Prediction method for failure diffusion of nodes of grid steel structure

Abstract

The invention relates to a grid steel structure node failure diffusion prediction method, which comprises the following steps of: forming a failure inducement and evolution path model according to a connection form of a bolt-sphere node and a welding node, a stress transfer path, node local yield and fatigue accumulation factors; a diffusion logic of node failure-force flow redistribution-adjacent node stress concentration-chain reaction is constructed; nodes are regarded as middle points of the graph, rod pieces are regarded as edges, and propagation weights between the nodes are defined; the weight parameters are dynamically assigned based on material attributes, node connection types, spatial geometric positions and historical stress states; multi-source monitoring sensors including strain gauges, displacement meters, thermometers and accelerometers are arranged at key nodes of the grid structure, so that node states are sensed; in combination with real-time monitoring data, extracting strain rate change, vibration characteristics, frequency drift and residual deformation key characteristics of each node; a node health state feature vector set is constructed by using time sequence analysis, and the gradual degradation trend of nodes is reflected.

Description

Technical Field

[0001] The present invention relates to a method for predicting the failure diffusion of grid steel structure joints. Background Art

[0002] Currently, the prediction of grid steel structure joint failures mainly relies on mechanical models based on finite element analysis, structural reliability assessment theories, empirical rule judgments, vibration mode analysis, and some data-driven algorithms that integrate structural health monitoring systems (SHM). Although the identification and safety assessment of local damages in grid structures have been achieved to a certain extent, there are still many deficiencies and obvious drawbacks in practical engineering applications, making it difficult to meet the requirements of modern large-scale space steel structures for accurate prediction, active control, and intelligent response of failure diffusion under high complexity and high dynamic change conditions. Specifically, it is reflected in the following aspects:

[0003] First, at the structural modeling level, traditional methods often simplify the space grid structure into a discrete element system with finite dimensions and rely on static mechanical analysis or calculation results under a single load combination to determine whether a certain node has entered the failure state. Such methods overly depend on the modeler's understanding of the structural force mode and cannot effectively consider the coupling relationship between nodes and the force flow redistribution mechanism under complex working conditions. Especially under the action of nonlinear multi-source loads such as earthquakes and wind vibrations, the structural response has obvious dynamics and uncertainty. The traditional model cannot dynamically update the state evolution process of each node in real time, let alone judge whether local failure will affect other nodes through path transmission. Second, in terms of the identification of the diffusion path, traditional methods mainly describe the failure conduction path through "hypothetical" or pre-set "propagation path libraries", lacking accurate modeling of the complex connection relationship between nodes in the actual structure and failing to introduce concepts such as propagation intensity, propagation speed, and path weight. As a result, when multiple nodes deteriorate simultaneously or complex boundary conditions change, the system cannot identify irregular diffusion behaviors such as the possible transfer, turning, and delay of failure, severely restricting the accuracy and engineering practicability of diffusion prediction. Third, in the time dimension, most existing methods rely on the "state value discrimination" mechanism, that is, when a certain sensor monitors that the strain or displacement of a certain node exceeds the pre-set threshold, it is judged that the node has failed or is about to fail. This method ignores the historical evolution process of the node state and also lacks the ability to judge the state trend, easily leading to the phenomena of "discovering failure only after it has occurred" or "misjudging short-term anomalies", especially when the node is in the slow deterioration stage, the traditional method is almost powerless and cannot achieve "early perception" and "active prediction". In addition, in terms of data processing and fusion, most current systems adopt a single-source data-driven mode, such as only collecting strain or displacement values, ignoring the prediction value brought by the coupling effect of multiple physical quantities. For example, information such as the vibration characteristics of nodes, frequency drift trends, and residual deformation rates is often ignored, and a comprehensive node state representation cannot be constructed, resulting in insufficient input information dimension, limited prediction model training ability, poor model generalization ability, and prone to misjudgment and missed judgment.

[0004] At the same time, in most current studies, there is a lack of description of the continuous evolution of the structural state space, and only "whether a certain node is abnormal currently" can be output, but it is impossible to simulate "whether this node will spread risks to surrounding nodes in a future period", that is, there is a lack of the ability to model the failure diffusion trend, restricting its role in the dynamic management of structural safety. Furthermore, most existing prediction systems adopt a static modeling method. Once the model is trained, it remains fixed and lacks the ability to update online, unable to adapt to the dynamic characteristics of the structure such as material property changes, connection loosening, and load fluctuations during long-term service, resulting in a continuous decline in prediction accuracy and continuous accumulation of errors over time, ultimately causing the system to fail. Summary of the Invention

[0005] The object of the present invention is to provide a prediction method for the failure diffusion of grid steel structure joints, so as to solve some of the drawbacks and deficiencies pointed out in the background art.

[0006] The following technical solutions adopted by the present invention to solve its above technical problems include the following steps:

[0007] S1. Establish a node failure evolution mechanism model:

[0008] S1.1. Based on the connection forms including bolt sphere joints and welded joints, stress transfer paths, local yield of nodes, and fatigue accumulation factors, form a failure inducement and evolution path model;

[0009] S1.2. Construct a diffusion logic of node failure → force flow redistribution → stress concentration in adjacent nodes → chain reaction;

[0010] S2. Regard the nodes as points in the graph and the members as edges, and define the propagation weights between the nodes; the weight parameters are dynamically assigned based on material properties, node connection types, spatial geometric positions, and historical stress states;

[0011] S3. Adopt a multi-dimensional structure monitoring system and construct nodes:

[0012] S3.1. Install multi-source monitoring sensors at the key nodes of the grid structure, including strain gauges, displacement gauges, thermometers, and accelerometers, to sense the node states; combine the real-time monitoring data to extract the key features of the strain rate change, vibration characteristics, frequency drift, and residual deformation of each node;

[0013] S3.2. Use time series analysis to construct a node health state feature vector set to reflect the gradual degradation trend of the nodes;

[0014] S4. Use the graph neural network GNN combined with the time series model to perform dynamic prediction of node failure in the time and space domains; the outputs of the series model include: ①. The failure risk level of a single node, ②. The range of potentially affected nodes, ③. The diffusion time window;

[0015] S5. Use the incremental learning algorithm to input new monitoring data and dynamically correct the model parameters; make the model evolve with the evolution of the structure;

[0016] S6. Display the prediction results through a three-dimensional digital twin visualization platform: represent the node failure risk level with different colors; dynamically demonstrate the predicted failure diffusion path and conduction time; provide functions for structural reinforcement, node replacement, and risk warning decision support.

[0017] Furthermore, the method of deploying a multi-dimensional structure monitoring system and constructing nodes:

[0018] A multi-source sensing system is deployed at the key structural nodes. The types of sensors include strain gauges, displacement gauges, accelerometers, and thermometers, which are used to sense stress responses, geometric deformations, dynamic disturbances, and environmental disturbances respectively. The collected data is synchronized and normalized in the spatio-temporal dimension. This is achieved by constructing a unified multi-sensor mapping function, which is defined as follows:

[0019]

[0020] Where:

[0021] Ψ(i,t) represents the total response projection value of the i-th sensor within time t; α i (t) is the data stability factor of the i-th sensor at time t, which is used to dynamically adjust the credible weight of the sampling signal of this sensor; Ω i (x,y,z) is its spatial deployment function on the three-dimensional spatial coordinates (x,y,z) of the structure; Δ i (t) is the signal relative delay function of the i-th sensor at time t, which reflects the error offset of different types of signals during the synchronization process; γ i is the delay compensation sensitivity factor, which controls the contribution intensity of the delay to the fusion result; T represents the selected time integration window.

[0022] Furthermore, the method for deploying the multi-dimensional structure monitoring system and constructing nodes:

[0023] Feature extraction is carried out for the evolutionary behavior of the nodes; taking the feature evolution rate as the core monitoring index, by analyzing the change of strain rate, acceleration spectrum offset, and displacement residual feature evolution trend in the time domain, state labels with predictive value are extracted. The feature enhancement processing formula is as follows:

[0024]

[0025] Where:

[0026] Θ k is the dynamic response intensity of the k-th type of feature, including strain rate, vibration amplitude, and frequency drift; φ j (t) represents the signal fluctuation function of the j-th node monitoring channel at time t, which reflects the real-time activity degree or noise change trend of the sensing channel; is the normalized projection factor of the k-th type of feature on the j-th channel; δ j is the instantaneous response deviation between the monitoring value of the j-th channel and the reference healthy value; represents the rate of change of the deviation with respect to time, which is used to measure the deterioration speed of this node under a specific physical quantity.

[0027] Furthermore, the method for deploying the multi-dimensional structure monitoring system and constructing nodes:

[0028] Project the states of all nodes into the three-dimensional structural space to form a continuously calculated state evolution field for judging the distribution and development direction of the failure risks of each node inside the structure; adopt a structural risk potential field model based on time evolution and node interaction to continuously deduce the diffusion trend; the potential field model is defined as follows:

[0029]

[0030] Wherein:

[0031] Σ(x, y, z, t) represents the risk potential value of the structure at the three-dimensional coordinate position (x, y, z) and time t; M is the total number of key nodes monitored in the structure; β m is the risk excitation coefficient of the m-th node; Θ m is the fusion feature strength of this node; is the gradient of the sensing response field of the m-th node in the three-dimensional structure, indicating the severity of the signal change sensed by the node in space; κ m is the diffusion impedance coefficient, which is used to measure the sensitivity of the risk of this node to spread to the neighborhood; then expresses the diffusion attenuation or growth trend of the structural risk over time.

[0032] Furthermore, the dynamic prediction method in the time and space dual domain includes:

[0033] Abstract the grid structure into a graph model, with each node in the structure as a vertex in the graph and the connecting members as edges, construct a graph topology structure including geometric position information, force flow transmission direction, connection stiffness and historical damage response, and assign weight coefficients representing physical correlation degree and structural influence to the edges and nodes respectively;

[0034] Learn the information interaction mode between the nodes and their neighborhoods in this graph through a graph neural network, and establish a spatial coupling update mechanism for describing the propagation influence of local failure states within the graph structure. Among them, the node feature propagation rule is described by the structure response propagation function Γ ij (t), and is expressed as:

[0035] Γ ij (t) = ∫0 τ [ξ ij ·(σ i (t) - σ j (t)) 2 + μ i ·χ ij ·ρ j (t)]dt

[0036] Wherein:

[0037] Γ ij(t) represents the cumulative value of the response influence from node i to adjacent node j within time t; ξ ij is the physical coupling coefficient related to the structural connection strength; σ i (t) and σ j (t) respectively represent the state response values of node i and node j at time t, used to reflect the change in the health state of the node; μ i is the influence enhancement coefficient of node i; χ ij is the propagation channel weight of the edge; ρ j (t) is the perturbation sensitivity function of node j; τ is the prediction step time window, controlling the time limit of the integration interval.

[0038] Furthermore, the dynamic prediction method in the time and space dual domain includes:

[0039] Taking the state sequence of each node within the historical monitoring period as input, using a time series model to dynamically model the evolution trend. The model input includes key parameters such as the strain rate change, frequency drift, and micro-vibration amplitude of the node, and introducing the sequence feature transformation function Φ k (t) for extracting the timing law, and the function is defined as follows:

[0040]

[0041] Where:

[0042] Φ k (t) represents the global change trend of the k-th type of feature at time t; ∈ m (t) is the feature response intensity function of the m-th node channel at time t; Δ m (t) is the offset response of this channel at time t, used to measure the degree of deviation of the node state from the healthy benchmark; represents the change rate of the state response; is the weighted normalization coefficient of the k-th type of feature in the m-th channel; lim_{n→∞} represents the global reduction limit of the statistical process under the multi-node network.

[0043] Furthermore, the dynamic prediction method in the time and space dual domain includes:

[0044] Through the linkage prediction mechanism of the time series model and the graph structure, the dual-domain joint output of node failure includes: the failure risk level of each node, output with a continuous scoring function Λ i to reflect the failure probability or risk level of the node in a future time period; the set of potentially affected nodes, extracting the range of nodes affected by the current high-risk nodes in the future through the joint modeling of the topological diffusion path and the propagation influence graph; the risk diffusion time window, using the state propagation delay function Υ ijCalculate the time required for risk propagation between computing nodes. The function is defined as follows:

[0045]

[0046] Where:

[0047] Υ ij represents the total time delay required for risk propagation from node i to node j; Γ ij (t) is the structural propagation response function between nodes; Φ i (t) and Φ j (t) are the temporal response evolution intensities of nodes i and j, used to adjust the risk propagation speed; η ij is the stability weight factor on the risk transfer path; δ is the predicted time interval for risk impact propagation, and [t0, t0+δ] indicates that the calculation occurs within this time period.

[0048] The prediction method for node failure diffusion of grid steel structures proposed by the present invention has significant engineering applicability, prediction foresight, and intelligent evolution ability. By integrating graph neural networks, time series modeling, multi-source structural monitoring, and risk potential field theory, it effectively solves the core technical bottlenecks in traditional methods, such as lagging node failure identification, unclear diffusion paths, and lack of time prediction ability. The specific beneficial effects include:

[0049] Model the grid structure nodes as graph structure units, combine the force flow relationship between nodes and the historical damage evolution law, and dynamically learn the structural topology characteristics and risk propagation paths through graph neural networks, significantly improving the prediction accuracy of the evolution trend from local damage to global failure, avoiding the misguidance of overall risk assessment by "point judgment", and achieving early warning.

[0050] Introduce time series modeling methods (such as LSTM / Transformer) to dynamically track state parameters such as node strain rate, frequency drift, and micro-vibration, and couple them with the spatial topology propagation model, realizing for the first time the joint reasoning ability of "where the structure risk occurs", "when it diffuses", and "where it will evolve", providing a new path for dynamic perception of structural safety. Description of the Drawings

[0051] Figure 1 It is a flowchart of the prediction method for node failure diffusion of the grid steel structure of the present invention.

[0052] Figure 2 It is a flowchart of the method for arranging a multi-dimensional structural monitoring system and constructing nodes of the present invention.

[0053] Figure 3 It is a flowchart of the dynamic prediction method in the time and space double domains of the present invention. Detailed Embodiments

[0054] The following will make a detailed description of the specific implementation manners of the present invention in conjunction with the accompanying drawings.

[0055] In the prediction method for the failure diffusion of the grid steel structure joints, step S1 is to establish a joint failure evolution mechanism model, the core of which lies in deeply depicting the physical and mechanical evolution path of the joint from local damage to overall diffusion. Among them, in S1.1, by classifying and modeling the typical joint forms in the grid structure, the response mechanisms of the main connection methods including bolt sphere joints and welded joints under different load conditions are clarified, and the microscopic damage behaviors such as local yield, plastic deformation, and fatigue crack accumulation that may occur during the stress-bearing process of these joint types are specifically analyzed. These factors together constitute the failure inducement library. By establishing the corresponding evolution path model, the physical damage process of the joint is expressed as a dynamic process from initial micro-damage to loss of function. Further, in S1.2, the causal logic path of failure diffusion is constructed, that is, when a local failure occurs in a certain joint, the internal force flow in the structure will redistribute, and the internal force originally transmitted through this joint will transfer to its adjacent joints, thus causing stress concentration phenomena in the surrounding joints. Especially when these adjacent joints are already in a high-stress or low-redundancy state, secondary failures are extremely likely to form in a short time. As the stress continuously transfers and accumulates, the structure will enter the evolution chain of failure → force flow reconstruction → stress concentration → chain failure, forming a typical diffusion-type failure behavior. This diffusion mechanism model not only reveals the internal process of the grid structure developing from point damage to surface or regional damage under complex loads, but also provides the evolution basis and propagation path input at the structural physical level for the subsequent prediction model.

[0056] Step S2 is to regard the nodes as points in the graph and the members as edges, thereby constructing a structural topology model based on graph theory. The core of this model lies in expressing the physical coupling relationship and failure propagation path between nodes with a graph structure. In this representation, each node not only represents the connection unit in the actual structure but also serves as an information unit in the graph. All adjacent nodes are connected through structural members, which are manifested as the existence of edges in the graph. To truly reflect the failure influence ability between nodes, it is necessary to assign a propagation weight to each edge. The definition of this propagation weight is not statically set but is dynamically calculated by combining multiple structural physical factors. First, the material properties are considered, including the yield strength, elastic modulus, fatigue limit, etc. of the steel used for the members and node components. These factors determine whether the connection path has high damping or high permeability during the stress perturbation transfer process. Secondly, the node connection type is considered. For example, compared with welded joints, bolted spherical joints have significant differences in stiffness, rotation angle release ability, and energy dissipation method, which will directly affect the propagation intensity of the failure impact between adjacent nodes. Thirdly, the spatial geometric position is also an important weight basis. The edges connected to the nodes closer to the main force flow channel or the node-dense area are more likely to become the main paths for failure diffusion, so higher topological importance needs to be assigned. Finally, the historical stress state is used to describe the loading memory and stress residue of the nodes during the long-term service of the structure. If a certain node has experienced large stresses or is in a high-fatigue area, the propagation weight of its connected edges will be dynamically adjusted to a sensitive path to identify potential high-risk areas in advance.

[0057] Step S3 is to deploy a multi-dimensional structure monitoring system and construct a node state feature model. Its core objective is to achieve state perception and degradation trend quantification at the node level. Specifically, in S3.1, a multi-source monitoring sensor system is deployed for key nodes with concentrated structural forces, complex connections, or high historical damage probabilities, including four types: strain gauges, displacement gauges, thermometers, and accelerometers. Among them, strain gauges are used to monitor the strain response of nodes under load in real time, capturing their deformation rate and stress concentration trend; displacement gauges are used to record the relative displacement of nodes, reflecting the geometric offset during the growth of the overall or local deflection of the structure; thermometers are used to dynamically monitor the impact of environmental factors on material properties, such as interference factors like thermal expansion and contraction caused by temperature fluctuations; accelerometers focus on the response characteristics of nodes under structural vibration, especially in working conditions such as wind load, earthquake, or mechanical shock, where the high-frequency micro-vibration behavior of nodes can be extracted to identify stiffness changes or connection loosening phenomena. The real-time data collected by these sensors is transmitted to the data processing module through edge computing or wireless communication. After synchronization, cleaning, and standardization processing, key physical feature indicators reflecting the health state of nodes are further extracted, including strain rate change, used to judge the stress evolution speed of nodes under continuous load; vibration characteristics, used to identify possible loosening points or stiffness weakening regions in the structure; frequency drift, characterizing the change trend of the dynamic performance of nodes, usually a precursor signal of connection stiffness decline or local damage; residual deformation, used to measure whether the node still retains irreversible geometric offset after the external force is removed, which is an important indicator of plastic deformation. In S3.2, based on the above results of high-frequency and high-dimensional physical quantity collection and feature extraction, time series analysis methods are used to continuously track and model the state of each node. By constructing a set of node health state feature vectors reflecting historical evolution trends, not only the current operating state of the node is characterized, but also its potential degradation trajectory in the future time is encoded and expressed.

[0058] Step S4 is to use the graph neural network GNN combined with the time series model to realize the dynamic prediction of node failure in both time and space domains. The core idea is to integrate the topological connection relationship of the structure with the time evolution process of the node state to solve the problem that the traditional method can only make static judgments and cannot predict the failure evolution path in advance. First, GNN, i.e., graph neural network, abstracts the entire grid structure into a graph model, regards the nodes as vertices in the graph, and the rods as edges, and combines the geometric distance between nodes, force transfer relationship, historical monitoring status and weight coefficient information to establish the spatial coupling relationship within the structure. The network can simulate how the mechanical disturbance caused by a node to the surrounding nodes spreads through the topological path when a node has a failure trend, and further learn the influence transmission law between the nodes. The graph neural network can not only capture the state characteristics of the local nodes, but also propagate the state information within the entire structure, thereby forming a spatial failure map at the structural level; at the same time, the time series model part is used to model each node in history. The state change trend in the historical monitoring cycle, such as strain rate fluctuation, frequency drift rate, vibration energy change and residual deformation evolution trajectory, etc. These state time series are input into sequence prediction frameworks such as LSTM, GRU or Transformer to explore the potential laws of node state evolution, identify early degradation characteristics and predict the state trend in the future period. Through the joint training of graph neural network and time series model, the spatial propagation mechanism and time evolution trend of node state are modeled synchronously, so that the model can not only evaluate the current structural state, but also have the ability to predict potential failures in the future. Finally, the model can output three types of core prediction results: first, the failure risk level of a single node, given in the form of risk score or classification, which is used to measure the possibility of damage or functional degradation of the node in the future time window; second, the scope of potential affected nodes, that is, through the structural topology map and the state transmission path between nodes, it is judged whether the current high-risk node may cause other surrounding nodes to enter the warning or critical state in the future;

[0059] Step S5 is to input new monitoring data using the incremental learning algorithm and dynamically correct the model parameters, so as to realize the evolution of the model with the structural evolution. Its core idea is to construct an intelligent prediction system with adaptive ability to solve the defect of the traditional model of "training once and remaining unchanged". Especially in actual engineering, the service environment, load pattern, node damage state and material properties of the grid structure will change continuously over time. If the prediction model is trained based on static historical data or early samples for a long time, it will be difficult to adapt to the current real state of the structure, resulting in a decrease in prediction accuracy or even misjudgment. Therefore, the present invention introduces an incremental learning mechanism, that is, continuously collecting real-time monitoring data feedback by sensors during the service of the structure, including multi-source signals such as node strain, displacement, temperature, vibration, etc., and inputting them into the existing model in the form of mini-batch or single-point update. By fusing, adjusting or replacing with the original model parameters, the weight matrix, bias term and structural connection weights in the graph neural network and time series model are dynamically corrected to ensure that the model always runs around the current state of the structure. This process does not require re-training the entire model, avoiding computational resource consumption and time delay, and can effectively inherit the existing learning ability and quickly adapt to new data, realizing the closed-loop mechanism of "learning - updating - prediction".

[0060] Step S6 is to display the prediction results through a 3D digital twin visualization platform. Its core purpose is to present the complex model calculation results in an intuitive and interactive manner, enhancing the engineering personnel's perception ability and decision-making efficiency of the current state of the structure and the future risk evolution trend. Specifically, the system first constructs a high-precision 3D digital twin model corresponding one-to-one with the actual grid structure based on the original BIM model of the structure, node coordinate information, and spatial component topological relationship. In this model, real-time monitoring data from the sensing system, failure prediction results output by the graph neural network and time series model, and structure risk level assessment information are integrated, and the failure risk levels of different nodes are visually expressed by means of color gradients, where green represents low risk, yellow represents medium risk, and red identifies high-risk or soon-to-fail nodes, enabling users to quickly locate the weak parts of the structure without referring to data reports. At the same time, the platform supports dynamic simulation of the predicted failure diffusion path, that is, gradually presenting the propagation trajectory of the risk spreading from the starting node to the surrounding area within a set time range. The prediction window can be precisely controlled and the conduction time of the failure between different nodes can be observed by cooperating with the time axis scrolling function, so as to help identify high-risk areas and conduction links. In addition, the system integrates decision-making assistance modules such as structure reinforcement suggestions, node replacement priority ranking, and risk warning trigger mechanisms. When the prediction model identifies continuous high-risk nodes or the diffusion rate exceeds the threshold, the system will automatically trigger a risk alarm and provide corresponding intervention strategy suggestions based on the evaluation of structural mechanics parameters and maintenance costs, finally realizing a closed-loop application process from "data - prediction - display - decision" and improving the engineering safety management efficiency and intelligent level.

[0061] Example 1:

[0062] For the roof steel grid structure of a large stadium, the structure adopts a square space grid layout with a span of 68 meters and a total of 984 vertical and horizontal nodes. Among them, 36 key nodes in Area A are identified and distributed in areas with stress concentration and complex geometric connections. It is decided to deploy a multi-source sensing system on these nodes to monitor the state changes in real time and build a health feature model. The sensor configuration scheme is as follows: 4 types of sensors are deployed on each key node, including 1 strain gauge, 1 displacement gauge, 1 accelerometer, and 1 thermometer. The sensor number is represented by i, that is, the i-th sensor corresponds to a certain type of measurement point on a certain node. For example, the strain gauge of the 12th node is i = 45. The installation coordinates of all sensors are located through the BIM model and are respectively assigned coordinate points (x i , y i , z i ). For example, the installation position of sensor 45 is 21.3, 12.6, 9.5. The recorded sampling frequency per second is 50Hz, and 500 data points will be collected within each data window period of T = 10s. This monitoring data is input into a unified fusion model for processing. The fusion calculation formula is as follows:

[0063] Ψ(i, t) = ∫₀ T [α i (t)·Ω i (x, y, z) + γ i ·Δ i (t)]dt

[0064] During the calculation process, the specific parameter values are as follows: α i (t) is the data stability factor, with a range of 0.85, 1.00. It is dynamically assigned after the system evaluates the sampling error, drift, and signal-to-noise ratio of the sensor signal in real time. Here, taking α 45 (t) = 0.93 as an example, Ω i (x, y, z) is the spatial layout function, and its value can be given by empirical fitting of the contribution degree of the response based on the structural influence area where the node is located. The range is 0.5, 1.2. The value of the node closer to the main path of the force flow is larger. Let the value at sensor 45 be 0.98; Δ i (t) is the signal delay function, which represents the time error caused by the clock offset or sampling period difference of different sensors. The absolute value range is usually between 0 and 0.8 seconds. Here, taking Δ 45 (t) = 0.15 s as an example; γ i is the delay compensation sensitivity factor, which is used to control the contribution intensity of this item to the total response fusion. It is usually an empirically selected coefficient, with a value range of 0.5, 2.0. A higher value should be set for dynamic sensors (such as accelerometers). In this example, γ 45 = 1.1. Substituting into the formula, we get:

[0065] Ψ(45, t) = ∫₀ 10 [0.93×0.98 + 1.1×0.15]dt = ∫₀ 10 [0.9114 + 0.165]dt = ∫₀ 10 1.0764dt

[0066] = 10.764

[0067] That is, within the data period of T = 10 s, the total response projection value of the 45th sensor is 10.764. This value is used as the unified input index after the multi-dimensional data fusion processing of the node state and is used for the subsequent feature extraction module. The entire system performs parallel calculations on the response values of 144 sensors at 36 key nodes to obtain a set of node-level state vectors. On this basis, combined with derived indicators such as frequency drift, strain growth trend, and high-frequency vibration energy density, a health portrait of each node is constructed through the time series analysis module, and then provided to the neural network to determine which nodes have a tendency to fail in the future, which nodes are more affected by their diffusion paths, and give suggestions for structural reinforcement or local node replacement in advance.

[0068] Taking this example, the Ψ value of node No. 45 is greater than 10 for three consecutive cycles, and the system determines that the abnormal trend of its state is obvious. The frequency fluctuation amplitude of the adjacent nodes No. 47 and No. 49 in the time series feature vector increases by more than 17%. The graph neural network predicts that there is a 31% probability that it will be affected to the warning state within the next 120 minutes. Therefore, the platform marks this node as orange (medium to high risk) through the three-dimensional twin model and automatically issues a local inspection prompt and warning record.

[0069] The next step in this embodiment is to extract the evolutionary behavior characteristics of the key node states. Around the potential failure development trend of the key nodes in the grid structure, the dynamic analysis of the data of each sensor is carried out by setting a time series window to identify the physical evolutionary signal characteristics with predictive value. In this example, the continuous sampling time period T = 10 seconds is selected, and the sampling frequency is 50 Hz, that is, each type of sensor provides 500 data points within a window. Taking node number N45 as an example, the numbers of the strain gauge, displacement gauge, and accelerometer configured for it are S45-1, S45-2, and S45-3 respectively. Since the temperature data has a relatively small direct impact on the structural response, it is only used for environmental interference elimination in the feature evolution analysis and does not directly enter the calculation. First, the dynamic response intensity of the k-th type of feature is extracted, and the feature enhancement processing formula is used:

[0070]

[0071] The specific explanation is as follows: Suppose n = 3 channels are selected in this calculation, that is, the 3 sensors of node N45; first, determine φ j (t), which represents the fluctuation degree of the signal of the channel within time t, and the numerical range is set to 0.7, 1.3, where 1 is the reference value of the signal stable state, >1 indicates enhanced fluctuation, and <1 indicates stable or insufficient fluctuation of the signal. In this example, assume that the strain channel φ1(t) = 1.15, the displacement channel φ2(t) = 1.05, and the acceleration channel φ3(t) = 1.22; then assign values indicating the contribution ratio of different physical quantity channels to the k-th type of feature (such as "strain rate" or "frequency drift"), and the numerical range is 0.3, 1.0, reflecting the normalized mapping ability of multi-source channels under a unified index. Suppose in the current feature calculation (for example, frequency drift is the target feature k), Then it is This value is the response change rate of the j-th channel of the node relative to the healthy reference value, and the numerical unit depends on the channel. For example, the strain is με / s, the displacement is mm / s, and the vibration is g / s. The change rate range is usually set within the interval of 0.02, 0.5, indicating from slow deterioration to rapid mutation. In this example, assume that the strain channel is 0.12, the displacement channel is 0.09, and the acceleration channel is 0.28. Substitute the above data into the formula for trial calculation as follows:

[0072]

[0073] That is, the enhanced response value of node N45 for the target feature k (here assumed to be frequency drift) is 0.1403, and this value will be compared with the risk threshold defined in the historical data. In this system, the high-risk threshold for the enhanced response of frequency drift is 0.12, which means that when the node feature response value exceeds this value, it will be identified as a "trend of abnormal growth" node and enter the status warning list, and the adjacent nodes corresponding to it will undergo a diffusion risk simulation by the model.

[0074] At the same time, it is marked as a secondary risk propagation candidate node. The calculation result this time shows that node N45 has exceeded the threshold. The system marks the status level of this node as orange (medium to high risk) and updates and displays the node color through the three-dimensional digital twin platform. In the prediction module, node N45 will still maintain an upward trend within the next 90 minutes. The system synchronously activates the enhanced feature tracking of nodes 47 and 49, completes the dynamic simulation 20 minutes in advance, and provides a visual intervention window for operation and maintenance management personnel.

[0075] In this embodiment, to further realize the visualization and dynamic deduction of the failure diffusion trend in the overall structure, the system maps the status data of each node to a three-dimensional structure coordinate system, constructs a continuous risk expression model in the structural space, that is, uses the structural risk potential field model to uniformly quantify the influence between nodes, the degree of response mutation, and the time diffusion behavior, so as to simulate how local damage evolves and diffuses in the structure and predict the propagation trend. The risk potential field model formula used is:

[0076]

[0077] In this example, the number of key nodes already arranged is M = 36. First, select node number N45 as the target calculation point. The node coordinates are 21.3, 12.6, 9.5, and the current time t = 30 minutes. At this time point, node N45 has obtained its fusion feature intensity Θ 45 = 0.1403 through the feature enhancement process in the previous stage. This value indicates that the node status has shown a strong evolution trend in the past cycle; then define the risk excitation coefficient β m , this coefficient reflects the driving ability of the risk of this node to spread to the surrounding area once it is abnormal. The value range is 0.6, 1.5. Among them, the core nodes close to the main flow area and connecting multi-sided components usually take high values, and the edge nodes take low values. Assume that node 45 is a core node, and assign β 45 = 1.32;

[0078] Then extract the sensing response gradient value of this node This gradient represents the degree of sudden change in the state of the node compared to adjacent nodes in three-dimensional space, reflecting whether it is a "response hot spot". According to the trend of sensor numerical changes, the gradient value range is set to 0.1, 0.6, and here we select Finally, the diffusion impedance coefficient κ is defined m , this coefficient is used to measure whether the risk of this node is likely to spread outward. It is usually limited by the number of connecting components, construction complexity, and force flow distribution characteristics. The lower the impedance coefficient value, the stronger the diffusion ability. The value range is 0.8, 2.0. The current node is a typical densely connected area, and let κ 45 = 0.92. Substituting all the data into the risk potential field model, we get:

[0079] Σ(21.3, 12.6, 9.5, 30) = 1.32·0.1403·exp(-0.92·0.48·30) = 0.1852·exp(-13.248)

[0080] ≈0.1852·1.78×10 -6 ≈3.3×10 -7

[0081] This calculation result indicates that at the current time t = 30 minutes, the risk potential value of node N45 at the three-dimensional coordinate point 21.3, 12.6, 9.5 has rapidly decayed, indicating that although its early state was abnormal, its diffusion ability was weak, and the risk did not form a propagation peak within the structure, reflecting that this node belongs to a "self-excited non-diffusing type" risk point. At the same time, the system synchronously calculates the risk potential values of its adjacent nodes N47 and N49. The results show that Σ 47 = 1.2×10 -5 , Σ 49 = 6.8×10 -6 , which is two orders of magnitude higher than N45, indicating that its diffusion path may have deflected, induced by N45 but not continued, and the diffusion peak area may have shifted to the vicinity of N47. The platform immediately marks node N47 as a red warning point on the three-dimensional digital twin interface and marks the path direction from N45 to N47 with a propagation trend arrow. At the same time, it is predicted that the diffusion front will expand to the N50 area at t + 20 minutes. Therefore, the system issues a local linkage monitoring prompt, requiring to increase the sampling frequency of N50 and start local visual inspection; This example verifies that the structural risk potential field model of the present invention can combine multi-source fusion states, spatial geometric distributions, and time evolution parameters to continuously simulate the internal risk diffusion trend of the structure. Its spatial distribution calculation and time response behavior have high consistency and physical rationality, support refined and trend-based structural intelligent warning strategies, and have extremely strong engineering application potential in large complex steel structures.

[0082] Example 2:

[0083] After the layout of the aforementioned monitoring system, extraction of state characteristics, and construction of the risk potential field model for the steel grid roof structure of a large stadium are completed, the system enters the prediction layer, that is, based on the structure diagram model and the time-series evolution mechanism, dynamic prediction of node failure in both the time and space domains is realized. The core of this stage is: through the graph neural network (GNN), learn the state influence relationship between each node in the structure and its neighborhood, explore the spatial propagation law of node failure within the graph structure, and combine the time-varying trend of state evolution to predict potential failure paths and the node groups that may be affected in real time.

[0084] In the implementation process, first, the actual structure needs to be abstracted into a graph model. In this example, a total of 36 key nodes are monitored, and the edges are formed by connecting rods between the nodes. Therefore, the number of vertices in the graph model is 36, and the number of edges depends on the member topology. Based on the existing design drawings, the average node connectivity is about 4.2. The system assigns three-dimensional spatial coordinates to each node and constructs a multi-dimensional structure topology attribute matrix including geometric adjacency, force flow transfer direction, node connection stiffness, and historical damage response parameters. At the same time, dynamic weight coefficients are defined for the edges and nodes respectively. The node weight is used to measure its state influence on the surrounding structure, denoted as μ i , with a value range of 1.0, 2.5, and the edge weight χ ij represents the member transfer efficiency, with a range of 0.6, 1.2, and the connection coupling strength coefficient ξ ij is set according to the actual member stiffness, connection method, and bearing capacity difference between nodes i and j, with a value range of 50, 300;

[0085] Taking node N27 and its adjacent node N31 as an example, numbered i = 27 and j = 31 respectively. In the past time period (assuming τ = 15 minutes), after extracting the monitoring data, the state response of node N27 is σ 27 (t) = 0.42, and for node N31 it is σ 31 (t) = 0.33. The difference between the two represents the local state perturbation gradient, indicating that N27 is in the risk leading state. In addition, the influence enhancement coefficient μ of node N27 27 = 1.85, the transfer weight χ of the connecting edge 27,31 = 1.1, the coupling coefficient ξ 27,31 = 180, and the perturbation sensitivity function value ρ of node N31 31 (t) averages 0.56 in this cycle. Substitute it into the following structure response propagation function for cumulative calculation of propagation influence:

[0086] Γ 27,31 (t) = ∫0 15 [180·(0.42 - 0.33) 2 + 1.85·1.1·0.56]dt = ∫0 15[180·0.0081 + 1.1386]dt

[0087] = ∫0 15 [1.458 + 1.1386]dt = ∫0 15 2.5966dt = 38.949

[0088] This result indicates that within the past 15 minutes, the cumulative value of the failure response impact of node N27 on its adjacent node N31 is 38.949, which is much higher than the system-set diffusion threshold of 25. The system marks N31 as a candidate node significantly affected by perturbations and activates the propagation calculation of its next adjacent node N34. Meanwhile, during the training process of the graph neural network, node 27 is marked as a high-impact core, and its weight is dynamically increased in the graph convolutional layer, further affecting the deviation of the propagation direction of the graph structure in the model. The system marks the state propagation direction from N27 to N31 with a dynamic red arrow in the three-dimensional digital twin interface and displays that the predicted propagation time window on this path is 21 minutes in the risk prediction module, prompting the operation and maintenance team to increase the encryption sampling frequency at N31, start local visual-aided inspection, and store the prediction results in the node time series feature database for subsequent use in combination with time series models.

[0089] This embodiment further enters the time dimension modeling stage of state evolution. The core lies in using the continuous data within the historical monitoring period of nodes to dynamically mine the state evolution trend of nodes and predict future change trajectories, so as to judge whether there is a trend of deviation towards the failure state. In the current period, the system sets a unified time window length of τ = 20 minutes and a sampling frequency of 50Hz for all 36 key nodes. That is, each type of sensor will provide 60,000 raw data points within this period, and each node is at least equipped with 3 types of sensing channels for state modeling, namely strain gauges, accelerometers, and displacement gauges, forming a three-dimensional time series input tensor. Before the system inputs this tensor into the LSTM network for training, it is necessary to standardize and extract the change trend of each type of feature in the time series and complete the dynamic reduction of features through the sequence feature transformation function Φ k (t). The function expression is:

[0090]

[0091] This time, taking node N31 as the research object, the corresponding three channel numbers are S31-1 (strain), S31-2 (vibration), and S31-3 (displacement). First, set m = 3, that is, n = 3 channels, and extract the feature response functions ∈ of each channel respectively m(t), with a value range of 0.6, 1.2, assigned according to the fluctuation amplitude of the sensor signal strength. Assume the current time t = 114 minutes, ∈1(t) = 1.05, ∈2(t) = 0.91, ∈3(t) = 1.18; then extract Δ m (t) and its derivative It represents the offset rate of the node state relative to the healthy reference value, usually in the range of 0.01, 0.5. In this example, the three channels are 0.24, 0.31, and 0.19 respectively; the weighted normalization coefficient reflects the contribution degree of different channels to the expression of the k-th type of feature, and the range is set to 0.3, 1.0, automatically adjusted during model training or set by expert experience. Assume the current is the frequency drift type of feature (k = 2), set Substitute the above data into the feature transformation function:

[0092]

[0093] That is, the time series transformation value of the frequency drift feature of node N31 at the current time t = 114 minutes is 0.1631. This value is much higher than the system-set frequency stability threshold of 0.09. The system judges that its trend is in a continuously increasing state and belongs to a typical "non-steady evolution node". Therefore, the mapping relationship between the LSTM prediction output of this node and the risk level is bound. Through the deduction of the historical evolution model, it is predicted that the drift amplitude of the node will continue to increase to the range of 0.185 - 0.192 within the future 30-minute window. The system automatically writes this trend change into the global time series database and triggers the dependent propagation prediction of the surrounding nodes N29 and N33. At the same time, the state trend arrow of N31 is dynamically marked on the twin visualization platform, transitioning from yellow to orange, and the predicted curve evolution trajectory is superimposed in the form of a point animation for the on-duty personnel to observe in real time.

[0094] Rely on the graph neural network to capture the propagation logic relationship between structural nodes, and at the same time combine the time series model to fit and predict the evolution trend of node states, so as to realize the output of the risk level of key node failures in the grid structure, the identification of potential diffusion paths, and the estimation of time windows. Among them, first, the graph neural network module calculates the risk propagation potential of each node to adjacent nodes in combination with the structural topology relationship, and then the time series model predicts the state evolution curve of each node in a specific future time period. Through the joint output of the two models, the system defines the failure risk level scoring function Λ i , and outputs a dynamic risk score for each node, with a range of 0, 1. The closer the value is to 1, the higher the possibility of node failure in the future. For example, at t = 150 minutes, the feature evolution trend of node N31 (calculated in the previous section) is Φ N31 (t) = 0.1631. Combining the calculation results of its spatial topology influence relationship and propagation function, the model outputs its risk level as Λ31 = 0.76, which is at the boundary of the orange - red level. Subsequently, the system starts predicting the range of potential impact nodes based on this result. By constructing a graph - path propagation map, it extracts the secondary nodes affected by N31. Combining the propagation response function and the state evolution trend, it determines that nodes N33 and N29 will be affected by it in the next time period. Then it enters the third key indicator: estimating the risk diffusion time window, and calls the risk propagation delay function Υ ij , which is used to predict the time required for the failure impact to spread from N31 to N33. The formula is:

[0095]

[0096] According to the known information, take the prediction start time t0 = 150 minutes and the prediction interval δ = 30 minutes, that is, the calculation interval is 150, 180 minutes. It is known from the previous section that Φ 31 (t)=0.1631, and the current state of node N33 is calculated in real - time to get Φ 33 (t)=0.119. The propagation response function Γ has been calculated in the previous stage as 31,33 (t)=38.949 (average response per unit time), and the risk transfer path stability coefficient η 31,33 is determined by the construction type of the connection path, the fatigue state of the components, and the connection method, and its value range is 0.6, 1.4. In this case, it is set to a medium - high level of η 31,33 = 1.2. Substitute all the data into the formula for estimation:

[0097]

[0098] The final result Υ 31,33 = 4969.5 represents the total response delay intensity value required for node N31 to spread the risk to N33. In the system, it is set that when this value exceeds 3000, it is a medium - high propagation rate level. The system marks this path as an "accelerated propagation path" according to the threshold, dynamically renders the heat lines on this path through a three - dimensional twin platform, and sends a "multi - point linkage warning" prompt to the duty - keeping end. At the same time, the prediction model feeds back the minimum diffusion time in this path as 14 minutes, which means that if no manual intervention is carried out, node N33 will reach the orange level and trigger the response limit after 14 minutes. This example demonstrates the three - element linkage prediction ability of "risk level - diffusion area - propagation time" in the dynamic prediction method of the present invention by fully substituting the time - series characteristic values, spatial propagation functions, and risk path stability coefficients. Especially in complex steel grid structures, there are high - order coupling relationships between nodes, and traditional independent modeling methods are difficult to identify the risk conduction law, while this method can realize the joint deduction of the spatial update and time trend of node states, and has extremely high practicality and early control ability in the engineering warning system.

Claims

1. A prediction method for the failure diffusion of grid steel structure joints, characterized in that It includes the following steps: S1. Establish a node failure evolution mechanism model: S1.

1. Based on connection forms including bolt sphere nodes and welded nodes, stress transfer paths, node local yield, and fatigue accumulation factors, form a failure cause and evolution path model; S1.

2. Construct the diffusion logic of node failure → force flow redistribution → stress concentration in adjacent nodes → chain reaction; S2. Regard the nodes as points in the graph and the members as edges, and define the propagation weights between nodes; the weight parameters are dynamically assigned based on material properties, node connection types, spatial geometric positions, and historical stress states; S3. Adopt a multi-dimensional structure monitoring system and construct nodes: S3.

1. Install multi-source monitoring sensors at key nodes of the space truss structure, including strain gauges, displacement gauges, thermometers, and accelerometers, to sense the node states; combine the real-time monitoring data to extract key features such as strain rate changes, vibration characteristics, frequency drift, and residual deformation of each node; S3.

2. Use time series analysis to construct a node health state feature vector set to reflect the gradual degradation trend of the nodes; S4. Use the graph neural network GNN combined with the time series model to perform dynamic prediction of node failure in both the time and space domains; The output of the sequence model includes: ①. The failure risk level of a single node, ②. The range of potentially affected nodes, ③. The diffusion time window; S5. Use the incremental learning algorithm to input new monitoring data and dynamically correct the model parameters; make the model evolve with the evolution of the structure; S6. Display the prediction results through a three-dimensional digital twin visualization platform: represent the node failure risk levels with different colors; dynamically demonstrate the predicted failure diffusion path and conduction time; provide decision-making support functions for structural reinforcement, node replacement, and risk warning.

2. The prediction method for the failure diffusion of the grid steel structure joints according to claim 1, characterized in that The method of installing a multi-dimensional structure monitoring system and constructing nodes: Install a multi-source sensing system at key nodes of the structure. The sensor types cover strain gauges, displacement gauges, accelerometers, and thermometers, which are used to sense stress responses, geometric deformations, dynamic disturbances, and environmental disturbances respectively; the collected data is synchronized and normalized in the time and space dimensions.

3. The prediction method for the failure diffusion of the grid steel structure node according to claim 2, wherein The method of installing a multi-dimensional structure monitoring system and constructing nodes: Extract features for the evolution behavior of nodes; Take the feature evolution rate as the core monitoring index, and extract state labels with predictive value by analyzing the evolution trends of strain rate changes, acceleration spectrum offset, and displacement residual features in the time domain; the feature enhancement processing formula is as follows: Where: Θ k is the dynamic response intensity of the k-th type of feature, including strain rate, vibration amplitude, and frequency drift; φ j (t) represents the signal fluctuation function of the j-th node monitoring channel at time t, reflecting the real-time activity degree or noise change trend of the sensing channel; is the normalized projection factor of the k-th type of feature on the j-th channel; δ j is the instantaneous response deviation between the monitoring value of the j-th channel and the reference healthy value; represents the rate of change of the deviation with respect to time, used to measure the deterioration speed of the node under a specific physical quantity.

4. The prediction method for the failure diffusion of the grid steel structure joints according to claim 3, characterized in that The method of installing a multi-dimensional structure monitoring system and constructing nodes: Project the states of all nodes into the three-dimensional space of the structure to form a continuously calculated state evolution field to judge the distribution and development direction of the failure risks of each node inside the structure; adopt a structure risk potential field model based on time evolution and node interaction to continuously deduce the diffusion trend; the potential field model is defined as follows: Where: Σ(x, y, z, t) represents the risk potential value of the structure at the three-dimensional coordinate position (x, y, z) and time t; M is the total number of key nodes monitored in the structure; β m is the risk excitation coefficient of the m-th node; Θ m is the intensity of the fusion feature of this node; is the gradient of the sensing response field of the m-th node in the three-dimensional structure, indicating the severity of the change in the signal sensed by the node in space; κ m is the diffusion impedance coefficient, which is used to measure the sensitivity of the risk of this node spreading to the neighborhood; then expresses the diffusion attenuation or growth trend of the structure risk over time.

5. The prediction method for the failure diffusion of the grid steel structure joint according to claim 1, wherein The dynamic prediction method in both the time and space domains includes: Abstract the grid structure into a graph model, where each node in the structure is used as a vertex in the graph, and the connecting members are used as edges. Construct a graph topology structure that includes geometric position information, force flow transfer direction, connection stiffness, and historical damage response, and assign weight coefficients representing physical correlation degrees and structural influence to the edges and nodes respectively. Learn the information interaction pattern between nodes and their neighborhoods in the graph through a graph neural network, and establish a spatial coupling update mechanism for describing the propagation influence of local failure states within the graph structure.

6. The prediction method for the failure diffusion of the grid steel structure joints according to claim 5, characterized in that The dynamic prediction method in the time and space dual domain includes: Taking the state sequence of each node within the historical monitoring period as the input, a time series model is used to dynamically model the evolution trend. The model inputs include key parameters such as the strain rate change, frequency drift, and micro-vibration amplitude of the node, and a sequence feature transformation function Φ k (t) is introduced to extract the time series law, and the function is defined as follows: Where: Φ k (t) represents the global change trend of the k-th type of feature at time t; ∈ m (t) is the feature response intensity function of the m-th node channel at time t; Δ m (t) is the offset response of this channel at time t, used to measure the degree to which the node state deviates from the healthy benchmark; represents the change rate of the state response; is the weighted normalization coefficient of the k-th type of feature in the m-th channel; lim_{n→∞} represents the global reduction limit of the statistical process under the multi-node network.

7. The prediction method for the failure diffusion of the grid steel structure joints according to claim 6, characterized in that The dynamic prediction method in the time and space dual domain includes: Through the linkage prediction mechanism of the time series model and the graph structure, the dual-domain joint output of node failure includes: the failure risk level of each node, with a continuous scoring function Λ i Output, reflecting the failure probability or risk level of the node in a future time period; the set of potentially affected nodes, and the range of nodes that will be affected by the current high-risk nodes in the future is extracted through the joint modeling of the topological diffusion path and the propagation influence graph spectrum.

Images