Predictive maintenance method for fatigue performance of bridge structures based on digital twins

Abstract

The present invention relates to the technical field of bridge structure health monitoring and maintenance, and discloses a predictive maintenance method for bridge structure fatigue resistance based on digital twins. The method comprises: constructing a digital twin model of the bridge structure and deploying a multimodal sensor network; establishing a knowledge graph of bridge fatigue mechanisms to structure expert experience and historical cases; integrating multi-source heterogeneous evidence based on DS evidence theory to resolve evidence conflicts; constructing a causal reasoning engine based on a Bayesian network to identify fatigue mechanism paths; and implementing an adaptive predictive maintenance decision system to generate an optimal maintenance strategy. The present invention addresses the problems of traditional bridge fatigue analysis methods, such as difficulty in processing multi-source heterogeneous data and insufficient fatigue prediction accuracy, and achieves high-precision prediction and precise maintenance of bridge structure fatigue resistance.

Description

Technical Field

[0001] The present invention relates to the technical field of bridge structure health monitoring and maintenance, and more specifically, to a predictive maintenance method for the fatigue resistance of bridge structures based on digital twins. Background Art

[0002] As a vital component of transportation infrastructure, bridge structures are constantly subjected to cyclical forces such as vehicle loads, wind loads, and temperature fluctuations during their long service life. This can lead to accumulated fatigue damage within the structure, which can ultimately cause degradation or even failure. According to statistics, fatigue damage has become one of the main causes of failure in large bridge structures, seriously threatening their safety and durability.

[0003] Existing bridge fatigue analysis and maintenance methods primarily rely on regular inspections, empirical judgment, and passive repairs. These methods typically rely on a single type of evidence or simple correlation analysis. With the development of sensor technology and data analysis techniques, some studies have begun to apply methods such as finite element analysis and machine learning to fatigue performance assessment. However, these methods still have the following problems: It is difficult to effectively integrate multi-source heterogeneous data, and it is impossible to comprehensively consider multi-dimensional evidence such as sensor data, structural response, environmental conditions, expert experience, and historical cases, resulting in inaccurate fatigue mechanism identification; expert knowledge is fragmented, lacking a systematic knowledge organization and utilization mechanism, making it difficult to support decision-making in complex scenarios; conflicts often exist between different sources of evidence, and there is a lack of effective conflict resolution methods, which reduces the reliability of judgments; fatigue predictions are mostly based on statistical models, which make it difficult to reveal the underlying causal mechanisms and result in insufficient prediction accuracy; maintenance decisions lack a systematic optimization mechanism and rely heavily on manual experience, resulting in high maintenance costs and poor results.

[0004] Therefore, there is an urgent need to develop a new method that can comprehensively utilize multi-source heterogeneous data, integrate expert experience, accurately predict fatigue performance and optimize maintenance decisions, so as to provide reliable guarantees for the safe operation of bridge structures and reduce maintenance costs. Summary of the Invention

[0005] The present invention provides a predictive maintenance method for the fatigue resistance of bridge structures based on digital twins, which solves the technical problems in the existing technology that bridge fatigue analysis is difficult to effectively process multi-source heterogeneous data, fatigue prediction accuracy is insufficient, and maintenance decision-making lacks an optimization mechanism.

[0006] The present invention provides a predictive maintenance method for fatigue resistance of bridge structures based on digital twins, comprising the following steps:

[0007] Build a digital twin model of the bridge structure and deploy a multimodal sensor network to achieve real-time data interaction between the physical structure and the virtual model;

[0008] Establish a knowledge map of bridge fatigue mechanisms, transform expert experience and historical cases into structured knowledge, and form a fatigue mechanism knowledge base;

[0009] Based on DS evidence theory, it realizes the fusion of multi-source heterogeneous evidence, eliminates the conflict between evidence from different sources, and improves the accuracy of fatigue mechanism identification;

[0010] Build a causal inference engine based on Bayesian networks to identify the causal path of fatigue damage development;

[0011] Implement an adaptive predictive maintenance decision system to generate the optimal maintenance strategy based on fatigue state prediction results.

[0012] Furthermore, the construction of the bridge structure digital twin model includes:

[0013] Use parametric modeling technology to construct a three-dimensional finite element model of the bridge;

[0014] Deploy multimodal sensing networks at key bridge locations, including strain sensors, accelerometers, and displacement sensors;

[0015] Establish a mapping relationship between sensor data and digital models, and use the Bayesian calibration method to optimize the parameters of the finite element model;

[0016] Build a two-way data channel based on edge computing technology to achieve real-time interaction between physical space and virtual space.

[0017] Furthermore, the establishment of the bridge fatigue mechanism knowledge graph includes:

[0018] Define the core concepts, attributes and relationships in the field of bridge fatigue based on ontology method;

[0019] Extract structured knowledge from fatigue research literature, maintenance reports, and expert knowledge through natural language processing technology;

[0020] Construct knowledge reasoning rules to realize automatic reasoning and expansion of knowledge;

[0021] Establish a case library and classify and store historical fatigue cases according to failure modes, causes, and performance characteristics.

[0022] Furthermore, the calculation of the evidence conflict coefficient in the fusion of multi-source heterogeneous evidence based on the DS evidence theory is obtained by summing the products of the basic probability assignments between all conflicting evidence sources, where mutual conflict refers to the situation where the intersection of propositions supported by different evidence sources is an empty set; the larger the coefficient value, the more serious the conflict between different evidence sources, and the more severe the need for stronger evidence adjustment.

[0023] Furthermore, the evidence weight adjustment in the fusion of multi-source heterogeneous evidence based on the DS evidence theory is achieved by calculating the average distance between each evidence source and other evidence sources, and converting the distance value into a weight value, so that the evidence with smaller differences from other evidence sources obtains a higher weight.

[0024] Furthermore, the evidence fusion in the multi-source heterogeneous evidence fusion based on the DS evidence theory is achieved by calculating the sum of the probability products of different evidence combinations supporting the same proposition and performing normalization processing to ensure that the final probability distribution meets the constraint condition that the sum of the probabilities is 1.

[0025] Furthermore, the construction of a causal inference engine based on a Bayesian network includes:

[0026] Construct an initial Bayesian network structure based on the causal relationship in the knowledge graph;

[0027] Combine monitoring data and expert knowledge to learn the conditional probability table of the Bayesian network;

[0028] Construct a posterior probability inference algorithm based on MCMC method;

[0029] Create a counterfactual intervention analysis module to calculate the intervention effects of different maintenance measures;

[0030] Create a fatigue mechanism path identification algorithm to find the most likely fatigue development path.

[0031] Furthermore, the intervention effect calculation of the counterfactual intervention analysis module is performed by taking a weighted average of the conditional probabilities under all possible covariate values, where the weights are the marginal distribution probabilities of the covariates, thereby evaluating the causal impact of specific intervention measures on the outcome variables.

[0032] Furthermore, the implementation of the adaptive predictive maintenance decision system includes:

[0033] Establish a fatigue damage prediction model based on LSTM neural network;

[0034] Define the maintenance decision objective function that comprehensively considers safety and reliability, maintenance cost and structural life;

[0035] Build a maintenance strategy optimization algorithm based on reinforcement learning;

[0036] Create a maintenance effectiveness evaluation model;

[0037] Establish a visual interface for maintenance decision-making.

[0038] Furthermore, the bridge structure fatigue performance predictive maintenance system based on digital twins implements the above-mentioned bridge structure fatigue performance predictive maintenance method based on digital twins, including:

[0039] Digital twin model building module for digital modeling of bridge structures and deployment of multimodal sensor networks;

[0040] Fatigue knowledge graph construction module, which is used to integrate expert experience and historical cases to form a structured fatigue mechanism knowledge base;

[0041] Multi-source evidence fusion module, used to resolve conflicts and fuse fatigue evidence from different sources based on DS evidence theory;

[0042] A causal reasoning engine module, which applies Bayesian networks and causal analysis techniques to identify potential fatigue mechanism pathways;

[0043] An adaptive predictive maintenance decision module is used to generate the optimal maintenance strategy through multi-objective optimization techniques.

[0044] The beneficial effects of the present invention are:

[0045] This solution achieves high-precision monitoring and simulation of the real-time status of bridges by constructing a digital twin model of the bridge structure, providing a data basis for fatigue performance assessment; establishes a knowledge graph of bridge fatigue mechanisms, realizes expert knowledge reuse and automated reasoning, and provides knowledge support for fatigue mechanism analysis; uses a multi-source heterogeneous evidence fusion mechanism based on DS evidence theory to resolve evidence conflicts and collaboratively analyze multi-dimensional evidence; uses a causal reasoning engine based on a Bayesian network to reveal the causal relationship of fatigue damage, providing a mechanism support for precise maintenance; creates an adaptive predictive maintenance decision system, generates the optimal maintenance strategy based on fatigue status prediction, and realizes the transition from passive maintenance to active prevention. In addition, the improved DS evidence theory can dynamically adjust the weight of evidence and improve the reliability of evidence fusion; through the Bayesian network and LSTM neural network, the accuracy of bridge structure fatigue status prediction is improved; and the maintenance strategy optimization based on reinforcement learning can adaptively adjust maintenance decisions, which reduces costs and improves structural reliability compared to traditional regular maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a flow chart of the predictive maintenance method for fatigue resistance of bridge structures based on digital twins of the present invention. DETAILED DESCRIPTION

[0047] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described in some examples may be combined in other examples.

[0048] At least one embodiment of the present invention discloses a method for predictive maintenance of fatigue resistance of bridge structures based on digital twins, such as Figure 1 As shown, the following steps are included:

[0049] Step 1: Build a digital twin model of the bridge structure and deploy a multimodal sensor network to achieve real-time data interaction between the physical structure and the virtual model;

[0050] Step 1.1: Use parametric modeling technology to construct a three-dimensional finite element model of the bridge based on the geometric parameters, material properties, and boundary conditions of the bridge structure. This model can accurately reflect the mechanical properties of the bridge structure.

[0051] Step 1.2: Deploy a multimodal sensor network at key locations on the bridge, including strain sensors, accelerometers, and displacement sensors, to enable real-time monitoring of the dynamic response of the bridge structure.

[0052] Step 1.3: Establish a mapping relationship between sensor data and digital models, and use the Bayesian calibration method to optimize the parameters of the finite element model so that the digital model can accurately reflect the dynamic characteristics of the actual structure;

[0053] Step 1.4: Configure data preprocessing algorithms, including outlier detection, missing value interpolation, and data denoising, to ensure the quality of acquired monitoring data.

[0054] In step 1.5, a two-way data channel is constructed between the digital twin model and the physical bridge to achieve real-time interaction between the physical space and the virtual space. This channel is based on edge computing technology to ensure the real-time and reliability of data transmission.

[0055] Step 2: Establish a knowledge graph of bridge fatigue mechanisms, transform expert experience and historical cases into structured knowledge, and form a fatigue mechanism knowledge base;

[0056] Step 2.1: Define the core concepts, attributes, and relationships in the field of bridge fatigue based on the ontology approach and establish a conceptual model of fatigue knowledge, including core elements such as material properties, environmental factors, load types, and fatigue damage modes;

[0057] Step 2.2: extract structured knowledge from fatigue research literature, maintenance reports, and expert knowledge using natural language processing technology, and convert it into triples (subject-relation-object) for storage;

[0058] Step 2.3: Construct knowledge reasoning rules based on description logic and first-order predicate logic to achieve automated reasoning and expansion of knowledge and enhance the expressive power of the knowledge graph.

[0059] Step 2.4: Create a knowledge fusion algorithm to resolve conflicts and overlaps between knowledge from different sources and ensure the consistency and integrity of the knowledge base.

[0060] Step 2.5: Establish a case library, classify and store historical fatigue cases according to failure modes, causes, and performance characteristics, and form a searchable case knowledge base to support case-based reasoning.

[0061] Step 3: Based on the DS evidence theory, multi-source heterogeneous evidence is integrated to resolve conflicts between evidence from different sources and improve the accuracy of fatigue mechanism identification.

[0062] Step 3.1: Establish a recognition framework for bridge fatigue identification ,in, 、 、 Respectively represent 、 、 Possible fatigue mechanisms, Indicates the Possible fatigue mechanisms, is the total number of fatigue mechanisms;

[0063] In step 3.2, a basic probability distribution function (BPA) is assigned to evidence from different sources (such as sensor data, expert knowledge, historical cases, etc.), which is expressed as ,in, Indicates the a source of evidence;

[0064] Step 3.3: Construct an evidence conflict detection algorithm and calculate the evidence conflict coefficient :

[0065] ;

[0066] in, represents the evidence conflict coefficient, represents the i-th proposition supported by the first source of evidence, represents the j-th proposition supported by the second source of evidence, Indicates the first source of evidence for the proposition The basic probability assignment of Indicates that the second source of evidence for the proposition The basic probability assignment of Indicates that the intersection of two propositions is an empty set, that is, the two propositions conflict with each other;

[0067] Step 3.4: Design an adaptive weight distribution model based on evidence distance to adjust the weights of conflicting evidence:

[0068] ;

[0069] in, Indicates the The weight of each source of evidence, Indicates the The average distance between a source of evidence and other sources of evidence, Shidi The average distance between a source of evidence and other sources of evidence, is the total number of sources of evidence;

[0070] Step 3.5: Apply the improved DS combination rule to perform evidence fusion and obtain the comprehensive BPA function :

[0071] ;

[0072] in, is a comprehensive BPA function, for any non-empty subset represents any non-empty subset in the recognition frame, is the evidence conflict coefficient, Indicates that the intersection of two propositions is equal to the proposition , represents the i-th proposition supported by the first source of evidence, represents the j-th proposition supported by the second source of evidence, Indicates the first source of evidence for the proposition The basic probability assignment of Indicates that the second source of evidence for the proposition The basic probability assignment of represents the normalization factor, ensuring that the final probability distribution satisfies the constraint that the sum of the probabilities is 1;

[0073] Taking fatigue damage identification in the support area of a bridge structure as an example, the specific application scenarios are as follows:

[0074] Identification framework construction: First, based on the typical mechanism of bridge fatigue damage, an identification framework is established ,in, Indicates fatigue cracks, Indicates bearing corrosion, Indicates material aging, Indicates welding defects, Indicates overload damage;

[0075] Evidence source allocation BPA: The system obtains data from three different evidence sources, namely: sensor monitoring data (denoted as ), expert evaluation (denoted as ) and historical case matching (denoted as Assume that sensor data indicates high-frequency micro-vibration characteristics in the support area. Experts assess that fatigue cracks or welding defects may exist in this area. Historical case matching indicates that similar cases are mostly caused by welding defects. Based on this information, the BPA of each evidence source is assigned as follows:

[0076] ;

[0077] ;

[0078] ;

[0079] Evidence conflict detection: Calculate the evidence conflict coefficient. and For example, the conflict coefficient between them is:

[0080] ;

[0081] in, express and The conflict coefficient between represents the empty set;

[0082] In this case, since there is no direct conflict between the sources of evidence (i.e. and The focal element intersection of the assigned probabilities is not empty), so However, in practice, if a source of evidence indicates that the fault type is , while the other shows that it cannot be , then conflicts arise.

[0083] Weight distribution: Calculate the weight of each evidence source based on the evidence distance. The evidence distance can be calculated using the Josselme distance formula:

[0084] ;

[0085] in, is a similarity matrix associated with the recognition frame, Represents a transpose operation.

[0086] Assume that the calculated average distance between the evidence sources is: , , , then the weight of each source of evidence is:

[0087] ;

[0088] ;

[0089] ;

[0090] Evidence fusion: Finally, the improved DS combination rule is used to fuse the various evidence sources:

[0091] ;

[0092] in, Indicates the BPA parameters after fusion, Represents a subset in the recognition framework;

[0093] For this example, the fused result might be:

[0094] ;

[0095] ;

[0096] ;

[0097] ;

[0098] This indicates a welding defect ( ) is the most likely cause of fatigue injury. Compared with the judgment of each single evidence source, the fusion result is more comprehensive and reliable.

[0099] Step 4: Build a causal reasoning engine based on Bayesian networks to identify the causal path of fatigue damage development;

[0100] Step 4.1: Based on the causal relationships in the knowledge graph, construct the initial Bayesian network structure and define nodes (variables) and edges (dependencies);

[0101] Step 4.2, combining monitoring data and expert knowledge, learn the conditional probability table (CPT) of the Bayesian network, expressed as ,in, is a node variable, yes The parent node set of

[0102] Step 4.3: Construct a posterior probability inference algorithm based on the MCMC (Markov Chain Monte Carlo) method to calculate the probability of fatigue mechanism given the observational evidence;

[0103] Step 4.4, create a counterfactual intervention analysis module and calculate the intervention effect through do-calculus:

[0104] ;

[0105] in, Indicates that the intervention variable The value is When the result variable The conditional probability of is the value of the intervention variable, is a covariate, is the intervention variable, is the outcome variable, is the set of covariates;

[0106] In step 4.5, a fatigue mechanism path identification algorithm is created to find the most likely fatigue development path based on the maximum a posteriori probability (MAP) criterion, providing a mechanism explanation for predictive maintenance.

[0107] Taking the causal analysis of fatigue damage of bridge main beams as an example, the specific application process is as follows:

[0108] Bayesian network structure construction: In the main beam fatigue analysis scenario, first extract relevant nodes and relationships from the knowledge graph and construct the following Bayesian network structure:

[0109] Root node (no parent node): load type (LT), environmental condition (EC), material quality (MQ);

[0110] Intermediate nodes: stress concentration (SC), corrosion degree (CD), structural response (SR);

[0111] Leaf node (result node): fatigue crack (FC);

[0112] The dependencies between these nodes are reflected as directed edges, for example: load type → stress concentration, environmental conditions → corrosion degree, material quality → stress concentration, stress concentration → structural response, corrosion degree → structural response, structural response → fatigue crack, etc.

[0113] Conditional probability table learning: Use historical data and expert knowledge collected from the digital twin model to calculate the conditional probability table of each node. For example, for the "Structural Response" node, its conditional probability table is It can be expressed as shown in Table 1:

[0114] Table 1: Conditional probability table :

[0115]

[0116] Posterior probability inference: When the digital twin model monitors the observed values of certain variables, the system uses the MCMC method to perform posterior probability inference. For example, if the sensor monitors: load type = heavy vehicle repeated passage (LT = H), environmental conditions = humidity (EC = W), structural response = high frequency vibration (SR = H), the probability of fatigue crack occurrence needs to be inferred. .

[0117] The MCMC sampling process is as follows:

[0118] Initialize the Markov chain: randomly assign initial values to unobserved variables;

[0119] Perform Gibbs sampling on each unobserved variable and update the state according to its conditional probability distribution;

[0120] Repeat the sampling steps multiple times (e.g., 10,000 times) to collect samples;

[0121] Statistically analyze the state distribution of fatigue crack variables in the sample and obtain the posterior probability;

[0122] After calculation, the posterior probability may be:

[0123] ,

[0124] It shows that under the given observation conditions, the probability of fatigue crack occurrence is 78%.

[0125] Counterfactual intervention analysis: To evaluate the effects of different maintenance measures, the system performs counterfactual intervention analysis. For example, suppose that in the current state, the system estimates that the probability of fatigue cracking is 78%. Now we want to evaluate the effect of increasing structural reinforcement (reducing stress concentration). Using do-calculus calculation:

[0126] ;

[0127] in, is the conditional probability after the intervention, The distribution before the intervention remains unchanged. After calculation, it is possible to obtain , indicating that by reducing stress concentration to a low level through reinforcement measures, the probability of fatigue crack occurrence can be reduced from 78% to 35%.

[0128] Fatigue mechanism path identification: Based on the constructed Bayesian network and the calculated conditional probabilities of each node, the system uses the maximum a posteriori probability criterion to identify the most likely fatigue development path. In the above example, the possible maximum a posteriori probability path is:

[0129] Repeated heavy vehicle traffic (LT=H) → high stress concentration (SC=H) → high structural response (SR=H) → fatigue cracking (FC=Y)

[0130] The joint probability of this path is the highest among all possible paths, providing the most likely explanation for the occurrence of fatigue damage and providing a basis for targeted maintenance decisions.

[0131] Step 5: Implement an adaptive predictive maintenance decision system to generate the optimal maintenance strategy based on fatigue state prediction results;

[0132] Step 5.1: Based on the integrated fatigue evidence and causal reasoning results, a fatigue damage prediction model is established, and deep learning methods (such as LSTM networks) are used to predict the fatigue damage trend of key parts;

[0133] Step 5.2, define the maintenance decision objective function, taking into account safety and reliability, maintenance cost and structural life:

[0134] ;

[0135] in, represents the objective function, represents the reliability index, represents the cost indicator, Represents the lifespan indicator, 、 、 are the corresponding weight coefficients respectively;

[0136] Step 5.3: Build a maintenance strategy optimization algorithm based on reinforcement learning, define the state space, action space, and reward function, and learn the optimal maintenance strategy through interaction with the environment;

[0137] Step 5.4: Create a maintenance effect evaluation model, simulate the effects of different maintenance plans based on the digital twin model, and update the maintenance strategy in real time;

[0138] Step 5.5: Establish a maintenance decision visualization interface to intuitively display fatigue status, predicted trends, and recommended maintenance plans to support engineers in making decisions.

[0139] Taking the fatigue maintenance of the main cable joint area of a large suspension bridge as an example, the specific application process is as follows:

[0140] Fatigue damage prediction model construction: Based on the real-time monitoring data and historical data collected by the digital twin model, an LSTM neural network model is established to predict the fatigue damage evolution in the main cable joint area. The LSTM network structure includes:

[0141] Input layer: receives various sensor data (strain, displacement, temperature, humidity, etc.) and traffic load information;

[0142] LSTM layer 1: contains 64 neurons, with tanh activation function;

[0143] Dropout layer: randomly discard 20% of neurons to prevent overfitting;

[0144] LSTM layer 2: contains 32 neurons, with tanh activation function;

[0145] Fully connected layer: maps the LSTM layer output to the prediction target;

[0146] Output layer: predict fatigue damage indicators at different time points in the future (1 day, 7 days, 30 days, 90 days);

[0147] The model was trained using 90% of the historical data as a training set and 10% as a validation set. The Adam optimizer was used with a learning rate of 0.001, a batch size of 64, and 500 training epochs. The trained model achieved a mean absolute error (MAE) of less than 5% on the validation set, demonstrating its ability to accurately predict fatigue damage trends.

[0148] Definition of maintenance decision objective function: Based on the bridge operation and management requirements, define the various indicators of the objective function:

[0149] Reliability Index : Based on the structural reliability index, the range is [0,1], and the larger the value, the safer the structure;

[0150] Cost indicators : Normalized representation of maintenance cost, ranging from [0,1], where smaller values indicate lower costs;

[0151] Lifespan Index : Normalized representation of remaining useful life, range [0,1], larger values indicate longer useful life;

[0152] The weight coefficient of each indicator is determined according to the importance of the bridge and the management strategy, for example: (Safety is the primary consideration), 、 .

[0153] Maintenance strategy optimization algorithm implementation: Implement reinforcement learning algorithm based on deep Q network (DQN), where:

[0154] State space: includes multi-dimensional state vectors such as fatigue damage degree, damage development rate, remaining life of the structure, and last maintenance time;

[0155] Action space: includes different maintenance strategies, such as no operation, inspection and monitoring, local reinforcement, replacement of important components, and comprehensive repair;

[0156] Reward function: based on the objective function Calculation, that is, the objective function value corresponding to the state after performing a certain action.

[0157] The Q network structure includes:

[0158] Input layer: receives the state vector (dimension is the size of the state space);

[0159] Hidden layer 1: 128 neurons, ReLU activation function;

[0160] Hidden layer 2: 64 neurons, ReLU activation function;

[0161] Output layer: the size of the action space and outputs the Q value of each action;

[0162] The training process uses an experience replay mechanism and a target network to improve learning stability. After 10,000 rounds of training, the Q network is able to recommend the optimal maintenance strategy based on the bridge status.

[0163] Specific application case: Decision-making process after detecting signs of fatigue damage in the main cable joint area:

[0164] The digital twin model collects strain and displacement data of the joint area in real time through sensors.

[0165] The DS evidence theory integrates sensor data, inspection records and historical cases to determine that the fatigue mechanism is "stress concentration caused by local corrosion of steel wire" with a confidence level of 83%.

[0166] The Bayesian causal network identified the fatigue development path: environmental corrosion (EC=H) → material degradation (MD=M) → stress concentration (SC=H) → fatigue crack development (FC=Y).

[0167] The LSTM prediction model predicts that if no measures are taken, the fatigue damage index will increase from the current 0.32 to 0.47 within 30 days, and may reach 0.68 within 90 days (exceeding the warning value of 0.65).

[0168] The reinforcement learning algorithm evaluates the Q-values of various maintenance options based on the current state:

[0169] No operation: Q value = -25.6;

[0170] Continue monitoring: Q value = -12.3;

[0171] Local anti-corrosion treatment: Q value = 18.5;

[0172] Cable clip replacement: Q value = 15.2;

[0173] Main cable section replacement: Q value = -8.7;

[0174] The system recommends the "local anti-corrosion treatment" solution (with the highest Q value) and predicts the following after implementation:

[0175] The reliability index will increase from 0.78 to 0.92;

[0176] The maintenance cost is 350,000 yuan (normalized value is 0.25);

[0177] Life expectancy in the region will increase by 7.5 years;

[0178] The comprehensive objective function value will increase from 0.52 to 0.83;

[0179] The maintenance decision visualization interface directly displays the above analysis results, including fatigue status heat map, trend prediction curve and comparison chart of various maintenance plans, to support engineers in making final decisions.

[0180] In actual applications, the predictive maintenance system updates data after each maintenance and continuously adjusts and optimizes its decision-making strategy through reinforcement learning, forming a closed-loop adaptive maintenance management mechanism, effectively improving the fatigue resistance and safety level of the bridge structure, while reducing the maintenance cost throughout the life cycle.

[0181] This implementation method integrates digital twin technology, knowledge graphs and causal reasoning to provide a systematic solution for predictive maintenance of bridge structure fatigue resistance, presenting many technical advantages: through multi-source evidence fusion and causal reasoning, it improves the accuracy of fatigue mechanism identification, reduces the misjudgment rate, and lays a solid foundation for maintenance decision-making; with the help of digital twins and adaptive maintenance decision-making systems, it can achieve accurate prediction of fatigue damage trends, generate optimal maintenance plans, and promote maintenance decisions from experience-driven to data-driven, reducing human judgment bias; using the improved DS evidence theory, it effectively eliminates conflicts between multi-source heterogeneous evidence, ensures the reliability of fusion results, and makes fatigue analysis more comprehensive and objective; establishes a fatigue knowledge graph, and structuredly stores expert experience and historical cases to achieve knowledge reuse and accumulation, and continuously empower fatigue analysis; adopts a predictive maintenance model, which reduces maintenance costs, extends the service life of bridges, and improves social and economic benefits compared to traditional periodic maintenance and fault repair; uses a causal reasoning engine to reveal the causal relationship and development path of fatigue damage, provide explainable analysis results, and enhance the credibility and transparency of maintenance decisions.

[0182] In one embodiment of the present invention, an example of the aforementioned predictive maintenance method for fatigue resistance of bridge structures based on digital twins is provided;

[0183] Application Scenario Description: This implementation method is applied to the predictive maintenance of a large sea-crossing cable-stayed bridge. The bridge is 1,680 meters long with a main span of 580 meters. It has been in operation for 12 years and sees an average daily traffic volume of approximately 45,000 vehicles, of which heavy vehicles account for approximately 18%. Located in a coastal environment, the bridge faces severe challenges from marine corrosion and typhoon loads, particularly fatigue damage risks in the bearing areas and cable anchorage areas.

[0184] Bridge management units face the following challenges:

[0185] Traditional regular inspections are difficult to detect early signs of fatigue damage;

[0186] Maintenance decisions mainly rely on engineers' experience and judgment, lacking reliable quantitative basis;

[0187] There are conflicts among multiple sources of information (test data, inspection records, historical cases, etc.), making it difficult to integrate;

[0188] The inability to accurately predict the development trend of fatigue damage leads to blind selection of maintenance timing and solutions;

[0189] Based on the above challenges, the management unit decided to apply the digital twin-based predictive maintenance method for bridge structure fatigue resistance proposed in this implementation method to carry out intelligent maintenance and management of the bridge.

[0190] The specific steps of the implementation process are as follows:

[0191] First, a detailed three-dimensional model of the bridge was constructed based on the design drawings and measured geometric data. Using finite element analysis software, the model was divided into approximately 87,500 elements and 98,600 nodes, establishing a mechanical model of the bridge structure. Subsequently, a multimodal sensor network was deployed at key locations on the bridge, as shown in Table 2:

[0192] Table 2: Layout of sensor network for cable-stayed bridges:

[0193]

[0194] Through edge computing devices and 5G networks, real-time sensor data collection and dynamic updating of the digital twin model were achieved. Using a Bayesian parameter calibration method, the finite element model was optimized using six months of monitoring data, resulting in a dynamic model with a deviation of less than 5% from the actual structural response.

[0195] To build a knowledge graph for bridge fatigue, the team analyzed literature and conducted expert interviews, defining 328 core concepts and 563 relationships that cover key knowledge areas related to bridge fatigue damage. Table 3 shows some of the core entities and relationships.

[0196] Table 3: Some core entities and relationships of fatigue knowledge graph:

[0197]

[0198] Using natural language processing techniques, we extracted structured knowledge from 287 research papers and 152 maintenance reports, generating approximately 25,600 knowledge triples. Furthermore, by constructing knowledge inference rules, we achieved the automatic derivation of implicit knowledge, enriching the knowledge graph.

[0199] For fatigue assessment of the cable anchorage area, the system collects information from different evidence sources. The multi-source evidence and its basic probability distribution in a fatigue analysis are shown in Table 4:

[0200] Table 4: Multi-source evidence fusion example data:

[0201]

[0202] By calculating the evidence distance matrix, the system identified obvious conflicts between the evidence sources, especially the conflict degree between the sensor data and the historical case matching, which was as high as 0.43. Using the evidence distance to adjust the evidence weight, the weights of each evidence source were calculated as follows: 、 、 、 .

[0203] The fused results show that the credibility of “joint stress concentration” is 0.42, the credibility of “anchor bolt loosening” is 0.28, and the credibility of “anchor plate weld fatigue” is 0.21, and the uncertainty is reduced to 0.09.

[0204] Based on the knowledge graph, a Bayesian causal network of the fatigue mechanism of the anchorage area of the inclined cable was constructed, including key variables such as environmental factors, load characteristics, manufacturing quality, material properties, stress state, and fatigue performance.

[0205] Using monitoring data and inspection results, the system calculated the conditional probability table between variables. When high-frequency vibration and small displacement changes were observed in the anchoring area, the system performed a posteriori probability inference. The inference results are shown in Table 5:

[0206] Table 5: Results of posterior probability calculation by the causal inference engine:

[0207]

[0208] Through counterfactual intervention analysis, the system evaluated the effects of different maintenance measures. For example, it calculated that adding damper intervention could reduce the probability of the main fatigue path from 0.62 to 0.28, optimizing the connection structure could reduce the probability to 0.35, and replacing the anchor device could reduce the probability to 0.08.

[0209] Based on the fused evidence and causal reasoning results, a LSTM fatigue damage prediction model was systematically constructed to predict the fatigue damage in the anchorage area of the inclined cable.

[0210] At the same time, the system compared and evaluated various maintenance strategies based on the reinforcement learning algorithm. The evaluation results are shown in Table 6:

[0211] Table 6: Maintenance strategy evaluation results:

[0212]

[0213] Based on the objective function evaluation, the system recommends the "additional damper" maintenance strategy, which strikes the optimal balance between cost and benefit. The system displays the recommended solution and its expected results through a visual interface, supporting managers in making the final decision.

[0214] Technical effectiveness verification: After 18 months of operation on the sea-crossing cable-stayed bridge, the predictive maintenance system significantly improved the fatigue resistance of the bridge structure, mainly reflected in the following two core technical effects:

[0215] Significantly improved fatigue mechanism identification accuracy: This implementation significantly improves the accuracy of fatigue mechanism identification through multi-source evidence fusion and causal reasoning technology. The identification results of 28 known fatigue damage cases were compared before and after the system was put into use, as shown in Table 7:

[0216] Table 7: Comparison of fatigue mechanism recognition accuracy:

[0217]

[0218] Data shows that compared with traditional methods based on single-evidence analysis, this method's fatigue mechanism identification accuracy increased from 57.1% to 85.7%, an average improvement of 28.6%. In particular, for fatigue in the more complex support and anchorage areas, the accuracy improvement exceeded 28%.

[0219] Maintenance costs are significantly reduced: The predictive maintenance model adopted by this implementation significantly reduces the maintenance costs of bridges compared to traditional planned maintenance and reactive maintenance models. The maintenance costs before and after the implementation of the system for 18 months are compared as shown in Table 8:

[0220] Table 8: Comparative analysis of maintenance costs (unit: 10,000 yuan):

[0221]

[0222] Table 8 shows that while monitoring system O&M costs increased by 82.2%, other maintenance costs decreased significantly, particularly emergency repair costs, which decreased by 63.9%. Overall, over the 18-month period, total bridge maintenance costs decreased from 9.16 million yuan to 6.25 million yuan, a decrease of 2.91 million yuan, or 31.8%.

[0223] In addition, the system has brought other quantitative benefits, including:

[0224] The average remaining life of key bridge components is expected to be extended by 25%;

[0225] Maintenance decision response time was shortened from an average of 3.5 days to 0.8 days, an improvement of 76.7%;

[0226] The risk of bridge safety accidents is reduced by approximately 45%;

[0227] The total life cycle cost of the bridge is expected to be reduced by about 28%.

[0228] In summary, this implementation method has achieved remarkable technical effects in practical applications. It not only improves the accuracy of fatigue mechanism identification, but also significantly reduces maintenance costs, extends the service life of the bridge, and provides a strong guarantee for the safe operation of the bridge structure.

[0229] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make more forms of equivalent embodiments based on the inspiration of this embodiment, all of which are protected by this embodiment.

Claims

1. A predictive maintenance method for fatigue performance of bridge structures based on digital twins, characterized by: The following steps are involved: Build a digital twin model of the bridge structure and deploy a multimodal sensor network to achieve real-time data interaction between the physical structure and the virtual model; Establish a knowledge map of bridge fatigue mechanisms, transform expert experience and historical cases into structured knowledge, and form a fatigue mechanism knowledge base; Based on DS evidence theory, it realizes the fusion of multi-source heterogeneous evidence, eliminates the conflict between evidence from different sources, and improves the accuracy of fatigue mechanism identification; Build a causal inference engine based on Bayesian networks to identify the causal path of fatigue damage development, including: Construct an initial Bayesian network structure based on the causal relationship in the knowledge graph; Combine monitoring data and expert knowledge to learn the conditional probability table of the Bayesian network; Construct a posterior probability inference algorithm based on MCMC method; Create a counterfactual intervention analysis module to calculate the intervention effects of different maintenance measures; Create a fatigue mechanism path identification algorithm to find the most likely fatigue development path; Implement an adaptive predictive maintenance decision system to generate the optimal maintenance strategy based on fatigue state prediction results.

2. The predictive maintenance method for fatigue resistance of bridge structures based on digital twins according to claim 1 is characterized in that: The construction of the bridge structure digital twin model includes: Use parametric modeling technology to construct a three-dimensional finite element model of the bridge; Deploy multimodal sensing networks at key bridge locations, including strain sensors, accelerometers, and displacement sensors; Establish a mapping relationship between sensor data and digital models, and use the Bayesian calibration method to optimize the parameters of the finite element model; Build a two-way data channel based on edge computing technology to achieve real-time interaction between physical space and virtual space.

3. The predictive maintenance method for fatigue resistance of bridge structures based on digital twins according to claim 1 is characterized in that: The establishment of the bridge fatigue mechanism knowledge graph includes: Define the core concepts, attributes and relationships in the field of bridge fatigue based on ontology method; Extract structured knowledge from fatigue research literature, maintenance reports, and expert knowledge through natural language processing technology; Construct knowledge reasoning rules to realize automatic reasoning and expansion of knowledge; Establish a case library and classify and store historical fatigue cases according to failure modes, causes, and performance characteristics.

4. The method for predictive maintenance of bridge structure fatigue resistance based on digital twin according to claim 1 is characterized in that: The calculation of the evidence conflict coefficient in the fusion of multi-source heterogeneous evidence based on the DS evidence theory is obtained by summing the products of the basic probability assignments between all conflicting evidence sources, where mutual conflict refers to the situation where the intersection of propositions supported by different evidence sources is an empty set; the larger the coefficient value, the more serious the conflict between different evidence sources, and the need for corresponding evidence adjustment.

5. The predictive maintenance method for fatigue resistance of bridge structures based on digital twins according to claim 1 is characterized in that: The evidence weight adjustment in the fusion of multi-source heterogeneous evidence based on DS evidence theory is achieved by calculating the average distance between each evidence source and other evidence sources, and converting the distance value into a weight value, so that the evidence with smaller difference from other evidence sources obtains a higher weight.

6. The predictive maintenance method for fatigue resistance of bridge structures based on digital twins according to claim 1 is characterized in that: The evidence fusion in the multi-source heterogeneous evidence fusion based on the DS evidence theory is achieved by calculating the sum of the probability products of different evidence combinations supporting the same proposition and performing normalization processing to ensure that the final probability distribution meets the constraint condition that the sum of the probabilities is 1.

7. The predictive maintenance method for fatigue resistance of bridge structures based on digital twins according to claim 1 is characterized in that: The intervention effect calculation of the counterfactual intervention analysis module is to evaluate the causal impact of the intervention measures on the outcome variables by taking a weighted average of the conditional probabilities under all possible covariate values, where the weights are the marginal distribution probabilities of the covariates.

8. The predictive maintenance method for fatigue resistance of bridge structures based on digital twins according to claim 1 is characterized in that: The adaptive predictive maintenance decision system includes: Establish a fatigue damage prediction model based on LSTM neural network; Define the maintenance decision objective function that comprehensively considers safety and reliability, maintenance cost and structural life; Build a maintenance strategy optimization algorithm based on reinforcement learning; Create a maintenance effectiveness evaluation model; Establish a visual interface for maintenance decision-making.

9. A bridge structure fatigue performance predictive maintenance system based on digital twins, which implements the bridge structure fatigue performance predictive maintenance method based on digital twins according to any one of claims 1 to 8, characterized in that: include: Digital twin model building module for digital modeling of bridge structures and deployment of multimodal sensor networks; Fatigue knowledge graph construction module, which is used to integrate expert experience and historical cases to form a structured fatigue mechanism knowledge base; Multi-source evidence fusion module, used to resolve conflicts and fuse fatigue evidence from different sources based on DS evidence theory; A causal reasoning engine module, which applies Bayesian networks and causal analysis techniques to identify potential fatigue mechanism pathways; An adaptive predictive maintenance decision module is used to generate the optimal maintenance strategy through multi-objective optimization techniques.

Images