial interpolation algorithm based on the location of the sparse key points and the corresponding stress prediction values, and to correct the stress distribution map using the initial stress scalar of the non-sparse key points.

9. The system according to claim 8, characterized in that, Determining the gradient threshold based on the gradient distribution of the initial stress scalar field includes: Calculate the gradient magnitudes at all locations in the initial stress scalar field, sort all gradient magnitudes in ascending order, and select the gradient magnitude at the first preset quantile as the gradient threshold.

10. The system according to claim 8, characterized in that, For each of the sparse keypoints, the visual representation of the keypoint is fused with the stress features at the corresponding position extracted from the initial stress scalar field to generate a multimodal feature representation, including: The stress scalar value, stress gradient X component, and stress gradient Y component at the corresponding positions of the sparse key points are extracted to form a stress feature vector, and the visual representation of the sparse key points is concatenated with the stress feature vector.

Images