A method and system for monitoring the condition of pile-slab road structures based on artificial intelligence

Abstract

This invention discloses an artificial intelligence-based method and system for monitoring the condition of pile-slab road structures. The method includes: determining the force transmission relationship between monitoring points based on finite element analysis of the structure, and constructing a directed force transmission graph; extracting response feature vectors under healthy conditions and clustering them to obtain multiple response patterns; mapping the response patterns corresponding to multiple historical responses of each monitoring point to the directed force transmission graph, and extracting the response style vector of each node in the directed force transmission graph using a graph convolutional neural network; calculating the response transmission intensity based on the response style vector and response similarity; inferring the response data under real-time load based on the response transmission intensity, and obtaining the condition monitoring results based on the inference bias. This method learns the force transmission path of the pile-slab road structure based on artificial intelligence, continuously monitors the evolution of structural mechanical properties, and achieves a leap from static threshold judgment to dynamic tracking, significantly improving the detection capability of early hidden damage.

Description

Technical Field

[0001] This invention relates to the field of road structure monitoring, specifically to an artificial intelligence-based method and system for monitoring the condition of pile-slab road structures. Background Technology

[0002] Pile-slab road structures, as a novel type of highway structure for composite foundation reinforcement, consist of precast pipe piles or cast-in-place piles at the bottom, a reinforced concrete pavement slab at the top, and the soil beneath the slab, forming a complex multi-medium coupled force system. They are widely used in engineering fields such as high-speed railway ballastless track subgrade, soft soil foundation treatment for highways, and bridge transition sections. In practical applications, pile-slab road structures are subjected to the cyclical impact of high-frequency heavy-load vehicles and alternating environmental factors over long periods, making them prone to hidden defects such as slab bottom voids, loose pile head connections, and micro-cracks in the pile body. These defects occur within the structure and are difficult to detect through traditional manual inspections. Existing monitoring methods based on single thresholds, primarily using localized point-based methods, also lack sufficient sensitivity and accuracy in identifying early-stage hidden damage. This results in damage detection often lagging behind macroscopically visible cracks or excessive deformation, increasing potential safety risks and the difficulty of subsequent maintenance. Summary of the Invention

[0003] To address the problems existing in the prior art, this invention provides an artificial intelligence-based method and system for monitoring the condition of pile-slab road structures. The aim is to combine the static force transmission path in the field of mechanics with data-driven dynamic graph modeling. First, physical priors are used to avoid false correlations caused by coupled vibrations. Then, artificial intelligence is used to analyze historical recurring patterns to learn the normal interaction benchmark of the internal structure of the pile-slab road. Finally, based on the predicted deviation between the current structural stress and the normal interaction benchmark, early hidden damage to the road structure is accurately identified.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] In a first aspect, the present invention provides a method for monitoring the condition of pile-slab road structures based on artificial intelligence, comprising:

[0006] Based on the finite element analysis of the pile-slab road structure, the force transmission relationship between all monitoring points is determined, and a directed force transmission graph of the response characteristics of the response load event is constructed.

[0007] Based on the historical response data of multiple monitoring points under the healthy state of pile-slab road structures, response feature vectors are extracted, and the response feature vectors are clustered to obtain multiple response patterns.

[0008] The response patterns corresponding to multiple historical responses at each monitoring point are mapped to the force transfer directed graph, and the response style vector of each node is extracted from the force transfer directed graph by a graph convolutional neural network.

[0009] The response propagation strength between nodes is calculated based on the response style vector and response similarity.

[0010] Based on the response transmission intensity, the response data under real-time load is inferred, and the state monitoring results of the pile-slab road structure are obtained based on the inference deviation.

[0011] Preferably, the finite element analysis based on the pile-slab road structure determines the force transmission relationship between all monitoring points and constructs a directed force transmission graph of response characteristics to load events, including:

[0012] Based on the finite element model of the pile-slab road structure, mechanical analysis is performed to determine the force transmission path inside the structure, and multiple force transmission levels of the structure are determined according to the force transmission sequence.

[0013] Obtain the peak data time of all monitoring points, and calculate the time difference between the peak data time of each monitoring point and the trigger time of the load event in turn;

[0014] The time differences are sorted according to their numerical values, and the time differences are divided into segments according to their maximum and minimum values ​​to obtain multiple consecutive non-overlapping difference segments, wherein the number of difference segments is the same as the number of force transmission levels.

[0015] Using monitoring points as nodes and force transmission direction as edge direction, a directed graph of force transmission characteristics is constructed, where:

[0016] Based on the difference range to which each monitoring point belongs, the monitoring points are mapped to the corresponding force transmission levels;

[0017] Within the same force transmission level, the upstream and downstream force transmission relationship between monitoring points is determined according to the chronological order of the peak data times of each monitoring point.

[0018] Between different force transmission levels, the connection relationship between nodes is determined according to the force transmission path.

[0019] Preferably, mapping the response patterns corresponding to multiple historical responses at each monitoring point to the force transmission directed graph includes:

[0020] Within each load event, the attention score among the response feature vectors of all monitoring points is calculated to obtain the attention score matrix;

[0021] Based on the attention score matrix, the feature contribution of each monitoring point to all other monitoring points in the force transmission directed graph is calculated to update the initial response feature vector of each monitoring point, thereby obtaining the enhanced feature vector of each monitoring point.

[0022] The similarity between the enhanced feature vector of each monitoring point and the center vector of multiple response patterns is calculated, and the multiple historical responses of the monitoring points are matched with the corresponding response patterns in turn.

[0023] The results of multiple response modes of each monitoring point in multiple load events are recorded to the corresponding nodes in the force transmission directed graph.

[0024] Preferably, the attention score matrix is ​​calculated based on the feature similarity between any two response feature vectors, and the feature similarity includes numerical feature similarity and physical correlation.

[0025] The physical correlation degree is calculated based on the node topological distance between the monitoring points that generate two response feature vectors in the directed force transmission graph.

[0026] Preferably, the step of extracting the response style vector of each node in the force transmission directed graph using a graph convolutional neural network includes:

[0027] Multiple response pattern results obtained by mapping the same node in different load events are used as derived description data and quantized into derived description vectors.

[0028] In the force transmission directed graph, the derived description vector of each node is subjected to multi-layer graph convolution operation based on a graph convolutional neural network, which aggregates the graph features of multiple neighboring nodes and outputs the response style vector of each node.

[0029] Preferably, the step of calculating the response propagation strength between nodes based on the response style vector and response similarity includes:

[0030] Based on the proportion of the connection relationship between two nodes in the directed graph of transmission and the time delay characteristic difference in historical responses that meet a preset threshold, the transmission frequency of each candidate node pair is statistically analyzed.

[0031] Calculate the similarity of the response style vectors of the candidate node pairs;

[0032] Response similarity is calculated based on the conditional cross-correlation coefficient between the response feature vectors of candidate node pairs;

[0033] The response transmission strength between nodes is obtained by weighting the response style vector based on the transmission frequency, summing it with the response similarity, and then normalizing the sum.

[0034] Preferably, the step of inferring the response data under real-time load based on the response transfer intensity and obtaining the state monitoring results of the pile-slab road structure based on the inference bias includes:

[0035] Collect real-time response data under load and identify the current force transmission starting point node based on time-series characteristics;

[0036] The inference response data of the downstream nodes is iteratively calculated from the starting node along the directed graph.

[0037] Calculate the inference deviation between the inference response data of the current monitoring node and the real-time monitoring response data of that node;

[0038] Obtain the reciprocal of the difference between the response transmission strength of the current node and its upstream neighboring node, and the response transmission strength of the current node and its downstream neighboring node. Then, use the reciprocal of the difference and the cumulative inference bias of all upstream nodes of the current node to weight and correct the inference bias of the current node.

[0039] The location of structural damage is determined sequentially based on the magnitude of the inference deviation after correction at each node.

[0040] Secondly, the present invention provides an artificial intelligence-based condition monitoring system for pile-slab road structures, comprising:

[0041] The response mapping module is used to determine the force transmission relationship between all monitoring points based on the finite element analysis of the pile-slab road structure, and to construct a directed force transmission graph of the response characteristics of the response load event.

[0042] The feature extraction module is used to extract response feature vectors based on historical response data from multiple monitoring points under the health status of pile-slab road structures, and to cluster the response feature vectors to obtain multiple response patterns.

[0043] The style learning module is used to map the response patterns corresponding to multiple historical responses of each monitoring point to the force transmission directed graph, and extract the response style vector of each node in the force transmission directed graph through a graph convolutional neural network.

[0044] The intensity determination module is used to calculate the response propagation intensity between nodes based on the response style vector and response similarity.

[0045] The condition monitoring module infers the response data under real-time load based on the response transmission intensity, and obtains the condition monitoring results of the pile-slab road structure based on the inference deviation.

[0046] Thirdly, the present invention provides an electronic device, comprising:

[0047] Memory, used to store executable instructions;

[0048] The processor, when running the executable instructions stored in the memory, implements the artificial intelligence-based piling-slab road structure condition monitoring method described above.

[0049] Fourthly, the present invention provides a computer-readable storage medium storing executable instructions, which, when executed by a processor, implement the artificial intelligence-based method for monitoring the condition of a pile-slab road structure as described above.

[0050] The present invention provides a method and system for monitoring the condition of pile-slab road structures based on artificial intelligence, which has the following beneficial effects:

[0051] 1. This invention constructs an internal force transmission directed graph that integrates the prior knowledge of physical force transmission paths and historical repetitive transmission modes, and performs real-time response prediction and deviation calculation based on the response transmission intensity. It can detect significant deviations in the early stages of damage, such as local voids under the slab or micro-cracks at the connection between the abutment and the pile, which lead to a slight reduction in force transmission intensity. This provides early warning and is significantly better than the passive mode of traditional threshold alarms or single statistical correlation methods that only trigger when there is significant settlement or cracks. This greatly reduces the risk of damage and maintenance costs of pile-slab road structures in long-term service.

[0052] 2. This invention introduces static physical dependency constraints combined with the temporal transmission habits in response data under load events, avoiding the false bidirectional strong correlation and dead loops generated by traditional statistical correlation methods in force transmission simulation under complex internal scenarios such as sheet-pile mutual push, ensuring the physical rationality of the directed force transmission graph and improving the reliability of monitoring results.

[0053] 3. This invention determines the force response transmission intensity within a pile-slab road structure by extracting response patterns through clustering and plotting response styles using graph convolution. It focuses on capturing relative time delay, attenuation patterns, and repetitive transmission habits, and utilizes attention enhancement and multi-layer convolution to deepen the response style, rather than directly using conventional statistical features. This approach is less affected by differences in loads and noise, ensuring that the normal force transmission benchmark maintains high accuracy under real-world, variable load conditions. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating an artificial intelligence-based method for monitoring the condition of pile-slab road structures according to the present invention.

[0055] Figure 2 This is a finite element model diagram of a pile-slab road structure provided in one embodiment of the present invention;

[0056] Figure 3 This is a structural block diagram of an artificial intelligence-based piling-slab road structure condition monitoring system according to the present invention. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] To facilitate understanding of this embodiment, a detailed description of an artificial intelligence-based method for monitoring the condition of pile-slab road structures disclosed in this embodiment of the invention will be provided first.

[0059] like Figure 1 As shown, the method includes:

[0060] Step S1: Based on the finite element analysis of the pile-slab road structure, determine the force transmission relationship between all monitoring points and construct a directed force transmission graph of the response characteristics of the response load event.

[0061] Based on the prior physical topology of the pile-slab structure, a constraint framework is provided for subsequent dynamic analysis.

[0062] Step S2: Extract response feature vectors from historical response data of multiple monitoring points under the healthy state of the pile-slab road structure, and cluster the response feature vectors to obtain multiple response patterns.

[0063] Step S3: Map the response patterns corresponding to multiple historical responses of each monitoring point to the force transmission directed graph, and extract the response style vector of each node in the force transmission directed graph through a graph convolutional neural network.

[0064] Step S4: Calculate the response propagation strength between nodes based on the response style vector and response similarity.

[0065] Step S5: Based on the response transmission intensity, infer the response data under real-time load, and obtain the state monitoring results of the pile-slab road structure based on the inference deviation.

[0066] The long-term service performance of pile-slab road structures depends on their concealed force transmission paths. Traditional monitoring methods rely on discrete sensor data, which makes it difficult to characterize the dynamic coupling relationships between components, leading to false alarms, ambiguous location, and the inability to warn of early, hidden damage. This embodiment integrates dynamic graph learning and graph convolutional neural networks, combining the force transmission path with real-time monitoring data to construct a dynamic graph model with historical style learning, achieving a technological leap from single-point anomaly alarms to path pattern diagnosis.

[0067] Understandably, damage to pile-slab structures often manifests first as abnormal force transmission paths, rather than a sudden and significant deviation at a single node. Conventional methods are prone to missing early, hidden damage. For example, early local voids under the slab may only cause a slight increase in delay and a minor decrease in amplitude at downstream nodes, with the data still within normal thresholds, but the path sequence / attenuation pattern is already abnormal. Furthermore, this embodiment not only identifies early damage alarms but also identifies which segment of the force transmission path has changed its transmission pattern, facilitating the location of defects in engineering projects.

[0068] The response of a pile-slab structure is not instantaneously transmitted to all nodes. The actual structure is a composite elastic body of reinforced concrete. The stress waves (longitudinal waves, transverse waves, and surface waves) generated by the load (such as a vehicle driving over the slab) propagate through the material at finite speeds (longitudinal wave speed in concrete is approximately 3500~4500 m / s, and shear wave speed is approximately 2000~2500 m / s). During propagation, reflection, refraction, attenuation, and mode transitions also occur. Starting from the point of load application, the response first reaches the nearest nodes and then gradually propagates towards more distant nodes. For example, when a vehicle directly hits the midpoint of a slab, the sensors in the slab record the peak value in milliseconds, while the connected pile caps take tens of milliseconds, and the pile tops may require 100~300 milliseconds. Under normal, healthy structures, the force transmission path is relatively fixed, and the response exhibits a predictable wavefront progression sequence.

[0069] Preferably, step S1, which involves determining the force transmission relationship between all monitoring points using finite element analysis based on the pile-slab road structure and constructing a directed force transmission graph of response characteristics to load events, includes:

[0070] S101, based on the finite element model of the pile-slab road structure, mechanical analysis is performed to determine the internal force transmission path of the structure, and multiple force transmission levels of the structure are determined according to the force transmission sequence.

[0071] like Figure 2 As shown, a spatial finite element beam-grid model of a truss system was established using the spatial finite element software Ansys. Both the superstructure and substructure of the pile-slab road were simulated using beam elements. The standard length of the pile-slab roadbed is 84 (7×12m) m, the length of the substructure pipe column is 20m, the length of the pier column is 8m, and the length of the pile column driven into the ground is 12m.

[0072] First, a refined model is used for stress analysis to reflect the structural stress characteristics. This mainly includes longitudinal shear lag effect, transverse frame effect, and pier-beam consolidation effect. By tracking the stress on the upper concrete beam-slab and the lower concrete pipe pile at each stage, the stress performance of the concrete beam-slab under all working conditions from the construction stage to bearing live loads and the service stage can be accurately obtained, providing support for the structural safety.

[0073] S102, obtain the peak data time of all monitoring points, and calculate the time difference between the peak data time of each monitoring point and the trigger time of the load event in sequence.

[0074] Finite element analysis (FEM) is based on design drawings, standard material parameters, and ideal boundary conditions, and the force transmission path it derives is a theoretical prediction. However, actual construction quality, material inhomogeneity, hidden defects (such as voids and microcracks), local variations in foundation conditions, and weakening due to long-term service can all cause deviations between the actual force transmission path of the structure and the ideal theoretical path. The measured peak time difference introduced in this method directly reflects the dynamic response of the structure under actual loads. By continuously comparing the differences between theoretical predictions and actual time series using time-series characteristics, it can dynamically reflect the current true mechanical state of the structure.

[0075] S103, sort the time differences according to their numerical values, divide the time differences into segments according to their maximum and minimum values, and obtain multiple consecutive non-overlapping difference segments, wherein the number of difference segments is the same as the number of force transmission levels.

[0076] S104, using monitoring points as nodes and force transmission direction as edge direction, constructs a directed graph of response characteristics, where:

[0077] Based on the difference range to which each monitoring point belongs, the monitoring points are mapped to the corresponding force transmission levels;

[0078] Within the same force transmission level, the upstream and downstream force transmission relationship between monitoring points is determined according to the chronological order of the peak data times of each monitoring point.

[0079] Between different force transmission levels, the connection relationship between nodes is determined according to the force transmission path.

[0080] Traditional methods for classifying monitoring points rely heavily on human experience, resulting in high subjectivity and low efficiency. This embodiment utilizes the statistical distribution characteristics of the peak time difference of each monitoring point, combined with the theoretical number of levels, to automatically segment and divide the data. This enables objective and rapid classification of monitoring points into different force transmission levels, transforming the abstract mechanical model into a data-driven directed force transmission graph. This provides an efficient and standardized method for interpreting data from large-scale monitoring networks.

[0081] Preferably, step S2, obtaining the response feature vector, includes:

[0082] In one specific implementation, the load event trigger point is first detected, such as a sudden acceleration change at a node on the plate exceeding a threshold, which serves as the starting point of the path. Within a time window of 2 to 5 seconds after the trigger, the occurrence time, amplitude, vibration frequency, vibration acceleration, strain, and other data of the peak response at all monitoring points are recorded and sorted by peak time to form a response data sequence.

[0083] As a simplified example, assume that the pile-slab structure has four key monitoring nodes: A. midpoint of the slab, B. edge of the slab / joint of the slab, C. pile cap, and D. top of the pile.

[0084] A typical load event where a vehicle directly runs over the midpoint A of the slab will generate the following data:

[0085] Node A: Peak occurs at 0ms, amplitude is 1.0, dominant frequency is 20Hz.

[0086] Node B: Peak occurs at 45ms, amplitude is 0.85, dominant frequency is 18Hz.

[0087] Node C: Peak occurs at 92ms, amplitude is 0.62, dominant frequency is 15Hz.

[0088] Node D: Peak occurs at 210ms, amplitude is 0.38, dominant frequency is 12Hz.

[0089] The response feature vector can then be represented as a quadruple: (node, relative delay, normalized amplitude, dominant frequency).

[0090] Then, the response characteristics of each monitoring point are statistically analyzed and discretized from all historical paths, and response events with the same or highly similar characteristics are clustered into a response pattern. Clustering is a technique well-known to those skilled in the art and is not limited here.

[0091] Preferably, step S3, which involves mapping the response patterns corresponding to multiple historical responses at each monitoring point to the directed force transmission graph, includes:

[0092] S301. Within each load event, calculate the attention score among the response feature vectors of all monitoring points to obtain the attention score matrix.

[0093] As a specific embodiment, the attention score matrix is ​​calculated based on the feature similarity between any two response feature vectors, wherein the feature similarity includes numerical feature similarity and physical correlation.

[0094] The physical correlation degree is calculated based on the node topological distance between the monitoring points that generate two response feature vectors in the directed force transmission graph.

[0095] S302, based on the attention score matrix, calculate the feature contribution of each monitoring point in the force transmission directed graph to all other monitoring points, update the initial response feature vector of each monitoring point, and obtain the enhanced feature vector of each monitoring point.

[0096] As a specific implementation method, the enhanced feature vector can be calculated as follows: The currently calculated monitoring point is designated as the target monitoring point. The enhanced feature vector of the target monitoring point is determined based on the sum of the initial response feature vector of the target monitoring point and the feature contributions of non-target monitoring points in the force transmission directed graph to the target monitoring point. The feature contributions of non-target monitoring points to the target monitoring point are jointly determined based on the initial response feature vector of the non-target monitoring points and the attention score matrix. For example, this can be expressed by the following formula:

[0097]

[0098] in, For nodes Enhanced feature vectors, , For each node ,node The initial response feature vector, For nodes and nodes Attention scores between and , To transmit the total number of nodes in a directed graph.

[0099] Actual road loads are random and variable, and the monitoring threshold range used by conventional methods is easily widened by extreme loads. This embodiment improves robustness to local noise by aggregating neighbor information through a self-attention mechanism, highlighting the highly consistent interactive characteristics within a single force transmission path. For example, sequentially adjacent node vectors have higher mutual attention, suppressing single-path noise and occasional fluctuations, making the derived vectors more representative of typical healthy patterns. Direct averaging would lose order information; self-attention automatically allows vectors with high correlation and closer topological distance to contribute more weight, achieving intelligent weighted aggregation.

[0100] S303, calculate the similarity between the enhanced feature vector of each monitoring point and the center vector of multiple response patterns respectively, and match the multiple historical responses of the monitoring points to the corresponding response patterns in sequence.

[0101] S304, record the results of multiple response modes of each monitoring point in multiple load events to the corresponding nodes in the force transmission directed graph.

[0102] Understandably, in a healthy path, upstream nodes bring downstream vector features closer together through decay information, while downstream nodes strengthen their predecessor associations through upstream nodes, making the overall sequence more consistent with the gradual decay chain. Even if there is some noise in a single path, it is corrected by the highly correlated healthy components after weighting, and it can also complement the information between adjacent nodes to enhance the feature information in historical responses.

[0103] Preferably, step S3, which involves extracting the response style vector of each node from the force transmission directed graph using a graph convolutional neural network, includes:

[0104] S321, take the multiple response mode results obtained by mapping the same node in different load events as the derived description data and quantize them into a derived description vector;

[0105] Using the original features directly will lose contextual association. A node should not only be considered in terms of its own features, but also in terms of the features it habitually connects to.

[0106] By combining derived data with graph convolution, multi-hop relationships can be automatically captured, generating more robust and comprehensive vector representations while highlighting highly repetitive patterns in a healthy state.

[0107] Understandably, the response pattern is equivalent to a response feature dictionary, summarizing the various response possibilities of each node under different load events. The same monitoring point will produce different stress responses under different load conditions, i.e., multiple response patterns. The different response patterns of a node are summarized into derived descriptive data. For example, if node m has a short-delay, strong attenuation response pattern in one load event and a short-delay, medium-amplitude characteristic in another load event, then the response style of node m could be a fast response with relatively rapid attenuation with force distance and a medium frequency.

[0108] S322, in the force transmission directed graph, the derived description vector of each node is subjected to multi-layer graph convolution operation based on graph convolutional neural network, the graph features of multiple neighboring nodes are aggregated, and the response style vector of each node is output.

[0109] The aforementioned steps only complete the initial mapping from physical nodes to individual response patterns, focusing only on the best-matching single feature information without aggregating multiple historical responses, resulting in relatively localized information. Graph convolution can automatically aggregate neighbor information, allowing vectors to contain both their own features and context, i.e., the style habits of nodes connected to them in the directed graph, forming a richer representation. Simultaneously, graph convolution can avoid the problem of a node being contaminated by abnormal noise under a certain load, leading to a decrease in vector consistency across all paths. Specifically, graph convolution, through multi-layer neighbor aggregation, can automatically propagate healthy repetition rules, such as long chain associations on the main force transmission path, allowing each node's representation to capture broader habitual patterns from multiple historical responses. This truly absorbs the diverse relationships of the same node in historical data and allows information from strongly repetitive patterns under healthy structures to spread rapidly, resulting in highly consistent vectors rich in propagation context, improving the accuracy of style vectors summarized across load events.

[0110] As a specific implementation method, multi-layer graph convolutional aggregation can be performed by: executing 2 to 3 convolutional layers. The first layer aggregates the neighbor features of directly related edges, and the second layer aggregates the features of two-hop neighbors to capture indirect force transmission habits. Each layer operates as follows: the new vector of each feature node = its own vector + a weighted average of the vectors of its direct neighbors, then connected via residuals and normalized to prevent the loss of deep information and ensure stable vector scale.

[0111] In this embodiment, a single-layer graph convolution update formula is provided:

[0112]

[0113] in, It is the first in a graph convolutional neural network Layer nodes The derived description vector is used to preserve the node's own information as a residual term. Represents a node In a transitive directed graph, the set of direct neighbors is a node. Represents a node In the transitive directed graph, the first... One direct neighbor. Represents normalized nodes and nodes Topological distance between them. It is the first in a graph convolutional neural network Layer nodes Derivative description vector, It is the first in a graph convolutional neural network Layer nodes Derived descriptive vectors.

[0114] The final output is the response style vector of each node, which represents the habitual role and contextual association of the node under multiple historical load events.

[0115] In this embodiment, a simplified graph convolutional neural network is employed. This network is designed as a lightweight, self-supervised system that relies on historical health data for parameter initialization and fine-tuning. Its core function is to extract a baseline of normal response patterns under healthy conditions, rather than supervising classification or regression tasks. It leverages the inherent structure of historical health data, such as the consistency of recurring patterns, to guide the aggregation process. The entire process avoids complex deep learning training loops. Through offline pre-training, historical data is processed in batches at once, resulting in a short training cycle. It can be used with either CPU or GPU, consuming minimal resources. During actual monitoring, it is directly invoked, involving only forward inference.

[0116] During offline training, the input to the graph convolutional neural network is the derived description vector of each node and the force transfer directed graph.

[0117] The derived description vector uses multiple response pattern results obtained by mapping the same node in different load events as derived description data and quantizes them into a single vector representation. This vector captures the diverse response possibilities of a node in historical load events, such as rapid response and decay characteristics, emphasizing contextual correlation rather than a single original feature.

[0118] Force-transfer directed graphs include directed edges between nodes and topological relationships, such as the set of direct neighbors for each node. They are used to guide neighbor aggregation in graph convolution operations, capturing multi-hop associations and force-transfer path patterns.

[0119] The output of a graph convolutional neural network is a response style vector for each node, which represents the node's habitual role, contextual association, and propagation pattern under multiple historical load events, including its own historical style and the aggregated information of its neighbors.

[0120] Info NCE loss is used as the loss function. Specifically, data with high repetition patterns in a healthy state are selected from the historical load event dataset as positive samples. For example, the frequency of identical patterns for nodes i and j in T events is counted. If the frequency is greater than a threshold and nodes A and B are directly connected in the force transmission directed graph, then it is identified as a positive sample. If node A and node B exhibit a "short-delay, strong-attenuation" pattern in 8 out of 10 historical load events, and they are direct neighbors, then... and It is a positive sample pair, and the graph convolutional neural network should bring them closer to learn the habitual collaborative response style of these nodes.

[0121] Negative samples should represent irrelevant or noisy node vectors, i.e., those vectors that rarely cooperate in history or are obviously anomalous. The goal is to simulate anomalous noise pollution, allowing the network to learn to resist interference from individual load events. This can be achieved by randomly sampling irrelevant nodes as negative samples, such as randomly selecting two nodes with large topological distances from the entire historical dataset, or selecting them from different load events to ensure they do not share habitual contexts, or randomly flipping some dimensions of the vectors to add artificial noise to construct negative samples.

[0122] After constructing positive and negative samples, the graph convolutional neural network is optimized using Info NCE loss as the loss function.

[0123]

[0124] in, This represents the loss function of a graph convolutional neural network. Represents a logarithmic function. Indicates the first The response style vector of each node, For the first Positive sample vectors of nodes , No. The node of the first One negative sample vector, Represents the similarity function. For temperature parameters, It is an exponential function. Info NCE loss is a technique well-known to those skilled in the art, and its principles will not be elaborated here.

[0125] Understandingly, a graph convolutional neural network (GNN) acts as a translator, converting information from a directed graph of response style descriptions aggregated at each node into a style vector. Through the GNN, information propagates across the directed graph: the original node aggregates its own historical style and the style information of its neighbors, forming a response style vector that synthesizes the characteristic role the node played in the historical load response and which preceding and following features it habitually connects to. This is equivalent to extracting information from the grammar rulebook of response features and the propagation path, generating a habitual propagation profile for each node, facilitating subsequent quantification of whether two nodes habitually collaborate.

[0126] In the monitoring scenario of pile-slab road structures, the responses between sensor nodes are understood as repetitive behavioral interactions. The node response time series is equivalent to historical behavioral big data, and the recurring response propagation patterns between nodes are equivalent to habitual intentions, i.e., stable mechanical force transmission habits under healthy structures. By introducing the concept of graph convolution processing, we can more comprehensively capture the recurring patterns and style features between nodes, enrich the expression of normal dynamic response benchmarks, and avoid implicit repetitive patterns that may be missed by single similarity calculations.

[0127] Preferably, the step S4 of calculating the response propagation strength between nodes based on the response style vector and response similarity includes:

[0128] S401, based on the proportion of the time delay characteristic difference between two nodes in the transmission directed graph and the historical response that meets a preset threshold, the transmission frequency of each candidate node pair is statistically analyzed.

[0129] S402, Calculate the similarity of the response style vectors of the candidate node pairs;

[0130] S403, calculate response similarity based on the conditional cross-correlation coefficient between the response feature vectors of candidate node pairs;

[0131] S404, the response style vector is weighted based on the transmission frequency and then normalized after being accumulated with the response similarity to obtain the response transmission strength between nodes.

[0132] As a specific implementation method, the response transfer strength can be calculated using the following formula:

[0133]

[0134] in, For nodes and nodes The intensity of response transfer between them Use the Sigmoid activation function to ensure the output is... Between 0 and 1. , Representing nodes respectively and nodes Response style vector, For nodes and nodes The cosine similarity between the response style vectors. For nodes and nodes The transmission frequency between nodes and nodes The value is 0 when nodes are not connected in the directed graph; when nodes are connected... and nodes When connecting nodes in a directed graph via any node, retrieve nodes from the historical response. and nodes The time difference between the data peak values ​​received by the sensors at the corresponding monitoring nodes is counted, and the count is incremented by one if the preset transmission threshold is met. In response to similarity, the conditional cross-correlation coefficient of the response feature vector is calculated. The conditional cross-correlation coefficient is a well-known technique in the art and will not be elaborated here. This is a balancing coefficient used to adjust the contribution of conventional conditional correlation similarity to the final intensity.

[0135] Vehicle random loads are strongly correlated with structural coupled vibrations, and synchronous vibrations often occur between the slab and pile. Traditional correlation analysis is prone to misinterpreting these as normal, making it difficult to detect early hidden damage (such as local voids under the slab or micro-cracks at the connection between the pile cap and the pile). Alarms are only triggered when significant settlement or cracks appear, resulting in high maintenance costs and impacting traffic safety. Simple conditional correlation in pile-slab road structures can create causal loops, failing to distinguish between true force transmission anomalies and normal coupling. Therefore, this embodiment aggregates the response similarity and response style vectors obtained from traditional conditional cross-correlation calculations, enabling the reflection of intensity changes in the force flow path from within the structure.

[0136] After assembly and entering service, pile-slab road structures experience tens of thousands of vehicle loads over a long period. Traditional monitoring methods treat this data as a jumbled time series, relying on simple data processing methods or neural network analysis. This embodiment, however, takes into account the memory nature of structural force flow. If the current pile-slab road structure is healthy, its force flow transmission response to loads should exhibit a high degree of consistency with historical normal conditions. This embodiment utilizes repetitive sensing technology to capture the normal transmission benchmark between repetitive behavior analysis nodes in historical monitoring. It analyzes whether the current response has undergone style migration in the frequency domain compared to the normal interaction benchmark. If a significant deviation exists, it indicates that invisible damage or displacement has occurred within the road structure.

[0137] Preferably, step S5, which involves inferring the response data under real-time load based on the response transfer intensity and obtaining the state monitoring results of the pile-slab road structure based on the inference deviation, includes:

[0138] S501 collects real-time response data under load and identifies the current force transmission starting point node based on time-series characteristics;

[0139] S502, iteratively calculate the inference response data of downstream nodes from the starting node along the directed graph;

[0140] S503, Calculate the inference deviation between the inference response data of the current monitoring node and the real-time monitoring response data of the node;

[0141] S504, obtain the response transmission strength between the current node and the upstream neighboring node, and the reciprocal of the difference between the current node and the downstream neighboring node, and correct the inference bias of the current node by weighting the reciprocal of the difference and the cumulative inference bias of all upstream nodes of the current node.

[0142] It should be understood that the reciprocal factor of the difference is used to achieve adaptive perception and enhanced diagnosis of the force transmission role of nodes. The difference in response transmission intensity between the current node and its upstream and downstream nodes directly quantifies the role and function of that node in the force transmission path. The larger the difference, the greater the likelihood that the node is a transmission hub, experiencing greater stress and more likely to suffer losses; a difference close to zero or negative indicates that it is more likely to be a damping energy dissipation point or a force transmission end. Introducing the reciprocal of this difference during correction can automatically adjust the diagnostic sensitivity for nodes with different roles. For critical transmission hubs, even minor self-damage can lead to global force transmission degradation; therefore, the algorithm uses this factor to give higher weight and attention to their deviations, improving the ability to capture early damage at critical locations.

[0143] On the other hand, by correcting based on the cumulative deviation upstream, the cascading effects caused by upstream damage propagation can be purposefully deducted from the total deviation of downstream nodes. This makes the corrected deviation more closely reflect the original anomaly of the node's own state, thus solving the problem of ambiguous localization in traditional methods where one damage leads to multiple alarms due to the damage propagation chain.

[0144] S505, determine the location of structural damage sequentially based on the magnitude of the inference deviation after correction at each node.

[0145] The correction process takes into account two fundamental physical and engineering principles: damage propagation and differences in node roles. This significantly improves the reliability of the diagnostic results and their engineering guidance value.

[0146] In this embodiment, the directed graph is used as a simulator for a runnable response propagation model, enabling the simulation of the expected performance of the entire structure under normal conditions using real data from a few source nodes. Compared to traditional methods that directly monitor whether the raw data exceeds a threshold, this embodiment monitors the deviation between the actual response and the expected response inferred from the health model. This deviation is more sensitive to damages such as local stiffness changes and connection failures, and can capture abnormal deviations in structural behavior earlier and more accurately. It also addresses the limitation of traditional monitoring methods that can only analyze damage points but not the causes of damage, achieving a leap from data alarms to mechanism diagnosis through reasoning based on physical models and deviation analysis.

[0147] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0148] Based on the same inventive concept, this application also provides a system for implementing the aforementioned artificial intelligence-based pile-slab road structure condition monitoring method. The solution provided by this system is similar to the implementation scheme described in the above method; therefore, the specific limitations in the embodiments of the artificial intelligence-based pile-slab road structure condition monitoring system provided below can be found in the limitations of the method described above, and will not be repeated here.

[0149] like Figure 3 As shown, the present invention also provides an artificial intelligence-based piling-slab road structure condition monitoring system, comprising:

[0150] The response mapping module is used to determine the force transmission relationship between all monitoring points based on the finite element analysis of the pile-slab road structure, and to construct a directed force transmission graph of the response characteristics of the response load event.

[0151] The feature extraction module is used to extract response feature vectors based on historical response data from multiple monitoring points under the health status of pile-slab road structures, and to cluster the response feature vectors to obtain multiple response patterns.

[0152] The style learning module is used to map the response patterns corresponding to multiple historical responses of each monitoring point to the force transmission directed graph, and extract the response style vector of each node in the force transmission directed graph through a graph convolutional neural network.

[0153] The intensity determination module is used to calculate the response propagation intensity between nodes based on the response style vector and response similarity.

[0154] The condition monitoring module infers the response data under real-time load based on the response transmission intensity, and obtains the condition monitoring results of the pile-slab road structure based on the inference deviation.

[0155] It should be noted that the condition monitoring system provided in this embodiment, when handling the condition monitoring of pile-slab road structures, only uses the above-mentioned division of functional modules as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the system can be divided into different functional modules. Each functional module can be composed of a single execution unit, or two or more execution units can be integrated into one functional module to realize all the functions of that functional module.

[0156] Those skilled in the art will understand that the above modules can be implemented in whole or in part through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0157] One embodiment of the present invention also provides an electronic device, comprising:

[0158] Memory, used to store executable instructions;

[0159] The processor, when executing the executable instructions stored in the memory, implements the artificial intelligence-based pile-slab road structure condition monitoring method described above. The processor can be a central processing unit, other general-purpose processors, digital signal processors, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or other chips, or combinations of the above types of chips.

[0160] The specific details of the computer equipment used for monitoring the condition of pile-slab road structures can be understood by referring to the relevant descriptions and effects of the methods mentioned above, and will not be repeated here.

[0161] Another embodiment of the present invention provides a computer-readable storage medium storing executable instructions, which, when executed by a processor, implement an artificial intelligence-based method for monitoring the condition of a pile-slab road structure as described above.

[0162] The storage medium can be any computer-readable storage medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory, magnetic disk, or optical disk.

[0163] This invention is not limited to the specific embodiments described above. Any modifications made by those skilled in the art based on the above concept without creative effort are within the scope of protection of this invention.

Claims

1. A method for monitoring the condition of pile-slab road structures based on artificial intelligence, characterized in that, Includes the following steps: Based on the finite element analysis of the pile-slab road structure, the force transmission relationship between all monitoring points is determined, and a directed force transmission graph of the response characteristics of the response load event is constructed. Based on the historical response data of multiple monitoring points under the healthy state of pile-slab road structures, response feature vectors are extracted, and the response feature vectors are clustered to obtain multiple response patterns. The response patterns corresponding to multiple historical responses at each monitoring point are mapped to the force transfer directed graph; the response style vector of each node is extracted from the force transfer directed graph using a graph convolutional neural network, specifically including: Multiple response pattern results obtained by mapping the same node in different load events are used as derived description data and quantized into derived description vectors. In the force transmission directed graph, the derived description vector of each node is subjected to multi-layer graph convolution operation based on graph convolutional neural network, the graph features of multiple neighboring nodes are aggregated, and the response style vector of each node is output. The response propagation strength between nodes is calculated based on the response style vector and response similarity, specifically including: Based on the connection relationship between two nodes in the force transmission directed graph and the proportion of time delay feature differences in historical responses that satisfy a preset threshold, the transmission frequency of each candidate node pair is statistically analyzed. Calculate the similarity of the response style vectors of the candidate node pairs; Response similarity is calculated based on the conditional cross-correlation coefficient between the response feature vectors of candidate node pairs; The response transmission strength between nodes is obtained by weighting the response style vector based on the transmission frequency, summing it with the response similarity, and then normalizing the result. Based on the response transmission intensity, the response data under real-time load is inferred, and the state monitoring results of the pile-slab road structure are obtained based on the inference deviation.

2. The method for monitoring the condition of pile-slab road structures based on artificial intelligence according to claim 1, characterized in that, The finite element analysis based on the pile-slab road structure determines the force transmission relationship between all monitoring points and constructs a directed force transmission graph of response characteristics for load events, including: Based on the finite element model of the pile-slab road structure, mechanical analysis is performed to determine the force transmission path inside the structure, and multiple force transmission levels of the structure are determined according to the force transmission sequence. Obtain the peak data time of all monitoring points, and calculate the time difference between the peak data time of each monitoring point and the trigger time of the load event in turn; The time differences are sorted according to their numerical values, and the time differences are divided into segments according to their maximum and minimum values ​​to obtain multiple consecutive non-overlapping difference segments, wherein the number of difference segments is the same as the number of force transmission levels. Using monitoring points as nodes and force transmission direction as edge direction, a directed graph of force transmission characteristics is constructed, where: Based on the difference range to which each monitoring point belongs, the monitoring points are mapped to the corresponding force transmission levels; Within the same force transmission level, the upstream and downstream force transmission relationship between monitoring points is determined according to the chronological order of the peak data times of each monitoring point. Between different force transmission levels, the connection relationship between nodes is determined according to the force transmission path.

3. The method for monitoring the condition of pile-slab road structures based on artificial intelligence according to claim 2, characterized in that, The step of mapping the response patterns corresponding to multiple historical responses at each monitoring point to the force transmission directed graph includes: Within each load event, the attention score among the response feature vectors of all monitoring points is calculated to obtain the attention score matrix; Based on the attention score matrix, the feature contribution of each monitoring point to all other monitoring points in the force transmission directed graph is calculated to update the initial response feature vector of each monitoring point, thereby obtaining the enhanced feature vector of each monitoring point. The similarity between the enhanced feature vector of each monitoring point and the center vector of multiple response patterns is calculated, and the multiple historical responses of the monitoring points are matched with the corresponding response patterns in turn. The results of multiple response modes of each monitoring point in multiple load events are recorded to the corresponding nodes in the force transmission directed graph.

4. The method for monitoring the condition of pile-slab road structures based on artificial intelligence according to claim 3, characterized in that, The attention score matrix is ​​calculated based on the feature similarity between any two response feature vectors, and the feature similarity includes numerical feature similarity and physical correlation. The physical correlation degree is calculated based on the node topological distance between the monitoring points that generate two response feature vectors in the directed force transmission graph.

5. The method for monitoring the condition of pile-slab road structures based on artificial intelligence according to claim 1, characterized in that, The step of inferring the response data under real-time load based on the response transfer intensity and obtaining the state monitoring results of the pile-slab road structure based on the inference bias includes: Collect real-time response data under load and identify the current force transmission starting point node based on time-series characteristics; The inference response data of the downstream nodes is iteratively calculated from the starting node along the directed graph. Calculate the inference deviation between the inference response data of the current monitoring node and the real-time monitoring response data of that node; Obtain the reciprocal of the difference between the response transmission strength of the current node and its upstream neighboring node, and the response transmission strength of the current node and its downstream neighboring node. Then, use the reciprocal of the difference and the cumulative inference bias of all upstream nodes of the current node to weight and correct the inference bias of the current node. The location of structural damage is determined sequentially based on the magnitude of the inference deviation after correction at each node.

6. A condition monitoring system for pile-slab road structures based on artificial intelligence, characterized in that, include: The response mapping module is used to determine the force transmission relationship between all monitoring points based on the finite element analysis of the pile-slab road structure, and to construct a directed force transmission graph of the response characteristics of the response load event. The feature extraction module is used to extract response feature vectors based on historical response data from multiple monitoring points under the health status of pile-slab road structures, and to cluster the response feature vectors to obtain multiple response patterns. The style learning module is used to map the response patterns corresponding to multiple historical responses of each monitoring point to the force transmission directed graph, and extract the response style vector of each node in the force transmission directed graph through a graph convolutional neural network. The style learning module is specifically executed through a graph convolutional neural network when extracting the response style vector of each node in the force transmission directed graph: Multiple response pattern results obtained by mapping the same node in different load events are used as derived description data and quantized into derived description vectors. In the force transmission directed graph, the derived description vector of each node is subjected to multi-layer graph convolution operation based on graph convolutional neural network, the graph features of multiple neighboring nodes are aggregated, and the response style vector of each node is output. The intensity determination module is used to calculate the response propagation intensity between nodes based on the response style vector and response similarity, specifically for executing: Based on the connection relationship between two nodes in the force transmission directed graph and the proportion of time delay feature differences in historical responses that satisfy a preset threshold, the transmission frequency of each candidate node pair is statistically analyzed. Calculate the similarity of the response style vectors of the candidate node pairs; Response similarity is calculated based on the conditional cross-correlation coefficient between the response feature vectors of candidate node pairs; The response transmission strength between nodes is obtained by weighting the response style vector based on the transmission frequency, summing it with the response similarity, and then normalizing the result. The condition monitoring module infers the response data under real-time load based on the response transmission intensity, and obtains the condition monitoring results of the pile-slab road structure based on the inference deviation.

7. An electronic device, characterized in that, The electronic device includes: Memory, used to store executable instructions; The processor, when running the executable instructions stored in the memory, implements the artificial intelligence-based piling-slab road structure condition monitoring method according to any one of claims 1 to 5.

8. A computer-readable storage medium storing executable instructions, characterized in that, When the executable instructions are executed by the processor, they implement the artificial intelligence-based piling-slab road structure condition monitoring method according to any one of claims 1 to 5.

Images