A method and system for monitoring uneven deformation of port terminal front panel

By fusing multi-source sensor data and reconstructing the panel deformation field using a physical information graph network, and combining dynamic mode decomposition and local differential curvature energy index, the problem of insufficient spatial resolution and causes in monitoring uneven deformation of the front panel of a port terminal was solved, achieving high-precision damage location and source tracing.

CN122946892BActive Publication Date: 2026-11-17TIANJIN RES INST FOR WATER TRANSPORT ENG M O T
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611391188.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-09-09
Publication Date
2026-11-17
Estimated Expiration
2046-09-09

AI Technical Summary

Technical Problem

Existing port terminal front panel monitoring methods cannot accurately capture the spatial continuous distribution of uneven deformation. The reconstruction results are prone to violating physical laws, making it difficult to separate uneven deformation components and trace the causes of damage.

Method used

By employing multi-source sensor data fusion and physical information graph network to reconstruct the spatiotemporal deformation field of the panel, combined with dynamic mode decomposition and local differential curvature energy index, non-uniform deformation is adaptively separated and the cause of damage is traced.

Benefits of technology

It improves the accuracy and robustness of non-uniform deformation identification, accurately locates damage and traces its causes, and provides a high-precision and high-reliability intelligent monitoring solution.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122946892B_ABST
    Figure CN122946892B_ABST
Patent Text Reader

Abstract

This invention relates to the field of port terminal structure monitoring technology, specifically disclosing a method and system for monitoring non-uniform deformation of a port terminal front panel. The invention acquires distributed strain, tilt angle, and spatiotemporal information of the panel's loads from ships, machinery, and trucks using multi-source sensors; it discretizes the panel into an unstructured graph using a physical information graph network and fuses multi-source data, reconstructing the full-field deformation in real time under the physical constraints of the thin plate bending control equation; it adaptively separates residual short-wavelength modes based on time characteristic values ​​and load correlation through dynamic modal decomposition, extracting the non-uniform residual deformation field; and it calculates the local differential curvature energy index based on Gaussian curvature and mean curvature deviation, combining this with load information for spatiotemporal clustering and causal tracing. This invention achieves accurate identification, real-time early warning, and damage cause tracing of non-uniform deformation of the terminal panel, improving the reliability and robustness of monitoring.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of deformation monitoring technology, and in particular to a method and system for monitoring uneven deformation of the front panel of a port terminal. Background Technology

[0002] As a core hub in the global logistics chain, the operational efficiency and safety performance of ports directly impact regional economic development. The quay deck, a key component in high-pile quay structures that directly bears operational loads, is constantly exposed to a complex and variable mechanical environment. With the continued trend towards larger ships, berthing and unberthing impact energy has significantly increased; the large-scale deployment of automated rail-mounted container cranes has greatly increased the frequency and concentration of wheel load cycles; simultaneously, the traffic density and axle load of heavy-duty inbound trucks are also constantly increasing. Against this backdrop, quay decks are highly susceptible to uneven deformation, manifesting as localized indentations, misalignment between decks, edge warping, and even voids. Therefore, monitoring technologies that accurately sense and diagnose uneven deformation of quay decks in real time and trace the causes of damage have broad application prospects.

[0003] Existing monitoring methods have the following drawbacks: First, traditional monitoring methods mostly rely on single-point settlement markers, displacement gauges, or leveling measurements, which can only obtain the vertical displacement of sparse discrete points. They struggle to accurately capture the continuous spatial distribution characteristics of non-uniform deformation, especially lacking spatial resolution for short-wavelength deformation patterns such as local depressions and warping. Second, conventional data-driven deformation field reconstruction methods, such as those based on pure mathematical interpolation or general neural network fitting, lack internal constraints on the mechanical behavior of panel structures. When sensor deployment is sparse, data noise is significant, or some sensors fail, the reconstruction results are prone to violating physical laws, resulting in false deformation or missing key anomaly areas. Third, existing analytical methods struggle to effectively distinguish the causes of non-uniform deformation. The total deformation signal usually contains multiple components, including temperature effects, overall settlement, and normal mechanical deformation. Dangerous non-uniform components are often submerged, making it impossible to accurately identify the onset and development trend of local damage. It is also difficult to trace whether the damage is caused by ship impact, repeated mechanical wheel pressure, or foundation softening, thus affecting the accuracy of early warnings and the targeted nature of maintenance decisions.

[0004] Therefore, this invention is proposed. Summary of the Invention

[0005] To address these issues, the present invention provides a method and system for monitoring non-uniform deformation of the front panel of a port terminal, which overcomes the problems of insufficient spatial resolution, deformation field reconstruction easily violating physical laws, and difficulty in effectively separating non-uniform deformation components and tracing the causes of damage in existing monitoring methods.

[0006] To achieve the above objectives, the present invention provides a method for monitoring non-uniform deformation of a port terminal front panel, comprising: S1, acquire multi-source sensor data deployed on the panel structure. The multi-source sensors include at least orthogonally deployed distributed fiber optic strain sensors, an inclinometer array, and a sensing device for providing spatiotemporal information on ship berthing and unberthing, rail machinery and heavy truck loads. S2, the panel structure is discretized into an unstructured graph, with nodes corresponding to physical spatial locations, edges encoding local stiffness and sensing topology, and the multi-source sensor data is fused into node features, which are then input into a trained physical information graph network to reconstruct the spatiotemporal deformation field of the panel in real time; wherein, the loss function of the physical information graph network during training includes a physical residual term constructed based on the thin plate bending control equation to constrain the reconstructed deformation field to satisfy the mechanical behavior of the panel; S3, Apply dynamic mode decomposition to the spatiotemporal deformation field, extract the modes and their time characteristic values ​​and spatial frequencies, and adaptively separate the residual shortwave modes based on the correlation between the modes and the known load time sequence and the spatial frequency characteristics, and reconstruct the non-uniform residual deformation field characterizing the non-uniform deformation. S4. Calculate the local differential curvature energy index at each location based on the non-uniform residual deformation field. The local differential curvature energy characterizes the local deviation energy of Gaussian curvature and average curvature relative to the reference curvature. Identify spatiotemporal clusters where the index exceeds the threshold, and combine them with the load spatiotemporal information to trace the source of damage and output the non-uniform deformation monitoring results.

[0007] Furthermore, in the multi-source sensor data, distributed fiber optic strain sensors are orthogonally arranged bidirectionally along the bottom surface of the panel to form a quasi-continuous strain field sensing; the inclinometer array is arranged at intervals of 5 to 10 meters to obtain the biaxial tilt angle of the panel surface; the sensing device includes a pose encoder installed on a track-mounted machine and / or a video analysis device for identifying the position and load of a heavy-duty truck to provide the spatiotemporal information of the load.

[0008] Furthermore, the physical information graph network is a message-passing graph neural network. The loss function used for training is composed of a data loss term, a physical residual term, and a smoothing regularization term, each multiplied by its corresponding weighting coefficient and then summed. The physical residual term is constructed as follows: sampling nodes are randomly selected in the unstructured graph. For each sampling node, the equation residual of the vertical displacement is calculated according to the thin plate bending control equation of the panel, and the sum of the squares of the equation residuals of all sampling nodes is taken as the physical residual term. The thin plate bending control equation includes terms related to the bending stiffness of the panel and the foundation coefficient, and the distributed load in the equation is determined by the load spatiotemporal information obtained by the sensing device. The data loss term is constructed as a loss term constraining the consistency between the reconstructed deformation field and the measured displacement, based on the displacement obtained by integrating the tilt angle measured by the inclinometer array. The smoothing regularization term is used to penalize the variation amplitude of the deformation field between adjacent nodes in the unstructured graph to suppress the discontinuity of the reconstructed deformation field.

[0009] Furthermore, the dynamic mode decomposition employs high-order dynamic mode decomposition and introduces a sliding window for online updating; the adaptive separation of residual shortwave modes based on the correlation between the modes and the known load time series and spatial frequency characteristics specifically includes: Calculate the temporal characteristic values ​​of each mode, and classify the modes with temporal characteristic value magnitude close to 1 and spatial frequency lower than the preset spatial threshold as quasi-static modes to characterize the overall settlement and temperature deformation. The correlation coefficient between the time coefficient of each mode and the load time sequence of ships and rail machinery provided by the sensing device is calculated. Modes with correlation coefficients higher than the preset correlation threshold are classified as load tracking modes to characterize interpretable working deformation. Modes with spatial frequencies higher than the preset spatial threshold, excluding quasi-static modes and load tracking modes, are classified as residual shortwave modes. The non-uniform residual deformation field is obtained by reconstructing using all the aforementioned residual shortwave modes.

[0010] Furthermore, the calculation method for the local differential curvature energy index is as follows: For each element in the discrete element of the panel, take the local region consisting of itself and its neighborhood; The sum of the square of the difference between the Gaussian curvature derived from the non-uniform residual deformation field and the reference Gaussian curvature, and the sum of the square of the difference between the average curvature and the reference average curvature within the local region, and the sum of the area integral of the local region and the area of ​​the local region, is the local differential curvature energy of the element. The reference Gaussian curvature and the reference average curvature are determined by the deformation field obtained by the panel in a healthy state.

[0011] Furthermore, the damage tracing based on the load spatiotemporal information includes: Spatial matching is performed between the spatiotemporal clusters of local differential curvature energy exceeding the threshold and the ship berthing position, tracked machinery trajectory, or heavy-duty truck path provided by the sensing device. Calculate the mutual information between the occurrence time of the spatiotemporal cluster and the occurrence time of the load event; When mutual information meets the preset significant conditions and spatial matching error is within the allowable range, the non-uniform deformation damage is traced back to the corresponding ship collision, rail-mounted machinery operation, or heavy-duty truck passage.

[0012] Furthermore, the real-time reconstruction panel spatiotemporal deformation field and the calculation of local differential curvature energy index are continuously executed online as the first level of processing; when the local differential curvature energy index exceeds the preset warning threshold, the steps of dynamic mode decomposition, non-uniform residual deformation field reconstruction and damage tracing are triggered as the second level of processing, and after the second level of processing is completed, the parameters of the physical information graph network are corrected and updated using the finite element model and historical accumulated data.

[0013] Furthermore, the edge-encoded local stiffness of the unstructured graph includes a function of the physical distance between nodes and the estimated elastic modulus; the residual shortwave mode is a modal component in the dynamic modal decomposition result whose time characteristic value is not quasi-static and is not significantly correlated with the known load time sequence, and whose spatial frequency is higher than a set threshold, used to characterize non-uniform deformations such as local depressions, warping, or voids.

[0014] The present invention also provides a system for monitoring uneven deformation of port terminal front panel, the system being used to implement the method for monitoring uneven deformation of port terminal front panel as described in any one of the claims, the system comprising: A multi-source data acquisition module is used to acquire multi-source sensor data deployed on the panel structure. The multi-source sensors include at least orthogonally arranged distributed fiber optic strain sensors, an inclinometer array, and a sensing device for providing spatiotemporal information on ship berthing and unberthing, rail machinery and heavy-duty truck loads. The deformation field reconstruction module is used to discretize the panel structure into an unstructured graph, with nodes corresponding to physical spatial locations, edges encoding local stiffness and sensing topology, and fusing the multi-source sensor data into node features. These features are then input into a trained physical information graph network to reconstruct the spatiotemporal deformation field of the panel in real time. The loss function of the physical information graph network during training includes a physical residual term constructed based on the thin plate bending control equation to constrain the reconstructed deformation field to meet the mechanical behavior of the panel. The non-uniform deformation separation module is used to apply dynamic mode decomposition to the spatiotemporal deformation field, extract the modes and their time characteristic values ​​and spatial frequencies, and adaptively separate the residual shortwave modes based on the correlation between the modes and the known load time sequence and the spatial frequency characteristics, and reconstruct the non-uniform residual deformation field characterizing the non-uniform deformation. The damage assessment and source tracing module is used to calculate the local differential curvature energy index at each location based on the non-uniform residual deformation field. The local differential curvature energy characterizes the local deviation energy of Gaussian curvature and mean curvature relative to the reference curvature. By identifying spatiotemporal clusters where the index exceeds the threshold, and combining the load spatiotemporal information, damage source tracing is performed, and non-uniform deformation monitoring results are output.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: Firstly, this invention introduces a physical information graph network into the reconstruction of the deformation field of a wharf panel, and embeds a physical residual term based on the thin plate bending control equation into the loss function of the network training. This solves the problem of traditional pure data-driven models producing spurious deformations and reconstructing results that violate mechanical laws when sensors are sparse or malfunctioning. Simultaneously, the panel is discretized into an unstructured graph, and local stiffness and sensing topology are encoded by edges, enabling the network to flexibly adapt to complex sensing layouts and structural geometric inhomogeneities. Furthermore, by combining dynamic mode decomposition, based on the correlation between modes and known load time sequences and spatial frequency characteristics, residual shortwave modes are adaptively separated to reconstruct the non-uniform residual deformation field. This solves the problem of frequency band aliasing and difficulty in distinguishing normal mechanical operation deformation and damage deformation in the total deformation signal. This improves the accuracy and robustness of non-uniform deformation identification under complex multi-factor coupling. Secondly, this invention defines a local differential curvature energy index based on the joint deviation of Gaussian curvature and average curvature. This quantifies the geometrical anomalies of non-uniform deformation, such as local depressions and warping, into calculable damage-sensitive parameters, solving the problem that traditional single-point settlement or tilt indices are overly sensitive to uniform deformation and cannot distinguish between overall deformation and local damage. Local differential curvature energy units exceeding a threshold are clustered in spatial and temporal dimensions to form spatiotemporal clusters. Mutual information calculation and significance testing are introduced to determine the causal correlation between the spatiotemporal clusters and the load events recorded by the sensing device, improving the accuracy of damage localization and the reliability of the source tracing results. Thirdly, this invention forms a complete technical chain from data acquisition, deformation reconstruction, non-uniformity separation to damage assessment and tracing. It achieves complementary sensing of strain, rotation, and load fields through multi-source sensor fusion deployment; realizes physically self-consistent real-time reconstruction of full-field deformation through a physical information graph network; achieves adaptive separation of non-uniform deformation through dynamic modal decomposition; and achieves precise damage location and tracing through local differential curvature energy indices. The entire method organically integrates physical model constraints with data-driven analysis, overcoming the technical bottlenecks of traditional monitoring methods such as insufficient spatial resolution, reconstruction results easily violating physical laws, difficulty in separating non-uniform deformation components, and inability to trace the cause of damage. It provides a high-precision, high-reliability, and traceable intelligent monitoring solution for the safe operation and maintenance of port terminal panels under multi-source loads such as ship impacts, track-mounted machinery crushing, and heavy truck crushing. Attached Figure Description

[0016] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 A flowchart illustrating a method for monitoring uneven deformation of a port terminal front panel, provided in an embodiment of the present invention. Figure 2 A flowchart illustrating the method for separating residual shortwave modes in a port terminal front panel non-uniform deformation monitoring method provided in this embodiment of the invention; Figure 3 This is a flowchart of a method for damage tracing by combining load spatiotemporal information in a method for monitoring uneven deformation of a port terminal front panel, as provided in an embodiment of the present invention. Figure 4 This is a structural block diagram of a port terminal front panel non-uniform deformation monitoring system provided in an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0019] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0020] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0021] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0022] Example 1 like Figures 1 to 3 As shown, the present invention provides a method for monitoring non-uniform deformation of a port terminal front panel, comprising: S1, acquire multi-source sensor data deployed on the panel structure. The multi-source sensors include at least orthogonally deployed distributed fiber optic strain sensors, an inclinometer array, and a sensing device for providing spatiotemporal information on ship berthing and unberthing, rail machinery and heavy truck loads. In the multi-source sensor data, distributed fiber optic strain sensors are orthogonally arranged bidirectionally along the bottom surface of the panel to form a quasi-continuous strain field sensing; the inclinometer array is arranged at intervals of 5 to 10 meters to obtain the biaxial tilt angle of the panel surface; the sensing device includes a pose encoder mounted on a track-mounted machine and / or a video analysis device for identifying the position and load of a heavy-duty truck to provide the spatiotemporal information of the load.

[0023] In one possible implementation, distributed fiber optic strain sensing cables are laid longitudinally and laterally on the underside of the front panel of the wharf, forming a bidirectional orthogonal strain sensing network. The sensing cables employ a tight-sheath structure, bonded to the concrete surface of the panel via epoxy resin bonding or embedded installation to accurately transmit the panel's bending strain. The longitudinal and transverse cables are spaced 0.8 meters apart, forming a quasi-continuous strain field sensing array covering the entire panel area. One end of the cable is connected to a distributed fiber optic demodulator based on Brillouin optical time-domain analysis technology. The demodulator continuously acquires the strain time history of each spatial point with a spatial resolution of 0.5 meters and a sampling frequency of 10 Hz, outputting real-time bidirectional strain distribution data of the panel's underside.

[0024] A dual-axis microelectromechanical system (MEMS) inclinometer array is deployed on the upper surface of the panel at a rectangular grid spacing of 7.5 meters. Each inclinometer is fixed to the panel surface with expansion bolts and measures the inclination of the panel along two orthogonal horizontal axes at its installation position in real time, with an angular resolution better than 0.001° and a sampling frequency of 10 Hz. The inclinometers aggregate the inclination data to an edge data acquisition unit via an RS-485 bus or wireless transmission module.

[0025] Distributed optical fibers are orthogonally deployed bidirectionally along the bottom surface of the panel, enabling the acquisition of quasi-continuous longitudinal and transverse strain distributions with a spatial resolution of 0.5 meters. They exhibit extremely high sensitivity to short-wavelength deformations such as localized bending and indentations. While the inclinometer array is spatially arranged at intervals of 5 to 10 meters, resulting in a much lower density of measurement points compared to optical fiber strain gauges, it directly measures the biaxial tilt angle of the panel surface. After selecting a reference node and integrating along the graph structure path, it can provide an absolute vertical displacement reference for discrete points on the panel. Deformation sensors alone can only observe the panel's response to unknown external forces, without determining the source of the deformation. This embodiment uses a pose encoder and video analysis device to acquire real-time data on ship berthing positions, wheel pressure trajectories and magnitudes of rail-mounted machinery, and the travel paths and axle loads of heavy-duty trucks, clearly quantifying the spatiotemporal distribution of the load acting on the panel.

[0026] To acquire spatiotemporal information on ship berthing and unberthing, rail-mounted machinery, and heavy-duty truck loads, this embodiment employs the following combination of sensing devices. For rail-mounted container cranes running along the dock's front rails, an absolute multi-turn encoder is installed on the driven wheel axle of the trolley traveling mechanism to record the crane's trolley position in real time. Simultaneously, another absolute encoder is installed at the end of the hoisting mechanism's drum shaft, combined with a load sensor, to acquire the lifting height and load weight of the spreader. Using the trolley position, the trolley position, and the known spatial coordinates of the track, the specific position of each crane wheel acting on the panel and the wheel pressure load value are calculated in real time. For heavy-duty trucks traveling on the panel, high-definition network cameras are installed at key locations in the dock entrance passage and panel area, with the camera's field of view covering the truck's travel path. A deep learning-based visual analysis device is deployed to perform real-time target detection and multi-target tracking on the video stream, automatically identifying heavy-duty trucks and recording their vehicle type, travel trajectory, and the truck's positioning coordinates on the panel in each frame. Simultaneously, this visual analysis device interfaces with the dock's weighbridge system to obtain the corresponding truck's total weight and axle load information. By aligning and mapping the spatiotemporal coordinates of the vehicle trajectory with the load information in time, the spatiotemporal distribution of the truck load acting on the panel is obtained.

[0027] All sensors and sensing devices are synchronized through the edge data collector, and GPS timestamps are uniformly applied using the network time protocol. After noise reduction and interpolation preprocessing, a time-aligned multi-source sensor dataset is formed.

[0028] This invention solves the problem that traditional single-point displacement or strain measurements cannot capture the continuous spatial distribution of non-uniform deformation by forming a quasi-continuous strain field perception through bidirectional orthogonal deployment of distributed fiber optic strain sensors along the bottom surface of the panel. This improves the spatial sensing density of the deformation field and the ability to distinguish local anomalies. By deploying a dual-axis inclinometer array at 5-10 meter intervals on the panel surface, the inherent limitation of distributed strain measurements being insensitive to rigid body displacement is overcome, providing independent angular constraints for subsequent deformation field reconstruction and improving the accuracy and robustness of full-field displacement inversion. By using a video analysis device to identify the position and load of heavy-duty trucks and correlate it with weighbridge data to obtain axle load information, the precise quantification of the spatiotemporal distribution of moving loads is achieved, improving the reliability of the input load distribution in the physical information graph network, thereby indirectly improving the accuracy of deformation field reconstruction.

[0029] S2, the panel structure is discretized into an unstructured graph, with nodes corresponding to physical spatial locations, edges encoding local stiffness and sensing topology, and the multi-source sensor data is fused into node features, which are then input into a trained physical information graph network to reconstruct the spatiotemporal deformation field of the panel in real time; wherein, the loss function of the physical information graph network during training includes a physical residual term constructed based on the thin plate bending control equation to constrain the reconstructed deformation field to satisfy the mechanical behavior of the panel; The physical information graph network is a message-passing graph neural network. The loss function used for training is composed of a data loss term, a physical residual term, and a smoothing regularization term, each multiplied by its corresponding weighting coefficient and then summed. The physical residual term is constructed as follows: sampling nodes are randomly selected in the unstructured graph. For each sampling node, the equation residual of the vertical displacement is calculated according to the thin plate bending control equation of the panel, and the sum of the squares of the equation residuals of all sampling nodes is taken as the physical residual term. The thin plate bending control equation includes terms related to the bending stiffness of the panel and the foundation coefficient, and the distributed load in the equation is determined by the load spatiotemporal information obtained by the sensing device. The data loss term is constructed based on the displacement obtained by integrating the tilt angle measured by the inclinometer array, and is used as a loss term to constrain the consistency between the reconstructed deformation field and the measured displacement. The smoothing regularization term is used to penalize the variation amplitude of the deformation field between adjacent nodes in the unstructured graph to suppress the discontinuity of the reconstructed deformation field.

[0030] In one possible implementation, the area of ​​the dock panel to be monitored is abstracted as a graph structure consisting of nodes and edges. Nodes are generated as follows: based on the panel's planar geometry, nodes are arranged in a uniform grid of 0.8 meters × 0.8 meters. Each node corresponds to a physical spatial location on the panel, and its planar coordinates are recorded. Simultaneously, a node is added at each inclinometer installation location to ensure that the inclinometer's measuring point coincides with a node in the graph. If complete coincidence is not possible, a correspondence is established by selecting the nearest grid node.

[0031] The edges are constructed as follows: For any two nodes, the straight-line distance between them is calculated. If this distance is less than 1.2 meters and there is no structural expansion joint between the two nodes, a connecting edge is established between them. Each edge is accompanied by an edge feature vector containing the following information: the actual distance between the two nodes, the design thickness of the panel in that area, the elastic modulus of the concrete determined according to the material test report, for example, 32 GPa, and an identifier indicating whether the edge is located on the distributed fiber optic sensing path (1 for yes, 0 for no). This identifier enables the network to know which areas in the graph structure have richer measured strain information.

[0032] At each sampling time, a multi-dimensional feature vector is constructed for each node in the graph, which is then used as input to the physical information graph network. This feature vector specifically contains the following types of information: The first category involves strain features extracted from distributed fiber optic strain data. Based on the current node coordinates, the longitudinal and transverse strain values ​​at that node are obtained through bilinear interpolation from the bidirectional strain field output by the fiber optic demodulator. The second type is the tilt feature obtained from the inclinometer array. If an inclinometer is located at the current node, the tilt value measured by the inclinometer along the two orthogonal horizontal axes is directly used as the node feature; if no inclinometer is located at the current node, the tilt value is filled with zero, and an additional indicator flag is added to distinguish whether the node has real tilt measurement data. The third category is load characteristics. Based on the spatiotemporal information of the ship, crane, or truck loads provided by the pose encoder or video analysis device, the equivalent distributed load value acting on the vicinity of the node at the current moment is estimated. If there is no load acting on the area, the load value is set to zero. The fourth category is physical parameter characteristics. This includes the flexural stiffness of the panel and the subgrade coefficient at the node. The flexural stiffness is calculated and determined based on the elastic modulus of the concrete in this area, the panel thickness, and Poisson's ratio; the subgrade coefficient is taken from the geological survey report, for example, 25 MN / m³. The fifth category is time characteristics. Sampling times throughout the day are encoded into two values ​​using sine and cosine functions respectively, reflecting the impact of daily temperature variations on deformation. All of the above features are standardized using the mean and standard deviation of the training dataset before being input into the network to eliminate dimensional differences.

[0033] The Physical Information Graph Network (PIG) ​​employs a three-layer message-passing graph neural network. Each layer sequentially performs three operations: edge feature update, neighbor node message aggregation, and node feature update. The first layer operation, edge feature update, concatenates the current feature vectors of the two nodes connected to the edge and the edge's own feature vector. This concatenation is input into a feedforward network with two fully connected layers (each containing 128 neurons) using the ReLU activation function, and the output is an updated edge representation vector. The second layer operation, message aggregation, collects the updated representation vectors of all edges pointing to that node and calculates their element-wise average as the aggregated message vector. The third layer operation, node feature update, concatenates the node's current feature vector with the aggregated message vector obtained in the previous step, inputs it into a gated recurrent unit for updating, and outputs the node's new feature vector.

[0034] After the three-layer message passing process is completed, the final feature vector of each node is fed into a linear output layer, which outputs the predicted vertical displacement of that node at the current sampling time. The predicted displacement values ​​of all nodes together constitute the spatiotemporal deformation field of the panel at the current time. The entire network contains approximately 250,000 trainable parameters.

[0035] The training of the Physical Information Graph Network employs a composite loss function, which is the sum of three terms: a data loss term, a physical residual term, and a smoothing and regularization term, each multiplied by its corresponding weighting coefficient. The weighting coefficients are determined using a grid search method; for example, the data loss term has a weight of 1.0, the physical residual term has a weight of 0.1, and the smoothing and regularization term has a weight of 0.01.

[0036] The data loss term is constructed as follows: For nodes equipped with inclinometers, a reference node with known or assumed zero displacement is first selected. Then, along the shortest path connecting the two nodes in the graph structure, the inclinometer measurements are numerically integrated using the trapezoidal rule to calculate the cumulative displacement from the reference node to the inclinometer node, which is used as the reference displacement value at that node. The data loss term is calculated as the average of the sum of squares of the differences between the network-predicted displacement and the reference displacement at all inclinometer nodes.

[0037] The physical residual term is crucial for ensuring that the network output conforms to the mechanical laws of the panel. Its construction method is as follows: In each training batch, a certain proportion, for example, 20% of the nodes are randomly selected as collocation points from all nodes. For each collocation point, the residual of the thin plate bending control equation needs to be calculated. This control equation includes the product of the panel bending stiffness and the fourth spatial derivative of the vertical displacement, the product of the foundation coefficient and the vertical displacement, and the distributed load term; theoretically, the sum of these three should be equal to zero. In this embodiment, the calculation of the fourth spatial derivative is achieved using a local polynomial fitting method on the graph structure: Taking the collocation point as the center, adjacent nodes within its two-order neighborhood are selected. Using the predicted vertical displacement values ​​and node coordinates of these nodes, a quadratic polynomial surface is fitted. Then, the fourth derivative of this polynomial is analytically obtained as an approximation of the fourth derivative at the collocation point. The distributed load value in the equation is determined by spatial interpolation of the load spatiotemporal information obtained by the sensing device in S1. The physical residual term is the average of the sum of squares of the control equation residuals at all collocation points.

[0038] The smoothing regularization term is constructed by calculating the average of the sum of squares of the differences between the predicted displacements of the two ends of all edges in the graph. This is used to suppress drastic jumps in the deformation field between adjacent nodes and to ensure the continuity of the reconstructed deformation field.

[0039] The training dataset consists of two parts: one part is sensor data collected by the real monitoring system during its historical operation, covering various load conditions and temperature conditions; the other part is data generated based on a refined 3D finite element model of the panel. This model applies load patterns and boundary conditions that conform to the statistical laws of actual measurements, calculates high-density full-field displacements as training labels, and outputs strain, tilt angle, and load distribution corresponding to the sensor deployment as network input features. Training uses the Adam optimizer with an initial learning rate set to 0.001, a batch size of 32 training samples, and iterative training on a graphics processor for approximately 200 epochs until the loss function converges.

[0040] After network training is completed, its model parameters are solidified and compressed, and deployed on edge computing devices. During the online operation phase, forward inference is performed every 0.1 seconds on the latest acquired and preprocessed multi-source sensor data, directly outputting the predicted vertical displacement values ​​of all nodes in the graph, forming the panel deformation field at the current moment. This deformation field is stored in the form of node coordinate-vertical displacement data pairs. On ordinary edge computing hardware, the inference time per frame is approximately 15 milliseconds, meeting the timeliness requirements of near real-time monitoring.

[0041] This invention addresses the problem of spurious deformations caused by sensor sparsity or failure in pure data-driven models when the reconstruction results violate mechanical laws, improving the physical consistency of the deformation field reconstruction and its robustness under missing data conditions by introducing a physical residual term based on the thin plate bending control equation into the loss function of the physical information graph network training. By using the tilt integral displacement of the inclinometer array as a data loss term, it compensates for the inability of distributed strain measurement to provide absolute displacement constraints, providing a displacement monitoring signal independent of the strain field for graph network training and improving the absolute accuracy of vertical displacement inversion. By setting a smoothing regularization term in the loss function to penalize the variation amplitude of the deformation field between adjacent nodes, it suppresses sawtooth fluctuations in the deformation field caused by measurement noise or missing local data, improving the spatial continuity and smoothness of the reconstructed deformation field and making it closer to the deformation morphology of the real panel. By using the spatiotemporal load information acquired by the sensing device as a distributed load input into the thin plate bending control equation, it ensures that the external force source in the physical residual term corresponds to the actual operating load in real time, solving the problem that load boundary conditions cannot be accurately given in traditional mechanical constraints and improving the timeliness and accuracy of physical constraints.

[0042] S3, Apply dynamic mode decomposition to the spatiotemporal deformation field, extract the modes and their time characteristic values ​​and spatial frequencies, and adaptively separate the residual shortwave modes based on the correlation between the modes and the known load time sequence and the spatial frequency characteristics, and reconstruct the non-uniform residual deformation field characterizing the non-uniform deformation. The dynamic mode decomposition employs high-order dynamic mode decomposition and introduces a sliding window for online updating; based on the correlation between the mode and the known load time series and the spatial frequency characteristics, the residual shortwave modes are adaptively separated, specifically including: Calculate the temporal characteristic values ​​of each mode, and classify the modes with temporal characteristic value magnitude close to 1 and spatial frequency lower than the preset spatial threshold as quasi-static modes to characterize the overall settlement and temperature deformation. The correlation coefficient between the time coefficient of each mode and the load time sequence of ships and rail machinery provided by the sensing device is calculated. Modes with correlation coefficients higher than the preset correlation threshold are classified as load tracking modes to characterize interpretable working deformation. Modes with spatial frequencies higher than the preset spatial threshold, excluding quasi-static modes and load tracking modes, are classified as residual shortwave modes. The non-uniform residual deformation field is obtained by reconstructing using all the aforementioned residual shortwave modes.

[0043] The edge-encoded local stiffness of the unstructured graph includes a function of the physical distance between nodes and the estimated elastic modulus; the residual short-wave mode is a modal component in the dynamic mode decomposition result whose time characteristic value is not quasi-static and is not significantly correlated with the known load time sequence, and whose spatial frequency is higher than a set threshold, and is used to characterize non-uniform deformations such as local depressions, warping or voids.

[0044] In one possible implementation, a sliding window mechanism is used to continuously process the latest data. The window length is set to 600 consecutive time steps (60 seconds), and the window sliding step size is set to 300 time steps (30 seconds). There is 50% overlap between adjacent windows. For the current window, the deformation fields of all time steps within the window are arranged into a snapshot matrix. Each column of the matrix corresponds to a time step and contains the displacement values ​​of all N nodes at that time step; therefore, the matrix size is N rows × 600 columns. Before constructing the matrix, the time series of each node is averaged to eliminate static bias.

[0045] Perform high-order dynamic mode decomposition on the snapshot matrix of the current window. The core parameter settings are as follows: select two embedding dimension parameters d and m, where d is set to 50, indicating that an enhanced snapshot is constructed using data from the current time step and the previous 49 time steps; m is set to 5, indicating that the first 5 dominant singular vectors are truncated when solving the system matrix to suppress noise. The specific process is as follows: Using `d` as the delay embedding dimension, the original snapshot matrix is ​​expanded into a Hankel block matrix, such that each column contains the node displacements of the current time step and the previous `d-1` time steps. This operation can capture high-order dynamic features in the time series. Singular value decomposition is performed on the enhanced data matrix, retaining the first `m` singular vectors and singular values ​​to obtain a low-dimensional approximate matrix. A linear operator matrix is ​​fitted in the low-dimensional space using the least squares method. This operator describes the evolution of the system state between adjacent time steps. Eigenvalue decomposition is performed on the fitted linear operator matrix to obtain a series of eigenvalues ​​λ. j (j=1,...,m) and their corresponding eigenvectors. Each eigenvalue λ j The temporal evolution characteristics of the corresponding modes are represented, and the modulus length determines the decay or growth trend of the mode over time. Each feature vector is mapped back to the original node space through inverse transformation to obtain the spatial distribution of each mode. The temporal coefficient of each mode is determined by the projection of the mode vector onto the original data, reflecting the activation intensity of the mode at each time step.

[0046] To determine the degree of spatial variation of each mode, a spatial frequency index is calculated for the spatial distribution vector of each mode. Specifically, the values ​​of the mode at each node are bilinearly interpolated according to the actual planar coordinates of the nodes, mapping them onto a uniform two-dimensional grid. Then, a two-dimensional Fourier transform is performed on this two-dimensional spatial distribution to obtain the spectrum. The spatial frequency index is taken as the cutoff frequency corresponding to 95% of the total energy accumulated from low to high frequencies in the spectrum, expressed in cycles per meter. A lower value indicates a smoother spatial variation of the mode, while a higher value indicates more drastic spatial fluctuations.

[0047] The sensing devices provide load timing information for ship berthing and unberthing, and rail-mounted machinery. Taking a rail-mounted container crane as an example, the position of the trolley and the lifting load can be obtained from the posture encoder and load sensor, and the load intensity time history curve of the crane wheel pressure acting on the panel can be obtained by combining them. The load time history is time-aligned with the time coefficient sequence of each mode within the window, and the Pearson correlation coefficient between the two is calculated, which reflects the strength of the linear correlation between the activation degree of the mode and the rhythm of load application.

[0048] Based on the temporal characteristic modulus, spatial frequency, and load correlation of each mode, the modes are divided into the following three categories: The first category is quasi-static modes. The criteria for judgment are: the temporal characteristic value modulus is between 0.995 and 1.005, and the spatial frequency is lower than a preset spatial threshold, which is 0.1 per meter in this embodiment. These modes hardly decay in time and change gently in space, and are classified as quasi-static modes. They are used to characterize the overall settlement of the panel and the overall bending deformation caused by diurnal and seasonal temperature changes. The second category is the load tracking mode. The judgment criterion is that the absolute value of the correlation coefficient between the time coefficient of the mode and the crane load time sequence provided by the sensing device is higher than a preset correlation threshold, which is 0.6 in this embodiment. The activation rhythm of this type of mode is highly synchronized with the known mechanical operation load, and is classified as a load tracking mode. It is used to characterize the interpretable working deformation caused by crane wheel pressure cycle or truck passage. The third category is residual shortwave modes. Modes that are neither quasi-static modes nor load-tracking modes, and whose spatial frequency is higher than a preset spatial threshold, are classified as residual shortwave modes in this embodiment with a frequency of 0.1 per meter. These modes do not significantly correlate with normal periodic loads in terms of temporal evolution, but exhibit high-frequency local fluctuation characteristics in space, and are used to characterize non-uniform deformations such as local depressions, panel edge warping, and separation of the panel from the foundation.

[0049] After modal classification, all modal components identified as residual shortwave modes are linearly superimposed to reconstruct the deformation field contributed solely by these modes. Specifically, for each time step within the window, the spatial vector of each residual shortwave mode is multiplied by its corresponding time coefficient, and then summed to obtain the non-uniform residual deformation field at that time. This deformation field has stripped away deformation components explainable by overall settlement, temperature effects, and normal mechanical operations, highlighting the dangerous non-uniform deformation distribution.

[0050] After reconstruction, the non-uniform residual deformation field at the middle of the window is used as the current output and stored in the result sequence. At the same time, the sliding window is advanced 300 times, and the above decomposition and separation process is repeated for the new window to achieve continuous online updates.

[0051] In this embodiment, the number of panel nodes N is 1500, the time window length is 600 time steps, the sliding step size is 300 time steps, the embedding dimension d is 50, the truncated rank m is 5, the preset spatial threshold is 0.1 per meter, and the correlation coefficient threshold is 0.6. These parameters were determined through numerical experiments and optimization using actual wharf monitoring data, achieving a balance between separation performance and computational overhead. A single window processing cycle, including HODMD decomposition, modal classification, and reconstruction, takes approximately 2.3 seconds on a conventional cloud computing instance, far less than the 30-second window update cycle, thus meeting the requirements for continuous online analysis.

[0052] This invention automatically filters out quasi-static modes based on whether the modal time characteristic value modulus is close to 1 and whether the spatial frequency is below a preset threshold. This solves the problem of slow bending deformation caused by overall settlement and temperature being mixed in the total deformation and interfering with damage identification. It reduces the false alarm rate of misjudging normal environmental effects as structural damage and improves the accuracy and reliability of non-uniform deformation judgment. By calculating the correlation coefficient between the time coefficient of each mode and the known load time sequence of cranes, ships, etc. provided by the sensing device, the load tracking mode is automatically identified and separated. This solves the problem of overlapping and indistinguishable frequency bands of periodic recoverable working deformation and damage deformation generated by normal mechanical operation. It improves the purity of the damage component in the non-uniform residual deformation field, so that it only reflects irreversible local abnormal deformation. By setting a spatial frequency threshold, high-frequency modes with drastic spatial changes are classified into residual short-wavelength modes. This can accurately capture the spatial characteristics of small-scale damage such as local depressions, panel edge warping, and foundation voids. This solves the problem of insufficient spatial resolution of short-wavelength non-uniform deformation by traditional deformation monitoring methods and improves the detection sensitivity of early local micro-damage.

[0053] S4. Calculate the local differential curvature energy index at each location based on the non-uniform residual deformation field. The local differential curvature energy characterizes the local deviation energy of Gaussian curvature and average curvature relative to the reference curvature. Identify spatiotemporal clusters where the index exceeds the threshold, and combine them with the load spatiotemporal information to trace the source of damage and output the non-uniform deformation monitoring results.

[0054] The calculation method for the local differential curvature energy index is as follows: For each element in the discrete element of the panel, take the local region consisting of itself and its neighborhood; The sum of the square of the difference between the Gaussian curvature derived from the non-uniform residual deformation field and the reference Gaussian curvature, and the sum of the square of the difference between the average curvature and the reference average curvature within the local region, and the sum of the area integral of the local region and the area of ​​the local region, is the local differential curvature energy of the element. The reference Gaussian curvature and the reference average curvature are determined by the deformation field obtained by the panel in a healthy state.

[0055] Damage source tracing based on the aforementioned load spatiotemporal information includes: Spatial matching is performed between the spatiotemporal clusters of local differential curvature energy exceeding the threshold and the ship berthing position, tracked machinery trajectory, or heavy-duty truck path provided by the sensing device. Calculate the mutual information between the occurrence time of the spatiotemporal cluster and the occurrence time of the load event; When mutual information meets the preset significant conditions and spatial matching error is within the allowable range, the non-uniform deformation damage is traced back to the corresponding ship collision, rail-mounted machinery operation, or heavy-duty truck passage.

[0056] The real-time reconstruction of the spatiotemporal deformation field of the panel and the calculation of the local differential curvature energy index are continuously executed online as the first level of processing. When the local differential curvature energy index exceeds the preset warning threshold, the steps of dynamic mode decomposition, non-uniform residual deformation field reconstruction and damage tracing are triggered as the second level of processing. After the second level of processing is completed, the parameters of the physical information graph network are corrected and updated using the finite element model and historical accumulated data.

[0057] In one possible implementation, the index of the latest output non-uniform residual deformation field is calculated every 0.1 seconds. It should be noted that in the first-level processing, the non-uniform residual deformation field is obtained by simplifying the total deformation field directly output by S2 through high-pass spatial filtering to ensure real-time performance; when the second-level processing is triggered, the accurate non-uniform residual deformation field obtained by S3 through complete dynamic mode decomposition is used for updating.

[0058] The panel region is discretized into a grid of cells for index calculation. The grid size is 0.4 m × 0.4 m, and the center coordinates of each cell correspond to the planar positions of the nodes in step S2. For cells near the panel edges or structural seams, they are trimmed according to the actual boundary shape. For each discrete cell, a 3 × 3 local region is formed by itself and its surrounding eight neighboring cells, containing a total of 9 cells. If a cell is located at the panel boundary and its neighborhood is incomplete, the local region is formed by the actual existing neighboring cells.

[0059] Within a local region, the vertical displacement values ​​of each node output by S2 or S3 are used to perform least-squares fitting on the node coordinates and displacements within the local region, resulting in a quadratic surface function. This quadratic surface is specifically a quadratic polynomial of displacement with respect to two horizontal coordinates, containing six coefficients: a constant term, a linear term, and a second-order term. The Gaussian curvature and mean curvature at the center point of the local region are analytically calculated using this quadratic surface function. The Gaussian curvature is the product of the two principal curvatures at that point, reflecting whether the local surface is bowl-shaped or saddle-shaped; the mean curvature is the arithmetic mean of the two principal curvatures, reflecting whether the local surface is convex upwards or concave downwards. The values ​​of the two principal curvatures are calculated using the second derivative coefficients of the quadratic surface according to the standard formulas of differential geometry.

[0060] The reference Gaussian curvature and reference average curvature are obtained from the panel's health status during the initial deployment of the monitoring system. Specifically, before the wharf is put into operation or for a period after comprehensive inspection confirming the structure is intact, deformation field data is continuously collected for 24 hours. The Gaussian curvature and average curvature of each unit at each sampling time are calculated. Then, the time average is taken for each unit to obtain the reference Gaussian curvature and reference average curvature of that unit, which are stored in the database as a long-term benchmark.

[0061] For each element, the square of the difference between its current Gaussian curvature and the reference Gaussian curvature, and the square of the difference between its current average curvature and the reference average curvature are calculated. These two values ​​are added together, multiplied by the area of ​​the local region, and integrated. Finally, the result is divided by the total area of ​​the local region to obtain the local differential curvature energy index value for that element. This index comprehensively considers both intrinsic and extrinsic curvature anomalies of the surface. It is insensitive to uniform tilt but strongly responds to non-uniform deformation modes such as local depressions and warps. The calculation results for all elements constitute the local differential curvature energy distribution map of the panel.

[0062] The warning threshold is determined using statistical methods. During the baseline data collection period for the healthy state, the time series of the local difference curvature energy index for all units is calculated, and the mean and standard deviation of the index values ​​for each unit are taken. The warning threshold is set as the mean plus three times the standard deviation. This threshold can be adjusted based on actual experience during the operational phase. When the local difference curvature energy index of any unit on the panel exceeds the warning threshold for a specified number of consecutive times, such as five consecutive sampling times, an abnormal event is detected, triggering the second level of processing.

[0063] After the first-level anomaly detection is triggered, the second-level processing is automatically started. The original multi-source sensor data for 30 seconds before and after the trigger time are packaged, and the high-order dynamic mode decomposition process described in S3 is called to perform fine decomposition of the deformation field during this period, accurately separate the quasi-static mode and the load tracking mode, reconstruct a high-quality non-uniform residual deformation field, and update the calculation results of the local differential curvature energy index for this period.

[0064] After obtaining accurate non-uniform residual deformation field and local differential curvature energy distribution through the second-level processing, spatiotemporal clustering is performed on cells exceeding the warning threshold. Spatially, adjacent cells exceeding the threshold are merged into the same cluster; temporally, clusters with overlapping spatial locations in consecutive time moments are concatenated into spatiotemporal clusters. Each spatiotemporal cluster records its spatial coverage, peak location, occurrence and duration time intervals, and peak intensity.

[0065] The identified spatiotemporal clusters are spatially matched with the load events recorded by the S1 sensing device. Specifically, this involves extracting the coordinates of the ship's berthing position, the coordinate sequence of the trolley trajectory of the rail-mounted crane, and the coordinates of the heavy-duty truck's travel path identified by the video analysis device within the same time period. The Euclidean distance between the geometric center of the spatiotemporal cluster and the location of each load event is calculated. If the distance is less than a preset matching threshold, such as 2 meters, the spatiotemporal cluster is considered to be spatially matched with the load event.

[0066] For spatially matched spatiotemporal clusters and load event pairs, the mutual information between them in the time dimension is further calculated to determine whether a causal relationship exists. Specifically, the time series of the local differential curvature energy peak of the spatiotemporal cluster is discretized into two states: exceeding a threshold and not exceeding a threshold. The intensity time series of the corresponding load event is also discretized into two states: above normal and normal. The mutual information value between the two discretized sequences is then calculated.

[0067] The significance of mutual information is determined using the substitution data method: the time series of load events is randomly shuffled 1000 times, and the mutual information value after each shuffle is calculated to form a zero distribution. If the actual mutual information value is greater than the 95th quantile of the zero distribution, the mutual information is determined to meet the preset significance condition, confirming a statistically significant causal relationship between the spatiotemporal cluster and the load event.

[0068] When mutual information meets the saliency condition and spatial matching error is within the allowable range, the non-uniform deformation damage is traced back to the corresponding load type: if the associated load is a ship berthing impact, the tracing result is damage caused by ship impact; if the associated load is the wheel pressure cycle of a rail crane, the tracing result is damage caused by rail machinery operation; if the associated load is the passage of a heavy truck, the tracing result is damage caused by heavy truck crushing.

[0069] After the second-level processing is completed, a high-fidelity finite element model is used to simulate and reproduce the abnormal event. The load spatiotemporal information recorded by the sensing device is used as the input boundary conditions of the finite element model to calculate the theoretical deformation response of the panel, and compared with the non-uniform residual deformation field separated in S3. If there is a systematic deviation between the two in spatial distribution and magnitude, it indicates that the structural parameters in the physical information graph network, such as bending stiffness and foundation coefficient, differ from the actual state. Based on the deviation, the parameter correction amount of each region is derived, and the node physical parameter characteristics of the physical information graph network in S2 are updated to complete the online correction of the model.

[0070] Finally, the output includes the non-uniform deformation monitoring results, which include damage location, damage mode, damage extent, damage tracing results, and recommended maintenance measures. These results are then pushed to the terminal operation and maintenance management platform in the form of a visual cloud map overlaid with a text report.

[0071] From deformation field reconstruction (S2) to non-uniform deformation separation (S3), and then to damage assessment and source tracing (S4), each step relies on a specific combination of multi-source data: S2 utilizes strain, tilt angle, and load information to achieve physically self-consistent full-field deformation reconstruction; S3 uses the time series of the deformation field and the load time series to separate residual short-wavelength modes; S4 calculates the curvature energy index based on the non-uniform residual deformation field and matches it again with the spatiotemporal location of the load to complete damage source tracing. This closed loop requires the sensing system to simultaneously possess spatial high resolution, absolute displacement reference, and external load sensing capabilities, which cannot be supported independently by a single type of sensor. Therefore, the combination of multi-source sensing methods described in this embodiment is an inevitable choice formed after a comprehensive analysis of the panel's mechanical behavior, damage mechanisms, and information processing links.

[0072] This invention addresses the problem that traditional single-point settlement or tilt indices are overly sensitive to uniform deformation and cannot distinguish between overall settlement and local damage by calculating a local differential curvature energy index based on the joint deviation of Gaussian curvature and average curvature. This improves the specificity of identifying non-uniform deformation patterns such as local depressions, warping, and voids, and reduces the probability of false alarms triggered by normal overall deformation. Furthermore, by spatially clustering local differential curvature energy units exceeding a threshold into spatiotemporal clusters, it solves the problem that isolated point anomalies are difficult to correlate with the actual damage area and are easily filtered out as noise, thus improving the spatial positioning accuracy and continuity of the damage area. By incorporating the mutual information calculation between the occurrence time of spatiotemporal clusters and the occurrence time of load events into causal correlation determination, and using the substitution data method to test significance, the problem of traditional monitoring being able to only issue abnormal alarms but not trace the cause of damage is solved, thus improving the reliability of damage source tracing. By setting up a two-level processing mechanism—the first level continuously calculates the local differential curvature energy index online and triggers the second level of fine decomposition and source tracing—the problem of large computational load and inability to run continuously in real time for high-precision analysis throughout the entire process is solved, thus improving the real-time response capability of the system, reducing the continuous occupation of cloud computing resources, and achieving an efficiency balance between real-time monitoring and in-depth diagnosis.

[0073] Example 2 like Figure 4 As shown, the present invention also provides a port terminal front panel non-uniform deformation monitoring system, the system being used to implement the port terminal front panel non-uniform deformation monitoring method described in Embodiment 1, the system comprising: A multi-source data acquisition module is used to acquire multi-source sensor data deployed on the panel structure. The multi-source sensors include at least orthogonally arranged distributed fiber optic strain sensors, an inclinometer array, and a sensing device for providing spatiotemporal information on ship berthing and unberthing, rail machinery and heavy-duty truck loads. The deformation field reconstruction module is used to discretize the panel structure into an unstructured graph, with nodes corresponding to physical spatial locations, edges encoding local stiffness and sensing topology, and fusing the multi-source sensor data into node features. These features are then input into a trained physical information graph network to reconstruct the spatiotemporal deformation field of the panel in real time. The loss function of the physical information graph network during training includes a physical residual term constructed based on the thin plate bending control equation to constrain the reconstructed deformation field to meet the mechanical behavior of the panel. The non-uniform deformation separation module is used to apply dynamic mode decomposition to the spatiotemporal deformation field, extract the modes and their time characteristic values ​​and spatial frequencies, and adaptively separate the residual shortwave modes based on the correlation between the modes and the known load time sequence and the spatial frequency characteristics, and reconstruct the non-uniform residual deformation field characterizing the non-uniform deformation. The damage assessment and source tracing module is used to calculate the local differential curvature energy index at each location based on the non-uniform residual deformation field. The local differential curvature energy characterizes the local deviation energy of Gaussian curvature and mean curvature relative to the reference curvature. By identifying spatiotemporal clusters where the index exceeds the threshold, and combining the load spatiotemporal information, damage source tracing is performed, and non-uniform deformation monitoring results are output.

[0074] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for monitoring uneven deformation of a port terminal front panel, characterized in that, include: S1, acquire multi-source sensor data deployed on the panel structure. The multi-source sensors include at least orthogonally deployed distributed fiber optic strain sensors, an inclinometer array, and a sensing device for providing spatiotemporal information on ship berthing and unberthing, rail machinery and heavy truck loads. S2, the panel structure is discretized into an unstructured graph, with nodes corresponding to physical spatial locations, edges encoding local stiffness and sensing topology, and the multi-source sensor data is fused into node features, which are then input into a trained physical information graph network to reconstruct the spatiotemporal deformation field of the panel in real time; wherein, the loss function of the physical information graph network during training includes a physical residual term constructed based on the thin plate bending control equation to constrain the reconstructed deformation field to satisfy the mechanical behavior of the panel; S3, Apply dynamic mode decomposition to the spatiotemporal deformation field, extract the modes and their time characteristic values ​​and spatial frequencies, and adaptively separate the residual shortwave modes based on the correlation between the modes and the known load time sequence and the spatial frequency characteristics, and reconstruct the non-uniform residual deformation field characterizing the non-uniform deformation. S4. Calculate the local differential curvature energy index at each location based on the non-uniform residual deformation field. The local differential curvature energy characterizes the local deviation energy of Gaussian curvature and average curvature relative to the reference curvature. Identify spatiotemporal clusters where the index exceeds the threshold, and combine them with the load spatiotemporal information to trace the source of damage and output the non-uniform deformation monitoring results.

2. The method for monitoring uneven deformation of the front panel of a port terminal according to claim 1, characterized in that, In the multi-source sensor data, distributed fiber optic strain sensors are orthogonally arranged bidirectionally along the bottom surface of the panel to form a quasi-continuous strain field sensing; the inclinometer array is arranged at intervals of 5 to 10 meters to obtain the biaxial tilt angle of the panel surface; the sensing device includes a pose encoder installed on a track-mounted machine and / or a video analysis device for identifying the position and load of a heavy-duty truck to provide the spatiotemporal information of the load.

3. The method for monitoring uneven deformation of the front panel of a port terminal according to claim 1, characterized in that, The physical information graph network is a message-passing graph neural network. The loss function used for training is composed of a data loss term, a physical residual term, and a smoothing regularization term, each multiplied by its corresponding weighting coefficient and then summed. The physical residual term is constructed as follows: sampling nodes are randomly selected in the unstructured graph. For each sampling node, the equation residual of the vertical displacement is calculated according to the thin plate bending control equation of the panel, and the sum of the squares of the equation residuals of all sampling nodes is taken as the physical residual term. The thin plate bending control equation includes terms related to the bending stiffness of the panel and the foundation coefficient, and the distributed load in the equation is determined by the load spatiotemporal information obtained by the sensing device. The data loss term is constructed based on the displacement obtained by integrating the tilt angle measured by the inclinometer array, and is used to constrain the consistency between the reconstructed deformation field and the measured displacement. The smoothing regularization term is used to penalize the variation amplitude of the deformation field between adjacent nodes in the unstructured graph to suppress the discontinuity of the reconstructed deformation field.

4. The method for monitoring uneven deformation of the front panel of a port terminal according to claim 1, characterized in that, The dynamic mode decomposition employs high-order dynamic mode decomposition and introduces a sliding window for online updating; the residual shortwave modes are adaptively separated based on the correlation between the modes and the known load time series and spatial frequency characteristics, specifically including: Calculate the temporal characteristic values ​​of each mode, and classify the modes with temporal characteristic value magnitude close to 1 and spatial frequency lower than the preset spatial threshold as quasi-static modes to characterize the overall settlement and temperature deformation. The correlation coefficient between the time coefficient of each mode and the load time sequence of ships and rail machinery provided by the sensing device is calculated. Modes with correlation coefficients higher than the preset correlation threshold are classified as load tracking modes to characterize interpretable working deformation. Modes with spatial frequencies higher than the preset spatial threshold, excluding quasi-static modes and load tracking modes, are classified as residual shortwave modes. The non-uniform residual deformation field is obtained by reconstructing using all the aforementioned residual shortwave modes.

5. The method for monitoring uneven deformation of the front panel of a port terminal according to claim 1, characterized in that, The calculation method for the local differential curvature energy index is as follows: For each element in the discrete element of the panel, take the local region consisting of itself and its neighborhood; The sum of the square of the difference between the Gaussian curvature derived from the non-uniform residual deformation field and the reference Gaussian curvature, and the sum of the square of the difference between the average curvature and the reference average curvature within the local region, and the result of the area integration over the local region divided by the area of ​​the local region, is the local differential curvature energy of the element. The reference Gaussian curvature and the reference average curvature are determined by the deformation field obtained by the panel in a healthy state.

6. The method for monitoring uneven deformation of the front panel of a port terminal according to claim 1, characterized in that, The damage tracing based on the load spatiotemporal information includes: Spatial matching is performed between the spatiotemporal clusters of local differential curvature energy exceeding the threshold and the ship berthing position, tracked machinery trajectory, or heavy-duty truck path provided by the sensing device. Calculate the mutual information between the occurrence time of the spatiotemporal cluster and the occurrence time of the load event; When mutual information meets the preset significant conditions and spatial matching error is within the allowable range, the non-uniform deformation damage is traced back to the corresponding ship collision, rail-mounted machinery operation, or heavy-duty truck passage.

7. The method for monitoring uneven deformation of the front panel of a port terminal according to claim 1, characterized in that, The real-time reconstructed panel spatiotemporal deformation field and the calculation of local differential curvature energy index are continuously executed online as the first-level processing. When the local differential curvature energy index exceeds the preset warning threshold, the steps of dynamic mode decomposition, non-uniform residual deformation field reconstruction and damage tracing are triggered as the second level of processing. After the second level of processing is completed, the parameters of the physical information graph network are corrected and updated using the finite element model and historical accumulated data.

8. The method for monitoring uneven deformation of the front panel of a port terminal according to claim 1, characterized in that, The edge-encoded local stiffness of the unstructured graph includes a function of the physical distance between nodes and the estimated elastic modulus; the residual shortwave mode is a modal component in the dynamic mode decomposition result whose time characteristic value is not quasi-static and is not significantly correlated with the known load time sequence, and whose spatial frequency is higher than a set threshold, and is used to characterize non-uniform deformations such as local depressions, warping or voids.

9. A monitoring system for uneven deformation of a port terminal front panel, characterized in that, The system is used to implement the method for monitoring uneven deformation of port terminal front panel as described in any one of claims 1 to 8, the system comprising: A multi-source data acquisition module is used to acquire multi-source sensor data deployed on the panel structure. The multi-source sensors include at least orthogonally arranged distributed fiber optic strain sensors, an inclinometer array, and a sensing device for providing spatiotemporal information on ship berthing and unberthing, rail machinery and heavy-duty truck loads. The deformation field reconstruction module is used to discretize the panel structure into an unstructured graph, with nodes corresponding to physical spatial locations, edges encoding local stiffness and sensing topology, and fusing the multi-source sensor data into node features. These features are then input into a trained physical information graph network to reconstruct the spatiotemporal deformation field of the panel in real time. The loss function of the physical information graph network during training includes a physical residual term constructed based on the thin plate bending control equation to constrain the reconstructed deformation field to meet the mechanical behavior of the panel. The non-uniform deformation separation module is used to apply dynamic mode decomposition to the spatiotemporal deformation field, extract the modes and their time characteristic values ​​and spatial frequencies, and adaptively separate the residual shortwave modes based on the correlation between the modes and the known load time sequence and the spatial frequency characteristics, and reconstruct the non-uniform residual deformation field characterizing the non-uniform deformation. The damage assessment and source tracing module is used to calculate the local differential curvature energy index at each location based on the non-uniform residual deformation field. The local differential curvature energy characterizes the local deviation energy of Gaussian curvature and mean curvature relative to the reference curvature. By identifying spatiotemporal clusters where the index exceeds the threshold, and combining the load spatiotemporal information, damage source tracing is performed, and non-uniform deformation monitoring results are output.

Citation Information

Patent Citations

  • Method for indirectly measuring cable tension based on bollard surface strain

    CN117852368A

  • Whole ship strain field rapid reconstruction method using ship motion and discrete strain information

    CN120449307A