A cloud processing method for bridge pile foundation health data

Abstract

This invention relates to the field of data processing technology, and discloses a cloud-based method for processing bridge pile foundation health data. The method constructs a spatial topology map based on the physical spatial distribution structure of the bridge pile foundations. Multidimensional time-series health data is divided into multiple distributed processing subtasks according to the connected components of the spatial topology map and assigned to corresponding cloud computing nodes. Each cloud computing node extracts local temporal features of the multidimensional time-series health data within the distributed processing subtasks, and performs graph convolution propagation along the edges of the spatial topology map to generate joint node features containing the spatial coupling relationships of the pile foundations. This invention maps the physical and mechanical transmission laws between pile foundations to an information aggregation process in the spatial topology map, restoring the mechanical transmission fluctuations of the pile foundation group, eliminating synchronous fluctuation interference caused by normal stress transmission, and avoiding the situation where the cloud processing system misjudges mechanical transmission fluctuations as simultaneous structural damage to multiple pile foundations.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, focusing on a method for distributed temporal analysis and spatial topology graph convolution calculation of bridge pile foundation health data in cloud computing nodes. This invention discloses a cloud processing method for bridge pile foundation health data. Background Technology

[0002] Existing cloud-based solutions for bridge pile foundation health data typically process the data for each pile foundation as an independent time series after receiving multidimensional data collected by front-end sensors. Specifically, the cloud system assigns the time series data (acceleration, strain, displacement, etc.) of a single pile foundation to independent cloud computing nodes based on the data arrival timestamp. Each computing node runs a sequence analysis model independently, extracting the data change trends of a single pile foundation and determining whether there are any abnormal fluctuations exceeding static thresholds. After processing, each node reports the health status label of a single pile foundation to the cloud aggregation layer, forming a discrete status list of the entire bridge pile foundation. During this process, there is no data interaction between different computing nodes, and the data processing logic for each pile foundation is physically isolated.

[0003] Bridge pile foundations constitute an interconnected mechanically coupled system under actual load. When a bridge is subjected to eccentric loading or impact, stress is spatially transmitted within the pile group through the abutment and soil. The existing technology, which treats each pile foundation as an isolated node for independent time-series analysis, severs the physical spatial connection between pile foundations. The cloud-based processing cannot perceive the stress transmission path and coupling state between different pile foundations. When stress transmission along spatial paths causes synchronous fluctuations in sensor data from multiple pile foundations, existing cloud-based processing systems, lacking spatial correlation analysis capabilities, may misinterpret this normal, clustered mechanical transmission fluctuation as simultaneous structural damage to multiple pile foundations, resulting in a misjudgment of clustered damage. Summary of the Invention

[0004] The purpose of this invention is to provide a cloud-based method for processing bridge pile foundation health data, which can solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A cloud-based method for processing bridge pile foundation health data includes: receiving multidimensional time-series health data uploaded by sensors of each bridge pile foundation in the cloud, and constructing a spatial topology map based on the physical spatial distribution structure of each bridge pile foundation; The cloud platform divides the multidimensional time-series health data into multiple distributed processing subtasks according to the connected components of the spatial topology graph, and assigns the multiple distributed processing subtasks to the corresponding cloud computing nodes. Each cloud computing node extracts local temporal features of the multidimensional time-series health data within the corresponding distributed processing subtask, and performs graph convolution propagation on the edges of the spatial topology graph to generate joint node features containing the spatial coupling relationship of the pile foundation; The cloud master node aggregates the joint features of the nodes output by each cloud computing node, allocates feature weights for different spatial topological regions through an attention mechanism, and outputs a vector of the overall health status of the bridge pile foundation group that integrates spatiotemporal features. The graph convolutional propagation maps the physical and mechanical transmission laws between pile foundations to the information aggregation process in the spatial topological graph.

[0006] Preferably, the step of constructing a spatial topology map based on the physical spatial distribution structure of each pile foundation of the bridge includes: acquiring the design drawing data of the bridge from the cloud and extracting the geographical coordinates of each pile foundation and the connection relationship of the pile caps; The cloud platform uses each pile foundation as the vertex of the spatial topology graph and establishes a first type of edge for the pile foundation vertices that are connected under the same pile cap. The cloud-based calculation determines the geographical distance between any two pile foundation vertices that do not have a pile cap connection relationship. When the geographical distance is less than the preset soil stress diffusion radius, a second type of edge is established between the two pile foundation vertices. The cloud platform configures a first weight coefficient for the first type of edge and a second weight coefficient based on the geographical distance decay for the second type of edge, and combines them to generate a weighted spatial topology graph adjacency matrix.

[0007] Preferably, the step of dividing the multidimensional time-series health data into multiple distributed processing subtasks according to the connected components of the spatial topology graph includes: the cloud obtaining the number of currently available cloud computing nodes and the real-time load status of each cloud computing node; The cloud uses a graph partitioning algorithm to divide the spatial topology graph. Under the conditions that the number of vertices in each subgraph matches the processing capacity of each cloud computing node and the sum of the weights of the cut edges is minimized, the spatial topology graph is divided into multiple subgraphs equal to the number of cloud computing nodes. The cloud platform packages the multi-dimensional time-series health data associated with the vertices of each subgraph into a distributed processing subtask, and schedules the distributed processing subtask to cloud computing nodes with idle resources that meet the number of vertices required by the subgraph based on the real-time load status of each cloud computing node.

[0008] Preferably, the step of extracting local temporal features of the multidimensional time-series health data within the corresponding distributed processing subtask includes: each cloud computing node dividing the multidimensional time-series health data within the distributed processing subtask according to a time window to obtain multi-scale time window data; Each of the cloud computing nodes uses a temporal feature extraction network containing multiple parallel one-dimensional convolutional kernels to process the multi-scale time window data, wherein one-dimensional convolutional kernels of different sizes are respectively used to extract temporal features at different time scales. Each cloud computing node concatenates the time-series features of different time scales along the channel dimension and reduces the dimensionality through pooling operations to generate the local time-series features of fixed dimensions. The local time-series features represent the local time-series evolution pattern of the pile foundation sensing data within the corresponding time window.

[0009] Preferably, the step of performing graph convolution propagation on the local temporal features along the edges of the spatial topology graph to generate node joint features containing the spatial coupling relationship of the pile foundation includes: each cloud computing node collecting local temporal features of adjacent vertices directly connected to each vertex within the distributed processing subtask; Each cloud computing node multiplies the local temporal features of the adjacent vertices with the weight coefficients of the corresponding edges and then sums them to obtain the neighborhood aggregation features; Each cloud computing node nonlinearly fuses the local temporal features of the vertex itself with the neighborhood aggregation features, and iteratively updates them through a multi-layer graph convolutional network, so that the feature information of a single vertex is transmitted to vertices outside the multi-layer along the path of the spatial topology graph, generating node joint features containing multi-order spatial physical and mechanical coupling relationships.

[0010] Preferably, the allocation of feature weights for different spatial topological regions through the attention mechanism includes: the cloud master node inputting the aggregated node joint features into the spatial attention network, wherein the node joint features include location encoding information; The spatial attention network generates a query vector by performing a linear transformation on the joint features of the nodes using a query matrix, and generates a key vector by performing a linear transformation on the joint features of the nodes using a key matrix. The cloud master node calculates the dot product of the query vector and all key vectors, and processes it through the softmax normalization function to generate the attention coefficients corresponding to the different spatial topological regions. The cloud master node multiplies the attention coefficient with the corresponding node joint features to obtain a weighted spatial feature map.

[0011] Preferably, configuring a first weighting coefficient for the first type of edge includes: obtaining real-time strain difference data of pile foundation sensors under the same pile cap connection relationship from the cloud. The cloud calculates the covariance matrix of the real-time strain difference data and extracts the principal eigenvalues ​​of the covariance matrix; The cloud platform inputs the main feature value into a pre-constructed pier stiffness degradation evaluation function and outputs the pier connection stiffness index at the current moment. The cloud platform dynamically corrects the initial first weight coefficient based on the pier connection stiffness index. When the pier connection stiffness index is lower than a set threshold, the value of the first weight coefficient is reduced. By adjusting the edge weights of the spatial topology graph, the degradation of mechanical transmission performance of the pier structure connection state over time is reflected.

[0012] Preferably, the step of segmenting the spatial topology graph using a graph partitioning algorithm further includes: the cloud acquiring the real-time network communication bandwidth between each cloud computing node during each iteration of the graph partitioning algorithm. When calculating the total weight of the severed edges, the cloud uses the reciprocal of the real-time network communication bandwidth between the cloud computing nodes where the two endpoints of the severed edges are located as a penalty factor, which is added to the weight value of the severed edges to generate a communication-aware segmentation cost. The cloud performs vertex swapping operations with the objective function of minimizing the communication-aware segmentation cost.

[0013] Preferably, the step of segmenting the multi-dimensional time-series health data within the distributed processing subtask according to a time window to obtain multi-scale time window data includes: each cloud computing node acquiring the sampling frequency of different pile foundation sensors within the distributed processing subtask; When there are pile foundation sensors with inconsistent sampling frequencies, each cloud computing node uses the target sampling frequency corresponding to the maximum window length in the multi-scale time window as a benchmark to perform upsampling interpolation processing on the time series health data of the pile foundation sensor with low sampling frequency, and downsampling filtering processing on the time series health data of the pile foundation sensor with high sampling frequency. Each cloud computing node aligns and segments the time-series health data of the pile foundation sensor, which has been processed with a unified sampling frequency, according to the multi-scale time window.

[0014] Preferably, the step of multiplying the local temporal features of the adjacent vertices with the weight coefficients of the corresponding edges and then summing them to obtain neighborhood aggregation features includes: each cloud computing node obtaining the spatial relative position coordinates of the adjacent vertices relative to the vertex and the longitudinal slope inclination angle of the bridge. Each cloud computing node divides the adjacent vertices into upstream force transmission vertices and downstream force transmission vertices based on the spatial relative position coordinates and the longitudinal slope inclination angle of the bridge; Each cloud computing node assigns a positive transmission aggregation coefficient to the upstream force transmission vertex and a negative damping aggregation coefficient to the downstream force transmission vertex. Each cloud computing node multiplies the local temporal features of the adjacent vertices by the corresponding positive propagation aggregation coefficient or negative damping aggregation coefficient and then sums them to generate neighborhood aggregation features with directional awareness capabilities.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs a spatial topology map based on the physical spatial distribution structure of bridge pile foundations. Multidimensional time-series health data is divided into distributed processing subtasks according to the connected components of the spatial topology map. After local time-series feature extraction is completed at each cloud computing node, these features are propagated along the edges of the spatial topology map using graph convolution. This processing mechanism transforms the physical and mechanical transmission laws between pile foundations into an information aggregation process within the spatial topology map. This allows individual computing nodes to acquire the stress state of adjacent pile foundations during processing, generating joint node features that include the spatial coupling relationships between pile foundations. The cloud master node allocates feature weights to different spatial topology regions through an attention mechanism, restoring single-point time-series data to reflect the spatiotemporal characteristics of the overall stress on the pile foundation group. This eliminates synchronous fluctuation interference caused by normal stress spatial transmission and avoids misjudging mechanical transmission fluctuations as simultaneous damage to multiple pile foundations.

[0016] 2. In constructing the spatial topology map, this invention establishes a first type of edge based on the connection relationship of the pier caps and a second type of edge based on the stress diffusion radius of the soil. The weight coefficients of the edges are dynamically adjusted in conjunction with the pier cap connection stiffness index, enabling the topology of the cloud-based spatial topology map to adaptively adjust as the mechanical properties of the pier caps degrade. In the distributed task partitioning stage, a graph partitioning algorithm is introduced, and a communication-aware segmentation cost is superimposed to allocate sub-tasks with the goal of minimizing cross-node communication latency, reducing cross-node communication overhead during graph convolution propagation. Before extracting temporal features, sensor data at different sampling frequencies undergo upsampling interpolation and downsampling filtering to ensure strict alignment of multi-channel data in the time dimension. In the graph convolution propagation stage, upstream and downstream vertices are distinguished based on their spatial relative position coordinates and the bridge's longitudinal slope inclination angle, and positive transmission and negative damping aggregation coefficients are assigned. This gives the neighborhood aggregation features directional sensing capabilities, further conforming to the physical laws of actual force transmission in bridges. Attached Figure Description

[0017] Figure 1 This is a flowchart of the cloud-based multidimensional time-series health data reception and preprocessing process of the present invention; Figure 2 This is a schematic diagram illustrating the construction principle of the spatial topology diagram of bridge pile foundations according to the present invention. Figure 3 This is a flowchart of the distributed subtask partitioning and cloud node scheduling process of the present invention; Figure 4 This is a flowchart of the local temporal feature extraction and graph convolution propagation feature generation of the present invention; Figure 5 This is the spatial topology graph mechanical transmission path construction and weight dynamic correction graph of the present invention; Figure 6 This is the global feature attention fusion and overall health state vector generation diagram of the present invention. Detailed Implementation

[0018] Please refer to Figure 1 This embodiment provides a cloud-based method for processing bridge pile foundation health data. The cloud receives multi-dimensional time-series health data uploaded by bridge pile foundation sensor clusters deployed at various pile foundation locations through an edge gateway communication module deployed on the edge gateway. The multi-dimensional time-series health data includes, but is not limited to, the time-series data of strain of the pile foundation structure, the time-series data of lateral and longitudinal acceleration, the time-series data of vertical and horizontal displacement, the time-series data of main reinforcement stress, and the time-series data of pore water pressure of the soil around the pile. Each set of time-series data carries a unique identifier code for the corresponding pile foundation, a data sampling timestamp, a sensor type code, and a sampling accuracy identifier. The cloud platform preprocesses the received multidimensional time-series health data to obtain standard time-series data. The preprocessing includes outlier removal, missing value completion, and data normalization. Outlier removal uses the 3σ criterion to remove sampled data that exceeds three times the standard deviation. Missing value completion uses linear interpolation to complete sampled points with consecutive missing lengths not exceeding a preset threshold. Data normalization uses the min-max normalization method to map all dimensions of sampled data to the [0,1] interval, eliminating the impact of differences in the dimensions of different sensors on the subsequent feature extraction process.

[0019] The system acquires physical spatial distribution data of bridge pile foundations in the cloud. This data includes the 3D geographic coordinates of each pile foundation in the bridge engineering coordinate system, the connection relationships between the pile foundation and the abutment, and the physical and mechanical parameters of the surrounding soil. A spatial topology map is constructed based on this data. Each pile foundation is used as a vertex in the topology map, with each vertex representing a unique entity. Vertex attributes include the 3D geographic coordinates, design parameters, and associated sensor data. Based on the physical and mechanical transmission paths between pile foundations, edges are established between vertices of pile foundations with mechanical coupling relationships. The weight coefficient of each edge represents the coupling strength between the two pile foundations. Finally, based on the vertices, edges, and corresponding weight coefficients, the system generates the adjacency matrix and degree matrix of the spatial topology map, completing its construction. (Adjacency matrix...) The definition satisfies the following formula:

[0020] Where N is the total number of bridge pile foundations, i.e., the total number of vertices in the spatial topology graph. Represents the set of real numbers, matrix elements This represents the edge weight coefficient between vertex i and vertex j, when there is no mechanical coupling between vertex i and vertex j. When there is a mechanical coupling relationship between vertex i and vertex j, For real numbers greater than 0, and the adjacency matrix... It is a symmetric matrix that satisfies This corresponds to the mutual transmission of mechanical forces between pile foundations, refer to... Figure 2 .

[0021] The mapping relationship between vertices and pile foundation physical properties used to construct the spatial topology map in this embodiment is shown in the following table: Table 1. Mapping of Spatial Topology Diagram Vertices and Pile Foundation Physical Properties

[0022] This table is used to establish the mapping relationship between the vertices of the spatial topology graph and the entities of the bridge pile foundation, clarify the physical entity attributes and associated data sources corresponding to each vertex, and ensure that the topological structure of the spatial topology graph is completely matched with the actual physical spatial distribution of the bridge pile foundation. This provides a basis for subsequent distributed task partitioning and graph convolution propagation based on the topology structure.

[0023] The cloud acquires node resource information from the cloud computing cluster, including the number of currently available cloud computing nodes, the computing resource quota, memory resource quota, and network communication resource quota for each available cloud computing node. Based on the connected components of the spatial topology graph, the multidimensional time-series health data is divided into multiple distributed processing subtasks, which are then assigned to corresponding cloud computing nodes. Using the connected components of the spatial topology graph as the basic unit, the cloud employs a graph partitioning algorithm to segment the spatial topology graph, generating multiple non-overlapping subgraphs. Each subgraph corresponds to a set of connected components; vertices within a subgraph are connected by edges, while edges between subgraphs are cut off. The cloud packages the preprocessed multidimensional time-series health data associated with all vertices of each subgraph to generate a distributed processing subtask. This subtask contains the subgraph's topological structure data, the pile-based multidimensional time-series health data corresponding to all vertices within the subgraph, and the association mapping data between the subgraph and the global spatial topology graph. Based on the real-time resource status of each available cloud computing node, the cloud schedules and allocates the generated distributed processing subtasks to the corresponding cloud computing nodes. Each cloud computing node is allocated at least one distributed processing subtask, and the number of vertices in the subgraph corresponding to the allocated subtask matches the available computing resource quota of that cloud computing node. Figure 3 .

[0024] Each cloud computing node extracts local temporal features of the multidimensional time-series health data within its corresponding distributed processing subtask. Upon receiving the corresponding subtask data packet, the cloud computing node assigned a distributed processing subtask parses the multidimensional time-series health data within the data packet, obtaining the pile foundation multidimensional time-series data corresponding to each vertex in the subgraph. The time-series data is arranged in units of time steps, forming a time-series data matrix of dimension T×D. The total length of the time steps for the time series data. The number of dimensions of the sensor data corresponding to each time step. A time-series feature extraction network is constructed on cloud computing nodes. This network includes one-dimensional convolutional layers, activation function layers, and pooling layers. The one-dimensional convolutional layers extract local evolution features from the time-series data. The activation function layers use the ReLU activation function to perform a non-linear transformation on the features output by the convolution. The pooling layers reduce the dimensionality of the features, generating a fixed-dimensional feature vector. The cloud computing nodes input the time-series data matrix corresponding to each vertex into the time-series feature extraction network. After convolution, non-linear activation, and pooling processing, the network outputs the local time-series features of the corresponding vertex. These local time-series features are fixed-dimensional vectors that represent the time-series change patterns and fluctuation characteristics of the sensor data of the corresponding pile foundation within the corresponding time window. The convolution operation for time-series feature extraction satisfies the following formula:

[0025] in, Let i be the initial temporal feature corresponding to vertex i. Let i be the multidimensional time-series health data matrix corresponding to vertex i. The weight matrix of the one-dimensional convolution kernel. is the bias vector for the convolution operation, and ReLU is the linear rectified activation function used to perform a nonlinear transformation on the result of the convolution operation.

[0026] Each cloud computing node performs graph convolution propagation on the edges of the spatial topology graph to generate joint node features containing the spatial coupling relationships between pile foundations. Graph convolution propagation maps the physical and mechanical transmission laws between pile foundations to an information aggregation process in the spatial topology graph. Based on the topology of the assigned subgraph, the cloud computing node obtains the set of neighboring vertices for each vertex within the subgraph. The set of neighboring vertices is the set of all vertices directly connected to the current vertex through edges. The cloud computing node constructs a graph convolutional network containing multiple graph convolutional layers. Each graph convolutional layer performs neighborhood feature aggregation and nonlinear transformation operations. During the processing of each graph convolutional layer, for each vertex within the subgraph, the cloud computing node collects the local temporal features corresponding to the neighboring vertices of that vertex. The local temporal features of the neighboring vertices are multiplied by the weight coefficients of the corresponding edges and then summed to generate the neighborhood aggregated features of that vertex. The cloud computing node concatenates the vertex's own local temporal features with the neighborhood aggregated features, inputs the result into a nonlinear activation function for transformation, and generates the updated features of that vertex in the current graph convolutional layer. After iterative updates through multiple graph convolutional layers, the feature information of a single vertex is transmitted along the edges of the spatial topology graph to vertices outside the layers. This results in each vertex's features containing the temporal feature information of all piles in its multi-order neighborhood, which in turn includes the spatial mechanical coupling relationship between piles. Ultimately, this generates a joint feature reference for each vertex. Figure 4 The neighborhood aggregation and feature update process of graph convolution satisfies the following formula:

[0027] in, Let i be the updated feature output of the (l+1)th graph convolutional layer. The initial input features are the input features of vertex i in the l-th graph convolutional layer. Local temporal features corresponding to vertex i ; To add a self-loop adjacency matrix, It is an identity matrix used to preserve the feature information of the vertices themselves; for The corresponding degree matrix satisfies ; Let L be the trainable weight matrix of the l-th graph convolutional layer. Let L be the bias vector of the l-th graph convolutional layer; For non-linear activation functions, either the Sigmoid function or the ReLU function can be used.

[0028] A master node is set up in the cloud. The master node establishes communication connections with all cloud computing nodes executing distributed processing subtasks. After generating joint node features, each cloud computing node uploads the generated joint node features along with the unique identifier and position code of the corresponding vertex to the master node. The master node aggregates the joint node features output by all cloud computing nodes, allocates feature weights for different spatial topological regions through an attention mechanism, and outputs an overall health state vector of the bridge pile foundation group that integrates spatiotemporal features. The master node receives the joint node features uploaded by all cloud computing nodes, sorts and concatenates them according to the unique identifier of the vertex, and generates a global node feature matrix. The global node feature matrix has a dimension of N×F, where N is the total number of vertices in the spatial topology graph, and F is the dimension of each joint node feature. The master node constructs a spatial attention network, which includes a linear transformation layer, an attention coefficient calculation layer, and a feature weighting layer. The spatial attention network performs a linear transformation on the global node feature matrix to generate query vector matrices, key vector matrices, and value vector matrices. Attention coefficients are calculated for different vertices through the dot product of the query and key vectors. These attention coefficients characterize the importance of the spatial topology region containing the corresponding vertex in the overall health assessment of the pile foundation group. The cloud master node multiplies the calculated attention coefficients by the corresponding node joint features, weighting the features of different spatial topology regions to strengthen the feature representation of key stress areas and weaken the interference features of non-critical regions. The cloud master node then performs global average pooling on the weighted global node feature matrix to generate a fixed-dimensional overall health state vector for the bridge pile foundation group. This overall health state vector integrates the temporal and spatial coupling features of all pile foundations, comprehensively reflecting the overall stress state and health level of the pile foundation group. The calculation process of the attention coefficients follows this formula:

[0029] in, Let be the attention coefficient of vertex j relative to vertex i, representing the importance of the features of vertex j to vertex i; Let i be the joint feature vector of the nodes corresponding to vertex i. Let j be the joint feature vector of the nodes corresponding to vertex j; The weight parameters of the query matrix are used to linearly transform the joint features of the nodes into a query vector; The weight parameters of the key matrix are used to linearly transform the joint features of the nodes into a key vector; is the dimension of the key vector, used to scale the dot product result to avoid the dot product value being too large, causing the gradient of the softmax function to vanish; softmax is the normalization exponential function, used to normalize all attention coefficients to the interval [0,1], and the sum of all attention coefficients is 1. This indicates transpose.

[0030] The generation process of the overall health state vector of the bridge pile foundation group satisfies the following formula:

[0031] in, The vector representing the overall health status of the bridge pile foundation group; The weight parameters of the value matrix are used to linearly transform the joint features of nodes into a value vector; GAP is the global average pooling operation, used to compress the weighted global feature matrix into a vector of fixed dimensions.

[0032] In this embodiment, by constructing a spatial topology graph that matches the physical spatial distribution of bridge pile foundations, the mechanical coupling relationship of the pile foundation group is transformed into the edges and weights of the topology graph. Based on the connected domains of the topology graph, distributed processing subtasks are divided. After each computing node completes the extraction of local temporal features, the aggregation of neighborhood features is achieved through graph convolution propagation. The physical and mechanical transmission law between pile foundations is mapped to the information aggregation process in the topology graph. Finally, the global features are weighted and fused through an attention mechanism to generate an overall health state vector that integrates spatiotemporal features. This enables the cloud processing process to perceive the spatial and mechanical coupling relationship of the pile foundation group and avoids the problem of misjudgment of cluster damage caused by isolated processing of single pile temporal data.

[0033] In a preferred embodiment, the cloud acquires BIM design drawings and as-built survey data of the bridge. From the design drawings, it extracts information such as the bridge's abutment layout, the number and number of piles corresponding to each abutment, the three-dimensional geographic coordinates of each pile in the bridge engineering coordinate system, and the design pile length, pile diameter, and concrete strength grade. From the as-built survey data, it extracts the actual construction coordinates of the piles, the actual construction dimensions of the abutments, and the connection status information, ensuring that the extracted spatial distribution data of the piles is consistent with the actual engineering status. The cloud uses each pile as a vertex in the spatial topology diagram, establishing first-type edges for pile vertices connected to the same abutment. The cloud establishes first-type edges between every pair of vertices corresponding to all piles belonging to the same abutment. These first-type edges correspond to the rigid mechanical transmission path between piles through the abutment. Since the piles within the same abutment redistribute and transfer loads through the abutment, the mechanical coupling strength corresponding to the first-type edges is higher than the mechanical coupling strength transmitted through the soil.

[0034] The cloud-based system calculates the geographical distance between any two pile foundation vertices that are not connected by a pile cap. When the geographical distance is less than a preset soil stress diffusion radius, a second type of edge is established between the two pile foundation vertices. For any two pile foundation vertices not belonging to the same pile cap, the cloud-based system calculates the Euclidean distance between the three-dimensional geographical coordinates of the corresponding pile foundations, which is taken as the geographical distance between the two pile foundations. Based on the internal friction angle, cohesion, and compression modulus of the soil surrounding the pile, the cloud-based system calculates the preset soil stress diffusion radius. The calculation of the soil stress diffusion radius is based on Terzaghi's foundation stress diffusion theory; the stress diffusion radius increases with the increase of the soil compression modulus and decreases with the increase of the soil void ratio. When the geographical distance between two pile foundations is less than or equal to the calculated soil stress diffusion radius, the cloud-based system establishes a second type of edge between the corresponding vertices of the two pile foundations. This second type of edge corresponds to the stress transmission path between the pile foundations through the surrounding soil, reflecting the influence of the stress diffusion of the surrounding soil on the stress state of adjacent pile foundations.

[0035] The cloud assigns a first weight coefficient to the first type of edges and a second weight coefficient based on geographical distance decay to the second type of edges, combining them to generate a weighted spatial topology graph adjacency matrix. The calculation process of the second weight coefficient for the second type of edges satisfies the following formula:

[0036] in, The second weight coefficient of the second type of edge between vertex i and vertex j; These are the initial baseline weight coefficients for the second type of edge; The geographical distance between the pile foundations corresponding to vertex i and vertex j; The preset soil stress diffusion radius is used; the exponential function is used to characterize the characteristic that the second weighting coefficient exhibits Gaussian decay as the geographical distance increases, which conforms to the physical law that soil stress decreases with increasing distance.

[0037] The correspondence between the pier connection stiffness index and the dynamic correction of the first weighting coefficient used in this embodiment is shown in the table below: Table 2 Comparison of Pier Connection Stiffness Index and First Weighting Coefficient with Dynamic Correction

[0038] This table establishes a mapping relationship between the pier cap connection stiffness index and the corrected value of the first weight coefficient. The cloud can directly match the corresponding corrected first weight coefficient based on the real-time calculated pier cap connection stiffness index, enabling dynamic adjustment of the first type of edge weights in the spatial topology graph. This allows the edge weights of the topology graph to reflect the degradation state of the pier cap structure connection stiffness in real time, ensuring that the information aggregation logic during graph convolution propagation remains consistent with the actual mechanical transmission laws. (Reference) Figure 5 .

[0039] The cloud-based system acquires real-time strain difference data uploaded by sensors from all pile foundations connected to the same pier cap. This data consists of a sequence of strain values ​​collected by strain sensors from each pile foundation within the same pier cap at the same timestamp. The cloud-based system standardizes this data to eliminate the influence of sensor dimension differences and calculates the covariance matrix of the standardized strain difference data. The covariance matrix has dimensions M×M, where M represents the number of pile foundations within the same pier cap. The cloud-based system performs eigenvalue decomposition on the covariance matrix, extracting the principal eigenvalues. These principal eigenvalues ​​are the largest eigenvalues ​​of the covariance matrix. The magnitude of the principal eigenvalue indicates the synchronicity of strain data among the pile foundations within the same pier cap. A larger principal eigenvalue indicates higher synchronicity of strain data among the pile foundations, better overall pier cap rigidity, and stronger load redistribution capacity. Conversely, a smaller principal eigenvalue indicates lower synchronicity of strain data among the pile foundations, degradation of pier cap connection stiffness, and decreased load redistribution capacity.

[0040] The cloud platform inputs the principal eigenvalues ​​into a pre-constructed pier stiffness degradation evaluation function and outputs the pier connection stiffness index at the current moment. The pier stiffness degradation evaluation function satisfies the following formula:

[0041] in, The current moment is the stiffness index of the pier connection, with a value range of [0,1]. The principal eigenvalues ​​of the covariance matrix of the strain difference data at the current moment; These are the baseline principal characteristic values ​​for the pier caps in a non-degradable state during the initial stage of bridge completion. The baseline principal characteristic values ​​are pre-calibrated using static load test data after bridge completion.

[0042] The cloud-based system dynamically adjusts the initial first weight coefficient based on the pile cap connection stiffness index. When the pile cap connection stiffness index falls below a set threshold, the value of the first weight coefficient is reduced. This adjusts the edge weights of the spatial topology graph to reflect the degradation of mechanical transmission performance of the pile cap structure connection over time. The cloud-based system pre-sets thresholds for the pile cap connection stiffness index, including multiple degradation thresholds corresponding to different stiffness degradation levels. The cloud-based system compares the real-time calculated pile cap connection stiffness index with the preset thresholds. When the pile cap connection stiffness index falls below the preset no-degradation threshold, the cloud-based system lowers the initial first weight coefficient according to the level of stiffness degradation and a preset correction ratio. The more severe the stiffness degradation, the greater the reduction in the first weight coefficient. The cloud-based system updates the corrected first weight coefficient in the adjacency matrix of the spatial topology graph, completing the dynamic update of the adjacency matrix. This allows subsequent graph convolution propagation processes to detect changes in mechanical transmission performance caused by pile cap stiffness degradation, ensuring that the node joint features accurately reflect the actual stress coupling state of the pile foundation group.

[0043] In this embodiment, a first type of edge is established through the connection relationship of the pile cap, and a second type of edge is established through the soil stress diffusion radius. This constructs a spatial topology graph that simultaneously includes two mechanical transmission paths: rigid pile cap transmission and soil stress diffusion. The principal eigenvalues ​​are extracted through covariance analysis of strain data, and a pile cap stiffness degradation evaluation function is constructed to achieve dynamic correction of the weights of the first type of edge. This allows the topological structure and edge weights of the spatial topology graph to adaptively adjust as the mechanical properties of the pile cap degrade, ensuring that the information aggregation process of graph convolution always remains consistent with the actual physical and mechanical transmission laws of the pile foundation group.

[0044] In another preferred embodiment, the cloud uses a cluster resource scheduling system to collect real-time operational status data of all computing nodes in the cloud computing cluster. This operational status data includes the number of currently available computing nodes, the real-time CPU utilization, memory utilization, disk I / O utilization, network uplink and downlink bandwidth utilization of each available computing node, and the remaining available computing resource quota, remaining available memory quota, and remaining available network bandwidth quota for each computing node. Based on the collected operational status data, the cloud calculates the real-time load rate of each available computing node. The real-time load rate is a weighted average of CPU utilization, memory utilization, and network bandwidth utilization, with weighting coefficients pre-set according to the computationally intensive and communication-intensive characteristics of the distributed processing subtasks. The cloud marks computing nodes with real-time load rates below a preset load threshold as available scheduling nodes for executing distributed processing subtasks.

[0045] The cloud-based graph partitioning algorithm divides the spatial topology into multiple subgraphs, equal to the number of cloud computing nodes, under the constraints that the number of vertices in each subgraph matches the processing capacity of each cloud computing node and minimizes the sum of the weights of the cut edges. The algorithm used is a multi-level graph partitioning algorithm, comprising a coarsening stage, an initial partitioning stage, and a refinement stage. In the coarsening stage, the cloud merges closely connected vertex pairs into supervertexes through vertex matching, gradually reducing the size of the topology and generating a multi-level coarsened graph. In the initial partitioning stage, the cloud divides the coarsest graph into multiple partitions equal to the number of available scheduling nodes, with the number of vertices in each partition matching the processing capacity of the corresponding scheduling node. In the refinement stage, the cloud reverses the merging process of the coarsening stage, mapping the partitions back to the original spatial topology. Vertex swapping operations are performed at each refinement level to optimize the partitioning results, minimizing the sum of the weights of the cut edges. The basic graph partitioning objective function satisfies the following formula:

[0046] in, Let S be the cut value of the partition, which is the sum of the weights of the cut edges; S is the set of vertices corresponding to the partitioned subgraph. The objective function aims to minimize the cut value, i.e., minimize the sum of the weights of the cut edges, thereby reducing the need for cross-subgraph information interaction during distributed processing.

[0047] During each iteration of the graph partitioning algorithm, the cloud acquires the real-time network communication bandwidth between each cloud computing node. In the refinement phase of the graph partitioning algorithm, before each vertex swapping operation iteration, the cloud uses the cluster network monitoring system to obtain the real-time network communication bandwidth between all available scheduling nodes. This real-time network communication bandwidth is the end-to-end available TCP bandwidth between two nodes, reflecting their data transmission capabilities. The cloud uses the reciprocal of the real-time network communication bandwidth between two nodes as a penalty factor. The magnitude of the penalty factor is inversely proportional to the communication bandwidth between the two nodes; the lower the bandwidth, the larger the penalty factor, indicating a higher cost for cross-node data transmission.

[0048] When calculating the sum of weights for the severed edges, the cloud uses the reciprocal of the real-time network communication bandwidth between the cloud computing nodes containing the two endpoints of the severed edge as a penalty factor, which is added to the weight value of the severed edge to generate the communication-aware segmentation cost. The calculation process of the communication-aware segmentation cost follows the formula:

[0049] in, For communication-aware segmentation cost; Let be the set of vertices corresponding to the subgraph assigned to the p-th cloud computing node. The set of vertices corresponding to the subgraph assigned to the q-th cloud computing node, p≠q indicates that vertex i and vertex j are assigned to different computing nodes and the corresponding edges are cut off; This represents the real-time network communication bandwidth between the p-th cloud computing node and the q-th cloud computing node. This represents the penalty factor between the two computing nodes. The objective function aims to minimize the communication-aware segmentation cost, thereby minimizing the weight of the severed edges while reducing the number of cross-node edge severances between nodes with low communication bandwidth.

[0050] The cloud performs vertex swapping operations with the objective function of minimizing the communication-aware segmentation cost. This reduces the number of edges cut between cloud computing nodes with low communication bandwidth, thereby lowering cross-node communication latency during distributed graph convolution propagation. In each iteration of the refinement phase, the cloud traverses all partition boundaries (vertices connected to vertices in other partitions) and calculates the change in communication-aware segmentation cost after swapping a vertex from the current partition to an adjacent partition. When a swap reduces the communication-aware segmentation cost, the cloud executes the swap and updates the partitioning result. The cloud repeats this vertex swapping iteration process until the decrease in communication-aware segmentation cost falls below a preset threshold or the maximum number of iterations is reached, completing the final graph partitioning result generation. The vertex set corresponding to each subgraph generated by the cloud is matched to the processing capacity of the allocated cloud computing nodes, with the highest edge connection density between vertices within the subgraph and the lowest communication cost for edges connecting vertices across subgraphs.

[0051] The cloud platform packages the multi-dimensional time-series health data associated with the vertices of each subgraph into a distributed processing subtask, and schedules the distributed processing subtask to cloud computing nodes with sufficient idle resources to meet the number of vertices in the subgraph based on the real-time load status of each cloud computing node. The matching relationship between the real-time status of cloud computing nodes and subtask allocation in this embodiment is shown in the table below: Table 3 Matching Table of Real-time Status of Cloud Computing Nodes and Subtask Allocation

[0052] This table records the real-time resource status and allocated subtask information of cloud computing nodes, establishing a matching relationship between the processing capacity of computing nodes and the computing load of subtasks. The cloud can dynamically adjust the scheduling strategy of subtasks based on the real-time data in the table to ensure load balancing in the distributed processing process and avoid resource waste problems such as some computing nodes being overloaded and some computing nodes being idle.

[0053] Based on the final graph partitioning results, the cloud platform packages the preprocessed multi-dimensional temporal health data associated with all vertices of each subgraph, the subgraph's topological structure data, and the mapping relationship data between vertices within the subgraph and global vertices, generating corresponding distributed processing subtasks. Each subtask carries a unique task identifier code and the identifier code of the corresponding computing node. Based on the real-time load status of each available scheduling node, the cloud platform prioritizes scheduling subtasks to the computing nodes with the most remaining available resources, in descending order of the number of vertices in the subgraph. This ensures that the number of vertices in the subgraph corresponding to a subtask allocated to each computing node does not exceed the maximum number of vertices that the computing node's remaining available resources can handle. After completing the subtask scheduling, the cloud platform sends task execution instructions to the corresponding computing nodes. Upon receiving the instructions, the computing nodes begin executing the local temporal feature extraction and graph convolutional propagation processing corresponding to the subtask.

[0054] In this embodiment, a multi-layer graph partitioning algorithm is used to segment the spatial topology graph. The basic objective is to minimize the sum of the weights of the cut edges. At the same time, a penalty factor for the real-time communication bandwidth between nodes is introduced to construct a communication-aware segmentation cost objective function. The partitioning results are optimized iteratively through vertex swapping, which achieves the matching of subtasks with the processing capabilities of computing nodes. It also reduces the cross-node communication latency during the distributed graph convolution propagation process and improves the execution efficiency and resource utilization of cloud-based distributed processing.

[0055] In another preferred embodiment, each cloud computing node segments the multi-dimensional time-series health data within the distributed processing subtask according to time windows to obtain multi-scale time window data. Upon receiving the distributed processing subtask, the cloud computing node parses the multi-dimensional time-series health data within the subtask, obtaining the sampling frequency, sampling timestamp, and time-series data length of each sensor corresponding to each vertex of the pile foundation. The cloud computing node pre-sets multiple sets of time windows of different lengths. The length of the time window is pre-set based on the sensor sampling frequency and the evolution characteristics of the bridge pile foundation health state, including short time windows, medium time windows, and long time windows. Short time windows are used to capture instantaneous impact fluctuation characteristics of the pile foundation sensor data, medium time windows are used to capture short-term load change characteristics, and long time windows are used to capture long-term structural degradation trend characteristics. The cloud computing node performs sliding segmentation of the time-series health data according to the set multi-scale time windows, with the sliding step size set to a preset proportion of the time window length to ensure overlap between adjacent time windows and avoid loss of time-series features.

[0056] Each cloud computing node acquires the sampling frequencies of different pile foundation sensors within the distributed processing subtask. When pile foundation sensors have inconsistent sampling frequencies, each cloud computing node uses the target sampling frequency corresponding to the maximum window length in the multi-scale time window as a benchmark. The target sampling frequency is the least common multiple of the sampling frequencies of all sensors, or a pre-set unified benchmark sampling frequency. For the time-series health data of pile foundation sensors with sampling frequencies lower than the target sampling frequency, each cloud computing node performs upsampling interpolation processing. The upsampling interpolation uses the cubic spline interpolation method, inserting new sampling points between the original sampling points to ensure that the sampling frequency of the interpolated time-series data is consistent with the target sampling frequency. The cubic spline interpolation function satisfies the following formula:

[0057] in, This is a cubic spline interpolation function used to generate interpolated time series data; and These are two adjacent sampling timestamps of the original time-series data; The spline coefficients belonging to the k-th interpolation interval are obtained by solving the boundary conditions of the original sampling point values ​​and the continuity of the second derivative; t is the sampling timestamp after interpolation, which is calculated by the target sampling frequency.

[0058] Each cloud computing node performs downsampling filtering on the pile foundation sensor time-series health data with sampling frequencies higher than the target sampling frequency. The downsampling filtering employs an anti-aliasing low-pass filter. First, the original time-series data is low-pass filtered to remove high-frequency components above the Nyquist frequency of the target sampling frequency. Then, the filtered time-series data is downsampled according to the target sampling frequency to avoid frequency aliasing during downsampling. Each cloud computing node then aligns and segments all pile foundation sensor time-series health data after unified sampling frequency processing according to the timestamps of multi-scale time windows. This ensures that all channel data within each time window are strictly aligned in the time dimension, and that the time step length of the data matrix corresponding to each time window is completely consistent, eliminating the impact of sampling frequency differences on subsequent time-series feature extraction.

[0059] The correspondence between the multi-scale time windows and the one-dimensional convolutional kernel configurations used in this embodiment is shown in the table below: Table 4. Correspondence between multi-scale time windows and one-dimensional convolutional kernel configurations

[0060] This table is used to establish the mapping relationship between multi-scale time windows and one-dimensional convolution kernel configurations. Time windows of different lengths correspond to one-dimensional convolution kernels of different sizes, enabling targeted extraction of temporal features at different time scales. Short time windows correspond to small-sized convolution kernels to capture high-frequency instantaneous fluctuation features, while long time windows correspond to large-sized convolution kernels to capture low-frequency long-term trend features, ensuring that local temporal features can fully cover the multi-scale temporal evolution patterns of pile foundation sensing data.

[0061] Each cloud computing node utilizes a temporal feature extraction network containing multiple parallel one-dimensional convolutional kernels to process multi-scale time window data. Different sized one-dimensional convolutional kernels extract temporal features at different time scales. The temporal feature extraction network constructed by the cloud computing nodes includes multiple parallel one-dimensional convolutional branches, each corresponding to a time window data scale. Each branch contains a one-dimensional convolutional kernel of a corresponding size, a ReLU activation function layer, and a max-pooling layer. The cloud computing nodes input the aligned and segmented multi-scale time window data into the corresponding parallel convolutional branches. The one-dimensional convolutional kernel within each branch performs convolution operations on the temporal data of the corresponding time window, extracting the temporal features at that time scale. The activation function layer performs a non-linear transformation on the features output by the convolution, and the pooling layer performs dimensionality reduction on the features, outputting a temporal feature vector at the corresponding scale.

[0062] Each cloud computing node concatenates temporal features from different time scales along the channel dimension and reduces dimensionality through pooling operations to generate fixed-dimensional local temporal features. These local temporal features characterize the local temporal evolution pattern of the pile foundation sensing data within the corresponding time window. The concatenation process of multi-scale features satisfies the following formula:

[0063] in, For vertex i, the concatenated multi-scale temporal features are shown. This is the short-time feature vector output by the short-time window branch; This is the mid-time series feature vector output by the mid-time window branch; This is the long-term temporal feature vector output by the long-term window branch; The concatenation operation at the channel dimension merges feature vectors of different scales into a feature vector containing multi-scale temporal information. Cloud computing nodes then perform global average pooling on the concatenated multi-scale temporal features to reduce dimensionality and generate fixed-dimensional local temporal features. This ensures that the dimensionality of the local temporal features corresponding to each vertex is completely consistent, providing a unified format of input features for subsequent graph convolution propagation.

[0064] For each vertex within the distributed processing subtask, each cloud computing node collects the local temporal features of its directly connected neighboring vertices. Based on the subgraph's topology data, the cloud computing nodes obtain a list of neighboring vertices for each vertex within the subgraph. This list includes the IDs of all vertices directly connected to the current vertex via edges, their corresponding local temporal features, the 3D geographic coordinates of their respective vertices, and the edge weight coefficients. Following this list, the cloud computing nodes collect the local temporal features of all neighboring vertices, forming a neighborhood feature set for the current vertex.

[0065] Each cloud computing node acquires the spatial relative coordinates of adjacent vertices with respect to the current vertex and the longitudinal slope inclination angle of the bridge. The cloud computing nodes also acquire the bridge's alignment design parameters and extract the longitudinal slope inclination angle, which is the angle between the bridge's centerline and the horizontal plane, representing the bridge's longitudinal gradient. The cloud computing nodes establish a bridge alignment coordinate system with the 3D geographic coordinates of the pile foundation corresponding to the current vertex as the origin. The X-axis of this coordinate system follows the bridge's forward direction, the Z-axis is vertically upward, and the Y-axis is horizontal. The cloud computing nodes then transform the 3D geographic coordinates of the pile foundations corresponding to adjacent vertices into this coordinate system, obtaining the spatial relative coordinates of the adjacent vertices with respect to the current vertex. Based on the sign of the X-axis component of the spatial relative position coordinates of adjacent vertices, and combined with the longitudinal slope inclination angle of the bridge, cloud computing nodes divide adjacent vertices into upstream force transmission vertices and downstream force transmission vertices. Adjacent vertices with negative X-axis components are located upstream of the current vertex and are classified as upstream force transmission vertices. The load of the upstream vertex will be transmitted downstream along the longitudinal slope of the bridge, producing a positive superposition effect on the force state of the current vertex. Adjacent vertices with positive X-axis components are located downstream of the current vertex and are classified as downstream force transmission vertices. The load of the downstream vertex has a damping and canceling effect on the force state of the current vertex.

[0066] Each cloud computing node assigns a positive transmission aggregation coefficient to the upstream force transmission vertex and a negative damping aggregation coefficient to the downstream force transmission vertex. The cloud computing nodes calculate the values ​​of the positive transmission aggregation coefficient and the negative damping aggregation coefficient based on the longitudinal slope inclination angle of the bridge. The larger the longitudinal slope inclination angle, the larger the value of the positive transmission aggregation coefficient and the larger the absolute value of the negative damping aggregation coefficient, consistent with the physical law of load transmission downstream along the longitudinal slope under gravity. The value range of the positive transmission aggregation coefficient is (0,1], and the value range of the negative damping aggregation coefficient is [-1,0].

[0067] Each cloud computing node multiplies the local temporal features of adjacent vertices by their corresponding positive propagation aggregation coefficient or negative damping aggregation coefficient, and then sums them to generate a neighborhood aggregation feature with directionality awareness. The calculation process of directional neighborhood aggregation satisfies the following formula:

[0068] in, The neighborhood aggregation feature corresponding to vertex i, which has directional awareness capability; Let i be the set of vertices upstream of vertex i that are subject to force transmission. Let i be the set of vertices downstream of vertex i that are subject to force transmission. The forward-conduction polymerization coefficient; It is the negative damping polymerization coefficient; The edge weight coefficient between vertex i and vertex j; Let be the local temporal feature corresponding to vertex j.

[0069] Each cloud computing node nonlinearly fuses the local temporal features of a vertex with its neighborhood aggregation features, and then iteratively updates these features through a multi-layer graph convolutional network. This allows the feature information of a single vertex to be transmitted along the path of the spatial topology graph to vertices outside the multi-layer network, generating joint node features that contain multi-order spatial physical and mechanical coupling relationships. The iterative update process of the multi-layer graph convolution satisfies the following formula:

[0070] in, Let be the weight matrix of the vertex features in the l-th graph convolutional layer; Let be the weight matrix of the neighborhood aggregation features in the l-th graph convolutional layer; The neighborhood aggregation feature is calculated by the l-th graph convolutional layer. Let L be the bias vector of the l-th graph convolutional layer; This is a non-linear activation function. After iterative updates through L layers of graph convolutional layers, the features of vertex i contain the feature information of all vertices in its L-order neighborhood, which means it includes multi-order spatial physical and mechanical coupling relationships, ultimately generating the joint node features corresponding to vertex i. ; For vertex i at the th Updated features of the output of the layer graph convolutional layer.

[0071] The cloud master node inputs the aggregated joint features of the nodes into the spatial attention network. These joint features include location encoding information. After generating the joint features, each cloud computing node adds location encoding information to each feature. This location encoding information is generated based on the 3D geographic coordinates of the corresponding vertex, using a trigonometric function location encoding method to map the 3D geographic coordinates into a location encoding vector with the same dimension as the joint feature. The location encoding vector is added to the joint feature vector, resulting in the joint feature containing the spatial location information of the corresponding pile foundation. Each cloud computing node then uploads the joint features with added location encoding information to the cloud master node.

[0072] The spatial attention network generates query vectors by linearly transforming the joint features of nodes using a query matrix, and generates key vectors by linearly transforming the joint features of nodes using a key matrix. The cloud master node calculates the dot product of the query vector and all key vectors, and then processes it using a softmax normalization function to generate attention coefficients corresponding to different spatial topological regions. The cloud master node multiplies the attention coefficients with the corresponding joint features of nodes to obtain a weighted spatial feature map, thereby strengthening the topological feature representation of regions carrying critical force paths. The cloud master node concatenates all the converged joint features of nodes to generate a global feature matrix of dimension N×F, where N is the total number of vertices and F is the dimension of the joint features of nodes. The spatial attention network contains three parallel linear transformation layers: a query linear layer, a key linear layer, and a value linear layer. The input to each of the three linear layers is the global feature matrix, and the outputs are the query matrix Q, the key matrix K, and the value matrix V, respectively. All three matrices have a dimension of N×d, where d is the dimension of the attention head. The cloud master node calculates the dot product of the query matrix Q and the transpose of the key matrix K, scales it by the square root of d, and then normalizes it using the softmax function to generate an attention weight matrix. Each element of the attention weight matrix corresponds to the attention coefficient between two vertices. The cloud master node multiplies the attention weight matrix with the value matrix V to generate a weighted global feature matrix. In the weighted global feature matrix, the features of the topological regions corresponding to key force paths are enhanced, while the interference features of non-critical regions are weakened. The cloud master node performs global average pooling on the weighted global feature matrix to generate a fixed-dimensional overall health status vector of the bridge pile foundation group. This overall health status vector can be directly input into the subsequent health status assessment model to generate the health level assessment results and damage warning information of the bridge pile foundation group. Figure 6 .

[0073] In this embodiment, multi-scale evolution features of pile foundation time-series data are extracted through multi-scale time window segmentation and parallel one-dimensional convolution kernels. Cubic spline interpolation and anti-aliasing low-pass filtering are used to align the time dimensions of sensor data from different sampling frequencies, ensuring the accuracy of time-series feature extraction. Upstream and downstream stress vertices are divided by the bridge's longitudinal slope inclination angle and relative position coordinates, and corresponding aggregation coefficients are assigned to generate neighborhood aggregation features with directional awareness capabilities. This makes the information aggregation process of graph convolution more consistent with the physical laws of bridge load transmission along the longitudinal slope. By adding a spatial attention network with position encoding, feature weight allocation for different spatial topological regions is achieved, strengthening the feature expression of key stress paths. The resulting overall health state vector comprehensively and accurately reflects the spatiotemporal coupled stress state and health level of the bridge pile foundation group.

Claims

1. A cloud-based method for processing bridge pile foundation health data, characterized in that, include: The cloud receives multi-dimensional time-series health data uploaded by sensors of each pile foundation of the bridge, and constructs a spatial topology map based on the physical spatial distribution structure of each pile foundation of the bridge. The cloud platform divides the multidimensional time-series health data into multiple distributed processing subtasks according to the connected components of the spatial topology graph, and assigns the multiple distributed processing subtasks to the corresponding cloud computing nodes. Each cloud computing node extracts local temporal features of the multidimensional time-series health data within the corresponding distributed processing subtask, and performs graph convolution propagation on the edges of the spatial topology graph to generate joint node features containing the spatial coupling relationship of the pile foundation; The cloud master node aggregates the joint features of the nodes output by each cloud computing node, allocates feature weights for different spatial topological regions through an attention mechanism, and outputs a vector of the overall health status of the bridge pile foundation group that integrates spatiotemporal features. The graph convolutional propagation maps the physical and mechanical transmission laws between pile foundations to the information aggregation process in the spatial topological graph.

2. The cloud-based processing method for bridge pile foundation health data according to claim 1, characterized in that, The step of constructing a spatial topology map based on the physical spatial distribution structure of each pile foundation of the bridge includes: acquiring the design drawing data of the bridge from the cloud and extracting the geographical coordinates of each pile foundation and the connection relationship of the pile caps; The cloud platform uses each pile foundation as the vertex of the spatial topology graph and establishes a first type of edge for the pile foundation vertices that are connected under the same pile cap. The cloud-based calculation determines the geographical distance between any two pile foundation vertices that do not have a pile cap connection relationship. When the geographical distance is less than the preset soil stress diffusion radius, a second type of edge is established between the two pile foundation vertices. The cloud platform configures a first weight coefficient for the first type of edge and a second weight coefficient based on the geographical distance decay for the second type of edge, and combines them to generate a weighted spatial topology graph adjacency matrix.

3. The cloud-based processing method for bridge pile foundation health data according to claim 1, characterized in that, The step of dividing the multidimensional time-series health data into multiple distributed processing subtasks according to the connected components of the spatial topology graph includes: the cloud obtaining the number of currently available cloud computing nodes and the real-time load status of each cloud computing node; The cloud uses a graph partitioning algorithm to divide the spatial topology graph. Under the conditions that the number of vertices in each subgraph matches the processing capacity of each cloud computing node and the sum of the weights of the cut edges is minimized, the spatial topology graph is divided into multiple subgraphs equal to the number of cloud computing nodes. The cloud platform packages the multi-dimensional time-series health data associated with the vertices of each subgraph into a distributed processing subtask, and schedules the distributed processing subtask to cloud computing nodes with idle resources that meet the number of vertices required by the subgraph based on the real-time load status of each cloud computing node.

4. The cloud processing method for bridge pile foundation health data according to claim 1, characterized in that, The step of extracting local temporal features of the multidimensional time-series health data within the corresponding distributed processing subtask includes: each cloud computing node segments the multidimensional time-series health data within the distributed processing subtask according to a time window to obtain multi-scale time window data. Each of the cloud computing nodes uses a temporal feature extraction network containing multiple parallel one-dimensional convolutional kernels to process the multi-scale time window data, wherein one-dimensional convolutional kernels of different sizes are respectively used to extract temporal features at different time scales. Each cloud computing node concatenates the time-series features of different time scales along the channel dimension and reduces the dimensionality through pooling operations to generate the local time-series features of fixed dimensions. The local time-series features represent the local time-series evolution pattern of the pile foundation sensing data within the corresponding time window.

5. The cloud-based processing method for bridge pile foundation health data according to claim 2, characterized in that, The step of performing graph convolution propagation on the local temporal features along the edges of the spatial topology graph to generate node joint features containing the spatial coupling relationship of the pile foundation includes: each cloud computing node collecting local temporal features of adjacent vertices directly connected to each vertex within the distributed processing subtask; Each cloud computing node multiplies the local temporal features of the adjacent vertices with the weight coefficients of the corresponding edges and then sums them to obtain the neighborhood aggregation features; Each cloud computing node nonlinearly fuses the local temporal features of the vertex itself with the neighborhood aggregation features, and iteratively updates them through a multi-layer graph convolutional network, so that the feature information of a single vertex is transmitted to vertices outside the multi-layer along the path of the spatial topology graph, generating node joint features containing multi-order spatial physical and mechanical coupling relationships.

6. The cloud-based processing method for bridge pile foundation health data according to claim 5, characterized in that, The method of allocating feature weights for different spatial topological regions through an attention mechanism includes: the cloud master node inputting the aggregated joint features of the nodes into the spatial attention network, wherein the joint features of the nodes contain location encoding information; The spatial attention network generates a query vector by performing a linear transformation on the joint features of the nodes using a query matrix, and generates a key vector by performing a linear transformation on the joint features of the nodes using a key matrix. The cloud master node calculates the dot product of the query vector and all key vectors, and processes it through the softmax normalization function to generate the attention coefficients corresponding to the different spatial topological regions. The cloud master node multiplies the attention coefficient with the corresponding node joint features to obtain a weighted spatial feature map.

7. The cloud-based processing method for bridge pile foundation health data according to claim 2, characterized in that, The step of configuring a first weight coefficient for the first type of edge includes: obtaining real-time strain difference data of pile foundation sensors under the same pile cap connection relationship from the cloud. The cloud calculates the covariance matrix of the real-time strain difference data and extracts the principal eigenvalues ​​of the covariance matrix; The cloud platform inputs the main feature value into a pre-constructed pier stiffness degradation evaluation function and outputs the pier connection stiffness index at the current moment. The cloud platform dynamically corrects the initial first weight coefficient based on the pier connection stiffness index. When the pier connection stiffness index is lower than a set threshold, the value of the first weight coefficient is reduced. By adjusting the edge weights of the spatial topology graph, the degradation of mechanical transmission performance of the pier structure connection state over time is reflected.

8. The cloud processing method for bridge pile foundation health data according to claim 3, characterized in that, The method of segmenting the spatial topology graph using a graph partitioning algorithm further includes: the cloud acquiring the real-time network communication bandwidth between each cloud computing node during each iteration of the graph partitioning algorithm. When calculating the total weight of the severed edges, the cloud uses the reciprocal of the real-time network communication bandwidth between the cloud computing nodes where the two endpoints of the severed edges are located as a penalty factor, which is added to the weight value of the severed edges to generate a communication-aware segmentation cost. The cloud performs vertex swapping operations with the objective function of minimizing the communication-aware segmentation cost.

9. The cloud-based processing method for bridge pile foundation health data according to claim 4, characterized in that, The step of segmenting the multi-dimensional time-series health data within the distributed processing subtask according to time windows to obtain multi-scale time window data includes: each cloud computing node acquiring the sampling frequency of different pile foundation sensors within the distributed processing subtask; When there are pile foundation sensors with inconsistent sampling frequencies, each cloud computing node uses the target sampling frequency corresponding to the maximum window length in the multi-scale time window as a benchmark to perform upsampling interpolation processing on the time series health data of the pile foundation sensor with low sampling frequency, and downsampling filtering processing on the time series health data of the pile foundation sensor with high sampling frequency. Each cloud computing node aligns and segments the time-series health data of the pile foundation sensor, which has been processed with a unified sampling frequency, according to the multi-scale time window.

10. The cloud-based processing method for bridge pile foundation health data according to claim 5, characterized in that, The step of multiplying the local temporal features of the adjacent vertices with the weight coefficients of the corresponding edges and summing them to obtain neighborhood aggregation features includes: each cloud computing node obtaining the spatial relative position coordinates of the adjacent vertices relative to the vertex and the longitudinal slope inclination angle of the bridge. Each cloud computing node divides the adjacent vertices into upstream force transmission vertices and downstream force transmission vertices based on the spatial relative position coordinates and the longitudinal slope inclination angle of the bridge; Each cloud computing node assigns a positive transmission aggregation coefficient to the upstream force transmission vertex and a negative damping aggregation coefficient to the downstream force transmission vertex; Each cloud computing node multiplies the local temporal features of the adjacent vertices by the corresponding positive propagation aggregation coefficient or negative damping aggregation coefficient and then sums them to generate neighborhood aggregation features with directional awareness capabilities.

Images