A masonry building existing building disease intelligent evaluation and protection decision system

Abstract

The present application belongs to the technical field of machine learning, and relates to a stone masonry existing building disease intelligent evaluation and protection decision system, comprising: an initial feature generation module, which establishes a structure dynamics attribute graph and generates an initial state feature set; a state evolution solving module, which performs coupling evolution of mechanics and implicit degradation state to generate a state transition matrix; a prediction residual calculation module, which performs forward fitting along a continuous time domain and calculates a mechanics response measurement residual in combination with an observation benchmark; a state inversion calculation module, which performs back propagation based on the residual to invert an unobserved area implicit physical degradation state field to generate a model uncertainty atlas; a double-target cost calculation module, which fuses data to calculate a double-target cost benchmark; and a feedback decision execution module, which distributes instructions and updates the initial state feature set accordingly. The present application solves the problem that the prior art is difficult to fuse physical mechanism and limited observation data, realizes inversion of an implicit degradation state in a structure and forward-looking risk decision-making.

Description

Technical Field

[0001] This invention belongs to the technical field of machine learning and relates to an intelligent assessment and protection decision-making system for existing stone masonry buildings. Background Technology

[0002] Long-term health monitoring and preventive protection of existing masonry buildings is an important research direction in the field of building operation and maintenance. These buildings are subject to the combined effects of natural environment and human activities over long periods, making them prone to hidden and progressive structural defects, such as internal mortar deterioration and microcrack propagation. By the time obvious deformation or cracks appear, the structure has often already suffered significant damage. Traditional delayed intervention measures are limited by their implementation difficulties and the risk of altering the original appearance of the building. Therefore, the industry urgently needs an intelligent operation and maintenance technology that supports early warning and objective decision-making.

[0003] Currently, commonly used detection methods in the industry mainly include periodic non-destructive inspection based on equipment scanning and visual recognition technology based on image acquisition. The former uses ground-penetrating radar or ultrasonic sensors to periodically inspect buildings, aiming to identify macroscopic defects such as voids or segregation within the masonry. The latter uses drones or ground-based equipment to acquire high-resolution images of building surfaces, combining them with computer vision algorithms to extract and classify features of visible surface defects such as cracks and weathering. Both of these technologies provide effective data support for the static identification of specific defects and offer intuitive references for building maintenance.

[0004] However, the aforementioned existing technologies have certain limitations in addressing the dynamic assessment of the progressive degradation process of masonry. Non-destructive testing based on periodic scanning is a static detection at discrete time points, making it difficult to capture the coupling relationship between the dynamic mechanical response of the structure under continuous external loads and the evolution of microscopic material properties, easily overlooking localized accelerated degradation processes caused by stress concentration. Simultaneously, visual image-based recognition methods mainly rely on surface optical features, making it difficult to effectively perceive the internal implicit physical degradation states affecting the overall load-bearing capacity of the structure, such as the decrease in mortar joint contact stiffness and changes in the friction coefficient. Existing solutions mostly focus on the superficial diagnosis of existing damage, lacking a dynamic computational framework that integrates physical evolution mechanisms with limited observational data. This leads to long-term protection decisions often relying on static assessment results and human experience, lacking quantitative consideration of future risk evolution trends. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides an intelligent assessment and protection decision-making system for existing stone masonry buildings.

[0006] A smart assessment and protection decision-making system for existing masonry buildings, comprising: The initial feature generation module is used to acquire the sensor monitoring signals of the surface of the existing stone masonry building, establish a structural dynamic property map based on the sensor monitoring signals, and generate an initial state feature set. The state evolution solution module is used to input the initial state feature set into a preset evolution solver, perform bidirectional coupled evolution calculation of mechanical state and implicit degenerate state, and generate state transition matrix; The prediction residual calculation module is used to perform forward fitting evolution by traversing the state transition matrix along the continuous time domain, extract global node evolution prediction data, and calculate the mechanical response measurement residual based on the global node evolution prediction data and the actual observation data. The state inversion calculation module is used to perform backpropagation calculations based on the mechanical response measurement residuals, invert the implicit physical degradation state field of the unobserved region, and generate a model uncertainty map; The dual-objective cost calculation module is used to integrate global node evolution prediction data and model uncertainty map, and combine them with a preset hybrid decision space table to calculate the dual-objective cost benchmark. The feedback decision execution module is used to select security intervention decision sequences based on a dual-objective cost benchmark, distribute detection and hardening instructions according to the security intervention decision sequences, and obtain perception and monitoring data to update the initial state feature set.

[0007] A further aspect of the present invention includes an initial feature generation module, used to perform the following steps: Acquire multi-source monitoring signals output by each independent sensor, extract the temporal micro-vibration components and displacement change rate characteristics of the multi-source monitoring signals, and combine them to generate a basic observation state set; The stone bearing units of the existing stone masonry building are set as topological nodes, and the mortar joint material interfaces between the stone bearing units are defined as topological edges. The structural dynamic property diagram is constructed based on the spatial arrangement and combination of topological nodes and topological edges. Extract the motion state parameters and mass inertia properties of the topological nodes, and determine the normal contact stiffness and tangential structural friction coefficient of the topological edges; An implicit physical degradation state vector is constructed for topological nodes and edges. The implicit physical degradation state vector is then connected with the basic observation state set, and the data dimensions are concatenated to generate an initial state feature set.

[0008] A further aspect of the present invention includes a state evolution solution module, which is used to perform the following steps: Receive the initial state feature set, extract the integration time step of the evolution solver, construct the discretized force balance transfer equation based on the integration time step, and adapt the discretized force balance transfer equation to the structural dynamics property map. Extract the nodal stress state data contained in the initial state feature set, and calculate the material degradation increment corresponding to the implicit physical degradation state vector; The implicit physical degradation state vector is updated by superimposing the numerical increment of material degradation and the historical degradation baseline of the system. The three-dimensional distribution of the force field response in the next time slice is calculated by subtracting the normal contact stiffness value of the topological edge from the updated implicit physical degradation state vector. The three-dimensional distribution law of the combined force field response and the numerical increment of material degradation are merged to generate a state transition matrix.

[0009] A further aspect of the present invention includes a residual prediction calculation module, used to perform the following steps: Based on the discretized force balance transfer equation, the state transition matrix is ​​derived, and continuous integration calculation steps are performed along the continuous time domain to obtain the expected mean parameter and discrete variance parameter of the global nodal displacement field vector at the end time slice of the evolution calculation. The expected mean parameter and discrete variance parameter are smoothed, and the predicted probability distribution tensor is reconstructed and used as the global node evolution prediction data. Extract the predicted mechanical characteristic response value corresponding to the observation position coordinates from the global nodal displacement field vector, and call the corresponding original displacement reading as the actual observation data; The absolute deviation of the predicted mechanical characteristic response value from the actual observed data is calculated, and the absolute deviation is defined as the mechanical response measurement residual.

[0010] A further aspect of the present invention includes a state inversion calculation module, used to perform the following steps: The network weight target partial derivative gradient flow triggered by the mechanical response measurement residual is used to extract the state back-correction amount; The reverse state correction is delivered along the network information back-tracing path defined in the structural dynamics property graph, and the reverse state correction is assigned to the non-observation nodes in the structural dynamics property graph that have not deployed sensors. Based on the state inverse correction amount, the implicit physical degradation state vector carried by the non-observation node is adjusted and corrected in reverse, and the implicit physical degradation state field is reconstructed and restored. Perform multiple concurrent projective simulation samplings on an implicit physical degenerate state field, and calculate the posterior variance generated by the multiple concurrent projective simulation samplings. Extract the contours of regions with posterior variance greater than a preset variance threshold to generate a model uncertainty map.

[0011] A further aspect of the present invention includes a dual-objective cost calculation module, used to perform the following steps: Extract the structural failure probability early warning index from the predicted probability distribution tensor and calculate the total variance probability integral of the region covered by the model uncertainty map. The structural failure probability early warning index is combined with the total variance probability integral to form multidimensional objective optimization constraint data; Using the total variance probability integral as the penalty benchmark when no exploration is performed, the penalty cost for information loss in the corresponding risk blocking intervention action set is quantitatively calculated. The structural failure probability warning index is used as the penalty benchmark when reinforcement is not implemented, and the structural risk penalty cost corresponding to the active exploration action set is quantitatively calculated. The basic engineering costs of each decision-making action are combined with their corresponding unresolved penalty costs to generate a dual-objective cost benchmark.

[0012] A further aspect of the present invention includes a feedback decision execution module, used to perform the following steps: Within a pre-defined hybrid decision space, a decision sequence is generated based on a random walk algorithm. Then, a genetic algorithm is invoked to calculate the overall operating cost of the decision sequence based on a dual-objective cost benchmark. Remove the sequence terms that do not satisfy the cost constraint to obtain the safe intervention decision sequence that satisfies the overall cost minimization constraint; Extract the preferred action from the security intervention decision sequence and determine the business type of the preferred action; If the preferred action belongs to the active exploration action set, package and assemble the three-dimensional spatial coordinate positioning data group, and send the three-dimensional spatial coordinate positioning data group to trigger the execution of the exploration command; If the first action to be executed belongs to the risk prevention and intervention action set, a reinforcement instruction will be output for the damaged area of ​​the masonry.

[0013] A further aspect of this invention involves extracting the temporal micro-vibration components and displacement change rate features of multi-source monitoring signals, and combining them to generate a basic observation state set, including: Acquire discrete sampling sequences of multi-source monitoring signals within a preset time window; Perform a summation of squares on the discrete sampled sequence, and divide the summation of squares by the total number of sample points to obtain the mean data; Perform a square root operation on the mean data to obtain the root mean square value of the time-series micro-vibration component; Perform a first-order difference operation on the displacement sequence in the multi-source monitoring signal to calculate the displacement increment between adjacent sampling time points and obtain the displacement change rate characteristics; The root mean square value and displacement rate of change characteristics are organized and combined according to the measurement point identifier to generate a basic observation state set.

[0014] A further aspect of the present invention includes the step of calculating and obtaining the material degradation numerical increment corresponding to the implicit physical degradation state vector, comprising: Extract the effective stress values ​​acting on the mortar joint interface from the nodal stress state data; Obtain the preset stress corrosion rate coefficient, multiply the effective stress value with the stress corrosion rate coefficient, and obtain the basic degradation rate. The numerical increment of the chemical corrosion factor is calculated by multiplying the base degradation rate by the integration time step. The numerical increment of the chemical corrosion factor is used as the numerical increment of material degradation corresponding to the implicit physical degradation state vector.

[0015] A further aspect of the present invention includes the step of subtracting the normal contact stiffness value of the topological edge based on the updated implicit physical degradation state vector, comprising: Extract the initial normal contact stiffness value of the topological edge; Traverse the updated implicit physical degradation state vector for multiple degradation dimensions that affect stiffness, and obtain the cumulative value of each degradation dimension. The difference between the value 1 and the cumulative value of each degradation dimension is calculated to obtain the retention ratio coefficient corresponding to each degradation dimension; The overall attenuation coefficient is obtained by multiplying the retention ratios corresponding to all degradation dimensions. The initial normal contact stiffness value is multiplied by the comprehensive attenuation coefficient to obtain the deducted normal contact stiffness value.

[0016] In summary, the present invention has the following beneficial technical effects: 1. By constructing a structural dynamics property graph and defining a bidirectional coupled evolution rule between the mechanical state and the implicit degradation state using a differentiable graphical Lagrangian evolution solver, a physical correlation between macroscopic mechanical response and microscopic material degradation is established at the computational model level. By embedding Lagrangian dynamics as a physical prior condition into a graph neural network framework, this approach not only achieves data-driven fitting but also simulates the physical processes of stress transfer, deformation, and stress-accelerated material degradation under load. This mechanism enables the system to follow physical laws when making long-term behavior predictions, effectively suppressing the accumulation of prediction errors caused by the lack of a physical mechanism in purely data-driven models, thereby improving the accuracy of structural state evolution predictions and the interpretability of the model.

[0017] 2. By utilizing the underlying automatic differential computation engine to backpropagate the mechanical response measurement residuals at the sensor deployment locations, the extracted state back-correction quantities are used to inversely adjust the implicit physical degradation state of areas without deployed sensors. This data assimilation mechanism can transform local macroscopic measurement errors into correction signals for the global microscopic physical state, enabling the model to calibrate its assessment of the overall internal state of the building based on limited surface observation feedback. This approach reduces the dependence of traditional detection techniques on the number of measurement points and physical coverage, achieving effective reconstruction of the implicit physical degradation field within the three-dimensional space of the building using limited observation data, thereby expanding the perception range of the structural internal health state.

[0018] 3. By generating a model uncertainty map representing the uncertainty of model predictions, and using it along with the structural failure probability as input to a dual-objective cost calculation module, a hybrid strategy of surveying and reinforcement is calculated. This mechanism introduces a quantitative consideration of model prediction uncertainty into the decision-making process, enabling the system to dynamically balance reducing structural risk with reducing model prediction uncertainty. When the model's risk assessment of a certain area has high uncertainty, the system prioritizes outputting proactive exploration and survey commands to supplement data, thereby reducing the probability of assessment bias caused by missing model parameters. This scheme can provide a reasonable intervention decision sequence based on the spatial distribution of risk and uncertainty, improving the targeting and scientific nature of operation and maintenance resource allocation. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. The drawings are used to provide a further understanding of the present invention.

[0020] Figure 1 This is a schematic diagram of the framework in the embodiments of this application.

[0021] Figure 2 This is a flowchart illustrating an embodiment of this application.

[0022] Figure 3 This is a schematic diagram of the extraction of time-series micro-vibration components from the basic observation state set in this application embodiment.

[0023] Figure 4 This is a diagram of the dynamic deduction and attenuation model of the normal contact stiffness of the topological edge in the embodiments of this application.

[0024] Figure 5 This is a graph showing the node evolution prediction probability distribution and prediction residual feature extraction in the embodiments of this application.

[0025] Figure 6 This is a diagram showing the convergence trajectory of the residual inversion of the implicitly degraded state field in the non-observation region in this embodiment of the application.

[0026] Figure 7 This is a posterior variance distribution map of the implicit physical degradation state of the unobserved area in the embodiments of this application.

[0027] Figure 8 This is a baseline chart comparing the quantitative differences in the dual-objective penalty costs in the embodiments of this application.

[0028] Figure 9 This is the optimization space diagram of the bi-objective genetic algorithm strategy in the feedback decision execution stage of this application embodiment. Detailed Implementation

[0029] The following is in conjunction with the appendix Figures 1-9 A preferred description of the present invention is provided below.

[0030] See attached document Figures 1-9 This invention proposes an intelligent assessment and protection decision-making system for existing masonry buildings, comprising the following modules: The initial feature generation module is used to acquire the sensor monitoring signals of the surface of the existing stone masonry building, establish a structural dynamic property map based on the sensor monitoring signals, and generate an initial state feature set. The state evolution solution module is used to input the initial state feature set into a preset evolution solver, perform bidirectional coupled evolution calculation of mechanical state and implicit degenerate state, and generate state transition matrix; The prediction residual calculation module is used to perform forward fitting evolution by traversing the state transition matrix along the continuous time domain, extract global node evolution prediction data, and calculate the mechanical response measurement residual based on the global node evolution prediction data and the actual observation data. The state inversion calculation module is used to perform backpropagation calculations based on the mechanical response measurement residuals, invert the implicit physical degradation state field of the unobserved region, and generate a model uncertainty map; The dual-objective cost calculation module is used to integrate global node evolution prediction data and model uncertainty map, and combine them with a preset hybrid decision space table to calculate the dual-objective cost benchmark. The feedback decision execution module is used to select security intervention decision sequences based on a dual-objective cost benchmark, distribute detection and hardening instructions according to the security intervention decision sequences, and obtain perception and monitoring data to update the initial state feature set.

[0031] In one embodiment of the present invention, the initial feature generation module is configured to perform the following steps: Acquire multi-source monitoring signals output by each independent sensor, extract the temporal micro-vibration components and displacement change rate characteristics of the multi-source monitoring signals, and combine them to generate a basic observation state set; The stone bearing units of the existing stone masonry building are set as topological nodes, and the mortar joint material interfaces between the stone bearing units are defined as topological edges. The structural dynamic property diagram is constructed based on the spatial arrangement and combination of topological nodes and topological edges. Extract the motion state parameters and mass inertia properties of the topological nodes, and determine the normal contact stiffness and tangential structural friction coefficient of the topological edges; An implicit physical degradation state vector is constructed for topological nodes and edges. The implicit physical degradation state vector is then connected with the basic observation state set, and the data dimensions are concatenated to generate an initial state feature set.

[0032] Specifically, the data acquisition program is initiated by deploying multiple independent sensors at key structural parts of existing masonry buildings. The independent sensors are triaxial physical accelerometers and fiber optic displacement measuring instruments. The data acquisition program continuously collects and records the voltage or light wavelength change signals output by each measuring point at a preset sampling frequency, such as 200 times per second, and converts these raw signals into digital time-series data streams to form multi-source monitoring signals.

[0033] The signal processing module uses a bandpass filter to remove high-frequency noise and low-frequency drift that are unrelated to the structure's natural vibration frequency from the multi-source monitoring signals, thereby extracting the effective temporal micro-vibration components characterizing the structure's dynamic response. Simultaneously, the module performs a first-order difference operation on the output sequence of the displacement measuring instrument to calculate the displacement increment between adjacent sampling time points, thus obtaining the displacement rate of change characteristics.

[0034] The temporal micro-vibration components and displacement rate features extracted from all measuring points are organized and combined into a structured basic observation state set based on the measuring point identifiers. In parallel, the system retrieves a pre-stored 3D digital geometric model of the existing masonry building. This model discretizes the building into a plurality of stone bearing units and the mortar interface connecting them. Based on this model, the system defines each stone bearing unit as a topological node in a graph network, and the mortar interface between adjacent stone bearing units as a topological edge, constructing an undirected graph-like structural dynamics property graph based on the actual spatial connection relationships.

[0035] For each topological node, the system extracts its volume from the geometric model and, in conjunction with the preset stone density parameters, calculates the node's mass inertia properties, while simultaneously initializing its six-degree-of-freedom motion state parameters as zero vectors.

[0036] For each topological edge, the system queries the engineering material parameter library based on the type and thickness of the mortar joint material and assigns it an initial normal contact stiffness value and tangential structural friction coefficient.

[0037] Based on this, the system constructs an implicit physical degradation state vector for all topological nodes and topological edges. This vector contains multiple degradation potential indices used to characterize the degradation trend at the microscopic level of the material. These indices are set to zero or estimated baseline values ​​based on historical environmental data in the initial state.

[0038] The system integrates and aligns the basic observation state set data of the topological nodes corresponding to the sensor measurement points in spatial location, as well as the six-degree-of-freedom motion state parameters, mass inertia properties, and implicit physical degradation state vectors of all topological nodes and topological edges, to generate a unified, graph-structured initial state feature set, which serves as the starting condition for subsequent evolution calculations.

[0039] To quantify the intensity of a time-series micro-vibration component, its root mean square value within a specific time window can be calculated as one of the characteristics. Let a discrete-time micro-vibration component signal of length N be... ,in Then its root mean square value The calculation formula is as follows: ; In the formula, Represents at discrete time points The collected acceleration values ​​are usually in units of 1. N represents the total number of sampling points used for calculation. It is a dimensionless positive integer, and its value determines the length of the time window for calculating the root mean square value. For example, it can be set to correspond to the amount of data in 1 second. The root mean square value of the calculated acceleration signal is given, and its units are... The units are consistent, for .

[0040] It should be noted that independent sensors are specific monitoring points located on the surface of a building structure to acquire dynamic response data of the structure under environmental loads. Multi-source monitoring signals mean that sensors do not cover all building units, but are deployed in a limited number of key locations, and their data sources include both acceleration and displacement physical quantities.

[0041] Temporal micro-vibration components specifically refer to the signal portion extracted from acceleration signals that reflects the minute vibrations of a structure under excitations such as wind loads and traffic vibrations; their frequency range is typically set between 0.5 Hz and 50 Hz. Displacement rate of change characteristics are indicators that measure the rate of structural deformation. Structural dynamic property diagrams are used to represent the elements of a real building and their interactions in calculations.

[0042] The six-degree-of-freedom motion state parameters are six-dimensional vectors containing three translational velocity components and three rotational velocity components, used to describe the motion tendency of the stone element as a rigid body. The normal contact stiffness value is a physical quantity characterizing the ability of mortar joint material to resist deformation under compression. Its setting is based on material mechanics test data. Assuming the building mortar joint is standard lime mortar, its initial value can be set in the range of 3 to 8 gigapascals.

[0043] The tangential structural friction coefficient is used to describe the critical condition for shear slip between stones, and its value typically ranges from 0.6 to 0.8. The implicit physical degradation state vector is one of the core definitions of this scheme. It is a multi-dimensional vector designed to quantify the slowly changing internal degradation processes of materials that cannot be directly measured. For example, this vector may contain three dimensions, representing the chemical corrosion factor, microcrack density parameter, and internal water saturation, respectively. Each dimension is normalized to the range of 0 to 1, where 0 represents no degradation and 1 represents complete failure. The initial values ​​of these indicators are set by expert evaluation based on the building's age, historical climate data, and known repair records.

[0044] In one embodiment of the invention, the evaluation process is quantified into a specific attenuation calculation formula. This is based on the microcrack density parameter. Taking the initial value as an example, its calculation formula is set as follows: ; in, The number of years since the building was constructed; An empirical aging coefficient is based on historical climate data, such as the annual average temperature difference rate of change. The validity weight parameter for known repaired records, with a value between 0 and 1; To repair the compensation coefficient. Through this quantification formula, the system can eliminate subjective intervention and automatically assign definite prior initial values ​​to the implicit physical degradation state vector.

[0045] The initial state feature set is a composite data structure that encapsulates the graph's topological relationships, the physical attributes of nodes, the connectivity attributes of edges, and the observed and implicit degenerate states of all units, providing a normalized input for subsequent physical neural network computations.

[0046] In one example of this embodiment, the masonry structure consists of three stone-bearing units, labeled as node 1, node 2, and node 3, and two mortar joint interfaces connecting them, labeled as edge 1-2 and edge 2-3, forming a linear structural dynamic property diagram. It is assumed that a triaxial physical accelerometer and displacement meter are deployed at node 1. Within a certain time window of one second, the data acquisition program acquires a set of time-series micro-vibration components containing 200 sampling points. Axial acceleration sequence Partial readings are Simultaneously obtained displacement sequence The last two readings are 0.15 mm and 0.17 mm.

[0047] The signal processing module first calculates The root mean square value of axial acceleration, such as Figure 3As shown, applying the aforementioned formula, the mean of the sum of squares of this sequence is ,but Meanwhile, the displacement rate of change is characterized as 4 mm / s. These calculations constitute part of the basic observation state set at node 1. For node 2, which has no sensors deployed, its geometric model is shown as a 1 m × 0.5 m × 0.5 m limestone block with a density of Based on this, the system calculates the mass in its mass-inertia properties to be 625 kg, and initializes its six-degree-of-freedom motion state parameters to a zero vector. For edge 1-2 connecting node 1 and node 2, the system assigns a normal contact stiffness of 5 gigapascals and a tangential structural friction coefficient of 0.7 from the parameter library, based on the setting that it is lime mortar. At this time, assuming that based on expert evaluation, the initial implicit physical degradation state vector of node 2 is set to... These correspond to chemical corrosion, microcracks, and moisture content, respectively.

[0048] The feature vector generated for node 2 will be concatenated with its mass inertia property, initial motion state, and implicit physical degradation state vector; the feature vector generated for edge 1-2 will be concatenated with its stiffness, friction coefficient, and its own implicit physical degradation state vector. All these vectors together constitute part of the initial state feature set.

[0049] In one embodiment of the present invention, the state evolution solution module is used to perform the following operations: Receive the initial state feature set, extract the integration time step of the evolution solver, construct the discretized force balance transfer equation based on the integration time step, and adapt the discretized force balance transfer equation to the structural dynamics property map. Extract the nodal stress state data contained in the initial state feature set, and calculate the material degradation increment corresponding to the implicit physical degradation state vector; The implicit physical degradation state vector is updated by superimposing the numerical increment of material degradation and the historical degradation baseline of the system. The three-dimensional distribution of the force field response in the next time slice is calculated by subtracting the normal contact stiffness value of the topological edge from the updated implicit physical degradation state vector. The three-dimensional distribution law of the combined force field response and the numerical increment of material degradation are merged to generate a state transition matrix.

[0050] Specifically, the initial state feature set is loaded into the input interface of the differentiable graph Lagrange evolution solver. This solver is configured with a fixed symplectic geometry to maintain the integration time step, ensuring energy conservation during subsequent dynamic evolution. Based on this symplectic geometry-maintained integration time step, the solver transforms the continuous-time Lagrange dynamic equations into a set of discretized force balance transfer equations. This set of equations is adapted to the topology of the structural dynamic property graph, specifically manifested as an iterative message-passing computation paradigm. In the specific neural network engineering implementation, each message-passing process of this solver includes: First, using a multilayer perceptron with two fully connected layers and the Swish activation function as an edge update network, extracting the current contact force and stiffness attributes of the topological edges to generate feature vectors; next, using a summation aggregation operator to collect messages passed from adjacent topological edges to the target topological node; finally, using another node update network with three fully connected layers and the ReLU activation function, combining the aggregated messages with the current node state, outputting force and torque. This network architecture design ensures that mechanical evolution calculations can be forward propagated within the deep learning framework as differentiable tensor operations.

[0051] Within each time step, each topological edge calculates the contact force and torque of the interaction based on the current relative displacement and velocity of the two topological nodes it connects to, as well as its own stiffness and friction properties, and transmits these forces and torques as messages to the connected topological nodes.

[0052] The system incorporates predefined mechanical evolution and chemical degradation functions into the core computational logic of the differentiable graph Lagrange evolution solver. The solver extracts nodal stress state data associated with each topological edge from the initial state feature set. This stress data is used as input to the chemical degradation function, which, based on a predefined material degradation physics model (e.g., stress corrosion model), calculates the increments of various indices in the implicit physical degradation state vector corresponding to each topological edge and node within the current time step—that is, the material degradation numerical increments.

[0053] The system adds the incremental value of the material degradation to the historical degradation baseline recorded by the system up to the previous time step, thereby updating the implicit physical degradation state vector to reflect the slight deterioration of the material properties.

[0054] Next, the system extracts the updated implicit physical degradation state vector and proportionally subtracts the normal contact stiffness value of the topological edge based on the dimension representing stiffness decay.

[0055] After the deduction is completed, the new normal contact stiffness value is immediately injected into the mechanical evolution function, which is then called to calculate the three-dimensional distribution of the force field response of the entire structure's dynamic property map in the next time slice under the new material parameters, specifically manifested as the predicted acceleration values ​​of all topological nodes.

[0056] The system combines the calculated three-dimensional distribution law of the force field response with the material degradation increment calculated in the same time step to form a high-dimensional matrix. This matrix completely describes the coupling evolution relationship between the mechanical state and the degradation state of the entire system from the current moment to the next moment, which is the state transition matrix.

[0057] To achieve bidirectional coupling between the mechanical state and the implicit degradation state, the chemical degradation function is used to calculate the numerical increment of stress-driven material deterioration; taking one degradation dimension as an example, such as the chemical corrosion factor, its increment... The calculation can be simplified to: ; In the formula, In time step The calculated numerical increment of the chemical corrosion factor is dimensionless. It is the stress corrosion rate coefficient, and its unit is . This coefficient needs to be calibrated through accelerated aging tests on materials. The effective stress acting on the joint interface is expressed in Pascals and is extracted and calculated from the nodal condition. It is the integral time step that maintains the symplectic geometric structure, in seconds.

[0058] The new stiffness value is a function that inversely affects the normal contact stiffness based on the updated degradation state. Calculated by the following formula: ; In the formula, This is the updated normal contact stiffness value, in Pascals. It is the initial normal contact stiffness value of the material. It is the updated version of the first The cumulative value of the dimension of an implicit physical degradation state vector is dimensionless. This represents the total number of degenerate dimensions affecting stiffness. The product form here... This represents the combined weakening effect of multiple degradation mechanisms on stiffness; in actual numerical calculations, to prevent the occurrence of singularity in the mechanical matrix of the system, when... When the value approaches 1, a very small residual stiffness retention factor can be assigned for truncation protection.

[0059] Differentiable graph Lagrange evolution solvers are a special computational framework that approximates the Lagrange dynamics equations in the form of graph neural networks. Every step of the forward computation process is derivative with respect to the input and parameters, which is the basis for subsequent backpropagation optimization.

[0060] The integral time step of the symplectic geometry is a key parameter for controlling the simulation accuracy. It is set based on maximizing computational efficiency while ensuring numerical stability. Assuming the application scenario is conventional wind load and traffic vibration simulation, its typical value can be set to 0.001 seconds to 0.01 seconds.

[0061] It should be noted that the mechanical evolution function is responsible for calculating the interaction forces between nodes based on the current geometry and material parameters. The chemical degradation function, on the other hand, encapsulates prior knowledge from materials science and describes how external stresses accelerate the microscopic degradation process within the material.

[0062] Nodal stress state data can be approximated from the displacement of nodes and the deformation of surrounding elements using stress recovery techniques in the finite element method. The state transition matrix is ​​a mathematical operator that, when applied to the system state vector at the current moment, outputs the system state vector at the next moment, and internally encodes complex nonlinear coupled dynamics.

[0063] In one example of this embodiment, the system receives an initial state feature set containing three nodes and two edges, and sets the symplectic geometry to maintain the integration time step. It is 0.005 seconds. At the current moment. The solver calculates the effective normal stress on edge 1-2 connecting node 1 and node 2. The value is 2 MPa. Assume the stress corrosion rate coefficient is set in the material library. for The system calls the chemical degradation function to calculate the numerical increment of material deterioration due to chemical corrosion factors. In the implicit physical degradation state vector of edge 1-2, the system history degradation cardinality of the first dimension is 0.05, and the updated value is... Assume that only chemical corrosion and microcracks affect stiffness, and that the cumulative value of the microcrack dimension is 0.1. For example... Figure 4 As shown, after deducting its normal contact stiffness value, its initial value... The stiffness is 5 gigapascals. The new stiffness value is calculated as follows: The weakened stiffness value of 4.275 gigapascals is injected into the mechanical evolution function. Subsequently, the solver uses this new stiffness value, combined with the current motion state of all nodes in the graph, to calculate the acceleration of all nodes in the next time slice, which is the three-dimensional distribution of the force field response. Finally, this acceleration distribution is compared with the incremental material degradation values ​​calculated for all elements, such as edge 1-2. They were encapsulated together, forming a sequence from time to time. arrive The state transition matrix.

[0064] In one embodiment of the present invention, the prediction residual calculation module is used to perform the following steps: Based on the discretized force balance transfer equation, the state transition matrix is ​​derived, and continuous integration calculation steps are performed along the continuous time domain to obtain the expected mean parameter and discrete variance parameter of the global nodal displacement field vector at the end time slice of the evolution calculation. The expected mean parameter and discrete variance parameter are smoothed, and the predicted probability distribution tensor is reconstructed and used as the global node evolution prediction data. Extract the predicted mechanical characteristic response value corresponding to the observation position coordinates from the global nodal displacement field vector, and call the corresponding original displacement reading as the actual observation data; The absolute deviation of the predicted mechanical characteristic response value from the actual observed data is calculated, and the absolute deviation is defined as the mechanical response measurement residual.

[0065] Specifically, the differentiable graph Lagrange evolution solver is initiated, and using the state transition matrix as the core operator, it performs a forward fitting evolution along a preset continuous time domain model, such as a continuous time window of length T. This evolution process is an iterative computation process. Within each symplectic geometrical conserving integration time step, the solver applies the state transition matrix to update the system state vector of the current entire structural dynamic property graph to deduce the state of the next time slice.

[0066] Due to the inherent uncertainties in model parameters and external loads, this evolution process can be executed multiple times to form a predictive ensemble, or a stochastic process can be incorporated into the model. When the evolution calculation reaches the final time slice T, for each topological node in the structural dynamics property graph, the system statistically analyzes the displacement output values ​​obtained from multiple predictive ensembles and calculates the first and second central moments of these output values, thereby obtaining the expected mean parameter and discrete variance parameter of the global nodal displacement field vector for that node.

[0067] The system extracts the expected mean and discrete variance parameters of all topological nodes in parallel and feeds them into the kernel density estimation algorithm. This algorithm performs nonparametric smoothing on the predicted displacement distribution of each node, transforming the discrete mean and variance into a continuous probability density function. By integrating the probability density functions of all nodes throughout the entire verification period, the system reconstructs a high-dimensional predicted probability distribution tensor, which probabilistically describes the overall stress state of the structure at any future time.

[0068] Simultaneously, the system extracts the predicted mechanical characteristic response value of the topological node corresponding to the deployment location coordinates of the independent sensor from the global node displacement field vector evolving to the end time slice T.

[0069] To obtain the actual physical observation benchmark, the system retrieves the original displacement measurement readings corresponding to the end time slice T, which were collected and stored by the displacement measuring instrument in step S1, and uses this data as the actual observation data.

[0070] The absolute deviation of the data is obtained by calculating the absolute value of the numerical difference between the predicted mechanical characteristic response value and the actual observed data, and this metric is finally defined as the mechanical response measurement residual of the model in this prediction stage.

[0071] Mechanical response measurement residual The calculation is a crucial step, and it is defined as follows: ; In the formula, This represents the measurement residual of the mechanical response, and its unit is meters. The model at the end time slice At that time, at the coordinates of the sensor's location The predicted mechanical characteristic response value at the location can be specifically taken as the expected mean parameter. These are actual observation data recorded by a displacement measuring instrument at the same location and time.

[0072] When using a kernel density estimation algorithm to reconstruct the predicted probability distribution, for the displacement of any node... Its probability density function Estimation can be performed using the Gaussian kernel function: ; In the formula, To predict displacement The probability density. The total number of predictions for the ensemble is a dimensionless integer. It is the first The displacement values ​​obtained from the sub-integration prediction. It is the bandwidth parameter of the kernel function, and its unit is the same as that of displacement. The same applies; it is used to control the smoothness of the estimate.

[0073] In this embodiment, the continuous time domain model refers to the time period from the initial moment to a specific future moment, which is the range within which the model performs dynamic extrapolation. The evolution calculation endpoint time slice is the end moment of this time period.

[0074] The global node displacement field vector is a set vector containing the three-dimensional displacement information of all topological nodes in the structure at a certain moment. The expected mean parameter describes the central tendency of the predicted displacement distribution, while the discrete variance parameter quantifies the uncertainty or dispersion of the prediction. The kernel density estimation algorithm is a statistical method for estimating the probability density function from a finite data sample. In this embodiment, a Gaussian kernel function is selected, with its bandwidth parameter... Adaptive settings are made based on the size and distribution of the dataset, such as the Silverman rule of thumb.

[0075] The prediction probability distribution tensor is a complex data structure capable of providing probabilistic predictions across the entire field and time domain. The predicted mechanical characteristic response value specifically refers to the predicted output directly comparable to measurements from physical sensors, such as the expected value of displacement prediction.

[0076] Actual observation data, i.e., the measured data from sensors, is the gold standard for evaluating the accuracy of model predictions. The mechanical response measurement residual is a direct product of this evaluation process and a key error signal driving model learning and optimization.

[0077] In one example of this embodiment, the system initiates a differentiable graph Lagrange evolution solver, extrapolating forward 1000 time steps (a total of 5 seconds) with a time step of 0.005 seconds. At the end of the evolution calculation, i.e., at the 5th second, the solver, for node 1 with deployed sensors, obtains a set of displacement prediction values ​​through 20 ensemble predictions. The system calculates the statistical properties of these 20 values, obtaining an expected mean parameter of 0.35 mm and a variance parameter of 0.01 mm squared. At this point, 0.35 mm is designated as the predicted mechanical characteristic response value.

[0078] The system retrieves the actual measurement data recorded by the displacement measuring instrument installed at node 1 at the 5th second. This data is 0.38 mm, and this is the actual observed data. Figure 5 As shown, the system then calculates the absolute deviation metric between the two data points, i.e. This 0.03mm difference is ultimately defined as the mechanical response measurement residual for this prediction step, which will be used in the subsequent backpropagation process to adjust model parameters and invert the physical state of the unknown region. Simultaneously, for all nodes, including node 1, their predicted expected mean and discrete variance parameters are fed into the kernel density estimation algorithm, assuming a bandwidth... With a value of 0.02 mm, the program will generate a smooth displacement probability density curve for each node. These curves together constitute a predictive probability distribution tensor that can show the overall deformation probability of the structure in the next 5 seconds.

[0079] In one embodiment of the present invention, the state inversion calculation module is used to perform the following steps: The network weight target partial derivative gradient flow triggered by the mechanical response measurement residual is used to extract the state back-correction amount; The reverse state correction is delivered along the network information back-tracing path defined in the structural dynamics property graph, and the reverse state correction is assigned to the non-observation nodes in the structural dynamics property graph that have not deployed sensors. Based on the state inverse correction amount, the implicit physical degradation state vector carried by the non-observation node is adjusted and corrected in reverse, and the implicit physical degradation state field is reconstructed and restored. Perform multiple concurrent projective simulation samplings on an implicit physical degenerate state field, and calculate the posterior variance generated by the multiple concurrent projective simulation samplings. Extract the contours of regions with posterior variance greater than a preset variance threshold to generate a model uncertainty map.

[0080] Specifically, the framework's underlying automatic differentiation engine performs backpropagation calculations based on the mechanical response measurement residuals. This engine automatically calculates the partial derivatives, or gradients, of the scalar loss function—the mechanical response measurement residuals—with respect to all variables involved in the forward computation along the reverse path of the forward computation graph.

[0081] The system then analyzes the obtained total gradient flow and actively separates it into two parts. One part is the partial derivative gradient flow of the network weights inside the differentiable graph Lagrange evolution solver, which is used for conventional model parameter training and optimization. The other part is the partial derivative of the initial implicit physical degradation state vector of each topological node and topological edge on the structural dynamics property graph, from which the state back-correction amount is extracted.

[0082] The state backpropagation correction is delivered from the sensor node that generated the residual to other nodes in the entire graph, following the information tracing path naturally formed during the backpropagation of the structural dynamics property graph. This correction is specifically injected into non-observation nodes that do not have independent sensors deployed. Based on the principle of gradient descent, the system uses this correction to make small-scale backpropagation adjustments and corrections to the original implicit physical degradation state vectors of these non-observation nodes.

[0083] This correction process is executed iteratively until the mechanical response measurement residual converges to below the preset tolerance threshold. Finally, the implicit physical degradation state vectors of all nodes after global adjustment and correction together constitute the reconstructed implicit physical degradation state field that fully maps the global entity's three-dimensional spatial grid distribution.

[0084] Building upon this foundation, the system initiates a nonlinear Monte Carlo variational approximation inversion testing mechanism to quantify the confidence level in the implicit physical degenerate state field inversion results. This mechanism first defines a parameterized probability distribution, i.e., a variational distribution, to approximate the true posterior probability distribution. Then, it performs numerous concurrent projection simulations on the reconstructed implicit physical degenerate state field ontology using sampling methods such as Markov chain Monte Carlo, generating thousands of mature implicit physical degenerate state field samples.

[0085] The system performs cluster statistics on the sampling results at each node location, calculating the absolute discrete variance of these samples at that point, which is the posterior variance. Finally, the system extracts the regions among all nodes whose posterior variance values ​​are higher than a certain quantile, smooths their spatial contours, and visualizes them in three dimensions, thereby creating a map representing the uncertainty of the model in which regions the inversion results are most unreliable.

[0086] When inversely adjusting and correcting the implicit physical degradation state vector of non-observable nodes, a gradient descent update rule is used. Let the nth node be an example of this. At the next iteration, the implicit physical degradation state vector of a certain non-observable node is: Then in the first The update formula is: ; In the formula, and All are implicit physical degradation state vectors. It is a positive learning rate hyperparameter, which has the same characteristics as... The opposite dimension is used to control the step size of each update and to ensure that the dimensions on both sides of the formula are homogeneous. That is, the state-reverse correction amount, which is the residual of the mechanical response measurement. For the state vector The partial derivative gradient, whose dimensions are... Consistent with the dimension of displacement. This formula ensures that the update of the state vector is performed in the direction that reduces the measurement residual.

[0087] When calculating the posterior variance, for any dimension of the implicit physical degradation state vector at any position in the graph... Its variance pass Samples obtained from Monte Carlo sampling Make an estimate: ; In the formula, Degenerate state dimension The posterior variance. It represents the total number of samples, which is a large positive integer. It is the first The value of this dimension obtained from the second sampling. Is this it? The arithmetic mean of a sample. This formula is a standard statistical estimator of variance, and its result has dimensions of _____. The square of the dimension can effectively measure the degree of uncertainty of the model inverting the state.

[0088] It should be noted that the automatic differentiation engine is a core function built into deep learning frameworks such as PyTorch or TensorFlow, which can automatically calculate the derivatives of complex functions according to the chain rule.

[0089] The state-backward correction is key to achieving inverse physics inversion in this method, as it transforms macroscopic observation error signals into corrections to microscopic physical states. Non-observational nodes refer to those nodes in the structural dynamics property diagram that do not have direct sensor data as monitoring signals.

[0090] The implicit physical degradation state field is a three-dimensional data field that provides an estimate of the potential degradation degree of the material at every point inside the building structure. The nonlinear Monte Carlo variational approximation inversion test mechanism is an advanced Bayesian inversion technique used to find easily approximated distributions when the model is complex and the posterior distribution is difficult to calculate directly, thereby quantifying the uncertainty of the parameters.

[0091] Posterior variance is a direct measure of this uncertainty; the larger the variance, the more ambiguous the model's understanding of the state. A model uncertainty map is a visualization that overlays this level of ambiguity onto the 3D model of the building in the form of a heatmap, visually indicating the model's blind spots.

[0092] In one example of this embodiment, based on the previously calculated mechanical response measurement residual of 0.03 mm, the system initiates the automatic differential calculation engine for backpropagation. The engine calculates the implicit physical degradation state vector of this residual for node 2, where no sensors are deployed. The gradient, i.e., the reverse state correction, is assumed to be a vector. The system's set learning rate for The implicit physical degradation state vector of node 2 before iteration is: The system performs an update, and the new state vector is... Please see. Figure 6 This process is repeated until the residual is less than a threshold, ultimately yielding the reconstructed implicit physical degradation state field. Subsequently, the system calculates the mean of the updated state vector for node 2. To activate the variational inversion mechanism, such as Figure 7 As shown, and proceed The system performed a second sampling. For the second dimension, the microcrack density parameter, the arithmetic mean of the 5000 sample values ​​was 0.088, and the sum of squared deviations divided by 4999 yielded 0.0025. This value of 0.0025 is the posterior variance of the microcrack density at node 2. Similarly, the system calculated the corresponding variance for node 3, which is far from the sensor, to be 0.05. When generating the model uncertainty map, the system renders the region containing node 2 with a cool color scheme representing low uncertainty, such as blue, while rendering the region containing node 3 with a warm color scheme representing high uncertainty, such as red, thus clearly indicating that the model's inversion reliability for the state of node 3 is much lower than that for node 2.

[0093] In one embodiment of the present invention, the dual-objective cost calculation module is used to perform the following steps: Extract the structural failure probability early warning index from the predicted probability distribution tensor and calculate the total variance probability integral of the region covered by the model uncertainty map. The structural failure probability early warning index is combined with the total variance probability integral to form multidimensional objective optimization constraint data; Using the total variance probability integral as the penalty benchmark when no exploration is performed, the penalty cost for information loss in the corresponding risk blocking intervention action set is quantitatively calculated. The structural failure probability warning index is used as the penalty benchmark when reinforcement is not implemented, and the structural risk penalty cost corresponding to the active exploration action set is quantitatively calculated. The basic engineering costs of each decision-making action are combined with their corresponding unresolved penalty costs to generate a dual-objective cost benchmark.

[0094] Specifically, the system first performs independent analysis on the predicted probability distribution tensor, extracts the predicted displacement distribution of key nodes related to the ultimate limit state of the structural bearing capacity, and calculates and obtains the structural failure probability early warning index by performing probability integration on the tail of the distribution.

[0095] The system reads the model uncertainty map, which displays the global posterior variance in the form of a 3D heatmap. The system sums or integrates the variance values ​​of all spatial grid points covered by the map to obtain a scalar value characterizing the overall prediction uncertainty level of the model, namely the total variance probability integral.

[0096] Subsequently, the system combines the structural failure probability early warning index with the total variance probability integral value to form multidimensional objective optimization constraint data that includes both structural risk and prediction uncertainty.

[0097] At the decision-making level, the system constructs a hybrid decision space table for surveying and reinforcement, which is a predefined, structured database containing all available intervention methods. This database is internally divided into two sets of actions: The collection of proactive exploration actions, each of which is packaged together with specific architectural space reconnaissance and survey location coordinates and designated signal detection and sensing types, such as ground penetrating radar or hyperspectral cameras.

[0098] The risk prevention and intervention action set consists of a solid structural engineering reinforcement point and a specific material injection reinforcement intervention timing, such as grouting within 24 hours after the predicted risk exceeds the threshold.

[0099] The core principle for calculating the system's execution cost benchmark is: the basic engineering cost of the assessed action plus its unresolved penalty cost. For the set of risk-blocking intervention actions, the system combines its basic engineering cost with the information gap penalty cost derived from the total variance probability integral, because direct reinforcement fails to eliminate the system's lack of internal real-state data collection; for the set of proactive exploration actions, the system combines its basic survey cost with the structural risk penalty cost derived from the structural failure probability early warning index, because exploration alone fails to directly eliminate the risk of physical collapse of the structure.

[0100] Finally, the system combines the above dimensions to generate a dual-objective cost benchmark to guide the subsequent optimization process.

[0101] In this embodiment, the structural failure probability early warning index The calculation formula is as follows: ; And the core calculation rules of the dual-objective cost benchmark, for a given decision action The cost Defined as a two-dimensional vector: ; in, ; ; In the formula, It is an action The combined cost vector is used to independently evaluate the cost of both dimensions. It is the cost of assessing the action in terms of mechanical risk. It is the cost of evaluating an action in the context of missing information. Representative decision-making actions, It is a collection of risk prevention and intervention actions. It is a collection of proactive exploration actions. It is to perform a specific action The economic cost of the required basic engineering or surveying. It is the structural risk penalty weighting coefficient, which has the dimension conversion property of transforming dimensionless probability into economic cost. This refers to the aforementioned failure probability. It is the information missing penalty weight coefficient, which has the dimensional transformation property of converting physical variance into economic cost; It is the total integral of the variance probability.

[0102] This set of cost formulas ensures that the system makes a scientific trade-off between the two interdependent goals of reducing physical risk and reducing prediction uncertainty.

[0103] Furthermore, the structural failure probability early warning index is a quantitative indicator that measures the likelihood of a structure collapsing or losing its basic functionality within a certain future timeframe. The total variance probability integral is an overall measure of the model's uncertainty regarding its predictive ability; a larger value indicates higher uncertainty in the model's predictions.

[0104] The survey and hardening dual-domain hybrid decision space table is a structured database of all potential operations, enabling the algorithm to traverse and evaluate them. The goal of the active exploration action set is to reduce model uncertainty by collecting new data through increased physical exploration.

[0105] The goal of the risk mitigation intervention package is to directly physically reinforce the structure to enhance its safety margin and reduce the probability of failure. The information gap penalty represents the expected long-term potential loss due to the uncertainty left by the failure to conduct exploration.

[0106] The structural risk penalty represents the physical collapse risk that must be borne for merely conducting surveys without substantial structural reinforcement. The dual-objective cost benchmark is not a single numerical value, but rather an evaluation system encompassing two independent dimensions: risk and information gaps, used to measure the overall merits of any decision sequence.

[0107] In one example of this embodiment, the system first parses the predicted probability distribution tensor. Structural safety codes define a critical threshold for displacement failure. The displacement prediction for node 3, which has a mean of 6mm and the highest risk, follows a normal distribution with a mean of 4mm and a standard deviation of 0.69mm. The system calculates the probability... The result is approximately 0.002, indicating an extremely low risk of physical collapse; this is the structural failure probability early warning indicator. Simultaneously, the system reads the model uncertainty map generated by S4, sums the posterior variances of all nodes, and obtains the total variance probability integral. This indicates a high degree of uncertainty in the current model's prediction results. Next, the system queries the dual-domain hybrid decision space table for surveying and reinforcement, which contains actions... That is, the basic exploration cost of using ground-penetrating radar to conduct surveys at node 3. For 5000; and actions This refers to "integral grouting reinforcement of the mortar joints connecting nodes 2 and 3," and its foundation engineering cost... The value is 15000. The system uniformly sets the risk penalty weighting coefficient. Weighting coefficient for information missing penalty All are 1,000,000.

[0108] For surveying operations The cost vector is calculated as follows: Due to the lack of physical reinforcement, the unresolved risk cost must be borne. The cost of information deficiency is its own basic exploration cost; Therefore, the cost vector is .

[0109] For reinforcement actions The risk cost is its own basic engineering cost. Because targeted investigations were not conducted, the company must bear the cost of unresolved information gaps. Therefore, the cost vector is .like Figure 8 The quantitative comparison shown shows that these two cost vectors constitute a dual-objective cost benchmark, which intuitively reflects that the overall cost of blindly carrying out expensive reinforcement is far higher than the cost of retaining minor physical risks and conducting targeted surveys.

[0110] In one embodiment of the present invention, the feedback decision execution module is used to perform the following steps: Within a pre-defined hybrid decision space, a decision sequence is generated based on a random walk algorithm. Then, a genetic algorithm is invoked to calculate the overall operating cost of the decision sequence based on a dual-objective cost benchmark. Remove the sequence terms that do not satisfy the cost constraint to obtain the safe intervention decision sequence that satisfies the overall cost minimization constraint; Extract the preferred action from the security intervention decision sequence and determine the business type of the preferred action; If the preferred action belongs to the active exploration action set, package and assemble the three-dimensional spatial coordinate positioning data group, and send the three-dimensional spatial coordinate positioning data group to trigger the execution of the exploration command; If the first action to be executed belongs to the risk prevention and intervention action set, output reinforcement instructions for the damaged areas of the masonry.

[0111] First, a random walk algorithm is initiated within the dual-domain hybrid decision space table for surveying and reinforcement. This algorithm explores possible combinations of intervention strategies by generating a series of random decision sequences.

[0112] For each generated decision sequence, the system calls a genetic algorithm for optimization. In the specific implementation of this genetic algorithm, the chromosome encoding scheme is first executed: each topological node and edge of the existing masonry building is mapped to a gene bit, using an integer encoding method. A gene bit value of 0 represents no action, 1 represents an active exploration and detection action, and 2 represents a risk blocking and reinforcement action, thus forming a chromosome representing the global decision sequence.

[0113] Secondly, the system defines a fitness function based on the aforementioned calculated dual-objective cost benchmark. Let the total risk penalty cost corresponding to a certain chromosome be... The total penalty for missing information is Then its fitness value is calculated as ,in To balance the dynamic weighting coefficients of the two objectives.

[0114] During the iterative evolution process, the genetic algorithm selects high-fitness individuals using roulette wheel selection, simulates combinations using a single-point crossover operator, and applies a random reset operator at a preset mutation rate to change some gene loci to escape local optima. Through multiple generations of iteration, the decision sequence population is continuously optimized, and those cost disadvantages that are dominated by other sequences in both the risk cost and information loss cost dimensions are gradually screened and calculated to eliminate.

[0115] After multiple generations of iterative evolution until convergence, sequence terms that do not meet the cost constraints are removed, resulting in a safe intervention decision sequence that meets the overall cost minimization constraint.

[0116] The system analyzes the security intervention decision sequence and extracts the preferred action based on a preset comprehensive preference strategy, such as selecting the two-dimensional optimization strategy with the lowest comprehensive penalty cost. The system immediately determines the business type of the preferred action.

[0117] If the determination result indicates that the preferred action belongs to the active exploration action set, the system automatically packages and assembles the corresponding three-dimensional spatial coordinate positioning data group and sends the data group to the external UAV detection equipment through the system interface, thereby triggering and issuing a fixed-point radar survey command.

[0118] Conversely, if the determination result is that the preferred action belongs to the risk blocking intervention action set, the system will execute the coding parsing step and output a grouting reinforcement instruction for the mortar joints of the masonry defect area.

[0119] After any of the aforementioned external actions are executed, the system will wirelessly collect supplementary newly measured feature data transmitted back by the aforementioned external UAV detection equipment, or structural response change data monitored by sensors after the reinforcement is completed. The system assimilates and transforms this new data, fusing it into the initial state feature set to update the attributes of the corresponding nodes or edges on the graph, thereby driving and maintaining the autonomous rolling self-healing evolutionary feedback update operation mechanism of the differentiable graph Lagrange evolution solver.

[0120] It should be noted that the random walk algorithm is used for global search in complex decision spaces. The genetic algorithm, on the other hand, is a classic evolutionary algorithm for solving multi-objective optimization problems. It does not seek a single optimal solution, but rather a set of Pareto optimal solutions that make different trade-offs among different objectives.

[0121] A safety intervention decision sequence is a collection of decision options, where improvement of any one option will inevitably lead to the deterioration of at least one other objective. The preferred action is the single action selected from the safety intervention decision sequence that should be executed most immediately. A three-dimensional spatial coordinate positioning data set is a data package containing information such as latitude, longitude, elevation, and the orientation of the detection equipment. A grouting reinforcement instruction is a work instruction sent to the construction unit, containing specific reinforcement locations, material specifications, and process procedures.

[0122] Newly added measured feature data refers to new information obtained through active detection, such as the size and location data of internal cavities obtained after analyzing ground-penetrating radar echo maps.

[0123] The feedback-update operating mechanism refers to the working mode in which this system continuously absorbs new information and corrects its own model through a cycle of prediction-decision-action-perception-re-prediction, thereby achieving continuous learning and adaptive optimization.

[0124] In one example of this embodiment, the system includes surveying actions. and reinforcement actions To find the best option within the decision space. For example... Figure 9 The optimization space distribution shown, after computation by the genetic algorithm, generates a safety intervention decision sequence that provides a two-dimensional comparison of solutions: Solution 1 is to perform exploration alone. The total evaluation cost is Solution 2 is to carry out reinforcement alone. The total evaluation cost is The system employs a two-dimensional overall cost optimization strategy. Since Solution 1 is superior to Solution 2 in both the risk and information gap cost dimensions, the system selects Solution 1 and executes the survey action. The system selects this as the primary action. It then determines that this action belongs to the active exploration action set. Based on this, the system packages and assembles the three-dimensional spatial coordinates of node 3's location, for example (X: 34.5 m, Y: 12.8 m, Z: 8.2 m), as a three-dimensional spatial coordinate positioning data set, and sends it to a standby UAV detection device via a wireless communication interface. After receiving the command, the UAV flies to the designated location and performs a ground-penetrating radar scan of the node 3 area. After the scan is complete, the UAV sends the returned radar data map back to the system. The system analyzes the map and discovers a [missing information - likely a location or feature] within node 3. The void is thus added as measured feature data. The system then performs assimilation transformation, using this void information as a physical constraint to update the internal damage parameters of node 3 in the initial state feature set. The system then uses the newly acquired information to begin a new round of dynamic evolution.

[0125] Each of the modules can be implemented in whole or in part through software, hardware, or a combination thereof. It supports hardware embedded in or independent of the processor in the computer device, and also supports software stored in the memory of the computer device, so that the processor can call and execute the operations corresponding to each of the above modules.

[0126] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A smart assessment and protection decision-making system for existing masonry building defects, characterized in that, include: The initial feature generation module is used to acquire the sensor monitoring signals of the surface of the existing stone masonry building, establish a structural dynamic property map based on the sensor monitoring signals, and generate an initial state feature set. The state evolution solution module is used to input the initial state feature set into a preset evolution solver, perform bidirectional coupled evolution calculation of mechanical state and implicit degenerate state, and generate state transition matrix; The prediction residual calculation module is used to perform forward fitting evolution by traversing the state transition matrix along the continuous time domain, extract global node evolution prediction data, and calculate the mechanical response measurement residual based on the global node evolution prediction data and the actual observation data. The state inversion calculation module is used to perform backpropagation calculations based on the mechanical response measurement residuals, invert the implicit physical degradation state field of the unobserved region, and generate a model uncertainty map; The dual-objective cost calculation module is used to integrate global node evolution prediction data and model uncertainty map, and combine them with a preset hybrid decision space table to calculate the dual-objective cost benchmark. The feedback decision execution module is used to select security intervention decision sequences based on a dual-objective cost benchmark, distribute detection and hardening instructions according to the security intervention decision sequences, and obtain perception and monitoring data to update the initial state feature set.

2. The intelligent assessment and protection decision-making system for existing stone masonry buildings according to claim 1, characterized in that, The initial feature generation module is used to perform the following steps: Acquire multi-source monitoring signals output by each independent sensor, extract the temporal micro-vibration components and displacement change rate characteristics of the multi-source monitoring signals, and combine them to generate a basic observation state set; The stone bearing units of the existing stone masonry building are set as topological nodes, and the mortar joint material interfaces between the stone bearing units are defined as topological edges. The structural dynamic property diagram is constructed based on the spatial arrangement and combination of topological nodes and topological edges. Extract the motion state parameters and mass inertia properties of the topological nodes, and determine the normal contact stiffness and tangential structural friction coefficient of the topological edges; An implicit physical degradation state vector is constructed for topological nodes and edges. The implicit physical degradation state vector is then connected with the basic observation state set, and the data dimensions are concatenated to generate an initial state feature set.

3. The intelligent assessment and protection decision-making system for existing stone masonry buildings according to claim 1, characterized in that, The state evolution solution module is used to perform the following steps: Receive the initial state feature set, extract the integration time step of the evolution solver, construct the discretized force balance transfer equation based on the integration time step, and adapt the discretized force balance transfer equation to the structural dynamics property map. Extract the nodal stress state data contained in the initial state feature set, and calculate the material degradation increment corresponding to the implicit physical degradation state vector; The implicit physical degradation state vector is updated by superimposing the numerical increment of material degradation and the historical degradation baseline of the system. The three-dimensional distribution of the force field response in the next time slice is calculated by subtracting the normal contact stiffness value of the topological edge from the updated implicit physical degradation state vector. The three-dimensional distribution law of the combined force field response and the numerical increment of material degradation are merged to generate a state transition matrix.

4. The intelligent assessment and protection decision-making system for existing masonry building defects according to claim 1, characterized in that, The prediction residual calculation module is used to perform the following steps: Based on the discretized force balance transfer equation, the state transition matrix is ​​derived, and continuous integration calculation steps are performed along the continuous time domain to obtain the expected mean parameter and discrete variance parameter of the global nodal displacement field vector at the end time slice of the evolution calculation. The expected mean parameter and discrete variance parameter are smoothed, and the predicted probability distribution tensor is reconstructed and used as the global node evolution prediction data. Extract the predicted mechanical characteristic response value corresponding to the observation position coordinates from the global nodal displacement field vector, and call the corresponding original displacement reading as the actual observation data; The absolute deviation of the predicted mechanical characteristic response value from the actual observed data is calculated, and the absolute deviation is defined as the mechanical response measurement residual.

5. The intelligent assessment and protection decision-making system for existing masonry building defects according to claim 1, characterized in that, The state inversion calculation module is used to perform the following steps: The network weight target partial derivative gradient flow triggered by the mechanical response measurement residual is used to extract the state back-correction amount; The reverse state correction is delivered along the network information back-tracing path defined in the structural dynamics property graph, and the reverse state correction is assigned to the non-observation nodes in the structural dynamics property graph that have not deployed sensors. Based on the state inverse correction amount, the implicit physical degradation state vector carried by the non-observation node is adjusted and corrected in reverse, and the implicit physical degradation state field is reconstructed and restored. Perform multiple concurrent projective simulation samplings on an implicit physical degenerate state field, and calculate the posterior variance generated by the multiple concurrent projective simulation samplings. Extract the contours of regions with posterior variance greater than a preset variance threshold to generate a model uncertainty map.

6. The intelligent assessment and protection decision-making system for existing stone masonry buildings according to claim 1, characterized in that, The dual-objective cost calculation module is used to perform the following steps: Extract the structural failure probability early warning index from the predicted probability distribution tensor and calculate the total variance probability integral of the region covered by the model uncertainty map. The structural failure probability early warning index is combined with the total variance probability integral to form multidimensional objective optimization constraint data; Using the total variance probability integral as the penalty benchmark when no exploration is performed, the penalty cost for information loss in the corresponding risk blocking intervention action set is quantitatively calculated. The structural failure probability warning index is used as the penalty benchmark when reinforcement is not implemented, and the structural risk penalty cost corresponding to the active exploration action set is quantitatively calculated. The basic engineering costs of each decision-making action are combined with their corresponding unresolved penalty costs to generate a dual-objective cost benchmark.

7. The intelligent assessment and protection decision-making system for existing masonry building defects according to claim 1, characterized in that, The feedback decision execution module is used to perform the following steps: Within a pre-defined hybrid decision space, a decision sequence is generated based on a random walk algorithm. Then, a genetic algorithm is invoked to calculate the overall operating cost of the decision sequence based on a dual-objective cost benchmark. Remove the sequence terms that do not satisfy the cost constraint to obtain the safe intervention decision sequence that satisfies the overall cost minimization constraint; Extract the preferred action from the security intervention decision sequence and determine the business type of the preferred action; If the preferred action belongs to the active exploration action set, package and assemble the three-dimensional spatial coordinate positioning data group, and send the three-dimensional spatial coordinate positioning data group to trigger the execution of the exploration command; If the first action to be executed belongs to the risk prevention and intervention action set, output reinforcement instructions for the damaged areas of the masonry.

8. The intelligent assessment and protection decision-making system for existing masonry building defects according to claim 2, characterized in that, The steps for extracting the temporal micro-vibration components and displacement rate characteristics of multi-source monitoring signals and combining them to generate a basic observation state set include: Acquire discrete sampling sequences of multi-source monitoring signals within a preset time window; Perform a summation of squares on the discrete sampled sequence, and divide the summation of squares by the total number of sample points to obtain the mean data; Perform a square root operation on the mean data to obtain the root mean square value of the time-series micro-vibration component; Perform a first-order difference operation on the displacement sequence in the multi-source monitoring signal to calculate the displacement increment between adjacent sampling time points and obtain the displacement change rate characteristics; The root mean square value and displacement rate of change characteristics are organized and combined according to the measurement point identifier to generate a basic observation state set.

9. The intelligent assessment and protection decision-making system for existing masonry building defects according to claim 3, characterized in that, The steps for calculating the material degradation increment corresponding to the implicit physical degradation state vector include: Extract the effective stress values ​​acting on the mortar joint interface from the nodal stress state data; Obtain the preset stress corrosion rate coefficient, multiply the effective stress value with the stress corrosion rate coefficient, and obtain the basic degradation rate. The numerical increment of the chemical corrosion factor is calculated by multiplying the base degradation rate by the integration time step. The numerical increment of the chemical corrosion factor is used as the numerical increment of material degradation corresponding to the implicit physical degradation state vector.

10. The intelligent assessment and protection decision-making system for existing masonry building defects according to claim 3, characterized in that, The steps of subtracting the normal contact stiffness value of the topological edge based on the updated implicit physical degradation state vector include: Extract the initial normal contact stiffness value of the topological edge; Traverse the updated implicit physical degradation state vector for multiple degradation dimensions that affect stiffness, and obtain the cumulative value of each degradation dimension. The difference between the value 1 and the cumulative value of each degradation dimension is calculated to obtain the retention ratio coefficient corresponding to each degradation dimension; The overall attenuation coefficient is obtained by multiplying the retention ratios corresponding to all degradation dimensions. The initial normal contact stiffness value is multiplied by the comprehensive attenuation coefficient to obtain the deducted normal contact stiffness value.

Images