Machine learning-based concrete strength sampling detection optimization method and system

Abstract

The application discloses a concrete strength sampling detection optimization method and system based on machine learning, relates to the technical field of building detection, and comprises the following steps: mapping concrete member data to a physical knowledge graph, generating an initial member graph containing node associated edges and weights based on physical constraints; extracting graph node features by using a heterogeneous graph attention network, dividing candidate test batch communities according to the features, and calculating a comprehensive risk index; detecting a signal-to-noise ratio by model calculation, generating a detection pseudo node in a low signal-to-noise ratio area, and establishing a target function that maximizes the risk index and contains a pseudo node repulsion constraint; solving the target function to obtain an optimal test batch division; and finally performing dynamic path planning under the constraint by using a graph reinforcement learning algorithm, and outputting a best detection scheme containing an ordered sampling sequence. The application realizes intelligent risk division of concrete test batches and dynamic planning of an optimal sampling detection path.

Description

Technical Field

[0001] This invention relates to the field of building inspection technology, and in particular to an optimized method and system for concrete strength sampling inspection based on machine learning. Background Technology

[0002] Accurate sampling and testing of concrete strength is a core step in ensuring the safety and quality acceptance of engineering projects. Traditional sampling methods mostly rely on manual experience or simple probability and statistical rules, treating each component inside the building as an isolated entity. This severely disrupts the complex topological relationships and underlying mechanical transmission paths of the overall building space, leading to extremely one-sided risk assessments. Furthermore, real engineering sites often face complex physical obstacles such as densely packed rebar, excessively thick plaster layers, or localized defects, causing significant signal attenuation in non-destructive testing instruments.

[0003] In recent years, some machine learning and graph clustering algorithms have been introduced into the field of building structure inspection. However, most of the existing intelligent sampling algorithms are pure data-driven black box models, which seriously lack the underlying integration of legal acceptance standards and physical mechanics knowledge. They cannot impose reasonable constraints on network topology and output discrete, disordered static points, which are difficult to guide efficient scheduling on site. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an optimized method for concrete strength sampling and testing based on machine learning, which solves the technical problems of inaccurate mechanical assessment and on-site physical blind spots in traditional concrete sampling and testing.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a machine learning-based method for optimizing concrete strength sampling and testing, which includes acquiring basic data of the concrete component to be tested and mapping it to a preset concrete physical knowledge graph.

[0008] Based on the physical constraints of the concrete physical knowledge graph, the heterogeneous association edges and edge weights between component nodes are determined to obtain the initial component graph.

[0009] The initial component graph is input into a heterogeneous graph attention network, and node feature representations are extracted using a specific attention head that matches the node type. Based on the feature representations, candidate test batch communities are divided and a comprehensive risk index is calculated.

[0010] The detection signal-to-noise ratio is quantitatively calculated using a detection sensitivity model, and a detection pseudo-node is generated within the candidate test batch community where the detection signal-to-noise ratio is lower than a preset sensitivity threshold. An objective function is established to maximize the comprehensive risk index, and a detection pseudo-node exclusion constraint is applied.

[0011] Solving the objective function determines the optimal batch partitioning result, and using a graph reinforcement learning algorithm to perform dynamic path planning under the constraint of the exclusion of the probe pseudo-nodes, outputting a detection scheme containing ordered sampling sequences;

[0012] The candidate test batch community is a preliminary set of components with homogeneous mechanical properties that are clustered in the network topology based on the feature representation and are to be merged for risk verification; the probe pseudo-node is a virtual physical location point within the community that meets the preset detection sensitivity condition and is mapped to the network topology of the initial component graph.

[0013] As a preferred embodiment of the machine learning-based concrete strength sampling and testing optimization method of the present invention, the basic data of the concrete component to be tested includes numerical testing data, entity attribute data, construction process data, and spatial topology data.

[0014] The numerical detection data is used to calculate feature similarity, including rebound value and carbonization depth;

[0015] The entity attribute data is used as an identifier for entity link matching, including component type and design strength level;

[0016] The construction process data is used to trigger physical constraint verification, including the pouring date and curing method;

[0017] The spatial topology data is used to determine the mechanical transmission path, including the connection between the floor and the structure.

[0018] As a preferred embodiment of the machine learning-based concrete strength sampling and detection optimization method of the present invention, wherein: obtaining the initial component diagram includes calculating the feature similarity between nodes based on the basic data of the component nodes;

[0019] When the feature similarity meets the preset association threshold, initial association edges and initial edge weights are generated between the corresponding component nodes to construct the initial network topology;

[0020] Based on the entity attribute data of each component node, the entity of the component node is linked to the corresponding entity in the concrete physics knowledge graph;

[0021] For connected node pairs with the initial associated edge, query the concrete physics knowledge graph to see if there are any preset physical attribute constraints between the entities mapped by the connected node pairs.

[0022] Based on the queried physical attribute constraints, the initial associated edges and their weights are verified, and the corresponding graph structure correction operation is performed.

[0023] If the physical property constraint determines that the connected node pair has a mechanical feature conflict due to incompatibility of component types, then a preset penalty coefficient is introduced to attenuate and reduce the weight of the initial edge.

[0024] If the physical property constraint determines that the connected node pair has a mechanical feature conflict that breaks the mechanical transmission path, then the corresponding initial associated edge is deleted.

[0025] The initial network topology after the graph structure correction operation is used as the initial component graph.

[0026] As a preferred embodiment of the machine learning-based concrete strength sampling and testing optimization method of the present invention, the method of dividing candidate inspection batch communities includes mapping the basic data in the initial component diagram to the initial node vector, and calling the corresponding specific attention head for each type of heterogeneous associated edge, and extracting the local feature vector of each component node by calculating the node-level attention weight.

[0027] Calculate the semantic-level attention weights corresponding to different heterogeneous association edge types, perform global weighted fusion on the local feature vectors, and output the feature representations of each component node;

[0028] Based on the feature representation, calculate the feature similarity between connected node pairs with heterogeneous edges in the initial component graph;

[0029] The feature similarity is weighted and fused with the corresponding edge weight in the initial component graph to calculate the comprehensive association weight between the connected node pairs.

[0030] With the goal of maximizing the modularity of the network topology, graph clustering is performed on each component node in the initial component graph based on the comprehensive association weight;

[0031] When the modularity of the network topology reaches the preset convergence condition, clustering stops, and the set of each component node formed in the current network topology is taken as the corresponding candidate check batch community.

[0032] As a preferred embodiment of the machine learning-based concrete strength sampling and testing optimization method of the present invention, the calculation of the comprehensive risk index includes: extracting the feature representation of each component node in the candidate inspection batch community, and calculating the mean and variance of the feature distribution of the candidate inspection batch community in the feature space;

[0033] Substituting the mean of the characteristic distribution into the preset probability density function of concrete strength, and combining it with the strength qualification threshold, the risk value of misjudgment by the producer is calculated. Risk value missed by the user ;

[0034] Based on the topological position of the candidate test batch communities in the initial component graph, the corresponding structure sensitivity penalty coefficient is obtained using the network topology centrality algorithm. ;

[0035] By analyzing the risk value of misjudgment by the aforementioned producer Risk value of missed detection by the user Weighted fusion is performed, and the structure-sensitive penalty coefficient is utilized. Perform a multiplicative correction with the variance of the characteristic distribution to output the comprehensive risk index.

[0036] As a preferred embodiment of the machine learning-based concrete strength sampling detection optimization method of the present invention, the detection sensitivity model is a mathematical evaluation model used to quantify the detection signal-to-noise ratio at a specified spatial location point based on the attenuation law of the detection signal according to the physical attenuation characteristic data of the component, including a feature mapping layer, an attenuation calculation layer and an evaluation quantization layer.

[0037] Based on the feature mapping layer, the physical attenuation feature data of each spatial location point within the candidate test batch community are obtained;

[0038] The physical attenuation characteristic data includes the surface carbonization depth. Reinforcing bar density and additional surface layer thickness ;

[0039] Based on the attenuation calculation layer, the physical attenuation characteristic data, combined with corresponding coefficients, are used as energy dissipation independent variables to measure the preset reference detection signal strength. Perform a physical simulation of negative exponential decay to calculate the effective arrival signal energy after penetrating the barrier;

[0040] Based on the aforementioned evaluation quantization layer, the effective arrival signal energy is compared with the ambient background noise intensity. Perform ratio calculations and output the detection signal-to-noise ratio for each of the aforementioned spatial locations. .

[0041] As a preferred embodiment of the machine learning-based concrete strength sampling and detection optimization method of the present invention, the method of applying the detection pseudo-node exclusion constraint includes generating the detection pseudo-node at the coordinates of the spatial location point where the detection signal-to-noise ratio is lower than the sensitivity threshold, and assigning spatial exclusion weight to the detection pseudo-node based on the difference between the detection signal-to-noise ratio and the sensitivity threshold.

[0042] For each component node in the initial component diagram, configure a 0-1 type decision variable to represent the sampling selection state of the component node;

[0043] With the goal of maximizing the comprehensive risk index, the comprehensive risk index is used as a weight to perform a weighted summation on the 0-1 type decision variables to establish a global optimization objective function;

[0044] Based on the spatial repulsion weight of the probe pseudo-node and the spatial topological distance between the component node and the probe pseudo-node, a probe pseudo-node repulsion constraint is constructed to restrict sampling points from falling into the detection-restricted area.

[0045] Under the premise of satisfying the exclusion constraint of the detected pseudo-nodes and the preset statutory sampling quantity, the global optimization objective function is solved to determine the optimal inspection batch division result;

[0046] The initial component graph and its feature representation are constructed into a state space. A joint reward function is constructed by combining the comprehensive risk index and the spatial exclusion weight of the detected pseudo-nodes. A hierarchical action space is configured to select the candidate test batch community and determine the corresponding component node.

[0047] The agent is driven by a graph reinforcement learning algorithm to perform sequential decision-making in the hierarchical action space according to the state space to maximize the joint reward function, and outputs a sampling detection scheme containing an ordered component node sampling sequence.

[0048] Secondly, the present invention provides a machine learning-based concrete strength sampling and testing optimization system, including a mapping unit, which acquires basic data of the concrete component to be tested and maps it to a preset concrete physical knowledge graph.

[0049] The graphing unit determines the heterogeneous association edges and edge weights between component nodes based on the physical constraints of the concrete physical knowledge graph, and obtains the initial component graph.

[0050] The clustering unit inputs the initial component graph into the heterogeneous graph attention network, extracts node feature representations using a specific attention head that matches the node type, divides candidate test batch communities based on the feature representations, and calculates a comprehensive risk index.

[0051] The modeling unit quantifies and calculates the detection signal-to-noise ratio through a detection sensitivity model, generates detection pseudo-nodes in the candidate test batch community where the detection signal-to-noise ratio is lower than a preset sensitivity threshold, establishes an objective function aimed at maximizing the comprehensive risk index, and applies detection pseudo-node exclusion constraints.

[0052] The decision unit solves the objective function to determine the optimal batch partitioning result, and uses a graph reinforcement learning algorithm to perform dynamic path planning under the constraint of the exclusion of the pseudo-nodes of the detector, and outputs a detection scheme containing an ordered sampling sequence.

[0053] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the machine learning-based concrete strength sampling and detection optimization method described in the first aspect of the present invention.

[0054] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the machine learning-based concrete strength sampling and detection optimization method as described in the first aspect of the present invention.

[0055] The beneficial effects of this invention are as follows: by deeply integrating the underlying association constraints of the concrete physical knowledge graph with the spatial exclusion mechanism based on the detection pseudo-nodes, a global optimization game model that takes into account both the common sense of structural mechanics and the blind spots of on-site physics is constructed. Then, the graph reinforcement learning algorithm is used to perform dynamic sequential optimization in the hierarchical action space. Finally, the traditional fragmented and blind discrete static point selection mode is transformed into a dynamic and orderly detection path planning that simultaneously satisfies extreme value risk coverage, legal compliance, and absolute measurability on-site. Attached Figure Description

[0056] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 This is a flowchart of a machine learning-based method for optimizing concrete strength sampling and testing.

[0058] Figure 2 This is a diagram of a computer device used for optimizing a machine learning-based method for sampling and testing concrete strength. Detailed Implementation

[0059] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0060] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0061] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0062] Reference Figures 1-2 This is one embodiment of the present invention, which provides a machine learning-based method for optimizing concrete strength sampling and testing, comprising the following steps:

[0063] S1: Obtain the basic data of the concrete component to be tested and map it to the preset concrete physical knowledge graph.

[0064] In order to comprehensively and accurately characterize the physical state and spatiotemporal relationship of the building entity, we first obtain multi-dimensional basic data of the concrete components to be inspected by connecting with Building Information Modeling (BIM), construction logs and on-site preliminary testing equipment.

[0065] The basic data of the concrete component to be tested includes numerical test data, entity attribute data, construction process data, and spatial topology data.

[0066] Specifically, the numerical detection data is used to calculate feature similarity, including rebound value and carbonization depth, which directly reflects the current apparent physical and mechanical performance degradation state of the component; the entity attribute data is used as an entity link matching identifier, including component type (such as beam, slab, column, wall, etc.) and design strength grade (such as C30, C40, etc.), providing an absolute legal label for entity alignment and classification; the construction process data is used to trigger physical constraint verification, including pouring date and curing method, which can effectively identify whether the components belong to the "same batch" and "same conditions", thereby meeting the underlying specification requirements for batch division; the spatial topology data is used to determine the mechanical transmission path, including the connection relationship between the floor and the structure (such as slab overlapping beam, beam supporting column, etc.), ensuring that the algorithm will not violate the basic mechanical common sense of building engineering when dividing communities in the future.

[0067] After obtaining the above multidimensional basic data, the entityization and attribute mounting operations are performed to map the discrete engineering data to the preset concrete physical knowledge graph.

[0068] Based on the entity attribute data (component type and design strength level), a corresponding entity node is instantiated for each real component to be tested on site in the knowledge graph. The numerical detection data and construction process data are transformed into structured attribute key-value pairs and accurately attached to the corresponding entity nodes, enriching the feature expression of individual components. According to the spatial topology data, topological edges representing the mechanical transmission path are established between related component nodes; at the same time, by comparing the construction process data and entity attribute data of connected nodes, a preset graph rule base is triggered to perform physical constraint verification (e.g., verifying whether the design strength of connected components poured on the same day is consistent), thereby constructing a low-level structured network with engineering semantics and physical rules.

[0069] By performing physical constraint verification based on the graph rule base, the prior knowledge of the domain is transformed into data boundaries. This avoids the technical defects of pure data-driven algorithms that are prone to violating common sense in engineering from the source of data, and lays the underlying physical logic foundation for the subsequent construction of a highly compliant inspection batch community division that conforms to the topology of physical space and the laws of mechanical transmission.

[0070] S2: Based on the physical constraints of the concrete physical knowledge graph, determine the heterogeneous association edges and edge weights between component nodes to obtain the initial component graph.

[0071] Based on the basic data of the component nodes, a feature vector is constructed. A preset distance metric algorithm (such as Euclidean distance, cosine similarity, or Gaussian kernel function) is used to quantify and calculate the feature similarity between any two component nodes in the multidimensional feature space.

[0072] A preset association threshold is set to filter the feature similarity. If the feature similarity between two nodes is greater than or equal to the association threshold, an initial association edge is established between the corresponding component nodes in the graph, and the value of the feature similarity is directly assigned to the current edge as the initial edge weight, thus constructing a purely data-driven initial network topology.

[0073] In order to incorporate physical prior knowledge into the network topology of pure data, entity alignment technology is used to link and map the entity entities of the component nodes to the corresponding entities in the concrete physical knowledge graph, based on the entity attribute data of each component node as a matching identifier.

[0074] For connected node pairs with the initial associated edge, query the concrete physical knowledge graph to see if there are preset physical attribute constraints between the entities mapped by the connected node pairs (e.g., classification restrictions between load-bearing and non-load-bearing components, combination restrictions between cross-floor components without load-bearing association, etc.).

[0075] Based on the retrieved physical attribute constraints, the initial associated edges and their weights are verified. If the initial associated edges conform to the physical attribute constraints set by the graph, they are retained. If there is a conflict, the corresponding graph structure correction operation is triggered, which is divided into two mechanisms: weight reduction and edge breaking.

[0076] If the physical property constraint determines that the connected nodes have mechanical feature conflicts due to incompatible component types, that is, the graph verification finds that although the numerical detection data of two connected nodes are similar, there is incompatibility in component type or design strength level (for example, the algorithm establishes a connection between a C30 non-load-bearing plate and a C40 load-bearing wall), then it is determined to be a mechanical feature conflict, and a preset penalty coefficient (such as an attenuation factor between 0 and 1) is introduced to attenuate and reduce the initial edge weights, thereby reducing the aggregation probability in subsequent network clustering.

[0077] If the physical attribute constraint determines that the connected nodes have a mechanical characteristic conflict that breaks the mechanical transmission path, that is, when the graph verification combined with the spatial topology data finds that the two connected nodes are separated by structural joints (such as expansion joints or settlement joints) in the actual building physical space, or are on completely different mechanical transmission paths (for example, components with the same number belonging to two independent buildings), then it is determined that such an association is not legal and reasonable in the engineering inspection batch division, and the corresponding initial association edge is directly deleted.

[0078] The initial network topology after the graph structure correction operation is used as the initial component graph for subsequent heterogeneous graph attention networks. This effectively corrects or cuts off false associations that violate common sense in engineering mechanics due to accidental data similarity, ensuring the objective rationality of the component network topology at the physical and specification levels.

[0079] S3: Input the initial component graph into the heterogeneous graph attention network, extract node feature representations using a specific attention head that matches the node type, divide candidate test batch communities based on the feature representations, and calculate a comprehensive risk index.

[0080] By leveraging heterogeneous graph attention networks to fuse the multidimensional physical attributes and spatial topology of nodes, adaptive partitioning of inspection batch communities and quantitative calculation of comprehensive risk indicators can be achieved.

[0081] The basic data in the initial component graph is mapped to initial node vectors, and a specific attention head is invoked for each type of heterogeneous associated edge. Local feature vectors of each component node are extracted by calculating node-level attention weights. Specifically, a feature encoder such as a multilayer perceptron (MLP) is used to uniformly map the heterogeneous basic data attached to the component nodes to a hidden layer space of the same dimension, generating initial node vectors. Since the initial component graph contains heterogeneous associated edges with various physical semantics (such as "spatial adjacency edges" and "edges cast in the same batch"), an independent attention calculation channel (specific attention head) is configured for each edge type. Within each channel, node-level attention weights between the target node and its neighboring nodes are calculated, thereby aggregating neighbor information under specific physical relationships and generating local feature vectors of the current component node from different relationship perspectives.

[0082] After acquiring node-level local features, a semantic-level attention mechanism is introduced to evaluate the relative importance of different heterogeneous edge types for the current detection task (e.g., "cast in the same batch" typically has a higher weight for strength evaluation than simple "spatial adjacency"). Semantic-level attention weights corresponding to different heterogeneous edge types are calculated, and the local feature vectors are globally weighted and fused to output the feature representation of each component node. Based on these feature representations, within the deep feature space generated by the heterogeneous graph attention network, the feature similarity between connected node pairs with heterogeneous edges in the initial component graph is calculated using a cosine similarity equidistance metric function.

[0083] To balance the implicit association features extracted by graph network deep learning with the explicit physical rules assigned by the prior knowledge graph verification, a fusion function is introduced to weight and fuse the feature similarity with the corresponding edge weights in the initial component graph, thereby calculating the comprehensive association weight between the connected node pairs.

[0084] Furthermore, in this embodiment, linear weighted fusion can be adopted, introducing a dynamically adjustable weight balancing hyperparameter α (with a value range of 0 < α < 1) to construct a linear weighted summation function.

[0085] The specific comprehensive correlation weight function can be expressed as:

[0086]

[0087] in, Indicates the overall correlation weight; This represents the preset weight balancing hyperparameters; Indicates feature similarity; This represents the initial edge weight.

[0088] In other embodiments, nonlinear mapping fusion can also be used to concatenate the feature similarity with the corresponding initial edge weights at the channel level, input them into a multilayer perceptron (MLP) containing learnable parameters, and output them through an activation function mapping.

[0089] Through the aforementioned weighted fusion mechanism, the comprehensive correlation weight output by the algorithm not only includes the deep implicit commonalities of multidimensional data mined by the graph neural network, but also strictly anchors the strength of the hard mechanical connection given by the underlying physical knowledge graph, thereby completely avoiding the "pseudo-correlation" or "oversmoothing" problems that are easy to be generated by a single data-driven model.

[0090] The comprehensive association weights are assigned to the corresponding edges of the initial component graph. Using a community detection algorithm (such as the Louvain algorithm or the Leiden algorithm) to maximize the network topology modularity, graph clustering is performed on each component node in the initial component graph based on the comprehensive association weights.

[0091] During the clustering iteration process, the modularity increment is monitored in real time. When the modularity of the network topology reaches the preset convergence condition (i.e., the increment is less than the preset threshold or the set maximum number of iterations is reached), the clustering stops, and the set of each component node formed in the current network topology is taken as the corresponding candidate check batch community.

[0092] The candidate test batch community is a preliminary set of components with homogeneous mechanical properties that are generated by clustering in the network topology based on the feature representation and are to be merged for risk verification.

[0093] Using the generated candidate inspection batch community as the calculation unit, the feature representation of each component node in the candidate inspection batch community is extracted, and the mean of the feature distribution of the candidate inspection batch community in the feature space (characterizing the overall expected level of concrete strength of the inspection batch) and the variance of the feature distribution (characterizing the dispersion or non-uniformity of the quality within the inspection batch) are calculated.

[0094] A preset concrete strength probability density function is constructed based on the normal distribution or other applicable statistical distribution laws. The mean of the characteristic distribution is substituted into the preset concrete strength probability density function, and combined with the strength qualification threshold, the risk value of misjudgment by the producer is calculated. Risk value missed by the user .

[0095] Based on the topological position of the candidate test batch communities in the initial component graph, the corresponding structure sensitivity penalty coefficient is obtained using the network topology centrality algorithm. Specifically, the candidate test batch communities are mapped back to the initial component graph containing global building topology information. For each component node in the graph... The betweenness centrality algorithm is used to quantify its pivotal importance in the global mechanics transmission path. Node The betweenness centrality can be expressed as:

[0096]

[0097] in, This represents the set of all component nodes in the initial component diagram; This represents any two source nodes and a target node in the initial component diagram; Represents a node To the node The total number of shortest topological paths (i.e., the most direct path of mechanical transmission) between them; This indicates that the nodes that must be visited in the shortest topological path are... The number of paths.

[0098] After obtaining the centrality of all nodes globally, for the k-th candidate check batch community (denoted as...) Extract all component nodes contained within the current community and calculate the overall centrality index of the current community. To ensure the safety of risk assessment (i.e., the weakest link principle, where a community's sensitivity depends on its most critical component), a maximum aggregation (or mean aggregation) strategy is adopted, which can be expressed as:

[0099]

[0100] in, Indicates belonging to the first One candidate test batch community Any component node within the set ; Indicates the first The overall betweenness centrality index of the candidate test batch community; This represents the output of the graph clustering partition. One candidate batch of community.

[0101] To transform the aforementioned topological centrality indices into coefficients that can be directly used for multiplicative correction of subsequent comprehensive risk indices, a normalized mapping function is introduced to calculate the... The structural sensitivity penalty coefficient corresponding to each candidate test batch community , can be represented as:

[0102]

[0103] in, This represents the structural sensitivity penalty coefficient (theoretical range is...). );

[0104] This represents the minimum value of the community centrality index for all candidate test batches in the initial component diagram; This represents the maximum value of the community centrality index for all candidate test batches in the initial component diagram; This indicates the system's preset risk penalty amplification base.

[0105] By analyzing the risk value of misjudgment by the aforementioned producer Risk value of missed detection by the user Weighted fusion is performed, and the structure-sensitive penalty coefficient is utilized. Perform a multiplicative correction with the variance of the characteristic distribution to output the comprehensive risk index.

[0106] The comprehensive risk indicator can be expressed as:

[0107]

[0108] in, This represents a comprehensive risk indicator; This indicates that the manufacturer misjudged the risk value; This indicates that the user missed a risk value. This indicates the preset risk weight for the producer; This indicates the preset user risk weight; Indicates the structural sensitivity penalty coefficient; Represents the variance of the characteristic distribution; This represents the preset variance penalty adjustment factor.

[0109] By constructing an adaptive clustering and multidimensional risk quantification mechanism based on graph neural networks, the subjective limitations of traditional manual division methods are overcome, and high-risk areas are objectively located based on the laws of building mechanics, thus providing reliable data support for subsequent sampling path optimization decisions.

[0110] S4: Quantitatively calculate the detection signal-to-noise ratio using a detection sensitivity model, generate detection pseudo-nodes within the candidate test batch community where the detection signal-to-noise ratio is lower than a preset sensitivity threshold, establish an objective function aimed at maximizing the comprehensive risk index, and apply detection pseudo-node exclusion constraints.

[0111] To objectively quantify the complex physical environmental interference at the engineering site into specific detection signal-to-noise ratio values, thereby providing a reliable mathematical basis for subsequent algorithms to accurately identify and avoid actual detection blind spots, a detection sensitivity model is established.

[0112] The detection sensitivity model is a mathematical evaluation model used to quantify the signal-to-noise ratio of a specified spatial location point by calculating the attenuation law of the detection signal based on the physical attenuation characteristic data of the component. It includes a feature mapping layer, an attenuation calculation layer, and an evaluation quantization layer.

[0113] Based on the feature mapping layer, the physical attenuation feature data of each spatial location point within the candidate test batch community is obtained. The physical attenuation feature data includes the surface carbonization depth. Reinforcing bar density and additional surface layer thickness .

[0114] Based on the attenuation calculation layer, the physical attenuation characteristic data, combined with corresponding coefficients, are used as energy dissipation independent variables to measure the preset reference detection signal strength. A negative exponential decay physical simulation was performed to calculate the effective arrival signal energy after penetrating the obstruction.

[0115] Based on the aforementioned evaluation quantization layer, the effective arrival signal energy is compared with the ambient background noise intensity. Perform ratio calculations and output the detection signal-to-noise ratio for each of the aforementioned spatial locations. .

[0116] The detection signal-to-noise ratio It can be represented as:

[0117]

[0118] in, Indicates the detection signal-to-noise ratio; Indicates the strength of the reference detection signal; This represents the preset carbonization degradation coefficient; This indicates the preset shielding and blocking coefficient; This indicates the preset surface absorption coefficient; Indicates the depth of surface carbonization; Indicates the density of the steel reinforcement arrangement; Indicates the thickness of the additional surface layer; This indicates the intensity of ambient background noise.

[0119] In obtaining the detection signal-to-noise ratio of each component node or region ( After that, compared with the preset sensitivity threshold. A traversal comparison is performed. At the spatial coordinates of the point where the detection signal-to-noise ratio is lower than the sensitivity threshold, a detection pseudo-node is generated. Based on the difference between the detection signal-to-noise ratio and the sensitivity threshold, a spatial repulsion weight is assigned to the detection pseudo-node, which can be expressed as:

[0120]

[0121] in, Indicates the detection of pseudo-nodes Spatial exclusion weight; This represents the preset weight amplification ratio coefficient; This indicates the preset sensitivity threshold. Indicates the detection of pseudo-nodes The detection signal-to-noise ratio.

[0122] Entering the detection point optimization stage, for each component node in the initial component diagram, a 0-1 type decision variable representing the sampling selection state of the component node is configured. (Right now When a node is selected as an actual sampling point Not selected 0). With the goal of maximizing the comprehensive risk index, the comprehensive risk index is used as a weight to perform a weighted summation of the 0-1 type decision variables, establishing a global optimization objective function. The objective function can be expressed as:

[0123]

[0124] in, This represents the value of the objective function that is maximized. This represents the set of all component nodes in the initial component diagram; This indicates that it belongs to the above candidate set. Any specific component node in ; Represents component nodes Comprehensive risk indicators; Represented as component nodes Configured 0-1 type decision variables.

[0125] To prevent the objective function from purely pursuing high-risk nodes and selecting nodes that are practically impossible to detect, a detection pseudo-node exclusion constraint is constructed based on the spatial exclusion weight of the detection pseudo-node and the spatial topological distance between the component node and the detection pseudo-node. This constraint restricts sampling points from falling into the detection-restricted region and can be expressed as:

[0126]

[0127] in, Represented as component nodes Configured 0-1 type decision variables; Indicates candidate nodes Detecting pseudo nodes Euclidean distance in actual building physical space; Represents the set of all probe pseudo-nodes; Represents the set of detected pseudo-nodes Any pseudo-node in the; Indicates the detection of pseudo-nodes The spatial repulsion weight assigned to it; This indicates the system's preset upper limit threshold for safety repulsion force.

[0128] Combining the above objective function with exclusion constraints and quantity constraints, we form a standard 0-1 integer programming model, which can be expressed as:

[0129]

[0130]

[0131] in, Represents a node Corresponding comprehensive risk indicators; Represented as component nodes Configured 0-1 type decision variables; This indicates the maximum sampling quantity allowed by legal regulations or project budget. Indicates the detection of pseudo-nodes The spatial repulsion weight assigned to it; Indicates candidate nodes Detecting pseudo nodes Euclidean distance in actual building physical space.

[0132] By employing a graph reinforcement learning algorithm, and under the premise of satisfying the exclusion constraint of the detected pseudo-nodes and the preset statutory sampling quantity, the global optimization objective function is solved to determine the optimal batch division result. The set of nodes with a decision variable value of 1 is the optimal sampling and detection path that balances high-risk coverage, compliance with regulations, and absolute measurability on-site.

[0133] By transforming the limitations and obstacles of detection in the real physical world into mathematical penalty force fields at the algorithm level, engineering conflicts that prevent theoretically high-risk locations from being implemented on-site are effectively avoided, thus ensuring the feasibility of the sampling and testing plan.

[0134] S5: Solve the objective function to determine the optimal batch partitioning result, and use the graph reinforcement learning algorithm to perform dynamic path planning under the constraint of the exclusion of the pseudo-nodes of the probe, and output a detection scheme containing an ordered sampling sequence.

[0135] For the established 0-1 integer programming model with the exclusion constraint of the probe pseudo-node, a graph reinforcement learning algorithm is used as the underlying solver.

[0136] First, the engineering sampling problem is transformed into a Markov decision process. Specifically, the initial component graph and its feature representation are constructed as a state space. A joint reward function is constructed by combining the comprehensive risk index and the spatial exclusion weight of the detected pseudo-nodes. A hierarchical action space is configured for selecting the candidate test batch community and determining the corresponding component node.

[0137] At any decision step ,state It includes the global topology of the current initial component graph, the feature representation of each node, the distribution of pseudo-nodes detected across the entire field, and a dynamically updated mask matrix. The mask matrix records the states of selected nodes and nodes locked due to violations of exclusion constraints.

[0138] To reduce the dimensionality of action search in large-scale building graph networks, a two-layer action space was designed, comprising a high-level action space and a low-level action space. The high-level action space is responsible for selecting the next target candidate test batch community at the macro level. The lower-level action space is in the selected community. Internally, select a specific component node. As actual sampling points.

[0139] Design a single-step reward function When the agent is in a state Next action (i.e., select node) When ), its reward value is calculated, which can be expressed as:

[0140]

[0141] in, Indicates the current action Selected target candidate component nodes; Indicates the selected node Corresponding comprehensive risk indicators; This indicates that the intelligent agent is in the decision-making step status Next, execute the action. The single-step reward value obtained; This indicates that the system is in the decision-making step The state at that time; This indicates that the intelligent agent is in the decision-making step The actions performed; Represents the set of all probe pseudo-nodes; This indicates the system's preset upper limit threshold for the safety repulsion force; This represents the preset maximum penalty constant.

[0142] The agent is driven by a graph reinforcement learning algorithm to perform sequential decision-making in the hierarchical action space according to the state space to maximize the joint reward function, and outputs a sampling detection scheme containing an ordered component node sampling sequence.

[0143] During the training phase, the agent interacts with the environment through a policy gradient algorithm. The agent uses a graph neural network to extract the current state. The characteristics of the plot are used to output the probability distribution of actions and select nodes, accumulating the total plot reward. Through backpropagation, the agent eventually learns the optimal strategy that maximizes the cumulative comprehensive risk index while avoiding detection blind spots.

[0144] During the inference phase, the converged agent performs continuous operations on the actual engineering graph network. This involves forward reasoning in decision-making. Since decision-making follows a step-by-step process... This process was carried out sequentially, automatically generating a detection path sequence that met the requirements of on-site movement efficiency. This will be output as the final sampling and testing plan.

[0145] By introducing graph reinforcement learning algorithms to solve the global optimization model with spatial exclusion constraints, the static discrete sampling point selection is transformed into dynamic sequential path planning. While ensuring the avoidance of blind spots in on-site detection, a direct and feasible detection solution is provided that balances high-risk capture rate and on-site operational efficiency.

[0146] This embodiment also provides a machine learning-based concrete strength sampling and testing optimization system, including: a mapping unit, which acquires basic data of the concrete component to be tested and maps it to a preset concrete physical knowledge graph.

[0147] The graphing unit determines the heterogeneous association edges and edge weights between component nodes based on the physical constraints of the concrete physical knowledge graph, and obtains the initial component graph.

[0148] The clustering unit inputs the initial component graph into the heterogeneous graph attention network, extracts node feature representations using a specific attention head that matches the node type, divides candidate test batch communities based on the feature representations, and calculates a comprehensive risk index.

[0149] The modeling unit quantifies and calculates the detection signal-to-noise ratio using a detection sensitivity model, generates detection pseudo-nodes within the candidate test batch community where the detection signal-to-noise ratio is lower than a preset sensitivity threshold, establishes an objective function aimed at maximizing the comprehensive risk index, and applies detection pseudo-node exclusion constraints.

[0150] The decision unit solves the objective function to determine the optimal batch partitioning result, and uses a graph reinforcement learning algorithm to perform dynamic path planning under the constraint of the exclusion of the pseudo-nodes of the detector, and outputs a detection scheme containing an ordered sampling sequence.

[0151] This embodiment also provides a computer device applicable to the optimization method of concrete strength sampling and testing based on machine learning, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the optimization method of concrete strength sampling and testing based on machine learning as proposed in the above embodiment.

[0152] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0153] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the machine learning-based concrete strength sampling detection optimization method proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0154] In summary, this invention achieves the following: mapping underlying building entity data into a knowledge graph constrained by physical rules; using a heterogeneous graph attention network to deeply mine multi-dimensional graph node features to accurately segment high-risk inspection batches; introducing a detection signal-to-noise ratio model to transform on-site physical blind spots into pseudo-node repulsive force fields; and finally combining graph reinforcement learning algorithms to solve a globally optimized objective function with spatial constraints. This overcomes the dual technical barriers of strong subjectivity and limited on-site measurement conditions inherent in traditional manual sampling. While ensuring the accurate identification of high-risk structural components, it outputs a dynamic and orderly sampling and inspection scheme that balances extremely high safety inspection rates with practical on-site feasibility.

[0155] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A machine learning-based optimization method for concrete strength sampling and testing, characterized in that: This includes acquiring the basic data of the concrete component to be tested and mapping it to a preset concrete physical knowledge graph; Based on the physical constraints of the concrete physical knowledge graph, the heterogeneous association edges and edge weights between component nodes are determined to obtain the initial component graph. The initial component graph is input into a heterogeneous graph attention network, and node feature representations are extracted using a specific attention head that matches the node type. Based on the feature representations, candidate test batch communities are divided and a comprehensive risk index is calculated. The detection signal-to-noise ratio is quantitatively calculated using a detection sensitivity model, and detection pseudo-nodes are generated in the candidate test batch community where the detection signal-to-noise ratio is lower than a preset sensitivity threshold. An objective function is established with the goal of maximizing the comprehensive risk index, and detection pseudo-node exclusion constraints are applied. Solving the objective function determines the optimal batch partitioning result, and using a graph reinforcement learning algorithm to perform dynamic path planning under the constraint of the exclusion of the probe pseudo-nodes, outputting a detection scheme containing ordered sampling sequences; The candidate test batch community is a preliminary set of components with homogeneous mechanical properties that are generated by clustering in the network topology based on the feature representation and are to be merged for risk verification. The detection pseudo-nodes are mapped to the network topology of the initial component graph and are used to represent virtual physical location points within the community that meet the preset detection sensitivity conditions.

2. The machine learning-based concrete strength sampling and testing optimization method as described in claim 1, characterized in that: The basic data of the concrete component to be tested includes numerical test data, entity attribute data, construction process data, and spatial topology data; The numerical detection data is used to calculate feature similarity, including rebound value and carbonization depth; The entity attribute data is used as an identifier for entity link matching, including component type and design strength level; The construction process data is used to trigger physical constraint verification, including the pouring date and curing method; The spatial topology data is used to determine the mechanical transmission path, including the connection between the floor and the structure.

3. The machine learning-based concrete strength sampling and testing optimization method as described in claim 2, characterized in that: The process of obtaining the initial component diagram includes calculating the feature similarity between nodes based on the basic data of the component nodes; When the feature similarity meets the preset association threshold, initial association edges and initial edge weights are generated between the corresponding component nodes to construct the initial network topology; Based on the entity attribute data of each component node, the entity of the component node is linked to the corresponding entity in the concrete physics knowledge graph; For connected node pairs with the initial associated edge, query the concrete physics knowledge graph to see if there are any preset physical attribute constraints between the entities mapped by the connected node pairs. Based on the queried physical attribute constraints, the initial associated edges and their weights are verified, and the corresponding graph structure correction operation is performed. If the physical property constraint determines that the connected node pair has a mechanical feature conflict due to incompatibility of component types, then a preset penalty coefficient is introduced to attenuate and reduce the weight of the initial edge. If the physical property constraint determines that the connected node pair has a mechanical feature conflict that breaks the mechanical transmission path, then the corresponding initial associated edge is deleted. The initial network topology after the graph structure correction operation is used as the initial component graph.

4. The machine learning-based concrete strength sampling and testing optimization method as described in claim 3, characterized in that: The process of dividing the candidate test batch community includes mapping the basic data in the initial component graph to an initial node vector, and calling the corresponding specific attention head for each type of heterogeneous associated edge, and extracting the local feature vector of each component node by calculating the node-level attention weight. Calculate the semantic-level attention weights corresponding to different heterogeneous association edge types, perform global weighted fusion on the local feature vectors, and output the feature representations of each component node; Based on the feature representation, calculate the feature similarity between connected node pairs with heterogeneous edges in the initial component graph; The feature similarity is weighted and fused with the corresponding edge weight in the initial component graph to calculate the comprehensive association weight between the connected node pairs. With the goal of maximizing the modularity of the network topology, graph clustering is performed on each component node in the initial component graph based on the comprehensive association weight; When the modularity of the network topology reaches the preset convergence condition, clustering stops, and the set of each component node formed in the current network topology is taken as the corresponding candidate check batch community.

5. The machine learning-based concrete strength sampling and testing optimization method as described in claim 4, characterized in that: The calculation of the comprehensive risk index includes extracting the feature representation of each component node in the candidate inspection batch community, and calculating the mean and variance of the feature distribution of the candidate inspection batch community in the feature space. Substituting the mean of the characteristic distribution into the preset probability density function of concrete strength, and combining it with the strength qualification threshold, the risk value of misjudgment by the producer is calculated. Risk value missed by the user ; Based on the topological position of the candidate test batch communities in the initial component graph, the corresponding structure sensitivity penalty coefficient is obtained using the network topology centrality algorithm. ; By analyzing the risk value of misjudgment by the aforementioned producer Risk value of missed judgment by the user Weighted fusion is performed, and the structure-sensitive penalty coefficient is utilized. Perform a multiplicative correction with the variance of the characteristic distribution to output the comprehensive risk index.

6. The machine learning-based concrete strength sampling and testing optimization method as described in claim 5, characterized in that: The detection sensitivity model is a mathematical evaluation model used to quantify the signal-to-noise ratio of a specified spatial location point based on the attenuation law of the detection signal according to the physical attenuation characteristic data of the component. It includes a feature mapping layer, an attenuation calculation layer, and an evaluation quantization layer. Based on the feature mapping layer, the physical attenuation feature data of each spatial location point within the candidate test batch community are obtained; The physical attenuation characteristic data includes the surface carbonization depth. Reinforcing bar density and additional surface layer thickness ; Based on the attenuation calculation layer, the physical attenuation characteristic data, combined with corresponding coefficients, are used as energy dissipation independent variables to measure the preset reference detection signal strength. Perform a physical simulation of negative exponential decay to calculate the effective arrival signal energy after penetrating the barrier; Based on the aforementioned evaluation quantization layer, the effective arrival signal energy is compared with the ambient background noise intensity. Perform ratio calculations and output the detection signal-to-noise ratio for each of the aforementioned spatial locations. .

7. The machine learning-based concrete strength sampling and testing optimization method as described in claim 6, characterized in that: The application of the detection pseudo-node exclusion constraint includes generating the detection pseudo-node at the spatial location coordinates where the detection signal-to-noise ratio is lower than the sensitivity threshold, and assigning a spatial exclusion weight to the detection pseudo-node based on the difference between the detection signal-to-noise ratio and the sensitivity threshold. For each component node in the initial component diagram, configure a 0-1 type decision variable to represent the sampling selection state of the component node; With the goal of maximizing the comprehensive risk index, the comprehensive risk index is used as a weight to perform a weighted summation on the 0-1 type decision variables to establish a global optimization objective function; Based on the spatial repulsion weight of the probe pseudo-node and the spatial topological distance between the component node and the probe pseudo-node, a probe pseudo-node repulsion constraint is constructed to restrict sampling points from falling into the detection-restricted area. Under the premise of satisfying the exclusion constraint of the detected pseudo-nodes and the preset statutory sampling quantity, the global optimization objective function is solved to determine the optimal inspection batch division result; The initial component graph and its feature representation are constructed into a state space. A joint reward function is constructed by combining the comprehensive risk index and the spatial exclusion weight of the detected pseudo-nodes. A hierarchical action space is configured to select the candidate test batch community and determine the corresponding component node. The agent is driven by a graph reinforcement learning algorithm to perform sequential decision-making in the hierarchical action space according to the state space to maximize the joint reward function, and outputs a sampling detection scheme containing an ordered component node sampling sequence.

8. A machine learning-based concrete strength sampling and testing optimization system, based on the machine learning-based concrete strength sampling and testing optimization method according to any one of claims 1 to 7, characterized in that: This includes a mapping unit that acquires the basic data of the concrete component to be tested and maps it to a preset concrete physical knowledge graph. The graphing unit determines the heterogeneous association edges and edge weights between component nodes based on the physical constraints of the concrete physical knowledge graph, and obtains the initial component graph. The clustering unit inputs the initial component graph into the heterogeneous graph attention network, extracts node feature representations using a specific attention head that matches the node type, divides candidate test batch communities based on the feature representations, and calculates a comprehensive risk index. The modeling unit quantifies and calculates the detection signal-to-noise ratio through a detection sensitivity model, generates detection pseudo-nodes in the candidate test batch community where the detection signal-to-noise ratio is lower than a preset sensitivity threshold, establishes an objective function aimed at maximizing the comprehensive risk index, and applies detection pseudo-node exclusion constraints. The decision unit solves the objective function to determine the optimal batch partitioning result, and uses a graph reinforcement learning algorithm to perform dynamic path planning under the constraint of the exclusion of the pseudo-nodes of the detector, and outputs a detection scheme containing an ordered sampling sequence.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the machine learning-based concrete strength sampling and detection optimization method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the machine learning-based concrete strength sampling and detection optimization method according to any one of claims 1 to 7.

Images