An internet-of-things-based traffic infrastructure digitalization wisdom management method and system

Abstract

The application relates to the technical field of intelligent management and maintenance of traffic infrastructure, and discloses a traffic infrastructure digitalization intelligent management and maintenance method and system based on the Internet of Things, wherein the method comprises the following steps: collecting multi-dimensional time sequence monitoring data and marking propagation mechanism attributes; calculating monitoring point space correlation weights, generating a space correlation graph structure, and decomposing the space correlation graph structure into three subgraphs, namely, an urgent layer, a medium-speed layer and a slow layer, according to degradation rate grades, configuring matching time window parameters for each subgraph; extracting degradation rate vectors according to each layer time window; aggregating neighborhood degradation state information through a hierarchical graph attention network; performing cross-layer gated message passing and outputting a comprehensive degradation trajectory prediction curve; identifying group degradation modes and aggregating the group degradation modes into hierarchical group management and maintenance areas; and generating an intensive intelligent management and maintenance scheduling scheme.

Description

Technical Field

[0001] This invention relates to the field of intelligent management and maintenance technology for transportation infrastructure, specifically to a digital intelligent management and maintenance method and system for transportation infrastructure based on the Internet of Things. Background Technology

[0002] Infrastructure such as bridge approach roads, road sections, and culvert groups in transportation corridors continuously output multi-dimensional time-series monitoring data, including structural stress, vibration frequency, crack width, and settlement displacement, through IoT sensors, providing data support for maintenance decisions. Existing maintenance systems typically model and predict degradation trajectories independently for each monitoring point and use a unified fixed retrospective period as a time window to determine the spatiotemporal correlation between abnormal events.

[0003] However, degradation propagation driven by different physical pathways has drastically different timescales: surface runoff can affect adjacent roadbeds within hours, subsurface seepage can propagate to adjacent bridge abutment foundations within weeks, and geological subsidence may take months to manifest in downstream facilities. Independent modeling cannot use monitoring signals from facilities that have already degraded as early warning evidence for neighboring facilities, resulting in the ineffective capture of precursors to cluster degradation. Under a uniform time window, rapidly propagating instantaneous correlation signals are smoothed and diluted in a long window, while slowly propagating gradual signals have not yet accumulated to a detectable level in a short window. This leads to a slow response of cluster warnings to rapid degradation events and insufficient identification of the root causes of slow propagation. Summary of the Invention

[0004] This invention provides a digital intelligent management and maintenance method and system for transportation infrastructure based on the Internet of Things, which solves the technical problems in related technologies where the monitoring data of transportation infrastructure fails to reflect the spatial correlation degradation and propagation law between monitoring points and cannot achieve cross-rate level collaborative early warning and intensive scheduling.

[0005] This invention discloses a digital intelligent management and maintenance method for transportation infrastructure based on the Internet of Things, including: collecting multi-dimensional time-series monitoring data from various monitoring points and labeling the degradation rate level of the spatial correlation between monitoring points;

[0006] Calculate the spatial correlation weights between monitoring points to generate the infrastructure spatial correlation graph structure. Divide the edge set into rapid, medium and slow sub-graphs according to the degradation rate level, and set the matching time window length for each sub-layer.

[0007] Perform sliding window segmentation processing according to the time window length of each sub-layer, and extract the degradation rate vector at each time scale;

[0008] The degradation rate vectors at each scale are input into the graph attention network of the corresponding sub-layer to aggregate the degradation state information in the neighborhood and output the degradation state embedding vector of each layer.

[0009] The gating fusion unit performs cross-layer message passing, controls the slow layer degradation signal to be passed step by step to the medium and fast layers, and outputs the comprehensive degradation trajectory prediction curve.

[0010] Based on the deviation comparison between the comprehensive degradation trajectory prediction curve and the independent prediction curve, spatially related early warning nodes are identified, aggregated into hierarchical group management areas, and an intensive intelligent management and scheduling scheme is generated.

[0011] Furthermore, the calculation of spatial correlation weights between monitoring points, for each monitoring point... With monitoring points Spatial correlation weight The calculation method is as follows:

[0012]

[0013] in, For monitoring points With monitoring points Geographical distance between them This is the distance decay function value. For geological attribute similarity, This is an indicator value for drainage zone affiliation; when two monitoring points belong to the same drainage zone... otherwise , , , The preset weighting coefficients and The distance decay function The calculation method is as follows:

[0014]

[0015] in, This represents the minimum geographical distance between all point pairs in the current set of monitoring points. The maximum geographical distance between all point pairs; the geological attribute similarity. The cosine similarity algorithm was used to calculate the similarity based on the geological attribute feature vectors of each monitoring point.

[0016] Furthermore, the degradation rate levels include rapid, medium, and slow levels. The rapid level corresponds to a conduction process on the order of several hours to several days, the medium level corresponds to a conduction process on the order of several days to several weeks, and the slow level corresponds to a conduction process on the order of several weeks to several months. The rapid, medium, and slow subgraphs share the same set of nodes, only their edge sets differ. The rapid subgraph is configured with a short time window and high sampling resolution, the medium subgraph is configured with a medium time window and medium sampling resolution, and the slow subgraph is configured with a long time window and low sampling resolution.

[0017] Furthermore, the extraction of the degradation rate vector includes: calculating the slope of the mean change, the gradient of variance increase, and the growth rate of peak frequency for each monitoring indicator within each sliding window segment, and concatenating them into a degradation rate vector; wherein, the slope of the mean change is the linear regression slope of the time-series mean of each monitoring indicator within the window, the gradient of variance increase is the ratio of the difference between the variances of the first half and the second half of the window to the window length, and the growth rate of peak frequency is the ratio of the number of peak events exceeding the threshold in the second half of the window to the first half minus one; the three statistical components are respectively standardized using Z-score to eliminate dimensional differences.

[0018] Furthermore, the graph attention networks of the three rate layers are independent and do not share parameters. The input layer of each graph attention network receives the degradation rate vector of the corresponding rate layer. The attention aggregation layer calculates the attention coefficients between node pairs on the edge set of the corresponding subgraph and weights and aggregates the neighborhood features. The output layer is a fully connected layer that generates the degradation state embedding vector of each node. The three graph attention networks use the labeled degradation state of the historical degradation events of each monitored node as training labels and use the mean squared error loss function for end-to-end training.

[0019] Further, the cross-layer message passing through the gating fusion unit includes: inputting the slow layer degradation state embedding vector into the cross-layer gating fusion unit to generate a gating signal to be passed to the intermediate layer; performing an element-wise multiplication operation on the gating signal and the intermediate layer degradation state embedding vector to obtain the intermediate layer update embedding; inputting the intermediate layer update embedding into the second-level gating fusion unit to generate a gating signal to be passed to the rapid layer; performing an element-wise multiplication operation on the gating signal and the rapid layer degradation state embedding vector to obtain the rapid layer update embedding; concatenating the slow layer degradation state embedding vector, the intermediate layer update embedding, and the rapid layer update embedding and inputting them into the fully connected prediction layer to output the comprehensive degradation trajectory prediction curve; wherein, the cross-layer gating fusion unit outputs a gating coefficient vector by linearly transforming the input embedding vector and then using the sigmoid activation function, with each element taking values ​​between 0 and 1.

[0020] Furthermore, the method for identifying spatially correlated early warning nodes based on deviation comparison includes: calculating the predicted degradation time deviation between the comprehensive degradation trajectory prediction curve of each monitoring node and the independent prediction curve of that node; marking the node as a spatially correlated early warning state when neighborhood propagation evidence causes the predicted degradation time of a node to be shifted forward by more than a preset proportion; retrieving node combinations in the set of nodes in the spatially correlated early warning state that simultaneously have connected paths in the slow-speed subgraph and the rapid-speed or medium-speed subgraph, and identifying the node combinations as cross-rate group degradation patterns; aggregating the correlated early warning nodes involving multi-rate propagation into hierarchical group management areas, each hierarchical group management area containing a set of early warning nodes sharing at least one cross-rate propagation path.

[0021] Furthermore, the generation of the intensive intelligent management and scheduling scheme includes calculating the urgency score of each node within the hierarchical group management area, and for monitoring nodes... Urgency rating The calculation method is as follows:

[0022]

[0023] in, For comprehensive degradation trajectory prediction curve prediction monitoring nodes The remaining time to reach the maintenance threshold. To maintain the reference time base corresponding to the threshold, For monitoring nodes The normalized coded value of the rate level. For monitoring nodes Frequency of occurrence as an intermediate node in all cross-rate propagation paths within the hierarchical group management area. This is the maximum value of that frequency. , , For preset weighting coefficients and .

[0024] Furthermore, the generation of the intensive intelligent maintenance scheduling scheme also includes: evaluating the reusable ratio of shared maintenance resources among nodes in the hierarchical group maintenance area, inputting the urgency score, root cause propagation rate level, reusable ratio of shared resources, and regional spatial range as constraints into the regional aggregation scheduling algorithm, and outputting an intensive intelligent maintenance scheduling scheme that includes regional maintenance time periods, repair urgency weights, shared equipment allocation paths, and batch execution sequences; wherein, in the batch execution sequence, nodes located at the root cause position of the slow layer are repaired first to block the degradation propagation path, and then downstream affected nodes are maintained.

[0025] This invention discloses a digital intelligent management and maintenance system for transportation infrastructure based on the Internet of Things (IoT), used to execute the above-mentioned digital intelligent management and maintenance method for transportation infrastructure based on the IoT, including: a data acquisition module, used to collect multi-dimensional time-series monitoring data of each monitoring point through an IoT sensor group and to label the degradation rate level of the spatial correlation between monitoring points;

[0026] The spatial association graph generation module is used to calculate the spatial association weights between monitoring points and generate the infrastructure spatial association graph structure. According to the degradation rate level, the edge set is divided into rapid layer subgraph, medium-speed layer subgraph and slow layer subgraph, and the matching time window length is set for each sub-layer.

[0027] The multi-scale feature extraction module is used to perform sliding window segmentation processing according to the time window length of each sub-layer, and extract the degradation rate vector at each time scale.

[0028] The hierarchical graph attention aggregation module is used to input the degradation rate vectors at each scale into the graph attention network of the corresponding sub-layer to aggregate the neighborhood degradation state information and output the degradation state embedding vector of each layer.

[0029] The cross-layer gating fusion module is used to perform cross-layer message passing through the gating fusion unit, control the slow layer degradation signal to be passed to the medium-speed layer and the fast layer step by step, and output the comprehensive degradation trajectory prediction curve.

[0030] The group early warning identification module is used to identify spatially related early warning nodes based on the deviation comparison between the comprehensive degradation trajectory prediction curve and the independent prediction curve, and aggregate them into hierarchical group management areas;

[0031] The scheduling scheme generation module is used to generate an intensive and intelligent management scheduling scheme based on the constraints of hierarchical group management areas.

[0032] The beneficial effects of this invention are as follows:

[0033] The edge sets of the infrastructure spatial correlation graph structure are decomposed into fast-propagation, medium-propagation, and slow-propagation subgraphs according to their degradation rate. A matching time window length and sampling resolution are configured for each subgraph. Fast-propagation signals maintain the sharpness of instantaneous correlation within a short time window without being smoothed out by a long time window, while slow-propagation signals achieve sufficient temporal accumulation within a long time window without being truncated by a short time window. This overcomes the problem of mutual interference between signals of different rates within a unified time window.

[0034] The hierarchical graph attention network independently aggregates neighborhood degradation evidence on its respective rate layer, enabling the degradation status of adjacent facilities to serve as early warning information for the node in prediction, thus overcoming the limitation of independent modeling in failing to capture early signs of group degradation.

[0035] Cross-layer gated message passing adopts a unidirectional cascade structure, allowing the root cause degradation signal of the slow layer to penetrate to the fast response end step by step, which helps to effectively identify the root cause of slow propagation.

[0036] By identifying spatially related early warning nodes through deviation comparison analysis and aggregating them into hierarchical group management areas, spatially related facilities are incorporated into a unified management unit. Priority is given to repairing root cause nodes to block the degradation propagation path, which helps to achieve regional preventive management and scheduling. Attached Figure Description

[0037] Figure 1 This is a flowchart of the Internet of Things-based digital intelligent management and maintenance method for transportation infrastructure provided in an embodiment of the present invention;

[0038] Figure 2This is a schematic diagram of the spatial association weight comparison of typical node pairs provided in the embodiments of the present invention;

[0039] Figure 3 This is a schematic diagram of the normalized vector components of the settlement displacement degradation rate of node 1 provided in an embodiment of the present invention;

[0040] Figure 4 This is a schematic diagram of the typical node embedding vector L2 norm distribution provided in an embodiment of the present invention;

[0041] Figure 5 This is a schematic diagram of the cross-layer gating fusion coefficient of node 3 provided in an embodiment of the present invention;

[0042] Figure 6 This is a schematic diagram comparing the independently predicted remaining lifetime of each node with the overall predicted forward shift provided in an embodiment of the present invention;

[0043] Figure 7 This is a schematic diagram comparing the forward shift ratio of the comprehensive prediction of each node with the early warning threshold provided in the embodiment of the present invention;

[0044] Figure 8 This is a schematic diagram of the repair urgency weight of the intensive intelligent maintenance and scheduling scheme provided in the embodiment of the present invention;

[0045] Figure 9 This is a schematic diagram of the scatter distribution of node pairs' geographical distance and spatial association weights provided in an embodiment of the present invention. Detailed Implementation

[0046] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0047] At least one embodiment of the present invention discloses a digital intelligent management and maintenance method for transportation infrastructure based on the Internet of Things, see [link to relevant documentation]. Figure 1 This includes the following steps:

[0048] Step 1: Obtain multi-dimensional time-series monitoring data of the Internet of Things and label the spatial correlation attributes between monitoring points.

[0049] By deploying IoT sensor clusters on various infrastructures along the transportation corridor, multi-dimensional time-series monitoring data such as structural stress, vibration frequency, crack width, and settlement displacement are continuously collected at each monitoring point. Simultaneously, the geographical coordinates, geological attribute labels, and drainage zoning codes of each monitoring point are acquired.

[0050] Based on the classification of physical connectivity between facilities, each spatial relationship between monitoring points is labeled with its propagation mode type and corresponding degradation rate level. The degradation rate levels include rapid, medium, and slow.

[0051] It should be noted that the above classification of degradation rate levels is based on the physical characteristics of degradation transmission between facilities. The rapid level corresponds to a transmission process on the order of several hours, such as surface runoff erosion; the medium-rapid level corresponds to a transmission process on the order of several weeks, such as underground seepage; and the slow level corresponds to a transmission process on the order of several months, such as geological subsidence. The degradation rate level marked for each spatial relationship is uniquely determined by the physical connectivity corresponding to that spatial relationship.

[0052] Step 2: Calculate the spatial correlation weight matrix to generate a hierarchical subgraph structure.

[0053] Based on the geographical proximity, geological attribute similarity, and drainage zoning relationship of each monitoring point, the spatial association weight between monitoring points is calculated, and an infrastructure spatial association graph structure with monitoring points as nodes and spatial association weights as edges is generated.

[0054] Based on the degradation rate level labeled on each edge, the edge set of the infrastructure spatial association graph structure is divided into rapid, medium, and slow subgraphs. A matching time window length and sampling resolution parameter are set for each subgraph.

[0055] It should be noted that the above spatial correlation weights Calculate as follows. Set up monitoring points. With monitoring points The geographical distance between them is Geological attribute similarity is The shared indicator value for drainage zones is ,but:

[0056]

[0057] in For monitoring points With monitoring points Spatial correlation weights between them This is the index of the first monitoring point. For the index of the second monitoring point, For monitoring points With monitoring points Geographical distance between them For monitoring points With monitoring points Similarity in geological properties between them For monitoring points With monitoring points The drainage zones between them share the same indicator value. , , The weighting coefficients of each term satisfy the following conditions: , and The monitoring point index is used to iterate through all monitoring point pairs. The maximum value of the reciprocal of the geographic distance between all monitoring points is used to perform range-based normalization on the reciprocal of the geographic distance, thus normalizing the range of values ​​for the geographic distance term. Geological attribute similarity The value is calculated based on the matching degree of the geological attribute labels of the two monitoring points, and the range is [value range missing]. The value is 1 if the labels are completely identical, and 0 if they are completely inconsistent. (Drainage zone shared indicator value) The value is 1 if the two monitoring points belong to the same drainage zone, and 0 otherwise. All three items are dimensionless quantities, and their values ​​range from [value range missing]. The weighted sum can be calculated directly.

[0058] It should be noted that the above time window lengths are set as follows: the rapid subgraph corresponds to a short time window and a high sampling resolution to maintain the temporal sharpness of the rapidly propagating signal; the slow subgraph corresponds to a long time window and a low sampling resolution to allow the progressive signal to accumulate sufficiently; the window length and sampling resolution of the medium-speed subgraph are between those of the rapid and slow subgraphs.

[0059] In this embodiment, to accurately match the time window parameters of different rate layers with the physical timescale of degradation propagation, the time window length is determined as follows: the time window length for the rapid rate submap is between 1 hour and 24 hours; the time window length for the medium rate submap is between 1 day and 4 weeks; and the time window length for the slow rate submap is between 1 month and 6 months. The sampling resolution is inversely configured with the time window length; the shorter the window, the smaller the sampling interval.

[0060] Step 3: Perform multi-scale sliding window segmentation on the multi-dimensional time series monitoring data to extract the degradation rate vector at each time scale.

[0061] For the multidimensional time-series monitoring data of each monitoring node, sliding window segmentation is performed according to three time window lengths corresponding to rapid, medium, and slow subgraphs. Within each time window, the degradation rate vector of each monitoring dimension is extracted to generate three sets of scale-specific node time-series degradation features.

[0062] It should be noted that the aforementioned degradation rate vector includes three components: the slope of the mean change, the gradient of the variance increase, and the peak frequency growth rate. Let the sampling sequence of a certain monitoring dimension within the current time window be... ,in For the first The monitoring values ​​at each sampling time, For sampling time index and , This represents the total number of sampling points within the window.

[0063] The slope of the mean change is the slope of the linear fit of the mean of the monitored quantities within that time window, specifically determined by the sampling sequence. About time index The slope of the regression line obtained by performing least squares linear regression is the slope of the mean change, reflecting the overall trend of change.

[0064] The variance increment gradient is the rate at which the variance of the monitored quantity increases over time within this time window. The current time window is divided into two equal segments, and the variance of the first segment is calculated separately. and subsequent sub-segment variance ,in The variance of the sampled values ​​within the preceding sub-segment. The variance of the sampled values ​​within the subsequent sub-segment is defined as the variance increasing gradient. This reflects the increasing trend in the amplitude of fluctuations.

[0065] The peak frequency growth rate is the number of times the monitored quantity exceeds a preset peak threshold within the current time window. The number of times corresponding to the previous window The rate of change of, where This represents the number of times the preset peak threshold is exceeded within the current time window. The peak frequency growth rate is defined as the number of times the preset peak threshold is exceeded within the previous time window. Adding 1 to the denominator avoids division by zero when the previous window count is zero, reflecting the evolution trend of the frequency of abnormal peaks.

[0066] Before splicing, the three components are Z-score standardized for the mean change slope, variance increase gradient, and peak frequency growth rate of each monitoring dimension to eliminate the influence of the differences in physical quantities of different monitoring dimensions on the subsequent graph attention network input. The degradation rate vector dimension of each monitoring node at each time scale is equal to the number of monitoring dimensions multiplied by 3.

[0067] Step 4: Input the degradation features at each scale into the graph attention network of the corresponding sub-layer, and output the degradation state embedding vector at each time scale.

[0068] Three sets of scale-specific node degradation features are input into the graph attention networks of the corresponding sub-layers. Each graph attention network takes the node degradation features of the rate layer subgraph as input, the edge set of the rate layer subgraph and the corresponding spatial association weights as adjacency constraints, dynamically calculates the degradation propagation probability weights and aggregates the degradation state information of neighboring nodes, and independently outputs the degradation state embedding vector of each node at that time scale, which integrates neighborhood degradation evidence.

[0069] The input layer of the graph attention network receives the degradation rate vector of each node, with a dimension equal to the dimension of the degradation rate vector at each time scale in step 3. The output layer of the graph attention network is a fully connected layer, outputting the degradation state embedding vector of each node. During training, the measured degradation amount of each monitoring node at the end of the corresponding time window is used as the supervision label, the mean squared error loss function is adopted, and the Adam optimization algorithm is used to update the network parameters.

[0070] It should be noted that the graph attention networks of the three rate layers operate independently without sharing parameters, ensuring that the attention coefficients of each graph attention network are adapted to the degradation propagation characteristics of the corresponding rate layer. The output layer dimension of the graph attention networks in the three rate layers is set to be the same, so that the dimension of the degradation state embedding vectors of the three rate layers is consistent, thus meeting the vector operation requirements of cross-layer gating fusion in step 5.

[0071] Step 5: Perform cross-layer message passing and gating fusion on the degradation state embedding vectors of the three time scales to output the comprehensive degradation trajectory prediction curve.

[0072] For rapid subplots, intermediate subplots, and slow subplots Figure 3 Degradation state embedding vectors at multiple time scales are used to perform cross-layer message passing. The propagation intensity of the root degradation signal from the slow-speed subgraph to the medium-speed and fast-speed subgraphs is controlled by gating fusion, and a comprehensive degradation trajectory prediction curve that fuses spatial propagation evidence from multiple time scales is output.

[0073] Step 5 may specifically include:

[0074] Step 501: Perform a concatenation operation on the degenerate state embedding vectors of the slow layer and the medium-speed layer at the same monitoring node, input the concatenation result into the first gating network, and output the propagation gating coefficients from the slow layer to the medium-speed layer. ,in It is a scalar value.

[0075] Step 502: Based on the propagation gating coefficient ,according to

[0076]

[0077] Calculate the mid-speed layer augmented embedding vector that incorporates information from the slow layer. ,in To enhance the embedding vector of the mid-speed layer after fusing information from the slow layer. Embedded vectors for the degenerate states of the intermediate-speed layer. This represents the propagation gating factor from the slow-speed layer to the intermediate-speed layer. This is the embedding vector for the degenerate state of the slow layer. Since the embedding vectors for the degenerate states of the three rate layers have the same dimension... and With consistent dimensions, scalar multiplication and vector addition are legal operations.

[0078] Step 503: Enhance the embedding vector of the medium-velocity layer for the same monitoring node The embedding vectors of the rapidly degenerate layers are concatenated, and the concatenation result is input into the second gating network, which outputs the propagation gating coefficients from the intermediate-velocity layers to the rapidly degenerate layers. ,in It is a scalar value.

[0079] Step 504: Based on the propagation gating coefficient ,according to

[0080]

[0081] Calculate the integrated degradation state embedding vector that fuses degradation evidence from three timescales. ,in This is an embedding vector for the comprehensive degradation state after fusing degradation evidence from three time scales. Embedded vectors for rapidly degenerate states in the rapid layer. This represents the propagation gating factor from the intermediate-velocity layer to the rapid-velocity layer. This enhances the embedding vector for the mid-speed layer after fusing information from the slow-speed layer. Because... and If the dimensions are the same, scalar multiplication and vector addition are legal operations.

[0082] Step 505: Embed the synthesized degenerate state into the vector The input is the regression prediction layer, and the output is the comprehensive degradation trajectory prediction curve for each monitoring node. The regression prediction layer is a linear mapping layer, and the input is the comprehensive degradation state embedding vector. The output is a sequence of estimated degradation values ​​at each prediction time, i.e., the comprehensive degradation trajectory prediction curve. During training, the measured degradation sequence within the corresponding time period of each monitoring node is used as the supervision label. The mean squared error loss function is adopted, and the Adam optimization algorithm is used to jointly update the parameters of the first gating network, the second gating network, and the regression prediction layer.

[0083] It should be noted that both the first and second gating networks mentioned above are single-layer fully connected networks followed by a sigmoid activation function. The input of the first gating network is the concatenation of the degenerate state embedding vectors of the slow layer and the medium-speed layer, and the output is the scalar gating coefficients. The input to the second gating network is the mid-speed layer augmented embedding vector. The concatenation of the embedding vectors of the rapidly degenerate layers outputs scalar gating coefficients. When the propagation gating coefficient approaches 1, it indicates that the degradation signal from the upstream slow layer has a strong transmission effect on the downstream fast layer; when the propagation gating coefficient approaches 0, it indicates that the cross-layer transmission relationship is weak, and the downstream layer mainly relies on the degradation evidence from its own rate layer. The direction of cross-layer message transmission is from the slow layer subgraph to the medium-speed layer subgraph and then to the fast layer subgraph, following the physical logic of degradation transmission from the root layer to the representation layer level by level.

[0084] In this embodiment, to enable the cross-layer fusion process to distinguish the cross-layer propagation differences of degradation signals from different monitoring dimensions, in steps 501 and 503, the propagation gating coefficient can be expanded from a scalar to a vector form with the same dimension as the degradation state embedding vector. At this time, the first gating network and the second gating network each output a gating vector with the same dimension as the degradation state embedding vector. Each dimension independently controls the cross-layer propagation intensity of the corresponding feature channel. The scalar multiplication in steps 502 and 504 is replaced with element-wise multiplication, allowing structural stress-related features and settlement displacement-related features to obtain differentiated gating weights during cross-layer transmission.

[0085] Step 6: Perform deviation comparison analysis, identify spatial correlation early warning status nodes, and generate hierarchical group management areas.

[0086] A deviation comparison analysis is performed between the comprehensive degradation trajectory prediction curve and the independent prediction curves for each monitoring point. The difference between the predicted time when the comprehensive degradation trajectory prediction curve reaches the preset degradation threshold and the predicted time when the independent prediction curve reaches the preset degradation threshold is calculated. When the predicted time corresponding to the comprehensive degradation trajectory prediction curve is shifted forward by more than a preset proportion compared to the predicted time corresponding to the independent prediction curve, the monitoring node is marked as a spatial correlation early warning state.

[0087] It should be noted that the aforementioned independent prediction curves refer to prediction curves obtained solely based on the historical time-series data of each monitoring point, fitting the degradation trend, and do not include evidence of neighborhood propagation. The preset percentage is a fixed percentage of the independently predicted remaining lifetime. When the combined degradation trajectory prediction curve causes the degradation arrival time to be shifted forward by more than the product of this percentage and the independently predicted remaining lifetime, it is determined that the evidence of neighborhood propagation has a significant degradation acceleration effect on that node.

[0088] Based on the comprehensive degradation trajectory prediction curve, the degradation patterns across rate levels triggered by the root cause of the slow subgraph are identified, and the associated early warning nodes involving multi-rate propagation are aggregated into hierarchical group management areas.

[0089] It should be noted that the identification method for the aforementioned cross-rate-level group degradation pattern is as follows: tracing the connectivity of spatially associated early warning nodes on the slow-speed subgraph, when multiple spatially associated early warning nodes are reachable from the same root node through the edge set of the slow-speed subgraph, and these spatially associated early warning nodes also have associated edges on the fast-speed or medium-speed subgraphs, this group of nodes is identified as a cross-rate-level group degradation pattern triggered by the root of the slow-speed subgraph. The aforementioned root node refers to the monitoring node in the slow-speed subgraph that is directly or indirectly connected to multiple spatially associated early warning nodes and has the largest absolute value of the slope of the mean change in its own slow-speed degradation rate vector. The spatial extent of the hierarchical group management area is determined by the geographic coordinate envelope of all associated early warning nodes and the root node within that hierarchical group management area.

[0090] In this embodiment of the application, to avoid mistakenly including spatially associated early warning nodes that are accidentally triggered by a single rate layer association into the hierarchical group management area, a cross-layer coverage constraint is also imposed on candidate early warning nodes when aggregating the hierarchical group management area. A candidate early warning node must have an association edge with other spatially associated early warning nodes in the same hierarchical group management area in at least two subgraphs of different rate layers in order to be included in the hierarchical group management area.

[0091] Step 7: Generate an intensive smart management and scheduling scheme based on the hierarchical group management area.

[0092] The spatial scope of the hierarchical group maintenance area, the urgency score of each node within the hierarchical group maintenance area, the root cause propagation rate level, and the reusable ratio of shared maintenance resources are used as constraints and input into the regional aggregation scheduling algorithm to generate an intensive intelligent maintenance scheduling scheme based on groups. The intensive intelligent maintenance scheduling scheme includes regional maintenance time periods, repair urgency weights corresponding to the root cause degradation rate level, shared equipment allocation paths, and batch execution sequences.

[0093] It should be noted that the above urgency score is calculated based on the remaining time for each node in the comprehensive degradation trajectory prediction curve to reach the preset degradation threshold. The shorter the remaining time, the higher the urgency score. The repair urgency weight is assigned differently according to the degradation rate level and the node role: slow-level root cause nodes, as the driving source of degradation in the hierarchical group maintenance area, have a higher repair urgency weight than non-root cause early warning nodes in the same hierarchical group maintenance area; rapid-level early warning nodes, due to their fast degradation propagation speed and short remaining response time, are prioritized for maintenance over medium-speed early warning nodes.

[0094] It should be noted that the aforementioned reusable ratio of shared maintenance resources refers to the degree of overlap in the equipment and manpower required for maintenance operations at different nodes within the same hierarchical group management area. When two adjacent spatially related early warning status nodes require the same type of maintenance equipment and their geographical distance is within the range of a single equipment relocation, the higher value of the reusable ratio of shared maintenance resources is taken, and the regional aggregation scheduling algorithm accordingly arranges the two spatially related early warning status nodes for continuous operations in the same batch.

[0095] In this embodiment, to improve the utilization efficiency of shared maintenance resources, the geographical distribution density of nodes within the same hierarchical group maintenance area is also included as a constraint when generating an intensive intelligent maintenance scheduling scheme. For node groups with compact spatial distribution, centralized maintenance in the same batch is prioritized to reduce equipment relocation costs. For node groups with discrete spatial distribution but sharing the same slow-layer subgraph root source, the repair of the root source node is prioritized to block the degradation propagation path before downstream spatially associated early warning status nodes are processed in batches.

[0096] This invention categorizes the edge sets of the infrastructure spatial association graph structure into rapidly degraded, moderately degraded, and slowly degraded subgraphs based on their degradation rate levels. Figure 3 Each sub-layer is configured with a matching time window length and sampling resolution, ensuring that fast-propagating signals maintain temporal sharpness under high-resolution conditions within short time windows, while slow-propagating signals accumulate sufficiently under low-resolution conditions within long time windows. This avoids mutual interference between signals of different rates under a unified fixed window. The hierarchical graph attention network independently aggregates neighborhood degradation evidence on its respective rate layer sub-graph. The attention coefficients of each graph attention network are adapted to the propagation characteristics of the corresponding rate layer, thus accurately capturing degradation propagation information on its respective time scale. The cross-layer gating fusion method allows the root cause degradation signals of the slow-progressing sub-graph to be progressively transmitted to the medium- and rapid-progressing sub-graphs according to the intensity controlled by the propagation gating coefficients. This transforms the monitoring signals of facilities with early degradation into early warning evidence of neighboring facilities, overcoming the limitation of the independent modeling mode in failing to identify group degradation precursors. Deviation comparison analysis compares the comprehensive degradation trajectory prediction curve, which integrates neighborhood propagation evidence, with the independent prediction curves, enabling the early identification of nodes whose degradation is accelerated by neighborhood propagation. The aggregation of hierarchical group management areas brings related early warning nodes with common root causes into a unified management unit. The integrated intelligent maintenance scheduling scheme distinguishes the urgency weights of root cause nodes and early warning nodes of different rate levels and arranges them in batches for execution, thus realizing regional preventive intelligent maintenance.

[0097] On the other hand, this invention also proposes a digital intelligent management and maintenance system for transportation infrastructure based on the Internet of Things, comprising:

[0098] The data acquisition module is used to acquire multi-dimensional time-series monitoring data from the Internet of Things (IoT) and label the spatial correlation attributes between monitoring points. It continuously collects multi-dimensional time-series monitoring data from each monitoring point through an IoT sensor cluster, obtaining the geographic coordinates, geological attribute labels, and drainage zoning codes for each monitoring point, and labeling the spatial correlation relationships between monitoring points with propagation mode types and degradation rate level labels.

[0099] The layered graph generation module is used to calculate the spatial association weight matrix and generate a layered subgraph structure. Based on geographic proximity, geological attribute similarity, and drainage zoning affiliation, spatial association weights are calculated to generate an infrastructure spatial association graph structure. The edge sets are then divided into rapid, medium, and slow-degradation subgraphs according to their degradation rate levels. Matching time window lengths and sampling resolution parameters are set for each sub-layer.

[0100] The multi-scale feature extraction module is used to perform multi-scale sliding window segmentation on multi-dimensional time-series monitoring data and extract degradation rate vectors. It segments the multi-dimensional time-series monitoring data of each monitoring node using sliding windows of three different time window lengths, extracts degradation rate vectors at each time scale, and generates three sets of scale-specific node time-series degradation features.

[0101] The hierarchical graph attention processing module is used to input degradation features at each scale into the graph attention network of the corresponding sub-layer and output degradation state embedding vectors. On the edge set of each rate layer subgraph, the degradation propagation probability weights are dynamically calculated through the graph attention network, the degradation state information of neighboring nodes is aggregated, and the degradation state embedding vectors at each time scale are output independently.

[0102] The cross-layer fusion prediction module performs cross-layer message passing and gating fusion on the degradation state embedding vectors at three time scales and outputs a comprehensive degradation trajectory prediction curve. It controls the step-by-step transmission of signals from the slow-speed subgraph to the medium-speed and rapid-speed subgraphs through gating fusion. The fused comprehensive degradation state embedding vector is then input into the regression prediction layer, outputting the comprehensive degradation trajectory prediction curve for each node.

[0103] The early warning identification module is used to perform deviation comparison analysis and identify spatially correlated early warning status nodes and hierarchical group management areas. It compares the deviation between the comprehensive degradation trajectory prediction curve and the independent prediction curve, marks the nodes whose predicted degradation time has been moved forward by more than a preset proportion as spatially correlated early warning status, identifies cross-rate level group degradation patterns, and aggregates them into hierarchical group management areas.

[0104] The scheduling generation module is used to generate an intensive smart maintenance scheduling scheme based on the constraints of hierarchical group maintenance areas. It uses the spatial range of the hierarchical group maintenance areas, node urgency scores, root cause propagation rate levels, and the reusable proportion of shared maintenance resources as constraints to generate an intensive smart maintenance scheduling scheme that includes regional maintenance periods, repair urgency weights, shared equipment allocation paths, and batch execution sequences.

[0105] See Figures 2-9 The following application examples are proposed:

[0106] By way of example, the system proposed in this invention may also include various features and combinations thereof in the method embodiments, which will not be elaborated here.

[0107] A transportation corridor maintenance center is responsible for the daily maintenance of a 38-kilometer-long mountainous highway corridor within its jurisdiction. The corridor includes three types of infrastructure: bridge approach roads, roadbed sections, and culvert clusters, with a total of eight IoT monitoring nodes (nodes 1 to 8). The corridor has complex geological conditions, with some sections having weak foundations, and the drainage system is divided into two drainage zones. In the B quarter of 20XX, a continuous heavy rainfall event occurred in the upstream area of ​​the corridor (near node 1), prompting the maintenance center to activate its intelligent maintenance system to assess and manage the degradation status of the entire route.

[0108] Implementation of Step 1:

[0109] The IoT sensor clusters at each monitoring node continuously collect time-series data on four monitoring dimensions: structural stress, vibration frequency, crack width, and settlement displacement. Simultaneously, the system reads the geographic coordinates, geological attribute labels, and drainage zoning codes of each node. Based on the physical connectivity between facilities, each spatial relationship between node pairs is labeled with a degradation rate level. Node 1 and Node 2 are directly connected via surface runoff and are labeled as rapid; Node 1 and Node 3 are connected via underground seepage and are labeled as medium-rate; Node 3 and Node 5 are connected via geological settlement and are labeled as slow.

[0110] Table 1. Annotation of basic attributes and spatial correlation of monitoring nodes

[0111]

[0112] Implementation of step 2:

[0113] The system takes the geographical distance, geological attribute similarity, and drainage zone co-ownership of each node as input, and calculates the result according to the formula. Calculate the spatial association weights, where , , Node 1 and Node 2 are geographically closest, and both have the geological attribute label "weak foundation". ), belonging to partition A ( Therefore, it has the highest weight. Node 3 and Node 5 have different geological attributes ( ), belonging to different drainage zones ( The weights are relatively low. After the calculation is complete, the system assigns all edges to the three sub-layers according to the degradation rate level, and sets matching time window parameters for each sub-layer.

[0114] Table 2 Spatial association weights and sublayer assignments for typical nodes

[0115]

[0116] Implementation of step 3:

[0117] The system performs sliding window segmentation processing on the four monitoring dimensions of node 1, dividing them into rapid (6-hour window), medium (2-week window), and slow (3-month window) layers. It extracts three components: the slope of the mean change, the gradient of the variance increase, and the peak frequency growth rate, concatenating them into a degradation rate vector. Taking the settlement displacement dimension of node 1 as an example, within the current rapid layer window, the settlement continues to rise under the drive of heavy rainfall, with a positive and relatively large slope of the mean change. Within the slow layer window, the cumulative trend of geological settlement is steadily increasing, while the peak frequency growth rate is low. After Z-score standardization, the degradation rate vectors of each monitoring dimension have unified dimensions and can be directly input into the graph attention network.

[0118] Table 3. Degradation rate vector components of node 1 at various time scales (settlement displacement dimension, after standardization)

[0119]

[0120] Implementation of step 4:

[0121] Three sets of scale-specific degradation features are input into the graph attention networks of their respective sub-layers. The rapid layer graph attention network, operating on the rapid layer sub-graph consisting of nodes 1-2-4, dynamically calculates the degradation propagation probability weight of node 1 to node 2 using spatial association weights as adjacency constraints. Since the degradation rate vectors of node 1's current settlement displacement and crack width are significantly higher than those of node 2, the rapid layer graph attention network assigns a higher attention coefficient to node 1, incorporating neighborhood degradation evidence from node 1 into the degradation state embedding vector of node 2. The slow layer graph attention network operates independently on the slow layer sub-graph consisting of nodes 3-5-6, capturing the gradual transmission signal of geological subsidence on a multi-month scale. Each of the three graph attention networks independently outputs a degradation state embedding vector, all with a dimension of 32 to meet the vector operation requirements of subsequent cross-layer fusion.

[0122] Table 4. Output state of attention network at each rate layer of a typical node (illustration of representative components of the embedding vector)

[0123]

[0124] Implementation of step 5:

[0125] Taking node 3 as an example, this illustrates the process of cross-layer message passing and gating fusion. The embedding vector of the slowly degenerate state of node 3 is shown. Embedded vectors of degenerate states in intermediate-velocity layers After concatenation, the input is the first gating network, and the output is the propagation gating coefficient. Due to the continued intensification of geological subsidence at node 3, A higher value indicates that the root cause degradation signal of the slow layer has a strong transmission effect on the medium-velocity layer, and the medium-velocity layer enhances the embedding vector. It incorporates significant prior evidence of geological subsidence. Subsequently, Embedded vectors of rapidly degenerate states in the rapid layer Concatenate the input to the second gating network, and output... This leads to the comprehensive degenerate state embedding vector. The comprehensive degradation state embedding vector is mapped through a regression prediction layer, outputting the comprehensive degradation trajectory prediction curve for node 3 over the next 90 days.

[0126] Table 5 Key parameters of the cross-layer gating fusion process at node 3

[0127]

[0128] Implementation of step 6:

[0129] The system compares the deviation between the comprehensive degradation trajectory prediction curve and the independent prediction curve for each node. The independent prediction remaining useful life of node 2 is 180 days. The comprehensive degradation trajectory prediction curve shows that its time to reach the preset degradation threshold has shifted 41 days earlier than the independent prediction, exceeding the preset proportion (20% of the independent prediction remaining useful life, i.e., 36 days). Therefore, node 2 is marked as a spatially associated early warning state. Tracing the connectivity of the slow layer subgraph, nodes 3 and 5 can both reach the same root node (node ​​3, whose slow layer degradation rate vector mean change slope has the largest absolute value) through the edge set of the slow layer subgraph. Furthermore, nodes 3 and 5 also have associated edges on the medium-speed layer subgraph, satisfying the cross-rate level group degradation pattern recognition condition. Simultaneously, node 2 has associated edges with other early warning nodes in both the rapid and medium-speed layer subgraphs, satisfying the cross-layer coverage constraint, and is included in the hierarchical group management area. Finally, nodes 1, 2, 3, and 5 are aggregated into a hierarchical group management area, the geographical scope of which is determined by the coordinate envelope of the four nodes.

[0130] Table 6. Results of Spatial Association Early Warning Status Identification

[0131]

[0132] It should be noted that the forward movement ratios of nodes 4 and 6 did not exceed the preset ratio, and they were not marked as spatial association warning states, and were not included in the hierarchical group management area.

[0133] Implementation of step 7:

[0134] The system inputs the spatial range of the hierarchical group maintenance areas (nodes 1, 2, 3, and 5), the urgency score of each node, the root cause propagation rate level (slow level, with node 3 as the root cause), and the reusable proportion of shared maintenance resources into the regional aggregation scheduling algorithm. Nodes 1 and 2 are geographically separated by 0.82 kilometers, allowing for coverage during a single equipment relocation, and have a high reusable proportion of shared maintenance resources; therefore, they are scheduled for continuous operation in the same batch. Node 3, as the root cause node in the slow layer, has the highest urgency weight for repair and is prioritized for repair to block the degradation propagation path to node 5. Node 5 is geographically separated from node 3 and is scheduled for processing in the second batch. The final generated intensive intelligent maintenance scheduling scheme is as follows.

[0135] Table 7. Integrated Smart Management and Scheduling Scheme

[0136]

[0137] The entire implementation process demonstrates a clear data flow logic. Step 1 collects multi-dimensional time-series monitoring data and spatial correlation attribute annotations, providing three types of raw inputs—geographical distance, geological attributes, and drainage zoning—for the weight calculation in Step 2. The hierarchical subgraph structure and time window parameters generated in Step 2 determine the scale configuration of the sliding window segmentation in Step 3. The three sets of scale-specific degradation rate vectors output in Step 3 serve as node feature inputs for the three independent graph attention networks in Step 4. The three sets of degradation state embedding vectors output in Step 4 are used to integrate multi-scale degradation evidence step by step through cross-layer gating fusion in Step 5, generating a comprehensive degradation trajectory prediction curve. Step 6 uses the deviation between this prediction curve and the independent prediction curves to identify early warning nodes affected by accelerated propagation in the neighborhood and aggregates them into hierarchical group management areas. Step 7 uses the spatial, urgency, and resource reuse information of this area as constraints to finally output an intensive intelligent management and scheduling scheme arranged in batches. The data is gradually transformed from raw sensor data through feature extraction, graph aggregation, cross-layer fusion, and deviation analysis into scheduling instructions that can directly guide on-site maintenance operations. The outputs and inputs of each step strictly correspond to each other, ensuring the integrity and logical consistency of the data flow.

[0138] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. An Internet of Things-based digital intelligent management method for traffic infrastructure, characterized in that, include: Collect multidimensional time-series monitoring data from each monitoring point and label the degradation rate level of the spatial correlation between monitoring points; Calculate the spatial correlation weights between monitoring points to generate the infrastructure spatial correlation graph structure. Divide the edge set into rapid, medium and slow sub-graphs according to the degradation rate level, and set the matching time window length for each sub-layer. Perform sliding window segmentation processing according to the time window length of each sub-layer, and extract the degradation rate vector at each time scale; The degradation rate vectors at each scale are input into the graph attention network of the corresponding sub-layer to aggregate the degradation state information in the neighborhood and output the degradation state embedding vector of each layer. The gating fusion unit performs cross-layer message passing, controls the slow layer degradation signal to be passed step by step to the medium and fast layers, and outputs the comprehensive degradation trajectory prediction curve. Based on the deviation comparison between the comprehensive degradation trajectory prediction curve and the independent prediction curve, spatially related early warning nodes are identified, aggregated into hierarchical group management areas, and an intensive intelligent management and scheduling scheme is generated.

2. The method for digital and intelligent management and maintenance of transportation infrastructure based on the Internet of Things according to claim 1, characterized in that, The spatial correlation weight between the monitoring points is calculated The spatial correlation weight between the monitoring points is calculated The spatial correlation weight between the monitoring points is calculated The spatial correlation weight between the monitoring points is calculated wherein, is a monitoring point and a monitoring point , is a distance attenuation function value, is a geological property similarity, is a drainage partition belonging relationship indicator value, when two monitoring points belong to the same drainage partition otherwise , , , is a preset weight coefficient and ; The distance decay function The calculation method is as follows: in, This represents the minimum geographical distance between all point pairs in the current set of monitoring points. The maximum geographical distance between all point pairs; the geological attribute similarity. The cosine similarity algorithm was used to calculate the similarity based on the geological attribute feature vectors of each monitoring point.

3. The method for digital and intelligent management and maintenance of transportation infrastructure based on the Internet of Things according to claim 1, characterized in that, The degradation rate levels include rapid, medium, and slow levels. The rapid level corresponds to a conduction process of several hours to several days, the medium level corresponds to a conduction process of several days to several weeks, and the slow level corresponds to a conduction process of several weeks to several months. The fast-speed, medium-speed, and slow-speed subgraphs share the same set of nodes, differing only in their set of edges; Rapid subplots are configured with short time windows and high sampling resolution, medium-speed subplots with medium time windows and medium sampling resolution, and slow subplots with long time windows and low sampling resolution.

4. The method for digital and intelligent management and maintenance of transportation infrastructure based on the Internet of Things according to claim 1, characterized in that, The extraction of the degradation rate vector includes: calculating the mean change slope, variance increment gradient, and peak frequency growth rate of each monitoring indicator within each sliding window segment and concatenating them into a degradation rate vector; wherein, the mean change slope is the linear regression slope of the time series mean of each monitoring indicator within the window, the variance increment gradient is the ratio of the difference between the variances of the first half and the second half of the window to the window length, and the peak frequency growth rate is the ratio of the number of peak events exceeding the threshold in the second half of the window to the first half minus one; the three statistical components are respectively standardized using Z-score to eliminate dimensional differences.

5. The method for digital and intelligent management and maintenance of transportation infrastructure based on the Internet of Things according to claim 1, characterized in that, The three rate-layer graph attention networks are independent and do not share parameters. The input layer of each graph attention network receives the degradation rate vector of the corresponding rate layer. The attention aggregation layer calculates the attention coefficients between node pairs on the edge set of the corresponding subgraph and weights and aggregates the neighborhood features. The output layer is a fully connected layer that generates the degradation state embedding vector of each node. The three graph attention networks use the labeled degradation state of the historical degradation events of each monitored node as training labels and use the mean squared error loss function for end-to-end training.

6. The method for digital and intelligent management and maintenance of transportation infrastructure based on the Internet of Things according to claim 1, characterized in that, The cross-layer message passing via the gating fusion unit includes: inputting the slow layer degradation state embedding vector into the cross-layer gating fusion unit to generate a gating signal to be passed to the intermediate layer; performing element-wise multiplication of the gating signal with the intermediate layer degradation state embedding vector to obtain the intermediate layer update embedding; inputting the intermediate layer update embedding into the second-level gating fusion unit to generate a gating signal to be passed to the rapid layer; performing element-wise multiplication of the gating signal with the rapid layer degradation state embedding vector to obtain the rapid layer update embedding; concatenating the slow layer degradation state embedding vector, the intermediate layer update embedding, and the rapid layer update embedding and inputting them into the fully connected prediction layer to output the comprehensive degradation trajectory prediction curve; wherein, the cross-layer gating fusion unit outputs a gating coefficient vector by linearly transforming the input embedding vector and then using the sigmoid activation function, with each element taking values ​​between 0 and 1.

7. The method for digital and intelligent management and maintenance of transportation infrastructure based on the Internet of Things according to claim 1, characterized in that, The method for identifying spatially correlated early warning nodes based on deviation comparison includes: calculating the predicted degradation time deviation between the comprehensive degradation trajectory prediction curve of each monitoring node and the independent prediction curve of that node; marking the node as a spatially correlated early warning state when neighborhood propagation evidence causes the predicted degradation time of a node to advance by more than a preset proportion; retrieving node combinations in the set of nodes in the spatially correlated early warning state that simultaneously have connected paths in the slow-speed subgraph and the rapid-speed or medium-speed subgraph, and identifying the node combinations as cross-rate group degradation patterns; aggregating correlated early warning nodes involving multi-rate propagation into hierarchical group management areas, each hierarchical group management area containing a set of early warning nodes sharing at least one cross-rate propagation path.

8. The method for digital and intelligent management and maintenance of transportation infrastructure based on the Internet of Things according to claim 1, characterized in that, The generation of the intensive intelligent management and scheduling scheme includes calculating the urgency score of each node within the hierarchical group management area, and for monitoring nodes... Urgency rating The calculation method is as follows: in, For comprehensive degradation trajectory prediction curve prediction monitoring nodes The remaining time to reach the maintenance threshold. To maintain the reference time base corresponding to the threshold, For monitoring nodes The normalized coded value of the rate level. For monitoring nodes Frequency of occurrence as an intermediate node in all cross-rate propagation paths within the hierarchical group management area. This is the maximum value of that frequency. , , For preset weighting coefficients and .

9. The method for digital and intelligent management and maintenance of transportation infrastructure based on the Internet of Things according to claim 8, characterized in that, The generation of the intensive intelligent maintenance scheduling scheme also includes: evaluating the reusable ratio of shared maintenance resources among nodes in the hierarchical group maintenance area, inputting the urgency score, root cause propagation rate level, reusable ratio of shared resources, and regional spatial range as constraints into the regional aggregation scheduling algorithm, and outputting an intensive intelligent maintenance scheduling scheme that includes regional maintenance time periods, repair urgency weights, shared equipment allocation paths, and batch execution sequences; wherein, in the batch execution sequence, nodes located at the root cause position of the slow layer are repaired first to block the degradation propagation path, and then downstream affected nodes are maintained.

10. A digital intelligent management and maintenance system for transportation infrastructure based on the Internet of Things (IoT), used to execute the digital intelligent management and maintenance method for transportation infrastructure based on the IoT as described in any one of claims 1 to 9, characterized in that, include: The data acquisition module is used to collect multi-dimensional time-series monitoring data from each monitoring point through an IoT sensor group and to label the degradation rate level of the spatial correlation between monitoring points; The spatial association graph generation module is used to calculate the spatial association weights between monitoring points and generate the infrastructure spatial association graph structure. According to the degradation rate level, the edge set is divided into rapid layer subgraph, medium-speed layer subgraph and slow layer subgraph, and the matching time window length is set for each sub-layer. The multi-scale feature extraction module is used to perform sliding window segmentation processing according to the time window length of each sub-layer, and extract the degradation rate vector at each time scale. The hierarchical graph attention aggregation module is used to input the degradation rate vectors at each scale into the graph attention network of the corresponding sub-layer to aggregate the neighborhood degradation state information and output the degradation state embedding vector of each layer. The cross-layer gating fusion module is used to perform cross-layer message passing through the gating fusion unit, control the slow layer degradation signal to be passed to the medium-speed layer and the fast layer step by step, and output the comprehensive degradation trajectory prediction curve. The group early warning identification module is used to identify spatially related early warning nodes based on the deviation comparison between the comprehensive degradation trajectory prediction curve and the independent prediction curve, and aggregate them into hierarchical group management areas; The scheduling scheme generation module is used to generate an intensive and intelligent management scheduling scheme based on the constraints of hierarchical group management areas.

Images