Bridge structure fault prediction method and system based on monitoring data

Abstract

The invention discloses a bridge structure fault prediction method and system based on monitoring data, and relates to the field of bridge structure health monitoring, and the method comprises the steps: obtaining an original monitoring data stream, carrying out the preprocessing, carrying out the feature extraction based on the preprocessed data, and obtaining a standardized feature data set; inputting the standardized feature data set into a causal feature mining framework based on physical information guidance, and constructing a causal directed acyclic graph between sensor nodes; based on a causal directed acyclic graph, quantifying contribution degrees of different factors to structure response by using a causal intervention effect, and screening out high-order damage sensitive feature vectors; and inputting the high-order damage sensitive feature vector into a digital twin model integrated with a phase-field method fracture model. According to the method, the effects of different maintenance strategies on crack development inhibition are quantitatively evaluated by constructing the probabilistic decision tree, and a graded early warning and specific maintenance decision report is generated based on a simulation prediction result and a dynamic risk threshold value.

Description

Technical Field

[0001] This invention relates to the field of bridge structural health monitoring, and in particular to a method and system for predicting bridge structural failures based on monitoring data. Background Technology

[0002] In the field of bridge structural health monitoring, intelligent fault prediction based on monitoring data is an important research direction for ensuring the safe operation of infrastructure. Existing technical solutions, such as a bridge condition assessment method that integrates deep neural networks and digital twins, represent a typical practice in this field. This method involves deploying a sensor network to collect structural response data, using deep neural networks to learn the complex mapping relationship between damage characteristics and condition, and combining digital twin technology to achieve visualization of structural condition and anomaly detection. By combining data-driven models and physical models, the automation and intuitiveness of condition assessment are improved.

[0003] Existing technologies have limitations in achieving interpretability of prediction results and providing accurate decision support based on predictions. The black-box nature of deep neural network models makes it difficult to elucidate the specific causal contributions of various environmental and operational factors to structural damage, resulting in a lack of clear causal logic to support the prediction results. Existing digital twins mostly focus on state reproduction and visualization, and their simulation process is often not deeply integrated with the causal mechanisms mined from the data. Therefore, it is difficult to reliably and quantitatively assess how different maintenance measures will affect the future evolution of damage. As a result, the output of existing methods usually remains at the level of risk warning and is difficult to be directly transformed into a clear decision-making basis to guide specific maintenance actions, leading to a disconnect between prediction and maintenance. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a bridge structure failure prediction method based on monitoring data to solve the problem in the prior art that the prediction results are unexplainable due to the lack of physical mechanism and causal analysis, and are difficult to directly transform into a basis for targeted maintenance decisions.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a bridge structure fault prediction method based on monitoring data, which includes: acquiring the original monitoring data stream and preprocessing it, extracting features based on the preprocessed data, and obtaining a standardized feature dataset.

[0008] The standardized feature dataset is input into a causal feature mining framework guided by physical information to construct a causal directed acyclic graph between sensor nodes.

[0009] Based on causal directed acyclic graphs, the contribution of different factors to structural response is quantified by causal intervention effects, and high-order damage-sensitive feature vectors are screened out.

[0010] By inputting high-order damage-sensitive feature vectors into a digital twin model that integrates a phase-field fracture model, and constructing a decision tree of damage state transition probability, the intervention effect of different maintenance strategies on crack propagation path is evaluated.

[0011] Based on real-time acquired multi-source data and impact loads, physical simulation of damage evolution is performed to output visualized pre-simulation data of the damage evolution process, and a multi-level early warning mechanism is activated to generate maintenance decision-making recommendation reports.

[0012] As a preferred embodiment of the bridge structure failure prediction method based on monitoring data described in this invention, the method includes the following steps: acquiring and preprocessing the original monitoring data stream, extracting features based on the preprocessed data, and obtaining a standardized feature dataset:

[0013] The raw monitoring data stream containing stress, vibration, displacement, temperature and images is acquired. Preprocessing operations, including missing value imputation, outlier removal and low-pass filtering, are performed on the raw monitoring data stream to obtain the preprocessed data stream.

[0014] Feature extraction is performed on the preprocessed data stream to extract the first three natural frequencies from the vibration component and the mean and variance from the strain component, thus obtaining a standardized feature dataset.

[0015] As a preferred embodiment of the bridge structure failure prediction method based on monitoring data described in this invention, the method involves: inputting a standardized feature dataset into a causal feature mining framework guided by physical information to construct a causal directed acyclic graph between sensor nodes, including the following steps:

[0016] The standardized feature dataset is input into a causal feature mining framework guided by physical information, and structural vibration differential equation constraints are embedded.

[0017] In the causal feature mining framework guided by physical information, the PC algorithm is used to perform causal discovery analysis on a standardized feature dataset;

[0018] Based on the results of causal discovery analysis, a causal directed acyclic graph is constructed among the sensor nodes of temperature node, load node, strain node, frequency node, and damage state node.

[0019] As a preferred embodiment of the bridge structure failure prediction method based on monitoring data described in this invention, the method includes the following steps: Based on a causal directed acyclic graph, the contribution of different factors to the structural response is quantified using causal intervention effects, and high-order damage-sensitive feature vectors are selected:

[0020] Starting from the causal directed acyclic graph, the temperature factor, load factor and fatigue accumulation factor are counterfactually intervened through the structural nesting model to remove the confounding effects between the factors and quantify the independent contribution of each factor to the structural response factor.

[0021] The causal directed acyclic graph is weighted using independent contribution values, and a random walk algorithm based on Monte Carlo simulation is executed to calculate the average first-hit probability from each input factor node to the damage state node, thus constructing a probability-weighted causal network.

[0022] The average first-hit probability of each path in the probability-weighted causal network is adaptively convolved with the real-time operational risk level to generate a dynamic causal sensitivity threshold that matches the current risk status.

[0023] Based on the dynamic causal sensitivity threshold, high-sensitivity causal paths are selected in the probability-weighted causal network, and the corresponding features in the standardized feature dataset mapped by the path are extracted to form a high-order damage-sensitive feature vector.

[0024] As a preferred embodiment of the bridge structure failure prediction method based on monitoring data described in this invention, the method includes the following steps: inputting a high-order damage-sensitive feature vector into a digital twin model integrating a phase-field fracture model; constructing a decision tree for damage state transition probabilities; and evaluating the intervention effect of different maintenance strategies on crack propagation paths.

[0025] Based on the average first-hit probability, adjust the initial distribution of the corresponding physical parameters in the phase-field fracture model;

[0026] In a digital twin model integrating a phase-field fracture model, crack grouting, carbon fiber cloth bonding, and external prestressing are defined, and each maintenance strategy is encoded as an intervention operation on a specific high-sensitivity causal path in a probability-weighted causal network.

[0027] Physical simulation of a digital twin model integrating a phase-field fracture model is performed. By comparing the crack evolution results before and after the application of maintenance strategy intervention, the changes in crack bifurcation, turning and healing patterns in the mesh are statistically analyzed, and a decision tree reflecting the probability of damage state transition under different interventions is constructed.

[0028] By traversing the decision tree of damage state transition probability, and analyzing the path transition probability distribution from the initial damage state node to each terminal state node under different maintenance strategy intervention operation sequences, the expected intervention effect of crack grouting strategy on the main crack propagation length is evaluated.

[0029] As a preferred embodiment of the bridge structure failure prediction method based on monitoring data described in this invention, the method includes the following steps: performing physical simulation of damage evolution based on real-time acquired multi-source data and impact loads to output visualized pre-simulation data of the damage evolution process.

[0030] Acquire real-time multi-source data from dynamic weighing, meteorological monitoring, and GPS, as well as recorded overload or impact events, as impact loads;

[0031] Real-time multi-source data and impact loads are input into a digital twin model that integrates a phase field method fracture model to perform physical simulation of damage evolution based on physical mechanisms.

[0032] From the physical simulation of damage evolution based on physical mechanisms, data on the changes of crack length, propagation angle, and stress intensity factor over time are extracted and output as visual pre-simulation data of the damage evolution process.

[0033] As a preferred embodiment of the bridge structure failure prediction method based on monitoring data described in this invention, the method includes the following steps: activating a multi-level early warning mechanism to generate a maintenance decision recommendation report.

[0034] The damage evolution process visualization simulation data is compared with the warning level thresholds preset based on historical safety data, structural design specifications and real-time operating environment to determine the degree of deviation of crack propagation rate, path and stress intensity factor from the warning level thresholds and obtain the comparison results.

[0035] Based on the comparison results, a multi-level early warning mechanism, including blue alert, yellow warning, orange alarm, and red emergency response, is triggered.

[0036] Based on the visualized pre-simulation data of the damage evolution process and the multi-level early warning mechanism, a maintenance decision recommendation report is generated.

[0037] Secondly, the present invention provides a bridge structure fault prediction system based on monitoring data, including a feature extraction module, which acquires the original monitoring data stream and preprocesses it, and extracts features based on the preprocessed data to obtain a standardized feature dataset.

[0038] The module constructs a causal directed acyclic graph between sensor nodes by inputting a standardized feature dataset into a causal feature mining framework guided by physical information.

[0039] The screening module, based on a causal directed acyclic graph, uses the causal intervention effect to quantify the contribution of different factors to the structural response and screens out high-order damage-sensitive feature vectors.

[0040] The evaluation module inputs high-order damage-sensitive feature vectors into a digital twin model that integrates a phase-field fracture model. By constructing a decision tree of damage state transition probability, it evaluates the intervention effect of different maintenance strategies on crack propagation paths.

[0041] The recommended report module performs physical simulation of damage evolution based on real-time acquired multi-source data and impact loads, outputs visualized pre-simulation data of the damage evolution process, and activates a multi-level early warning mechanism to generate maintenance decision recommendation reports.

[0042] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the bridge structure fault prediction method based on monitoring data as described in the first aspect of the present invention.

[0043] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the bridge structure fault prediction method based on monitoring data as described in the first aspect of the present invention.

[0044] The beneficial effects of this invention are as follows: By collecting and fusing multi-dimensional data such as stress, vibration, and temperature from multiple sensors, a standardized feature dataset is constructed. Using a causal inference framework that integrates constraints from physical mechanics equations, a causal relationship network between load, environment, and structural response is mined from the data. The independent and joint effects of each factor on structural damage are quantified, thereby selecting high-order features that are extremely sensitive to damage. These features are then input into a digital twin that integrates a phase-field fracture mechanics model. The twin not only simulates the evolution of cracks under real-time loads and environments, but also quantifies the effectiveness of different maintenance strategies in suppressing crack development by constructing a probabilistic decision tree. Based on simulation prediction results and dynamic risk thresholds, graded early warning and specific maintenance decision reports are generated. Attached Figure Description

[0045] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart of a bridge structure failure prediction method based on monitoring data.

[0047] Figure 2 This is a schematic diagram of a bridge structure fault prediction system based on monitoring data. Detailed Implementation

[0048] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0049] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0050] Secondly, the term "one embodiment" or "example" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the invention. The appearance of an embodiment in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that mutually excludes other embodiments.

[0051] Reference Figures 1-2 This is one embodiment of the present invention, which provides a bridge structure failure prediction method based on monitoring data, including the following steps:

[0052] S1. Obtain the raw monitoring data stream and preprocess it. Based on the preprocessed data, extract features to obtain a standardized feature dataset.

[0053] S1.1 Acquire the raw monitoring data stream containing stress, vibration, displacement, temperature and images, and perform preprocessing operations on the raw monitoring data stream, including missing value filling, outlier removal and low-pass filtering, to obtain the preprocessed data stream.

[0054] Furthermore, fiber optic strain sensors are deployed at key stress locations such as the mid-span of the main girder, supports, and piers to collect stress signals; piezoelectric accelerometers are deployed to collect vibration signals; hydrostatic levels are deployed to collect displacement signals; digital temperature sensors are deployed to collect temperature signals; and high-definition network cameras are deployed to capture images of the bridge's surface. All sensors are triggered by a synchronous clock signal and operate continuously according to a preset sampling frequency, thus forming a raw monitoring data stream containing stress, vibration, displacement, temperature, and images. Preprocessing operations are performed on the raw monitoring data stream. For gaps in the data sequence caused by brief communication interruptions, missing values ​​are filled using linear interpolation based on adjacent data points. For abnormal data points that deviate significantly from the normal range due to instantaneous strong electromagnetic interference or occasional sensor malfunctions, the Laida criterion based on statistical principles is used for identification and removal. For the high-frequency random noise commonly present in vibration acceleration signals, a Butterworth low-pass filter with a specific cutoff frequency is used for filtering to retain the low-frequency effective components reflecting the main dynamic characteristics of the structure, resulting in a preprocessed data stream.

[0055] Specifically, the synchronous acquisition of data from multiple heterogeneous sensors lays a solid foundation for fusion analysis. A hierarchical and targeted preprocessing strategy, rather than a single general filter, is adopted to retain the effective information in different physical signals and remove their unique noise. For example, linear interpolation ensures the continuity of data, the Laida criterion effectively eliminates sudden interference, and low-pass filtering with adjustable cutoff frequency accurately separates the structural response and high-frequency noise in the vibration signal, thus improving data quality.

[0056] S1.2. Perform feature extraction based on the preprocessed data stream, extract the first three natural frequencies from the vibration part, and extract the mean and variance from the strain part to obtain a standardized feature dataset.

[0057] Furthermore, feature extraction is performed based on the preprocessed data stream. For the vibration acceleration signal portion of the preprocessed data stream, the Fast Fourier Transform algorithm is applied to transform it from the time domain to the frequency domain to obtain the power spectral density estimate of the signal. The peak detection algorithm is used to identify the three most prominent spectral peaks in the power spectrum, and the frequency values ​​corresponding to these three spectral peaks are extracted as the first three natural frequencies of the structure. For the strain signal portion of the preprocessed data stream, the average value of all strain samples in the entire time series is obtained as the mean feature to characterize the average stress level or static deformation of the measuring point. The variance of the strain value relative to its mean is obtained to characterize the fluctuation amplitude of the stress or the dynamic response intensity of the measuring point. The extracted first three natural frequencies, strain mean, and strain variance are arranged and combined in a predetermined order. The influence of differences in dimensions and magnitudes between different features is eliminated by methods such as max-min normalization, thereby forming a standardized feature dataset.

[0058] Specifically, characteristic indicators with clear physical meaning and sensitivity to damage are selectively extracted from the preprocessed multi-source data. Natural frequencies are extracted from vibration signals because these features are a comprehensive reflection of the overall stiffness and mass of the structure and are extremely sensitive to overall structural damage such as stiffness degradation, reflecting the evolution trend of structural performance from a macroscopic perspective. The mean and variance are extracted from strain signals, which characterize the local stress state from both static and dynamic dimensions. The mean reflects the stable stress level, while the variance reflects the severity of stress changes. The combination of the two can more comprehensively capture the signs of local damage initiation or development, constructing a comprehensive feature descriptor that takes into account both global and local, static and dynamic aspects.

[0059] S2. Input the standardized feature dataset into the causal feature mining framework guided by physical information to construct a causal directed acyclic graph between sensor nodes.

[0060] S2.1 Input the standardized feature dataset into the causal feature mining framework guided by physical information and embed the structural vibration differential equation constraint.

[0061] Furthermore, the governing equations describing the dynamic behavior of the structure, such as the equations of motion of a multi-degree-of-freedom system, or their discretized form or frequency response function relationship, are transformed into rules that impose a priori constraints on the possible causal relationships between variables in the standardized feature dataset. For example, according to vibration theory, the natural frequency of a structure is related to the stiffness matrix and the mass matrix, while temperature changes affect the elastic modulus of the material and thus the stiffness. The framework will prioritize or forcibly retain the directed edges from the temperature node to the frequency node, while excluding those causal assumptions that violate the basic principles of mechanics, such as the reverse causal relationship from the frequency node to the temperature node.

[0062] Specifically, by incorporating mathematical constraints into the data-driven causal discovery process, the risk of spurious correlations or incorrect causal directions due to limited samples or noise interference when simply mining causal relationships from data is reduced. This improves the physical credibility of the constructed causal graph, transforming it into an adjustable soft constraint or Bayesian prior, guiding the search algorithm to explore within a solution space that conforms to physical laws, thus achieving an organic combination of data-driven and mechanism-driven approaches.

[0063] S2.2 In the causal feature mining framework guided by physical information, the PC algorithm is used to perform causal discovery analysis on the standardized feature dataset.

[0064] Furthermore, within the physics-guided causal feature mining framework that already incorporates constraints from the structural vibration differential equation, the PC algorithm performs causal discovery analysis on the standardized feature dataset. By performing a series of statistical conditional independence tests, the completely undirected graph is gradually reduced, inferring the causal skeleton and orientation between variables. The embedded physical constraints play a crucial role. For example, during conditional independence tests, if the potential causal relationship between two variables is supported or prohibited by the physical equation, the significance level of the test or the edge orientation rule will be adjusted accordingly. In the iterative edge deletion phase, if the causal relationship represented by an edge to be deleted is strongly supported by the structural vibration differential equation, the significance level required to delete the edge will be set higher, thus increasing the likelihood of retaining the edge. In the edge orientation determination phase, physical constraints can provide additional prior information for the V-structure orientation rule, helping to distinguish ambiguous causal directions.

[0065] Specifically, by using statistical tests guided and corrected by physical knowledge, stable and reliable causal dependencies between variables can be identified from observational data. This makes up for the shortcomings of pure physical models that may not be able to cover all complex interactions in reality, and also corrects the conclusions that may be produced by pure data-driven methods that violate common sense in physics. The resulting causal structure is the optimal inference under the combined effect of data evidence and physical principles.

[0066] S2.3. Based on the results of causal discovery analysis, construct a causal directed acyclic graph between sensor nodes for temperature node, load node, strain node, frequency node, and damage state node.

[0067] Furthermore, the causal dependencies between variables output by the PC algorithm under physical constraints are mapped to a directed acyclic graph data structure. Each node represents a specific feature or an abstract damage state in the standardized feature dataset. For example, the temperature node corresponds to the temperature feature, the load node corresponds to the vehicle load feature, the strain node corresponds to the strain mean or variance feature, the frequency node corresponds to the extracted first three natural frequencies, and the damage state node may be a latent variable or comprehensive index defined based on strain or frequency anomalies. For example, an edge from the temperature node to the frequency node indicates that temperature change is considered one of the causes of natural frequency drift, an edge from the load node to the strain node indicates that vehicle load is the main cause of strain change, and an edge from multiple response nodes (such as abnormal strain nodes and abnormal frequency nodes) to the damage state node indicates that these anomalies collectively point to potential structural damage.

[0068] Specifically, the abstract relationships obtained from causal discovery analysis are transformed into a structured graphical model. The causal directed acyclic graph clearly reveals the causal transmission path from environmental and load inputs to various structural responses and finally to the comprehensive damage state. This provides a clear logical blueprint for understanding the damage mechanism and subsequent contribution quantification and sensitive feature screening. The damage state is incorporated into the causal graph as a latent variable node that can be derived from multiple observable features.

[0069] S3. Based on the causal directed acyclic graph, the contribution of different factors to the structural response is quantified by the causal intervention effect, and high-order damage-sensitive feature vectors are screened out.

[0070] S3.1 Starting from the causal directed acyclic graph, the temperature factor, load factor and fatigue accumulation factor are counterfactually intervened through the structural nesting model to remove the confounding effects between the factors and quantify their independent contribution values ​​to the structural response factors.

[0071] Furthermore, based on the causal dependency structure revealed by the causal directed acyclic graph, the structural nesting model constructs a parameterized conditional mean model to characterize the conditional expectation of structural response factors given temperature, load, fatigue accumulation, and other covariates. It then performs counterfactual intervention by simulating a series of virtual experiments: for the temperature factor, its value is fixed at a specific counterfactual level, such as maintaining it at the annual average temperature; while the values ​​of the load and fatigue accumulation factors are fixed at their historical averages or typical patterns in the observed data, respectively. Under this intervention condition, the results are used to... The nested model recalculates the predicted values ​​of structural response factors and compares them with the predicted values ​​of structural response factors when all factors act together under real observation conditions. The average difference between the predicted values ​​is quantified as the independent contribution value of the temperature factor to the structural response factor. This process is repeated for the load factor and fatigue accumulation factor. The independent contribution values ​​of each factor are obtained by fixing them at the counterfactual level and fixing the other factors at the observation background. The nested model simulates the counterfactual scenario where only the factors change while everything else remains unchanged, thus removing the confounding effects caused by common causes or chain reactions among the factors.

[0072] Specifically, decomposing the pure impact of each input factor on the structural response at the causal level is crucial for identifying key drivers of damage and guiding targeted maintenance. Now, a high-level causal inference model, the structural nesting model, used to process observational data, is deeply integrated with specific issues in bridge structural health monitoring. By utilizing structural information provided by causal directed acyclic graphs (DAGs), the model form is correctly specified and the set of covariates to be controlled is determined, ensuring the logical correctness of counterfactual interventions. For example, if a causal DAG shows that fatigue accumulation may be simultaneously affected by load and historical temperature, then when simulating the independent contribution of load factors, the structural nesting model appropriately treats fatigue accumulation as an intermediate variable affected by the load, more accurately separating the direct and indirect effects of the load.

[0073] S3.2. Use independent contribution values ​​to perform a weighted transformation on the causal directed acyclic graph, execute a random walk algorithm based on Monte Carlo simulation, calculate the average first-hit probability from each input factor node to the damage state node, and construct a probability-weighted causal network.

[0074] Furthermore, each directed edge in the causal directed acyclic graph (DAG) pointing from a parent node to a child node is assigned a weight. This weight is equal to the proportion of the parent node's independent contribution value to the sum of the independent contribution values ​​of all parent nodes pointing to that child node. This transforms the DAG into a weighted directed network. A Monte Carlo simulation-based random walk algorithm is then executed, defining the starting point from the current node... Random walk to any of its child nodes transition probability The transition probability is equal to the connection and edge weight Occupy node The proportion of the sum of edge weights to all its child nodes, such as temperature factor nodes, is used to perform a large number of random walk simulations. Each walk moves in the network according to the above transition probability rules until a certain damaged state node is reached or the preset maximum number of steps is reached. By statistically analyzing the proportion of the number of times each damaged state node is reached for the first time in multiple walks starting from a certain input factor node, the average first-hit probability from the input factor node to each damaged state node is calculated. Repeating this process for all input factor nodes can construct a probability-weighted causal network with input factor nodes and damaged state nodes as the two ends and the average first-hit probability as the connection strength.

[0075] Specifically, by simulating random walks, the cumulative effect of factors indirectly influencing the damage state through multiple possible paths in the network was captured, thus providing a more comprehensive assessment of the overall probability of each factor causing damage. The causal directed acyclic graph was treated as a state transition network of a stochastic process, and Monte Carlo simulation was used to explore the probability that factors would eventually lead to damage through complex causal chains. This approach reveals the true influence of factors in complex systems more effectively than relying solely on direct causal strength or simple path analysis.

[0076] The expressions for random walk and hit probability are:

[0077] ;

[0078] in, To start from the current node Random walk to its child nodes The transition probability, The node where the current random walk is located. For the current node A direct child node, For the node To the node The weight of the directed edge. For loop variable, Let be the set of all direct children of a node in a causal directed acyclic graph;

[0079] S3.3. Adaptively convolution the average first-hit probability of each path in the probability-weighted causal network with the real-time operational risk level to generate a dynamic causal sensitivity threshold that matches the current risk status.

[0080] Furthermore, the average first-hit probability corresponding to each path from the input factor node to the damage state node in the probability-weighted causal network is adaptively convolved with a convolution kernel that is dynamically adjusted according to the real-time operational risk level. The real-time operational risk level is comprehensively evaluated by indicators such as real-time monitored traffic flow, maximum axle load, and wind speed. The convolution kernel can be designed as a function that reflects the sensitivity requirements for early and weak causal signals under different risk levels. For example, under high risk levels, the convolution kernel may be designed to give higher weights to even lower average first-hit probabilities, thereby lowering the threshold and capturing more potential risk paths. Under low risk levels, a more stringent convolution kernel is used. Through adaptive convolution, the original average first-hit probability spectrum is transformed into a dynamic causal sensitivity score closely related to the current structural safety status, and a dynamic causal sensitivity threshold is set.

[0081] Specifically, the system dynamically adjusts its warning sensitivity based on the bridge's real-time operational risk status, achieving adaptive warning sensitivity. This increases warning sensitivity to prevent problems before they occur when risks are high, and reduces false alarm rates to conserve maintenance resources when risks are low. By combining the static probability metric of causal networks with the dynamic operational environment and utilizing the signal processing concept of convolution operations, the system achieves context-aware adjustment of causal influence scoring, ensuring that the selected features are not only based on their inherent causal importance.

[0082] S3.4. Based on the dynamic causal sensitivity threshold, select high-sensitivity causal paths in the probability-weighted causal network, and extract the corresponding features in the standardized feature dataset mapped by the path to form a high-order damage sensitivity feature vector.

[0083] Furthermore, based on the dynamic causal sensitivity threshold, a screening process is performed in the probability-weighted causal network to identify paths from input factor nodes to damage state nodes. Paths whose corresponding dynamic causal sensitivity scores exceed the threshold are defined as high-sensitivity causal paths. All nodes on these high-sensitivity causal paths are traced, and these nodes correspond to specific features in the standardized feature dataset. For example, a high-sensitivity path may pass through nodes such as daily temperature range and strain variance at mid-span of the main beam. The corresponding features in the standardized feature dataset mapped by these nodes are extracted, and these features are combined to form a high-order damage sensitivity feature vector.

[0084] Specifically, a dual-driven feature selection mechanism based on the global influence of causal networks and real-time risk dynamic adjustment was implemented. The selected high-order damage-sensitive feature vectors are not only statistically related to damage, but also have a significant driving force on damage occurrence in terms of causal logic. Furthermore, the selection criteria are adaptively adjusted according to changes in operational risks, thereby ensuring the time-varying relevance and high predictive value of the feature vectors. The feature selection criteria have been upgraded from traditional statistical relevance or importance ranking to a dynamic, networked metric that integrates causal transmission probability and real-time risk preference. This avoids selecting features with only superficial relevance and prevents the underreporting of weak but important early damage signals during calm periods.

[0085] S4. Input the high-order damage-sensitive feature vector into the digital twin model that integrates the phase-field fracture model, and evaluate the intervention effect of different maintenance strategies on the crack propagation path by constructing a decision tree of damage state transition probability.

[0086] S4.1 Adjust the initial distribution of the corresponding physical parameters in the phase-field fracture model based on the average first-hit probability.

[0087] Furthermore, the average first-hit probability quantifies the overall likelihood that different input factor nodes will trigger specific damage state nodes. In digital twin models, input factors are typically mapped to physical parameters that influence crack evolution. For example, load level corresponds to stress boundary conditions, and temperature factor corresponds to the temperature coefficient of the material's elastic modulus. For a certain physical parameter, if the average first-hit probability of its corresponding input factor node reaching the critical damage state node is high, then when initializing the phase-field fracture model, the initial probability distribution of this parameter is set to a value range that is more likely to trigger damage. For example, the mean of the initial stress distribution is appropriately increased, or its variance is appropriately increased, to reflect the high-risk weight of this factor in the causal network. By setting the prior distribution of the parameter, the physical simulation is guided to explore high-risk conditions with a higher probability.

[0088] Specifically, the macroscopic risk measure derived from causal analysis is seamlessly embedded into the initialization process of the mesoscopic simulation model based on physical mechanisms. This transforms the simulation from a blind search based on uniform or empirical assumptions into a biased exploration based on data-driven causal cognition, thereby significantly improving the efficiency of the simulation in capturing real high-risk damage patterns. It also establishes a mapping bridge between causal influence and the initial uncertainty of physical parameters. For example, causal analysis may reveal that the average first hit probability from the load node of an overweight vehicle to the crack node at the lower edge of the main beam is very high. By setting the initial load distribution to a distribution biased towards high values, the simulation will encounter high-stress conditions more frequently in multiple runs, thus enabling more efficient study of crack evolution under such high-risk causal drive.

[0089] S4.2 In the digital twin model integrating the phase field method fracture model, define crack grouting, carbon fiber cloth bonding, and external prestressing, and encode each maintenance strategy as an intervention operation on a specific high-sensitivity causal path in the probability-weighted causal network.

[0090] Furthermore, the analysis identifies highly sensitive causal paths in the probabilistic weighted causal network. These paths describe key causal transmission chains from specific input factors to damage states. For example, a highly sensitive path might show dynamic overload → local stress concentration → microcrack initiation → macrocrack propagation. For each maintenance strategy, the analysis examines how it affects this causal chain at the physical level. The physical function of crack grouting is mainly to fill crack voids and restore material continuity, thereby directly blocking the microcrack initiation → macrocrack propagation link, and may even partially reverse local stress concentration. The crack grouting strategy is encoded as a cutting operation on the edge representing microcrack initiation → macrocrack propagation in the probabilistic weighted causal network. That is, in the digital twin simulation, the causal effect weight of this edge is set to zero or a minimum value during the intervention period to simulate the effect of crack propagation being suppressed after the crack is filled. The physical function of bonding carbon fiber cloth is mainly to provide additional tensile strength and stiffness. It does not directly eliminate cracks, but reduces the degree of local stress concentration by sharing the load, thereby weakening the causal effect from dynamic overload to local stress concentration. Therefore, the strategy of bonding carbon fiber cloth is encoded as a weakening operation on the side representing dynamic overload → local stress concentration. That is, in the digital twin simulation, the causal effect weight of this side is adjusted accordingly based on the composite section properties after carbon fiber cloth reinforcement. The physical function of applying external prestress is mainly to introduce reverse bending moment or pressure to actively offset the load effect. It may simultaneously weaken multiple links such as dynamic overload → local stress concentration and local stress concentration → microcrack initiation. The strategy of applying external prestress is encoded as a comprehensive modulation operation on multiple related sides involving load transfer and stress accumulation, that is, simultaneously adjusting the causal effect weight of multiple related sides.

[0091] Specifically, abstract maintenance engineering measures are transformed into precise interventions on identified, specific causal mechanisms. This makes simulation in digital twins no longer a simple change of a material parameter, but a simulation of targeted treatment of the causal chain. A hierarchical and interpretable mapping framework is established, from physical maintenance methods to causal network interventions, and then to the adjustment of digital twin model parameters. This not only makes the simulation closer to engineering reality, but more importantly, it allows for the prediction of the most effective intervention target (i.e., which causal path) for different maintenance strategies based on causal analysis before simulation. This enables the mechanistic and personalized design and evaluation of maintenance strategies. For example, if causal analysis shows that the propagation of a crack is mainly dominated by the stress amplitude caused by fatigue load, rather than the peak stress, then increasing stiffness (attaching carbon fiber cloth) may not be as effective as improving fatigue performance (grouting cracks to reduce the stress concentration factor). This insight can be reflected in the selection of coding strategies. Each maintenance strategy is successfully encoded as a targeted intervention operation on a specific high-sensitivity causal path in a probability-weighted causal network.

[0092] S4.3 Perform physical simulation of the digital twin model integrating the phase field method fracture model. By comparing the crack evolution results before and after the application of maintenance strategy intervention, statistically analyze the changes in crack bifurcation, turning and healing modes in the mesh, and construct a decision tree that reflects the probability of damage state transition under different interventions.

[0093] Furthermore, by comparing the simulation results under the two scenarios, the key pattern changes of crack morphology in the computational grid are statistically analyzed, including but not limited to whether the crack bifurcates, whether the dominant direction of crack propagation is deflected, and whether there is a local healing trend near the crack tip due to material strengthening or stress redistribution. Based on a large number of simulation comparisons, for each maintenance strategy intervention, the frequency of crack evolution results transitioning to different subsequent states (such as stable propagation, bifurcated propagation, deflected propagation, and propagation stagnation) under a specific initial damage state is statistically analyzed, and this frequency is normalized into a probability, thereby constructing a decision tree that reflects the damage state transition probability from the current state to various possible subsequent states under different interventions.

[0094] Specifically, deterministic physical simulation results are transformed into probabilistic state transition rules, constructing a stochastic process model that considers uncertainty and includes human intervention options. Through a large number of simulations, the probability distribution of the physical system's behavior under intervention is learned. For example, for the intervention of crack grouting, through one hundred simulations, it may be found that in a certain initial microcrack state, seventy simulations show that crack propagation is effectively suppressed (state transition to propagation stagnation), twenty simulations show that propagation slows down (stable propagation), and ten simulations show that the crack changes direction due to uneven grouting (directed propagation). The statistical data constitute the probability of transitioning from the microcrack state to each sub-state under grouting intervention in the decision tree. This makes the decision support not only based on physical mechanisms but also quantifies the uncertainty of the intervention effect, providing a richer basis for risk decision-making.

[0095] S4.4. Traverse the decision tree of damage state transition probability. By analyzing the path transition probability distribution from the initial damage state node to each terminal state node under different maintenance strategy intervention operation sequences, evaluate the expected intervention effect of crack grouting strategy on the main crack propagation length.

[0096] Furthermore, the analysis results of causal directed acyclic graphs are deeply integrated into the construction and reasoning process of the decision tree model, rather than simply concatenating them. High-order damage-sensitive feature vectors, as outputs of the causal directed acyclic graph, provide the decision tree with input features that have been causally filtered and have clear physical meaning. This reduces the huge computational burden and overfitting risk brought about by directly constructing the decision tree from massive, high-dimensional raw monitoring data streams. When constructing the decision tree for damage state transition probabilities, the high-order damage-sensitive feature vectors are obtained through preset mapping rules, ensuring that the starting point of the decision analysis has a solid causal data foundation. When selecting maintenance strategy intervention operations at each node of the decision tree, the optional actions (such as crack grouting and carbon fiber cloth bonding) are encoded as interventions on specific high-sensitivity causal paths in the probability-weighted causal network. This makes each branch of the decision tree correspond to the active regulation of the identified key causal mechanisms. The process of traversing the decision tree for damage state transition probabilities and evaluating the intervention effect is essentially simulating and quantifying the expected impact of different causal intervention strategies on the damage evolution outcome within a probability space guided by causal knowledge structure.

[0097] Specifically, it overcomes the limitations of traditional decision trees, which lack in-depth causal thinking and rely solely on statistical correlations. By introducing additional structured information generated by causal analysis, it enables complex decision analysis to be driven by interpretable causal logic. This ensures computational efficiency while allowing the evaluation of maintenance strategies to be based on understanding and changing the underlying mechanisms of causes, rather than merely associating phenomena.

[0098] S5. Based on real-time acquired multi-source data and impact load, perform physical simulation of damage evolution based on physical mechanisms, and output visualized pre-simulation data of damage evolution process.

[0099] S5.1 Acquire real-time multi-source data from dynamic weighing, meteorological monitoring and global positioning, as well as recorded overload or impact events as impact loads.

[0100] Furthermore, dynamic weighing sensors installed on the bridge deck continuously acquire axle load, total weight, and speed data of passing vehicles to form a real-time load spectrum; real-time data on temperature, humidity, wind speed, and wind direction are acquired through weather stations deployed at the bridge site; and real-time three-dimensional displacement data of key bridge points are acquired through GPS receivers. By analyzing abnormal peak values ​​in the dynamic weighing data or combining them with video surveillance recordings, overloaded vehicle incidents or sudden impact events such as vehicles colliding with guardrails can be identified and recorded.

[0101] Specifically, by using the time, location, and equivalent load amplitude of the event as the impact load input, a digital information flow reflecting the real service environment of the bridge in real time is constructed. This synchronously integrates the originally dispersed load, environment, and response information, providing high-fidelity boundary conditions and driving sources for physical simulation. It is not just about collecting data, but also about interpreting and encapsulating the data in an event-driven and condition-driven manner. The impact load is separated from ordinary load data and treated as an independent input event with instantaneous high-energy characteristics. This ensures that subsequent physical simulations can accurately simulate the transient response of the bridge under extreme or accidental actions and capture the damage triggering and evolution mechanisms that cannot be revealed by static load analysis.

[0102] S5.2 Input real-time multi-source data and impact loads into a digital twin model that integrates a phase field method fracture model, and perform physical simulation of damage evolution based on physical mechanisms.

[0103] Furthermore, temperature and humidity data from real-time multi-source data are used to update the temperature and humidity fields of material parameters in the digital twin model, thereby dynamically correcting the material's elastic modulus, strength, and other properties. Real-time load spectra provided by dynamic weighing data are applied as time-varying force boundary conditions to the corresponding locations in the digital twin model. Displacement calculated from global positioning data is used as displacement boundary conditions or for model verification. Impact load events are applied to the model as transient pulse forces or displacement boundary conditions acting at specific locations at specific times. Driven by all these real-time updated boundary conditions and material properties, the digital twin model integrating the phase-field fracture model obtains the evolution of stress, strain fields, and phase-field variables inside the structure by solving the coupled mechanical equilibrium equations and phase-field evolution equations, thereby simulating the damage evolution process such as crack initiation, propagation, bifurcation, or closure.

[0104] Specifically, it achieves real-time synchronization and interaction between the digital twin model and the physical world. Based on real-time monitoring data, it broadcasts the damage evolution of the structure under future loads and environmental effects, realizing online and dynamic prediction. It seamlessly and collaboratively embeds real-time multi-source data and impact loads into the phase-field method fracture model based on strict physical laws, forming a hybrid simulation paradigm where data drives boundary conditions and physical laws drive internal responses. For example, when an overweight vehicle is identified by the dynamic weighing system and used as an impact load input, the digital twin model will immediately apply a high-amplitude moving load at the corresponding lane and time.

[0105] S5.3 Extract data on the changes of crack length, propagation angle and stress intensity factor over time from the physical simulation of damage evolution based on physical mechanisms, and output them as visual pre-simulation data of the damage evolution process.

[0106] Furthermore, during the simulation, the geometric profile of the crack is determined by monitoring the isosurface of the phase field variables (e.g., taking the phase field value as a specific threshold). The maximum length of the crack, the length of the main branches, and the angle of the overall crack propagation direction relative to the initial defect are obtained as the propagation angle. Based on the stress field and displacement field near the crack tip obtained from the simulation, the Type I, Type II, and Type III stress intensity factors at the crack tip are calculated by methods such as interactive integration. The values ​​of physical quantities (crack length, propagation angle, and stress intensity factors of each type) at each simulation time step are recorded to form a time-varying sequence of data. The time-series data, together with information such as the structural geometry and load history in the simulation, are packaged and output as structured damage evolution process visualization pre-simulation data.

[0107] Specifically, the high-dimensional and complex full-field physical simulation results are extracted into the time history of several macroscopic damage indicators that are most critical and intuitive for engineering judgment. This realizes the transformation from physical field data to engineering decision-making information. The selected output indicators (crack length, angle, and stress intensity factor) are all core parameters in fracture mechanics for evaluating crack stability and predicting its propagation path. This allows the output simulation data to not only show how the damage looks (visualization), but also to quantitatively answer how severe the damage is and how it will develop in the future (quantifiable prediction). For example, whether the stress intensity factor exceeds the fracture toughness of the material over time can be directly used to determine whether the crack will become unstable and propagate, providing an accurate quantitative threshold for early warning.

[0108] S6. Activate the multi-level early warning mechanism to generate maintenance decision-making recommendation reports.

[0109] S6.1 Compare the visual simulation data of the damage evolution process with the pre-set warning level thresholds based on historical safety data, structural design specifications and real-time operating environment to determine the degree of deviation of crack propagation rate, path and stress intensity factor from the warning level thresholds and obtain the comparison results.

[0110] Furthermore, the preset warning level threshold is a multi-dimensional dynamic standard. It integrates the critical damage development value statistically derived from historical bridge accidents or long-term monitoring data, the limits such as material fracture toughness or allowable crack width specified in relevant structural design codes, and the floating coefficient adjusted according to the current real-time traffic flow, wind speed, and other operational environmental risk factors. The comparison process is as follows: from the visual simulation data of the damage evolution process, the rate of change of crack length over time is extracted as the crack propagation rate; the deflection angle of the crack front relative to the main force direction is extracted as the path deviation index; and the maximum value and trend of the stress intensity factor at the crack tip are extracted. The index is compared item by item with the corresponding warning level threshold, and the percentage or level of exceeding or approaching the threshold is calculated. For example, if the crack propagation rate exceeds the historically statistical safe propagation rate but does not reach the rapid propagation rate, and the stress intensity factor is still lower than the standard fracture toughness, it is judged as a slight deviation; if the propagation rate and path deviation both reach a high risk level, and the stress intensity factor is close to the material fracture toughness, it is judged as a serious deviation.

[0111] Specifically, it realizes quantitative, multi-index comprehensive risk assessment based on physical simulation prediction, transforming abstract damage evolution data into specific risk level judgments, providing a precise basis for graded response. The warning threshold is not a static, single value, but a triple dynamic standard that integrates historical experience, standard specifications, and real-time environment. For example, under severe weather conditions such as heavy rain or strong winds (high risk in real-time operating environment), even with the same crack propagation rate and stress intensity factor, the warning level triggered may be higher than under clear weather conditions, because environmental factors may accelerate damage or reduce the safety reserve of the structure.

[0112] S6.2. Based on the comparison results, trigger a multi-level early warning mechanism, including blue alert, yellow warning, orange alarm, and red emergency response.

[0113] Furthermore, the triggering rules are determined comprehensively based on the degree of deviation of indicators such as crack propagation rate, path, and stress intensity factor in the comparison results. Blue alerts usually correspond to indicators that have only shown initial abnormalities but have not exceeded the main safety thresholds. For example, the crack rate has increased slightly but is still at a historical low, or the stress intensity factor has started to rise but still has a large margin from the standard value. At this time, the warning mechanism may only trigger background log recording and trend monitoring. Yellow alerts correspond to indicators that have exceeded the low-level warning threshold, indicating a clear risk growth trend. For example, the crack propagation rate continues to exceed the safe propagation rate. The warning mechanism will send a notification to maintenance management personnel to remind them to strengthen manual inspections. Orange alerts correspond to indicators that have deviated severely and are approaching or have reached a high-risk level that may affect the short-term safety of the structure. For example, the stress intensity factor reaches a certain proportion of the standard fracture toughness, and the propagation path points to the critical stress-bearing parts. The warning mechanism will trigger a higher-level alarm and may recommend preparing emergency repair resources or implementing temporary traffic control. The red emergency response indicator has exceeded the safety threshold, indicating that the structure is facing the risk of immediate failure. For example, the stress intensity factor has exceeded the fracture toughness and the crack is accelerating. The early warning mechanism will immediately activate the highest level of emergency response, including automatically sending emergency notices, recommending immediate closure of traffic and activation of emergency plans.

[0114] Specifically, a tiered response system was established that precisely matches the severity of risks. This avoids the frequent false alarms or slow response that may result from single-threshold warnings. The classification of warning levels is directly linked to the physical state and engineering consequences of damage evolution and mapped to different management response actions. For example, a yellow warning corresponds to the engineering meaning that damage is developing and requires planned intervention, thus triggering notification and preparation actions. On the other hand, a red emergency means that instability is imminent and immediate rescue is required, thus triggering the most urgent actions. This makes the warning information not only an alarm signal but also a clear action instruction, greatly improving the practicality of warnings and decision-making efficiency.

[0115] S6.3. Based on the visualized pre-simulation data of the damage evolution process and the multi-level early warning mechanism, generate a maintenance decision recommendation report.

[0116] Furthermore, the report will clearly define the current warning level and its main basis. For example, it will specify that the yellow warning was triggered because the crack propagation rate at the lower edge of the main beam exceeded the yellow warning threshold. The report will cite key prediction results from the visual simulation data of the damage evolution process in detail, such as showing the predicted crack length curve and the stress intensity factor over time, and marking the predicted time point when the next warning level will be reached. Based on the analysis results of the decision tree of the damage state transition probability, the report will recommend one or more maintenance strategies that are technically and economically optimized. For example, for the current microcrack state, the decision tree analysis may indicate that crack grouting has the best overall effect in terms of suppressing propagation and cost, while bonding carbon fiber cloth is better in terms of long-term durability. The report will clearly list this analysis conclusion. Combining the multi-level warning mechanism, the report will give specific action suggestions, time windows, and resource preparation suggestions. For example, for a yellow warning, it may be recommended to arrange crack grouting and sealing treatment within the next two weeks; for an orange warning, it may be recommended to complete temporary reinforcement and start special testing within 48 hours.

[0117] Specifically, the complex monitoring data, simulation predictions, and risk analysis results are transformed into a clear, actionable engineering decision document with explicit recommendations. This completes the value chain loop from data to insight to action plan. The maintenance decision recommendation report is not simply a forwarding of early warning information, but a crystallization of wisdom that deeply integrates causal analysis, physical simulation, and decision tree evaluation. It explains where the risk comes from, how it will develop, and provides comprehensive decision support on what to do, when to do it, and what the expected results are. For example, the report may point out that the current cracks are mainly caused by a specific overload pattern, so the recommended maintenance measures should focus on improving the local bending resistance of that part, while supplementing it with recommendations for overload management. This is a comprehensive recommendation that goes from treating the symptoms (repairing cracks) to addressing the root cause (managing loads).

[0118] This embodiment also provides a bridge structure fault prediction system based on monitoring data, including: a feature extraction module, which acquires the original monitoring data stream and preprocesses it, and extracts features based on the preprocessed data to obtain a standardized feature dataset;

[0119] The module constructs a causal directed acyclic graph between sensor nodes by inputting a standardized feature dataset into a causal feature mining framework guided by physical information.

[0120] The screening module, based on a causal directed acyclic graph, uses the causal intervention effect to quantify the contribution of different factors to the structural response and screens out high-order damage-sensitive feature vectors.

[0121] The evaluation module inputs high-order damage-sensitive feature vectors into a digital twin model that integrates a phase-field fracture model. By constructing a decision tree of damage state transition probability, it evaluates the intervention effect of different maintenance strategies on crack propagation paths.

[0122] The recommended report module performs physical simulation of damage evolution based on real-time acquired multi-source data and impact loads, outputs visualized pre-simulation data of the damage evolution process, and activates a multi-level early warning mechanism to generate maintenance decision recommendation reports.

[0123] This embodiment also provides a computer device applicable to the bridge structure fault prediction method based on monitoring data, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the bridge structure fault prediction method based on monitoring data as proposed in the above embodiment.

[0124] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0125] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the bridge structure fault prediction method based on monitoring data as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0126] In summary, this invention collects and fuses multi-dimensional data such as stress, vibration, and temperature from multiple sensors to construct a standardized feature dataset. Utilizing a causal inference framework that integrates constraints from physical mechanics equations, it mines the causal relationship network between load, environment, and structural response from the data, and quantifies the independent and combined effects of each factor on structural damage. This allows for the selection of high-order features that are highly sensitive to damage, which are then input into a digital twin integrating a phase-field fracture mechanics model. The twin not only simulates the evolution of cracks under real-time loads and environments but also quantifies the effectiveness of different maintenance strategies in suppressing crack development by constructing a probabilistic decision tree. Based on simulation prediction results and dynamic risk thresholds, it generates graded early warning and specific maintenance decision reports.

[0127] It should be noted that 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting bridge structural failures based on monitoring data, characterized in that: This includes acquiring and preprocessing the raw monitoring data stream, extracting features based on the preprocessed data, and obtaining a standardized feature dataset. The standardized feature dataset is input into a causal feature mining framework guided by physical information to construct a causal directed acyclic graph between sensor nodes. Based on causal directed acyclic graphs, the contribution of different factors to structural response is quantified by causal intervention effects, and high-order damage-sensitive feature vectors are screened out. By inputting high-order damage-sensitive feature vectors into a digital twin model that integrates a phase-field fracture model, and constructing a decision tree of damage state transition probability, the intervention effect of different maintenance strategies on crack propagation path is evaluated. Based on real-time acquired multi-source data and impact loads, physical simulation of damage evolution is performed to output visualized pre-simulation data of the damage evolution process, and a multi-level early warning mechanism is activated to generate maintenance decision-making recommendation reports.

2. The bridge structure fault prediction method based on monitoring data as described in claim 1, characterized in that: The process involves acquiring and preprocessing the raw monitoring data stream, extracting features from the preprocessed data, and obtaining a standardized feature dataset. This includes the following steps: The raw monitoring data stream containing stress, vibration, displacement, temperature and images is acquired. Preprocessing operations including missing value imputation, outlier removal and low-pass filtering are performed on the raw monitoring data stream to obtain the preprocessed data stream. Feature extraction is performed on the preprocessed data stream to extract the first three natural frequencies from the vibration component and the mean and variance from the strain component, thus obtaining a standardized feature dataset.

3. The bridge structure fault prediction method based on monitoring data as described in claim 2, characterized in that: The standardized feature dataset is input into a causal feature mining framework guided by physical information to construct a causal directed acyclic graph between sensor nodes, including the following steps: The standardized feature dataset is input into a causal feature mining framework guided by physical information, and structural vibration differential equation constraints are embedded. In the causal feature mining framework guided by physical information, the PC algorithm is used to perform causal discovery analysis on a standardized feature dataset; Based on the results of causal discovery analysis, a causal directed acyclic graph is constructed among the sensor nodes of temperature node, load node, strain node, frequency node, and damage state node.

4. The bridge structure fault prediction method based on monitoring data as described in claim 3, characterized in that: Based on causal directed acyclic graphs, the contribution of different factors to the structural response is quantified using causal intervention effects, and high-order damage-sensitive feature vectors are screened out, including the following steps: Starting from the causal directed acyclic graph, the temperature factor, load factor and fatigue accumulation factor are counterfactually intervened through the structural nesting model to remove the confounding effects between the factors and quantify the independent contribution of each factor to the structural response factor. The causal directed acyclic graph is weighted using independent contribution values, and a random walk algorithm based on Monte Carlo simulation is executed to calculate the average first-hit probability from each input factor node to the damage state node, thus constructing a probability-weighted causal network. The average first-hit probability of each path in the probability-weighted causal network is adaptively convolved with the real-time operational risk level to generate a dynamic causal sensitivity threshold that matches the current risk status. Based on the dynamic causal sensitivity threshold, high-sensitivity causal paths are selected in the probability-weighted causal network, and the corresponding features in the standardized feature dataset mapped by the path are extracted to form a high-order damage-sensitive feature vector.

5. The bridge structure fault prediction method based on monitoring data as described in claim 4, characterized in that: The high-order damage-sensitive feature vector is input into a digital twin model integrating a phase-field fracture model. By constructing a decision tree for damage state transition probabilities, the intervention effect of different maintenance strategies on crack propagation paths is evaluated, including the following steps: Based on the average first-hit probability, adjust the initial distribution of the corresponding physical parameters in the phase-field fracture model; In a digital twin model integrating a phase-field fracture model, crack grouting, carbon fiber cloth bonding, and external prestressing are defined, and each maintenance strategy is encoded as an intervention operation on a specific high-sensitivity causal path in a probability-weighted causal network. Physical simulation of a digital twin model integrating a phase-field fracture model is performed. By comparing the crack evolution results before and after the application of maintenance strategy intervention, the changes in crack bifurcation, turning and healing patterns in the mesh are statistically analyzed, and a decision tree reflecting the probability of damage state transition under different interventions is constructed. By traversing the decision tree of damage state transition probability, and analyzing the path transition probability distribution from the initial damage state node to each terminal state node under different maintenance strategy intervention operation sequences, the expected intervention effect of crack grouting strategy on the main crack propagation length is evaluated.

6. The bridge structure fault prediction method based on monitoring data as described in claim 5, characterized in that: Physical simulation of damage evolution based on real-time acquired multi-source data and impact load is performed to output visualized pre-simulation data of the damage evolution process, including the following steps: Real-time vehicle weight information, real-time temperature and humidity information, and real-time traffic flow information are acquired as multi-source data to identify and record impact load information; Real-time multi-source data and impact loads are input into a digital twin model that integrates a phase field method fracture model to perform physical simulation of damage evolution based on physical mechanisms. From the physical simulation of damage evolution based on physical mechanisms, data on the changes of crack length, propagation angle, and stress intensity factor over time are extracted and output as visual pre-simulation data of the damage evolution process.

7. The bridge structure fault prediction method based on monitoring data as described in claim 6, characterized in that: Activating a multi-level early warning mechanism to generate maintenance decision recommendation reports includes the following steps: The damage evolution process visualization simulation data is compared with the warning level thresholds preset based on historical safety data, structural design specifications and real-time operating environment to determine the degree of deviation of crack propagation rate, path and stress intensity factor from the warning level thresholds and obtain the comparison results. Based on the comparison results, a multi-level early warning mechanism, including blue alert, yellow warning, orange alarm, and red emergency response, is triggered. Based on the visualized pre-simulation data of the damage evolution process and the multi-level early warning mechanism, a maintenance decision recommendation report is generated.

8. A bridge structure fault prediction system based on monitoring data, based on the bridge structure fault prediction method based on monitoring data according to any one of claims 1 to 7, characterized in that: This includes a feature extraction module, which acquires and preprocesses the raw monitoring data stream, extracts features based on the preprocessed data, and obtains a standardized feature dataset. The module constructs a causal directed acyclic graph between sensor nodes by inputting a standardized feature dataset into a causal feature mining framework guided by physical information. The screening module, based on a causal directed acyclic graph, uses the causal intervention effect to quantify the contribution of different factors to the structural response and screens out high-order damage-sensitive feature vectors. The evaluation module inputs high-order damage-sensitive feature vectors into a digital twin model that integrates a phase-field fracture model. By constructing a decision tree of damage state transition probability, it evaluates the intervention effect of different maintenance strategies on crack propagation paths. The recommended report module performs physical simulation of damage evolution based on real-time acquired multi-source data and impact loads, outputs visualized pre-simulation data of the damage evolution process, and activates a multi-level early warning mechanism to generate maintenance decision recommendation reports.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the bridge structure fault prediction method based on monitoring data as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the bridge structure fault prediction method based on monitoring data as described in any one of claims 1 to 7.

Images